Linux Audio

Check our new training course

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