Linux Audio

Check our new training course

Loading...
v4.6
 
   1/*
   2 * mm/mmap.c
   3 *
   4 * Written by obz.
   5 *
   6 * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
   7 */
   8
   9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11#include <linux/kernel.h>
  12#include <linux/slab.h>
  13#include <linux/backing-dev.h>
  14#include <linux/mm.h>
  15#include <linux/vmacache.h>
  16#include <linux/shm.h>
  17#include <linux/mman.h>
  18#include <linux/pagemap.h>
  19#include <linux/swap.h>
  20#include <linux/syscalls.h>
  21#include <linux/capability.h>
  22#include <linux/init.h>
  23#include <linux/file.h>
  24#include <linux/fs.h>
  25#include <linux/personality.h>
  26#include <linux/security.h>
  27#include <linux/hugetlb.h>
 
  28#include <linux/profile.h>
  29#include <linux/export.h>
  30#include <linux/mount.h>
  31#include <linux/mempolicy.h>
  32#include <linux/rmap.h>
  33#include <linux/mmu_notifier.h>
  34#include <linux/mmdebug.h>
  35#include <linux/perf_event.h>
  36#include <linux/audit.h>
  37#include <linux/khugepaged.h>
  38#include <linux/uprobes.h>
  39#include <linux/rbtree_augmented.h>
  40#include <linux/notifier.h>
  41#include <linux/memory.h>
  42#include <linux/printk.h>
  43#include <linux/userfaultfd_k.h>
  44#include <linux/moduleparam.h>
  45#include <linux/pkeys.h>
 
 
  46
  47#include <asm/uaccess.h>
  48#include <asm/cacheflush.h>
  49#include <asm/tlb.h>
  50#include <asm/mmu_context.h>
  51
 
 
 
  52#include "internal.h"
  53
  54#ifndef arch_mmap_check
  55#define arch_mmap_check(addr, len, flags)	(0)
  56#endif
  57
  58#ifndef arch_rebalance_pgtables
  59#define arch_rebalance_pgtables(addr, len)		(addr)
  60#endif
  61
  62#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
  63const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
  64const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
  65int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
  66#endif
  67#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
  68const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
  69const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
  70int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
  71#endif
  72
  73static bool ignore_rlimit_data = true;
  74core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
  75
  76static 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 */
  95pgprot_t protection_map[16] = {
  96	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
  97	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
  98};
  99
 
 
 
 
 
 
 
 100pgprot_t vm_get_page_prot(unsigned long vm_flags)
 101{
 102	return __pgprot(pgprot_val(protection_map[vm_flags &
 103				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
 104			pgprot_val(arch_vm_get_page_prot(vm_flags)));
 
 
 105}
 106EXPORT_SYMBOL(vm_get_page_prot);
 107
 108static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
 109{
 110	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
 111}
 112
 113/* Update vma->vm_page_prot to reflect vma->vm_flags. */
 114void vma_set_page_prot(struct vm_area_struct *vma)
 115{
 116	unsigned long vm_flags = vma->vm_flags;
 
 117
 118	vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
 119	if (vma_wants_writenotify(vma)) {
 120		vm_flags &= ~VM_SHARED;
 121		vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
 122						     vm_flags);
 123	}
 
 
 124}
 125
 126/*
 127 * Requires inode->i_mapping->i_mmap_rwsem
 128 */
 129static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 130		struct file *file, struct address_space *mapping)
 131{
 132	if (vma->vm_flags & VM_DENYWRITE)
 133		atomic_inc(&file_inode(file)->i_writecount);
 134	if (vma->vm_flags & VM_SHARED)
 135		mapping_unmap_writable(mapping);
 136
 137	flush_dcache_mmap_lock(mapping);
 138	vma_interval_tree_remove(vma, &mapping->i_mmap);
 139	flush_dcache_mmap_unlock(mapping);
 140}
 141
 142/*
 143 * Unlink a file-based vm structure from its interval tree, to hide
 144 * vma from rmap and vmtruncate before freeing its page tables.
 145 */
 146void unlink_file_vma(struct vm_area_struct *vma)
 147{
 148	struct file *file = vma->vm_file;
 149
 150	if (file) {
 151		struct address_space *mapping = file->f_mapping;
 152		i_mmap_lock_write(mapping);
 153		__remove_shared_vm_struct(vma, file, mapping);
 154		i_mmap_unlock_write(mapping);
 155	}
 156}
 157
 158/*
 159 * Close a vm structure and free it, returning the next.
 160 */
 161static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
 162{
 163	struct vm_area_struct *next = vma->vm_next;
 164
 165	might_sleep();
 166	if (vma->vm_ops && vma->vm_ops->close)
 167		vma->vm_ops->close(vma);
 168	if (vma->vm_file)
 169		fput(vma->vm_file);
 170	mpol_put(vma_policy(vma));
 171	kmem_cache_free(vm_area_cachep, vma);
 172	return next;
 173}
 174
 175static unsigned long do_brk(unsigned long addr, unsigned long len);
 176
 177SYSCALL_DEFINE1(brk, unsigned long, brk)
 178{
 179	unsigned long retval;
 180	unsigned long newbrk, oldbrk;
 181	struct mm_struct *mm = current->mm;
 
 182	unsigned long min_brk;
 183	bool populate;
 
 
 184
 185	down_write(&mm->mmap_sem);
 
 
 
 186
 187#ifdef CONFIG_COMPAT_BRK
 188	/*
 189	 * CONFIG_COMPAT_BRK can still be overridden by setting
 190	 * randomize_va_space to 2, which will still cause mm->start_brk
 191	 * to be arbitrarily shifted
 192	 */
 193	if (current->brk_randomized)
 194		min_brk = mm->start_brk;
 195	else
 196		min_brk = mm->end_data;
 197#else
 198	min_brk = mm->start_brk;
 199#endif
 200	if (brk < min_brk)
 201		goto out;
 202
 203	/*
 204	 * Check against rlimit here. If this check is done later after the test
 205	 * of oldbrk with newbrk then it can escape the test and let the data
 206	 * segment grow beyond its set limit the in case where the limit is
 207	 * not page aligned -Ram Gupta
 208	 */
 209	if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
 210			      mm->end_data, mm->start_data))
 211		goto out;
 212
 213	newbrk = PAGE_ALIGN(brk);
 214	oldbrk = PAGE_ALIGN(mm->brk);
 215	if (oldbrk == newbrk)
 216		goto set_brk;
 
 
 217
 218	/* Always allow shrinking brk. */
 
 
 
 219	if (brk <= mm->brk) {
 220		if (!do_munmap(mm, newbrk, oldbrk-newbrk))
 221			goto set_brk;
 222		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 223	}
 224
 225	/* Check against existing mmap mappings. */
 226	if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
 
 227		goto out;
 228
 229	/* Ok, looks good - let it rip. */
 230	if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
 231		goto out;
 232
 233set_brk:
 234	mm->brk = brk;
 
 
 235	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
 236	up_write(&mm->mmap_sem);
 
 
 
 
 237	if (populate)
 238		mm_populate(oldbrk, newbrk - oldbrk);
 239	return brk;
 240
 241out:
 242	retval = mm->brk;
 243	up_write(&mm->mmap_sem);
 244	return retval;
 245}
 246
 247static long vma_compute_subtree_gap(struct vm_area_struct *vma)
 248{
 249	unsigned long max, subtree_gap;
 250	max = vma->vm_start;
 251	if (vma->vm_prev)
 252		max -= vma->vm_prev->vm_end;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 253	if (vma->vm_rb.rb_left) {
 254		subtree_gap = rb_entry(vma->vm_rb.rb_left,
 255				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 256		if (subtree_gap > max)
 257			max = subtree_gap;
 258	}
 259	if (vma->vm_rb.rb_right) {
 260		subtree_gap = rb_entry(vma->vm_rb.rb_right,
 261				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 262		if (subtree_gap > max)
 263			max = subtree_gap;
 264	}
 265	return max;
 266}
 267
 268#ifdef CONFIG_DEBUG_VM_RB
 269static int browse_rb(struct mm_struct *mm)
 270{
 271	struct rb_root *root = &mm->mm_rb;
 272	int i = 0, j, bug = 0;
 273	struct rb_node *nd, *pn = NULL;
 274	unsigned long prev = 0, pend = 0;
 275
 276	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 277		struct vm_area_struct *vma;
 278		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 279		if (vma->vm_start < prev) {
 280			pr_emerg("vm_start %lx < prev %lx\n",
 281				  vma->vm_start, prev);
 282			bug = 1;
 283		}
 284		if (vma->vm_start < pend) {
 285			pr_emerg("vm_start %lx < pend %lx\n",
 286				  vma->vm_start, pend);
 287			bug = 1;
 288		}
 289		if (vma->vm_start > vma->vm_end) {
 290			pr_emerg("vm_start %lx > vm_end %lx\n",
 291				  vma->vm_start, vma->vm_end);
 292			bug = 1;
 293		}
 294		spin_lock(&mm->page_table_lock);
 295		if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
 296			pr_emerg("free gap %lx, correct %lx\n",
 297			       vma->rb_subtree_gap,
 298			       vma_compute_subtree_gap(vma));
 299			bug = 1;
 300		}
 301		spin_unlock(&mm->page_table_lock);
 302		i++;
 303		pn = nd;
 304		prev = vma->vm_start;
 305		pend = vma->vm_end;
 306	}
 307	j = 0;
 308	for (nd = pn; nd; nd = rb_prev(nd))
 309		j++;
 310	if (i != j) {
 311		pr_emerg("backwards %d, forwards %d\n", j, i);
 312		bug = 1;
 313	}
 314	return bug ? -1 : i;
 315}
 316
 317static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
 318{
 319	struct rb_node *nd;
 320
 321	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 322		struct vm_area_struct *vma;
 323		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 324		VM_BUG_ON_VMA(vma != ignore &&
 325			vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
 326			vma);
 327	}
 328}
 329
 330static void validate_mm(struct mm_struct *mm)
 331{
 332	int bug = 0;
 333	int i = 0;
 334	unsigned long highest_address = 0;
 335	struct vm_area_struct *vma = mm->mmap;
 336
 337	while (vma) {
 338		struct anon_vma *anon_vma = vma->anon_vma;
 339		struct anon_vma_chain *avc;
 340
 341		if (anon_vma) {
 342			anon_vma_lock_read(anon_vma);
 343			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 344				anon_vma_interval_tree_verify(avc);
 345			anon_vma_unlock_read(anon_vma);
 346		}
 347
 348		highest_address = vma->vm_end;
 349		vma = vma->vm_next;
 350		i++;
 351	}
 352	if (i != mm->map_count) {
 353		pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
 354		bug = 1;
 355	}
 356	if (highest_address != mm->highest_vm_end) {
 357		pr_emerg("mm->highest_vm_end %lx, found %lx\n",
 358			  mm->highest_vm_end, highest_address);
 359		bug = 1;
 360	}
 361	i = browse_rb(mm);
 362	if (i != mm->map_count) {
 363		if (i != -1)
 364			pr_emerg("map_count %d rb %d\n", mm->map_count, i);
 365		bug = 1;
 366	}
 367	VM_BUG_ON_MM(bug, mm);
 368}
 369#else
 370#define validate_mm_rb(root, ignore) do { } while (0)
 371#define validate_mm(mm) do { } while (0)
 372#endif
 373
 374RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
 375		     unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
 
 376
 377/*
 378 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
 379 * vma->vm_prev->vm_end values changed, without modifying the vma's position
 380 * in the rbtree.
 381 */
 382static void vma_gap_update(struct vm_area_struct *vma)
 383{
 384	/*
 385	 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
 386	 * function that does exacltly what we want.
 387	 */
 388	vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
 389}
 390
 391static inline void vma_rb_insert(struct vm_area_struct *vma,
 392				 struct rb_root *root)
 393{
 394	/* All rb_subtree_gap values must be consistent prior to insertion */
 395	validate_mm_rb(root, NULL);
 396
 397	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 398}
 399
 400static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
 401{
 402	/*
 403	 * All rb_subtree_gap values must be consistent prior to erase,
 404	 * with the possible exception of the vma being erased.
 405	 */
 406	validate_mm_rb(root, vma);
 407
 408	/*
 409	 * Note rb_erase_augmented is a fairly large inline function,
 410	 * so make sure we instantiate it only once with our desired
 411	 * augmented rbtree callbacks.
 412	 */
 413	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 414}
 415
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 416/*
 417 * vma has some anon_vma assigned, and is already inserted on that
 418 * anon_vma's interval trees.
 419 *
 420 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
 421 * vma must be removed from the anon_vma's interval trees using
 422 * anon_vma_interval_tree_pre_update_vma().
 423 *
 424 * After the update, the vma will be reinserted using
 425 * anon_vma_interval_tree_post_update_vma().
 426 *
 427 * The entire update must be protected by exclusive mmap_sem and by
 428 * the root anon_vma's mutex.
 429 */
 430static inline void
 431anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
 432{
 433	struct anon_vma_chain *avc;
 434
 435	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 436		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
 437}
 438
 439static inline void
 440anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
 441{
 442	struct anon_vma_chain *avc;
 443
 444	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 445		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
 446}
 447
 448static int find_vma_links(struct mm_struct *mm, unsigned long addr,
 449		unsigned long end, struct vm_area_struct **pprev,
 450		struct rb_node ***rb_link, struct rb_node **rb_parent)
 451{
 452	struct rb_node **__rb_link, *__rb_parent, *rb_prev;
 453
 454	__rb_link = &mm->mm_rb.rb_node;
 455	rb_prev = __rb_parent = NULL;
 456
 457	while (*__rb_link) {
 458		struct vm_area_struct *vma_tmp;
 459
 460		__rb_parent = *__rb_link;
 461		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
 462
 463		if (vma_tmp->vm_end > addr) {
 464			/* Fail if an existing vma overlaps the area */
 465			if (vma_tmp->vm_start < end)
 466				return -ENOMEM;
 467			__rb_link = &__rb_parent->rb_left;
 468		} else {
 469			rb_prev = __rb_parent;
 470			__rb_link = &__rb_parent->rb_right;
 471		}
 472	}
 473
 474	*pprev = NULL;
 475	if (rb_prev)
 476		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 477	*rb_link = __rb_link;
 478	*rb_parent = __rb_parent;
 479	return 0;
 480}
 481
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 482static unsigned long count_vma_pages_range(struct mm_struct *mm,
 483		unsigned long addr, unsigned long end)
 484{
 485	unsigned long nr_pages = 0;
 486	struct vm_area_struct *vma;
 487
 488	/* Find first overlaping mapping */
 489	vma = find_vma_intersection(mm, addr, end);
 490	if (!vma)
 491		return 0;
 492
 493	nr_pages = (min(end, vma->vm_end) -
 494		max(addr, vma->vm_start)) >> PAGE_SHIFT;
 495
 496	/* Iterate over the rest of the overlaps */
 497	for (vma = vma->vm_next; vma; vma = vma->vm_next) {
 498		unsigned long overlap_len;
 499
 500		if (vma->vm_start > end)
 501			break;
 502
 503		overlap_len = min(end, vma->vm_end) - vma->vm_start;
 504		nr_pages += overlap_len >> PAGE_SHIFT;
 505	}
 506
 507	return nr_pages;
 508}
 509
 510void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 511		struct rb_node **rb_link, struct rb_node *rb_parent)
 512{
 513	/* Update tracking information for the gap following the new vma. */
 514	if (vma->vm_next)
 515		vma_gap_update(vma->vm_next);
 516	else
 517		mm->highest_vm_end = vma->vm_end;
 518
 519	/*
 520	 * vma->vm_prev wasn't known when we followed the rbtree to find the
 521	 * correct insertion point for that vma. As a result, we could not
 522	 * update the vma vm_rb parents rb_subtree_gap values on the way down.
 523	 * So, we first insert the vma with a zero rb_subtree_gap value
 524	 * (to be consistent with what we did on the way down), and then
 525	 * immediately update the gap to the correct value. Finally we
 526	 * rebalance the rbtree after all augmented values have been set.
 527	 */
 528	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 529	vma->rb_subtree_gap = 0;
 530	vma_gap_update(vma);
 531	vma_rb_insert(vma, &mm->mm_rb);
 532}
 533
 534static void __vma_link_file(struct vm_area_struct *vma)
 535{
 536	struct file *file;
 537
 538	file = vma->vm_file;
 539	if (file) {
 540		struct address_space *mapping = file->f_mapping;
 541
 542		if (vma->vm_flags & VM_DENYWRITE)
 543			atomic_dec(&file_inode(file)->i_writecount);
 544		if (vma->vm_flags & VM_SHARED)
 545			atomic_inc(&mapping->i_mmap_writable);
 546
 547		flush_dcache_mmap_lock(mapping);
 548		vma_interval_tree_insert(vma, &mapping->i_mmap);
 549		flush_dcache_mmap_unlock(mapping);
 550	}
 551}
 552
 553static void
 554__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 555	struct vm_area_struct *prev, struct rb_node **rb_link,
 556	struct rb_node *rb_parent)
 557{
 558	__vma_link_list(mm, vma, prev, rb_parent);
 559	__vma_link_rb(mm, vma, rb_link, rb_parent);
 560}
 561
 562static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 563			struct vm_area_struct *prev, struct rb_node **rb_link,
 564			struct rb_node *rb_parent)
 565{
 566	struct address_space *mapping = NULL;
 567
 568	if (vma->vm_file) {
 569		mapping = vma->vm_file->f_mapping;
 570		i_mmap_lock_write(mapping);
 571	}
 572
 573	__vma_link(mm, vma, prev, rb_link, rb_parent);
 574	__vma_link_file(vma);
 575
 576	if (mapping)
 577		i_mmap_unlock_write(mapping);
 578
 579	mm->map_count++;
 580	validate_mm(mm);
 581}
 582
 583/*
 584 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
 585 * mm's list and rbtree.  It has already been inserted into the interval tree.
 586 */
 587static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 588{
 589	struct vm_area_struct *prev;
 590	struct rb_node **rb_link, *rb_parent;
 591
 592	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
 593			   &prev, &rb_link, &rb_parent))
 594		BUG();
 595	__vma_link(mm, vma, prev, rb_link, rb_parent);
 596	mm->map_count++;
 597}
 598
 599static inline void
 600__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
 601		struct vm_area_struct *prev)
 602{
 603	struct vm_area_struct *next;
 604
 605	vma_rb_erase(vma, &mm->mm_rb);
 606	prev->vm_next = next = vma->vm_next;
 607	if (next)
 608		next->vm_prev = prev;
 609
 610	/* Kill the cache */
 611	vmacache_invalidate(mm);
 612}
 613
 614/*
 615 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
 616 * is already present in an i_mmap tree without adjusting the tree.
 617 * The following helper function should be used when such adjustments
 618 * are necessary.  The "insert" vma (if any) is to be inserted
 619 * before we drop the necessary locks.
 620 */
 621int vma_adjust(struct vm_area_struct *vma, unsigned long start,
 622	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
 
 623{
 624	struct mm_struct *mm = vma->vm_mm;
 625	struct vm_area_struct *next = vma->vm_next;
 626	struct vm_area_struct *importer = NULL;
 627	struct address_space *mapping = NULL;
 628	struct rb_root *root = NULL;
 629	struct anon_vma *anon_vma = NULL;
 630	struct file *file = vma->vm_file;
 631	bool start_changed = false, end_changed = false;
 632	long adjust_next = 0;
 633	int remove_next = 0;
 634
 635	if (next && !insert) {
 636		struct vm_area_struct *exporter = NULL;
 637
 638		if (end >= next->vm_end) {
 639			/*
 640			 * vma expands, overlapping all the next, and
 641			 * perhaps the one after too (mprotect case 6).
 
 
 642			 */
 643again:			remove_next = 1 + (end > next->vm_end);
 644			end = next->vm_end;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 645			exporter = next;
 646			importer = vma;
 
 
 
 
 
 
 
 
 647		} else if (end > next->vm_start) {
 648			/*
 649			 * vma expands, overlapping part of the next:
 650			 * mprotect case 5 shifting the boundary up.
 651			 */
 652			adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
 653			exporter = next;
 654			importer = vma;
 
 655		} else if (end < vma->vm_end) {
 656			/*
 657			 * vma shrinks, and !insert tells it's not
 658			 * split_vma inserting another: so it must be
 659			 * mprotect case 4 shifting the boundary down.
 660			 */
 661			adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
 662			exporter = vma;
 663			importer = next;
 
 664		}
 665
 666		/*
 667		 * Easily overlooked: when mprotect shifts the boundary,
 668		 * make sure the expanding vma has anon_vma set if the
 669		 * shrinking vma had, to cover any anon pages imported.
 670		 */
 671		if (exporter && exporter->anon_vma && !importer->anon_vma) {
 672			int error;
 673
 674			importer->anon_vma = exporter->anon_vma;
 675			error = anon_vma_clone(importer, exporter);
 676			if (error)
 677				return error;
 678		}
 679	}
 
 
 680
 681	if (file) {
 682		mapping = file->f_mapping;
 683		root = &mapping->i_mmap;
 684		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
 685
 686		if (adjust_next)
 687			uprobe_munmap(next, next->vm_start, next->vm_end);
 688
 689		i_mmap_lock_write(mapping);
 690		if (insert) {
 691			/*
 692			 * Put into interval tree now, so instantiated pages
 693			 * are visible to arm/parisc __flush_dcache_page
 694			 * throughout; but we cannot insert into address
 695			 * space until vma start or end is updated.
 696			 */
 697			__vma_link_file(insert);
 698		}
 699	}
 700
 701	vma_adjust_trans_huge(vma, start, end, adjust_next);
 702
 703	anon_vma = vma->anon_vma;
 704	if (!anon_vma && adjust_next)
 705		anon_vma = next->anon_vma;
 706	if (anon_vma) {
 707		VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
 708			  anon_vma != next->anon_vma, next);
 709		anon_vma_lock_write(anon_vma);
 710		anon_vma_interval_tree_pre_update_vma(vma);
 711		if (adjust_next)
 712			anon_vma_interval_tree_pre_update_vma(next);
 713	}
 714
 715	if (root) {
 716		flush_dcache_mmap_lock(mapping);
 717		vma_interval_tree_remove(vma, root);
 718		if (adjust_next)
 719			vma_interval_tree_remove(next, root);
 720	}
 721
 722	if (start != vma->vm_start) {
 723		vma->vm_start = start;
 724		start_changed = true;
 725	}
 726	if (end != vma->vm_end) {
 727		vma->vm_end = end;
 728		end_changed = true;
 729	}
 730	vma->vm_pgoff = pgoff;
 731	if (adjust_next) {
 732		next->vm_start += adjust_next << PAGE_SHIFT;
 733		next->vm_pgoff += adjust_next;
 734	}
 735
 736	if (root) {
 737		if (adjust_next)
 738			vma_interval_tree_insert(next, root);
 739		vma_interval_tree_insert(vma, root);
 740		flush_dcache_mmap_unlock(mapping);
 741	}
 742
 743	if (remove_next) {
 744		/*
 745		 * vma_merge has merged next into vma, and needs
 746		 * us to remove next before dropping the locks.
 747		 */
 748		__vma_unlink(mm, next, vma);
 
 
 
 
 
 
 
 
 
 
 
 
 749		if (file)
 750			__remove_shared_vm_struct(next, file, mapping);
 751	} else if (insert) {
 752		/*
 753		 * split_vma has split insert from vma, and needs
 754		 * us to insert it before dropping the locks
 755		 * (it may either follow vma or precede it).
 756		 */
 757		__insert_vm_struct(mm, insert);
 758	} else {
 759		if (start_changed)
 760			vma_gap_update(vma);
 761		if (end_changed) {
 762			if (!next)
 763				mm->highest_vm_end = end;
 764			else if (!adjust_next)
 765				vma_gap_update(next);
 766		}
 767	}
 768
 769	if (anon_vma) {
 770		anon_vma_interval_tree_post_update_vma(vma);
 771		if (adjust_next)
 772			anon_vma_interval_tree_post_update_vma(next);
 773		anon_vma_unlock_write(anon_vma);
 774	}
 775	if (mapping)
 776		i_mmap_unlock_write(mapping);
 777
 778	if (root) {
 
 779		uprobe_mmap(vma);
 780
 781		if (adjust_next)
 782			uprobe_mmap(next);
 783	}
 784
 785	if (remove_next) {
 786		if (file) {
 787			uprobe_munmap(next, next->vm_start, next->vm_end);
 788			fput(file);
 789		}
 790		if (next->anon_vma)
 791			anon_vma_merge(vma, next);
 792		mm->map_count--;
 793		mpol_put(vma_policy(next));
 794		kmem_cache_free(vm_area_cachep, next);
 795		/*
 796		 * In mprotect's case 6 (see comments on vma_merge),
 797		 * we must remove another next too. It would clutter
 798		 * up the code too much to do both in one go.
 799		 */
 800		next = vma->vm_next;
 801		if (remove_next == 2)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 802			goto again;
 
 803		else if (next)
 804			vma_gap_update(next);
 805		else
 806			mm->highest_vm_end = end;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 807	}
 808	if (insert && file)
 809		uprobe_mmap(insert);
 810
 811	validate_mm(mm);
 812
 813	return 0;
 814}
 815
 816/*
 817 * If the vma has a ->close operation then the driver probably needs to release
 818 * per-vma resources, so we don't attempt to merge those.
 819 */
 820static inline int is_mergeable_vma(struct vm_area_struct *vma,
 821				struct file *file, unsigned long vm_flags,
 822				struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 823{
 824	/*
 825	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
 826	 * match the flags but dirty bit -- the caller should mark
 827	 * merged VMA as dirty. If dirty bit won't be excluded from
 828	 * comparison, we increase pressue on the memory system forcing
 829	 * the kernel to generate new VMAs when old one could be
 830	 * extended instead.
 831	 */
 832	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
 833		return 0;
 834	if (vma->vm_file != file)
 835		return 0;
 836	if (vma->vm_ops && vma->vm_ops->close)
 837		return 0;
 838	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
 839		return 0;
 840	return 1;
 841}
 842
 843static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
 844					struct anon_vma *anon_vma2,
 845					struct vm_area_struct *vma)
 846{
 847	/*
 848	 * The list_is_singular() test is to avoid merging VMA cloned from
 849	 * parents. This can improve scalability caused by anon_vma lock.
 850	 */
 851	if ((!anon_vma1 || !anon_vma2) && (!vma ||
 852		list_is_singular(&vma->anon_vma_chain)))
 853		return 1;
 854	return anon_vma1 == anon_vma2;
 855}
 856
 857/*
 858 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 859 * in front of (at a lower virtual address and file offset than) the vma.
 860 *
 861 * We cannot merge two vmas if they have differently assigned (non-NULL)
 862 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 863 *
 864 * We don't check here for the merged mmap wrapping around the end of pagecache
 865 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
 866 * wrap, nor mmaps which cover the final page at index -1UL.
 867 */
 868static int
 869can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
 870		     struct anon_vma *anon_vma, struct file *file,
 871		     pgoff_t vm_pgoff,
 872		     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 873{
 874	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
 875	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 876		if (vma->vm_pgoff == vm_pgoff)
 877			return 1;
 878	}
 879	return 0;
 880}
 881
 882/*
 883 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 884 * beyond (at a higher virtual address and file offset than) the vma.
 885 *
 886 * We cannot merge two vmas if they have differently assigned (non-NULL)
 887 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 888 */
 889static int
 890can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
 891		    struct anon_vma *anon_vma, struct file *file,
 892		    pgoff_t vm_pgoff,
 893		    struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 894{
 895	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
 896	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 897		pgoff_t vm_pglen;
 898		vm_pglen = vma_pages(vma);
 899		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
 900			return 1;
 901	}
 902	return 0;
 903}
 904
 905/*
 906 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
 907 * whether that can be merged with its predecessor or its successor.
 908 * Or both (it neatly fills a hole).
 909 *
 910 * In most cases - when called for mmap, brk or mremap - [addr,end) is
 911 * certain not to be mapped by the time vma_merge is called; but when
 912 * called for mprotect, it is certain to be already mapped (either at
 913 * an offset within prev, or at the start of next), and the flags of
 914 * this area are about to be changed to vm_flags - and the no-change
 915 * case has already been eliminated.
 916 *
 917 * The following mprotect cases have to be considered, where AAAA is
 918 * the area passed down from mprotect_fixup, never extending beyond one
 919 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
 920 *
 921 *     AAAA             AAAA                AAAA          AAAA
 922 *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
 923 *    cannot merge    might become    might become    might become
 924 *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
 925 *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
 926 *    mremap move:                                    PPPPNNNNNNNN 8
 927 *        AAAA
 928 *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
 929 *    might become    case 1 below    case 2 below    case 3 below
 930 *
 931 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
 932 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 933 */
 934struct vm_area_struct *vma_merge(struct mm_struct *mm,
 935			struct vm_area_struct *prev, unsigned long addr,
 936			unsigned long end, unsigned long vm_flags,
 937			struct anon_vma *anon_vma, struct file *file,
 938			pgoff_t pgoff, struct mempolicy *policy,
 939			struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 940{
 941	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
 942	struct vm_area_struct *area, *next;
 943	int err;
 944
 945	/*
 946	 * We later require that vma->vm_flags == vm_flags,
 947	 * so this tests vma->vm_flags & VM_SPECIAL, too.
 948	 */
 949	if (vm_flags & VM_SPECIAL)
 950		return NULL;
 951
 952	if (prev)
 953		next = prev->vm_next;
 954	else
 955		next = mm->mmap;
 956	area = next;
 957	if (next && next->vm_end == end)		/* cases 6, 7, 8 */
 958		next = next->vm_next;
 959
 
