Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v3.1
   1/*
   2 * mm/mmap.c
   3 *
   4 * Written by obz.
   5 *
   6 * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
   7 */
   8
 
   9#include <linux/slab.h>
  10#include <linux/backing-dev.h>
  11#include <linux/mm.h>
 
  12#include <linux/shm.h>
  13#include <linux/mman.h>
  14#include <linux/pagemap.h>
  15#include <linux/swap.h>
  16#include <linux/syscalls.h>
  17#include <linux/capability.h>
  18#include <linux/init.h>
  19#include <linux/file.h>
  20#include <linux/fs.h>
  21#include <linux/personality.h>
  22#include <linux/security.h>
  23#include <linux/hugetlb.h>
  24#include <linux/profile.h>
  25#include <linux/module.h>
  26#include <linux/mount.h>
  27#include <linux/mempolicy.h>
  28#include <linux/rmap.h>
  29#include <linux/mmu_notifier.h>
  30#include <linux/perf_event.h>
  31#include <linux/audit.h>
  32#include <linux/khugepaged.h>
 
 
 
 
 
  33
  34#include <asm/uaccess.h>
  35#include <asm/cacheflush.h>
  36#include <asm/tlb.h>
  37#include <asm/mmu_context.h>
  38
  39#include "internal.h"
  40
  41#ifndef arch_mmap_check
  42#define arch_mmap_check(addr, len, flags)	(0)
  43#endif
  44
  45#ifndef arch_rebalance_pgtables
  46#define arch_rebalance_pgtables(addr, len)		(addr)
  47#endif
  48
  49static void unmap_region(struct mm_struct *mm,
  50		struct vm_area_struct *vma, struct vm_area_struct *prev,
  51		unsigned long start, unsigned long end);
  52
  53/*
  54 * WARNING: the debugging will use recursive algorithms so never enable this
  55 * unless you know what you are doing.
  56 */
  57#undef DEBUG_MM_RB
  58
  59/* description of effects of mapping type and prot in current implementation.
  60 * this is due to the limited x86 page protection hardware.  The expected
  61 * behavior is in parens:
  62 *
  63 * map_type	prot
  64 *		PROT_NONE	PROT_READ	PROT_WRITE	PROT_EXEC
  65 * MAP_SHARED	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  66 *		w: (no) no	w: (no) no	w: (yes) yes	w: (no) no
  67 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  68 *		
  69 * MAP_PRIVATE	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  70 *		w: (no) no	w: (no) no	w: (copy) copy	w: (no) no
  71 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  72 *
  73 */
  74pgprot_t protection_map[16] = {
  75	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
  76	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
  77};
  78
  79pgprot_t vm_get_page_prot(unsigned long vm_flags)
  80{
  81	return __pgprot(pgprot_val(protection_map[vm_flags &
  82				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
  83			pgprot_val(arch_vm_get_page_prot(vm_flags)));
  84}
  85EXPORT_SYMBOL(vm_get_page_prot);
  86
  87int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
  88int sysctl_overcommit_ratio __read_mostly = 50;	/* default is 50% */
 
  89int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
 
 
  90/*
  91 * Make sure vm_committed_as in one cacheline and not cacheline shared with
  92 * other variables. It can be updated by several CPUs frequently.
  93 */
  94struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
  95
  96/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  97 * Check that a process has enough memory to allocate a new virtual
  98 * mapping. 0 means there is enough memory for the allocation to
  99 * succeed and -ENOMEM implies there is not.
 100 *
 101 * We currently support three overcommit policies, which are set via the
 102 * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
 103 *
 104 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
 105 * Additional code 2002 Jul 20 by Robert Love.
 106 *
 107 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
 108 *
 109 * Note this is a helper function intended to be used by LSMs which
 110 * wish to use this logic.
 111 */
 112int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
 113{
 114	unsigned long free, allowed;
 115
 116	vm_acct_memory(pages);
 117
 118	/*
 119	 * Sometimes we want to use more memory than we have
 120	 */
 121	if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
 122		return 0;
 123
 124	if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
 125		free = global_page_state(NR_FREE_PAGES);
 126		free += global_page_state(NR_FILE_PAGES);
 127
 128		/*
 129		 * shmem pages shouldn't be counted as free in this
 130		 * case, they can't be purged, only swapped out, and
 131		 * that won't affect the overall amount of available
 132		 * memory in the system.
 133		 */
 134		free -= global_page_state(NR_SHMEM);
 135
 136		free += nr_swap_pages;
 137
 138		/*
 139		 * Any slabs which are created with the
 140		 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
 141		 * which are reclaimable, under pressure.  The dentry
 142		 * cache and most inode caches should fall into this
 143		 */
 144		free += global_page_state(NR_SLAB_RECLAIMABLE);
 145
 146		/*
 147		 * Leave reserved pages. The pages are not for anonymous pages.
 148		 */
 149		if (free <= totalreserve_pages)
 150			goto error;
 151		else
 152			free -= totalreserve_pages;
 153
 154		/*
 155		 * Leave the last 3% for root
 156		 */
 157		if (!cap_sys_admin)
 158			free -= free / 32;
 159
 160		if (free > pages)
 161			return 0;
 162
 163		goto error;
 164	}
 165
 166	allowed = (totalram_pages - hugetlb_total_pages())
 167	       	* sysctl_overcommit_ratio / 100;
 168	/*
 169	 * Leave the last 3% for root
 170	 */
 171	if (!cap_sys_admin)
 172		allowed -= allowed / 32;
 173	allowed += total_swap_pages;
 174
 175	/* Don't let a single process grow too big:
 176	   leave 3% of the size of this process for other processes */
 177	if (mm)
 178		allowed -= mm->total_vm / 32;
 
 
 
 179
 180	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
 181		return 0;
 182error:
 183	vm_unacct_memory(pages);
 184
 185	return -ENOMEM;
 186}
 187
 188/*
 189 * Requires inode->i_mapping->i_mmap_mutex
 190 */
 191static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 192		struct file *file, struct address_space *mapping)
 193{
 194	if (vma->vm_flags & VM_DENYWRITE)
 195		atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
 196	if (vma->vm_flags & VM_SHARED)
 197		mapping->i_mmap_writable--;
 198
 199	flush_dcache_mmap_lock(mapping);
 200	if (unlikely(vma->vm_flags & VM_NONLINEAR))
 201		list_del_init(&vma->shared.vm_set.list);
 202	else
 203		vma_prio_tree_remove(vma, &mapping->i_mmap);
 204	flush_dcache_mmap_unlock(mapping);
 205}
 206
 207/*
 208 * Unlink a file-based vm structure from its prio_tree, to hide
 209 * vma from rmap and vmtruncate before freeing its page tables.
 210 */
 211void unlink_file_vma(struct vm_area_struct *vma)
 212{
 213	struct file *file = vma->vm_file;
 214
 215	if (file) {
 216		struct address_space *mapping = file->f_mapping;
 217		mutex_lock(&mapping->i_mmap_mutex);
 218		__remove_shared_vm_struct(vma, file, mapping);
 219		mutex_unlock(&mapping->i_mmap_mutex);
 220	}
 221}
 222
 223/*
 224 * Close a vm structure and free it, returning the next.
 225 */
 226static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
 227{
 228	struct vm_area_struct *next = vma->vm_next;
 229
 230	might_sleep();
 231	if (vma->vm_ops && vma->vm_ops->close)
 232		vma->vm_ops->close(vma);
 233	if (vma->vm_file) {
 234		fput(vma->vm_file);
 235		if (vma->vm_flags & VM_EXECUTABLE)
 236			removed_exe_file_vma(vma->vm_mm);
 237	}
 238	mpol_put(vma_policy(vma));
 239	kmem_cache_free(vm_area_cachep, vma);
 240	return next;
 241}
 242
 
 
 243SYSCALL_DEFINE1(brk, unsigned long, brk)
 244{
 245	unsigned long rlim, retval;
 246	unsigned long newbrk, oldbrk;
 247	struct mm_struct *mm = current->mm;
 248	unsigned long min_brk;
 
 249
 250	down_write(&mm->mmap_sem);
 251
 252#ifdef CONFIG_COMPAT_BRK
 253	/*
 254	 * CONFIG_COMPAT_BRK can still be overridden by setting
 255	 * randomize_va_space to 2, which will still cause mm->start_brk
 256	 * to be arbitrarily shifted
 257	 */
 258	if (current->brk_randomized)
 259		min_brk = mm->start_brk;
 260	else
 261		min_brk = mm->end_data;
 262#else
 263	min_brk = mm->start_brk;
 264#endif
 265	if (brk < min_brk)
 266		goto out;
 267
 268	/*
 269	 * Check against rlimit here. If this check is done later after the test
 270	 * of oldbrk with newbrk then it can escape the test and let the data
 271	 * segment grow beyond its set limit the in case where the limit is
 272	 * not page aligned -Ram Gupta
 273	 */
 274	rlim = rlimit(RLIMIT_DATA);
 275	if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
 276			(mm->end_data - mm->start_data) > rlim)
 277		goto out;
 278
 279	newbrk = PAGE_ALIGN(brk);
 280	oldbrk = PAGE_ALIGN(mm->brk);
 281	if (oldbrk == newbrk)
 282		goto set_brk;
 283
 284	/* Always allow shrinking brk. */
 285	if (brk <= mm->brk) {
 286		if (!do_munmap(mm, newbrk, oldbrk-newbrk))
 287			goto set_brk;
 288		goto out;
 289	}
 290
 291	/* Check against existing mmap mappings. */
 292	if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
 293		goto out;
 294
 295	/* Ok, looks good - let it rip. */
 296	if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
 297		goto out;
 
 298set_brk:
 299	mm->brk = brk;
 
 
 
 
 
 
 300out:
 301	retval = mm->brk;
 302	up_write(&mm->mmap_sem);
 303	return retval;
 304}
 305
 306#ifdef DEBUG_MM_RB
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 307static int browse_rb(struct rb_root *root)
 308{
 309	int i = 0, j;
 310	struct rb_node *nd, *pn = NULL;
 311	unsigned long prev = 0, pend = 0;
 312
 313	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 314		struct vm_area_struct *vma;
 315		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 316		if (vma->vm_start < prev)
 317			printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
 318		if (vma->vm_start < pend)
 
 
 319			printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
 320		if (vma->vm_start > vma->vm_end)
 321			printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
 
 
 
 
 
 
 
 
 
 
 
 322		i++;
 323		pn = nd;
 324		prev = vma->vm_start;
 325		pend = vma->vm_end;
 326	}
 327	j = 0;
 328	for (nd = pn; nd; nd = rb_prev(nd)) {
 329		j++;
 
 
 
 330	}
 331	if (i != j)
 332		printk("backwards %d, forwards %d\n", j, i), i = 0;
 333	return i;
 334}
 335
 336void validate_mm(struct mm_struct *mm)
 
 
 
 
 
 
 
 
 
 
 
 
 337{
 338	int bug = 0;
 339	int i = 0;
 340	struct vm_area_struct *tmp = mm->mmap;
 341	while (tmp) {
 342		tmp = tmp->vm_next;
 
 
 
 
 
 
 
 343		i++;
 344	}
 345	if (i != mm->map_count)
 346		printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
 
 
 
 
 
 
 
 347	i = browse_rb(&mm->mm_rb);
 348	if (i != mm->map_count)
 349		printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
 
 
 350	BUG_ON(bug);
 351}
 352#else
 
 353#define validate_mm(mm) do { } while (0)
 354#endif
 355
 356static struct vm_area_struct *
 357find_vma_prepare(struct mm_struct *mm, unsigned long addr,
 358		struct vm_area_struct **pprev, struct rb_node ***rb_link,
 359		struct rb_node ** rb_parent)
 
 
 
 
 
 360{
 361	struct vm_area_struct * vma;
 362	struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 363
 364	__rb_link = &mm->mm_rb.rb_node;
 365	rb_prev = __rb_parent = NULL;
 366	vma = NULL;
 367
 368	while (*__rb_link) {
 369		struct vm_area_struct *vma_tmp;
 370
 371		__rb_parent = *__rb_link;
 372		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
 373
 374		if (vma_tmp->vm_end > addr) {
 375			vma = vma_tmp;
 376			if (vma_tmp->vm_start <= addr)
 377				break;
 378			__rb_link = &__rb_parent->rb_left;
 379		} else {
 380			rb_prev = __rb_parent;
 381			__rb_link = &__rb_parent->rb_right;
 382		}
 383	}
 384
 385	*pprev = NULL;
 386	if (rb_prev)
 387		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 388	*rb_link = __rb_link;
 389	*rb_parent = __rb_parent;
 390	return vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 391}
 392
 393void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 394		struct rb_node **rb_link, struct rb_node *rb_parent)
 395{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 396	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 397	rb_insert_color(&vma->vm_rb, &mm->mm_rb);
 
 
 398}
 399
 400static void __vma_link_file(struct vm_area_struct *vma)
 401{
 402	struct file *file;
 403
 404	file = vma->vm_file;
 405	if (file) {
 406		struct address_space *mapping = file->f_mapping;
 407
 408		if (vma->vm_flags & VM_DENYWRITE)
 409			atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
 410		if (vma->vm_flags & VM_SHARED)
 411			mapping->i_mmap_writable++;
 412
 413		flush_dcache_mmap_lock(mapping);
 414		if (unlikely(vma->vm_flags & VM_NONLINEAR))
 415			vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 416		else
 417			vma_prio_tree_insert(vma, &mapping->i_mmap);
 418		flush_dcache_mmap_unlock(mapping);
 419	}
 420}
 421
 422static void
 423__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 424	struct vm_area_struct *prev, struct rb_node **rb_link,
 425	struct rb_node *rb_parent)
 426{
 427	__vma_link_list(mm, vma, prev, rb_parent);
 428	__vma_link_rb(mm, vma, rb_link, rb_parent);
 429}
 430
 431static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 432			struct vm_area_struct *prev, struct rb_node **rb_link,
 433			struct rb_node *rb_parent)
 434{
 435	struct address_space *mapping = NULL;
 436
 437	if (vma->vm_file)
 438		mapping = vma->vm_file->f_mapping;
 439
 440	if (mapping)
 441		mutex_lock(&mapping->i_mmap_mutex);
 442
 443	__vma_link(mm, vma, prev, rb_link, rb_parent);
 444	__vma_link_file(vma);
 445
 446	if (mapping)
 447		mutex_unlock(&mapping->i_mmap_mutex);
 448
 449	mm->map_count++;
 450	validate_mm(mm);
 451}
 452
 453/*
 454 * Helper for vma_adjust in the split_vma insert case:
 455 * insert vm structure into list and rbtree and anon_vma,
 456 * but it has already been inserted into prio_tree earlier.
 457 */
 458static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 459{
 460	struct vm_area_struct *__vma, *prev;
 461	struct rb_node **rb_link, *rb_parent;
 462
 463	__vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
 464	BUG_ON(__vma && __vma->vm_start < vma->vm_end);
 
 465	__vma_link(mm, vma, prev, rb_link, rb_parent);
 466	mm->map_count++;
 467}
 468
 469static inline void
 470__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
 471		struct vm_area_struct *prev)
 472{
 473	struct vm_area_struct *next = vma->vm_next;
 474
 475	prev->vm_next = next;
 
 476	if (next)
 477		next->vm_prev = prev;
 478	rb_erase(&vma->vm_rb, &mm->mm_rb);
 479	if (mm->mmap_cache == vma)
 480		mm->mmap_cache = prev;
 481}
 482
 483/*
 484 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
 485 * is already present in an i_mmap tree without adjusting the tree.
 486 * The following helper function should be used when such adjustments
 487 * are necessary.  The "insert" vma (if any) is to be inserted
 488 * before we drop the necessary locks.
 489 */
 490int vma_adjust(struct vm_area_struct *vma, unsigned long start,
 491	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
 492{
 493	struct mm_struct *mm = vma->vm_mm;
 494	struct vm_area_struct *next = vma->vm_next;
 495	struct vm_area_struct *importer = NULL;
 496	struct address_space *mapping = NULL;
 497	struct prio_tree_root *root = NULL;
 498	struct anon_vma *anon_vma = NULL;
 499	struct file *file = vma->vm_file;
 
 500	long adjust_next = 0;
 501	int remove_next = 0;
 502
 503	if (next && !insert) {
 504		struct vm_area_struct *exporter = NULL;
 505
 506		if (end >= next->vm_end) {
 507			/*
 508			 * vma expands, overlapping all the next, and
 509			 * perhaps the one after too (mprotect case 6).
 510			 */
 511again:			remove_next = 1 + (end > next->vm_end);
 512			end = next->vm_end;
 513			exporter = next;
 514			importer = vma;
 515		} else if (end > next->vm_start) {
 516			/*
 517			 * vma expands, overlapping part of the next:
 518			 * mprotect case 5 shifting the boundary up.
 519			 */
 520			adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
 521			exporter = next;
 522			importer = vma;
 523		} else if (end < vma->vm_end) {
 524			/*
 525			 * vma shrinks, and !insert tells it's not
 526			 * split_vma inserting another: so it must be
 527			 * mprotect case 4 shifting the boundary down.
 528			 */
 529			adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
 530			exporter = vma;
 531			importer = next;
 532		}
 533
 534		/*
 535		 * Easily overlooked: when mprotect shifts the boundary,
 536		 * make sure the expanding vma has anon_vma set if the
 537		 * shrinking vma had, to cover any anon pages imported.
 538		 */
 539		if (exporter && exporter->anon_vma && !importer->anon_vma) {
 540			if (anon_vma_clone(importer, exporter))
 541				return -ENOMEM;
 542			importer->anon_vma = exporter->anon_vma;
 543		}
 544	}
 545
 546	if (file) {
 547		mapping = file->f_mapping;
 548		if (!(vma->vm_flags & VM_NONLINEAR))
 549			root = &mapping->i_mmap;
 
 
 
 
 
 
 
 550		mutex_lock(&mapping->i_mmap_mutex);
 551		if (insert) {
 552			/*
 553			 * Put into prio_tree now, so instantiated pages
 554			 * are visible to arm/parisc __flush_dcache_page
 555			 * throughout; but we cannot insert into address
 556			 * space until vma start or end is updated.
 557			 */
 558			__vma_link_file(insert);
 559		}
 560	}
 561
 562	vma_adjust_trans_huge(vma, start, end, adjust_next);
 563
 564	/*
 565	 * When changing only vma->vm_end, we don't really need anon_vma
 566	 * lock. This is a fairly rare case by itself, but the anon_vma
 567	 * lock may be shared between many sibling processes.  Skipping
 568	 * the lock for brk adjustments makes a difference sometimes.
 569	 */
 570	if (vma->anon_vma && (importer || start != vma->vm_start)) {
 571		anon_vma = vma->anon_vma;
 572		anon_vma_lock(anon_vma);
 
 573	}
 574
 575	if (root) {
 576		flush_dcache_mmap_lock(mapping);
 577		vma_prio_tree_remove(vma, root);
 578		if (adjust_next)
 579			vma_prio_tree_remove(next, root);
 580	}
 581
 582	vma->vm_start = start;
 583	vma->vm_end = end;
 
 
 
 
 
 
 584	vma->vm_pgoff = pgoff;
 585	if (adjust_next) {
 586		next->vm_start += adjust_next << PAGE_SHIFT;
 587		next->vm_pgoff += adjust_next;
 588	}
 589
 590	if (root) {
 591		if (adjust_next)
 592			vma_prio_tree_insert(next, root);
 593		vma_prio_tree_insert(vma, root);
 594		flush_dcache_mmap_unlock(mapping);
 595	}
 596
 597	if (remove_next) {
 598		/*
 599		 * vma_merge has merged next into vma, and needs
 600		 * us to remove next before dropping the locks.
 601		 */
 602		__vma_unlink(mm, next, vma);
 603		if (file)
 604			__remove_shared_vm_struct(next, file, mapping);
 605	} else if (insert) {
 606		/*
 607		 * split_vma has split insert from vma, and needs
 608		 * us to insert it before dropping the locks
 609		 * (it may either follow vma or precede it).
 610		 */
 611		__insert_vm_struct(mm, insert);
 
