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