 
 
 
 
 960	/*
 961	 * Can it merge with the predecessor?
 962	 */
 963	if (prev && prev->vm_end == addr &&
 964			mpol_equal(vma_policy(prev), policy) &&
 965			can_vma_merge_after(prev, vm_flags,
 966					    anon_vma, file, pgoff,
 967					    vm_userfaultfd_ctx)) {
 968		/*
 969		 * OK, it can.  Can we now merge in the successor as well?
 970		 */
 971		if (next && end == next->vm_start &&
 972				mpol_equal(policy, vma_policy(next)) &&
 973				can_vma_merge_before(next, vm_flags,
 974						     anon_vma, file,
 975						     pgoff+pglen,
 976						     vm_userfaultfd_ctx) &&
 977				is_mergeable_anon_vma(prev->anon_vma,
 978						      next->anon_vma, NULL)) {
 979							/* cases 1, 6 */
 980			err = vma_adjust(prev, prev->vm_start,
 981				next->vm_end, prev->vm_pgoff, NULL);
 
 982		} else					/* cases 2, 5, 7 */
 983			err = vma_adjust(prev, prev->vm_start,
 984				end, prev->vm_pgoff, NULL);
 985		if (err)
 986			return NULL;
 987		khugepaged_enter_vma_merge(prev, vm_flags);
 988		return prev;
 989	}
 990
 991	/*
 992	 * Can this new request be merged in front of next?
 993	 */
 994	if (next && end == next->vm_start &&
 995			mpol_equal(policy, vma_policy(next)) &&
 996			can_vma_merge_before(next, vm_flags,
 997					     anon_vma, file, pgoff+pglen,
 998					     vm_userfaultfd_ctx)) {
 999		if (prev && addr < prev->vm_end)	/* case 4 */
1000			err = vma_adjust(prev, prev->vm_start,
1001				addr, prev->vm_pgoff, NULL);
1002		else					/* cases 3, 8 */
1003			err = vma_adjust(area, addr, next->vm_end,
1004				next->vm_pgoff - pglen, NULL);
 
 
 
 
 
 
 
1005		if (err)
1006			return NULL;
1007		khugepaged_enter_vma_merge(area, vm_flags);
1008		return area;
1009	}
1010
1011	return NULL;
1012}
1013
1014/*
1015 * Rough compatbility check to quickly see if it's even worth looking
1016 * at sharing an anon_vma.
1017 *
1018 * They need to have the same vm_file, and the flags can only differ
1019 * in things that mprotect may change.
1020 *
1021 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1022 * we can merge the two vma's. For example, we refuse to merge a vma if
1023 * there is a vm_ops->close() function, because that indicates that the
1024 * driver is doing some kind of reference counting. But that doesn't
1025 * really matter for the anon_vma sharing case.
1026 */
1027static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1028{
1029	return a->vm_end == b->vm_start &&
1030		mpol_equal(vma_policy(a), vma_policy(b)) &&
1031		a->vm_file == b->vm_file &&
1032		!((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1033		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1034}
1035
1036/*
1037 * Do some basic sanity checking to see if we can re-use the anon_vma
1038 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1039 * the same as 'old', the other will be the new one that is trying
1040 * to share the anon_vma.
1041 *
1042 * NOTE! This runs with mm_sem held for reading, so it is possible that
1043 * the anon_vma of 'old' is concurrently in the process of being set up
1044 * by another page fault trying to merge _that_. But that's ok: if it
1045 * is being set up, that automatically means that it will be a singleton
1046 * acceptable for merging, so we can do all of this optimistically. But
1047 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1048 *
1049 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1050 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1051 * is to return an anon_vma that is "complex" due to having gone through
1052 * a fork).
1053 *
1054 * We also make sure that the two vma's are compatible (adjacent,
1055 * and with the same memory policies). That's all stable, even with just
1056 * a read lock on the mm_sem.
1057 */
1058static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1059{
1060	if (anon_vma_compatible(a, b)) {
1061		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1062
1063		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1064			return anon_vma;
1065	}
1066	return NULL;
1067}
1068
1069/*
1070 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1071 * neighbouring vmas for a suitable anon_vma, before it goes off
1072 * to allocate a new anon_vma.  It checks because a repetitive
1073 * sequence of mprotects and faults may otherwise lead to distinct
1074 * anon_vmas being allocated, preventing vma merge in subsequent
1075 * mprotect.
1076 */
1077struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1078{
1079	struct anon_vma *anon_vma;
1080	struct vm_area_struct *near;
 
 
 
 
 
 
 
 
 
 
1081
1082	near = vma->vm_next;
1083	if (!near)
1084		goto try_prev;
1085
1086	anon_vma = reusable_anon_vma(near, vma, near);
1087	if (anon_vma)
1088		return anon_vma;
1089try_prev:
1090	near = vma->vm_prev;
1091	if (!near)
1092		goto none;
1093
1094	anon_vma = reusable_anon_vma(near, near, vma);
1095	if (anon_vma)
1096		return anon_vma;
1097none:
1098	/*
 
 
1099	 * There's no absolute need to look only at touching neighbours:
1100	 * we could search further afield for "compatible" anon_vmas.
1101	 * But it would probably just be a waste of time searching,
1102	 * or lead to too many vmas hanging off the same anon_vma.
1103	 * We're trying to allow mprotect remerging later on,
1104	 * not trying to minimize memory used for anon_vmas.
1105	 */
1106	return NULL;
1107}
1108
1109/*
1110 * If a hint addr is less than mmap_min_addr change hint to be as
1111 * low as possible but still greater than mmap_min_addr
1112 */
1113static inline unsigned long round_hint_to_min(unsigned long hint)
1114{
1115	hint &= PAGE_MASK;
1116	if (((void *)hint != NULL) &&
1117	    (hint < mmap_min_addr))
1118		return PAGE_ALIGN(mmap_min_addr);
1119	return hint;
1120}
1121
1122static inline int mlock_future_check(struct mm_struct *mm,
1123				     unsigned long flags,
1124				     unsigned long len)
1125{
1126	unsigned long locked, lock_limit;
1127
1128	/*  mlock MCL_FUTURE? */
1129	if (flags & VM_LOCKED) {
1130		locked = len >> PAGE_SHIFT;
1131		locked += mm->locked_vm;
1132		lock_limit = rlimit(RLIMIT_MEMLOCK);
1133		lock_limit >>= PAGE_SHIFT;
1134		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1135			return -EAGAIN;
1136	}
1137	return 0;
1138}
1139
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1140/*
1141 * The caller must hold down_write(&current->mm->mmap_sem).
1142 */
1143unsigned long do_mmap(struct file *file, unsigned long addr,
1144			unsigned long len, unsigned long prot,
1145			unsigned long flags, vm_flags_t vm_flags,
1146			unsigned long pgoff, unsigned long *populate)
1147{
1148	struct mm_struct *mm = current->mm;
 
1149	int pkey = 0;
1150
1151	*populate = 0;
1152
1153	if (!len)
1154		return -EINVAL;
1155
1156	/*
1157	 * Does the application expect PROT_READ to imply PROT_EXEC?
1158	 *
1159	 * (the exception is when the underlying filesystem is noexec
1160	 *  mounted, in which case we dont add PROT_EXEC.)
1161	 */
1162	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1163		if (!(file && path_noexec(&file->f_path)))
1164			prot |= PROT_EXEC;
1165
 
 
 
 
1166	if (!(flags & MAP_FIXED))
1167		addr = round_hint_to_min(addr);
1168
1169	/* Careful about overflows.. */
1170	len = PAGE_ALIGN(len);
1171	if (!len)
1172		return -ENOMEM;
1173
1174	/* offset overflow? */
1175	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1176		return -EOVERFLOW;
1177
1178	/* Too many mappings? */
1179	if (mm->map_count > sysctl_max_map_count)
1180		return -ENOMEM;
1181
1182	/* Obtain the address to map to. we verify (or select) it and ensure
1183	 * that it represents a valid section of the address space.
1184	 */
1185	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1186	if (offset_in_page(addr))
1187		return addr;
1188
 
 
 
 
 
1189	if (prot == PROT_EXEC) {
1190		pkey = execute_only_pkey(mm);
1191		if (pkey < 0)
1192			pkey = 0;
1193	}
1194
1195	/* Do simple checking here so the lower-level routines won't have
1196	 * to. we assume access permissions have been handled by the open
1197	 * of the memory object, so we don't do any here.
1198	 */
1199	vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1200			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1201
1202	if (flags & MAP_LOCKED)
1203		if (!can_do_mlock())
1204			return -EPERM;
1205
1206	if (mlock_future_check(mm, vm_flags, len))
1207		return -EAGAIN;
1208
1209	if (file) {
1210		struct inode *inode = file_inode(file);
 
 
 
 
 
 
1211
1212		switch (flags & MAP_TYPE) {
1213		case MAP_SHARED:
1214			if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1215				return -EACCES;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1216
1217			/*
1218			 * Make sure we don't allow writing to an append-only
1219			 * file..
1220			 */
1221			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1222				return -EACCES;
1223
1224			/*
1225			 * Make sure there are no mandatory locks on the file.
1226			 */
1227			if (locks_verify_locked(file))
1228				return -EAGAIN;
1229
1230			vm_flags |= VM_SHARED | VM_MAYSHARE;
1231			if (!(file->f_mode & FMODE_WRITE))
1232				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1233
1234			/* fall through */
1235		case MAP_PRIVATE:
1236			if (!(file->f_mode & FMODE_READ))
1237				return -EACCES;
1238			if (path_noexec(&file->f_path)) {
1239				if (vm_flags & VM_EXEC)
1240					return -EPERM;
1241				vm_flags &= ~VM_MAYEXEC;
1242			}
1243
1244			if (!file->f_op->mmap)
1245				return -ENODEV;
1246			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1247				return -EINVAL;
1248			break;
1249
1250		default:
1251			return -EINVAL;
1252		}
1253	} else {
1254		switch (flags & MAP_TYPE) {
1255		case MAP_SHARED:
1256			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1257				return -EINVAL;
1258			/*
1259			 * Ignore pgoff.
1260			 */
1261			pgoff = 0;
1262			vm_flags |= VM_SHARED | VM_MAYSHARE;
1263			break;
1264		case MAP_PRIVATE:
1265			/*
1266			 * Set pgoff according to addr for anon_vma.
1267			 */
1268			pgoff = addr >> PAGE_SHIFT;
1269			break;
1270		default:
1271			return -EINVAL;
1272		}
1273	}
1274
1275	/*
1276	 * Set 'VM_NORESERVE' if we should not account for the
1277	 * memory use of this mapping.
1278	 */
1279	if (flags & MAP_NORESERVE) {
1280		/* We honor MAP_NORESERVE if allowed to overcommit */
1281		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1282			vm_flags |= VM_NORESERVE;
1283
1284		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1285		if (file && is_file_hugepages(file))
1286			vm_flags |= VM_NORESERVE;
1287	}
1288
1289	addr = mmap_region(file, addr, len, vm_flags, pgoff);
1290	if (!IS_ERR_VALUE(addr) &&
1291	    ((vm_flags & VM_LOCKED) ||
1292	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1293		*populate = len;
1294	return addr;
1295}
1296
1297SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1298		unsigned long, prot, unsigned long, flags,
1299		unsigned long, fd, unsigned long, pgoff)
1300{
1301	struct file *file = NULL;
1302	unsigned long retval;
1303
1304	if (!(flags & MAP_ANONYMOUS)) {
1305		audit_mmap_fd(fd, flags);
1306		file = fget(fd);
1307		if (!file)
1308			return -EBADF;
1309		if (is_file_hugepages(file))
1310			len = ALIGN(len, huge_page_size(hstate_file(file)));
1311		retval = -EINVAL;
1312		if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1313			goto out_fput;
 
1314	} else if (flags & MAP_HUGETLB) {
1315		struct user_struct *user = NULL;
1316		struct hstate *hs;
1317
1318		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1319		if (!hs)
1320			return -EINVAL;
1321
1322		len = ALIGN(len, huge_page_size(hs));
1323		/*
1324		 * VM_NORESERVE is used because the reservations will be
1325		 * taken when vm_ops->mmap() is called
1326		 * A dummy user value is used because we are not locking
1327		 * memory so no accounting is necessary
1328		 */
1329		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1330				VM_NORESERVE,
1331				&user, HUGETLB_ANONHUGE_INODE,
1332				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1333		if (IS_ERR(file))
1334			return PTR_ERR(file);
1335	}
1336
1337	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1338
1339	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1340out_fput:
1341	if (file)
1342		fput(file);
1343	return retval;
1344}
1345
 
 
 
 
 
 
 
1346#ifdef __ARCH_WANT_SYS_OLD_MMAP
1347struct mmap_arg_struct {
1348	unsigned long addr;
1349	unsigned long len;
1350	unsigned long prot;
1351	unsigned long flags;
1352	unsigned long fd;
1353	unsigned long offset;
1354};
1355
1356SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1357{
1358	struct mmap_arg_struct a;
1359
1360	if (copy_from_user(&a, arg, sizeof(a)))
1361		return -EFAULT;
1362	if (offset_in_page(a.offset))
1363		return -EINVAL;
1364
1365	return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1366			      a.offset >> PAGE_SHIFT);
1367}
1368#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1369
1370/*
1371 * Some shared mappigns will want the pages marked read-only
1372 * to track write events. If so, we'll downgrade vm_page_prot
1373 * to the private version (using protection_map[] without the
1374 * VM_SHARED bit).
1375 */
1376int vma_wants_writenotify(struct vm_area_struct *vma)
1377{
1378	vm_flags_t vm_flags = vma->vm_flags;
1379	const struct vm_operations_struct *vm_ops = vma->vm_ops;
1380
1381	/* If it was private or non-writable, the write bit is already clear */
1382	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1383		return 0;
1384
1385	/* The backer wishes to know when pages are first written to? */
1386	if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1387		return 1;
1388
1389	/* The open routine did something to the protections that pgprot_modify
1390	 * won't preserve? */
1391	if (pgprot_val(vma->vm_page_prot) !=
1392	    pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1393		return 0;
1394
1395	/* Do we need to track softdirty? */
1396	if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1397		return 1;
1398
1399	/* Specialty mapping? */
1400	if (vm_flags & VM_PFNMAP)
1401		return 0;
1402
1403	/* Can the mapping track the dirty pages? */
1404	return vma->vm_file && vma->vm_file->f_mapping &&
1405		mapping_cap_account_dirty(vma->vm_file->f_mapping);
1406}
1407
1408/*
1409 * We account for memory if it's a private writeable mapping,
1410 * not hugepages and VM_NORESERVE wasn't set.
1411 */
1412static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1413{
1414	/*
1415	 * hugetlb has its own accounting separate from the core VM
1416	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1417	 */
1418	if (file && is_file_hugepages(file))
1419		return 0;
1420
1421	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1422}
1423
1424unsigned long mmap_region(struct file *file, unsigned long addr,
1425		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
 
1426{
1427	struct mm_struct *mm = current->mm;
1428	struct vm_area_struct *vma, *prev;
1429	int error;
1430	struct rb_node **rb_link, *rb_parent;
1431	unsigned long charged = 0;
1432
1433	/* Check against address space limit. */
1434	if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1435		unsigned long nr_pages;
1436
1437		/*
1438		 * MAP_FIXED may remove pages of mappings that intersects with
1439		 * requested mapping. Account for the pages it would unmap.
1440		 */
1441		nr_pages = count_vma_pages_range(mm, addr, addr + len);
1442
1443		if (!may_expand_vm(mm, vm_flags,
1444					(len >> PAGE_SHIFT) - nr_pages))
1445			return -ENOMEM;
1446	}
1447
1448	/* Clear old maps */
1449	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1450			      &rb_parent)) {
1451		if (do_munmap(mm, addr, len))
1452			return -ENOMEM;
1453	}
1454
1455	/*
1456	 * Private writable mapping: check memory availability
1457	 */
1458	if (accountable_mapping(file, vm_flags)) {
1459		charged = len >> PAGE_SHIFT;
1460		if (security_vm_enough_memory_mm(mm, charged))
1461			return -ENOMEM;
1462		vm_flags |= VM_ACCOUNT;
1463	}
1464
1465	/*
1466	 * Can we just expand an old mapping?
1467	 */
1468	vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1469			NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1470	if (vma)
1471		goto out;
1472
1473	/*
1474	 * Determine the object being mapped and call the appropriate
1475	 * specific mapper. the address has already been validated, but
1476	 * not unmapped, but the maps are removed from the list.
1477	 */
1478	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1479	if (!vma) {
1480		error = -ENOMEM;
1481		goto unacct_error;
1482	}
1483
1484	vma->vm_mm = mm;
1485	vma->vm_start = addr;
1486	vma->vm_end = addr + len;
1487	vma->vm_flags = vm_flags;
1488	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1489	vma->vm_pgoff = pgoff;
1490	INIT_LIST_HEAD(&vma->anon_vma_chain);
1491
1492	if (file) {
1493		if (vm_flags & VM_DENYWRITE) {
1494			error = deny_write_access(file);
1495			if (error)
1496				goto free_vma;
1497		}
1498		if (vm_flags & VM_SHARED) {
1499			error = mapping_map_writable(file->f_mapping);
1500			if (error)
1501				goto allow_write_and_free_vma;
1502		}
1503
1504		/* ->mmap() can change vma->vm_file, but must guarantee that
1505		 * vma_link() below can deny write-access if VM_DENYWRITE is set
1506		 * and map writably if VM_SHARED is set. This usually means the
1507		 * new file must not have been exposed to user-space, yet.
1508		 */
1509		vma->vm_file = get_file(file);
1510		error = file->f_op->mmap(file, vma);
1511		if (error)
1512			goto unmap_and_free_vma;
1513
1514		/* Can addr have changed??
1515		 *
1516		 * Answer: Yes, several device drivers can do it in their
1517		 *         f_op->mmap method. -DaveM
1518		 * Bug: If addr is changed, prev, rb_link, rb_parent should
1519		 *      be updated for vma_link()
1520		 */
1521		WARN_ON_ONCE(addr != vma->vm_start);
1522
1523		addr = vma->vm_start;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1524		vm_flags = vma->vm_flags;
1525	} else if (vm_flags & VM_SHARED) {
1526		error = shmem_zero_setup(vma);
1527		if (error)
1528			goto free_vma;
 
 
 
 
 
 
 
 
 
 
 
1529	}
1530
1531	vma_link(mm, vma, prev, rb_link, rb_parent);
1532	/* Once vma denies write, undo our temporary denial count */
1533	if (file) {
 
1534		if (vm_flags & VM_SHARED)
1535			mapping_unmap_writable(file->f_mapping);
1536		if (vm_flags & VM_DENYWRITE)
1537			allow_write_access(file);
1538	}
1539	file = vma->vm_file;
1540out:
1541	perf_event_mmap(vma);
1542
1543	vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1544	if (vm_flags & VM_LOCKED) {
1545		if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1546					vma == get_gate_vma(current->mm)))
1547			mm->locked_vm += (len >> PAGE_SHIFT);
1548		else
1549			vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
 
 
1550	}
1551
1552	if (file)
1553		uprobe_mmap(vma);
1554
1555	/*
1556	 * New (or expanded) vma always get soft dirty status.
1557	 * Otherwise user-space soft-dirty page tracker won't
1558	 * be able to distinguish situation when vma area unmapped,
1559	 * then new mapped in-place (which must be aimed as
1560	 * a completely new data area).
1561	 */
1562	vma->vm_flags |= VM_SOFTDIRTY;
1563
1564	vma_set_page_prot(vma);
1565
1566	return addr;
1567
1568unmap_and_free_vma:
 
1569	vma->vm_file = NULL;
1570	fput(file);
1571
1572	/* Undo any partial mapping done by a device driver. */
1573	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1574	charged = 0;
1575	if (vm_flags & VM_SHARED)
1576		mapping_unmap_writable(file->f_mapping);
1577allow_write_and_free_vma:
1578	if (vm_flags & VM_DENYWRITE)
1579		allow_write_access(file);
1580free_vma:
1581	kmem_cache_free(vm_area_cachep, vma);
1582unacct_error:
1583	if (charged)
1584		vm_unacct_memory(charged);
1585	return error;
1586}
1587
1588unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1589{
1590	/*
1591	 * We implement the search by looking for an rbtree node that
1592	 * immediately follows a suitable gap. That is,
1593	 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1594	 * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1595	 * - gap_end - gap_start >= length
1596	 */
1597
1598	struct mm_struct *mm = current->mm;
1599	struct vm_area_struct *vma;
1600	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1601
1602	/* Adjust search length to account for worst case alignment overhead */
1603	length = info->length + info->align_mask;
1604	if (length < info->length)
1605		return -ENOMEM;
1606
1607	/* Adjust search limits by the desired length */
1608	if (info->high_limit < length)
1609		return -ENOMEM;
1610	high_limit = info->high_limit - length;
1611
1612	if (info->low_limit > high_limit)
1613		return -ENOMEM;
1614	low_limit = info->low_limit + length;
1615
1616	/* Check if rbtree root looks promising */
1617	if (RB_EMPTY_ROOT(&mm->mm_rb))
1618		goto check_highest;
1619	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1620	if (vma->rb_subtree_gap < length)
1621		goto check_highest;
1622
1623	while (true) {
1624		/* Visit left subtree if it looks promising */
1625		gap_end = vma->vm_start;
1626		if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1627			struct vm_area_struct *left =
1628				rb_entry(vma->vm_rb.rb_left,
1629					 struct vm_area_struct, vm_rb);
1630			if (left->rb_subtree_gap >= length) {
1631				vma = left;
1632				continue;
1633			}
1634		}
1635
1636		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1637check_current:
1638		/* Check if current node has a suitable gap */
1639		if (gap_start > high_limit)
1640			return -ENOMEM;
1641		if (gap_end >= low_limit && gap_end - gap_start >= length)
 
1642			goto found;
1643
1644		/* Visit right subtree if it looks promising */
1645		if (vma->vm_rb.rb_right) {
1646			struct vm_area_struct *right =
1647				rb_entry(vma->vm_rb.rb_right,
1648					 struct vm_area_struct, vm_rb);
1649			if (right->rb_subtree_gap >= length) {
1650				vma = right;
1651				continue;
1652			}
1653		}
1654
1655		/* Go back up the rbtree to find next candidate node */
1656		while (true) {
1657			struct rb_node *prev = &vma->vm_rb;
1658			if (!rb_parent(prev))
1659				goto check_highest;
1660			vma = rb_entry(rb_parent(prev),
1661				       struct vm_area_struct, vm_rb);
1662			if (prev == vma->vm_rb.rb_left) {
1663				gap_start = vma->vm_prev->vm_end;
1664				gap_end = vma->vm_start;
1665				goto check_current;
1666			}
1667		}
1668	}
1669
1670check_highest:
1671	/* Check highest gap, which does not precede any rbtree node */
1672	gap_start = mm->highest_vm_end;
1673	gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1674	if (gap_start > high_limit)
1675		return -ENOMEM;
1676
1677found:
1678	/* We found a suitable gap. Clip it with the original low_limit. */
1679	if (gap_start < info->low_limit)
1680		gap_start = info->low_limit;
1681
1682	/* Adjust gap address to the desired alignment */
1683	gap_start += (info->align_offset - gap_start) & info->align_mask;
1684
1685	VM_BUG_ON(gap_start + info->length > info->high_limit);
1686	VM_BUG_ON(gap_start + info->length > gap_end);
1687	return gap_start;
1688}
1689
1690unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1691{
1692	struct mm_struct *mm = current->mm;
1693	struct vm_area_struct *vma;
1694	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1695
1696	/* Adjust search length to account for worst case alignment overhead */
1697	length = info->length + info->align_mask;
1698	if (length < info->length)
1699		return -ENOMEM;
1700
1701	/*
1702	 * Adjust search limits by the desired length.
1703	 * See implementation comment at top of unmapped_area().
1704	 */
1705	gap_end = info->high_limit;
1706	if (gap_end < length)
1707		return -ENOMEM;
1708	high_limit = gap_end - length;
1709
1710	if (info->low_limit > high_limit)
1711		return -ENOMEM;
1712	low_limit = info->low_limit + length;
1713
1714	/* Check highest gap, which does not precede any rbtree node */
1715	gap_start = mm->highest_vm_end;
1716	if (gap_start <= high_limit)
1717		goto found_highest;
1718
1719	/* Check if rbtree root looks promising */
1720	if (RB_EMPTY_ROOT(&mm->mm_rb))
1721		return -ENOMEM;
1722	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1723	if (vma->rb_subtree_gap < length)
1724		return -ENOMEM;
1725
1726	while (true) {
1727		/* Visit right subtree if it looks promising */
1728		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1729		if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1730			struct vm_area_struct *right =
1731				rb_entry(vma->vm_rb.rb_right,
1732					 struct vm_area_struct, vm_rb);
1733			if (right->rb_subtree_gap >= length) {
1734				vma = right;
1735				continue;
1736			}
1737		}
1738
1739check_current:
1740		/* Check if current node has a suitable gap */
1741		gap_end = vma->vm_start;
1742		if (gap_end < low_limit)
1743			return -ENOMEM;
1744		if (gap_start <= high_limit && gap_end - gap_start >= length)
 
1745			goto found;
1746
1747		/* Visit left subtree if it looks promising */
1748		if (vma->vm_rb.rb_left) {
1749			struct vm_area_struct *left =
1750				rb_entry(vma->vm_rb.rb_left,
1751					 struct vm_area_struct, vm_rb);
1752			if (left->rb_subtree_gap >= length) {
1753				vma = left;
1754				continue;
1755			}
1756		}
1757
1758		/* Go back up the rbtree to find next candidate node */
1759		while (true) {
1760			struct rb_node *prev = &vma->vm_rb;
1761			if (!rb_parent(prev))
1762				return -ENOMEM;
1763			vma = rb_entry(rb_parent(prev),
1764				       struct vm_area_struct, vm_rb);
1765			if (prev == vma->vm_rb.rb_right) {
1766				gap_start = vma->vm_prev ?
1767					vma->vm_prev->vm_end : 0;
1768				goto check_current;
1769			}
1770		}
1771	}
1772
1773found:
1774	/* We found a suitable gap. Clip it with the original high_limit. */
1775	if (gap_end > info->high_limit)
1776		gap_end = info->high_limit;
1777
1778found_highest:
1779	/* Compute highest gap address at the desired alignment */
1780	gap_end -= info->length;
1781	gap_end -= (gap_end - info->align_offset) & info->align_mask;
1782
1783	VM_BUG_ON(gap_end < info->low_limit);
1784	VM_BUG_ON(gap_end < gap_start);
1785	return gap_end;
1786}
1787
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1788/* Get an address range which is currently unmapped.
1789 * For shmat() with addr=0.
1790 *
1791 * Ugly calling convention alert:
1792 * Return value with the low bits set means error value,
1793 * ie
1794 *	if (ret & ~PAGE_MASK)
1795 *		error = ret;
1796 *
1797 * This function "knows" that -ENOMEM has the bits set.
1798 */
1799#ifndef HAVE_ARCH_UNMAPPED_AREA
1800unsigned long
1801arch_get_unmapped_area(struct file *filp, unsigned long addr,
1802		unsigned long len, unsigned long pgoff, unsigned long flags)
1803{
1804	struct mm_struct *mm = current->mm;
1805	struct vm_area_struct *vma;
1806	struct vm_unmapped_area_info info;
 