 
 
 
 
 
 
 
 
 612	}
 613
 614	if (anon_vma)
 615		anon_vma_unlock(anon_vma);
 
 
 
 
 616	if (mapping)
 617		mutex_unlock(&mapping->i_mmap_mutex);
 618
 
 
 
 
 
 
 
 619	if (remove_next) {
 620		if (file) {
 
 621			fput(file);
 622			if (next->vm_flags & VM_EXECUTABLE)
 623				removed_exe_file_vma(mm);
 624		}
 625		if (next->anon_vma)
 626			anon_vma_merge(vma, next);
 627		mm->map_count--;
 628		mpol_put(vma_policy(next));
 629		kmem_cache_free(vm_area_cachep, next);
 630		/*
 631		 * In mprotect's case 6 (see comments on vma_merge),
 632		 * we must remove another next too. It would clutter
 633		 * up the code too much to do both in one go.
 634		 */
 635		if (remove_next == 2) {
 636			next = vma->vm_next;
 637			goto again;
 638		}
 
 
 
 639	}
 
 
 640
 641	validate_mm(mm);
 642
 643	return 0;
 644}
 645
 646/*
 647 * If the vma has a ->close operation then the driver probably needs to release
 648 * per-vma resources, so we don't attempt to merge those.
 649 */
 650static inline int is_mergeable_vma(struct vm_area_struct *vma,
 651			struct file *file, unsigned long vm_flags)
 652{
 653	/* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
 654	if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
 
 
 
 
 
 
 
 655		return 0;
 656	if (vma->vm_file != file)
 657		return 0;
 658	if (vma->vm_ops && vma->vm_ops->close)
 659		return 0;
 660	return 1;
 661}
 662
 663static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
 664					struct anon_vma *anon_vma2,
 665					struct vm_area_struct *vma)
 666{
 667	/*
 668	 * The list_is_singular() test is to avoid merging VMA cloned from
 669	 * parents. This can improve scalability caused by anon_vma lock.
 670	 */
 671	if ((!anon_vma1 || !anon_vma2) && (!vma ||
 672		list_is_singular(&vma->anon_vma_chain)))
 673		return 1;
 674	return anon_vma1 == anon_vma2;
 675}
 676
 677/*
 678 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 679 * in front of (at a lower virtual address and file offset than) the vma.
 680 *
 681 * We cannot merge two vmas if they have differently assigned (non-NULL)
 682 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 683 *
 684 * We don't check here for the merged mmap wrapping around the end of pagecache
 685 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
 686 * wrap, nor mmaps which cover the final page at index -1UL.
 687 */
 688static int
 689can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
 690	struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
 691{
 692	if (is_mergeable_vma(vma, file, vm_flags) &&
 693	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 694		if (vma->vm_pgoff == vm_pgoff)
 695			return 1;
 696	}
 697	return 0;
 698}
 699
 700/*
 701 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 702 * beyond (at a higher virtual address and file offset than) the vma.
 703 *
 704 * We cannot merge two vmas if they have differently assigned (non-NULL)
 705 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 706 */
 707static int
 708can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
 709	struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
 710{
 711	if (is_mergeable_vma(vma, file, vm_flags) &&
 712	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 713		pgoff_t vm_pglen;
 714		vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
 715		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
 716			return 1;
 717	}
 718	return 0;
 719}
 720
 721/*
 722 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
 723 * whether that can be merged with its predecessor or its successor.
 724 * Or both (it neatly fills a hole).
 725 *
 726 * In most cases - when called for mmap, brk or mremap - [addr,end) is
 727 * certain not to be mapped by the time vma_merge is called; but when
 728 * called for mprotect, it is certain to be already mapped (either at
 729 * an offset within prev, or at the start of next), and the flags of
 730 * this area are about to be changed to vm_flags - and the no-change
 731 * case has already been eliminated.
 732 *
 733 * The following mprotect cases have to be considered, where AAAA is
 734 * the area passed down from mprotect_fixup, never extending beyond one
 735 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
 736 *
 737 *     AAAA             AAAA                AAAA          AAAA
 738 *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
 739 *    cannot merge    might become    might become    might become
 740 *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
 741 *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
 742 *    mremap move:                                    PPPPNNNNNNNN 8
 743 *        AAAA
 744 *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
 745 *    might become    case 1 below    case 2 below    case 3 below
 746 *
 747 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
 748 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
 749 */
 750struct vm_area_struct *vma_merge(struct mm_struct *mm,
 751			struct vm_area_struct *prev, unsigned long addr,
 752			unsigned long end, unsigned long vm_flags,
 753		     	struct anon_vma *anon_vma, struct file *file,
 754			pgoff_t pgoff, struct mempolicy *policy)
 755{
 756	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
 757	struct vm_area_struct *area, *next;
 758	int err;
 759
 760	/*
 761	 * We later require that vma->vm_flags == vm_flags,
 762	 * so this tests vma->vm_flags & VM_SPECIAL, too.
 763	 */
 764	if (vm_flags & VM_SPECIAL)
 765		return NULL;
 766
 767	if (prev)
 768		next = prev->vm_next;
 769	else
 770		next = mm->mmap;
 771	area = next;
 772	if (next && next->vm_end == end)		/* cases 6, 7, 8 */
 773		next = next->vm_next;
 774
 775	/*
 776	 * Can it merge with the predecessor?
 777	 */
 778	if (prev && prev->vm_end == addr &&
 779  			mpol_equal(vma_policy(prev), policy) &&
 780			can_vma_merge_after(prev, vm_flags,
 781						anon_vma, file, pgoff)) {
 782		/*
 783		 * OK, it can.  Can we now merge in the successor as well?
 784		 */
 785		if (next && end == next->vm_start &&
 786				mpol_equal(policy, vma_policy(next)) &&
 787				can_vma_merge_before(next, vm_flags,
 788					anon_vma, file, pgoff+pglen) &&
 789				is_mergeable_anon_vma(prev->anon_vma,
 790						      next->anon_vma, NULL)) {
 791							/* cases 1, 6 */
 792			err = vma_adjust(prev, prev->vm_start,
 793				next->vm_end, prev->vm_pgoff, NULL);
 794		} else					/* cases 2, 5, 7 */
 795			err = vma_adjust(prev, prev->vm_start,
 796				end, prev->vm_pgoff, NULL);
 797		if (err)
 798			return NULL;
 799		khugepaged_enter_vma_merge(prev);
 800		return prev;
 801	}
 802
 803	/*
 804	 * Can this new request be merged in front of next?
 805	 */
 806	if (next && end == next->vm_start &&
 807 			mpol_equal(policy, vma_policy(next)) &&
 808			can_vma_merge_before(next, vm_flags,
 809					anon_vma, file, pgoff+pglen)) {
 810		if (prev && addr < prev->vm_end)	/* case 4 */
 811			err = vma_adjust(prev, prev->vm_start,
 812				addr, prev->vm_pgoff, NULL);
 813		else					/* cases 3, 8 */
 814			err = vma_adjust(area, addr, next->vm_end,
 815				next->vm_pgoff - pglen, NULL);
 816		if (err)
 817			return NULL;
 818		khugepaged_enter_vma_merge(area);
 819		return area;
 820	}
 821
 822	return NULL;
 823}
 824
 825/*
 826 * Rough compatbility check to quickly see if it's even worth looking
 827 * at sharing an anon_vma.
 828 *
 829 * They need to have the same vm_file, and the flags can only differ
 830 * in things that mprotect may change.
 831 *
 832 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
 833 * we can merge the two vma's. For example, we refuse to merge a vma if
 834 * there is a vm_ops->close() function, because that indicates that the
 835 * driver is doing some kind of reference counting. But that doesn't
 836 * really matter for the anon_vma sharing case.
 837 */
 838static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
 839{
 840	return a->vm_end == b->vm_start &&
 841		mpol_equal(vma_policy(a), vma_policy(b)) &&
 842		a->vm_file == b->vm_file &&
 843		!((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
 844		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
 845}
 846
 847/*
 848 * Do some basic sanity checking to see if we can re-use the anon_vma
 849 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
 850 * the same as 'old', the other will be the new one that is trying
 851 * to share the anon_vma.
 852 *
 853 * NOTE! This runs with mm_sem held for reading, so it is possible that
 854 * the anon_vma of 'old' is concurrently in the process of being set up
 855 * by another page fault trying to merge _that_. But that's ok: if it
 856 * is being set up, that automatically means that it will be a singleton
 857 * acceptable for merging, so we can do all of this optimistically. But
 858 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
 859 *
 860 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
 861 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
 862 * is to return an anon_vma that is "complex" due to having gone through
 863 * a fork).
 864 *
 865 * We also make sure that the two vma's are compatible (adjacent,
 866 * and with the same memory policies). That's all stable, even with just
 867 * a read lock on the mm_sem.
 868 */
 869static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
 870{
 871	if (anon_vma_compatible(a, b)) {
 872		struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
 873
 874		if (anon_vma && list_is_singular(&old->anon_vma_chain))
 875			return anon_vma;
 876	}
 877	return NULL;
 878}
 879
 880/*
 881 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
 882 * neighbouring vmas for a suitable anon_vma, before it goes off
 883 * to allocate a new anon_vma.  It checks because a repetitive
 884 * sequence of mprotects and faults may otherwise lead to distinct
 885 * anon_vmas being allocated, preventing vma merge in subsequent
 886 * mprotect.
 887 */
 888struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
 889{
 890	struct anon_vma *anon_vma;
 891	struct vm_area_struct *near;
 892
 893	near = vma->vm_next;
 894	if (!near)
 895		goto try_prev;
 896
 897	anon_vma = reusable_anon_vma(near, vma, near);
 898	if (anon_vma)
 899		return anon_vma;
 900try_prev:
 901	near = vma->vm_prev;
 902	if (!near)
 903		goto none;
 904
 905	anon_vma = reusable_anon_vma(near, near, vma);
 906	if (anon_vma)
 907		return anon_vma;
 908none:
 909	/*
 910	 * There's no absolute need to look only at touching neighbours:
 911	 * we could search further afield for "compatible" anon_vmas.
 912	 * But it would probably just be a waste of time searching,
 913	 * or lead to too many vmas hanging off the same anon_vma.
 914	 * We're trying to allow mprotect remerging later on,
 915	 * not trying to minimize memory used for anon_vmas.
 916	 */
 917	return NULL;
 918}
 919
 920#ifdef CONFIG_PROC_FS
 921void vm_stat_account(struct mm_struct *mm, unsigned long flags,
 922						struct file *file, long pages)
 923{
 924	const unsigned long stack_flags
 925		= VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
 926
 
 
 927	if (file) {
 928		mm->shared_vm += pages;
 929		if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
 930			mm->exec_vm += pages;
 931	} else if (flags & stack_flags)
 932		mm->stack_vm += pages;
 933	if (flags & (VM_RESERVED|VM_IO))
 934		mm->reserved_vm += pages;
 935}
 936#endif /* CONFIG_PROC_FS */
 937
 938/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 939 * The caller must hold down_write(&current->mm->mmap_sem).
 940 */
 941
 942unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
 943			unsigned long len, unsigned long prot,
 944			unsigned long flags, unsigned long pgoff)
 
 945{
 946	struct mm_struct * mm = current->mm;
 947	struct inode *inode;
 948	vm_flags_t vm_flags;
 949	int error;
 950	unsigned long reqprot = prot;
 951
 952	/*
 953	 * Does the application expect PROT_READ to imply PROT_EXEC?
 954	 *
 955	 * (the exception is when the underlying filesystem is noexec
 956	 *  mounted, in which case we dont add PROT_EXEC.)
 957	 */
 958	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 959		if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
 960			prot |= PROT_EXEC;
 961
 962	if (!len)
 963		return -EINVAL;
 964
 965	if (!(flags & MAP_FIXED))
 966		addr = round_hint_to_min(addr);
 967
 968	/* Careful about overflows.. */
 969	len = PAGE_ALIGN(len);
 970	if (!len)
 971		return -ENOMEM;
 972
 973	/* offset overflow? */
 974	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
 975               return -EOVERFLOW;
 976
 977	/* Too many mappings? */
 978	if (mm->map_count > sysctl_max_map_count)
 979		return -ENOMEM;
 980
 981	/* Obtain the address to map to. we verify (or select) it and ensure
 982	 * that it represents a valid section of the address space.
 983	 */
 984	addr = get_unmapped_area(file, addr, len, pgoff, flags);
 985	if (addr & ~PAGE_MASK)
 986		return addr;
 987
 988	/* Do simple checking here so the lower-level routines won't have
 989	 * to. we assume access permissions have been handled by the open
 990	 * of the memory object, so we don't do any here.
 991	 */
 992	vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
 993			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
 994
 995	if (flags & MAP_LOCKED)
 996		if (!can_do_mlock())
 997			return -EPERM;
 998
 999	/* mlock MCL_FUTURE? */
1000	if (vm_flags & VM_LOCKED) {
1001		unsigned long locked, lock_limit;
1002		locked = len >> PAGE_SHIFT;
1003		locked += mm->locked_vm;
1004		lock_limit = rlimit(RLIMIT_MEMLOCK);
1005		lock_limit >>= PAGE_SHIFT;
1006		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1007			return -EAGAIN;
1008	}
1009
1010	inode = file ? file->f_path.dentry->d_inode : NULL;
1011
1012	if (file) {
 
 
1013		switch (flags & MAP_TYPE) {
1014		case MAP_SHARED:
1015			if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1016				return -EACCES;
1017
1018			/*
1019			 * Make sure we don't allow writing to an append-only
1020			 * file..
1021			 */
1022			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1023				return -EACCES;
1024
1025			/*
1026			 * Make sure there are no mandatory locks on the file.
1027			 */
1028			if (locks_verify_locked(inode))
1029				return -EAGAIN;
1030
1031			vm_flags |= VM_SHARED | VM_MAYSHARE;
1032			if (!(file->f_mode & FMODE_WRITE))
1033				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1034
1035			/* fall through */
1036		case MAP_PRIVATE:
1037			if (!(file->f_mode & FMODE_READ))
1038				return -EACCES;
1039			if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1040				if (vm_flags & VM_EXEC)
1041					return -EPERM;
1042				vm_flags &= ~VM_MAYEXEC;
1043			}
1044
1045			if (!file->f_op || !file->f_op->mmap)
1046				return -ENODEV;
 
 
1047			break;
1048
1049		default:
1050			return -EINVAL;
1051		}
1052	} else {
1053		switch (flags & MAP_TYPE) {
1054		case MAP_SHARED:
 
 
1055			/*
1056			 * Ignore pgoff.
1057			 */
1058			pgoff = 0;
1059			vm_flags |= VM_SHARED | VM_MAYSHARE;
1060			break;
1061		case MAP_PRIVATE:
1062			/*
1063			 * Set pgoff according to addr for anon_vma.
1064			 */
1065			pgoff = addr >> PAGE_SHIFT;
1066			break;
1067		default:
1068			return -EINVAL;
1069		}
1070	}
1071
1072	error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1073	if (error)
1074		return error;
 
 
 
 
 
1075
1076	return mmap_region(file, addr, len, flags, vm_flags, pgoff);
 
 
 
 
 
 
 
 
 
 
1077}
1078EXPORT_SYMBOL(do_mmap_pgoff);
1079
1080SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1081		unsigned long, prot, unsigned long, flags,
1082		unsigned long, fd, unsigned long, pgoff)
1083{
1084	struct file *file = NULL;
1085	unsigned long retval = -EBADF;
1086
1087	if (!(flags & MAP_ANONYMOUS)) {
1088		audit_mmap_fd(fd, flags);
1089		if (unlikely(flags & MAP_HUGETLB))
1090			return -EINVAL;
1091		file = fget(fd);
1092		if (!file)
1093			goto out;
 
 
 
 
 
1094	} else if (flags & MAP_HUGETLB) {
1095		struct user_struct *user = NULL;
 
 
 
 
 
 
 
1096		/*
1097		 * VM_NORESERVE is used because the reservations will be
1098		 * taken when vm_ops->mmap() is called
1099		 * A dummy user value is used because we are not locking
1100		 * memory so no accounting is necessary
1101		 */
1102		len = ALIGN(len, huge_page_size(&default_hstate));
1103		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1104						&user, HUGETLB_ANONHUGE_INODE);
 
1105		if (IS_ERR(file))
1106			return PTR_ERR(file);
1107	}
1108
1109	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1110
1111	down_write(&current->mm->mmap_sem);
1112	retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1113	up_write(&current->mm->mmap_sem);
1114
1115	if (file)
1116		fput(file);
1117out:
1118	return retval;
1119}
1120
1121#ifdef __ARCH_WANT_SYS_OLD_MMAP
1122struct mmap_arg_struct {
1123	unsigned long addr;
1124	unsigned long len;
1125	unsigned long prot;
1126	unsigned long flags;
1127	unsigned long fd;
1128	unsigned long offset;
1129};
1130
1131SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1132{
1133	struct mmap_arg_struct a;
1134
1135	if (copy_from_user(&a, arg, sizeof(a)))
1136		return -EFAULT;
1137	if (a.offset & ~PAGE_MASK)
1138		return -EINVAL;
1139
1140	return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1141			      a.offset >> PAGE_SHIFT);
1142}
1143#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1144
1145/*
1146 * Some shared mappigns will want the pages marked read-only
1147 * to track write events. If so, we'll downgrade vm_page_prot
1148 * to the private version (using protection_map[] without the
1149 * VM_SHARED bit).
1150 */
1151int vma_wants_writenotify(struct vm_area_struct *vma)
1152{
1153	vm_flags_t vm_flags = vma->vm_flags;
1154
1155	/* If it was private or non-writable, the write bit is already clear */
1156	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1157		return 0;
1158
1159	/* The backer wishes to know when pages are first written to? */
1160	if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1161		return 1;
1162
1163	/* The open routine did something to the protections already? */
1164	if (pgprot_val(vma->vm_page_prot) !=
1165	    pgprot_val(vm_get_page_prot(vm_flags)))
1166		return 0;
1167
1168	/* Specialty mapping? */
1169	if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1170		return 0;
1171
1172	/* Can the mapping track the dirty pages? */
1173	return vma->vm_file && vma->vm_file->f_mapping &&
1174		mapping_cap_account_dirty(vma->vm_file->f_mapping);
1175}
1176
1177/*
1178 * We account for memory if it's a private writeable mapping,
1179 * not hugepages and VM_NORESERVE wasn't set.
1180 */
1181static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1182{
1183	/*
1184	 * hugetlb has its own accounting separate from the core VM
1185	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1186	 */
1187	if (file && is_file_hugepages(file))
1188		return 0;
1189
1190	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1191}
1192
1193unsigned long mmap_region(struct file *file, unsigned long addr,
1194			  unsigned long len, unsigned long flags,
1195			  vm_flags_t vm_flags, unsigned long pgoff)
1196{
1197	struct mm_struct *mm = current->mm;
1198	struct vm_area_struct *vma, *prev;
1199	int correct_wcount = 0;
1200	int error;
1201	struct rb_node **rb_link, *rb_parent;
1202	unsigned long charged = 0;
1203	struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1204
1205	/* Clear old maps */
1206	error = -ENOMEM;
1207munmap_back:
1208	vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1209	if (vma && vma->vm_start < addr + len) {
1210		if (do_munmap(mm, addr, len))
1211			return -ENOMEM;
1212		goto munmap_back;
1213	}
1214
1215	/* Check against address space limit. */
1216	if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1217		return -ENOMEM;
1218
1219	/*
1220	 * Set 'VM_NORESERVE' if we should not account for the
1221	 * memory use of this mapping.
1222	 */
1223	if ((flags & MAP_NORESERVE)) {
1224		/* We honor MAP_NORESERVE if allowed to overcommit */
1225		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1226			vm_flags |= VM_NORESERVE;
1227
1228		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1229		if (file && is_file_hugepages(file))
1230			vm_flags |= VM_NORESERVE;
1231	}
1232
1233	/*
1234	 * Private writable mapping: check memory availability
1235	 */
1236	if (accountable_mapping(file, vm_flags)) {
1237		charged = len >> PAGE_SHIFT;
1238		if (security_vm_enough_memory(charged))
1239			return -ENOMEM;
1240		vm_flags |= VM_ACCOUNT;
1241	}
1242
1243	/*
1244	 * Can we just expand an old mapping?
1245	 */
1246	vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1247	if (vma)
1248		goto out;
1249
1250	/*
1251	 * Determine the object being mapped and call the appropriate
1252	 * specific mapper. the address has already been validated, but
1253	 * not unmapped, but the maps are removed from the list.
1254	 */
1255	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1256	if (!vma) {
1257		error = -ENOMEM;
1258		goto unacct_error;
1259	}
1260
1261	vma->vm_mm = mm;
1262	vma->vm_start = addr;
1263	vma->vm_end = addr + len;
1264	vma->vm_flags = vm_flags;
1265	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1266	vma->vm_pgoff = pgoff;
1267	INIT_LIST_HEAD(&vma->anon_vma_chain);
1268
1269	if (file) {
1270		error = -EINVAL;
1271		if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1272			goto free_vma;
1273		if (vm_flags & VM_DENYWRITE) {
1274			error = deny_write_access(file);
1275			if (error)
1276				goto free_vma;
1277			correct_wcount = 1;
1278		}
1279		vma->vm_file = file;
1280		get_file(file);
1281		error = file->f_op->mmap(file, vma);
1282		if (error)
1283			goto unmap_and_free_vma;
1284		if (vm_flags & VM_EXECUTABLE)
1285			added_exe_file_vma(mm);
1286
1287		/* Can addr have changed??
1288		 *
1289		 * Answer: Yes, several device drivers can do it in their
1290		 *         f_op->mmap method. -DaveM
 