1807
1808	if (len > TASK_SIZE - mmap_min_addr)
1809		return -ENOMEM;
1810
1811	if (flags & MAP_FIXED)
1812		return addr;
1813
1814	if (addr) {
1815		addr = PAGE_ALIGN(addr);
1816		vma = find_vma(mm, addr);
1817		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1818		    (!vma || addr + len <= vma->vm_start))
 
1819			return addr;
1820	}
1821
1822	info.flags = 0;
1823	info.length = len;
1824	info.low_limit = mm->mmap_base;
1825	info.high_limit = TASK_SIZE;
1826	info.align_mask = 0;
 
1827	return vm_unmapped_area(&info);
1828}
1829#endif
1830
1831/*
1832 * This mmap-allocator allocates new areas top-down from below the
1833 * stack's low limit (the base):
1834 */
1835#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1836unsigned long
1837arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1838			  const unsigned long len, const unsigned long pgoff,
1839			  const unsigned long flags)
1840{
1841	struct vm_area_struct *vma;
1842	struct mm_struct *mm = current->mm;
1843	unsigned long addr = addr0;
1844	struct vm_unmapped_area_info info;
 
1845
1846	/* requested length too big for entire address space */
1847	if (len > TASK_SIZE - mmap_min_addr)
1848		return -ENOMEM;
1849
1850	if (flags & MAP_FIXED)
1851		return addr;
1852
1853	/* requesting a specific address */
1854	if (addr) {
1855		addr = PAGE_ALIGN(addr);
1856		vma = find_vma(mm, addr);
1857		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1858				(!vma || addr + len <= vma->vm_start))
 
1859			return addr;
1860	}
1861
1862	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1863	info.length = len;
1864	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1865	info.high_limit = mm->mmap_base;
1866	info.align_mask = 0;
 
1867	addr = vm_unmapped_area(&info);
1868
1869	/*
1870	 * A failed mmap() very likely causes application failure,
1871	 * so fall back to the bottom-up function here. This scenario
1872	 * can happen with large stack limits and large mmap()
1873	 * allocations.
1874	 */
1875	if (offset_in_page(addr)) {
1876		VM_BUG_ON(addr != -ENOMEM);
1877		info.flags = 0;
1878		info.low_limit = TASK_UNMAPPED_BASE;
1879		info.high_limit = TASK_SIZE;
1880		addr = vm_unmapped_area(&info);
1881	}
1882
1883	return addr;
1884}
1885#endif
1886
1887unsigned long
1888get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1889		unsigned long pgoff, unsigned long flags)
1890{
1891	unsigned long (*get_area)(struct file *, unsigned long,
1892				  unsigned long, unsigned long, unsigned long);
1893
1894	unsigned long error = arch_mmap_check(addr, len, flags);
1895	if (error)
1896		return error;
1897
1898	/* Careful about overflows.. */
1899	if (len > TASK_SIZE)
1900		return -ENOMEM;
1901
1902	get_area = current->mm->get_unmapped_area;
1903	if (file && file->f_op->get_unmapped_area)
1904		get_area = file->f_op->get_unmapped_area;
 
 
 
 
 
 
 
 
 
 
 
1905	addr = get_area(file, addr, len, pgoff, flags);
1906	if (IS_ERR_VALUE(addr))
1907		return addr;
1908
1909	if (addr > TASK_SIZE - len)
1910		return -ENOMEM;
1911	if (offset_in_page(addr))
1912		return -EINVAL;
1913
1914	addr = arch_rebalance_pgtables(addr, len);
1915	error = security_mmap_addr(addr);
1916	return error ? error : addr;
1917}
1918
1919EXPORT_SYMBOL(get_unmapped_area);
1920
1921/* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1922struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1923{
1924	struct rb_node *rb_node;
1925	struct vm_area_struct *vma;
1926
1927	/* Check the cache first. */
1928	vma = vmacache_find(mm, addr);
1929	if (likely(vma))
1930		return vma;
1931
1932	rb_node = mm->mm_rb.rb_node;
1933
1934	while (rb_node) {
1935		struct vm_area_struct *tmp;
1936
1937		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1938
1939		if (tmp->vm_end > addr) {
1940			vma = tmp;
1941			if (tmp->vm_start <= addr)
1942				break;
1943			rb_node = rb_node->rb_left;
1944		} else
1945			rb_node = rb_node->rb_right;
1946	}
1947
1948	if (vma)
1949		vmacache_update(addr, vma);
1950	return vma;
1951}
1952
1953EXPORT_SYMBOL(find_vma);
1954
1955/*
1956 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1957 */
1958struct vm_area_struct *
1959find_vma_prev(struct mm_struct *mm, unsigned long addr,
1960			struct vm_area_struct **pprev)
1961{
1962	struct vm_area_struct *vma;
1963
1964	vma = find_vma(mm, addr);
1965	if (vma) {
1966		*pprev = vma->vm_prev;
1967	} else {
1968		struct rb_node *rb_node = mm->mm_rb.rb_node;
1969		*pprev = NULL;
1970		while (rb_node) {
1971			*pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1972			rb_node = rb_node->rb_right;
1973		}
1974	}
1975	return vma;
1976}
1977
1978/*
1979 * Verify that the stack growth is acceptable and
1980 * update accounting. This is shared with both the
1981 * grow-up and grow-down cases.
1982 */
1983static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
 
1984{
1985	struct mm_struct *mm = vma->vm_mm;
1986	struct rlimit *rlim = current->signal->rlim;
1987	unsigned long new_start, actual_size;
1988
1989	/* address space limit tests */
1990	if (!may_expand_vm(mm, vma->vm_flags, grow))
1991		return -ENOMEM;
1992
1993	/* Stack limit test */
1994	actual_size = size;
1995	if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
1996		actual_size -= PAGE_SIZE;
1997	if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1998		return -ENOMEM;
1999
2000	/* mlock limit tests */
2001	if (vma->vm_flags & VM_LOCKED) {
2002		unsigned long locked;
2003		unsigned long limit;
2004		locked = mm->locked_vm + grow;
2005		limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2006		limit >>= PAGE_SHIFT;
2007		if (locked > limit && !capable(CAP_IPC_LOCK))
2008			return -ENOMEM;
2009	}
2010
2011	/* Check to ensure the stack will not grow into a hugetlb-only region */
2012	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2013			vma->vm_end - size;
2014	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2015		return -EFAULT;
2016
2017	/*
2018	 * Overcommit..  This must be the final test, as it will
2019	 * update security statistics.
2020	 */
2021	if (security_vm_enough_memory_mm(mm, grow))
2022		return -ENOMEM;
2023
2024	return 0;
2025}
2026
2027#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2028/*
2029 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2030 * vma is the last one with address > vma->vm_end.  Have to extend vma.
2031 */
2032int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2033{
2034	struct mm_struct *mm = vma->vm_mm;
 
 
2035	int error = 0;
2036
2037	if (!(vma->vm_flags & VM_GROWSUP))
2038		return -EFAULT;
2039
2040	/* Guard against wrapping around to address 0. */
2041	if (address < PAGE_ALIGN(address+4))
2042		address = PAGE_ALIGN(address+4);
2043	else
2044		return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2045
2046	/* We must make sure the anon_vma is allocated. */
2047	if (unlikely(anon_vma_prepare(vma)))
2048		return -ENOMEM;
2049
2050	/*
2051	 * vma->vm_start/vm_end cannot change under us because the caller
2052	 * is required to hold the mmap_sem in read mode.  We need the
2053	 * anon_vma lock to serialize against concurrent expand_stacks.
2054	 */
2055	anon_vma_lock_write(vma->anon_vma);
2056
2057	/* Somebody else might have raced and expanded it already */
2058	if (address > vma->vm_end) {
2059		unsigned long size, grow;
2060
2061		size = address - vma->vm_start;
2062		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2063
2064		error = -ENOMEM;
2065		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2066			error = acct_stack_growth(vma, size, grow);
2067			if (!error) {
2068				/*
2069				 * vma_gap_update() doesn't support concurrent
2070				 * updates, but we only hold a shared mmap_sem
2071				 * lock here, so we need to protect against
2072				 * concurrent vma expansions.
2073				 * anon_vma_lock_write() doesn't help here, as
2074				 * we don't guarantee that all growable vmas
2075				 * in a mm share the same root anon vma.
2076				 * So, we reuse mm->page_table_lock to guard
2077				 * against concurrent vma expansions.
2078				 */
2079				spin_lock(&mm->page_table_lock);
2080				if (vma->vm_flags & VM_LOCKED)
2081					mm->locked_vm += grow;
2082				vm_stat_account(mm, vma->vm_flags, grow);
2083				anon_vma_interval_tree_pre_update_vma(vma);
2084				vma->vm_end = address;
2085				anon_vma_interval_tree_post_update_vma(vma);
2086				if (vma->vm_next)
2087					vma_gap_update(vma->vm_next);
2088				else
2089					mm->highest_vm_end = address;
2090				spin_unlock(&mm->page_table_lock);
2091
2092				perf_event_mmap(vma);
2093			}
2094		}
2095	}
2096	anon_vma_unlock_write(vma->anon_vma);
2097	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2098	validate_mm(mm);
2099	return error;
2100}
2101#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2102
2103/*
2104 * vma is the first one with address < vma->vm_start.  Have to extend vma.
2105 */
2106int expand_downwards(struct vm_area_struct *vma,
2107				   unsigned long address)
2108{
2109	struct mm_struct *mm = vma->vm_mm;
2110	int error;
 
2111
2112	address &= PAGE_MASK;
2113	error = security_mmap_addr(address);
2114	if (error)
2115		return error;
 
 
 
 
 
 
 
 
2116
2117	/* We must make sure the anon_vma is allocated. */
2118	if (unlikely(anon_vma_prepare(vma)))
2119		return -ENOMEM;
2120
2121	/*
2122	 * vma->vm_start/vm_end cannot change under us because the caller
2123	 * is required to hold the mmap_sem in read mode.  We need the
2124	 * anon_vma lock to serialize against concurrent expand_stacks.
2125	 */
2126	anon_vma_lock_write(vma->anon_vma);
2127
2128	/* Somebody else might have raced and expanded it already */
2129	if (address < vma->vm_start) {
2130		unsigned long size, grow;
2131
2132		size = vma->vm_end - address;
2133		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2134
2135		error = -ENOMEM;
2136		if (grow <= vma->vm_pgoff) {
2137			error = acct_stack_growth(vma, size, grow);
2138			if (!error) {
2139				/*
2140				 * vma_gap_update() doesn't support concurrent
2141				 * updates, but we only hold a shared mmap_sem
2142				 * lock here, so we need to protect against
2143				 * concurrent vma expansions.
2144				 * anon_vma_lock_write() doesn't help here, as
2145				 * we don't guarantee that all growable vmas
2146				 * in a mm share the same root anon vma.
2147				 * So, we reuse mm->page_table_lock to guard
2148				 * against concurrent vma expansions.
2149				 */
2150				spin_lock(&mm->page_table_lock);
2151				if (vma->vm_flags & VM_LOCKED)
2152					mm->locked_vm += grow;
2153				vm_stat_account(mm, vma->vm_flags, grow);
2154				anon_vma_interval_tree_pre_update_vma(vma);
2155				vma->vm_start = address;
2156				vma->vm_pgoff -= grow;
2157				anon_vma_interval_tree_post_update_vma(vma);
2158				vma_gap_update(vma);
2159				spin_unlock(&mm->page_table_lock);
2160
2161				perf_event_mmap(vma);
2162			}
2163		}
2164	}
2165	anon_vma_unlock_write(vma->anon_vma);
2166	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2167	validate_mm(mm);
2168	return error;
2169}
2170
2171/*
2172 * Note how expand_stack() refuses to expand the stack all the way to
2173 * abut the next virtual mapping, *unless* that mapping itself is also
2174 * a stack mapping. We want to leave room for a guard page, after all
2175 * (the guard page itself is not added here, that is done by the
2176 * actual page faulting logic)
2177 *
2178 * This matches the behavior of the guard page logic (see mm/memory.c:
2179 * check_stack_guard_page()), which only allows the guard page to be
2180 * removed under these circumstances.
2181 */
 
 
 
 
 
2182#ifdef CONFIG_STACK_GROWSUP
2183int expand_stack(struct vm_area_struct *vma, unsigned long address)
2184{
2185	struct vm_area_struct *next;
2186
2187	address &= PAGE_MASK;
2188	next = vma->vm_next;
2189	if (next && next->vm_start == address + PAGE_SIZE) {
2190		if (!(next->vm_flags & VM_GROWSUP))
2191			return -ENOMEM;
2192	}
2193	return expand_upwards(vma, address);
2194}
2195
2196struct vm_area_struct *
2197find_extend_vma(struct mm_struct *mm, unsigned long addr)
2198{
2199	struct vm_area_struct *vma, *prev;
2200
2201	addr &= PAGE_MASK;
2202	vma = find_vma_prev(mm, addr, &prev);
2203	if (vma && (vma->vm_start <= addr))
2204		return vma;
 
2205	if (!prev || expand_stack(prev, addr))
2206		return NULL;
2207	if (prev->vm_flags & VM_LOCKED)
2208		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2209	return prev;
2210}
2211#else
2212int expand_stack(struct vm_area_struct *vma, unsigned long address)
2213{
2214	struct vm_area_struct *prev;
2215
2216	address &= PAGE_MASK;
2217	prev = vma->vm_prev;
2218	if (prev && prev->vm_end == address) {
2219		if (!(prev->vm_flags & VM_GROWSDOWN))
2220			return -ENOMEM;
2221	}
2222	return expand_downwards(vma, address);
2223}
2224
2225struct vm_area_struct *
2226find_extend_vma(struct mm_struct *mm, unsigned long addr)
2227{
2228	struct vm_area_struct *vma;
2229	unsigned long start;
2230
2231	addr &= PAGE_MASK;
2232	vma = find_vma(mm, addr);
2233	if (!vma)
2234		return NULL;
2235	if (vma->vm_start <= addr)
2236		return vma;
2237	if (!(vma->vm_flags & VM_GROWSDOWN))
2238		return NULL;
2239	start = vma->vm_start;
2240	if (expand_stack(vma, addr))
2241		return NULL;
2242	if (vma->vm_flags & VM_LOCKED)
2243		populate_vma_page_range(vma, addr, start, NULL);
2244	return vma;
2245}
2246#endif
2247
2248EXPORT_SYMBOL_GPL(find_extend_vma);
2249
2250/*
2251 * Ok - we have the memory areas we should free on the vma list,
2252 * so release them, and do the vma updates.
2253 *
2254 * Called with the mm semaphore held.
2255 */
2256static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2257{
2258	unsigned long nr_accounted = 0;
2259
2260	/* Update high watermark before we lower total_vm */
2261	update_hiwater_vm(mm);
2262	do {
2263		long nrpages = vma_pages(vma);
2264
2265		if (vma->vm_flags & VM_ACCOUNT)
2266			nr_accounted += nrpages;
2267		vm_stat_account(mm, vma->vm_flags, -nrpages);
2268		vma = remove_vma(vma);
2269	} while (vma);
2270	vm_unacct_memory(nr_accounted);
2271	validate_mm(mm);
2272}
2273
2274/*
2275 * Get rid of page table information in the indicated region.
2276 *
2277 * Called with the mm semaphore held.
2278 */
2279static void unmap_region(struct mm_struct *mm,
2280		struct vm_area_struct *vma, struct vm_area_struct *prev,
2281		unsigned long start, unsigned long end)
2282{
2283	struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2284	struct mmu_gather tlb;
2285
2286	lru_add_drain();
2287	tlb_gather_mmu(&tlb, mm, start, end);
2288	update_hiwater_rss(mm);
2289	unmap_vmas(&tlb, vma, start, end);
2290	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2291				 next ? next->vm_start : USER_PGTABLES_CEILING);
2292	tlb_finish_mmu(&tlb, start, end);
2293}
2294
2295/*
2296 * Create a list of vma's touched by the unmap, removing them from the mm's
2297 * vma list as we go..
2298 */
2299static void
2300detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2301	struct vm_area_struct *prev, unsigned long end)
2302{
2303	struct vm_area_struct **insertion_point;
2304	struct vm_area_struct *tail_vma = NULL;
2305
2306	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2307	vma->vm_prev = NULL;
2308	do {
2309		vma_rb_erase(vma, &mm->mm_rb);
2310		mm->map_count--;
2311		tail_vma = vma;
2312		vma = vma->vm_next;
2313	} while (vma && vma->vm_start < end);
2314	*insertion_point = vma;
2315	if (vma) {
2316		vma->vm_prev = prev;
2317		vma_gap_update(vma);
2318	} else
2319		mm->highest_vm_end = prev ? prev->vm_end : 0;
2320	tail_vma->vm_next = NULL;
2321
2322	/* Kill the cache */
2323	vmacache_invalidate(mm);
 
 
 
 
 
 
 
 
 
 
 
2324}
2325
2326/*
2327 * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2328 * munmap path where it doesn't make sense to fail.
2329 */
2330static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2331	      unsigned long addr, int new_below)
2332{
2333	struct vm_area_struct *new;
2334	int err;
2335
2336	if (is_vm_hugetlb_page(vma) && (addr &
2337					~(huge_page_mask(hstate_vma(vma)))))
2338		return -EINVAL;
 
 
2339
2340	new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2341	if (!new)
2342		return -ENOMEM;
2343
2344	/* most fields are the same, copy all, and then fixup */
2345	*new = *vma;
2346
2347	INIT_LIST_HEAD(&new->anon_vma_chain);
2348
2349	if (new_below)
2350		new->vm_end = addr;
2351	else {
2352		new->vm_start = addr;
2353		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2354	}
2355
2356	err = vma_dup_policy(vma, new);
2357	if (err)
2358		goto out_free_vma;
2359
2360	err = anon_vma_clone(new, vma);
2361	if (err)
2362		goto out_free_mpol;
2363
2364	if (new->vm_file)
2365		get_file(new->vm_file);
2366
2367	if (new->vm_ops && new->vm_ops->open)
2368		new->vm_ops->open(new);
2369
2370	if (new_below)
2371		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2372			((addr - new->vm_start) >> PAGE_SHIFT), new);
2373	else
2374		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2375
2376	/* Success. */
2377	if (!err)
2378		return 0;
2379
2380	/* Clean everything up if vma_adjust failed. */
2381	if (new->vm_ops && new->vm_ops->close)
2382		new->vm_ops->close(new);
2383	if (new->vm_file)
2384		fput(new->vm_file);
2385	unlink_anon_vmas(new);
2386 out_free_mpol:
2387	mpol_put(vma_policy(new));
2388 out_free_vma:
2389	kmem_cache_free(vm_area_cachep, new);
2390	return err;
2391}
2392
2393/*
2394 * Split a vma into two pieces at address 'addr', a new vma is allocated
2395 * either for the first part or the tail.
2396 */
2397int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2398	      unsigned long addr, int new_below)
2399{
2400	if (mm->map_count >= sysctl_max_map_count)
2401		return -ENOMEM;
2402
2403	return __split_vma(mm, vma, addr, new_below);
2404}
2405
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2406/* Munmap is split into 2 main parts -- this part which finds
2407 * what needs doing, and the areas themselves, which do the
2408 * work.  This now handles partial unmappings.
2409 * Jeremy Fitzhardinge <jeremy@goop.org>
2410 */
2411int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
 
2412{
2413	unsigned long end;
2414	struct vm_area_struct *vma, *prev, *last;
2415
2416	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2417		return -EINVAL;
2418
2419	len = PAGE_ALIGN(len);
 
2420	if (len == 0)
2421		return -EINVAL;
2422
2423	/* Find the first overlapping VMA */
2424	vma = find_vma(mm, start);
 
 
 
 
 
 
 
2425	if (!vma)
2426		return 0;
2427	prev = vma->vm_prev;
2428	/* we have  start < vma->vm_end  */
2429
2430	/* if it doesn't overlap, we have nothing.. */
2431	end = start + len;
2432	if (vma->vm_start >= end)
2433		return 0;
2434
2435	/*
2436	 * If we need to split any vma, do it now to save pain later.
2437	 *
2438	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2439	 * unmapped vm_area_struct will remain in use: so lower split_vma
2440	 * places tmp vma above, and higher split_vma places tmp vma below.
2441	 */
2442	if (start > vma->vm_start) {
2443		int error;
2444
2445		/*
2446		 * Make sure that map_count on return from munmap() will
2447		 * not exceed its limit; but let map_count go just above
2448		 * its limit temporarily, to help free resources as expected.
2449		 */
2450		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2451			return -ENOMEM;
2452
2453		error = __split_vma(mm, vma, start, 0);
2454		if (error)
2455			return error;
2456		prev = vma;
2457	}
2458
2459	/* Does it split the last one? */
2460	last = find_vma(mm, end);
2461	if (last && end > last->vm_start) {
2462		int error = __split_vma(mm, last, end, 1);
2463		if (error)
2464			return error;
2465	}
2466	vma = prev ? prev->vm_next : mm->mmap;
2467
2468	/*
2469	 * unlock any mlock()ed ranges before detaching vmas
2470	 */
2471	if (mm->locked_vm) {
2472		struct vm_area_struct *tmp = vma;
2473		while (tmp && tmp->vm_start < end) {
2474			if (tmp->vm_flags & VM_LOCKED) {
2475				mm->locked_vm -= vma_pages(tmp);
2476				munlock_vma_pages_all(tmp);
2477			}
2478			tmp = tmp->vm_next;
2479		}
 
2480	}
2481
2482	/*
2483	 * Remove the vma's, and unmap the actual pages
2484	 */
2485	detach_vmas_to_be_unmapped(mm, vma, prev, end);
2486	unmap_region(mm, vma, prev, start, end);
 
 
 
 
 
 
 
2487
2488	arch_unmap(mm, vma, start, end);
2489
2490	/* Fix up all other VM information */
2491	remove_vma_list(mm, vma);
2492
2493	return 0;
2494}
2495
2496int vm_munmap(unsigned long start, size_t len)
 
 
 
 
 
 
2497{
2498	int ret;
2499	struct mm_struct *mm = current->mm;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2500
2501	down_write(&mm->mmap_sem);
2502	ret = do_munmap(mm, start, len);
2503	up_write(&mm->mmap_sem);
2504	return ret;
2505}
 
 
 
 
 
2506EXPORT_SYMBOL(vm_munmap);
2507
2508SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2509{
 
2510	profile_munmap(addr);
2511	return vm_munmap(addr, len);
2512}
2513
2514
2515/*
2516 * Emulation of deprecated remap_file_pages() syscall.
2517 */
2518SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2519		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2520{
2521
2522	struct mm_struct *mm = current->mm;
2523	struct vm_area_struct *vma;
2524	unsigned long populate = 0;
2525	unsigned long ret = -EINVAL;
2526	struct file *file;
2527
2528	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2529		     current->comm, current->pid);
2530
2531	if (prot)
2532		return ret;
2533	start = start & PAGE_MASK;
2534	size = size & PAGE_MASK;
2535
2536	if (start + size <= start)
2537		return ret;
2538
2539	/* Does pgoff wrap? */
2540	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2541		return ret;
2542
2543	down_write(&mm->mmap_sem);
 
 
2544	vma = find_vma(mm, start);
2545
2546	if (!vma || !(vma->vm_flags & VM_SHARED))
2547		goto out;
2548
2549	if (start < vma->vm_start)
2550		goto out;
2551
2552	if (start + size > vma->vm_end) {
2553		struct vm_area_struct *next;
2554
2555		for (next = vma->vm_next; next; next = next->vm_next) {
2556			/* hole between vmas ? */
2557			if (next->vm_start != next->vm_prev->vm_end)
2558				goto out;
2559
2560			if (next->vm_file != vma->vm_file)
2561				goto out;
2562
2563			if (next->vm_flags != vma->vm_flags)
2564				goto out;
2565
2566			if (start + size <= next->vm_end)
2567				break;
2568		}
2569
2570		if (!next)
2571			goto out;
2572	}
2573
2574	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2575	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2576	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2577
2578	flags &= MAP_NONBLOCK;
2579	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2580	if (vma->vm_flags & VM_LOCKED) {
2581		struct vm_area_struct *tmp;
2582		flags |= MAP_LOCKED;
2583
2584		/* drop PG_Mlocked flag for over-mapped range */
2585		for (tmp = vma; tmp->vm_start >= start + size;
2586				tmp = tmp->vm_next) {
2587			munlock_vma_pages_range(tmp,
2588					max(tmp->vm_start, start),
2589					min(tmp->vm_end, start + size));
2590		}
2591	}
2592
2593	file = get_file(vma->vm_file);
2594	ret = do_mmap_pgoff(vma->vm_file, start, size,
2595			prot, flags, pgoff, &populate);
2596	fput(file);
2597out:
2598	up_write(&mm->mmap_sem);
2599	if (populate)
2600		mm_populate(ret, populate);
2601	if (!IS_ERR_VALUE(ret))
2602		ret = 0;
2603	return ret;
2604}
2605
2606static inline void verify_mm_writelocked(struct mm_struct *mm)
2607{
2608#ifdef CONFIG_DEBUG_VM
2609	if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2610		WARN_ON(1);
2611		up_read(&mm->mmap_sem);
2612	}
2613#endif
2614}
2615
2616/*
2617 *  this is really a simplified "do_mmap".  it only handles
2618 *  anonymous maps.  eventually we may be able to do some
2619 *  brk-specific accounting here.
2620 */
2621static unsigned long do_brk(unsigned long addr, unsigned long len)
2622{
2623	struct mm_struct *mm = current->mm;
2624	struct vm_area_struct *vma, *prev;
2625	unsigned long flags;
2626	struct rb_node **rb_link, *rb_parent;
2627	pgoff_t pgoff = addr >> PAGE_SHIFT;
2628	int error;
 
2629
2630	len = PAGE_ALIGN(len);
2631	if (!len)
2632		return addr;
2633
2634	flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2635
2636	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2637	if (offset_in_page(error))
2638		return error;
2639
2640	error = mlock_future_check(mm, mm->def_flags, len);
2641	if (error)
2642		return error;
2643
2644	/*
2645	 * mm->mmap_sem is required to protect against another thread
2646	 * changing the mappings in case we sleep.
2647	 */
2648	verify_mm_writelocked(mm);
2649
2650	/*
2651	 * Clear old maps.  this also does some error checking for us
2652	 */
2653	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2654			      &rb_parent)) {
2655		if (do_munmap(mm, addr, len))
2656			return -ENOMEM;
2657	}
2658
2659	/* Check against address space limits *after* clearing old maps... */
2660	if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2661		return -ENOMEM;
2662
2663	if (mm->map_count > sysctl_max_map_count)
2664		return -ENOMEM;
2665
2666	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2667		return -ENOMEM;
2668
2669	/* Can we just expand an old private anonymous mapping? */
2670	vma = vma_merge(mm, prev, addr, addr + len, flags,
2671			NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2672	if (vma)
2673		goto out;
2674
2675	/*
2676	 * create a vma struct for an anonymous mapping
2677	 */
2678	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2679	if (!vma) {
2680		vm_unacct_memory(len >> PAGE_SHIFT);
2681		return -ENOMEM;
2682	}
2683
2684	INIT_LIST_HEAD(&vma->anon_vma_chain);
2685	vma->vm_mm = mm;
2686	vma->vm_start = addr;
2687	vma->vm_end = addr + len;
2688	vma->vm_pgoff = pgoff;
2689	vma->vm_flags = flags;
2690	vma->vm_page_prot = vm_get_page_prot(flags);
2691	vma_link(mm, vma, prev, rb_link, rb_parent);
2692out:
2693	perf_event_mmap(vma);
2694	mm->total_vm += len >> PAGE_SHIFT;
2695	mm->data_vm += len >> PAGE_SHIFT;
2696	if (flags & VM_LOCKED)
2697		mm->locked_vm += (len >> PAGE_SHIFT);
2698	vma->vm_flags |= VM_SOFTDIRTY;
2699	return addr;
2700}
2701
2702unsigned long vm_brk(unsigned long addr, unsigned long len)
2703{
2704	struct mm_struct *mm = current->mm;
2705	unsigned long ret;
 
2706	bool populate;
 
2707
2708	down_write(&mm->mmap_sem);
2709	ret = do_brk(addr, len);
 
 
 
 
 
 
 
 
2710	populate = ((mm->def_flags & VM_LOCKED) != 0);
2711	up_write(&mm->mmap_sem);
2712	if (populate)
 
2713		mm_populate(addr, len);
2714	return ret;
2715}
 
 
 
 
 
 
2716EXPORT_SYMBOL(vm_brk);
2717
2718/* Release all mmaps. */
2719void exit_mmap(struct mm_struct *mm)
2720{
2721	struct mmu_gather tlb;
2722	struct vm_area_struct *vma;
2723	unsigned long nr_accounted = 0;
2724
2725	/* mm's last user has gone, and its about to be pulled down */
2726	mmu_notifier_release(mm);
2727
2728	if (mm->locked_vm) {
2729		vma = mm->mmap;
2730		while (vma) {
2731			if (vma->vm_flags & VM_LOCKED)
2732				munlock_vma_pages_all(vma);
2733			vma = vma->vm_next;
2734		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2735	}
2736
 
 
 
2737	arch_exit_mmap(mm);
2738
2739	vma = mm->mmap;
2740	if (!vma)	/* Can happen if dup_mmap() received an OOM */
2741		return;
2742
2743	lru_add_drain();
2744	flush_cache_mm(mm);
2745	tlb_gather_mmu(&tlb, mm, 0, -1);
2746	/* update_hiwater_rss(mm) here? but nobody should be looking */
2747	/* Use -1 here to ensure all VMAs in the mm are unmapped */
2748	unmap_vmas(&tlb, vma, 0, -1);
2749
2750	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2751	tlb_finish_mmu(&tlb, 0, -1);
2752
2753	/*
2754	 * Walk the list again, actually closing and freeing it,
2755	 * with preemption enabled, without holding any MM locks.
2756	 */
2757	while (vma) {
2758		if (vma->vm_flags & VM_ACCOUNT)
2759			nr_accounted += vma_pages(vma);
2760		vma = remove_vma(vma);
 
2761	}
2762	vm_unacct_memory(nr_accounted);
2763}
2764
2765/* Insert vm structure into process list sorted by address
2766 * and into the inode's i_mmap tree.  If vm_file is non-NULL
2767 * then i_mmap_rwsem is taken here.
2768 */
2769int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2770{
2771	struct vm_area_struct *prev;
2772	struct rb_node **rb_link, *rb_parent;
2773
2774	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2775			   &prev, &rb_link, &rb_parent))
2776		return -ENOMEM;
2777	if ((vma->vm_flags & VM_ACCOUNT) &&
2778	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
2779		return -ENOMEM;
2780
2781	/*
2782	 * The vm_pgoff of a purely anonymous vma should be irrelevant
2783	 * until its first write fault, when page's anon_vma and index
2784	 * are set.  But now set the vm_pgoff it will almost certainly
2785	 * end up with (unless mremap moves it elsewhere before that
2786	 * first wfault), so /proc/pid/maps tells a consistent story.
2787	 *
2788	 * By setting it to reflect the virtual start address of the
2789	 * vma, merges and splits can happen in a seamless way, just
2790	 * using the existing file pgoff checks and manipulations.
2791	 * Similarly in do_mmap_pgoff and in do_brk.
2792	 */
2793	if (vma_is_anonymous(vma)) {
2794		BUG_ON(vma->anon_vma);
2795		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2796	}
2797
2798	vma_link(mm, vma, prev, rb_link, rb_parent);
2799	return 0;
2800}
2801
2802/*
2803 * Copy the vma structure to a new location in the same mm,
2804 * prior to moving page table entries, to effect an mremap move.
2805 */
2806struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2807	unsigned long addr, unsigned long len, pgoff_t pgoff,
2808	bool *need_rmap_locks)
2809{
2810	struct vm_area_struct *vma = *vmap;
2811	unsigned long vma_start = vma->vm_start;
2812	struct mm_struct *mm = vma->vm_mm;
2813	struct vm_area_struct *new_vma, *prev;
2814	struct rb_node **rb_link, *rb_parent;
2815	bool faulted_in_anon_vma = true;
2816
2817	/*
2818	 * If anonymous vma has not yet been faulted, update new pgoff
2819	 * to match new location, to increase its chance of merging.
2820	 */
2821	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2822		pgoff = addr >> PAGE_SHIFT;
2823		faulted_in_anon_vma = false;
2824	}
2825
2826	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2827		return NULL;	/* should never get here */
2828	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2829			    vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2830			    vma->vm_userfaultfd_ctx);
2831	if (new_vma) {
2832		/*
2833		 * Source vma may have been merged into new_vma
2834		 */
2835		if (unlikely(vma_start >= new_vma->vm_start &&
2836			     vma_start < new_vma->vm_end)) {
2837			/*
2838			 * The only way we can get a vma_merge with
2839			 * self during an mremap is if the vma hasn't
2840			 * been faulted in yet and we were allowed to
2841			 * reset the dst vma->vm_pgoff to the
2842			 * destination address of the mremap to allow
2843			 * the merge to happen. mremap must change the
2844			 * vm_pgoff linearity between src and dst vmas
2845			 * (in turn preventing a vma_merge) to be
2846			 * safe. It is only safe to keep the vm_pgoff
2847			 * linear if there are no pages mapped yet.
2848			 */
2849			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
2850			*vmap = vma = new_vma;
2851		}
2852		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2853	} else {
2854		new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2855		if (!new_vma)
2856			goto out;
2857		*new_vma = *vma;
2858		new_vma->vm_start = addr;
2859		new_vma->vm_end = addr + len;
2860		new_vma->vm_pgoff = pgoff;
2861		if (vma_dup_policy(vma, new_vma))
2862			goto out_free_vma;
2863		INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2864		if (anon_vma_clone(new_vma, vma))
2865			goto out_free_mempol;
2866		if (new_vma->vm_file)
2867			get_file(new_vma->vm_file);
2868		if (new_vma->vm_ops && new_vma->vm_ops->open)
2869			new_vma->vm_ops->open(new_vma);
2870		vma_link(mm, new_vma, prev, rb_link, rb_parent);
2871		*need_rmap_locks = false;
2872	}
2873	return new_vma;
2874
2875out_free_mempol:
2876	mpol_put(vma_policy(new_vma));
2877out_free_vma:
2878	kmem_cache_free(vm_area_cachep, new_vma);
2879out:
2880	return NULL;
2881}
2882
2883/*
2884 * Return true if the calling process may expand its vm space by the passed
2885 * number of pages
2886 */
2887bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
2888{
2889	if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
2890		return false;
2891
2892	if (is_data_mapping(flags) &&
2893	    mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
2894		if (ignore_rlimit_data)
2895			pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Will be forbidden soon.\n",
2896				     current->comm, current->pid,
2897				     (mm->data_vm + npages) << PAGE_SHIFT,
2898				     rlimit(RLIMIT_DATA));
2899		else
 
 
 
 
 
 
2900			return false;
2901	}
2902
2903	return true;
2904}
2905
2906void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
2907{
2908	mm->total_vm += npages;
2909
2910	if (is_exec_mapping(flags))
2911		mm->exec_vm += npages;
2912	else if (is_stack_mapping(flags))
2913		mm->stack_vm += npages;
2914	else if (is_data_mapping(flags))
2915		mm->data_vm += npages;
2916}
2917
2918static int special_mapping_fault(struct vm_area_struct *vma,
2919				 struct vm_fault *vmf);
2920
2921/*
2922 * Having a close hook prevents vma merging regardless of flags.
2923 */
2924static void special_mapping_close(struct vm_area_struct *vma)
2925{
2926}
2927
2928static const char *special_mapping_name(struct vm_area_struct *vma)
2929{
2930	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
2931}
2932
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2933static const struct vm_operations_struct special_mapping_vmops = {
2934	.close = special_mapping_close,
2935	.fault = special_mapping_fault,
 
2936	.name = special_mapping_name,
 
 
 
2937};
2938
2939static const struct vm_operations_struct legacy_special_mapping_vmops = {
2940	.close = special_mapping_close,
2941	.fault = special_mapping_fault,
2942};
2943
2944static int special_mapping_fault(struct vm_area_struct *vma,
2945				struct vm_fault *vmf)
2946{
 
2947	pgoff_t pgoff;
2948	struct page **pages;
2949
2950	if (vma->vm_ops == &legacy_special_mapping_vmops) {
2951		pages = vma->vm_private_data;
2952	} else {
2953		struct vm_special_mapping *sm = vma->vm_private_data;
2954
2955		if (sm->fault)
2956			return sm->fault(sm, vma, vmf);
2957
2958		pages = sm->pages;
2959	}
2960
2961	for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
2962		pgoff--;
2963
2964	if (*pages) {
2965		struct page *page = *pages;
2966		get_page(page);
2967		vmf->page = page;
2968		return 0;
2969	}
2970
2971	return VM_FAULT_SIGBUS;
2972}
2973
2974static struct vm_area_struct *__install_special_mapping(
2975	struct mm_struct *mm,
2976	unsigned long addr, unsigned long len,
2977	unsigned long vm_flags, void *priv,
2978	const struct vm_operations_struct *ops)
2979{
2980	int ret;
2981	struct vm_area_struct *vma;
2982
2983	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2984	if (unlikely(vma == NULL))
2985		return ERR_PTR(-ENOMEM);
2986
2987	INIT_LIST_HEAD(&vma->anon_vma_chain);
2988	vma->vm_mm = mm;
2989	vma->vm_start = addr;
2990	vma->vm_end = addr + len;
2991
2992	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
2993	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2994
2995	vma->vm_ops = ops;
2996	vma->vm_private_data = priv;
2997
2998	ret = insert_vm_struct(mm, vma);
2999	if (ret)
3000		goto out;
3001
3002	vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3003
3004	perf_event_mmap(vma);
3005
3006	return vma;
3007
3008out:
3009	kmem_cache_free(vm_area_cachep, vma);
3010	return ERR_PTR(ret);
3011}
3012
 
 
 
 
 
 
 
 
3013/*
3014 * Called with mm->mmap_sem held for writing.
3015 * Insert a new vma covering the given region, with the given flags.
3016 * Its pages are supplied by the given array of struct page *.
3017 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3018 * The region past the last page supplied will always produce SIGBUS.
3019 * The array pointer and the pages it points to are assumed to stay alive
3020 * for as long as this mapping might exist.
3021 */
3022struct vm_area_struct *_install_special_mapping(
3023	struct mm_struct *mm,
3024	unsigned long addr, unsigned long len,
3025	unsigned long vm_flags, const struct vm_special_mapping *spec)
3026{
3027	return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3028					&special_mapping_vmops);
3029}
3030
3031int install_special_mapping(struct mm_struct *mm,
3032			    unsigned long addr, unsigned long len,
3033			    unsigned long vm_flags, struct page **pages)
3034{
3035	struct vm_area_struct *vma = __install_special_mapping(
3036		mm, addr, len, vm_flags, (void *)pages,
3037		&legacy_special_mapping_vmops);
3038
3039	return PTR_ERR_OR_ZERO(vma);
3040}
3041
3042static DEFINE_MUTEX(mm_all_locks_mutex);
3043
3044static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3045{
3046	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3047		/*
3048		 * The LSB of head.next can't change from under us
3049		 * because we hold the mm_all_locks_mutex.
3050		 */
3051		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3052		/*
3053		 * We can safely modify head.next after taking the
3054		 * anon_vma->root->rwsem. If some other vma in this mm shares
3055		 * the same anon_vma we won't take it again.
3056		 *
3057		 * No need of atomic instructions here, head.next
3058		 * can't change from under us thanks to the
3059		 * anon_vma->root->rwsem.
3060		 */
3061		if (__test_and_set_bit(0, (unsigned long *)
3062				       &anon_vma->root->rb_root.rb_node))
3063			BUG();
3064	}
3065}
3066
3067static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3068{
3069	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3070		/*
3071		 * AS_MM_ALL_LOCKS can't change from under us because
3072		 * we hold the mm_all_locks_mutex.
3073		 *
3074		 * Operations on ->flags have to be atomic because
3075		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3076		 * mm_all_locks_mutex, there may be other cpus
3077		 * changing other bitflags in parallel to us.
3078		 */
3079		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3080			BUG();
3081		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3082	}
3083}
3084
3085/*
3086 * This operation locks against the VM for all pte/vma/mm related
3087 * operations that could ever happen on a certain mm. This includes
3088 * vmtruncate, try_to_unmap, and all page faults.
3089 *
3090 * The caller must take the mmap_sem in write mode before calling
3091 * mm_take_all_locks(). The caller isn't allowed to release the
3092 * mmap_sem until mm_drop_all_locks() returns.
3093 *
3094 * mmap_sem in write mode is required in order to block all operations
3095 * that could modify pagetables and free pages without need of
3096 * altering the vma layout. It's also needed in write mode to avoid new
3097 * anon_vmas to be associated with existing vmas.
3098 *
3099 * A single task can't take more than one mm_take_all_locks() in a row
3100 * or it would deadlock.
3101 *
3102 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3103 * mapping->flags avoid to take the same lock twice, if more than one
3104 * vma in this mm is backed by the same anon_vma or address_space.
3105 *
3106 * We take locks in following order, accordingly to comment at beginning
3107 * of mm/rmap.c:
3108 *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3109 *     hugetlb mapping);
3110 *   - all i_mmap_rwsem locks;
3111 *   - all anon_vma->rwseml
3112 *
3113 * We can take all locks within these types randomly because the VM code
3114 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3115 * mm_all_locks_mutex.
3116 *
3117 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3118 * that may have to take thousand of locks.
3119 *
3120 * mm_take_all_locks() can fail if it's interrupted by signals.
3121 */
3122int mm_take_all_locks(struct mm_struct *mm)
3123{
3124	struct vm_area_struct *vma;
3125	struct anon_vma_chain *avc;
3126
3127	BUG_ON(down_read_trylock(&mm->mmap_sem));
3128
3129	mutex_lock(&mm_all_locks_mutex);
3130
3131	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3132		if (signal_pending(current))
3133			goto out_unlock;
3134		if (vma->vm_file && vma->vm_file->f_mapping &&
3135				is_vm_hugetlb_page(vma))
3136			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3137	}
3138
3139	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3140		if (signal_pending(current))
3141			goto out_unlock;
3142		if (vma->vm_file && vma->vm_file->f_mapping &&
3143				!is_vm_hugetlb_page(vma))
3144			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3145	}
3146
3147	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3148		if (signal_pending(current))
3149			goto out_unlock;
3150		if (vma->anon_vma)
3151			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3152				vm_lock_anon_vma(mm, avc->anon_vma);
3153	}
3154
3155	return 0;
3156
3157out_unlock:
3158	mm_drop_all_locks(mm);
3159	return -EINTR;
3160}
3161
3162static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3163{
3164	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3165		/*
3166		 * The LSB of head.next can't change to 0 from under
3167		 * us because we hold the mm_all_locks_mutex.
3168		 *
3169		 * We must however clear the bitflag before unlocking
3170		 * the vma so the users using the anon_vma->rb_root will
3171		 * never see our bitflag.
3172		 *
3173		 * No need of atomic instructions here, head.next
3174		 * can't change from under us until we release the
3175		 * anon_vma->root->rwsem.
3176		 */
3177		if (!__test_and_clear_bit(0, (unsigned long *)
3178					  &anon_vma->root->rb_root.rb_node))
3179			BUG();
3180		anon_vma_unlock_write(anon_vma);
3181	}
3182}
3183
3184static void vm_unlock_mapping(struct address_space *mapping)
3185{
3186	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3187		/*
3188		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3189		 * because we hold the mm_all_locks_mutex.
3190		 */
3191		i_mmap_unlock_write(mapping);
3192		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3193					&mapping->flags))
3194			BUG();
3195	}
3196}
3197
3198/*
3199 * The mmap_sem cannot be released by the caller until
3200 * mm_drop_all_locks() returns.
3201 */
3202void mm_drop_all_locks(struct mm_struct *mm)
3203{
3204	struct vm_area_struct *vma;
3205	struct anon_vma_chain *avc;
3206
3207	BUG_ON(down_read_trylock(&mm->mmap_sem));
3208	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3209
3210	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3211		if (vma->anon_vma)
3212			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3213				vm_unlock_anon_vma(avc->anon_vma);
3214		if (vma->vm_file && vma->vm_file->f_mapping)
3215			vm_unlock_mapping(vma->vm_file->f_mapping);
3216	}
3217
3218	mutex_unlock(&mm_all_locks_mutex);
3219}
3220
3221/*
3222 * initialise the VMA slab
3223 */
3224void __init mmap_init(void)
3225{
3226	int ret;
3227
3228	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3229	VM_BUG_ON(ret);
3230}
3231
3232/*
3233 * Initialise sysctl_user_reserve_kbytes.
3234 *
3235 * This is intended to prevent a user from starting a single memory hogging
3236 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3237 * mode.
3238 *
3239 * The default value is min(3% of free memory, 128MB)
3240 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3241 */
3242static int init_user_reserve(void)
3243{
3244	unsigned long free_kbytes;
3245
3246	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3247
3248	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3249	return 0;
3250}
3251subsys_initcall(init_user_reserve);
3252
3253/*
3254 * Initialise sysctl_admin_reserve_kbytes.
3255 *
3256 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3257 * to log in and kill a memory hogging process.
3258 *
3259 * Systems with more than 256MB will reserve 8MB, enough to recover
3260 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3261 * only reserve 3% of free pages by default.
3262 */
3263static int init_admin_reserve(void)
3264{
3265	unsigned long free_kbytes;
3266
3267	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3268
3269	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3270	return 0;
3271}
3272subsys_initcall(init_admin_reserve);
3273
3274/*
3275 * Reinititalise user and admin reserves if memory is added or removed.
3276 *
3277 * The default user reserve max is 128MB, and the default max for the
3278 * admin reserve is 8MB. These are usually, but not always, enough to
3279 * enable recovery from a memory hogging process using login/sshd, a shell,
3280 * and tools like top. It may make sense to increase or even disable the
3281 * reserve depending on the existence of swap or variations in the recovery
3282 * tools. So, the admin may have changed them.
3283 *
3284 * If memory is added and the reserves have been eliminated or increased above
3285 * the default max, then we'll trust the admin.
3286 *
3287 * If memory is removed and there isn't enough free memory, then we
3288 * need to reset the reserves.
3289 *
3290 * Otherwise keep the reserve set by the admin.
3291 */
3292static int reserve_mem_notifier(struct notifier_block *nb,
3293			     unsigned long action, void *data)
3294{
3295	unsigned long tmp, free_kbytes;
3296
3297	switch (action) {
3298	case MEM_ONLINE:
3299		/* Default max is 128MB. Leave alone if modified by operator. */
3300		tmp = sysctl_user_reserve_kbytes;
3301		if (0 < tmp && tmp < (1UL << 17))
3302			init_user_reserve();
3303
3304		/* Default max is 8MB.  Leave alone if modified by operator. */
3305		tmp = sysctl_admin_reserve_kbytes;
3306		if (0 < tmp && tmp < (1UL << 13))
3307			init_admin_reserve();
3308
3309		break;
3310	case MEM_OFFLINE:
3311		free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3312
3313		if (sysctl_user_reserve_kbytes > free_kbytes) {
3314			init_user_reserve();
3315			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3316				sysctl_user_reserve_kbytes);
3317		}
3318
3319		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3320			init_admin_reserve();
3321			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3322				sysctl_admin_reserve_kbytes);
3323		}
3324		break;
3325	default:
3326		break;
3327	}
3328	return NOTIFY_OK;
3329}
3330
3331static struct notifier_block reserve_mem_nb = {
3332	.notifier_call = reserve_mem_notifier,
3333};
3334
3335static int __meminit init_reserve_notifier(void)
3336{
3337	if (register_hotmemory_notifier(&reserve_mem_nb))
3338		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3339
3340	return 0;
3341}
3342subsys_initcall(init_reserve_notifier);
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * mm/mmap.c
   4 *
   5 * Written by obz.
   6 *
   7 * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
   8 */
   9
  10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12#include <linux/kernel.h>
  13#include <linux/slab.h>
  14#include <linux/backing-dev.h>
  15#include <linux/mm.h>
  16#include <linux/vmacache.h>
  17#include <linux/shm.h>
  18#include <linux/mman.h>
  19#include <linux/pagemap.h>
  20#include <linux/swap.h>
  21#include <linux/syscalls.h>
  22#include <linux/capability.h>
  23#include <linux/init.h>
  24#include <linux/file.h>
  25#include <linux/fs.h>
  26#include <linux/personality.h>
  27#include <linux/security.h>
  28#include <linux/hugetlb.h>
  29#include <linux/shmem_fs.h>
  30#include <linux/profile.h>
  31#include <linux/export.h>
  32#include <linux/mount.h>
  33#include <linux/mempolicy.h>
  34#include <linux/rmap.h>
  35#include <linux/mmu_notifier.h>
  36#include <linux/mmdebug.h>
  37#include <linux/perf_event.h>
  38#include <linux/audit.h>
  39#include <linux/khugepaged.h>
  40#include <linux/uprobes.h>
  41#include <linux/rbtree_augmented.h>
  42#include <linux/notifier.h>
  43#include <linux/memory.h>
  44#include <linux/printk.h>
  45#include <linux/userfaultfd_k.h>
  46#include <linux/moduleparam.h>
  47#include <linux/pkeys.h>
  48#include <linux/oom.h>
  49#include <linux/sched/mm.h>
  50
  51#include <linux/uaccess.h>
  52#include <asm/cacheflush.h>
  53#include <asm/tlb.h>
  54#include <asm/mmu_context.h>
  55
  56#define CREATE_TRACE_POINTS
  57#include <trace/events/mmap.h>
  58
  59#include "internal.h"
  60
  61#ifndef arch_mmap_check
  62#define arch_mmap_check(addr, len, flags)	(0)
  63#endif
  64
 
 
 
 
  65#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
  66const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
  67const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
  68int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
  69#endif
  70#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
  71const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
  72const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
  73int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
  74#endif
  75
  76static bool ignore_rlimit_data;
  77core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
  78
  79static void unmap_region(struct mm_struct *mm,
  80		struct vm_area_struct *vma, struct vm_area_struct *prev,
  81		unsigned long start, unsigned long end);
  82
  83/* description of effects of mapping type and prot in current implementation.
  84 * this is due to the limited x86 page protection hardware.  The expected
  85 * behavior is in parens:
  86 *
  87 * map_type	prot
  88 *		PROT_NONE	PROT_READ	PROT_WRITE	PROT_EXEC
  89 * MAP_SHARED	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  90 *		w: (no) no	w: (no) no	w: (yes) yes	w: (no) no
  91 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  92 *
  93 * MAP_PRIVATE	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  94 *		w: (no) no	w: (no) no	w: (copy) copy	w: (no) no
  95 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  96 *
  97 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
  98 * MAP_PRIVATE (with Enhanced PAN supported):
  99 *								r: (no) no
 100 *								w: (no) no
 101 *								x: (yes) yes
 102 */
 103pgprot_t protection_map[16] __ro_after_init = {
 104	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
 105	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
 106};
 107
 108#ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
 109static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
 110{
 111	return prot;
 112}
 113#endif
 114
 115pgprot_t vm_get_page_prot(unsigned long vm_flags)
 116{
 117	pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
 118				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
 119			pgprot_val(arch_vm_get_page_prot(vm_flags)));
 120
 121	return arch_filter_pgprot(ret);
 122}
 123EXPORT_SYMBOL(vm_get_page_prot);
 124
 125static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
 126{
 127	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
 128}
 129
 130/* Update vma->vm_page_prot to reflect vma->vm_flags. */
 131void vma_set_page_prot(struct vm_area_struct *vma)
 132{
 133	unsigned long vm_flags = vma->vm_flags;
 134	pgprot_t vm_page_prot;
 135
 136	vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
 137	if (vma_wants_writenotify(vma, vm_page_prot)) {
 138		vm_flags &= ~VM_SHARED;
 139		vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
 
 140	}
 141	/* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
 142	WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
 143}
 144
 145/*
 146 * Requires inode->i_mapping->i_mmap_rwsem
 147 */
 148static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 149		struct file *file, struct address_space *mapping)
 150{
 151	if (vma->vm_flags & VM_DENYWRITE)
 152		allow_write_access(file);
 153	if (vma->vm_flags & VM_SHARED)
 154		mapping_unmap_writable(mapping);
 155
 156	flush_dcache_mmap_lock(mapping);
 157	vma_interval_tree_remove(vma, &mapping->i_mmap);
 158	flush_dcache_mmap_unlock(mapping);
 159}
 160
 161/*
 162 * Unlink a file-based vm structure from its interval tree, to hide
 163 * vma from rmap and vmtruncate before freeing its page tables.
 164 */
 165void unlink_file_vma(struct vm_area_struct *vma)
 166{
 167	struct file *file = vma->vm_file;
 168
 169	if (file) {
 170		struct address_space *mapping = file->f_mapping;
 171		i_mmap_lock_write(mapping);
 172		__remove_shared_vm_struct(vma, file, mapping);
 173		i_mmap_unlock_write(mapping);
 174	}
 175}
 176
 177/*
 178 * Close a vm structure and free it, returning the next.
 179 */
 180static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
 181{
 182	struct vm_area_struct *next = vma->vm_next;
 183
 184	might_sleep();
 185	if (vma->vm_ops && vma->vm_ops->close)
 186		vma->vm_ops->close(vma);
 187	if (vma->vm_file)
 188		fput(vma->vm_file);
 189	mpol_put(vma_policy(vma));
 190	vm_area_free(vma);
 191	return next;
 192}
 193
 194static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
 195		struct list_head *uf);
 196SYSCALL_DEFINE1(brk, unsigned long, brk)
 197{
 198	unsigned long newbrk, oldbrk, origbrk;
 
 199	struct mm_struct *mm = current->mm;
 200	struct vm_area_struct *next;
 201	unsigned long min_brk;
 202	bool populate;
 203	bool downgraded = false;
 204	LIST_HEAD(uf);
 205
 206	if (mmap_write_lock_killable(mm))
 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_lock to read.
 247	 */
 248	if (brk <= mm->brk) {
 249		int ret;
 250
 251		/*
 252		 * mm->brk must to be protected by write mmap_lock so update it
 253		 * before downgrading mmap_lock. 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
 277success:
 278	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
 279	if (downgraded)
 280		mmap_read_unlock(mm);
 281	else
 282		mmap_write_unlock(mm);
 283	userfaultfd_unmap_complete(mm, &uf);
 284	if (populate)
 285		mm_populate(oldbrk, newbrk - oldbrk);
 286	return brk;
 287
 288out:
 289	mmap_write_unlock(mm);
 290	return origbrk;
 