 
1291		 */
 
 
1292		addr = vma->vm_start;
1293		pgoff = vma->vm_pgoff;
1294		vm_flags = vma->vm_flags;
1295	} else if (vm_flags & VM_SHARED) {
1296		error = shmem_zero_setup(vma);
1297		if (error)
1298			goto free_vma;
1299	}
1300
1301	if (vma_wants_writenotify(vma)) {
1302		pgprot_t pprot = vma->vm_page_prot;
1303
1304		/* Can vma->vm_page_prot have changed??
1305		 *
1306		 * Answer: Yes, drivers may have changed it in their
1307		 *         f_op->mmap method.
1308		 *
1309		 * Ensures that vmas marked as uncached stay that way.
1310		 */
1311		vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1312		if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1313			vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1314	}
1315
1316	vma_link(mm, vma, prev, rb_link, rb_parent);
1317	file = vma->vm_file;
1318
1319	/* Once vma denies write, undo our temporary denial count */
1320	if (correct_wcount)
1321		atomic_inc(&inode->i_writecount);
 
1322out:
1323	perf_event_mmap(vma);
1324
1325	mm->total_vm += len >> PAGE_SHIFT;
1326	vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1327	if (vm_flags & VM_LOCKED) {
1328		if (!mlock_vma_pages_range(vma, addr, addr + len))
 
1329			mm->locked_vm += (len >> PAGE_SHIFT);
1330	} else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1331		make_pages_present(addr, addr + len);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1332	return addr;
1333
1334unmap_and_free_vma:
1335	if (correct_wcount)
1336		atomic_inc(&inode->i_writecount);
1337	vma->vm_file = NULL;
1338	fput(file);
1339
1340	/* Undo any partial mapping done by a device driver. */
1341	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1342	charged = 0;
1343free_vma:
1344	kmem_cache_free(vm_area_cachep, vma);
1345unacct_error:
1346	if (charged)
1347		vm_unacct_memory(charged);
1348	return error;
1349}
1350
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1351/* Get an address range which is currently unmapped.
1352 * For shmat() with addr=0.
1353 *
1354 * Ugly calling convention alert:
1355 * Return value with the low bits set means error value,
1356 * ie
1357 *	if (ret & ~PAGE_MASK)
1358 *		error = ret;
1359 *
1360 * This function "knows" that -ENOMEM has the bits set.
1361 */
1362#ifndef HAVE_ARCH_UNMAPPED_AREA
1363unsigned long
1364arch_get_unmapped_area(struct file *filp, unsigned long addr,
1365		unsigned long len, unsigned long pgoff, unsigned long flags)
1366{
1367	struct mm_struct *mm = current->mm;
1368	struct vm_area_struct *vma;
1369	unsigned long start_addr;
1370
1371	if (len > TASK_SIZE)
1372		return -ENOMEM;
1373
1374	if (flags & MAP_FIXED)
1375		return addr;
1376
1377	if (addr) {
1378		addr = PAGE_ALIGN(addr);
1379		vma = find_vma(mm, addr);
1380		if (TASK_SIZE - len >= addr &&
1381		    (!vma || addr + len <= vma->vm_start))
1382			return addr;
1383	}
1384	if (len > mm->cached_hole_size) {
1385	        start_addr = addr = mm->free_area_cache;
1386	} else {
1387	        start_addr = addr = TASK_UNMAPPED_BASE;
1388	        mm->cached_hole_size = 0;
1389	}
1390
1391full_search:
1392	for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1393		/* At this point:  (!vma || addr < vma->vm_end). */
1394		if (TASK_SIZE - len < addr) {
1395			/*
1396			 * Start a new search - just in case we missed
1397			 * some holes.
1398			 */
1399			if (start_addr != TASK_UNMAPPED_BASE) {
1400				addr = TASK_UNMAPPED_BASE;
1401			        start_addr = addr;
1402				mm->cached_hole_size = 0;
1403				goto full_search;
1404			}
1405			return -ENOMEM;
1406		}
1407		if (!vma || addr + len <= vma->vm_start) {
1408			/*
1409			 * Remember the place where we stopped the search:
1410			 */
1411			mm->free_area_cache = addr + len;
1412			return addr;
1413		}
1414		if (addr + mm->cached_hole_size < vma->vm_start)
1415		        mm->cached_hole_size = vma->vm_start - addr;
1416		addr = vma->vm_end;
1417	}
1418}
1419#endif	
1420
1421void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1422{
1423	/*
1424	 * Is this a new hole at the lowest possible address?
1425	 */
1426	if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1427		mm->free_area_cache = addr;
1428		mm->cached_hole_size = ~0UL;
1429	}
1430}
1431
1432/*
1433 * This mmap-allocator allocates new areas top-down from below the
1434 * stack's low limit (the base):
1435 */
1436#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1437unsigned long
1438arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1439			  const unsigned long len, const unsigned long pgoff,
1440			  const unsigned long flags)
1441{
1442	struct vm_area_struct *vma;
1443	struct mm_struct *mm = current->mm;
1444	unsigned long addr = addr0;
 
1445
1446	/* requested length too big for entire address space */
1447	if (len > TASK_SIZE)
1448		return -ENOMEM;
1449
1450	if (flags & MAP_FIXED)
1451		return addr;
1452
1453	/* requesting a specific address */
1454	if (addr) {
1455		addr = PAGE_ALIGN(addr);
1456		vma = find_vma(mm, addr);
1457		if (TASK_SIZE - len >= addr &&
1458				(!vma || addr + len <= vma->vm_start))
1459			return addr;
1460	}
1461
1462	/* check if free_area_cache is useful for us */
1463	if (len <= mm->cached_hole_size) {
1464 	        mm->cached_hole_size = 0;
1465 		mm->free_area_cache = mm->mmap_base;
1466 	}
1467
1468	/* either no address requested or can't fit in requested address hole */
1469	addr = mm->free_area_cache;
1470
1471	/* make sure it can fit in the remaining address space */
1472	if (addr > len) {
1473		vma = find_vma(mm, addr-len);
1474		if (!vma || addr <= vma->vm_start)
1475			/* remember the address as a hint for next time */
1476			return (mm->free_area_cache = addr-len);
1477	}
1478
1479	if (mm->mmap_base < len)
1480		goto bottomup;
1481
1482	addr = mm->mmap_base-len;
1483
1484	do {
1485		/*
1486		 * Lookup failure means no vma is above this address,
1487		 * else if new region fits below vma->vm_start,
1488		 * return with success:
1489		 */
1490		vma = find_vma(mm, addr);
1491		if (!vma || addr+len <= vma->vm_start)
1492			/* remember the address as a hint for next time */
1493			return (mm->free_area_cache = addr);
1494
1495 		/* remember the largest hole we saw so far */
1496 		if (addr + mm->cached_hole_size < vma->vm_start)
1497 		        mm->cached_hole_size = vma->vm_start - addr;
1498
1499		/* try just below the current vma->vm_start */
1500		addr = vma->vm_start-len;
1501	} while (len < vma->vm_start);
1502
1503bottomup:
1504	/*
1505	 * A failed mmap() very likely causes application failure,
1506	 * so fall back to the bottom-up function here. This scenario
1507	 * can happen with large stack limits and large mmap()
1508	 * allocations.
1509	 */
1510	mm->cached_hole_size = ~0UL;
1511  	mm->free_area_cache = TASK_UNMAPPED_BASE;
1512	addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1513	/*
1514	 * Restore the topdown base:
1515	 */
1516	mm->free_area_cache = mm->mmap_base;
1517	mm->cached_hole_size = ~0UL;
1518
1519	return addr;
1520}
1521#endif
1522
1523void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1524{
1525	/*
1526	 * Is this a new hole at the highest possible address?
1527	 */
1528	if (addr > mm->free_area_cache)
1529		mm->free_area_cache = addr;
1530
1531	/* dont allow allocations above current base */
1532	if (mm->free_area_cache > mm->mmap_base)
1533		mm->free_area_cache = mm->mmap_base;
1534}
1535
1536unsigned long
1537get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1538		unsigned long pgoff, unsigned long flags)
1539{
1540	unsigned long (*get_area)(struct file *, unsigned long,
1541				  unsigned long, unsigned long, unsigned long);
1542
1543	unsigned long error = arch_mmap_check(addr, len, flags);
1544	if (error)
1545		return error;
1546
1547	/* Careful about overflows.. */
1548	if (len > TASK_SIZE)
1549		return -ENOMEM;
1550
1551	get_area = current->mm->get_unmapped_area;
1552	if (file && file->f_op && file->f_op->get_unmapped_area)
1553		get_area = file->f_op->get_unmapped_area;
1554	addr = get_area(file, addr, len, pgoff, flags);
1555	if (IS_ERR_VALUE(addr))
1556		return addr;
1557
1558	if (addr > TASK_SIZE - len)
1559		return -ENOMEM;
1560	if (addr & ~PAGE_MASK)
1561		return -EINVAL;
1562
1563	return arch_rebalance_pgtables(addr, len);
 
 
1564}
1565
1566EXPORT_SYMBOL(get_unmapped_area);
1567
1568/* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1569struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1570{
1571	struct vm_area_struct *vma = NULL;
 
1572
1573	if (mm) {
1574		/* Check the cache first. */
1575		/* (Cache hit rate is typically around 35%.) */
1576		vma = mm->mmap_cache;
1577		if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1578			struct rb_node * rb_node;
1579
1580			rb_node = mm->mm_rb.rb_node;
1581			vma = NULL;
1582
1583			while (rb_node) {
1584				struct vm_area_struct * vma_tmp;
1585
1586				vma_tmp = rb_entry(rb_node,
1587						struct vm_area_struct, vm_rb);
1588
1589				if (vma_tmp->vm_end > addr) {
1590					vma = vma_tmp;
1591					if (vma_tmp->vm_start <= addr)
1592						break;
1593					rb_node = rb_node->rb_left;
1594				} else
1595					rb_node = rb_node->rb_right;
1596			}
1597			if (vma)
1598				mm->mmap_cache = vma;
1599		}
1600	}
 
 
 
1601	return vma;
1602}
1603
1604EXPORT_SYMBOL(find_vma);
1605
1606/* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
 
 
1607struct vm_area_struct *
1608find_vma_prev(struct mm_struct *mm, unsigned long addr,
1609			struct vm_area_struct **pprev)
1610{
1611	struct vm_area_struct *vma = NULL, *prev = NULL;
1612	struct rb_node *rb_node;
1613	if (!mm)
1614		goto out;
1615
1616	/* Guard against addr being lower than the first VMA */
1617	vma = mm->mmap;
1618
1619	/* Go through the RB tree quickly. */
1620	rb_node = mm->mm_rb.rb_node;
1621
1622	while (rb_node) {
1623		struct vm_area_struct *vma_tmp;
1624		vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1625
1626		if (addr < vma_tmp->vm_end) {
1627			rb_node = rb_node->rb_left;
1628		} else {
1629			prev = vma_tmp;
1630			if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1631				break;
 
 
1632			rb_node = rb_node->rb_right;
1633		}
1634	}
1635
1636out:
1637	*pprev = prev;
1638	return prev ? prev->vm_next : vma;
1639}
1640
1641/*
1642 * Verify that the stack growth is acceptable and
1643 * update accounting. This is shared with both the
1644 * grow-up and grow-down cases.
1645 */
1646static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1647{
1648	struct mm_struct *mm = vma->vm_mm;
1649	struct rlimit *rlim = current->signal->rlim;
1650	unsigned long new_start;
1651
1652	/* address space limit tests */
1653	if (!may_expand_vm(mm, grow))
1654		return -ENOMEM;
1655
1656	/* Stack limit test */
1657	if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1658		return -ENOMEM;
1659
1660	/* mlock limit tests */
1661	if (vma->vm_flags & VM_LOCKED) {
1662		unsigned long locked;
1663		unsigned long limit;
1664		locked = mm->locked_vm + grow;
1665		limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1666		limit >>= PAGE_SHIFT;
1667		if (locked > limit && !capable(CAP_IPC_LOCK))
1668			return -ENOMEM;
1669	}
1670
1671	/* Check to ensure the stack will not grow into a hugetlb-only region */
1672	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1673			vma->vm_end - size;
1674	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1675		return -EFAULT;
1676
1677	/*
1678	 * Overcommit..  This must be the final test, as it will
1679	 * update security statistics.
1680	 */
1681	if (security_vm_enough_memory_mm(mm, grow))
1682		return -ENOMEM;
1683
1684	/* Ok, everything looks good - let it rip */
1685	mm->total_vm += grow;
1686	if (vma->vm_flags & VM_LOCKED)
1687		mm->locked_vm += grow;
1688	vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1689	return 0;
1690}
1691
1692#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1693/*
1694 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1695 * vma is the last one with address > vma->vm_end.  Have to extend vma.
1696 */
1697int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1698{
1699	int error;
1700
1701	if (!(vma->vm_flags & VM_GROWSUP))
1702		return -EFAULT;
1703
1704	/*
1705	 * We must make sure the anon_vma is allocated
1706	 * so that the anon_vma locking is not a noop.
1707	 */
1708	if (unlikely(anon_vma_prepare(vma)))
1709		return -ENOMEM;
1710	vma_lock_anon_vma(vma);
1711
1712	/*
1713	 * vma->vm_start/vm_end cannot change under us because the caller
1714	 * is required to hold the mmap_sem in read mode.  We need the
1715	 * anon_vma lock to serialize against concurrent expand_stacks.
1716	 * Also guard against wrapping around to address 0.
1717	 */
1718	if (address < PAGE_ALIGN(address+4))
1719		address = PAGE_ALIGN(address+4);
1720	else {
1721		vma_unlock_anon_vma(vma);
1722		return -ENOMEM;
1723	}
1724	error = 0;
1725
1726	/* Somebody else might have raced and expanded it already */
1727	if (address > vma->vm_end) {
1728		unsigned long size, grow;
1729
1730		size = address - vma->vm_start;
1731		grow = (address - vma->vm_end) >> PAGE_SHIFT;
1732
1733		error = -ENOMEM;
1734		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1735			error = acct_stack_growth(vma, size, grow);
1736			if (!error) {
 
 
 
 
 
 
 
 
 
 
 
 
 
1737				vma->vm_end = address;
 
 
 
 
 
 
 
1738				perf_event_mmap(vma);
1739			}
1740		}
1741	}
1742	vma_unlock_anon_vma(vma);
1743	khugepaged_enter_vma_merge(vma);
 
1744	return error;
1745}
1746#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1747
1748/*
1749 * vma is the first one with address < vma->vm_start.  Have to extend vma.
1750 */
1751int expand_downwards(struct vm_area_struct *vma,
1752				   unsigned long address)
1753{
1754	int error;
1755
1756	/*
1757	 * We must make sure the anon_vma is allocated
1758	 * so that the anon_vma locking is not a noop.
1759	 */
1760	if (unlikely(anon_vma_prepare(vma)))
1761		return -ENOMEM;
1762
1763	address &= PAGE_MASK;
1764	error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1765	if (error)
1766		return error;
1767
1768	vma_lock_anon_vma(vma);
1769
1770	/*
1771	 * vma->vm_start/vm_end cannot change under us because the caller
1772	 * is required to hold the mmap_sem in read mode.  We need the
1773	 * anon_vma lock to serialize against concurrent expand_stacks.
1774	 */
1775
1776	/* Somebody else might have raced and expanded it already */
1777	if (address < vma->vm_start) {
1778		unsigned long size, grow;
1779
1780		size = vma->vm_end - address;
1781		grow = (vma->vm_start - address) >> PAGE_SHIFT;
1782
1783		error = -ENOMEM;
1784		if (grow <= vma->vm_pgoff) {
1785			error = acct_stack_growth(vma, size, grow);
1786			if (!error) {
 
 
 
 
 
 
 
 
 
 
 
 
 
1787				vma->vm_start = address;
1788				vma->vm_pgoff -= grow;
 
 
 
 
1789				perf_event_mmap(vma);
1790			}
1791		}
1792	}
1793	vma_unlock_anon_vma(vma);
1794	khugepaged_enter_vma_merge(vma);
 
1795	return error;
1796}
1797
 
 
 
 
 
 
 
 
 
 
 
1798#ifdef CONFIG_STACK_GROWSUP
1799int expand_stack(struct vm_area_struct *vma, unsigned long address)
1800{
 
 
 
 
 
 
 
 
1801	return expand_upwards(vma, address);
1802}
1803
1804struct vm_area_struct *
1805find_extend_vma(struct mm_struct *mm, unsigned long addr)
1806{
1807	struct vm_area_struct *vma, *prev;
1808
1809	addr &= PAGE_MASK;
1810	vma = find_vma_prev(mm, addr, &prev);
1811	if (vma && (vma->vm_start <= addr))
1812		return vma;
1813	if (!prev || expand_stack(prev, addr))
1814		return NULL;
1815	if (prev->vm_flags & VM_LOCKED) {
1816		mlock_vma_pages_range(prev, addr, prev->vm_end);
1817	}
1818	return prev;
1819}
1820#else
1821int expand_stack(struct vm_area_struct *vma, unsigned long address)
1822{
 
 
 
 
 
 
 
 
1823	return expand_downwards(vma, address);
1824}
1825
1826struct vm_area_struct *
1827find_extend_vma(struct mm_struct * mm, unsigned long addr)
1828{
1829	struct vm_area_struct * vma;
1830	unsigned long start;
1831
1832	addr &= PAGE_MASK;
1833	vma = find_vma(mm,addr);
1834	if (!vma)
1835		return NULL;
1836	if (vma->vm_start <= addr)
1837		return vma;
1838	if (!(vma->vm_flags & VM_GROWSDOWN))
1839		return NULL;
1840	start = vma->vm_start;
1841	if (expand_stack(vma, addr))
1842		return NULL;
1843	if (vma->vm_flags & VM_LOCKED) {
1844		mlock_vma_pages_range(vma, addr, start);
1845	}
1846	return vma;
1847}
1848#endif
1849
1850/*
1851 * Ok - we have the memory areas we should free on the vma list,
1852 * so release them, and do the vma updates.
1853 *
1854 * Called with the mm semaphore held.
1855 */
1856static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1857{
 