 291}
 292
 293static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
 294{
 295	unsigned long gap, prev_end;
 296
 297	/*
 298	 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
 299	 * allow two stack_guard_gaps between them here, and when choosing
 300	 * an unmapped area; whereas when expanding we only require one.
 301	 * That's a little inconsistent, but keeps the code here simpler.
 302	 */
 303	gap = vm_start_gap(vma);
 304	if (vma->vm_prev) {
 305		prev_end = vm_end_gap(vma->vm_prev);
 306		if (gap > prev_end)
 307			gap -= prev_end;
 308		else
 309			gap = 0;
 310	}
 311	return gap;
 312}
 313
 314#ifdef CONFIG_DEBUG_VM_RB
 315static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
 316{
 317	unsigned long max = vma_compute_gap(vma), subtree_gap;
 318	if (vma->vm_rb.rb_left) {
 319		subtree_gap = rb_entry(vma->vm_rb.rb_left,
 320				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 321		if (subtree_gap > max)
 322			max = subtree_gap;
 323	}
 324	if (vma->vm_rb.rb_right) {
 325		subtree_gap = rb_entry(vma->vm_rb.rb_right,
 326				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 327		if (subtree_gap > max)
 328			max = subtree_gap;
 329	}
 330	return max;
 331}
 332
 
 333static int browse_rb(struct mm_struct *mm)
 334{
 335	struct rb_root *root = &mm->mm_rb;
 336	int i = 0, j, bug = 0;
 337	struct rb_node *nd, *pn = NULL;
 338	unsigned long prev = 0, pend = 0;
 339
 340	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 341		struct vm_area_struct *vma;
 342		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 343		if (vma->vm_start < prev) {
 344			pr_emerg("vm_start %lx < prev %lx\n",
 345				  vma->vm_start, prev);
 346			bug = 1;
 347		}
 348		if (vma->vm_start < pend) {
 349			pr_emerg("vm_start %lx < pend %lx\n",
 350				  vma->vm_start, pend);
 351			bug = 1;
 352		}
 353		if (vma->vm_start > vma->vm_end) {
 354			pr_emerg("vm_start %lx > vm_end %lx\n",
 355				  vma->vm_start, vma->vm_end);
 356			bug = 1;
 357		}
 358		spin_lock(&mm->page_table_lock);
 359		if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
 360			pr_emerg("free gap %lx, correct %lx\n",
 361			       vma->rb_subtree_gap,
 362			       vma_compute_subtree_gap(vma));
 363			bug = 1;
 364		}
 365		spin_unlock(&mm->page_table_lock);
 366		i++;
 367		pn = nd;
 368		prev = vma->vm_start;
 369		pend = vma->vm_end;
 370	}
 371	j = 0;
 372	for (nd = pn; nd; nd = rb_prev(nd))
 373		j++;
 374	if (i != j) {
 375		pr_emerg("backwards %d, forwards %d\n", j, i);
 376		bug = 1;
 377	}
 378	return bug ? -1 : i;
 379}
 380
 381static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
 382{
 383	struct rb_node *nd;
 384
 385	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 386		struct vm_area_struct *vma;
 387		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 388		VM_BUG_ON_VMA(vma != ignore &&
 389			vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
 390			vma);
 391	}
 392}
 393
 394static void validate_mm(struct mm_struct *mm)
 395{
 396	int bug = 0;
 397	int i = 0;
 398	unsigned long highest_address = 0;
 399	struct vm_area_struct *vma = mm->mmap;
 400
 401	while (vma) {
 402		struct anon_vma *anon_vma = vma->anon_vma;
 403		struct anon_vma_chain *avc;
 404
 405		if (anon_vma) {
 406			anon_vma_lock_read(anon_vma);
 407			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 408				anon_vma_interval_tree_verify(avc);
 409			anon_vma_unlock_read(anon_vma);
 410		}
 411
 412		highest_address = vm_end_gap(vma);
 413		vma = vma->vm_next;
 414		i++;
 415	}
 416	if (i != mm->map_count) {
 417		pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
 418		bug = 1;
 419	}
 420	if (highest_address != mm->highest_vm_end) {
 421		pr_emerg("mm->highest_vm_end %lx, found %lx\n",
 422			  mm->highest_vm_end, highest_address);
 423		bug = 1;
 424	}
 425	i = browse_rb(mm);
 426	if (i != mm->map_count) {
 427		if (i != -1)
 428			pr_emerg("map_count %d rb %d\n", mm->map_count, i);
 429		bug = 1;
 430	}
 431	VM_BUG_ON_MM(bug, mm);
 432}
 433#else
 434#define validate_mm_rb(root, ignore) do { } while (0)
 435#define validate_mm(mm) do { } while (0)
 436#endif
 437
 438RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
 439			 struct vm_area_struct, vm_rb,
 440			 unsigned long, rb_subtree_gap, vma_compute_gap)
 441
 442/*
 443 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
 444 * vma->vm_prev->vm_end values changed, without modifying the vma's position
 445 * in the rbtree.
 446 */
 447static void vma_gap_update(struct vm_area_struct *vma)
 448{
 449	/*
 450	 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
 451	 * a callback function that does exactly what we want.
 452	 */
 453	vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
 454}
 455
 456static inline void vma_rb_insert(struct vm_area_struct *vma,
 457				 struct rb_root *root)
 458{
 459	/* All rb_subtree_gap values must be consistent prior to insertion */
 460	validate_mm_rb(root, NULL);
 461
 462	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 463}
 464
 465static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
 466{
 467	/*
 
 
 
 
 
 
 468	 * Note rb_erase_augmented is a fairly large inline function,
 469	 * so make sure we instantiate it only once with our desired
 470	 * augmented rbtree callbacks.
 471	 */
 472	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 473}
 474
 475static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
 476						struct rb_root *root,
 477						struct vm_area_struct *ignore)
 478{
 479	/*
 480	 * All rb_subtree_gap values must be consistent prior to erase,
 481	 * with the possible exception of
 482	 *
 483	 * a. the "next" vma being erased if next->vm_start was reduced in
 484	 *    __vma_adjust() -> __vma_unlink()
 485	 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
 486	 *    vma_rb_erase()
 487	 */
 488	validate_mm_rb(root, ignore);
 489
 490	__vma_rb_erase(vma, root);
 491}
 492
 493static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
 494					 struct rb_root *root)
 495{
 496	vma_rb_erase_ignore(vma, root, vma);
 497}
 498
 499/*
 500 * vma has some anon_vma assigned, and is already inserted on that
 501 * anon_vma's interval trees.
 502 *
 503 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
 504 * vma must be removed from the anon_vma's interval trees using
 505 * anon_vma_interval_tree_pre_update_vma().
 506 *
 507 * After the update, the vma will be reinserted using
 508 * anon_vma_interval_tree_post_update_vma().
 509 *
 510 * The entire update must be protected by exclusive mmap_lock and by
 511 * the root anon_vma's mutex.
 512 */
 513static inline void
 514anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
 515{
 516	struct anon_vma_chain *avc;
 517
 518	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 519		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
 520}
 521
 522static inline void
 523anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
 524{
 525	struct anon_vma_chain *avc;
 526
 527	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 528		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
 529}
 530
 531static int find_vma_links(struct mm_struct *mm, unsigned long addr,
 532		unsigned long end, struct vm_area_struct **pprev,
 533		struct rb_node ***rb_link, struct rb_node **rb_parent)
 534{
 535	struct rb_node **__rb_link, *__rb_parent, *rb_prev;
 536
 537	__rb_link = &mm->mm_rb.rb_node;
 538	rb_prev = __rb_parent = NULL;
 539
 540	while (*__rb_link) {
 541		struct vm_area_struct *vma_tmp;
 542
 543		__rb_parent = *__rb_link;
 544		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
 545
 546		if (vma_tmp->vm_end > addr) {
 547			/* Fail if an existing vma overlaps the area */
 548			if (vma_tmp->vm_start < end)
 549				return -ENOMEM;
 550			__rb_link = &__rb_parent->rb_left;
 551		} else {
 552			rb_prev = __rb_parent;
 553			__rb_link = &__rb_parent->rb_right;
 554		}
 555	}
 556
 557	*pprev = NULL;
 558	if (rb_prev)
 559		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 560	*rb_link = __rb_link;
 561	*rb_parent = __rb_parent;
 562	return 0;
 563}
 564
 565/*
 566 * vma_next() - Get the next VMA.
 567 * @mm: The mm_struct.
 568 * @vma: The current vma.
 569 *
 570 * If @vma is NULL, return the first vma in the mm.
 571 *
 572 * Returns: The next VMA after @vma.
 573 */
 574static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
 575					 struct vm_area_struct *vma)
 576{
 577	if (!vma)
 578		return mm->mmap;
 579
 580	return vma->vm_next;
 581}
 582
 583/*
 584 * munmap_vma_range() - munmap VMAs that overlap a range.
 585 * @mm: The mm struct
 586 * @start: The start of the range.
 587 * @len: The length of the range.
 588 * @pprev: pointer to the pointer that will be set to previous vm_area_struct
 589 * @rb_link: the rb_node
 590 * @rb_parent: the parent rb_node
 591 *
 592 * Find all the vm_area_struct that overlap from @start to
 593 * @end and munmap them.  Set @pprev to the previous vm_area_struct.
 594 *
 595 * Returns: -ENOMEM on munmap failure or 0 on success.
 596 */
 597static inline int
 598munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
 599		 struct vm_area_struct **pprev, struct rb_node ***link,
 600		 struct rb_node **parent, struct list_head *uf)
 601{
 602
 603	while (find_vma_links(mm, start, start + len, pprev, link, parent))
 604		if (do_munmap(mm, start, len, uf))
 605			return -ENOMEM;
 606
 607	return 0;
 608}
 609static unsigned long count_vma_pages_range(struct mm_struct *mm,
 610		unsigned long addr, unsigned long end)
 611{
 612	unsigned long nr_pages = 0;
 613	struct vm_area_struct *vma;
 614
 615	/* Find first overlapping mapping */
 616	vma = find_vma_intersection(mm, addr, end);
 617	if (!vma)
 618		return 0;
 619
 620	nr_pages = (min(end, vma->vm_end) -
 621		max(addr, vma->vm_start)) >> PAGE_SHIFT;
 622
 623	/* Iterate over the rest of the overlaps */
 624	for (vma = vma->vm_next; vma; vma = vma->vm_next) {
 625		unsigned long overlap_len;
 626
 627		if (vma->vm_start > end)
 628			break;
 629
 630		overlap_len = min(end, vma->vm_end) - vma->vm_start;
 631		nr_pages += overlap_len >> PAGE_SHIFT;
 632	}
 633
 634	return nr_pages;
 635}
 636
 637void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 638		struct rb_node **rb_link, struct rb_node *rb_parent)
 639{
 640	/* Update tracking information for the gap following the new vma. */
 641	if (vma->vm_next)
 642		vma_gap_update(vma->vm_next);
 643	else
 644		mm->highest_vm_end = vm_end_gap(vma);
 645
 646	/*
 647	 * vma->vm_prev wasn't known when we followed the rbtree to find the
 648	 * correct insertion point for that vma. As a result, we could not
 649	 * update the vma vm_rb parents rb_subtree_gap values on the way down.
 650	 * So, we first insert the vma with a zero rb_subtree_gap value
 651	 * (to be consistent with what we did on the way down), and then
 652	 * immediately update the gap to the correct value. Finally we
 653	 * rebalance the rbtree after all augmented values have been set.
 654	 */
 655	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 656	vma->rb_subtree_gap = 0;
 657	vma_gap_update(vma);
 658	vma_rb_insert(vma, &mm->mm_rb);
 659}
 660
 661static void __vma_link_file(struct vm_area_struct *vma)
 662{
 663	struct file *file;
 664
 665	file = vma->vm_file;
 666	if (file) {
 667		struct address_space *mapping = file->f_mapping;
 668
 669		if (vma->vm_flags & VM_DENYWRITE)
 670			put_write_access(file_inode(file));
 671		if (vma->vm_flags & VM_SHARED)
 672			mapping_allow_writable(mapping);
 673
 674		flush_dcache_mmap_lock(mapping);
 675		vma_interval_tree_insert(vma, &mapping->i_mmap);
 676		flush_dcache_mmap_unlock(mapping);
 677	}
 678}
 679
 680static void
 681__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 682	struct vm_area_struct *prev, struct rb_node **rb_link,
 683	struct rb_node *rb_parent)
 684{
 685	__vma_link_list(mm, vma, prev);
 686	__vma_link_rb(mm, vma, rb_link, rb_parent);
 687}
 688
 689static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 690			struct vm_area_struct *prev, struct rb_node **rb_link,
 691			struct rb_node *rb_parent)
 692{
 693	struct address_space *mapping = NULL;
 694
 695	if (vma->vm_file) {
 696		mapping = vma->vm_file->f_mapping;
 697		i_mmap_lock_write(mapping);
 698	}
 699
 700	__vma_link(mm, vma, prev, rb_link, rb_parent);
 701	__vma_link_file(vma);
 702
 703	if (mapping)
 704		i_mmap_unlock_write(mapping);
 705
 706	mm->map_count++;
 707	validate_mm(mm);
 708}
 709
 710/*
 711 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
 712 * mm's list and rbtree.  It has already been inserted into the interval tree.
 713 */
 714static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 715{
 716	struct vm_area_struct *prev;
 717	struct rb_node **rb_link, *rb_parent;
 718
 719	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
 720			   &prev, &rb_link, &rb_parent))
 721		BUG();
 722	__vma_link(mm, vma, prev, rb_link, rb_parent);
 723	mm->map_count++;
 724}
 725
 726static __always_inline void __vma_unlink(struct mm_struct *mm,
 727						struct vm_area_struct *vma,
 728						struct vm_area_struct *ignore)
 729{
 730	vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
 731	__vma_unlink_list(mm, vma);
 
 
 
 
 
 732	/* Kill the cache */
 733	vmacache_invalidate(mm);
 734}
 735
 736/*
 737 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
 738 * is already present in an i_mmap tree without adjusting the tree.
 739 * The following helper function should be used when such adjustments
 740 * are necessary.  The "insert" vma (if any) is to be inserted
 741 * before we drop the necessary locks.
 742 */
 743int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
 744	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
 745	struct vm_area_struct *expand)
 746{
 747	struct mm_struct *mm = vma->vm_mm;
 748	struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
 
 749	struct address_space *mapping = NULL;
 750	struct rb_root_cached *root = NULL;
 751	struct anon_vma *anon_vma = NULL;
 752	struct file *file = vma->vm_file;
 753	bool start_changed = false, end_changed = false;
 754	long adjust_next = 0;
 755	int remove_next = 0;
 756
 757	if (next && !insert) {
 758		struct vm_area_struct *exporter = NULL, *importer = NULL;
 759
 760		if (end >= next->vm_end) {
 761			/*
 762			 * vma expands, overlapping all the next, and
 763			 * perhaps the one after too (mprotect case 6).
 764			 * The only other cases that gets here are
 765			 * case 1, case 7 and case 8.
 766			 */
 767			if (next == expand) {
 768				/*
 769				 * The only case where we don't expand "vma"
 770				 * and we expand "next" instead is case 8.
 771				 */
 772				VM_WARN_ON(end != next->vm_end);
 773				/*
 774				 * remove_next == 3 means we're
 775				 * removing "vma" and that to do so we
 776				 * swapped "vma" and "next".
 777				 */
 778				remove_next = 3;
 779				VM_WARN_ON(file != next->vm_file);
 780				swap(vma, next);
 781			} else {
 782				VM_WARN_ON(expand != vma);
 783				/*
 784				 * case 1, 6, 7, remove_next == 2 is case 6,
 785				 * remove_next == 1 is case 1 or 7.
 786				 */
 787				remove_next = 1 + (end > next->vm_end);
 788				VM_WARN_ON(remove_next == 2 &&
 789					   end != next->vm_next->vm_end);
 790				/* trim end to next, for case 6 first pass */
 791				end = next->vm_end;
 792			}
 793
 794			exporter = next;
 795			importer = vma;
 796
 797			/*
 798			 * If next doesn't have anon_vma, import from vma after
 799			 * next, if the vma overlaps with it.
 800			 */
 801			if (remove_next == 2 && !next->anon_vma)
 802				exporter = next->vm_next;
 803
 804		} else if (end > next->vm_start) {
 805			/*
 806			 * vma expands, overlapping part of the next:
 807			 * mprotect case 5 shifting the boundary up.
 808			 */
 809			adjust_next = (end - next->vm_start);
 810			exporter = next;
 811			importer = vma;
 812			VM_WARN_ON(expand != importer);
 813		} else if (end < vma->vm_end) {
 814			/*
 815			 * vma shrinks, and !insert tells it's not
 816			 * split_vma inserting another: so it must be
 817			 * mprotect case 4 shifting the boundary down.
 818			 */
 819			adjust_next = -(vma->vm_end - end);
 820			exporter = vma;
 821			importer = next;
 822			VM_WARN_ON(expand != importer);
 823		}
 824
 825		/*
 826		 * Easily overlooked: when mprotect shifts the boundary,
 827		 * make sure the expanding vma has anon_vma set if the
 828		 * shrinking vma had, to cover any anon pages imported.
 829		 */
 830		if (exporter && exporter->anon_vma && !importer->anon_vma) {
 831			int error;
 832
 833			importer->anon_vma = exporter->anon_vma;
 834			error = anon_vma_clone(importer, exporter);
 835			if (error)
 836				return error;
 837		}
 838	}
 839again:
 840	vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
 841
 842	if (file) {
 843		mapping = file->f_mapping;
 844		root = &mapping->i_mmap;
 845		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
 846
 847		if (adjust_next)
 848			uprobe_munmap(next, next->vm_start, next->vm_end);
 849
 850		i_mmap_lock_write(mapping);
 851		if (insert) {
 852			/*
 853			 * Put into interval tree now, so instantiated pages
 854			 * are visible to arm/parisc __flush_dcache_page
 855			 * throughout; but we cannot insert into address
 856			 * space until vma start or end is updated.
 857			 */
 858			__vma_link_file(insert);
 859		}
 860	}
 861
 
 
 862	anon_vma = vma->anon_vma;
 863	if (!anon_vma && adjust_next)
 864		anon_vma = next->anon_vma;
 865	if (anon_vma) {
 866		VM_WARN_ON(adjust_next && next->anon_vma &&
 867			   anon_vma != next->anon_vma);
 868		anon_vma_lock_write(anon_vma);
 869		anon_vma_interval_tree_pre_update_vma(vma);
 870		if (adjust_next)
 871			anon_vma_interval_tree_pre_update_vma(next);
 872	}
 873
 874	if (file) {
 875		flush_dcache_mmap_lock(mapping);
 876		vma_interval_tree_remove(vma, root);
 877		if (adjust_next)
 878			vma_interval_tree_remove(next, root);
 879	}
 880
 881	if (start != vma->vm_start) {
 882		vma->vm_start = start;
 883		start_changed = true;
 884	}
 885	if (end != vma->vm_end) {
 886		vma->vm_end = end;
 887		end_changed = true;
 888	}
 889	vma->vm_pgoff = pgoff;
 890	if (adjust_next) {
 891		next->vm_start += adjust_next;
 892		next->vm_pgoff += adjust_next >> PAGE_SHIFT;
 893	}
 894
 895	if (file) {
 896		if (adjust_next)
 897			vma_interval_tree_insert(next, root);
 898		vma_interval_tree_insert(vma, root);
 899		flush_dcache_mmap_unlock(mapping);
 900	}
 901
 902	if (remove_next) {
 903		/*
 904		 * vma_merge has merged next into vma, and needs
 905		 * us to remove next before dropping the locks.
 906		 */
 907		if (remove_next != 3)
 908			__vma_unlink(mm, next, next);
 909		else
 910			/*
 911			 * vma is not before next if they've been
 912			 * swapped.
 913			 *
 914			 * pre-swap() next->vm_start was reduced so
 915			 * tell validate_mm_rb to ignore pre-swap()
 916			 * "next" (which is stored in post-swap()
 917			 * "vma").
 918			 */
 919			__vma_unlink(mm, next, vma);
 920		if (file)
 921			__remove_shared_vm_struct(next, file, mapping);
 922	} else if (insert) {
 923		/*
 924		 * split_vma has split insert from vma, and needs
 925		 * us to insert it before dropping the locks
 926		 * (it may either follow vma or precede it).
 927		 */
 928		__insert_vm_struct(mm, insert);
 929	} else {
 930		if (start_changed)
 931			vma_gap_update(vma);
 932		if (end_changed) {
 933			if (!next)
 934				mm->highest_vm_end = vm_end_gap(vma);
 935			else if (!adjust_next)
 936				vma_gap_update(next);
 937		}
 938	}
 939
 940	if (anon_vma) {
 941		anon_vma_interval_tree_post_update_vma(vma);
 942		if (adjust_next)
 943			anon_vma_interval_tree_post_update_vma(next);
 944		anon_vma_unlock_write(anon_vma);
 945	}
 
 
 946
 947	if (file) {
 948		i_mmap_unlock_write(mapping);
 949		uprobe_mmap(vma);
 950
 951		if (adjust_next)
 952			uprobe_mmap(next);
 953	}
 954
 955	if (remove_next) {
 956		if (file) {
 957			uprobe_munmap(next, next->vm_start, next->vm_end);
 958			fput(file);
 959		}
 960		if (next->anon_vma)
 961			anon_vma_merge(vma, next);
 962		mm->map_count--;
 963		mpol_put(vma_policy(next));
 964		vm_area_free(next);
 965		/*
 966		 * In mprotect's case 6 (see comments on vma_merge),
 967		 * we must remove another next too. It would clutter
 968		 * up the code too much to do both in one go.
 969		 */
 970		if (remove_next != 3) {
 971			/*
 972			 * If "next" was removed and vma->vm_end was
 973			 * expanded (up) over it, in turn
 974			 * "next->vm_prev->vm_end" changed and the
 975			 * "vma->vm_next" gap must be updated.
 976			 */
 977			next = vma->vm_next;
 978		} else {
 979			/*
 980			 * For the scope of the comment "next" and
 981			 * "vma" considered pre-swap(): if "vma" was
 982			 * removed, next->vm_start was expanded (down)
 983			 * over it and the "next" gap must be updated.
 984			 * Because of the swap() the post-swap() "vma"
 985			 * actually points to pre-swap() "next"
 986			 * (post-swap() "next" as opposed is now a
 987			 * dangling pointer).
 988			 */
 989			next = vma;
 990		}
 991		if (remove_next == 2) {
 992			remove_next = 1;
 993			end = next->vm_end;
 994			goto again;
 995		}
 996		else if (next)
 997			vma_gap_update(next);
 998		else {
 999			/*
1000			 * If remove_next == 2 we obviously can't
1001			 * reach this path.
1002			 *
1003			 * If remove_next == 3 we can't reach this
1004			 * path because pre-swap() next is always not
1005			 * NULL. pre-swap() "next" is not being
1006			 * removed and its next->vm_end is not altered
1007			 * (and furthermore "end" already matches
1008			 * next->vm_end in remove_next == 3).
1009			 *
1010			 * We reach this only in the remove_next == 1
1011			 * case if the "next" vma that was removed was
1012			 * the highest vma of the mm. However in such
1013			 * case next->vm_end == "end" and the extended
1014			 * "vma" has vma->vm_end == next->vm_end so
1015			 * mm->highest_vm_end doesn't need any update
1016			 * in remove_next == 1 case.
1017			 */
1018			VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1019		}
1020	}
1021	if (insert && file)
1022		uprobe_mmap(insert);
1023
1024	validate_mm(mm);
1025
1026	return 0;
1027}
1028
1029/*
1030 * If the vma has a ->close operation then the driver probably needs to release
1031 * per-vma resources, so we don't attempt to merge those.
1032 */
1033static inline int is_mergeable_vma(struct vm_area_struct *vma,
1034				struct file *file, unsigned long vm_flags,
1035				struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1036{
1037	/*
1038	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1039	 * match the flags but dirty bit -- the caller should mark
1040	 * merged VMA as dirty. If dirty bit won't be excluded from
1041	 * comparison, we increase pressure on the memory system forcing
1042	 * the kernel to generate new VMAs when old one could be
1043	 * extended instead.
1044	 */
1045	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1046		return 0;
1047	if (vma->vm_file != file)
1048		return 0;
1049	if (vma->vm_ops && vma->vm_ops->close)
1050		return 0;
1051	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1052		return 0;
1053	return 1;
1054}
1055
1056static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1057					struct anon_vma *anon_vma2,
1058					struct vm_area_struct *vma)
1059{
1060	/*
1061	 * The list_is_singular() test is to avoid merging VMA cloned from
1062	 * parents. This can improve scalability caused by anon_vma lock.
1063	 */
1064	if ((!anon_vma1 || !anon_vma2) && (!vma ||
1065		list_is_singular(&vma->anon_vma_chain)))
1066		return 1;
1067	return anon_vma1 == anon_vma2;
1068}
1069
1070/*
1071 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1072 * in front of (at a lower virtual address and file offset than) the vma.
1073 *
1074 * We cannot merge two vmas if they have differently assigned (non-NULL)
1075 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1076 *
1077 * We don't check here for the merged mmap wrapping around the end of pagecache
1078 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1079 * wrap, nor mmaps which cover the final page at index -1UL.
1080 */
1081static int
1082can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1083		     struct anon_vma *anon_vma, struct file *file,
1084		     pgoff_t vm_pgoff,
1085		     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1086{
1087	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1088	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1089		if (vma->vm_pgoff == vm_pgoff)
1090			return 1;
1091	}
1092	return 0;
1093}
1094
1095/*
1096 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1097 * beyond (at a higher virtual address and file offset than) the vma.
1098 *
1099 * We cannot merge two vmas if they have differently assigned (non-NULL)
1100 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1101 */
1102static int
1103can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1104		    struct anon_vma *anon_vma, struct file *file,
1105		    pgoff_t vm_pgoff,
1106		    struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1107{
1108	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1109	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1110		pgoff_t vm_pglen;
1111		vm_pglen = vma_pages(vma);
1112		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1113			return 1;
1114	}
1115	return 0;
1116}
1117
1118/*
1119 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1120 * whether that can be merged with its predecessor or its successor.
1121 * Or both (it neatly fills a hole).
1122 *
1123 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1124 * certain not to be mapped by the time vma_merge is called; but when
1125 * called for mprotect, it is certain to be already mapped (either at
1126 * an offset within prev, or at the start of next), and the flags of
1127 * this area are about to be changed to vm_flags - and the no-change
1128 * case has already been eliminated.
1129 *
1130 * The following mprotect cases have to be considered, where AAAA is
1131 * the area passed down from mprotect_fixup, never extending beyond one
1132 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1133 *
1134 *     AAAA             AAAA                   AAAA
1135 *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
1136 *    cannot merge    might become       might become
1137 *                    PPNNNNNNNNNN       PPPPPPPPPPNN
1138 *    mmap, brk or    case 4 below       case 5 below
1139 *    mremap move:
1140 *                        AAAA               AAAA
1141 *                    PPPP    NNNN       PPPPNNNNXXXX
1142 *                    might become       might become
1143 *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
1144 *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
1145 *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
1146 *
1147 * It is important for case 8 that the vma NNNN overlapping the
1148 * region AAAA is never going to extended over XXXX. Instead XXXX must
1149 * be extended in region AAAA and NNNN must be removed. This way in
1150 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1151 * rmap_locks, the properties of the merged vma will be already
1152 * correct for the whole merged range. Some of those properties like
1153 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1154 * be correct for the whole merged range immediately after the
1155 * rmap_locks are released. Otherwise if XXXX would be removed and
1156 * NNNN would be extended over the XXXX range, remove_migration_ptes
1157 * or other rmap walkers (if working on addresses beyond the "end"
1158 * parameter) may establish ptes with the wrong permissions of NNNN
1159 * instead of the right permissions of XXXX.
1160 */
1161struct vm_area_struct *vma_merge(struct mm_struct *mm,
1162			struct vm_area_struct *prev, unsigned long addr,
1163			unsigned long end, unsigned long vm_flags,
1164			struct anon_vma *anon_vma, struct file *file,
1165			pgoff_t pgoff, struct mempolicy *policy,
1166			struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1167{
1168	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1169	struct vm_area_struct *area, *next;
1170	int err;
1171
1172	/*
1173	 * We later require that vma->vm_flags == vm_flags,
1174	 * so this tests vma->vm_flags & VM_SPECIAL, too.
1175	 */
1176	if (vm_flags & VM_SPECIAL)
1177		return NULL;
1178
1179	next = vma_next(mm, prev);
 
 
 
1180	area = next;
1181	if (area && area->vm_end == end)		/* cases 6, 7, 8 */
1182		next = next->vm_next;
1183
1184	/* verify some invariant that must be enforced by the caller */
1185	VM_WARN_ON(prev && addr <= prev->vm_start);
1186	VM_WARN_ON(area && end > area->vm_end);
1187	VM_WARN_ON(addr >= end);
1188
1189	/*
1190	 * Can it merge with the predecessor?
1191	 */
1192	if (prev && prev->vm_end == addr &&
1193			mpol_equal(vma_policy(prev), policy) &&
1194			can_vma_merge_after(prev, vm_flags,
1195					    anon_vma, file, pgoff,
1196					    vm_userfaultfd_ctx)) {
1197		/*
1198		 * OK, it can.  Can we now merge in the successor as well?
1199		 */
1200		if (next && end == next->vm_start &&
1201				mpol_equal(policy, vma_policy(next)) &&
1202				can_vma_merge_before(next, vm_flags,
1203						     anon_vma, file,
1204						     pgoff+pglen,
1205						     vm_userfaultfd_ctx) &&
1206				is_mergeable_anon_vma(prev->anon_vma,
1207						      next->anon_vma, NULL)) {
1208							/* cases 1, 6 */
1209			err = __vma_adjust(prev, prev->vm_start,
1210					 next->vm_end, prev->vm_pgoff, NULL,
1211					 prev);
1212		} else					/* cases 2, 5, 7 */
1213			err = __vma_adjust(prev, prev->vm_start,
1214					 end, prev->vm_pgoff, NULL, prev);
1215		if (err)
1216			return NULL;
1217		khugepaged_enter_vma_merge(prev, vm_flags);
1218		return prev;
1219	}
1220
1221	/*
1222	 * Can this new request be merged in front of next?
1223	 */
1224	if (next && end == next->vm_start &&
1225			mpol_equal(policy, vma_policy(next)) &&
1226			can_vma_merge_before(next, vm_flags,
1227					     anon_vma, file, pgoff+pglen,
1228					     vm_userfaultfd_ctx)) {
1229		if (prev && addr < prev->vm_end)	/* case 4 */
1230			err = __vma_adjust(prev, prev->vm_start,
1231					 addr, prev->vm_pgoff, NULL, next);
1232		else {					/* cases 3, 8 */
1233			err = __vma_adjust(area, addr, next->vm_end,
1234					 next->vm_pgoff - pglen, NULL, next);
1235			/*
1236			 * In case 3 area is already equal to next and
1237			 * this is a noop, but in case 8 "area" has
1238			 * been removed and next was expanded over it.
1239			 */
1240			area = next;
1241		}
1242		if (err)
1243			return NULL;
1244		khugepaged_enter_vma_merge(area, vm_flags);
1245		return area;
1246	}
1247
1248	return NULL;
1249}
1250
1251/*
1252 * Rough compatibility check to quickly see if it's even worth looking
1253 * at sharing an anon_vma.
1254 *
1255 * They need to have the same vm_file, and the flags can only differ
1256 * in things that mprotect may change.
1257 *
1258 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1259 * we can merge the two vma's. For example, we refuse to merge a vma if
1260 * there is a vm_ops->close() function, because that indicates that the
1261 * driver is doing some kind of reference counting. But that doesn't
1262 * really matter for the anon_vma sharing case.
1263 */
1264static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1265{
1266	return a->vm_end == b->vm_start &&
1267		mpol_equal(vma_policy(a), vma_policy(b)) &&
1268		a->vm_file == b->vm_file &&
1269		!((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1270		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1271}
1272
1273/*
1274 * Do some basic sanity checking to see if we can re-use the anon_vma
1275 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1276 * the same as 'old', the other will be the new one that is trying
1277 * to share the anon_vma.
1278 *
1279 * NOTE! This runs with mm_sem held for reading, so it is possible that
1280 * the anon_vma of 'old' is concurrently in the process of being set up
1281 * by another page fault trying to merge _that_. But that's ok: if it
1282 * is being set up, that automatically means that it will be a singleton
1283 * acceptable for merging, so we can do all of this optimistically. But
1284 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1285 *
1286 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1287 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1288 * is to return an anon_vma that is "complex" due to having gone through
1289 * a fork).
1290 *
1291 * We also make sure that the two vma's are compatible (adjacent,
1292 * and with the same memory policies). That's all stable, even with just
1293 * a read lock on the mm_sem.
1294 */
1295static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1296{
1297	if (anon_vma_compatible(a, b)) {
1298		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1299
1300		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1301			return anon_vma;
1302	}
1303	return NULL;
1304}
1305
1306/*
1307 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1308 * neighbouring vmas for a suitable anon_vma, before it goes off
1309 * to allocate a new anon_vma.  It checks because a repetitive
1310 * sequence of mprotects and faults may otherwise lead to distinct
1311 * anon_vmas being allocated, preventing vma merge in subsequent
1312 * mprotect.
1313 */
1314struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1315{
1316	struct anon_vma *anon_vma = NULL;
1317
1318	/* Try next first. */
1319	if (vma->vm_next) {
1320		anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1321		if (anon_vma)
1322			return anon_vma;
1323	}
1324
1325	/* Try prev next. */
1326	if (vma->vm_prev)
1327		anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1328
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1329	/*
1330	 * We might reach here with anon_vma == NULL if we can't find
1331	 * any reusable anon_vma.
1332	 * There's no absolute need to look only at touching neighbours:
1333	 * we could search further afield for "compatible" anon_vmas.
1334	 * But it would probably just be a waste of time searching,
1335	 * or lead to too many vmas hanging off the same anon_vma.
1336	 * We're trying to allow mprotect remerging later on,
1337	 * not trying to minimize memory used for anon_vmas.
1338	 */
1339	return anon_vma;
1340}
1341
1342/*
1343 * If a hint addr is less than mmap_min_addr change hint to be as
1344 * low as possible but still greater than mmap_min_addr
1345 */
1346static inline unsigned long round_hint_to_min(unsigned long hint)
1347{
1348	hint &= PAGE_MASK;
1349	if (((void *)hint != NULL) &&
1350	    (hint < mmap_min_addr))
1351		return PAGE_ALIGN(mmap_min_addr);
1352	return hint;
1353}
1354
1355int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1356		       unsigned long len)
 
1357{
1358	unsigned long locked, lock_limit;
1359
1360	/*  mlock MCL_FUTURE? */
1361	if (flags & VM_LOCKED) {
1362		locked = len >> PAGE_SHIFT;
1363		locked += mm->locked_vm;
1364		lock_limit = rlimit(RLIMIT_MEMLOCK);
1365		lock_limit >>= PAGE_SHIFT;
1366		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1367			return -EAGAIN;
1368	}
1369	return 0;
1370}
1371
1372static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1373{
1374	if (S_ISREG(inode->i_mode))
1375		return MAX_LFS_FILESIZE;
1376
1377	if (S_ISBLK(inode->i_mode))
1378		return MAX_LFS_FILESIZE;
1379
1380	if (S_ISSOCK(inode->i_mode))
1381		return MAX_LFS_FILESIZE;
1382
1383	/* Special "we do even unsigned file positions" case */
1384	if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1385		return 0;
1386
1387	/* Yes, random drivers might want more. But I'm tired of buggy drivers */
1388	return ULONG_MAX;
1389}
1390
1391static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1392				unsigned long pgoff, unsigned long len)
1393{
1394	u64 maxsize = file_mmap_size_max(file, inode);
1395
1396	if (maxsize && len > maxsize)
1397		return false;
1398	maxsize -= len;
1399	if (pgoff > maxsize >> PAGE_SHIFT)
1400		return false;
1401	return true;
1402}
1403
1404/*
1405 * The caller must write-lock current->mm->mmap_lock.
1406 */
1407unsigned long do_mmap(struct file *file, unsigned long addr,
1408			unsigned long len, unsigned long prot,
1409			unsigned long flags, unsigned long pgoff,
1410			unsigned long *populate, struct list_head *uf)
1411{
1412	struct mm_struct *mm = current->mm;
1413	vm_flags_t vm_flags;
1414	int pkey = 0;
1415
1416	*populate = 0;
1417
1418	if (!len)
1419		return -EINVAL;
1420
1421	/*
1422	 * Does the application expect PROT_READ to imply PROT_EXEC?
1423	 *
1424	 * (the exception is when the underlying filesystem is noexec
1425	 *  mounted, in which case we dont add PROT_EXEC.)
1426	 */
1427	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1428		if (!(file && path_noexec(&file->f_path)))
1429			prot |= PROT_EXEC;
1430
1431	/* force arch specific MAP_FIXED handling in get_unmapped_area */
1432	if (flags & MAP_FIXED_NOREPLACE)
1433		flags |= MAP_FIXED;
1434
1435	if (!(flags & MAP_FIXED))
1436		addr = round_hint_to_min(addr);
1437
1438	/* Careful about overflows.. */
1439	len = PAGE_ALIGN(len);
1440	if (!len)
1441		return -ENOMEM;
1442
1443	/* offset overflow? */
1444	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1445		return -EOVERFLOW;
1446
1447	/* Too many mappings? */
1448	if (mm->map_count > sysctl_max_map_count)
1449		return -ENOMEM;
1450
1451	/* Obtain the address to map to. we verify (or select) it and ensure
1452	 * that it represents a valid section of the address space.
1453	 */
1454	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1455	if (IS_ERR_VALUE(addr))
1456		return addr;
1457
1458	if (flags & MAP_FIXED_NOREPLACE) {
1459		if (find_vma_intersection(mm, addr, addr + len))
1460			return -EEXIST;
1461	}
1462
1463	if (prot == PROT_EXEC) {
1464		pkey = execute_only_pkey(mm);
1465		if (pkey < 0)
1466			pkey = 0;
1467	}
1468
1469	/* Do simple checking here so the lower-level routines won't have
1470	 * to. we assume access permissions have been handled by the open
1471	 * of the memory object, so we don't do any here.
1472	 */
1473	vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1474			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1475
1476	if (flags & MAP_LOCKED)
1477		if (!can_do_mlock())
1478			return -EPERM;
1479
1480	if (mlock_future_check(mm, vm_flags, len))
1481		return -EAGAIN;
1482
1483	if (file) {
1484		struct inode *inode = file_inode(file);
1485		unsigned long flags_mask;
1486
1487		if (!file_mmap_ok(file, inode, pgoff, len))
1488			return -EOVERFLOW;
1489
1490		flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1491
1492		switch (flags & MAP_TYPE) {
1493		case MAP_SHARED:
1494			/*
1495			 * Force use of MAP_SHARED_VALIDATE with non-legacy
1496			 * flags. E.g. MAP_SYNC is dangerous to use with
1497			 * MAP_SHARED as you don't know which consistency model
1498			 * you will get. We silently ignore unsupported flags
1499			 * with MAP_SHARED to preserve backward compatibility.
1500			 */
1501			flags &= LEGACY_MAP_MASK;
1502			fallthrough;
1503		case MAP_SHARED_VALIDATE:
1504			if (flags & ~flags_mask)
1505				return -EOPNOTSUPP;
1506			if (prot & PROT_WRITE) {
1507				if (!(file->f_mode & FMODE_WRITE))
1508					return -EACCES;
1509				if (IS_SWAPFILE(file->f_mapping->host))
1510					return -ETXTBSY;
1511			}
1512
1513			/*
1514			 * Make sure we don't allow writing to an append-only
1515			 * file..
1516			 */
1517			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1518				return -EACCES;
1519
1520			/*
1521			 * Make sure there are no mandatory locks on the file.
1522			 */
1523			if (locks_verify_locked(file))
1524				return -EAGAIN;
1525
1526			vm_flags |= VM_SHARED | VM_MAYSHARE;
1527			if (!(file->f_mode & FMODE_WRITE))
1528				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1529			fallthrough;
 
1530		case MAP_PRIVATE:
1531			if (!(file->f_mode & FMODE_READ))
1532				return -EACCES;
1533			if (path_noexec(&file->f_path)) {
1534				if (vm_flags & VM_EXEC)
1535					return -EPERM;
1536				vm_flags &= ~VM_MAYEXEC;
1537			}
1538
1539			if (!file->f_op->mmap)
1540				return -ENODEV;
1541			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1542				return -EINVAL;
1543			break;
1544
1545		default:
1546			return -EINVAL;
1547		}
1548	} else {
1549		switch (flags & MAP_TYPE) {
1550		case MAP_SHARED:
1551			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1552				return -EINVAL;
1553			/*
1554			 * Ignore pgoff.
1555			 */
1556			pgoff = 0;
1557			vm_flags |= VM_SHARED | VM_MAYSHARE;
1558			break;
1559		case MAP_PRIVATE:
1560			/*
1561			 * Set pgoff according to addr for anon_vma.
1562			 */
1563			pgoff = addr >> PAGE_SHIFT;
1564			break;
1565		default:
1566			return -EINVAL;
1567		}
1568	}
1569
1570	/*
1571	 * Set 'VM_NORESERVE' if we should not account for the
1572	 * memory use of this mapping.
1573	 */
1574	if (flags & MAP_NORESERVE) {
1575		/* We honor MAP_NORESERVE if allowed to overcommit */
1576		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1577			vm_flags |= VM_NORESERVE;
1578
1579		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1580		if (file && is_file_hugepages(file))
1581			vm_flags |= VM_NORESERVE;
1582	}
1583
1584	addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1585	if (!IS_ERR_VALUE(addr) &&
1586	    ((vm_flags & VM_LOCKED) ||
1587	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1588		*populate = len;
1589	return addr;
1590}
1591
1592unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1593			      unsigned long prot, unsigned long flags,
1594			      unsigned long fd, unsigned long pgoff)
1595{
1596	struct file *file = NULL;
1597	unsigned long retval;
1598
1599	if (!(flags & MAP_ANONYMOUS)) {
1600		audit_mmap_fd(fd, flags);
1601		file = fget(fd);
1602		if (!file)
1603			return -EBADF;
1604		if (is_file_hugepages(file)) {
1605			len = ALIGN(len, huge_page_size(hstate_file(file)));
1606		} else if (unlikely(flags & MAP_HUGETLB)) {
1607			retval = -EINVAL;
1608			goto out_fput;
1609		}
1610	} else if (flags & MAP_HUGETLB) {
1611		struct ucounts *ucounts = NULL;
1612		struct hstate *hs;
1613
1614		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1615		if (!hs)
1616			return -EINVAL;
1617
1618		len = ALIGN(len, huge_page_size(hs));
1619		/*
1620		 * VM_NORESERVE is used because the reservations will be
1621		 * taken when vm_ops->mmap() is called
1622		 * A dummy user value is used because we are not locking
1623		 * memory so no accounting is necessary
1624		 */
1625		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1626				VM_NORESERVE,
1627				&ucounts, HUGETLB_ANONHUGE_INODE,
1628				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1629		if (IS_ERR(file))
1630			return PTR_ERR(file);
1631	}
1632
1633	flags &= ~MAP_DENYWRITE;
1634
1635	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1636out_fput:
1637	if (file)
1638		fput(file);
1639	return retval;
1640}
1641
1642SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1643		unsigned long, prot, unsigned long, flags,
1644		unsigned long, fd, unsigned long, pgoff)
1645{
1646	return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1647}
1648
1649#ifdef __ARCH_WANT_SYS_OLD_MMAP
1650struct mmap_arg_struct {
1651	unsigned long addr;
1652	unsigned long len;
1653	unsigned long prot;
1654	unsigned long flags;
1655	unsigned long fd;
1656	unsigned long offset;
1657};
1658
1659SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1660{
1661	struct mmap_arg_struct a;
1662
1663	if (copy_from_user(&a, arg, sizeof(a)))
1664		return -EFAULT;
1665	if (offset_in_page(a.offset))
1666		return -EINVAL;
1667
1668	return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1669			       a.offset >> PAGE_SHIFT);
1670}
1671#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1672
1673/*
1674 * Some shared mappings will want the pages marked read-only
1675 * to track write events. If so, we'll downgrade vm_page_prot
1676 * to the private version (using protection_map[] without the
1677 * VM_SHARED bit).
1678 */
1679int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1680{
1681	vm_flags_t vm_flags = vma->vm_flags;
1682	const struct vm_operations_struct *vm_ops = vma->vm_ops;
1683
1684	/* If it was private or non-writable, the write bit is already clear */
1685	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1686		return 0;
1687
1688	/* The backer wishes to know when pages are first written to? */
1689	if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1690		return 1;
1691
1692	/* The open routine did something to the protections that pgprot_modify
1693	 * won't preserve? */
1694	if (pgprot_val(vm_page_prot) !=
1695	    pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1696		return 0;
1697
1698	/* Do we need to track softdirty? */
1699	if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1700		return 1;
1701
1702	/* Specialty mapping? */
1703	if (vm_flags & VM_PFNMAP)
1704		return 0;
1705
1706	/* Can the mapping track the dirty pages? */
1707	return vma->vm_file && vma->vm_file->f_mapping &&
1708		mapping_can_writeback(vma->vm_file->f_mapping);
1709}
1710
1711/*
1712 * We account for memory if it's a private writeable mapping,
1713 * not hugepages and VM_NORESERVE wasn't set.
1714 */
1715static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1716{
1717	/*
1718	 * hugetlb has its own accounting separate from the core VM
1719	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1720	 */
1721	if (file && is_file_hugepages(file))
1722		return 0;
1723
1724	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1725}
1726
1727unsigned long mmap_region(struct file *file, unsigned long addr,
1728		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1729		struct list_head *uf)
1730{
1731	struct mm_struct *mm = current->mm;
1732	struct vm_area_struct *vma, *prev, *merge;
1733	int error;
1734	struct rb_node **rb_link, *rb_parent;
1735	unsigned long charged = 0;
1736
1737	/* Check against address space limit. */
1738	if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1739		unsigned long nr_pages;
1740
1741		/*
1742		 * MAP_FIXED may remove pages of mappings that intersects with
1743		 * requested mapping. Account for the pages it would unmap.
1744		 */
1745		nr_pages = count_vma_pages_range(mm, addr, addr + len);
1746
1747		if (!may_expand_vm(mm, vm_flags,
1748					(len >> PAGE_SHIFT) - nr_pages))
1749			return -ENOMEM;
1750	}
1751
1752	/* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1753	if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1754		return -ENOMEM;
 
 
 
 
1755	/*
1756	 * Private writable mapping: check memory availability
1757	 */
1758	if (accountable_mapping(file, vm_flags)) {
1759		charged = len >> PAGE_SHIFT;
1760		if (security_vm_enough_memory_mm(mm, charged))
1761			return -ENOMEM;
1762		vm_flags |= VM_ACCOUNT;
1763	}
1764
1765	/*
1766	 * Can we just expand an old mapping?
1767	 */
1768	vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1769			NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1770	if (vma)
1771		goto out;
1772
1773	/*
1774	 * Determine the object being mapped and call the appropriate
1775	 * specific mapper. the address has already been validated, but
1776	 * not unmapped, but the maps are removed from the list.
1777	 */
1778	vma = vm_area_alloc(mm);
1779	if (!vma) {
1780		error = -ENOMEM;
1781		goto unacct_error;
1782	}
1783
 
1784	vma->vm_start = addr;
1785	vma->vm_end = addr + len;
1786	vma->vm_flags = vm_flags;
1787	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1788	vma->vm_pgoff = pgoff;
 
1789
1790	if (file) {
1791		if (vm_flags & VM_DENYWRITE) {
1792			error = deny_write_access(file);
1793			if (error)
1794				goto free_vma;
1795		}
1796		if (vm_flags & VM_SHARED) {
1797			error = mapping_map_writable(file->f_mapping);
1798			if (error)
1799				goto allow_write_and_free_vma;
1800		}
1801
1802		/* ->mmap() can change vma->vm_file, but must guarantee that
1803		 * vma_link() below can deny write-access if VM_DENYWRITE is set
1804		 * and map writably if VM_SHARED is set. This usually means the
1805		 * new file must not have been exposed to user-space, yet.
1806		 */
1807		vma->vm_file = get_file(file);
1808		error = call_mmap(file, vma);
1809		if (error)
1810			goto unmap_and_free_vma;
1811
1812		/* Can addr have changed??
1813		 *
1814		 * Answer: Yes, several device drivers can do it in their
1815		 *         f_op->mmap method. -DaveM
1816		 * Bug: If addr is changed, prev, rb_link, rb_parent should
1817		 *      be updated for vma_link()
1818		 */
1819		WARN_ON_ONCE(addr != vma->vm_start);
1820
1821		addr = vma->vm_start;
1822
1823		/* If vm_flags changed after call_mmap(), we should try merge vma again
1824		 * as we may succeed this time.
1825		 */
1826		if (unlikely(vm_flags != vma->vm_flags && prev)) {
1827			merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1828				NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1829			if (merge) {
1830				/* ->mmap() can change vma->vm_file and fput the original file. So
1831				 * fput the vma->vm_file here or we would add an extra fput for file
1832				 * and cause general protection fault ultimately.
1833				 */
1834				fput(vma->vm_file);
1835				vm_area_free(vma);
1836				vma = merge;
1837				/* Update vm_flags to pick up the change. */
1838				vm_flags = vma->vm_flags;
1839				goto unmap_writable;
1840			}
1841		}
1842
1843		vm_flags = vma->vm_flags;
1844	} else if (vm_flags & VM_SHARED) {
1845		error = shmem_zero_setup(vma);
1846		if (error)
1847			goto free_vma;
1848	} else {
1849		vma_set_anonymous(vma);
1850	}
1851
1852	/* Allow architectures to sanity-check the vm_flags */
1853	if (!arch_validate_flags(vma->vm_flags)) {
1854		error = -EINVAL;
1855		if (file)
1856			goto unmap_and_free_vma;
1857		else
1858			goto free_vma;
1859	}
1860
1861	vma_link(mm, vma, prev, rb_link, rb_parent);
1862	/* Once vma denies write, undo our temporary denial count */
1863	if (file) {
1864unmap_writable:
1865		if (vm_flags & VM_SHARED)
1866			mapping_unmap_writable(file->f_mapping);
1867		if (vm_flags & VM_DENYWRITE)
1868			allow_write_access(file);
1869	}
1870	file = vma->vm_file;
1871out:
1872	perf_event_mmap(vma);
1873
1874	vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1875	if (vm_flags & VM_LOCKED) {
1876		if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1877					is_vm_hugetlb_page(vma) ||
1878					vma == get_gate_vma(current->mm))
 
1879			vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1880		else
1881			mm->locked_vm += (len >> PAGE_SHIFT);
1882	}
1883
1884	if (file)
1885		uprobe_mmap(vma);
1886
1887	/*
1888	 * New (or expanded) vma always get soft dirty status.
1889	 * Otherwise user-space soft-dirty page tracker won't
1890	 * be able to distinguish situation when vma area unmapped,
1891	 * then new mapped in-place (which must be aimed as
1892	 * a completely new data area).
1893	 */
1894	vma->vm_flags |= VM_SOFTDIRTY;
1895
1896	vma_set_page_prot(vma);
1897
1898	return addr;
1899
1900unmap_and_free_vma:
1901	fput(vma->vm_file);
1902	vma->vm_file = NULL;
 