 
1858	/* Update high watermark before we lower total_vm */
1859	update_hiwater_vm(mm);
1860	do {
1861		long nrpages = vma_pages(vma);
1862
1863		mm->total_vm -= nrpages;
 
1864		vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1865		vma = remove_vma(vma);
1866	} while (vma);
 
1867	validate_mm(mm);
1868}
1869
1870/*
1871 * Get rid of page table information in the indicated region.
1872 *
1873 * Called with the mm semaphore held.
1874 */
1875static void unmap_region(struct mm_struct *mm,
1876		struct vm_area_struct *vma, struct vm_area_struct *prev,
1877		unsigned long start, unsigned long end)
1878{
1879	struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1880	struct mmu_gather tlb;
1881	unsigned long nr_accounted = 0;
1882
1883	lru_add_drain();
1884	tlb_gather_mmu(&tlb, mm, 0);
1885	update_hiwater_rss(mm);
1886	unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1887	vm_unacct_memory(nr_accounted);
1888	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1889				 next ? next->vm_start : 0);
1890	tlb_finish_mmu(&tlb, start, end);
1891}
1892
1893/*
1894 * Create a list of vma's touched by the unmap, removing them from the mm's
1895 * vma list as we go..
1896 */
1897static void
1898detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1899	struct vm_area_struct *prev, unsigned long end)
1900{
1901	struct vm_area_struct **insertion_point;
1902	struct vm_area_struct *tail_vma = NULL;
1903	unsigned long addr;
1904
1905	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1906	vma->vm_prev = NULL;
1907	do {
1908		rb_erase(&vma->vm_rb, &mm->mm_rb);
1909		mm->map_count--;
1910		tail_vma = vma;
1911		vma = vma->vm_next;
1912	} while (vma && vma->vm_start < end);
1913	*insertion_point = vma;
1914	if (vma)
1915		vma->vm_prev = prev;
 
 
 
1916	tail_vma->vm_next = NULL;
1917	if (mm->unmap_area == arch_unmap_area)
1918		addr = prev ? prev->vm_end : mm->mmap_base;
1919	else
1920		addr = vma ?  vma->vm_start : mm->mmap_base;
1921	mm->unmap_area(mm, addr);
1922	mm->mmap_cache = NULL;		/* Kill the cache. */
1923}
1924
1925/*
1926 * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1927 * munmap path where it doesn't make sense to fail.
1928 */
1929static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1930	      unsigned long addr, int new_below)
1931{
1932	struct mempolicy *pol;
1933	struct vm_area_struct *new;
1934	int err = -ENOMEM;
1935
1936	if (is_vm_hugetlb_page(vma) && (addr &
1937					~(huge_page_mask(hstate_vma(vma)))))
1938		return -EINVAL;
1939
1940	new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1941	if (!new)
1942		goto out_err;
1943
1944	/* most fields are the same, copy all, and then fixup */
1945	*new = *vma;
1946
1947	INIT_LIST_HEAD(&new->anon_vma_chain);
1948
1949	if (new_below)
1950		new->vm_end = addr;
1951	else {
1952		new->vm_start = addr;
1953		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1954	}
1955
1956	pol = mpol_dup(vma_policy(vma));
1957	if (IS_ERR(pol)) {
1958		err = PTR_ERR(pol);
1959		goto out_free_vma;
1960	}
1961	vma_set_policy(new, pol);
1962
1963	if (anon_vma_clone(new, vma))
1964		goto out_free_mpol;
1965
1966	if (new->vm_file) {
1967		get_file(new->vm_file);
1968		if (vma->vm_flags & VM_EXECUTABLE)
1969			added_exe_file_vma(mm);
1970	}
1971
1972	if (new->vm_ops && new->vm_ops->open)
1973		new->vm_ops->open(new);
1974
1975	if (new_below)
1976		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1977			((addr - new->vm_start) >> PAGE_SHIFT), new);
1978	else
1979		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1980
1981	/* Success. */
1982	if (!err)
1983		return 0;
1984
1985	/* Clean everything up if vma_adjust failed. */
1986	if (new->vm_ops && new->vm_ops->close)
1987		new->vm_ops->close(new);
1988	if (new->vm_file) {
1989		if (vma->vm_flags & VM_EXECUTABLE)
1990			removed_exe_file_vma(mm);
1991		fput(new->vm_file);
1992	}
1993	unlink_anon_vmas(new);
1994 out_free_mpol:
1995	mpol_put(pol);
1996 out_free_vma:
1997	kmem_cache_free(vm_area_cachep, new);
1998 out_err:
1999	return err;
2000}
2001
2002/*
2003 * Split a vma into two pieces at address 'addr', a new vma is allocated
2004 * either for the first part or the tail.
2005 */
2006int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2007	      unsigned long addr, int new_below)
2008{
2009	if (mm->map_count >= sysctl_max_map_count)
2010		return -ENOMEM;
2011
2012	return __split_vma(mm, vma, addr, new_below);
2013}
2014
2015/* Munmap is split into 2 main parts -- this part which finds
2016 * what needs doing, and the areas themselves, which do the
2017 * work.  This now handles partial unmappings.
2018 * Jeremy Fitzhardinge <jeremy@goop.org>
2019 */
2020int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2021{
2022	unsigned long end;
2023	struct vm_area_struct *vma, *prev, *last;
2024
2025	if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2026		return -EINVAL;
2027
2028	if ((len = PAGE_ALIGN(len)) == 0)
2029		return -EINVAL;
2030
2031	/* Find the first overlapping VMA */
2032	vma = find_vma(mm, start);
2033	if (!vma)
2034		return 0;
2035	prev = vma->vm_prev;
2036	/* we have  start < vma->vm_end  */
2037
2038	/* if it doesn't overlap, we have nothing.. */
2039	end = start + len;
2040	if (vma->vm_start >= end)
2041		return 0;
2042
2043	/*
2044	 * If we need to split any vma, do it now to save pain later.
2045	 *
2046	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2047	 * unmapped vm_area_struct will remain in use: so lower split_vma
2048	 * places tmp vma above, and higher split_vma places tmp vma below.
2049	 */
2050	if (start > vma->vm_start) {
2051		int error;
2052
2053		/*
2054		 * Make sure that map_count on return from munmap() will
2055		 * not exceed its limit; but let map_count go just above
2056		 * its limit temporarily, to help free resources as expected.
2057		 */
2058		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2059			return -ENOMEM;
2060
2061		error = __split_vma(mm, vma, start, 0);
2062		if (error)
2063			return error;
2064		prev = vma;
2065	}
2066
2067	/* Does it split the last one? */
2068	last = find_vma(mm, end);
2069	if (last && end > last->vm_start) {
2070		int error = __split_vma(mm, last, end, 1);
2071		if (error)
2072			return error;
2073	}
2074	vma = prev? prev->vm_next: mm->mmap;
2075
2076	/*
2077	 * unlock any mlock()ed ranges before detaching vmas
2078	 */
2079	if (mm->locked_vm) {
2080		struct vm_area_struct *tmp = vma;
2081		while (tmp && tmp->vm_start < end) {
2082			if (tmp->vm_flags & VM_LOCKED) {
2083				mm->locked_vm -= vma_pages(tmp);
2084				munlock_vma_pages_all(tmp);
2085			}
2086			tmp = tmp->vm_next;
2087		}
2088	}
2089
2090	/*
2091	 * Remove the vma's, and unmap the actual pages
2092	 */
2093	detach_vmas_to_be_unmapped(mm, vma, prev, end);
2094	unmap_region(mm, vma, prev, start, end);
2095
2096	/* Fix up all other VM information */
2097	remove_vma_list(mm, vma);
2098
2099	return 0;
2100}
2101
2102EXPORT_SYMBOL(do_munmap);
2103
2104SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2105{
2106	int ret;
2107	struct mm_struct *mm = current->mm;
2108
2109	profile_munmap(addr);
2110
2111	down_write(&mm->mmap_sem);
2112	ret = do_munmap(mm, addr, len);
2113	up_write(&mm->mmap_sem);
2114	return ret;
2115}
 
 
 
 
 
 
 
2116
2117static inline void verify_mm_writelocked(struct mm_struct *mm)
2118{
2119#ifdef CONFIG_DEBUG_VM
2120	if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2121		WARN_ON(1);
2122		up_read(&mm->mmap_sem);
2123	}
2124#endif
2125}
2126
2127/*
2128 *  this is really a simplified "do_mmap".  it only handles
2129 *  anonymous maps.  eventually we may be able to do some
2130 *  brk-specific accounting here.
2131 */
2132unsigned long do_brk(unsigned long addr, unsigned long len)
2133{
2134	struct mm_struct * mm = current->mm;
2135	struct vm_area_struct * vma, * prev;
2136	unsigned long flags;
2137	struct rb_node ** rb_link, * rb_parent;
2138	pgoff_t pgoff = addr >> PAGE_SHIFT;
2139	int error;
2140
2141	len = PAGE_ALIGN(len);
2142	if (!len)
2143		return addr;
2144
2145	error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2146	if (error)
2147		return error;
2148
2149	flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2150
2151	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2152	if (error & ~PAGE_MASK)
2153		return error;
2154
2155	/*
2156	 * mlock MCL_FUTURE?
2157	 */
2158	if (mm->def_flags & VM_LOCKED) {
2159		unsigned long locked, lock_limit;
2160		locked = len >> PAGE_SHIFT;
2161		locked += mm->locked_vm;
2162		lock_limit = rlimit(RLIMIT_MEMLOCK);
2163		lock_limit >>= PAGE_SHIFT;
2164		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2165			return -EAGAIN;
2166	}
2167
2168	/*
2169	 * mm->mmap_sem is required to protect against another thread
2170	 * changing the mappings in case we sleep.
2171	 */
2172	verify_mm_writelocked(mm);
2173
2174	/*
2175	 * Clear old maps.  this also does some error checking for us
2176	 */
2177 munmap_back:
2178	vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2179	if (vma && vma->vm_start < addr + len) {
2180		if (do_munmap(mm, addr, len))
2181			return -ENOMEM;
2182		goto munmap_back;
2183	}
2184
2185	/* Check against address space limits *after* clearing old maps... */
2186	if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2187		return -ENOMEM;
2188
2189	if (mm->map_count > sysctl_max_map_count)
2190		return -ENOMEM;
2191
2192	if (security_vm_enough_memory(len >> PAGE_SHIFT))
2193		return -ENOMEM;
2194
2195	/* Can we just expand an old private anonymous mapping? */
2196	vma = vma_merge(mm, prev, addr, addr + len, flags,
2197					NULL, NULL, pgoff, NULL);
2198	if (vma)
2199		goto out;
2200
2201	/*
2202	 * create a vma struct for an anonymous mapping
2203	 */
2204	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2205	if (!vma) {
2206		vm_unacct_memory(len >> PAGE_SHIFT);
2207		return -ENOMEM;
2208	}
2209
2210	INIT_LIST_HEAD(&vma->anon_vma_chain);
2211	vma->vm_mm = mm;
2212	vma->vm_start = addr;
2213	vma->vm_end = addr + len;
2214	vma->vm_pgoff = pgoff;
2215	vma->vm_flags = flags;
2216	vma->vm_page_prot = vm_get_page_prot(flags);
2217	vma_link(mm, vma, prev, rb_link, rb_parent);
2218out:
2219	perf_event_mmap(vma);
2220	mm->total_vm += len >> PAGE_SHIFT;
2221	if (flags & VM_LOCKED) {
2222		if (!mlock_vma_pages_range(vma, addr, addr + len))
2223			mm->locked_vm += (len >> PAGE_SHIFT);
2224	}
2225	return addr;
2226}
2227
2228EXPORT_SYMBOL(do_brk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2229
2230/* Release all mmaps. */
2231void exit_mmap(struct mm_struct *mm)
2232{
2233	struct mmu_gather tlb;
2234	struct vm_area_struct *vma;
2235	unsigned long nr_accounted = 0;
2236	unsigned long end;
2237
2238	/* mm's last user has gone, and its about to be pulled down */
2239	mmu_notifier_release(mm);
2240
2241	if (mm->locked_vm) {
2242		vma = mm->mmap;
2243		while (vma) {
2244			if (vma->vm_flags & VM_LOCKED)
2245				munlock_vma_pages_all(vma);
2246			vma = vma->vm_next;
2247		}
2248	}
2249
2250	arch_exit_mmap(mm);
2251
2252	vma = mm->mmap;
2253	if (!vma)	/* Can happen if dup_mmap() received an OOM */
2254		return;
2255
2256	lru_add_drain();
2257	flush_cache_mm(mm);
2258	tlb_gather_mmu(&tlb, mm, 1);
2259	/* update_hiwater_rss(mm) here? but nobody should be looking */
2260	/* Use -1 here to ensure all VMAs in the mm are unmapped */
2261	end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2262	vm_unacct_memory(nr_accounted);
2263
2264	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2265	tlb_finish_mmu(&tlb, 0, end);
2266
2267	/*
2268	 * Walk the list again, actually closing and freeing it,
2269	 * with preemption enabled, without holding any MM locks.
2270	 */
2271	while (vma)
 
 
2272		vma = remove_vma(vma);
 
 
2273
2274	BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
 
2275}
2276
2277/* Insert vm structure into process list sorted by address
2278 * and into the inode's i_mmap tree.  If vm_file is non-NULL
2279 * then i_mmap_mutex is taken here.
2280 */
2281int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2282{
2283	struct vm_area_struct * __vma, * prev;
2284	struct rb_node ** rb_link, * rb_parent;
2285
2286	/*
2287	 * The vm_pgoff of a purely anonymous vma should be irrelevant
2288	 * until its first write fault, when page's anon_vma and index
2289	 * are set.  But now set the vm_pgoff it will almost certainly
2290	 * end up with (unless mremap moves it elsewhere before that
2291	 * first wfault), so /proc/pid/maps tells a consistent story.
2292	 *
2293	 * By setting it to reflect the virtual start address of the
2294	 * vma, merges and splits can happen in a seamless way, just
2295	 * using the existing file pgoff checks and manipulations.
2296	 * Similarly in do_mmap_pgoff and in do_brk.
2297	 */
2298	if (!vma->vm_file) {
2299		BUG_ON(vma->anon_vma);
2300		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2301	}
2302	__vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2303	if (__vma && __vma->vm_start < vma->vm_end)
2304		return -ENOMEM;
2305	if ((vma->vm_flags & VM_ACCOUNT) &&
2306	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
2307		return -ENOMEM;
 
2308	vma_link(mm, vma, prev, rb_link, rb_parent);
2309	return 0;
2310}
2311
2312/*
2313 * Copy the vma structure to a new location in the same mm,
2314 * prior to moving page table entries, to effect an mremap move.
2315 */
2316struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2317	unsigned long addr, unsigned long len, pgoff_t pgoff)
 
2318{
2319	struct vm_area_struct *vma = *vmap;
2320	unsigned long vma_start = vma->vm_start;
2321	struct mm_struct *mm = vma->vm_mm;
2322	struct vm_area_struct *new_vma, *prev;
2323	struct rb_node **rb_link, *rb_parent;
2324	struct mempolicy *pol;
2325
2326	/*
2327	 * If anonymous vma has not yet been faulted, update new pgoff
2328	 * to match new location, to increase its chance of merging.
2329	 */
2330	if (!vma->vm_file && !vma->anon_vma)
2331		pgoff = addr >> PAGE_SHIFT;
 
 
2332
2333	find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
 
2334	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2335			vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2336	if (new_vma) {
2337		/*
2338		 * Source vma may have been merged into new_vma
2339		 */
2340		if (vma_start >= new_vma->vm_start &&
2341		    vma_start < new_vma->vm_end)
2342			*vmap = new_vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2343	} else {
2344		new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2345		if (new_vma) {
2346			*new_vma = *vma;
2347			pol = mpol_dup(vma_policy(vma));
2348			if (IS_ERR(pol))
 
 
2349				goto out_free_vma;
2350			INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2351			if (anon_vma_clone(new_vma, vma))
2352				goto out_free_mempol;
2353			vma_set_policy(new_vma, pol);
2354			new_vma->vm_start = addr;
2355			new_vma->vm_end = addr + len;
2356			new_vma->vm_pgoff = pgoff;
2357			if (new_vma->vm_file) {
2358				get_file(new_vma->vm_file);
2359				if (vma->vm_flags & VM_EXECUTABLE)
2360					added_exe_file_vma(mm);
2361			}
2362			if (new_vma->vm_ops && new_vma->vm_ops->open)
2363				new_vma->vm_ops->open(new_vma);
2364			vma_link(mm, new_vma, prev, rb_link, rb_parent);
 