1903
1904	/* Undo any partial mapping done by a device driver. */
1905	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1906	charged = 0;
1907	if (vm_flags & VM_SHARED)
1908		mapping_unmap_writable(file->f_mapping);
1909allow_write_and_free_vma:
1910	if (vm_flags & VM_DENYWRITE)
1911		allow_write_access(file);
1912free_vma:
1913	vm_area_free(vma);
1914unacct_error:
1915	if (charged)
1916		vm_unacct_memory(charged);
1917	return error;
1918}
1919
1920static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1921{
1922	/*
1923	 * We implement the search by looking for an rbtree node that
1924	 * immediately follows a suitable gap. That is,
1925	 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1926	 * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1927	 * - gap_end - gap_start >= length
1928	 */
1929
1930	struct mm_struct *mm = current->mm;
1931	struct vm_area_struct *vma;
1932	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1933
1934	/* Adjust search length to account for worst case alignment overhead */
1935	length = info->length + info->align_mask;
1936	if (length < info->length)
1937		return -ENOMEM;
1938
1939	/* Adjust search limits by the desired length */
1940	if (info->high_limit < length)
1941		return -ENOMEM;
1942	high_limit = info->high_limit - length;
1943
1944	if (info->low_limit > high_limit)
1945		return -ENOMEM;
1946	low_limit = info->low_limit + length;
1947
1948	/* Check if rbtree root looks promising */
1949	if (RB_EMPTY_ROOT(&mm->mm_rb))
1950		goto check_highest;
1951	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1952	if (vma->rb_subtree_gap < length)
1953		goto check_highest;
1954
1955	while (true) {
1956		/* Visit left subtree if it looks promising */
1957		gap_end = vm_start_gap(vma);
1958		if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1959			struct vm_area_struct *left =
1960				rb_entry(vma->vm_rb.rb_left,
1961					 struct vm_area_struct, vm_rb);
1962			if (left->rb_subtree_gap >= length) {
1963				vma = left;
1964				continue;
1965			}
1966		}
1967
1968		gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1969check_current:
1970		/* Check if current node has a suitable gap */
1971		if (gap_start > high_limit)
1972			return -ENOMEM;
1973		if (gap_end >= low_limit &&
1974		    gap_end > gap_start && gap_end - gap_start >= length)
1975			goto found;
1976
1977		/* Visit right subtree if it looks promising */
1978		if (vma->vm_rb.rb_right) {
1979			struct vm_area_struct *right =
1980				rb_entry(vma->vm_rb.rb_right,
1981					 struct vm_area_struct, vm_rb);
1982			if (right->rb_subtree_gap >= length) {
1983				vma = right;
1984				continue;
1985			}
1986		}
1987
1988		/* Go back up the rbtree to find next candidate node */
1989		while (true) {
1990			struct rb_node *prev = &vma->vm_rb;
1991			if (!rb_parent(prev))
1992				goto check_highest;
1993			vma = rb_entry(rb_parent(prev),
1994				       struct vm_area_struct, vm_rb);
1995			if (prev == vma->vm_rb.rb_left) {
1996				gap_start = vm_end_gap(vma->vm_prev);
1997				gap_end = vm_start_gap(vma);
1998				goto check_current;
1999			}
2000		}
2001	}
2002
2003check_highest:
2004	/* Check highest gap, which does not precede any rbtree node */
2005	gap_start = mm->highest_vm_end;
2006	gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
2007	if (gap_start > high_limit)
2008		return -ENOMEM;
2009
2010found:
2011	/* We found a suitable gap. Clip it with the original low_limit. */
2012	if (gap_start < info->low_limit)
2013		gap_start = info->low_limit;
2014
2015	/* Adjust gap address to the desired alignment */
2016	gap_start += (info->align_offset - gap_start) & info->align_mask;
2017
2018	VM_BUG_ON(gap_start + info->length > info->high_limit);
2019	VM_BUG_ON(gap_start + info->length > gap_end);
2020	return gap_start;
2021}
2022
2023static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2024{
2025	struct mm_struct *mm = current->mm;
2026	struct vm_area_struct *vma;
2027	unsigned long length, low_limit, high_limit, gap_start, gap_end;
2028
2029	/* Adjust search length to account for worst case alignment overhead */
2030	length = info->length + info->align_mask;
2031	if (length < info->length)
2032		return -ENOMEM;
2033
2034	/*
2035	 * Adjust search limits by the desired length.
2036	 * See implementation comment at top of unmapped_area().
2037	 */
2038	gap_end = info->high_limit;
2039	if (gap_end < length)
2040		return -ENOMEM;
2041	high_limit = gap_end - length;
2042
2043	if (info->low_limit > high_limit)
2044		return -ENOMEM;
2045	low_limit = info->low_limit + length;
2046
2047	/* Check highest gap, which does not precede any rbtree node */
2048	gap_start = mm->highest_vm_end;
2049	if (gap_start <= high_limit)
2050		goto found_highest;
2051
2052	/* Check if rbtree root looks promising */
2053	if (RB_EMPTY_ROOT(&mm->mm_rb))
2054		return -ENOMEM;
2055	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2056	if (vma->rb_subtree_gap < length)
2057		return -ENOMEM;
2058
2059	while (true) {
2060		/* Visit right subtree if it looks promising */
2061		gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2062		if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2063			struct vm_area_struct *right =
2064				rb_entry(vma->vm_rb.rb_right,
2065					 struct vm_area_struct, vm_rb);
2066			if (right->rb_subtree_gap >= length) {
2067				vma = right;
2068				continue;
2069			}
2070		}
2071
2072check_current:
2073		/* Check if current node has a suitable gap */
2074		gap_end = vm_start_gap(vma);
2075		if (gap_end < low_limit)
2076			return -ENOMEM;
2077		if (gap_start <= high_limit &&
2078		    gap_end > gap_start && gap_end - gap_start >= length)
2079			goto found;
2080
2081		/* Visit left subtree if it looks promising */
2082		if (vma->vm_rb.rb_left) {
2083			struct vm_area_struct *left =
2084				rb_entry(vma->vm_rb.rb_left,
2085					 struct vm_area_struct, vm_rb);
2086			if (left->rb_subtree_gap >= length) {
2087				vma = left;
2088				continue;
2089			}
2090		}
2091
2092		/* Go back up the rbtree to find next candidate node */
2093		while (true) {
2094			struct rb_node *prev = &vma->vm_rb;
2095			if (!rb_parent(prev))
2096				return -ENOMEM;
2097			vma = rb_entry(rb_parent(prev),
2098				       struct vm_area_struct, vm_rb);
2099			if (prev == vma->vm_rb.rb_right) {
2100				gap_start = vma->vm_prev ?
2101					vm_end_gap(vma->vm_prev) : 0;
2102				goto check_current;
2103			}
2104		}
2105	}
2106
2107found:
2108	/* We found a suitable gap. Clip it with the original high_limit. */
2109	if (gap_end > info->high_limit)
2110		gap_end = info->high_limit;
2111
2112found_highest:
2113	/* Compute highest gap address at the desired alignment */
2114	gap_end -= info->length;
2115	gap_end -= (gap_end - info->align_offset) & info->align_mask;
2116
2117	VM_BUG_ON(gap_end < info->low_limit);
2118	VM_BUG_ON(gap_end < gap_start);
2119	return gap_end;
2120}
2121
2122/*
2123 * Search for an unmapped address range.
2124 *
2125 * We are looking for a range that:
2126 * - does not intersect with any VMA;
2127 * - is contained within the [low_limit, high_limit) interval;
2128 * - is at least the desired size.
2129 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2130 */
2131unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2132{
2133	unsigned long addr;
2134
2135	if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2136		addr = unmapped_area_topdown(info);
2137	else
2138		addr = unmapped_area(info);
2139
2140	trace_vm_unmapped_area(addr, info);
2141	return addr;
2142}
2143
2144#ifndef arch_get_mmap_end
2145#define arch_get_mmap_end(addr)	(TASK_SIZE)
2146#endif
2147
2148#ifndef arch_get_mmap_base
2149#define arch_get_mmap_base(addr, base) (base)
2150#endif
2151
2152/* Get an address range which is currently unmapped.
2153 * For shmat() with addr=0.
2154 *
2155 * Ugly calling convention alert:
2156 * Return value with the low bits set means error value,
2157 * ie
2158 *	if (ret & ~PAGE_MASK)
2159 *		error = ret;
2160 *
2161 * This function "knows" that -ENOMEM has the bits set.
2162 */
2163#ifndef HAVE_ARCH_UNMAPPED_AREA
2164unsigned long
2165arch_get_unmapped_area(struct file *filp, unsigned long addr,
2166		unsigned long len, unsigned long pgoff, unsigned long flags)
2167{
2168	struct mm_struct *mm = current->mm;
2169	struct vm_area_struct *vma, *prev;
2170	struct vm_unmapped_area_info info;
2171	const unsigned long mmap_end = arch_get_mmap_end(addr);
2172
2173	if (len > mmap_end - mmap_min_addr)
2174		return -ENOMEM;
2175
2176	if (flags & MAP_FIXED)
2177		return addr;
2178
2179	if (addr) {
2180		addr = PAGE_ALIGN(addr);
2181		vma = find_vma_prev(mm, addr, &prev);
2182		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2183		    (!vma || addr + len <= vm_start_gap(vma)) &&
2184		    (!prev || addr >= vm_end_gap(prev)))
2185			return addr;
2186	}
2187
2188	info.flags = 0;
2189	info.length = len;
2190	info.low_limit = mm->mmap_base;
2191	info.high_limit = mmap_end;
2192	info.align_mask = 0;
2193	info.align_offset = 0;
2194	return vm_unmapped_area(&info);
2195}
2196#endif
2197
2198/*
2199 * This mmap-allocator allocates new areas top-down from below the
2200 * stack's low limit (the base):
2201 */
2202#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2203unsigned long
2204arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2205			  unsigned long len, unsigned long pgoff,
2206			  unsigned long flags)
2207{
2208	struct vm_area_struct *vma, *prev;
2209	struct mm_struct *mm = current->mm;
 
2210	struct vm_unmapped_area_info info;
2211	const unsigned long mmap_end = arch_get_mmap_end(addr);
2212
2213	/* requested length too big for entire address space */
2214	if (len > mmap_end - mmap_min_addr)
2215		return -ENOMEM;
2216
2217	if (flags & MAP_FIXED)
2218		return addr;
2219
2220	/* requesting a specific address */
2221	if (addr) {
2222		addr = PAGE_ALIGN(addr);
2223		vma = find_vma_prev(mm, addr, &prev);
2224		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2225				(!vma || addr + len <= vm_start_gap(vma)) &&
2226				(!prev || addr >= vm_end_gap(prev)))
2227			return addr;
2228	}
2229
2230	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2231	info.length = len;
2232	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2233	info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2234	info.align_mask = 0;
2235	info.align_offset = 0;
2236	addr = vm_unmapped_area(&info);
2237
2238	/*
2239	 * A failed mmap() very likely causes application failure,
2240	 * so fall back to the bottom-up function here. This scenario
2241	 * can happen with large stack limits and large mmap()
2242	 * allocations.
2243	 */
2244	if (offset_in_page(addr)) {
2245		VM_BUG_ON(addr != -ENOMEM);
2246		info.flags = 0;
2247		info.low_limit = TASK_UNMAPPED_BASE;
2248		info.high_limit = mmap_end;
2249		addr = vm_unmapped_area(&info);
2250	}
2251
2252	return addr;
2253}
2254#endif
2255
2256unsigned long
2257get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2258		unsigned long pgoff, unsigned long flags)
2259{
2260	unsigned long (*get_area)(struct file *, unsigned long,
2261				  unsigned long, unsigned long, unsigned long);
2262
2263	unsigned long error = arch_mmap_check(addr, len, flags);
2264	if (error)
2265		return error;
2266
2267	/* Careful about overflows.. */
2268	if (len > TASK_SIZE)
2269		return -ENOMEM;
2270
2271	get_area = current->mm->get_unmapped_area;
2272	if (file) {
2273		if (file->f_op->get_unmapped_area)
2274			get_area = file->f_op->get_unmapped_area;
2275	} else if (flags & MAP_SHARED) {
2276		/*
2277		 * mmap_region() will call shmem_zero_setup() to create a file,
2278		 * so use shmem's get_unmapped_area in case it can be huge.
2279		 * do_mmap() will clear pgoff, so match alignment.
2280		 */
2281		pgoff = 0;
2282		get_area = shmem_get_unmapped_area;
2283	}
2284
2285	addr = get_area(file, addr, len, pgoff, flags);
2286	if (IS_ERR_VALUE(addr))
2287		return addr;
2288
2289	if (addr > TASK_SIZE - len)
2290		return -ENOMEM;
2291	if (offset_in_page(addr))
2292		return -EINVAL;
2293
 
2294	error = security_mmap_addr(addr);
2295	return error ? error : addr;
2296}
2297
2298EXPORT_SYMBOL(get_unmapped_area);
2299
2300/* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2301struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2302{
2303	struct rb_node *rb_node;
2304	struct vm_area_struct *vma;
2305
2306	/* Check the cache first. */
2307	vma = vmacache_find(mm, addr);
2308	if (likely(vma))
2309		return vma;
2310
2311	rb_node = mm->mm_rb.rb_node;
2312
2313	while (rb_node) {
2314		struct vm_area_struct *tmp;
2315
2316		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2317
2318		if (tmp->vm_end > addr) {
2319			vma = tmp;
2320			if (tmp->vm_start <= addr)
2321				break;
2322			rb_node = rb_node->rb_left;
2323		} else
2324			rb_node = rb_node->rb_right;
2325	}
2326
2327	if (vma)
2328		vmacache_update(addr, vma);
2329	return vma;
2330}
2331
2332EXPORT_SYMBOL(find_vma);
2333
2334/*
2335 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2336 */
2337struct vm_area_struct *
2338find_vma_prev(struct mm_struct *mm, unsigned long addr,
2339			struct vm_area_struct **pprev)
2340{
2341	struct vm_area_struct *vma;
2342
2343	vma = find_vma(mm, addr);
2344	if (vma) {
2345		*pprev = vma->vm_prev;
2346	} else {
2347		struct rb_node *rb_node = rb_last(&mm->mm_rb);
2348
2349		*pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
 
 
 
2350	}
2351	return vma;
2352}
2353
2354/*
2355 * Verify that the stack growth is acceptable and
2356 * update accounting. This is shared with both the
2357 * grow-up and grow-down cases.
2358 */
2359static int acct_stack_growth(struct vm_area_struct *vma,
2360			     unsigned long size, unsigned long grow)
2361{
2362	struct mm_struct *mm = vma->vm_mm;
2363	unsigned long new_start;
 
2364
2365	/* address space limit tests */
2366	if (!may_expand_vm(mm, vma->vm_flags, grow))
2367		return -ENOMEM;
2368
2369	/* Stack limit test */
2370	if (size > rlimit(RLIMIT_STACK))
 
 
 
2371		return -ENOMEM;
2372
2373	/* mlock limit tests */
2374	if (vma->vm_flags & VM_LOCKED) {
2375		unsigned long locked;
2376		unsigned long limit;
2377		locked = mm->locked_vm + grow;
2378		limit = rlimit(RLIMIT_MEMLOCK);
2379		limit >>= PAGE_SHIFT;
2380		if (locked > limit && !capable(CAP_IPC_LOCK))
2381			return -ENOMEM;
2382	}
2383
2384	/* Check to ensure the stack will not grow into a hugetlb-only region */
2385	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2386			vma->vm_end - size;
2387	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2388		return -EFAULT;
2389
2390	/*
2391	 * Overcommit..  This must be the final test, as it will
2392	 * update security statistics.
2393	 */
2394	if (security_vm_enough_memory_mm(mm, grow))
2395		return -ENOMEM;
2396
2397	return 0;
2398}
2399
2400#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2401/*
2402 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2403 * vma is the last one with address > vma->vm_end.  Have to extend vma.
2404 */
2405int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2406{
2407	struct mm_struct *mm = vma->vm_mm;
2408	struct vm_area_struct *next;
2409	unsigned long gap_addr;
2410	int error = 0;
2411
2412	if (!(vma->vm_flags & VM_GROWSUP))
2413		return -EFAULT;
2414
2415	/* Guard against exceeding limits of the address space. */
2416	address &= PAGE_MASK;
2417	if (address >= (TASK_SIZE & PAGE_MASK))
 
2418		return -ENOMEM;
2419	address += PAGE_SIZE;
2420
2421	/* Enforce stack_guard_gap */
2422	gap_addr = address + stack_guard_gap;
2423
2424	/* Guard against overflow */
2425	if (gap_addr < address || gap_addr > TASK_SIZE)
2426		gap_addr = TASK_SIZE;
2427
2428	next = vma->vm_next;
2429	if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2430		if (!(next->vm_flags & VM_GROWSUP))
2431			return -ENOMEM;
2432		/* Check that both stack segments have the same anon_vma? */
2433	}
2434
2435	/* We must make sure the anon_vma is allocated. */
2436	if (unlikely(anon_vma_prepare(vma)))
2437		return -ENOMEM;
2438
2439	/*
2440	 * vma->vm_start/vm_end cannot change under us because the caller
2441	 * is required to hold the mmap_lock in read mode.  We need the
2442	 * anon_vma lock to serialize against concurrent expand_stacks.
2443	 */
2444	anon_vma_lock_write(vma->anon_vma);
2445
2446	/* Somebody else might have raced and expanded it already */
2447	if (address > vma->vm_end) {
2448		unsigned long size, grow;
2449
2450		size = address - vma->vm_start;
2451		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2452
2453		error = -ENOMEM;
2454		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2455			error = acct_stack_growth(vma, size, grow);
2456			if (!error) {
2457				/*
2458				 * vma_gap_update() doesn't support concurrent
2459				 * updates, but we only hold a shared mmap_lock
2460				 * lock here, so we need to protect against
2461				 * concurrent vma expansions.
2462				 * anon_vma_lock_write() doesn't help here, as
2463				 * we don't guarantee that all growable vmas
2464				 * in a mm share the same root anon vma.
2465				 * So, we reuse mm->page_table_lock to guard
2466				 * against concurrent vma expansions.
2467				 */
2468				spin_lock(&mm->page_table_lock);
2469				if (vma->vm_flags & VM_LOCKED)
2470					mm->locked_vm += grow;
2471				vm_stat_account(mm, vma->vm_flags, grow);
2472				anon_vma_interval_tree_pre_update_vma(vma);
2473				vma->vm_end = address;
2474				anon_vma_interval_tree_post_update_vma(vma);
2475				if (vma->vm_next)
2476					vma_gap_update(vma->vm_next);
2477				else
2478					mm->highest_vm_end = vm_end_gap(vma);
2479				spin_unlock(&mm->page_table_lock);
2480
2481				perf_event_mmap(vma);
2482			}
2483		}
2484	}
2485	anon_vma_unlock_write(vma->anon_vma);
2486	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2487	validate_mm(mm);
2488	return error;
2489}
2490#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2491
2492/*
2493 * vma is the first one with address < vma->vm_start.  Have to extend vma.
2494 */
2495int expand_downwards(struct vm_area_struct *vma,
2496				   unsigned long address)
2497{
2498	struct mm_struct *mm = vma->vm_mm;
2499	struct vm_area_struct *prev;
2500	int error = 0;
2501
2502	address &= PAGE_MASK;
2503	if (address < mmap_min_addr)
2504		return -EPERM;
2505
2506	/* Enforce stack_guard_gap */
2507	prev = vma->vm_prev;
2508	/* Check that both stack segments have the same anon_vma? */
2509	if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2510			vma_is_accessible(prev)) {
2511		if (address - prev->vm_end < stack_guard_gap)
2512			return -ENOMEM;
2513	}
2514
2515	/* We must make sure the anon_vma is allocated. */
2516	if (unlikely(anon_vma_prepare(vma)))
2517		return -ENOMEM;
2518
2519	/*
2520	 * vma->vm_start/vm_end cannot change under us because the caller
2521	 * is required to hold the mmap_lock in read mode.  We need the
2522	 * anon_vma lock to serialize against concurrent expand_stacks.
2523	 */
2524	anon_vma_lock_write(vma->anon_vma);
2525
2526	/* Somebody else might have raced and expanded it already */
2527	if (address < vma->vm_start) {
2528		unsigned long size, grow;
2529
2530		size = vma->vm_end - address;
2531		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2532
2533		error = -ENOMEM;
2534		if (grow <= vma->vm_pgoff) {
2535			error = acct_stack_growth(vma, size, grow);
2536			if (!error) {
2537				/*
2538				 * vma_gap_update() doesn't support concurrent
2539				 * updates, but we only hold a shared mmap_lock
2540				 * lock here, so we need to protect against
2541				 * concurrent vma expansions.
2542				 * anon_vma_lock_write() doesn't help here, as
2543				 * we don't guarantee that all growable vmas
2544				 * in a mm share the same root anon vma.
2545				 * So, we reuse mm->page_table_lock to guard
2546				 * against concurrent vma expansions.
2547				 */
2548				spin_lock(&mm->page_table_lock);
2549				if (vma->vm_flags & VM_LOCKED)
2550					mm->locked_vm += grow;
2551				vm_stat_account(mm, vma->vm_flags, grow);
2552				anon_vma_interval_tree_pre_update_vma(vma);
2553				vma->vm_start = address;
2554				vma->vm_pgoff -= grow;
2555				anon_vma_interval_tree_post_update_vma(vma);
2556				vma_gap_update(vma);
2557				spin_unlock(&mm->page_table_lock);
2558
2559				perf_event_mmap(vma);
2560			}
2561		}
2562	}
2563	anon_vma_unlock_write(vma->anon_vma);
2564	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2565	validate_mm(mm);
2566	return error;
2567}
2568
2569/* enforced gap between the expanding stack and other mappings. */
2570unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2571
2572static int __init cmdline_parse_stack_guard_gap(char *p)
2573{
2574	unsigned long val;
2575	char *endptr;
2576
2577	val = simple_strtoul(p, &endptr, 10);
2578	if (!*endptr)
2579		stack_guard_gap = val << PAGE_SHIFT;
2580
2581	return 0;
2582}
2583__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2584
2585#ifdef CONFIG_STACK_GROWSUP
2586int expand_stack(struct vm_area_struct *vma, unsigned long address)
2587{
 
 
 
 
 
 
 
 
2588	return expand_upwards(vma, address);
2589}
2590
2591struct vm_area_struct *
2592find_extend_vma(struct mm_struct *mm, unsigned long addr)
2593{
2594	struct vm_area_struct *vma, *prev;
2595
2596	addr &= PAGE_MASK;
2597	vma = find_vma_prev(mm, addr, &prev);
2598	if (vma && (vma->vm_start <= addr))
2599		return vma;
2600	/* don't alter vm_end if the coredump is running */
2601	if (!prev || expand_stack(prev, addr))
2602		return NULL;
2603	if (prev->vm_flags & VM_LOCKED)
2604		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2605	return prev;
2606}
2607#else
2608int expand_stack(struct vm_area_struct *vma, unsigned long address)
2609{
 
 
 
 
 
 
 
 
2610	return expand_downwards(vma, address);
2611}
2612
2613struct vm_area_struct *
2614find_extend_vma(struct mm_struct *mm, unsigned long addr)
2615{
2616	struct vm_area_struct *vma;
2617	unsigned long start;
2618
2619	addr &= PAGE_MASK;
2620	vma = find_vma(mm, addr);
2621	if (!vma)
2622		return NULL;
2623	if (vma->vm_start <= addr)
2624		return vma;
2625	if (!(vma->vm_flags & VM_GROWSDOWN))
2626		return NULL;
2627	start = vma->vm_start;
2628	if (expand_stack(vma, addr))
2629		return NULL;
2630	if (vma->vm_flags & VM_LOCKED)
2631		populate_vma_page_range(vma, addr, start, NULL);
2632	return vma;
2633}
2634#endif
2635
2636EXPORT_SYMBOL_GPL(find_extend_vma);
2637
2638/*
2639 * Ok - we have the memory areas we should free on the vma list,
2640 * so release them, and do the vma updates.
2641 *
2642 * Called with the mm semaphore held.
2643 */
2644static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2645{
2646	unsigned long nr_accounted = 0;
2647
2648	/* Update high watermark before we lower total_vm */
2649	update_hiwater_vm(mm);
2650	do {
2651		long nrpages = vma_pages(vma);
2652
2653		if (vma->vm_flags & VM_ACCOUNT)
2654			nr_accounted += nrpages;
2655		vm_stat_account(mm, vma->vm_flags, -nrpages);
2656		vma = remove_vma(vma);
2657	} while (vma);
2658	vm_unacct_memory(nr_accounted);
2659	validate_mm(mm);
2660}
2661
2662/*
2663 * Get rid of page table information in the indicated region.
2664 *
2665 * Called with the mm semaphore held.
2666 */
2667static void unmap_region(struct mm_struct *mm,
2668		struct vm_area_struct *vma, struct vm_area_struct *prev,
2669		unsigned long start, unsigned long end)
2670{
2671	struct vm_area_struct *next = vma_next(mm, prev);
2672	struct mmu_gather tlb;
2673
2674	lru_add_drain();
2675	tlb_gather_mmu(&tlb, mm);
2676	update_hiwater_rss(mm);
2677	unmap_vmas(&tlb, vma, start, end);
2678	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2679				 next ? next->vm_start : USER_PGTABLES_CEILING);
2680	tlb_finish_mmu(&tlb);
2681}
2682
2683/*
2684 * Create a list of vma's touched by the unmap, removing them from the mm's
2685 * vma list as we go..
2686 */
2687static bool
2688detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2689	struct vm_area_struct *prev, unsigned long end)
2690{
2691	struct vm_area_struct **insertion_point;
2692	struct vm_area_struct *tail_vma = NULL;
2693
2694	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2695	vma->vm_prev = NULL;
2696	do {
2697		vma_rb_erase(vma, &mm->mm_rb);
2698		mm->map_count--;
2699		tail_vma = vma;
2700		vma = vma->vm_next;
2701	} while (vma && vma->vm_start < end);
2702	*insertion_point = vma;
2703	if (vma) {
2704		vma->vm_prev = prev;
2705		vma_gap_update(vma);
2706	} else
2707		mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2708	tail_vma->vm_next = NULL;
2709
2710	/* Kill the cache */
2711	vmacache_invalidate(mm);
2712
2713	/*
2714	 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2715	 * VM_GROWSUP VMA. Such VMAs can change their size under
2716	 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2717	 */
2718	if (vma && (vma->vm_flags & VM_GROWSDOWN))
2719		return false;
2720	if (prev && (prev->vm_flags & VM_GROWSUP))
2721		return false;
2722	return true;
2723}
2724
2725/*
2726 * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2727 * has already been checked or doesn't make sense to fail.
2728 */
2729int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2730		unsigned long addr, int new_below)
2731{
2732	struct vm_area_struct *new;
2733	int err;
2734
2735	if (vma->vm_ops && vma->vm_ops->may_split) {
2736		err = vma->vm_ops->may_split(vma, addr);
2737		if (err)
2738			return err;
2739	}
2740
2741	new = vm_area_dup(vma);
2742	if (!new)
2743		return -ENOMEM;
2744
 
 
 
 
 
2745	if (new_below)
2746		new->vm_end = addr;
2747	else {
2748		new->vm_start = addr;
2749		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2750	}
2751
2752	err = vma_dup_policy(vma, new);
2753	if (err)
2754		goto out_free_vma;
2755
2756	err = anon_vma_clone(new, vma);
2757	if (err)
2758		goto out_free_mpol;
2759
2760	if (new->vm_file)
2761		get_file(new->vm_file);
2762
2763	if (new->vm_ops && new->vm_ops->open)
2764		new->vm_ops->open(new);
2765
2766	if (new_below)
2767		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2768			((addr - new->vm_start) >> PAGE_SHIFT), new);
2769	else
2770		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2771
2772	/* Success. */
2773	if (!err)
2774		return 0;
2775
2776	/* Clean everything up if vma_adjust failed. */
2777	if (new->vm_ops && new->vm_ops->close)
2778		new->vm_ops->close(new);
2779	if (new->vm_file)
2780		fput(new->vm_file);
2781	unlink_anon_vmas(new);
2782 out_free_mpol:
2783	mpol_put(vma_policy(new));
2784 out_free_vma:
2785	vm_area_free(new);
2786	return err;
2787}
2788
2789/*
2790 * Split a vma into two pieces at address 'addr', a new vma is allocated
2791 * either for the first part or the tail.
2792 */
2793int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2794	      unsigned long addr, int new_below)
2795{
2796	if (mm->map_count >= sysctl_max_map_count)
2797		return -ENOMEM;
2798
2799	return __split_vma(mm, vma, addr, new_below);
2800}
2801
2802static inline void
2803unlock_range(struct vm_area_struct *start, unsigned long limit)
2804{
2805	struct mm_struct *mm = start->vm_mm;
2806	struct vm_area_struct *tmp = start;
2807
2808	while (tmp && tmp->vm_start < limit) {
2809		if (tmp->vm_flags & VM_LOCKED) {
2810			mm->locked_vm -= vma_pages(tmp);
2811			munlock_vma_pages_all(tmp);
2812		}
2813
2814		tmp = tmp->vm_next;
2815	}
2816}
2817
2818/* Munmap is split into 2 main parts -- this part which finds
2819 * what needs doing, and the areas themselves, which do the
2820 * work.  This now handles partial unmappings.
2821 * Jeremy Fitzhardinge <jeremy@goop.org>
2822 */
2823int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2824		struct list_head *uf, bool downgrade)
2825{
2826	unsigned long end;
2827	struct vm_area_struct *vma, *prev, *last;
2828
2829	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2830		return -EINVAL;
2831
2832	len = PAGE_ALIGN(len);
2833	end = start + len;
2834	if (len == 0)
2835		return -EINVAL;
2836
2837	/*
2838	 * arch_unmap() might do unmaps itself.  It must be called
2839	 * and finish any rbtree manipulation before this code
2840	 * runs and also starts to manipulate the rbtree.
2841	 */
2842	arch_unmap(mm, start, end);
2843
2844	/* Find the first overlapping VMA where start < vma->vm_end */
2845	vma = find_vma_intersection(mm, start, end);
2846	if (!vma)
2847		return 0;
2848	prev = vma->vm_prev;
 
 
 
 
 
 
2849
2850	/*
2851	 * If we need to split any vma, do it now to save pain later.
2852	 *
2853	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2854	 * unmapped vm_area_struct will remain in use: so lower split_vma
2855	 * places tmp vma above, and higher split_vma places tmp vma below.
2856	 */
2857	if (start > vma->vm_start) {
2858		int error;
2859
2860		/*
2861		 * Make sure that map_count on return from munmap() will
2862		 * not exceed its limit; but let map_count go just above
2863		 * its limit temporarily, to help free resources as expected.
2864		 */
2865		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2866			return -ENOMEM;
2867
2868		error = __split_vma(mm, vma, start, 0);
2869		if (error)
2870			return error;
2871		prev = vma;
2872	}
2873
2874	/* Does it split the last one? */
2875	last = find_vma(mm, end);
2876	if (last && end > last->vm_start) {
2877		int error = __split_vma(mm, last, end, 1);
2878		if (error)
2879			return error;
2880	}
2881	vma = vma_next(mm, prev);
2882
2883	if (unlikely(uf)) {
2884		/*
2885		 * If userfaultfd_unmap_prep returns an error the vmas
2886		 * will remain split, but userland will get a
2887		 * highly unexpected error anyway. This is no
2888		 * different than the case where the first of the two
2889		 * __split_vma fails, but we don't undo the first
2890		 * split, despite we could. This is unlikely enough
2891		 * failure that it's not worth optimizing it for.
2892		 */
2893		int error = userfaultfd_unmap_prep(vma, start, end, uf);
2894		if (error)
2895			return error;
2896	}
2897
2898	/*
2899	 * unlock any mlock()ed ranges before detaching vmas
2900	 */
2901	if (mm->locked_vm)
2902		unlock_range(vma, end);
2903
2904	/* Detach vmas from rbtree */
2905	if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2906		downgrade = false;
2907
2908	if (downgrade)
2909		mmap_write_downgrade(mm);
2910
2911	unmap_region(mm, vma, prev, start, end);
2912
2913	/* Fix up all other VM information */
2914	remove_vma_list(mm, vma);
2915
2916	return downgrade ? 1 : 0;
2917}
2918
2919int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2920	      struct list_head *uf)
2921{
2922	return __do_munmap(mm, start, len, uf, false);
2923}
2924
2925static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2926{
2927	int ret;
2928	struct mm_struct *mm = current->mm;
2929	LIST_HEAD(uf);
2930
2931	if (mmap_write_lock_killable(mm))
2932		return -EINTR;
2933
2934	ret = __do_munmap(mm, start, len, &uf, downgrade);
2935	/*
2936	 * Returning 1 indicates mmap_lock is downgraded.
2937	 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2938	 * it to 0 before return.
2939	 */
2940	if (ret == 1) {
2941		mmap_read_unlock(mm);
2942		ret = 0;
2943	} else
2944		mmap_write_unlock(mm);
2945
2946	userfaultfd_unmap_complete(mm, &uf);
 
 
2947	return ret;
2948}
2949
2950int vm_munmap(unsigned long start, size_t len)
2951{
2952	return __vm_munmap(start, len, false);
2953}
2954EXPORT_SYMBOL(vm_munmap);
2955
2956SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2957{
2958	addr = untagged_addr(addr);
2959	profile_munmap(addr);
2960	return __vm_munmap(addr, len, true);
2961}
2962
2963
2964/*
2965 * Emulation of deprecated remap_file_pages() syscall.
2966 */
2967SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2968		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2969{
2970
2971	struct mm_struct *mm = current->mm;
2972	struct vm_area_struct *vma;
2973	unsigned long populate = 0;
2974	unsigned long ret = -EINVAL;
2975	struct file *file;
2976
2977	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2978		     current->comm, current->pid);
2979
2980	if (prot)
2981		return ret;
2982	start = start & PAGE_MASK;
2983	size = size & PAGE_MASK;
2984
2985	if (start + size <= start)
2986		return ret;
2987
2988	/* Does pgoff wrap? */
2989	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2990		return ret;
2991
2992	if (mmap_write_lock_killable(mm))
2993		return -EINTR;
2994
2995	vma = find_vma(mm, start);
2996
2997	if (!vma || !(vma->vm_flags & VM_SHARED))
2998		goto out;
2999
3000	if (start < vma->vm_start)
3001		goto out;
3002
3003	if (start + size > vma->vm_end) {
3004		struct vm_area_struct *next;
3005
3006		for (next = vma->vm_next; next; next = next->vm_next) {
3007			/* hole between vmas ? */
3008			if (next->vm_start != next->vm_prev->vm_end)
3009				goto out;
3010
3011			if (next->vm_file != vma->vm_file)
3012				goto out;
3013
3014			if (next->vm_flags != vma->vm_flags)
3015				goto out;
3016
3017			if (start + size <= next->vm_end)
3018				break;
3019		}
3020
3021		if (!next)
3022			goto out;
3023	}
3024
3025	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3026	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3027	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3028
3029	flags &= MAP_NONBLOCK;
3030	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3031	if (vma->vm_flags & VM_LOCKED)
 
3032		flags |= MAP_LOCKED;
3033
 
 
 
 
 
 
 
 
 
3034	file = get_file(vma->vm_file);
3035	ret = do_mmap(vma->vm_file, start, size,
3036			prot, flags, pgoff, &populate, NULL);
3037	fput(file);
3038out:
3039	mmap_write_unlock(mm);
3040	if (populate)
3041		mm_populate(ret, populate);
3042	if (!IS_ERR_VALUE(ret))
3043		ret = 0;
3044	return ret;
3045}
3046
 
 
 
 
 
 
 
 
 
 
3047/*
3048 *  this is really a simplified "do_mmap".  it only handles
3049 *  anonymous maps.  eventually we may be able to do some
3050 *  brk-specific accounting here.
3051 */
3052static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3053{
3054	struct mm_struct *mm = current->mm;
3055	struct vm_area_struct *vma, *prev;
 
3056	struct rb_node **rb_link, *rb_parent;
3057	pgoff_t pgoff = addr >> PAGE_SHIFT;
3058	int error;
3059	unsigned long mapped_addr;
3060
3061	/* Until we need other flags, refuse anything except VM_EXEC. */
3062	if ((flags & (~VM_EXEC)) != 0)
3063		return -EINVAL;
3064	flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
 
3065
3066	mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3067	if (IS_ERR_VALUE(mapped_addr))
3068		return mapped_addr;
3069
3070	error = mlock_future_check(mm, mm->def_flags, len);
3071	if (error)
3072		return error;
3073
3074	/* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3075	if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3076		return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
3077
3078	/* Check against address space limits *after* clearing old maps... */
3079	if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3080		return -ENOMEM;
3081
3082	if (mm->map_count > sysctl_max_map_count)
3083		return -ENOMEM;
3084
3085	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3086		return -ENOMEM;
3087
3088	/* Can we just expand an old private anonymous mapping? */
3089	vma = vma_merge(mm, prev, addr, addr + len, flags,
3090			NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3091	if (vma)
3092		goto out;
3093
3094	/*
3095	 * create a vma struct for an anonymous mapping
3096	 */
3097	vma = vm_area_alloc(mm);
3098	if (!vma) {
3099		vm_unacct_memory(len >> PAGE_SHIFT);
3100		return -ENOMEM;
3101	}
3102
3103	vma_set_anonymous(vma);
 
3104	vma->vm_start = addr;
3105	vma->vm_end = addr + len;
3106	vma->vm_pgoff = pgoff;
3107	vma->vm_flags = flags;
3108	vma->vm_page_prot = vm_get_page_prot(flags);
3109	vma_link(mm, vma, prev, rb_link, rb_parent);
3110out:
3111	perf_event_mmap(vma);
3112	mm->total_vm += len >> PAGE_SHIFT;
3113	mm->data_vm += len >> PAGE_SHIFT;
3114	if (flags & VM_LOCKED)
3115		mm->locked_vm += (len >> PAGE_SHIFT);
3116	vma->vm_flags |= VM_SOFTDIRTY;
3117	return 0;
3118}
3119
3120int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3121{
3122	struct mm_struct *mm = current->mm;
3123	unsigned long len;
3124	int ret;
3125	bool populate;
3126	LIST_HEAD(uf);
3127
3128	len = PAGE_ALIGN(request);
3129	if (len < request)
3130		return -ENOMEM;
3131	if (!len)
3132		return 0;
3133
3134	if (mmap_write_lock_killable(mm))
3135		return -EINTR;
3136
3137	ret = do_brk_flags(addr, len, flags, &uf);
3138	populate = ((mm->def_flags & VM_LOCKED) != 0);
3139	mmap_write_unlock(mm);
3140	userfaultfd_unmap_complete(mm, &uf);
3141	if (populate && !ret)
3142		mm_populate(addr, len);
3143	return ret;
3144}
3145EXPORT_SYMBOL(vm_brk_flags);
3146
3147int vm_brk(unsigned long addr, unsigned long len)
3148{
3149	return vm_brk_flags(addr, len, 0);
3150}
3151EXPORT_SYMBOL(vm_brk);
3152
3153/* Release all mmaps. */
3154void exit_mmap(struct mm_struct *mm)
3155{
3156	struct mmu_gather tlb;
3157	struct vm_area_struct *vma;
3158	unsigned long nr_accounted = 0;
3159
3160	/* mm's last user has gone, and its about to be pulled down */
3161	mmu_notifier_release(mm);
3162
3163	if (unlikely(mm_is_oom_victim(mm))) {
3164		/*
3165		 * Manually reap the mm to free as much memory as possible.
3166		 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3167		 * this mm from further consideration.  Taking mm->mmap_lock for
3168		 * write after setting MMF_OOM_SKIP will guarantee that the oom
3169		 * reaper will not run on this mm again after mmap_lock is
3170		 * dropped.
3171		 *
3172		 * Nothing can be holding mm->mmap_lock here and the above call
3173		 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3174		 * __oom_reap_task_mm() will not block.
3175		 *
3176		 * This needs to be done before calling munlock_vma_pages_all(),
3177		 * which clears VM_LOCKED, otherwise the oom reaper cannot
3178		 * reliably test it.
3179		 */
3180		(void)__oom_reap_task_mm(mm);
3181
3182		set_bit(MMF_OOM_SKIP, &mm->flags);
3183		mmap_write_lock(mm);
3184		mmap_write_unlock(mm);
3185	}
3186
3187	if (mm->locked_vm)
3188		unlock_range(mm->mmap, ULONG_MAX);
3189
3190	arch_exit_mmap(mm);
3191
3192	vma = mm->mmap;
3193	if (!vma)	/* Can happen if dup_mmap() received an OOM */
3194		return;
3195
3196	lru_add_drain();
3197	flush_cache_mm(mm);
3198	tlb_gather_mmu_fullmm(&tlb, mm);
3199	/* update_hiwater_rss(mm) here? but nobody should be looking */
3200	/* Use -1 here to ensure all VMAs in the mm are unmapped */
3201	unmap_vmas(&tlb, vma, 0, -1);
 
3202	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3203	tlb_finish_mmu(&tlb);
3204
3205	/*
3206	 * Walk the list again, actually closing and freeing it,
3207	 * with preemption enabled, without holding any MM locks.
3208	 */
3209	while (vma) {
3210		if (vma->vm_flags & VM_ACCOUNT)
3211			nr_accounted += vma_pages(vma);
3212		vma = remove_vma(vma);
3213		cond_resched();
3214	}
3215	vm_unacct_memory(nr_accounted);
3216}
3217
3218/* Insert vm structure into process list sorted by address
3219 * and into the inode's i_mmap tree.  If vm_file is non-NULL
3220 * then i_mmap_rwsem is taken here.
3221 */
3222int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3223{
3224	struct vm_area_struct *prev;
3225	struct rb_node **rb_link, *rb_parent;
3226
3227	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3228			   &prev, &rb_link, &rb_parent))
3229		return -ENOMEM;
3230	if ((vma->vm_flags & VM_ACCOUNT) &&
3231	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
3232		return -ENOMEM;
3233
3234	/*
3235	 * The vm_pgoff of a purely anonymous vma should be irrelevant
3236	 * until its first write fault, when page's anon_vma and index
3237	 * are set.  But now set the vm_pgoff it will almost certainly
3238	 * end up with (unless mremap moves it elsewhere before that
3239	 * first wfault), so /proc/pid/maps tells a consistent story.
3240	 *
3241	 * By setting it to reflect the virtual start address of the
3242	 * vma, merges and splits can happen in a seamless way, just
3243	 * using the existing file pgoff checks and manipulations.
3244	 * Similarly in do_mmap and in do_brk_flags.
3245	 */
3246	if (vma_is_anonymous(vma)) {
3247		BUG_ON(vma->anon_vma);
3248		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3249	}
3250
3251	vma_link(mm, vma, prev, rb_link, rb_parent);
3252	return 0;
3253}
3254
3255/*
3256 * Copy the vma structure to a new location in the same mm,
3257 * prior to moving page table entries, to effect an mremap move.
3258 */
3259struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3260	unsigned long addr, unsigned long len, pgoff_t pgoff,
3261	bool *need_rmap_locks)
3262{
3263	struct vm_area_struct *vma = *vmap;
3264	unsigned long vma_start = vma->vm_start;
3265	struct mm_struct *mm = vma->vm_mm;
3266	struct vm_area_struct *new_vma, *prev;
3267	struct rb_node **rb_link, *rb_parent;
3268	bool faulted_in_anon_vma = true;
3269
3270	/*
3271	 * If anonymous vma has not yet been faulted, update new pgoff
3272	 * to match new location, to increase its chance of merging.
3273	 */
3274	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3275		pgoff = addr >> PAGE_SHIFT;
3276		faulted_in_anon_vma = false;
3277	}
3278
3279	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3280		return NULL;	/* should never get here */
3281	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3282			    vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3283			    vma->vm_userfaultfd_ctx);
3284	if (new_vma) {
3285		/*
3286		 * Source vma may have been merged into new_vma
3287		 */
3288		if (unlikely(vma_start >= new_vma->vm_start &&
3289			     vma_start < new_vma->vm_end)) {
3290			/*
3291			 * The only way we can get a vma_merge with
3292			 * self during an mremap is if the vma hasn't
3293			 * been faulted in yet and we were allowed to
3294			 * reset the dst vma->vm_pgoff to the
3295			 * destination address of the mremap to allow
3296			 * the merge to happen. mremap must change the
3297			 * vm_pgoff linearity between src and dst vmas
3298			 * (in turn preventing a vma_merge) to be
3299			 * safe. It is only safe to keep the vm_pgoff
3300			 * linear if there are no pages mapped yet.
3301			 */
3302			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3303			*vmap = vma = new_vma;
3304		}
3305		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3306	} else {
3307		new_vma = vm_area_dup(vma);
3308		if (!new_vma)
3309			goto out;
 
3310		new_vma->vm_start = addr;
3311		new_vma->vm_end = addr + len;
3312		new_vma->vm_pgoff = pgoff;
3313		if (vma_dup_policy(vma, new_vma))
3314			goto out_free_vma;
 
3315		if (anon_vma_clone(new_vma, vma))
3316			goto out_free_mempol;
3317		if (new_vma->vm_file)
3318			get_file(new_vma->vm_file);
3319		if (new_vma->vm_ops && new_vma->vm_ops->open)
3320			new_vma->vm_ops->open(new_vma);
3321		vma_link(mm, new_vma, prev, rb_link, rb_parent);
3322		*need_rmap_locks = false;
3323	}
3324	return new_vma;
3325
3326out_free_mempol:
3327	mpol_put(vma_policy(new_vma));
3328out_free_vma:
3329	vm_area_free(new_vma);
3330out:
3331	return NULL;
3332}
3333
3334/*
3335 * Return true if the calling process may expand its vm space by the passed
3336 * number of pages
3337 */
3338bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3339{
3340	if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3341		return false;
3342
3343	if (is_data_mapping(flags) &&
3344	    mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3345		/* Workaround for Valgrind */
3346		if (rlimit(RLIMIT_DATA) == 0 &&
3347		    mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3348			return true;
3349
3350		pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3351			     current->comm, current->pid,
3352			     (mm->data_vm + npages) << PAGE_SHIFT,
3353			     rlimit(RLIMIT_DATA),
3354			     ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3355
3356		if (!ignore_rlimit_data)
3357			return false;
3358	}
3359
3360	return true;
3361}
3362
3363void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3364{
3365	mm->total_vm += npages;
3366
3367	if (is_exec_mapping(flags))
3368		mm->exec_vm += npages;
3369	else if (is_stack_mapping(flags))
3370		mm->stack_vm += npages;
3371	else if (is_data_mapping(flags))
3372		mm->data_vm += npages;
3373}
3374
3375static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
 
3376
3377/*
3378 * Having a close hook prevents vma merging regardless of flags.
3379 */
3380static void special_mapping_close(struct vm_area_struct *vma)
3381{
3382}
3383
3384static const char *special_mapping_name(struct vm_area_struct *vma)
3385{
3386	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3387}
3388
3389static int special_mapping_mremap(struct vm_area_struct *new_vma)
3390{
3391	struct vm_special_mapping *sm = new_vma->vm_private_data;
3392
3393	if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3394		return -EFAULT;
3395
3396	if (sm->mremap)
3397		return sm->mremap(sm, new_vma);
3398
3399	return 0;
3400}
3401
3402static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3403{
3404	/*
3405	 * Forbid splitting special mappings - kernel has expectations over
3406	 * the number of pages in mapping. Together with VM_DONTEXPAND
3407	 * the size of vma should stay the same over the special mapping's
3408	 * lifetime.
3409	 */
3410	return -EINVAL;
3411}
3412
3413static const struct vm_operations_struct special_mapping_vmops = {
3414	.close = special_mapping_close,
3415	.fault = special_mapping_fault,
3416	.mremap = special_mapping_mremap,
3417	.name = special_mapping_name,
3418	/* vDSO code relies that VVAR can't be accessed remotely */
3419	.access = NULL,
3420	.may_split = special_mapping_split,
3421};
3422
3423static const struct vm_operations_struct legacy_special_mapping_vmops = {
3424	.close = special_mapping_close,
3425	.fault = special_mapping_fault,
3426};
3427
3428static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
 
3429{
3430	struct vm_area_struct *vma = vmf->vma;
3431	pgoff_t pgoff;
3432	struct page **pages;
3433
3434	if (vma->vm_ops == &legacy_special_mapping_vmops) {
3435		pages = vma->vm_private_data;
3436	} else {
3437		struct vm_special_mapping *sm = vma->vm_private_data;
3438
3439		if (sm->fault)
3440			return sm->fault(sm, vmf->vma, vmf);
3441
3442		pages = sm->pages;
3443	}
3444
3445	for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3446		pgoff--;
3447
3448	if (*pages) {
3449		struct page *page = *pages;
3450		get_page(page);
3451		vmf->page = page;
3452		return 0;
3453	}
3454
3455	return VM_FAULT_SIGBUS;
3456}
3457
3458static struct vm_area_struct *__install_special_mapping(
3459	struct mm_struct *mm,
3460	unsigned long addr, unsigned long len,
3461	unsigned long vm_flags, void *priv,
3462	const struct vm_operations_struct *ops)
3463{
3464	int ret;
3465	struct vm_area_struct *vma;
3466
3467	vma = vm_area_alloc(mm);
3468	if (unlikely(vma == NULL))
3469		return ERR_PTR(-ENOMEM);
3470
 
 
3471	vma->vm_start = addr;
3472	vma->vm_end = addr + len;
3473
3474	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3475	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3476
3477	vma->vm_ops = ops;
3478	vma->vm_private_data = priv;
3479
3480	ret = insert_vm_struct(mm, vma);
3481	if (ret)
3482		goto out;
3483
3484	vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3485
3486	perf_event_mmap(vma);
3487
3488	return vma;
3489
3490out:
3491	vm_area_free(vma);
3492	return ERR_PTR(ret);
3493}
3494
3495bool vma_is_special_mapping(const struct vm_area_struct *vma,
3496	const struct vm_special_mapping *sm)
3497{
3498	return vma->vm_private_data == sm &&
3499		(vma->vm_ops == &special_mapping_vmops ||
3500		 vma->vm_ops == &legacy_special_mapping_vmops);
3501}
3502
3503/*
3504 * Called with mm->mmap_lock held for writing.
3505 * Insert a new vma covering the given region, with the given flags.
3506 * Its pages are supplied by the given array of struct page *.
3507 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3508 * The region past the last page supplied will always produce SIGBUS.
3509 * The array pointer and the pages it points to are assumed to stay alive
3510 * for as long as this mapping might exist.
3511 */
3512struct vm_area_struct *_install_special_mapping(
3513	struct mm_struct *mm,
3514	unsigned long addr, unsigned long len,
3515	unsigned long vm_flags, const struct vm_special_mapping *spec)
3516{
3517	return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3518					&special_mapping_vmops);
3519}
3520
3521int install_special_mapping(struct mm_struct *mm,
3522			    unsigned long addr, unsigned long len,
3523			    unsigned long vm_flags, struct page **pages)
3524{
3525	struct vm_area_struct *vma = __install_special_mapping(
3526		mm, addr, len, vm_flags, (void *)pages,
3527		&legacy_special_mapping_vmops);
3528
3529	return PTR_ERR_OR_ZERO(vma);
3530}
3531
3532static DEFINE_MUTEX(mm_all_locks_mutex);
3533
3534static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3535{
3536	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3537		/*
3538		 * The LSB of head.next can't change from under us
3539		 * because we hold the mm_all_locks_mutex.
3540		 */
3541		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3542		/*
3543		 * We can safely modify head.next after taking the
3544		 * anon_vma->root->rwsem. If some other vma in this mm shares
3545		 * the same anon_vma we won't take it again.
3546		 *
3547		 * No need of atomic instructions here, head.next
3548		 * can't change from under us thanks to the
3549		 * anon_vma->root->rwsem.
3550		 */
3551		if (__test_and_set_bit(0, (unsigned long *)
3552				       &anon_vma->root->rb_root.rb_root.rb_node))
3553			BUG();
3554	}
3555}
3556
3557static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3558{
3559	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3560		/*
3561		 * AS_MM_ALL_LOCKS can't change from under us because
3562		 * we hold the mm_all_locks_mutex.
3563		 *
3564		 * Operations on ->flags have to be atomic because
3565		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3566		 * mm_all_locks_mutex, there may be other cpus
3567		 * changing other bitflags in parallel to us.
3568		 */
3569		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3570			BUG();
3571		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3572	}
3573}
3574
3575/*
3576 * This operation locks against the VM for all pte/vma/mm related
3577 * operations that could ever happen on a certain mm. This includes
3578 * vmtruncate, try_to_unmap, and all page faults.
3579 *
3580 * The caller must take the mmap_lock in write mode before calling
3581 * mm_take_all_locks(). The caller isn't allowed to release the
3582 * mmap_lock until mm_drop_all_locks() returns.
3583 *
3584 * mmap_lock in write mode is required in order to block all operations
3585 * that could modify pagetables and free pages without need of
3586 * altering the vma layout. It's also needed in write mode to avoid new
3587 * anon_vmas to be associated with existing vmas.
3588 *
3589 * A single task can't take more than one mm_take_all_locks() in a row
3590 * or it would deadlock.
3591 *
3592 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3593 * mapping->flags avoid to take the same lock twice, if more than one
3594 * vma in this mm is backed by the same anon_vma or address_space.
3595 *
3596 * We take locks in following order, accordingly to comment at beginning
3597 * of mm/rmap.c:
3598 *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3599 *     hugetlb mapping);
3600 *   - all i_mmap_rwsem locks;
3601 *   - all anon_vma->rwseml
3602 *
3603 * We can take all locks within these types randomly because the VM code
3604 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3605 * mm_all_locks_mutex.
3606 *
3607 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3608 * that may have to take thousand of locks.
3609 *
3610 * mm_take_all_locks() can fail if it's interrupted by signals.
3611 */
3612int mm_take_all_locks(struct mm_struct *mm)
3613{
3614	struct vm_area_struct *vma;
3615	struct anon_vma_chain *avc;
3616
3617	BUG_ON(mmap_read_trylock(mm));
3618
3619	mutex_lock(&mm_all_locks_mutex);
3620
3621	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3622		if (signal_pending(current))
3623			goto out_unlock;
3624		if (vma->vm_file && vma->vm_file->f_mapping &&
3625				is_vm_hugetlb_page(vma))
3626			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3627	}
3628
3629	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3630		if (signal_pending(current))
3631			goto out_unlock;
3632		if (vma->vm_file && vma->vm_file->f_mapping &&
3633				!is_vm_hugetlb_page(vma))
3634			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3635	}
3636
3637	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3638		if (signal_pending(current))
3639			goto out_unlock;
3640		if (vma->anon_vma)
3641			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3642				vm_lock_anon_vma(mm, avc->anon_vma);
3643	}
3644
3645	return 0;
3646
3647out_unlock:
3648	mm_drop_all_locks(mm);
3649	return -EINTR;
3650}
3651
3652static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3653{
3654	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3655		/*
3656		 * The LSB of head.next can't change to 0 from under
3657		 * us because we hold the mm_all_locks_mutex.
3658		 *
3659		 * We must however clear the bitflag before unlocking
3660		 * the vma so the users using the anon_vma->rb_root will
3661		 * never see our bitflag.
3662		 *
3663		 * No need of atomic instructions here, head.next
3664		 * can't change from under us until we release the
3665		 * anon_vma->root->rwsem.
3666		 */
3667		if (!__test_and_clear_bit(0, (unsigned long *)
3668					  &anon_vma->root->rb_root.rb_root.rb_node))
3669			BUG();
3670		anon_vma_unlock_write(anon_vma);
3671	}
3672}
3673
3674static void vm_unlock_mapping(struct address_space *mapping)
3675{
3676	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3677		/*
3678		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3679		 * because we hold the mm_all_locks_mutex.
3680		 */
3681		i_mmap_unlock_write(mapping);
3682		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3683					&mapping->flags))
3684			BUG();
3685	}
3686}
3687
3688/*
3689 * The mmap_lock cannot be released by the caller until
3690 * mm_drop_all_locks() returns.
3691 */
3692void mm_drop_all_locks(struct mm_struct *mm)
3693{
3694	struct vm_area_struct *vma;
3695	struct anon_vma_chain *avc;
3696
3697	BUG_ON(mmap_read_trylock(mm));
3698	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3699
3700	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3701		if (vma->anon_vma)
3702			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3703				vm_unlock_anon_vma(avc->anon_vma);
3704		if (vma->vm_file && vma->vm_file->f_mapping)
3705			vm_unlock_mapping(vma->vm_file->f_mapping);
3706	}
3707
3708	mutex_unlock(&mm_all_locks_mutex);
3709}
3710
3711/*
3712 * initialise the percpu counter for VM
3713 */
3714void __init mmap_init(void)
3715{
3716	int ret;
3717
3718	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3719	VM_BUG_ON(ret);
3720}
3721
3722/*
3723 * Initialise sysctl_user_reserve_kbytes.
3724 *
3725 * This is intended to prevent a user from starting a single memory hogging
3726 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3727 * mode.
3728 *
3729 * The default value is min(3% of free memory, 128MB)
3730 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3731 */
3732static int init_user_reserve(void)
3733{
3734	unsigned long free_kbytes;
3735
3736	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3737
3738	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3739	return 0;
3740}
3741subsys_initcall(init_user_reserve);
3742
3743/*
3744 * Initialise sysctl_admin_reserve_kbytes.
3745 *
3746 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3747 * to log in and kill a memory hogging process.
3748 *
3749 * Systems with more than 256MB will reserve 8MB, enough to recover
3750 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3751 * only reserve 3% of free pages by default.
3752 */
3753static int init_admin_reserve(void)
3754{
3755	unsigned long free_kbytes;
3756
3757	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3758
3759	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3760	return 0;
3761}
3762subsys_initcall(init_admin_reserve);
3763
3764/*
3765 * Reinititalise user and admin reserves if memory is added or removed.
3766 *
3767 * The default user reserve max is 128MB, and the default max for the
3768 * admin reserve is 8MB. These are usually, but not always, enough to
3769 * enable recovery from a memory hogging process using login/sshd, a shell,
3770 * and tools like top. It may make sense to increase or even disable the
3771 * reserve depending on the existence of swap or variations in the recovery
3772 * tools. So, the admin may have changed them.
3773 *
3774 * If memory is added and the reserves have been eliminated or increased above
3775 * the default max, then we'll trust the admin.
3776 *
3777 * If memory is removed and there isn't enough free memory, then we
3778 * need to reset the reserves.
3779 *
3780 * Otherwise keep the reserve set by the admin.
3781 */
3782static int reserve_mem_notifier(struct notifier_block *nb,
3783			     unsigned long action, void *data)
3784{
3785	unsigned long tmp, free_kbytes;
3786
3787	switch (action) {
3788	case MEM_ONLINE:
3789		/* Default max is 128MB. Leave alone if modified by operator. */
3790		tmp = sysctl_user_reserve_kbytes;
3791		if (0 < tmp && tmp < (1UL << 17))
3792			init_user_reserve();
3793
3794		/* Default max is 8MB.  Leave alone if modified by operator. */
3795		tmp = sysctl_admin_reserve_kbytes;
3796		if (0 < tmp && tmp < (1UL << 13))
3797			init_admin_reserve();
3798
3799		break;
3800	case MEM_OFFLINE:
3801		free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3802
3803		if (sysctl_user_reserve_kbytes > free_kbytes) {
3804			init_user_reserve();
3805			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3806				sysctl_user_reserve_kbytes);
3807		}
3808
3809		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3810			init_admin_reserve();
3811			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3812				sysctl_admin_reserve_kbytes);
3813		}
3814		break;
3815	default:
3816		break;
3817	}
3818	return NOTIFY_OK;
3819}
3820
3821static struct notifier_block reserve_mem_nb = {
3822	.notifier_call = reserve_mem_notifier,
3823};
3824
3825static int __meminit init_reserve_notifier(void)
3826{
3827	if (register_hotmemory_notifier(&reserve_mem_nb))
3828		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3829
3830	return 0;
3831}
3832subsys_initcall(init_reserve_notifier);