2365		}
2366	}
2367	return new_vma;
2368
2369 out_free_mempol:
2370	mpol_put(pol);
2371 out_free_vma:
2372	kmem_cache_free(vm_area_cachep, new_vma);
2373	return NULL;
2374}
2375
2376/*
2377 * Return true if the calling process may expand its vm space by the passed
2378 * number of pages
2379 */
2380int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2381{
2382	unsigned long cur = mm->total_vm;	/* pages */
2383	unsigned long lim;
2384
2385	lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2386
2387	if (cur + npages > lim)
2388		return 0;
2389	return 1;
2390}
2391
2392
2393static int special_mapping_fault(struct vm_area_struct *vma,
2394				struct vm_fault *vmf)
2395{
2396	pgoff_t pgoff;
2397	struct page **pages;
2398
2399	/*
2400	 * special mappings have no vm_file, and in that case, the mm
2401	 * uses vm_pgoff internally. So we have to subtract it from here.
2402	 * We are allowed to do this because we are the mm; do not copy
2403	 * this code into drivers!
2404	 */
2405	pgoff = vmf->pgoff - vma->vm_pgoff;
2406
2407	for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2408		pgoff--;
2409
2410	if (*pages) {
2411		struct page *page = *pages;
2412		get_page(page);
2413		vmf->page = page;
2414		return 0;
2415	}
2416
2417	return VM_FAULT_SIGBUS;
2418}
2419
2420/*
2421 * Having a close hook prevents vma merging regardless of flags.
2422 */
2423static void special_mapping_close(struct vm_area_struct *vma)
2424{
2425}
2426
2427static const struct vm_operations_struct special_mapping_vmops = {
2428	.close = special_mapping_close,
2429	.fault = special_mapping_fault,
2430};
2431
2432/*
2433 * Called with mm->mmap_sem held for writing.
2434 * Insert a new vma covering the given region, with the given flags.
2435 * Its pages are supplied by the given array of struct page *.
2436 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2437 * The region past the last page supplied will always produce SIGBUS.
2438 * The array pointer and the pages it points to are assumed to stay alive
2439 * for as long as this mapping might exist.
2440 */
2441int install_special_mapping(struct mm_struct *mm,
2442			    unsigned long addr, unsigned long len,
2443			    unsigned long vm_flags, struct page **pages)
2444{
2445	int ret;
2446	struct vm_area_struct *vma;
2447
2448	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2449	if (unlikely(vma == NULL))
2450		return -ENOMEM;
2451
2452	INIT_LIST_HEAD(&vma->anon_vma_chain);
2453	vma->vm_mm = mm;
2454	vma->vm_start = addr;
2455	vma->vm_end = addr + len;
2456
2457	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2458	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2459
2460	vma->vm_ops = &special_mapping_vmops;
2461	vma->vm_private_data = pages;
2462
2463	ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2464	if (ret)
2465		goto out;
2466
2467	ret = insert_vm_struct(mm, vma);
2468	if (ret)
2469		goto out;
2470
2471	mm->total_vm += len >> PAGE_SHIFT;
2472
2473	perf_event_mmap(vma);
2474
2475	return 0;
2476
2477out:
2478	kmem_cache_free(vm_area_cachep, vma);
2479	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
2480}
2481
2482static DEFINE_MUTEX(mm_all_locks_mutex);
2483
2484static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2485{
2486	if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2487		/*
2488		 * The LSB of head.next can't change from under us
2489		 * because we hold the mm_all_locks_mutex.
2490		 */
2491		mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2492		/*
2493		 * We can safely modify head.next after taking the
2494		 * anon_vma->root->mutex. If some other vma in this mm shares
2495		 * the same anon_vma we won't take it again.
2496		 *
2497		 * No need of atomic instructions here, head.next
2498		 * can't change from under us thanks to the
2499		 * anon_vma->root->mutex.
2500		 */
2501		if (__test_and_set_bit(0, (unsigned long *)
2502				       &anon_vma->root->head.next))
2503			BUG();
2504	}
2505}
2506
2507static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2508{
2509	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2510		/*
2511		 * AS_MM_ALL_LOCKS can't change from under us because
2512		 * we hold the mm_all_locks_mutex.
2513		 *
2514		 * Operations on ->flags have to be atomic because
2515		 * even if AS_MM_ALL_LOCKS is stable thanks to the
2516		 * mm_all_locks_mutex, there may be other cpus
2517		 * changing other bitflags in parallel to us.
2518		 */
2519		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2520			BUG();
2521		mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2522	}
2523}
2524
2525/*
2526 * This operation locks against the VM for all pte/vma/mm related
2527 * operations that could ever happen on a certain mm. This includes
2528 * vmtruncate, try_to_unmap, and all page faults.
2529 *
2530 * The caller must take the mmap_sem in write mode before calling
2531 * mm_take_all_locks(). The caller isn't allowed to release the
2532 * mmap_sem until mm_drop_all_locks() returns.
2533 *
2534 * mmap_sem in write mode is required in order to block all operations
2535 * that could modify pagetables and free pages without need of
2536 * altering the vma layout (for example populate_range() with
2537 * nonlinear vmas). It's also needed in write mode to avoid new
2538 * anon_vmas to be associated with existing vmas.
2539 *
2540 * A single task can't take more than one mm_take_all_locks() in a row
2541 * or it would deadlock.
2542 *
2543 * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2544 * mapping->flags avoid to take the same lock twice, if more than one
2545 * vma in this mm is backed by the same anon_vma or address_space.
2546 *
2547 * We can take all the locks in random order because the VM code
2548 * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2549 * takes more than one of them in a row. Secondly we're protected
2550 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2551 *
2552 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2553 * that may have to take thousand of locks.
2554 *
2555 * mm_take_all_locks() can fail if it's interrupted by signals.
2556 */
2557int mm_take_all_locks(struct mm_struct *mm)
2558{
2559	struct vm_area_struct *vma;
2560	struct anon_vma_chain *avc;
2561	int ret = -EINTR;
2562
2563	BUG_ON(down_read_trylock(&mm->mmap_sem));
2564
2565	mutex_lock(&mm_all_locks_mutex);
2566
2567	for (vma = mm->mmap; vma; vma = vma->vm_next) {
2568		if (signal_pending(current))
2569			goto out_unlock;
2570		if (vma->vm_file && vma->vm_file->f_mapping)
2571			vm_lock_mapping(mm, vma->vm_file->f_mapping);
2572	}
2573
2574	for (vma = mm->mmap; vma; vma = vma->vm_next) {
2575		if (signal_pending(current))
2576			goto out_unlock;
2577		if (vma->anon_vma)
2578			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2579				vm_lock_anon_vma(mm, avc->anon_vma);
2580	}
2581
2582	ret = 0;
2583
2584out_unlock:
2585	if (ret)
2586		mm_drop_all_locks(mm);
2587
2588	return ret;
2589}
2590
2591static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2592{
2593	if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2594		/*
2595		 * The LSB of head.next can't change to 0 from under
2596		 * us because we hold the mm_all_locks_mutex.
2597		 *
2598		 * We must however clear the bitflag before unlocking
2599		 * the vma so the users using the anon_vma->head will
2600		 * never see our bitflag.
2601		 *
2602		 * No need of atomic instructions here, head.next
2603		 * can't change from under us until we release the
2604		 * anon_vma->root->mutex.
2605		 */
2606		if (!__test_and_clear_bit(0, (unsigned long *)
2607					  &anon_vma->root->head.next))
2608			BUG();
2609		anon_vma_unlock(anon_vma);
2610	}
2611}
2612
2613static void vm_unlock_mapping(struct address_space *mapping)
2614{
2615	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2616		/*
2617		 * AS_MM_ALL_LOCKS can't change to 0 from under us
2618		 * because we hold the mm_all_locks_mutex.
2619		 */
2620		mutex_unlock(&mapping->i_mmap_mutex);
2621		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2622					&mapping->flags))
2623			BUG();
2624	}
2625}
2626
2627/*
2628 * The mmap_sem cannot be released by the caller until
2629 * mm_drop_all_locks() returns.
2630 */
2631void mm_drop_all_locks(struct mm_struct *mm)
2632{
2633	struct vm_area_struct *vma;
2634	struct anon_vma_chain *avc;
2635
2636	BUG_ON(down_read_trylock(&mm->mmap_sem));
2637	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2638
2639	for (vma = mm->mmap; vma; vma = vma->vm_next) {
2640		if (vma->anon_vma)
2641			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2642				vm_unlock_anon_vma(avc->anon_vma);
2643		if (vma->vm_file && vma->vm_file->f_mapping)
2644			vm_unlock_mapping(vma->vm_file->f_mapping);
2645	}
2646
2647	mutex_unlock(&mm_all_locks_mutex);
2648}
2649
2650/*
2651 * initialise the VMA slab
2652 */
2653void __init mmap_init(void)
2654{
2655	int ret;
2656
2657	ret = percpu_counter_init(&vm_committed_as, 0);
2658	VM_BUG_ON(ret);
2659}
v3.15
   1/*
   2 * mm/mmap.c
   3 *
   4 * Written by obz.
   5 *
   6 * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
   7 */
   8
   9#include <linux/kernel.h>
  10#include <linux/slab.h>
  11#include <linux/backing-dev.h>
  12#include <linux/mm.h>
  13#include <linux/vmacache.h>
  14#include <linux/shm.h>
  15#include <linux/mman.h>
  16#include <linux/pagemap.h>
  17#include <linux/swap.h>
  18#include <linux/syscalls.h>
  19#include <linux/capability.h>
  20#include <linux/init.h>
  21#include <linux/file.h>
  22#include <linux/fs.h>
  23#include <linux/personality.h>
  24#include <linux/security.h>
  25#include <linux/hugetlb.h>
  26#include <linux/profile.h>
  27#include <linux/export.h>
  28#include <linux/mount.h>
  29#include <linux/mempolicy.h>
  30#include <linux/rmap.h>
  31#include <linux/mmu_notifier.h>
  32#include <linux/perf_event.h>
  33#include <linux/audit.h>
  34#include <linux/khugepaged.h>
  35#include <linux/uprobes.h>
  36#include <linux/rbtree_augmented.h>
  37#include <linux/sched/sysctl.h>
  38#include <linux/notifier.h>
  39#include <linux/memory.h>
  40
  41#include <asm/uaccess.h>
  42#include <asm/cacheflush.h>
  43#include <asm/tlb.h>
  44#include <asm/mmu_context.h>
  45
  46#include "internal.h"
  47
  48#ifndef arch_mmap_check
  49#define arch_mmap_check(addr, len, flags)	(0)
  50#endif
  51
  52#ifndef arch_rebalance_pgtables
  53#define arch_rebalance_pgtables(addr, len)		(addr)
  54#endif
  55
  56static void unmap_region(struct mm_struct *mm,
  57		struct vm_area_struct *vma, struct vm_area_struct *prev,
  58		unsigned long start, unsigned long end);
  59
 
 
 
 
 
 
  60/* description of effects of mapping type and prot in current implementation.
  61 * this is due to the limited x86 page protection hardware.  The expected
  62 * behavior is in parens:
  63 *
  64 * map_type	prot
  65 *		PROT_NONE	PROT_READ	PROT_WRITE	PROT_EXEC
  66 * MAP_SHARED	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  67 *		w: (no) no	w: (no) no	w: (yes) yes	w: (no) no
  68 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  69 *		
  70 * MAP_PRIVATE	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  71 *		w: (no) no	w: (no) no	w: (copy) copy	w: (no) no
  72 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  73 *
  74 */
  75pgprot_t protection_map[16] = {
  76	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
  77	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
  78};
  79
  80pgprot_t vm_get_page_prot(unsigned long vm_flags)
  81{
  82	return __pgprot(pgprot_val(protection_map[vm_flags &
  83				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
  84			pgprot_val(arch_vm_get_page_prot(vm_flags)));
  85}
  86EXPORT_SYMBOL(vm_get_page_prot);
  87
  88int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
  89int sysctl_overcommit_ratio __read_mostly = 50;	/* default is 50% */
  90unsigned long sysctl_overcommit_kbytes __read_mostly;
  91int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
  92unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
  93unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
  94/*
  95 * Make sure vm_committed_as in one cacheline and not cacheline shared with
  96 * other variables. It can be updated by several CPUs frequently.
  97 */
  98struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
  99
 100/*
 101 * The global memory commitment made in the system can be a metric
 102 * that can be used to drive ballooning decisions when Linux is hosted
 103 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
 104 * balancing memory across competing virtual machines that are hosted.
 105 * Several metrics drive this policy engine including the guest reported
 106 * memory commitment.
 107 */
 108unsigned long vm_memory_committed(void)
 109{
 110	return percpu_counter_read_positive(&vm_committed_as);
 111}
 112EXPORT_SYMBOL_GPL(vm_memory_committed);
 113
 114/*
 115 * Check that a process has enough memory to allocate a new virtual
 116 * mapping. 0 means there is enough memory for the allocation to
 117 * succeed and -ENOMEM implies there is not.
 118 *
 119 * We currently support three overcommit policies, which are set via the
 120 * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
 121 *
 122 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
 123 * Additional code 2002 Jul 20 by Robert Love.
 124 *
 125 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
 126 *
 127 * Note this is a helper function intended to be used by LSMs which
 128 * wish to use this logic.
 129 */
 130int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
 131{
 132	unsigned long free, allowed, reserve;
 133
 134	vm_acct_memory(pages);
 135
 136	/*
 137	 * Sometimes we want to use more memory than we have
 138	 */
 139	if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
 140		return 0;
 141
 142	if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
 143		free = global_page_state(NR_FREE_PAGES);
 144		free += global_page_state(NR_FILE_PAGES);
 145
 146		/*
 147		 * shmem pages shouldn't be counted as free in this
 148		 * case, they can't be purged, only swapped out, and
 149		 * that won't affect the overall amount of available
 150		 * memory in the system.
 151		 */
 152		free -= global_page_state(NR_SHMEM);
 153
 154		free += get_nr_swap_pages();
 155
 156		/*
 157		 * Any slabs which are created with the
 158		 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
 159		 * which are reclaimable, under pressure.  The dentry
 160		 * cache and most inode caches should fall into this
 161		 */
 162		free += global_page_state(NR_SLAB_RECLAIMABLE);
 163
 164		/*
 165		 * Leave reserved pages. The pages are not for anonymous pages.
 166		 */
 167		if (free <= totalreserve_pages)
 168			goto error;
 169		else
 170			free -= totalreserve_pages;
 171
 172		/*
 173		 * Reserve some for root
 174		 */
 175		if (!cap_sys_admin)
 176			free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
 177
 178		if (free > pages)
 179			return 0;
 180
 181		goto error;
 182	}
 183
 184	allowed = vm_commit_limit();
 
 185	/*
 186	 * Reserve some for root
 187	 */
 188	if (!cap_sys_admin)
 189		allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
 
 190
 191	/*
 192	 * Don't let a single process grow so big a user can't recover
 193	 */
 194	if (mm) {
 195		reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
 196		allowed -= min(mm->total_vm / 32, reserve);
 197	}
 198
 199	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
 200		return 0;
 201error:
 202	vm_unacct_memory(pages);
 203
 204	return -ENOMEM;
 205}
 206
 207/*
 208 * Requires inode->i_mapping->i_mmap_mutex
 209 */
 210static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 211		struct file *file, struct address_space *mapping)
 212{
 213	if (vma->vm_flags & VM_DENYWRITE)
 214		atomic_inc(&file_inode(file)->i_writecount);
 215	if (vma->vm_flags & VM_SHARED)
 216		mapping->i_mmap_writable--;
 217
 218	flush_dcache_mmap_lock(mapping);
 219	if (unlikely(vma->vm_flags & VM_NONLINEAR))
 220		list_del_init(&vma->shared.nonlinear);
 221	else
 222		vma_interval_tree_remove(vma, &mapping->i_mmap);
 223	flush_dcache_mmap_unlock(mapping);
 224}
 225
 226/*
 227 * Unlink a file-based vm structure from its interval tree, to hide
 228 * vma from rmap and vmtruncate before freeing its page tables.
 229 */
 230void unlink_file_vma(struct vm_area_struct *vma)
 231{
 232	struct file *file = vma->vm_file;
 233
 234	if (file) {
 235		struct address_space *mapping = file->f_mapping;
 236		mutex_lock(&mapping->i_mmap_mutex);
 237		__remove_shared_vm_struct(vma, file, mapping);
 238		mutex_unlock(&mapping->i_mmap_mutex);
 239	}
 240}
 241
 242/*
 243 * Close a vm structure and free it, returning the next.
 244 */
 245static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
 246{
 247	struct vm_area_struct *next = vma->vm_next;
 248
 249	might_sleep();
 250	if (vma->vm_ops && vma->vm_ops->close)
 251		vma->vm_ops->close(vma);
 252	if (vma->vm_file)
 253		fput(vma->vm_file);
 
 
 
 254	mpol_put(vma_policy(vma));
 255	kmem_cache_free(vm_area_cachep, vma);
 256	return next;
 257}
 258
 259static unsigned long do_brk(unsigned long addr, unsigned long len);
 260
 261SYSCALL_DEFINE1(brk, unsigned long, brk)
 262{
 263	unsigned long rlim, retval;
 264	unsigned long newbrk, oldbrk;
 265	struct mm_struct *mm = current->mm;
 266	unsigned long min_brk;
 267	bool populate;
 268
 269	down_write(&mm->mmap_sem);
 270
 271#ifdef CONFIG_COMPAT_BRK
 272	/*
 273	 * CONFIG_COMPAT_BRK can still be overridden by setting
 274	 * randomize_va_space to 2, which will still cause mm->start_brk
 275	 * to be arbitrarily shifted
 276	 */
 277	if (current->brk_randomized)
 278		min_brk = mm->start_brk;
 279	else
 280		min_brk = mm->end_data;
 281#else
 282	min_brk = mm->start_brk;
 283#endif
 284	if (brk < min_brk)
 285		goto out;
 286
 287	/*
 288	 * Check against rlimit here. If this check is done later after the test
 289	 * of oldbrk with newbrk then it can escape the test and let the data
 290	 * segment grow beyond its set limit the in case where the limit is
 291	 * not page aligned -Ram Gupta
 292	 */
 293	rlim = rlimit(RLIMIT_DATA);
 294	if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
 295			(mm->end_data - mm->start_data) > rlim)
 296		goto out;
 297
 298	newbrk = PAGE_ALIGN(brk);
 299	oldbrk = PAGE_ALIGN(mm->brk);
 300	if (oldbrk == newbrk)
 301		goto set_brk;
 302
 303	/* Always allow shrinking brk. */
 304	if (brk <= mm->brk) {
 305		if (!do_munmap(mm, newbrk, oldbrk-newbrk))
 306			goto set_brk;
 307		goto out;
 308	}
 309
 310	/* Check against existing mmap mappings. */
 311	if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
 312		goto out;
 313
 314	/* Ok, looks good - let it rip. */
 315	if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
 316		goto out;
 317
 318set_brk:
 319	mm->brk = brk;
 320	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
 321	up_write(&mm->mmap_sem);
 322	if (populate)
 323		mm_populate(oldbrk, newbrk - oldbrk);
 324	return brk;
 325
 326out:
 327	retval = mm->brk;
 328	up_write(&mm->mmap_sem);
 329	return retval;
 330}
 331
 332static long vma_compute_subtree_gap(struct vm_area_struct *vma)
 333{
 334	unsigned long max, subtree_gap;
 335	max = vma->vm_start;
 336	if (vma->vm_prev)
 337		max -= vma->vm_prev->vm_end;
 338	if (vma->vm_rb.rb_left) {
 339		subtree_gap = rb_entry(vma->vm_rb.rb_left,
 340				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 341		if (subtree_gap > max)
 342			max = subtree_gap;
 343	}
 344	if (vma->vm_rb.rb_right) {
 345		subtree_gap = rb_entry(vma->vm_rb.rb_right,
 346				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 347		if (subtree_gap > max)
 348			max = subtree_gap;
 349	}
 350	return max;
 351}
 352
 353#ifdef CONFIG_DEBUG_VM_RB
 354static int browse_rb(struct rb_root *root)
 355{
 356	int i = 0, j, bug = 0;
 357	struct rb_node *nd, *pn = NULL;
 358	unsigned long prev = 0, pend = 0;
 359
 360	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 361		struct vm_area_struct *vma;
 362		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 363		if (vma->vm_start < prev) {
 364			printk("vm_start %lx prev %lx\n", vma->vm_start, prev);
 365			bug = 1;
 366		}
 367		if (vma->vm_start < pend) {
 368			printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
 369			bug = 1;
 370		}
 371		if (vma->vm_start > vma->vm_end) {
 372			printk("vm_end %lx < vm_start %lx\n",
 373				vma->vm_end, vma->vm_start);
 374			bug = 1;
 375		}
 376		if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
 377			printk("free gap %lx, correct %lx\n",
 378			       vma->rb_subtree_gap,
 379			       vma_compute_subtree_gap(vma));
 380			bug = 1;
 381		}
 382		i++;
 383		pn = nd;
 384		prev = vma->vm_start;
 385		pend = vma->vm_end;
 386	}
 387	j = 0;
 388	for (nd = pn; nd; nd = rb_prev(nd))
 389		j++;
 390	if (i != j) {
 391		printk("backwards %d, forwards %d\n", j, i);
 392		bug = 1;
 393	}
 394	return bug ? -1 : i;
 
 
 395}
 396
 397static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
 398{
 399	struct rb_node *nd;
 400
 401	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 402		struct vm_area_struct *vma;
 403		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 404		BUG_ON(vma != ignore &&
 405		       vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
 406	}
 407}
 408
 409static void validate_mm(struct mm_struct *mm)
 410{
 411	int bug = 0;
 412	int i = 0;
 413	unsigned long highest_address = 0;
 414	struct vm_area_struct *vma = mm->mmap;
 415	while (vma) {
 416		struct anon_vma_chain *avc;
 417		vma_lock_anon_vma(vma);
 418		list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 419			anon_vma_interval_tree_verify(avc);
 420		vma_unlock_anon_vma(vma);
 421		highest_address = vma->vm_end;
 422		vma = vma->vm_next;
 423		i++;
 424	}
 425	if (i != mm->map_count) {
 426		printk("map_count %d vm_next %d\n", mm->map_count, i);
 427		bug = 1;
 428	}
 429	if (highest_address != mm->highest_vm_end) {
 430		printk("mm->highest_vm_end %lx, found %lx\n",
 431		       mm->highest_vm_end, highest_address);
 432		bug = 1;
 433	}
 434	i = browse_rb(&mm->mm_rb);
 435	if (i != mm->map_count) {
 436		printk("map_count %d rb %d\n", mm->map_count, i);
 437		bug = 1;
 438	}
 439	BUG_ON(bug);
 440}
 441#else
 442#define validate_mm_rb(root, ignore) do { } while (0)
 443#define validate_mm(mm) do { } while (0)
 444#endif
 445
 446RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
 447		     unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
 448
 449/*
 450 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
 451 * vma->vm_prev->vm_end values changed, without modifying the vma's position
 452 * in the rbtree.
 453 */
 454static void vma_gap_update(struct vm_area_struct *vma)
 455{
 456	/*
 457	 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
 458	 * function that does exacltly what we want.
 459	 */
 460	vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
 461}
 462
 463static inline void vma_rb_insert(struct vm_area_struct *vma,
 464				 struct rb_root *root)
 465{
 466	/* All rb_subtree_gap values must be consistent prior to insertion */
 467	validate_mm_rb(root, NULL);
 468
 469	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 470}
 471
 472static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
 473{
 474	/*
 475	 * All rb_subtree_gap values must be consistent prior to erase,
 476	 * with the possible exception of the vma being erased.
 477	 */
 478	validate_mm_rb(root, vma);
 479
 480	/*
 481	 * Note rb_erase_augmented is a fairly large inline function,
 482	 * so make sure we instantiate it only once with our desired
 483	 * augmented rbtree callbacks.
 484	 */
 485	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 486}
 487
 488/*
 489 * vma has some anon_vma assigned, and is already inserted on that
 490 * anon_vma's interval trees.
 491 *
 492 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
 493 * vma must be removed from the anon_vma's interval trees using
 494 * anon_vma_interval_tree_pre_update_vma().
 495 *
 496 * After the update, the vma will be reinserted using
 497 * anon_vma_interval_tree_post_update_vma().
 498 *
 499 * The entire update must be protected by exclusive mmap_sem and by
 500 * the root anon_vma's mutex.
 501 */
 502static inline void
 503anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
 504{
 505	struct anon_vma_chain *avc;
 506
 507	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 508		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
 509}
 510
 511static inline void
 512anon_vma_interval_tree_post_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_insert(avc, &avc->anon_vma->rb_root);
 518}
 519
 520static int find_vma_links(struct mm_struct *mm, unsigned long addr,
 521		unsigned long end, struct vm_area_struct **pprev,
 522		struct rb_node ***rb_link, struct rb_node **rb_parent)
 523{
 524	struct rb_node **__rb_link, *__rb_parent, *rb_prev;
 525
 526	__rb_link = &mm->mm_rb.rb_node;
 527	rb_prev = __rb_parent = NULL;
 
 528
 529	while (*__rb_link) {
 530		struct vm_area_struct *vma_tmp;
 531
 532		__rb_parent = *__rb_link;
 533		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
 534
 535		if (vma_tmp->vm_end > addr) {
 536			/* Fail if an existing vma overlaps the area */
 537			if (vma_tmp->vm_start < end)
 538				return -ENOMEM;
 539			__rb_link = &__rb_parent->rb_left;
 540		} else {
 541			rb_prev = __rb_parent;
 542			__rb_link = &__rb_parent->rb_right;
 543		}
 544	}
 545
 546	*pprev = NULL;
 547	if (rb_prev)
 548		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 549	*rb_link = __rb_link;
 550	*rb_parent = __rb_parent;
 551	return 0;
 552}
 553
 554static unsigned long count_vma_pages_range(struct mm_struct *mm,
 555		unsigned long addr, unsigned long end)
 556{
 557	unsigned long nr_pages = 0;
 558	struct vm_area_struct *vma;
 559
 560	/* Find first overlaping mapping */
 561	vma = find_vma_intersection(mm, addr, end);
 562	if (!vma)
 563		return 0;
 564
 565	nr_pages = (min(end, vma->vm_end) -
 566		max(addr, vma->vm_start)) >> PAGE_SHIFT;
 567
 568	/* Iterate over the rest of the overlaps */
 569	for (vma = vma->vm_next; vma; vma = vma->vm_next) {
 570		unsigned long overlap_len;
 571
 572		if (vma->vm_start > end)
 573			break;
 574
 575		overlap_len = min(end, vma->vm_end) - vma->vm_start;
 576		nr_pages += overlap_len >> PAGE_SHIFT;
 577	}
 578
 579	return nr_pages;
 580}
 581
 582void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 583		struct rb_node **rb_link, struct rb_node *rb_parent)
 584{
 585	/* Update tracking information for the gap following the new vma. */
 586	if (vma->vm_next)
 587		vma_gap_update(vma->vm_next);
 588	else
 589		mm->highest_vm_end = vma->vm_end;
 590
 591	/*
 592	 * vma->vm_prev wasn't known when we followed the rbtree to find the
 593	 * correct insertion point for that vma. As a result, we could not
 594	 * update the vma vm_rb parents rb_subtree_gap values on the way down.
 595	 * So, we first insert the vma with a zero rb_subtree_gap value
 596	 * (to be consistent with what we did on the way down), and then
 597	 * immediately update the gap to the correct value. Finally we
 598	 * rebalance the rbtree after all augmented values have been set.
 599	 */
 600	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 601	vma->rb_subtree_gap = 0;
 602	vma_gap_update(vma);
 603	vma_rb_insert(vma, &mm->mm_rb);
 604}
 605
 606static void __vma_link_file(struct vm_area_struct *vma)
 607{
 608	struct file *file;
 609
 610	file = vma->vm_file;
 611	if (file) {
 612		struct address_space *mapping = file->f_mapping;
 613
 614		if (vma->vm_flags & VM_DENYWRITE)
 615			atomic_dec(&file_inode(file)->i_writecount);
 616		if (vma->vm_flags & VM_SHARED)
 617			mapping->i_mmap_writable++;
 618
 619		flush_dcache_mmap_lock(mapping);
 620		if (unlikely(vma->vm_flags & VM_NONLINEAR))
 621			vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 622		else
 623			vma_interval_tree_insert(vma, &mapping->i_mmap);
 624		flush_dcache_mmap_unlock(mapping);
 625	}
 626}
 627
 628static void
 629__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 630	struct vm_area_struct *prev, struct rb_node **rb_link,
 631	struct rb_node *rb_parent)
 632{
 633	__vma_link_list(mm, vma, prev, rb_parent);
 634	__vma_link_rb(mm, vma, rb_link, rb_parent);
 635}
 636
 637static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 638			struct vm_area_struct *prev, struct rb_node **rb_link,
 639			struct rb_node *rb_parent)
 640{
 641	struct address_space *mapping = NULL;
 642
 643	if (vma->vm_file)
 644		mapping = vma->vm_file->f_mapping;
 645
 646	if (mapping)
 647		mutex_lock(&mapping->i_mmap_mutex);
 648
 649	__vma_link(mm, vma, prev, rb_link, rb_parent);
 650	__vma_link_file(vma);
 651
 652	if (mapping)
 653		mutex_unlock(&mapping->i_mmap_mutex);
 654
 655	mm->map_count++;
 656	validate_mm(mm);
 657}
 658
 659/*
 660 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
 661 * mm's list and rbtree.  It has already been inserted into the interval tree.
 
 662 */
 663static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 664{
 665	struct vm_area_struct *prev;
 666	struct rb_node **rb_link, *rb_parent;
 667
 668	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
 669			   &prev, &rb_link, &rb_parent))
 670		BUG();
 671	__vma_link(mm, vma, prev, rb_link, rb_parent);
 672	mm->map_count++;
 673}
 674
 675static inline void
 676__vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
 677		struct vm_area_struct *prev)
 678{
 679	struct vm_area_struct *next;
 680
 681	vma_rb_erase(vma, &mm->mm_rb);
 682	prev->vm_next = next = vma->vm_next;
 683	if (next)
 684		next->vm_prev = prev;
 685
 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{
 700	struct mm_struct *mm = vma->vm_mm;
 701	struct vm_area_struct *next = vma->vm_next;
 702	struct vm_area_struct *importer = NULL;
 703	struct address_space *mapping = NULL;
 704	struct rb_root *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;
 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			 */
 719again:			remove_next = 1 + (end > next->vm_end);
 720			end = next->vm_end;
 721			exporter = next;
 722			importer = vma;
 723		} else if (end > next->vm_start) {
 724			/*
 725			 * vma expands, overlapping part of the next:
 726			 * mprotect case 5 shifting the boundary up.
 727			 */
 728			adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
 729			exporter = next;
 730			importer = vma;
 731		} else if (end < vma->vm_end) {
 732			/*
 733			 * vma shrinks, and !insert tells it's not
 734			 * split_vma inserting another: so it must be
 735			 * mprotect case 4 shifting the boundary down.
 736			 */
 737			adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
 738			exporter = vma;
 739			importer = next;
 740		}
 741
 742		/*
 743		 * Easily overlooked: when mprotect shifts the boundary,
 744		 * make sure the expanding vma has anon_vma set if the
 745		 * shrinking vma had, to cover any anon pages imported.
 746		 */
 747		if (exporter && exporter->anon_vma && !importer->anon_vma) {
 748			if (anon_vma_clone(importer, exporter))
 749				return -ENOMEM;
 750			importer->anon_vma = exporter->anon_vma;
 751		}
 752	}
 753
 754	if (file) {
 755		mapping = file->f_mapping;
 756		if (!(vma->vm_flags & VM_NONLINEAR)) {
 757			root = &mapping->i_mmap;
 758			uprobe_munmap(vma, vma->vm_start, vma->vm_end);
 759
 760			if (adjust_next)
 761				uprobe_munmap(next, next->vm_start,
 762							next->vm_end);
 763		}
 764
 765		mutex_lock(&mapping->i_mmap_mutex);
 766		if (insert) {
 767			/*
 768			 * Put into interval tree now, so instantiated pages
 769			 * are visible to arm/parisc __flush_dcache_page
 770			 * throughout; but we cannot insert into address
 771			 * space until vma start or end is updated.
 772			 */
 773			__vma_link_file(insert);
 774		}
 775	}
 776
 777	vma_adjust_trans_huge(vma, start, end, adjust_next);
 778
 779	anon_vma = vma->anon_vma;
 780	if (!anon_vma && adjust_next)
 781		anon_vma = next->anon_vma;
 782	if (anon_vma) {
 783		VM_BUG_ON(adjust_next && next->anon_vma &&
 784			  anon_vma != next->anon_vma);
 785		anon_vma_lock_write(anon_vma);
 786		anon_vma_interval_tree_pre_update_vma(vma);
 787		if (adjust_next)
 788			anon_vma_interval_tree_pre_update_vma(next);
 789	}
 790
 791	if (root) {
 792		flush_dcache_mmap_lock(mapping);
 793		vma_interval_tree_remove(vma, root);
 794		if (adjust_next)
 795			vma_interval_tree_remove(next, root);
 796	}
 797
 798	if (start != vma->vm_start) {
 799		vma->vm_start = start;
 800		start_changed = true;
 801	}
 802	if (end != vma->vm_end) {
 803		vma->vm_end = end;
 804		end_changed = true;
 805	}
 806	vma->vm_pgoff = pgoff;
 807	if (adjust_next) {
 808		next->vm_start += adjust_next << PAGE_SHIFT;
 809		next->vm_pgoff += adjust_next;
 810	}
 811
 812	if (root) {
 813		if (adjust_next)
 814			vma_interval_tree_insert(next, root);
 815		vma_interval_tree_insert(vma, root);
 816		flush_dcache_mmap_unlock(mapping);
 817	}
 818
 819	if (remove_next) {
 820		/*
 821		 * vma_merge has merged next into vma, and needs
 822		 * us to remove next before dropping the locks.
 823		 */
 824		__vma_unlink(mm, next, vma);
 825		if (file)
 826			__remove_shared_vm_struct(next, file, mapping);
 827	} else if (insert) {
 828		/*
 829		 * split_vma has split insert from vma, and needs
 830		 * us to insert it before dropping the locks
 831		 * (it may either follow vma or precede it).
 832		 */
 833		__insert_vm_struct(mm, insert);
 834	} else {
 835		if (start_changed)
 836			vma_gap_update(vma);
 837		if (end_changed) {
 838			if (!next)
 839				mm->highest_vm_end = end;
 840			else if (!adjust_next)
 841				vma_gap_update(next);
 842		}
 843	}
 844
 845	if (anon_vma) {
 846		anon_vma_interval_tree_post_update_vma(vma);
 847		if (adjust_next)
 848			anon_vma_interval_tree_post_update_vma(next);
 849		anon_vma_unlock_write(anon_vma);
 850	}
 851	if (mapping)
 852		mutex_unlock(&mapping->i_mmap_mutex);
 853
 854	if (root) {
 855		uprobe_mmap(vma);
 856
 857		if (adjust_next)
 858			uprobe_mmap(next);
 859	}
 860
 861	if (remove_next) {
 862		if (file) {
 863			uprobe_munmap(next, next->vm_start, next->vm_end);
 864			fput(file);
 
 
 865		}
 866		if (next->anon_vma)
 867			anon_vma_merge(vma, next);
 868		mm->map_count--;
 869		mpol_put(vma_policy(next));
 870		kmem_cache_free(vm_area_cachep, next);
 871		/*
 872		 * In mprotect's case 6 (see comments on vma_merge),
 873		 * we must remove another next too. It would clutter
 874		 * up the code too much to do both in one go.
 875		 */
 876		next = vma->vm_next;
 877		if (remove_next == 2)
 878			goto again;
 879		else if (next)
 880			vma_gap_update(next);
 881		else
 882			mm->highest_vm_end = end;
 883	}
 884	if (insert && file)
 885		uprobe_mmap(insert);
 886
 887	validate_mm(mm);
 888
 889	return 0;
 890}
 891
 892/*
 893 * If the vma has a ->close operation then the driver probably needs to release
 894 * per-vma resources, so we don't attempt to merge those.
 895 */
 896static inline int is_mergeable_vma(struct vm_area_struct *vma,
 897			struct file *file, unsigned long vm_flags)
 898{
 899	/*
 900	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
 901	 * match the flags but dirty bit -- the caller should mark
 902	 * merged VMA as dirty. If dirty bit won't be excluded from
 903	 * comparison, we increase pressue on the memory system forcing
 904	 * the kernel to generate new VMAs when old one could be
 905	 * extended instead.
 906	 */
 907	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
 908		return 0;
 909	if (vma->vm_file != file)
 910		return 0;
 911	if (vma->vm_ops && vma->vm_ops->close)
 912		return 0;
 913	return 1;
 914}
 915
 916static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
 917					struct anon_vma *anon_vma2,
 918					struct vm_area_struct *vma)
 919{
 920	/*
 921	 * The list_is_singular() test is to avoid merging VMA cloned from
 922	 * parents. This can improve scalability caused by anon_vma lock.
 923	 */
 924	if ((!anon_vma1 || !anon_vma2) && (!vma ||
 925		list_is_singular(&vma->anon_vma_chain)))
 926		return 1;
 927	return anon_vma1 == anon_vma2;
 928}
 929
 930/*
 931 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 932 * in front of (at a lower virtual address and file offset than) the vma.
 933 *
 934 * We cannot merge two vmas if they have differently assigned (non-NULL)
 935 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 936 *
 937 * We don't check here for the merged mmap wrapping around the end of pagecache
 938 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
 939 * wrap, nor mmaps which cover the final page at index -1UL.
 940 */
 941static int
 942can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
 943	struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
 944{
 945	if (is_mergeable_vma(vma, file, vm_flags) &&
 946	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 947		if (vma->vm_pgoff == vm_pgoff)
 948			return 1;
 949	}
 950	return 0;
 951}
 952
 953/*
 954 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 955 * beyond (at a higher virtual address and file offset than) the vma.
 956 *
 957 * We cannot merge two vmas if they have differently assigned (non-NULL)
 958 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 959 */
 960static int
 961can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
 962	struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
 963{
 964	if (is_mergeable_vma(vma, file, vm_flags) &&
 965	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 966		pgoff_t vm_pglen;
 967		vm_pglen = vma_pages(vma);
 968		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
 969			return 1;
 970	}
 971	return 0;
 972}
 973
 974/*
 975 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
 976 * whether that can be merged with its predecessor or its successor.
 977 * Or both (it neatly fills a hole).
 978 *
 979 * In most cases - when called for mmap, brk or mremap - [addr,end) is
 980 * certain not to be mapped by the time vma_merge is called; but when
 981 * called for mprotect, it is certain to be already mapped (either at
 982 * an offset within prev, or at the start of next), and the flags of
 983 * this area are about to be changed to vm_flags - and the no-change
 984 * case has already been eliminated.
 985 *
 986 * The following mprotect cases have to be considered, where AAAA is
 987 * the area passed down from mprotect_fixup, never extending beyond one
 988 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
 989 *
 990 *     AAAA             AAAA                AAAA          AAAA
 991 *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
 992 *    cannot merge    might become    might become    might become
 993 *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
 994 *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
 995 *    mremap move:                                    PPPPNNNNNNNN 8
 996 *        AAAA
 997 *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
 998 *    might become    case 1 below    case 2 below    case 3 below
 999 *
1000 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1001 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1002 */
1003struct vm_area_struct *vma_merge(struct mm_struct *mm,
1004			struct vm_area_struct *prev, unsigned long addr,
1005			unsigned long end, unsigned long vm_flags,
1006		     	struct anon_vma *anon_vma, struct file *file,
1007			pgoff_t pgoff, struct mempolicy *policy)
1008{
1009	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1010	struct vm_area_struct *area, *next;
1011	int err;
1012
1013	/*
1014	 * We later require that vma->vm_flags == vm_flags,
1015	 * so this tests vma->vm_flags & VM_SPECIAL, too.
1016	 */
1017	if (vm_flags & VM_SPECIAL)
1018		return NULL;
1019
1020	if (prev)
1021		next = prev->vm_next;
1022	else
1023		next = mm->mmap;
1024	area = next;
1025	if (next && next->vm_end == end)		/* cases 6, 7, 8 */
1026		next = next->vm_next;
1027
1028	/*
1029	 * Can it merge with the predecessor?
1030	 */
1031	if (prev && prev->vm_end == addr &&
1032  			mpol_equal(vma_policy(prev), policy) &&
1033			can_vma_merge_after(prev, vm_flags,
1034						anon_vma, file, pgoff)) {
1035		/*
1036		 * OK, it can.  Can we now merge in the successor as well?
1037		 */
1038		if (next && end == next->vm_start &&
1039				mpol_equal(policy, vma_policy(next)) &&
1040				can_vma_merge_before(next, vm_flags,
1041					anon_vma, file, pgoff+pglen) &&
1042				is_mergeable_anon_vma(prev->anon_vma,
1043						      next->anon_vma, NULL)) {
1044							/* cases 1, 6 */
1045			err = vma_adjust(prev, prev->vm_start,
1046				next->vm_end, prev->vm_pgoff, NULL);
1047		} else					/* cases 2, 5, 7 */
1048			err = vma_adjust(prev, prev->vm_start,
1049				end, prev->vm_pgoff, NULL);
1050		if (err)
1051			return NULL;
1052		khugepaged_enter_vma_merge(prev);
1053		return prev;
1054	}
1055
1056	/*
1057	 * Can this new request be merged in front of next?
1058	 */
1059	if (next && end == next->vm_start &&
1060 			mpol_equal(policy, vma_policy(next)) &&
1061			can_vma_merge_before(next, vm_flags,
1062					anon_vma, file, pgoff+pglen)) {
1063		if (prev && addr < prev->vm_end)	/* case 4 */
1064			err = vma_adjust(prev, prev->vm_start,
1065				addr, prev->vm_pgoff, NULL);
1066		else					/* cases 3, 8 */
1067			err = vma_adjust(area, addr, next->vm_end,
1068				next->vm_pgoff - pglen, NULL);
1069		if (err)
1070			return NULL;
1071		khugepaged_enter_vma_merge(area);
1072		return area;
1073	}
1074
1075	return NULL;
1076}
1077
1078/*
1079 * Rough compatbility check to quickly see if it's even worth looking
1080 * at sharing an anon_vma.
1081 *
1082 * They need to have the same vm_file, and the flags can only differ
1083 * in things that mprotect may change.
1084 *
1085 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1086 * we can merge the two vma's. For example, we refuse to merge a vma if
1087 * there is a vm_ops->close() function, because that indicates that the
1088 * driver is doing some kind of reference counting. But that doesn't
1089 * really matter for the anon_vma sharing case.
1090 */
1091static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1092{
1093	return a->vm_end == b->vm_start &&
1094		mpol_equal(vma_policy(a), vma_policy(b)) &&
1095		a->vm_file == b->vm_file &&
1096		!((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1097		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1098}
1099
1100/*
1101 * Do some basic sanity checking to see if we can re-use the anon_vma
1102 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1103 * the same as 'old', the other will be the new one that is trying
1104 * to share the anon_vma.
1105 *
1106 * NOTE! This runs with mm_sem held for reading, so it is possible that
1107 * the anon_vma of 'old' is concurrently in the process of being set up
1108 * by another page fault trying to merge _that_. But that's ok: if it
1109 * is being set up, that automatically means that it will be a singleton
1110 * acceptable for merging, so we can do all of this optimistically. But
1111 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1112 *
1113 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1114 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1115 * is to return an anon_vma that is "complex" due to having gone through
1116 * a fork).
1117 *
1118 * We also make sure that the two vma's are compatible (adjacent,
1119 * and with the same memory policies). That's all stable, even with just
1120 * a read lock on the mm_sem.
1121 */
1122static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1123{
1124	if (anon_vma_compatible(a, b)) {
1125		struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1126
1127		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1128			return anon_vma;
1129	}
1130	return NULL;
1131}
1132
1133/*
1134 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1135 * neighbouring vmas for a suitable anon_vma, before it goes off
1136 * to allocate a new anon_vma.  It checks because a repetitive
1137 * sequence of mprotects and faults may otherwise lead to distinct
1138 * anon_vmas being allocated, preventing vma merge in subsequent
1139 * mprotect.
1140 */
1141struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1142{
1143	struct anon_vma *anon_vma;
1144	struct vm_area_struct *near;
1145
1146	near = vma->vm_next;
1147	if (!near)
1148		goto try_prev;
1149
1150	anon_vma = reusable_anon_vma(near, vma, near);
1151	if (anon_vma)
1152		return anon_vma;
1153try_prev:
1154	near = vma->vm_prev;
1155	if (!near)
1156		goto none;
1157
1158	anon_vma = reusable_anon_vma(near, near, vma);
1159	if (anon_vma)
1160		return anon_vma;
1161none:
1162	/*
1163	 * There's no absolute need to look only at touching neighbours:
1164	 * we could search further afield for "compatible" anon_vmas.
1165	 * But it would probably just be a waste of time searching,
1166	 * or lead to too many vmas hanging off the same anon_vma.
1167	 * We're trying to allow mprotect remerging later on,
1168	 * not trying to minimize memory used for anon_vmas.
1169	 */
1170	return NULL;
1171}
1172
1173#ifdef CONFIG_PROC_FS
1174void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1175						struct file *file, long pages)
1176{
1177	const unsigned long stack_flags
1178		= VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1179
1180	mm->total_vm += pages;
1181
1182	if (file) {
1183		mm->shared_vm += pages;
1184		if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1185			mm->exec_vm += pages;
1186	} else if (flags & stack_flags)
1187		mm->stack_vm += pages;
 
 
1188}
1189#endif /* CONFIG_PROC_FS */
1190
1191/*
1192 * If a hint addr is less than mmap_min_addr change hint to be as
1193 * low as possible but still greater than mmap_min_addr
1194 */
1195static inline unsigned long round_hint_to_min(unsigned long hint)
1196{
1197	hint &= PAGE_MASK;
1198	if (((void *)hint != NULL) &&
1199	    (hint < mmap_min_addr))
1200		return PAGE_ALIGN(mmap_min_addr);
1201	return hint;
1202}
1203
1204static inline int mlock_future_check(struct mm_struct *mm,
1205				     unsigned long flags,
1206				     unsigned long len)
1207{
1208	unsigned long locked, lock_limit;
1209
1210	/*  mlock MCL_FUTURE? */
1211	if (flags & VM_LOCKED) {
1212		locked = len >> PAGE_SHIFT;
1213		locked += mm->locked_vm;
1214		lock_limit = rlimit(RLIMIT_MEMLOCK);
1215		lock_limit >>= PAGE_SHIFT;
1216		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1217			return -EAGAIN;
1218	}
1219	return 0;
1220}
1221
1222/*
1223 * The caller must hold down_write(&current->mm->mmap_sem).
1224 */
1225
1226unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1227			unsigned long len, unsigned long prot,
1228			unsigned long flags, unsigned long pgoff,
1229			unsigned long *populate)
1230{
1231	struct mm_struct * mm = current->mm;
 
1232	vm_flags_t vm_flags;
1233
1234	*populate = 0;
1235
1236	/*
1237	 * Does the application expect PROT_READ to imply PROT_EXEC?
1238	 *
1239	 * (the exception is when the underlying filesystem is noexec
1240	 *  mounted, in which case we dont add PROT_EXEC.)
1241	 */
1242	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1243		if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1244			prot |= PROT_EXEC;
1245
1246	if (!len)
1247		return -EINVAL;
1248
1249	if (!(flags & MAP_FIXED))
1250		addr = round_hint_to_min(addr);
1251
1252	/* Careful about overflows.. */
1253	len = PAGE_ALIGN(len);
1254	if (!len)
1255		return -ENOMEM;
1256
1257	/* offset overflow? */
1258	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1259               return -EOVERFLOW;
1260
1261	/* Too many mappings? */
1262	if (mm->map_count > sysctl_max_map_count)
1263		return -ENOMEM;
1264
1265	/* Obtain the address to map to. we verify (or select) it and ensure
1266	 * that it represents a valid section of the address space.
1267	 */
1268	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1269	if (addr & ~PAGE_MASK)
1270		return addr;
1271
1272	/* Do simple checking here so the lower-level routines won't have
1273	 * to. we assume access permissions have been handled by the open
1274	 * of the memory object, so we don't do any here.
1275	 */
1276	vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1277			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1278
1279	if (flags & MAP_LOCKED)
1280		if (!can_do_mlock())
1281			return -EPERM;
1282
1283	if (mlock_future_check(mm, vm_flags, len))
1284		return -EAGAIN;
 
 
 
 
 
 
 
 
 
 
1285
1286	if (file) {
1287		struct inode *inode = file_inode(file);
1288
1289		switch (flags & MAP_TYPE) {
1290		case MAP_SHARED:
1291			if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1292				return -EACCES;
1293
1294			/*
1295			 * Make sure we don't allow writing to an append-only
1296			 * file..
1297			 */
1298			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1299				return -EACCES;
1300
1301			/*
1302			 * Make sure there are no mandatory locks on the file.
1303			 */
1304			if (locks_verify_locked(file))
1305				return -EAGAIN;
1306
1307			vm_flags |= VM_SHARED | VM_MAYSHARE;
1308			if (!(file->f_mode & FMODE_WRITE))
1309				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1310
1311			/* fall through */
1312		case MAP_PRIVATE:
1313			if (!(file->f_mode & FMODE_READ))
1314				return -EACCES;
1315			if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1316				if (vm_flags & VM_EXEC)
1317					return -EPERM;
1318				vm_flags &= ~VM_MAYEXEC;
1319			}
1320
1321			if (!file->f_op->mmap)
1322				return -ENODEV;
1323			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1324				return -EINVAL;
1325			break;
1326
1327		default:
1328			return -EINVAL;
1329		}
1330	} else {
1331		switch (flags & MAP_TYPE) {
1332		case MAP_SHARED:
1333			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1334				return -EINVAL;
1335			/*
1336			 * Ignore pgoff.
1337			 */
1338			pgoff = 0;
1339			vm_flags |= VM_SHARED | VM_MAYSHARE;
1340			break;
1341		case MAP_PRIVATE:
1342			/*
1343			 * Set pgoff according to addr for anon_vma.
1344			 */
1345			pgoff = addr >> PAGE_SHIFT;
1346			break;
1347		default:
1348			return -EINVAL;
1349		}
1350	}
1351
1352	/*
1353	 * Set 'VM_NORESERVE' if we should not account for the
1354	 * memory use of this mapping.
1355	 */
1356	if (flags & MAP_NORESERVE) {
1357		/* We honor MAP_NORESERVE if allowed to overcommit */
1358		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1359			vm_flags |= VM_NORESERVE;
1360
1361		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1362		if (file && is_file_hugepages(file))
1363			vm_flags |= VM_NORESERVE;
1364	}
1365
1366	addr = mmap_region(file, addr, len, vm_flags, pgoff);
1367	if (!IS_ERR_VALUE(addr) &&
1368	    ((vm_flags & VM_LOCKED) ||
1369	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1370		*populate = len;
1371	return addr;
1372}
 
1373
1374SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1375		unsigned long, prot, unsigned long, flags,
1376		unsigned long, fd, unsigned long, pgoff)
1377{
1378	struct file *file = NULL;
1379	unsigned long retval = -EBADF;
1380
1381	if (!(flags & MAP_ANONYMOUS)) {
1382		audit_mmap_fd(fd, flags);
 
 
1383		file = fget(fd);
1384		if (!file)
1385			goto out;
1386		if (is_file_hugepages(file))
1387			len = ALIGN(len, huge_page_size(hstate_file(file)));
1388		retval = -EINVAL;
1389		if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1390			goto out_fput;
1391	} else if (flags & MAP_HUGETLB) {
1392		struct user_struct *user = NULL;
1393		struct hstate *hs;
1394
1395		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1396		if (!hs)
1397			return -EINVAL;
1398
1399		len = ALIGN(len, huge_page_size(hs));
1400		/*
1401		 * VM_NORESERVE is used because the reservations will be
1402		 * taken when vm_ops->mmap() is called
1403		 * A dummy user value is used because we are not locking
1404		 * memory so no accounting is necessary
1405		 */
1406		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1407				VM_NORESERVE,
1408				&user, HUGETLB_ANONHUGE_INODE,
1409				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1410		if (IS_ERR(file))
1411			return PTR_ERR(file);
1412	}
1413
1414	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1415
1416	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1417out_fput:
 
 
1418	if (file)
1419		fput(file);
1420out:
1421	return retval;
1422}
1423
1424#ifdef __ARCH_WANT_SYS_OLD_MMAP
1425struct mmap_arg_struct {
1426	unsigned long addr;
1427	unsigned long len;
1428	unsigned long prot;
1429	unsigned long flags;
1430	unsigned long fd;
1431	unsigned long offset;
1432};
1433
1434SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1435{
1436	struct mmap_arg_struct a;
1437
1438	if (copy_from_user(&a, arg, sizeof(a)))
1439		return -EFAULT;
1440	if (a.offset & ~PAGE_MASK)
1441		return -EINVAL;
1442
1443	return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1444			      a.offset >> PAGE_SHIFT);
1445}
1446#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1447
1448/*
1449 * Some shared mappigns will want the pages marked read-only
1450 * to track write events. If so, we'll downgrade vm_page_prot
1451 * to the private version (using protection_map[] without the
1452 * VM_SHARED bit).
1453 */
1454int vma_wants_writenotify(struct vm_area_struct *vma)
1455{
1456	vm_flags_t vm_flags = vma->vm_flags;
1457
1458	/* If it was private or non-writable, the write bit is already clear */
1459	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1460		return 0;
1461
1462	/* The backer wishes to know when pages are first written to? */
1463	if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1464		return 1;
1465
1466	/* The open routine did something to the protections already? */
1467	if (pgprot_val(vma->vm_page_prot) !=
1468	    pgprot_val(vm_get_page_prot(vm_flags)))
1469		return 0;
1470
1471	/* Specialty mapping? */
1472	if (vm_flags & VM_PFNMAP)
1473		return 0;
1474
1475	/* Can the mapping track the dirty pages? */
1476	return vma->vm_file && vma->vm_file->f_mapping &&
1477		mapping_cap_account_dirty(vma->vm_file->f_mapping);
1478}
1479
1480/*
1481 * We account for memory if it's a private writeable mapping,
1482 * not hugepages and VM_NORESERVE wasn't set.
1483 */
1484static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1485{
1486	/*
1487	 * hugetlb has its own accounting separate from the core VM
1488	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1489	 */
1490	if (file && is_file_hugepages(file))
1491		return 0;
1492
1493	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1494}
1495
1496unsigned long mmap_region(struct file *file, unsigned long addr,
1497		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
 
1498{
1499	struct mm_struct *mm = current->mm;
1500	struct vm_area_struct *vma, *prev;
 
1501	int error;
1502	struct rb_node **rb_link, *rb_parent;
1503	unsigned long charged = 0;
1504
1505	/* Check against address space limit. */
1506	if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1507		unsigned long nr_pages;
1508
1509		/*
1510		 * MAP_FIXED may remove pages of mappings that intersects with
1511		 * requested mapping. Account for the pages it would unmap.
1512		 */
1513		if (!(vm_flags & MAP_FIXED))
1514			return -ENOMEM;
1515
1516		nr_pages = count_vma_pages_range(mm, addr, addr + len);
1517
1518		if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1519			return -ENOMEM;
1520	}
1521
1522	/* Clear old maps */
1523	error = -ENOMEM;
1524munmap_back:
1525	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
 
1526		if (do_munmap(mm, addr, len))
1527			return -ENOMEM;
1528		goto munmap_back;
1529	}
1530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1531	/*
1532	 * Private writable mapping: check memory availability
1533	 */
1534	if (accountable_mapping(file, vm_flags)) {
1535		charged = len >> PAGE_SHIFT;
1536		if (security_vm_enough_memory_mm(mm, charged))
1537			return -ENOMEM;
1538		vm_flags |= VM_ACCOUNT;
1539	}
1540
1541	/*
1542	 * Can we just expand an old mapping?
1543	 */
1544	vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1545	if (vma)
1546		goto out;
1547
1548	/*
1549	 * Determine the object being mapped and call the appropriate
1550	 * specific mapper. the address has already been validated, but
1551	 * not unmapped, but the maps are removed from the list.
1552	 */
1553	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1554	if (!vma) {
1555		error = -ENOMEM;
1556		goto unacct_error;
1557	}
1558
1559	vma->vm_mm = mm;
1560	vma->vm_start = addr;
1561	vma->vm_end = addr + len;
1562	vma->vm_flags = vm_flags;
1563	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1564	vma->vm_pgoff = pgoff;
1565	INIT_LIST_HEAD(&vma->anon_vma_chain);
1566
1567	if (file) {
 
 
 
1568		if (vm_flags & VM_DENYWRITE) {
1569			error = deny_write_access(file);
1570			if (error)
1571				goto free_vma;
 
1572		}
1573		vma->vm_file = get_file(file);
 
1574		error = file->f_op->mmap(file, vma);
1575		if (error)
1576			goto unmap_and_free_vma;
 
 
1577
1578		/* Can addr have changed??
1579		 *
1580		 * Answer: Yes, several device drivers can do it in their
1581		 *         f_op->mmap method. -DaveM
1582		 * Bug: If addr is changed, prev, rb_link, rb_parent should
1583		 *      be updated for vma_link()
1584		 */
1585		WARN_ON_ONCE(addr != vma->vm_start);
1586
1587		addr = vma->vm_start;
 
1588		vm_flags = vma->vm_flags;
1589	} else if (vm_flags & VM_SHARED) {
1590		error = shmem_zero_setup(vma);
1591		if (error)
1592			goto free_vma;
1593	}
1594
1595	if (vma_wants_writenotify(vma)) {
1596		pgprot_t pprot = vma->vm_page_prot;
1597
1598		/* Can vma->vm_page_prot have changed??
1599		 *
1600		 * Answer: Yes, drivers may have changed it in their
1601		 *         f_op->mmap method.
1602		 *
1603		 * Ensures that vmas marked as uncached stay that way.
1604		 */
1605		vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1606		if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1607			vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1608	}
1609
1610	vma_link(mm, vma, prev, rb_link, rb_parent);
 
 
1611	/* Once vma denies write, undo our temporary denial count */
1612	if (vm_flags & VM_DENYWRITE)
1613		allow_write_access(file);
1614	file = vma->vm_file;
1615out:
1616	perf_event_mmap(vma);
1617
 
1618	vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1619	if (vm_flags & VM_LOCKED) {
1620		if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1621					vma == get_gate_vma(current->mm)))
1622			mm->locked_vm += (len >> PAGE_SHIFT);
1623		else
1624			vma->vm_flags &= ~VM_LOCKED;
1625	}
1626
1627	if (file)
1628		uprobe_mmap(vma);
1629
1630	/*
1631	 * New (or expanded) vma always get soft dirty status.
1632	 * Otherwise user-space soft-dirty page tracker won't
1633	 * be able to distinguish situation when vma area unmapped,
1634	 * then new mapped in-place (which must be aimed as
1635	 * a completely new data area).
1636	 */
1637	vma->vm_flags |= VM_SOFTDIRTY;
1638
1639	return addr;
1640
1641unmap_and_free_vma:
1642	if (vm_flags & VM_DENYWRITE)
1643		allow_write_access(file);
1644	vma->vm_file = NULL;
1645	fput(file);
1646
1647	/* Undo any partial mapping done by a device driver. */
1648	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1649	charged = 0;
1650free_vma:
1651	kmem_cache_free(vm_area_cachep, vma);
1652unacct_error:
1653	if (charged)
1654		vm_unacct_memory(charged);
1655	return error;
1656}
1657
1658unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1659{
1660	/*
1661	 * We implement the search by looking for an rbtree node that
1662	 * immediately follows a suitable gap. That is,
1663	 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1664	 * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1665	 * - gap_end - gap_start >= length
1666	 */
1667
1668	struct mm_struct *mm = current->mm;
1669	struct vm_area_struct *vma;
1670	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1671
1672	/* Adjust search length to account for worst case alignment overhead */
1673	length = info->length + info->align_mask;
1674	if (length < info->length)
1675		return -ENOMEM;
1676
1677	/* Adjust search limits by the desired length */
1678	if (info->high_limit < length)
1679		return -ENOMEM;
1680	high_limit = info->high_limit - length;
1681
1682	if (info->low_limit > high_limit)
1683		return -ENOMEM;
1684	low_limit = info->low_limit + length;
1685
1686	/* Check if rbtree root looks promising */
1687	if (RB_EMPTY_ROOT(&mm->mm_rb))
1688		goto check_highest;
1689	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1690	if (vma->rb_subtree_gap < length)
1691		goto check_highest;
1692
1693	while (true) {
1694		/* Visit left subtree if it looks promising */
1695		gap_end = vma->vm_start;
1696		if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1697			struct vm_area_struct *left =
1698				rb_entry(vma->vm_rb.rb_left,
1699					 struct vm_area_struct, vm_rb);
1700			if (left->rb_subtree_gap >= length) {
1701				vma = left;
1702				continue;
1703			}
1704		}
1705
1706		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1707check_current:
1708		/* Check if current node has a suitable gap */
1709		if (gap_start > high_limit)
1710			return -ENOMEM;
1711		if (gap_end >= low_limit && gap_end - gap_start >= length)
1712			goto found;
1713
1714		/* Visit right subtree if it looks promising */
1715		if (vma->vm_rb.rb_right) {
1716			struct vm_area_struct *right =
1717				rb_entry(vma->vm_rb.rb_right,
1718					 struct vm_area_struct, vm_rb);
1719			if (right->rb_subtree_gap >= length) {
1720				vma = right;
1721				continue;
1722			}
1723		}
1724
1725		/* Go back up the rbtree to find next candidate node */
1726		while (true) {
1727			struct rb_node *prev = &vma->vm_rb;
1728			if (!rb_parent(prev))
1729				goto check_highest;
1730			vma = rb_entry(rb_parent(prev),
1731				       struct vm_area_struct, vm_rb);
1732			if (prev == vma->vm_rb.rb_left) {
1733				gap_start = vma->vm_prev->vm_end;
1734				gap_end = vma->vm_start;
1735				goto check_current;
1736			}
1737		}
1738	}
1739
1740check_highest:
1741	/* Check highest gap, which does not precede any rbtree node */
1742	gap_start = mm->highest_vm_end;
1743	gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1744	if (gap_start > high_limit)
1745		return -ENOMEM;
1746
1747found:
1748	/* We found a suitable gap. Clip it with the original low_limit. */
1749	if (gap_start < info->low_limit)
1750		gap_start = info->low_limit;
1751
1752	/* Adjust gap address to the desired alignment */
1753	gap_start += (info->align_offset - gap_start) & info->align_mask;
1754
1755	VM_BUG_ON(gap_start + info->length > info->high_limit);
1756	VM_BUG_ON(gap_start + info->length > gap_end);
1757	return gap_start;
1758}
1759
1760unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1761{
1762	struct mm_struct *mm = current->mm;
1763	struct vm_area_struct *vma;
1764	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1765
1766	/* Adjust search length to account for worst case alignment overhead */
1767	length = info->length + info->align_mask;
1768	if (length < info->length)
1769		return -ENOMEM;
1770
1771	/*
1772	 * Adjust search limits by the desired length.
1773	 * See implementation comment at top of unmapped_area().
1774	 */
1775	gap_end = info->high_limit;
1776	if (gap_end < length)
1777		return -ENOMEM;
1778	high_limit = gap_end - length;
1779
1780	if (info->low_limit > high_limit)
1781		return -ENOMEM;
1782	low_limit = info->low_limit + length;
1783
1784	/* Check highest gap, which does not precede any rbtree node */
1785	gap_start = mm->highest_vm_end;
1786	if (gap_start <= high_limit)
1787		goto found_highest;
1788
1789	/* Check if rbtree root looks promising */
1790	if (RB_EMPTY_ROOT(&mm->mm_rb))
1791		return -ENOMEM;
1792	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1793	if (vma->rb_subtree_gap < length)
1794		return -ENOMEM;
1795
1796	while (true) {
1797		/* Visit right subtree if it looks promising */
1798		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1799		if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1800			struct vm_area_struct *right =
1801				rb_entry(vma->vm_rb.rb_right,
1802					 struct vm_area_struct, vm_rb);
1803			if (right->rb_subtree_gap >= length) {
1804				vma = right;
1805				continue;
1806			}
1807		}
1808
1809check_current:
1810		/* Check if current node has a suitable gap */
1811		gap_end = vma->vm_start;
1812		if (gap_end < low_limit)
1813			return -ENOMEM;
1814		if (gap_start <= high_limit && gap_end - gap_start >= length)
1815			goto found;
1816
1817		/* Visit left subtree if it looks promising */
1818		if (vma->vm_rb.rb_left) {
1819			struct vm_area_struct *left =
1820				rb_entry(vma->vm_rb.rb_left,
1821					 struct vm_area_struct, vm_rb);
1822			if (left->rb_subtree_gap >= length) {
1823				vma = left;
1824				continue;
1825			}
1826		}
1827
1828		/* Go back up the rbtree to find next candidate node */
1829		while (true) {
1830			struct rb_node *prev = &vma->vm_rb;
1831			if (!rb_parent(prev))
1832				return -ENOMEM;
1833			vma = rb_entry(rb_parent(prev),
1834				       struct vm_area_struct, vm_rb);
1835			if (prev == vma->vm_rb.rb_right) {
1836				gap_start = vma->vm_prev ?
1837					vma->vm_prev->vm_end : 0;
1838				goto check_current;
1839			}
1840		}
1841	}
1842
1843found:
1844	/* We found a suitable gap. Clip it with the original high_limit. */
1845	if (gap_end > info->high_limit)
1846		gap_end = info->high_limit;
1847
1848found_highest:
1849	/* Compute highest gap address at the desired alignment */
1850	gap_end -= info->length;
1851	gap_end -= (gap_end - info->align_offset) & info->align_mask;
1852
1853	VM_BUG_ON(gap_end < info->low_limit);
1854	VM_BUG_ON(gap_end < gap_start);
1855	return gap_end;
1856}
1857
1858/* Get an address range which is currently unmapped.
1859 * For shmat() with addr=0.
1860 *
1861 * Ugly calling convention alert:
1862 * Return value with the low bits set means error value,
1863 * ie
1864 *	if (ret & ~PAGE_MASK)
1865 *		error = ret;
1866 *
1867 * This function "knows" that -ENOMEM has the bits set.
1868 */
1869#ifndef HAVE_ARCH_UNMAPPED_AREA
1870unsigned long
1871arch_get_unmapped_area(struct file *filp, unsigned long addr,
1872		unsigned long len, unsigned long pgoff, unsigned long flags)
1873{
1874	struct mm_struct *mm = current->mm;
1875	struct vm_area_struct *vma;
1876	struct vm_unmapped_area_info info;
1877
1878	if (len > TASK_SIZE - mmap_min_addr)
1879		return -ENOMEM;
1880
1881	if (flags & MAP_FIXED)
1882		return addr;
1883
1884	if (addr) {
1885		addr = PAGE_ALIGN(addr);
1886		vma = find_vma(mm, addr);
1887		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1888		    (!vma || addr + len <= vma->vm_start))
1889			return addr;
1890	}
 
 
 
 
 
 
1891
1892	info.flags = 0;
1893	info.length = len;
1894	info.low_limit = mm->mmap_base;
1895	info.high_limit = TASK_SIZE;
1896	info.align_mask = 0;
1897	return vm_unmapped_area(&info);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1898}
1899#endif	
1900
 
 
 
 
 
 
 
 
 
 
 
1901/*
1902 * This mmap-allocator allocates new areas top-down from below the
1903 * stack's low limit (the base):
1904 */
1905#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1906unsigned long
1907arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1908			  const unsigned long len, const unsigned long pgoff,
1909			  const unsigned long flags)
1910{
1911	struct vm_area_struct *vma;
1912	struct mm_struct *mm = current->mm;
1913	unsigned long addr = addr0;
1914	struct vm_unmapped_area_info info;
1915
1916	/* requested length too big for entire address space */
1917	if (len > TASK_SIZE - mmap_min_addr)
1918		return -ENOMEM;
1919
1920	if (flags & MAP_FIXED)
1921		return addr;
1922
1923	/* requesting a specific address */
1924	if (addr) {
1925		addr = PAGE_ALIGN(addr);
1926		vma = find_vma(mm, addr);
1927		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1928				(!vma || addr + len <= vma->vm_start))
1929			return addr;
1930	}
1931
1932	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1933	info.length = len;
1934	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1935	info.high_limit = mm->mmap_base;
1936	info.align_mask = 0;
1937	addr = vm_unmapped_area(&info);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1938
 
1939	/*
1940	 * A failed mmap() very likely causes application failure,
1941	 * so fall back to the bottom-up function here. This scenario
1942	 * can happen with large stack limits and large mmap()
1943	 * allocations.
1944	 */
1945	if (addr & ~PAGE_MASK) {
1946		VM_BUG_ON(addr != -ENOMEM);
1947		info.flags = 0;
1948		info.low_limit = TASK_UNMAPPED_BASE;
1949		info.high_limit = TASK_SIZE;
1950		addr = vm_unmapped_area(&info);
1951	}
 
1952
1953	return addr;
1954}
1955#endif
1956
 
 
 
 
 
 
 
 
 
 
 
 
 
1957unsigned long
1958get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1959		unsigned long pgoff, unsigned long flags)
1960{
1961	unsigned long (*get_area)(struct file *, unsigned long,
1962				  unsigned long, unsigned long, unsigned long);
1963
1964	unsigned long error = arch_mmap_check(addr, len, flags);
1965	if (error)
1966		return error;
1967
1968	/* Careful about overflows.. */
1969	if (len > TASK_SIZE)
1970		return -ENOMEM;
1971
1972	get_area = current->mm->get_unmapped_area;
1973	if (file && file->f_op->get_unmapped_area)
1974		get_area = file->f_op->get_unmapped_area;
1975	addr = get_area(file, addr, len, pgoff, flags);
1976	if (IS_ERR_VALUE(addr))
1977		return addr;
1978
1979	if (addr > TASK_SIZE - len)
1980		return -ENOMEM;
1981	if (addr & ~PAGE_MASK)
1982		return -EINVAL;
1983
1984	addr = arch_rebalance_pgtables(addr, len);
1985	error = security_mmap_addr(addr);
1986	return error ? error : addr;
1987}
1988
1989EXPORT_SYMBOL(get_unmapped_area);
1990
1991/* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1992struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1993{
1994	struct rb_node *rb_node;
1995	struct vm_area_struct *vma;
1996
1997	/* Check the cache first. */
1998	vma = vmacache_find(mm, addr);
1999	if (likely(vma))
2000		return vma;
2001
2002	rb_node = mm->mm_rb.rb_node;
2003	vma = NULL;
2004
2005	while (rb_node) {
2006		struct vm_area_struct *tmp;
2007
2008		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2009
2010		if (tmp->vm_end > addr) {
2011			vma = tmp;
2012			if (tmp->vm_start <= addr)
2013				break;
2014			rb_node = rb_node->rb_left;
2015		} else
2016			rb_node = rb_node->rb_right;
 
 
 
 
 
 
 
2017	}
2018
2019	if (vma)
2020		vmacache_update(addr, vma);
2021	return vma;
2022}
2023
2024EXPORT_SYMBOL(find_vma);
2025
2026/*
2027 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2028 */
2029struct vm_area_struct *
2030find_vma_prev(struct mm_struct *mm, unsigned long addr,
2031			struct vm_area_struct **pprev)
2032{
2033	struct vm_area_struct *vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
2034
2035	vma = find_vma(mm, addr);
2036	if (vma) {
2037		*pprev = vma->vm_prev;
2038	} else {
2039		struct rb_node *rb_node = mm->mm_rb.rb_node;
2040		*pprev = NULL;
2041		while (rb_node) {
2042			*pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2043			rb_node = rb_node->rb_right;
2044		}
2045	}
2046	return vma;
 
 
 
2047}
2048
2049/*
2050 * Verify that the stack growth is acceptable and
2051 * update accounting. This is shared with both the
2052 * grow-up and grow-down cases.
2053 */
2054static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2055{
2056	struct mm_struct *mm = vma->vm_mm;
2057	struct rlimit *rlim = current->signal->rlim;
2058	unsigned long new_start;
2059
2060	/* address space limit tests */
2061	if (!may_expand_vm(mm, grow))
2062		return -ENOMEM;
2063
2064	/* Stack limit test */
2065	if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2066		return -ENOMEM;
2067
2068	/* mlock limit tests */
2069	if (vma->vm_flags & VM_LOCKED) {
2070		unsigned long locked;
2071		unsigned long limit;
2072		locked = mm->locked_vm + grow;
2073		limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2074		limit >>= PAGE_SHIFT;
2075		if (locked > limit && !capable(CAP_IPC_LOCK))
2076			return -ENOMEM;
2077	}
2078
2079	/* Check to ensure the stack will not grow into a hugetlb-only region */
2080	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2081			vma->vm_end - size;
2082	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2083		return -EFAULT;
2084
2085	/*
2086	 * Overcommit..  This must be the final test, as it will
2087	 * update security statistics.
2088	 */
2089	if (security_vm_enough_memory_mm(mm, grow))
2090		return -ENOMEM;
2091
2092	/* Ok, everything looks good - let it rip */
 
2093	if (vma->vm_flags & VM_LOCKED)
2094		mm->locked_vm += grow;
2095	vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2096	return 0;
2097}
2098
2099#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2100/*
2101 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2102 * vma is the last one with address > vma->vm_end.  Have to extend vma.
2103 */
2104int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2105{
2106	int error;
2107
2108	if (!(vma->vm_flags & VM_GROWSUP))
2109		return -EFAULT;
2110
2111	/*
2112	 * We must make sure the anon_vma is allocated
2113	 * so that the anon_vma locking is not a noop.
2114	 */
2115	if (unlikely(anon_vma_prepare(vma)))
2116		return -ENOMEM;
2117	vma_lock_anon_vma(vma);
2118
2119	/*
2120	 * vma->vm_start/vm_end cannot change under us because the caller
2121	 * is required to hold the mmap_sem in read mode.  We need the
2122	 * anon_vma lock to serialize against concurrent expand_stacks.
2123	 * Also guard against wrapping around to address 0.
2124	 */
2125	if (address < PAGE_ALIGN(address+4))
2126		address = PAGE_ALIGN(address+4);
2127	else {
2128		vma_unlock_anon_vma(vma);
2129		return -ENOMEM;
2130	}
2131	error = 0;
2132
2133	/* Somebody else might have raced and expanded it already */
2134	if (address > vma->vm_end) {
2135		unsigned long size, grow;
2136
2137		size = address - vma->vm_start;
2138		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2139
2140		error = -ENOMEM;
2141		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2142			error = acct_stack_growth(vma, size, grow);
2143			if (!error) {
2144				/*
2145				 * vma_gap_update() doesn't support concurrent
2146				 * updates, but we only hold a shared mmap_sem
2147				 * lock here, so we need to protect against
2148				 * concurrent vma expansions.
2149				 * vma_lock_anon_vma() doesn't help here, as
2150				 * we don't guarantee that all growable vmas
2151				 * in a mm share the same root anon vma.
2152				 * So, we reuse mm->page_table_lock to guard
2153				 * against concurrent vma expansions.
2154				 */
2155				spin_lock(&vma->vm_mm->page_table_lock);
2156				anon_vma_interval_tree_pre_update_vma(vma);
2157				vma->vm_end = address;
2158				anon_vma_interval_tree_post_update_vma(vma);
2159				if (vma->vm_next)
2160					vma_gap_update(vma->vm_next);
2161				else
2162					vma->vm_mm->highest_vm_end = address;
2163				spin_unlock(&vma->vm_mm->page_table_lock);
2164
2165				perf_event_mmap(vma);
2166			}
2167		}
2168	}
2169	vma_unlock_anon_vma(vma);
2170	khugepaged_enter_vma_merge(vma);
2171	validate_mm(vma->vm_mm);
2172	return error;
2173}
2174#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2175
2176/*
2177 * vma is the first one with address < vma->vm_start.  Have to extend vma.
2178 */
2179int expand_downwards(struct vm_area_struct *vma,
2180				   unsigned long address)
2181{
2182	int error;
2183
2184	/*
2185	 * We must make sure the anon_vma is allocated
2186	 * so that the anon_vma locking is not a noop.
2187	 */
2188	if (unlikely(anon_vma_prepare(vma)))
2189		return -ENOMEM;
2190
2191	address &= PAGE_MASK;
2192	error = security_mmap_addr(address);
2193	if (error)
2194		return error;
2195
2196	vma_lock_anon_vma(vma);
2197
2198	/*
2199	 * vma->vm_start/vm_end cannot change under us because the caller
2200	 * is required to hold the mmap_sem in read mode.  We need the
2201	 * anon_vma lock to serialize against concurrent expand_stacks.
2202	 */
2203
2204	/* Somebody else might have raced and expanded it already */
2205	if (address < vma->vm_start) {
2206		unsigned long size, grow;
2207
2208		size = vma->vm_end - address;
2209		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2210
2211		error = -ENOMEM;
2212		if (grow <= vma->vm_pgoff) {
2213			error = acct_stack_growth(vma, size, grow);
2214			if (!error) {
2215				/*
2216				 * vma_gap_update() doesn't support concurrent
2217				 * updates, but we only hold a shared mmap_sem
2218				 * lock here, so we need to protect against
2219				 * concurrent vma expansions.
2220				 * vma_lock_anon_vma() doesn't help here, as
2221				 * we don't guarantee that all growable vmas
2222				 * in a mm share the same root anon vma.
2223				 * So, we reuse mm->page_table_lock to guard
2224				 * against concurrent vma expansions.
2225				 */
2226				spin_lock(&vma->vm_mm->page_table_lock);
2227				anon_vma_interval_tree_pre_update_vma(vma);
2228				vma->vm_start = address;
2229				vma->vm_pgoff -= grow;
2230				anon_vma_interval_tree_post_update_vma(vma);
2231				vma_gap_update(vma);
2232				spin_unlock(&vma->vm_mm->page_table_lock);
2233
2234				perf_event_mmap(vma);
2235			}
2236		}
2237	}
2238	vma_unlock_anon_vma(vma);
2239	khugepaged_enter_vma_merge(vma);
2240	validate_mm(vma->vm_mm);
2241	return error;
2242}
2243
2244/*
2245 * Note how expand_stack() refuses to expand the stack all the way to
2246 * abut the next virtual mapping, *unless* that mapping itself is also
2247 * a stack mapping. We want to leave room for a guard page, after all
2248 * (the guard page itself is not added here, that is done by the
2249 * actual page faulting logic)
2250 *
2251 * This matches the behavior of the guard page logic (see mm/memory.c:
2252 * check_stack_guard_page()), which only allows the guard page to be
2253 * removed under these circumstances.
2254 */
2255#ifdef CONFIG_STACK_GROWSUP
2256int expand_stack(struct vm_area_struct *vma, unsigned long address)
2257{
2258	struct vm_area_struct *next;
2259
2260	address &= PAGE_MASK;
2261	next = vma->vm_next;
2262	if (next && next->vm_start == address + PAGE_SIZE) {
2263		if (!(next->vm_flags & VM_GROWSUP))
2264			return -ENOMEM;
2265	}
2266	return expand_upwards(vma, address);
2267}
2268
2269struct vm_area_struct *
2270find_extend_vma(struct mm_struct *mm, unsigned long addr)
2271{
2272	struct vm_area_struct *vma, *prev;
2273
2274	addr &= PAGE_MASK;
2275	vma = find_vma_prev(mm, addr, &prev);
2276	if (vma && (vma->vm_start <= addr))
2277		return vma;
2278	if (!prev || expand_stack(prev, addr))
2279		return NULL;
2280	if (prev->vm_flags & VM_LOCKED)
2281		__mlock_vma_pages_range(prev, addr, prev->vm_end, NULL);
 
2282	return prev;
2283}
2284#else
2285int expand_stack(struct vm_area_struct *vma, unsigned long address)
2286{
2287	struct vm_area_struct *prev;
2288
2289	address &= PAGE_MASK;
2290	prev = vma->vm_prev;
2291	if (prev && prev->vm_end == address) {
2292		if (!(prev->vm_flags & VM_GROWSDOWN))
2293			return -ENOMEM;
2294	}
2295	return expand_downwards(vma, address);
2296}
2297
2298struct vm_area_struct *
2299find_extend_vma(struct mm_struct * mm, unsigned long addr)
2300{
2301	struct vm_area_struct * vma;
2302	unsigned long start;
2303
2304	addr &= PAGE_MASK;
2305	vma = find_vma(mm,addr);
2306	if (!vma)
2307		return NULL;
2308	if (vma->vm_start <= addr)
2309		return vma;
2310	if (!(vma->vm_flags & VM_GROWSDOWN))
2311		return NULL;
2312	start = vma->vm_start;
2313	if (expand_stack(vma, addr))
2314		return NULL;
2315	if (vma->vm_flags & VM_LOCKED)
2316		__mlock_vma_pages_range(vma, addr, start, NULL);
 
2317	return vma;
2318}
2319#endif
2320
2321/*
2322 * Ok - we have the memory areas we should free on the vma list,
2323 * so release them, and do the vma updates.
2324 *
2325 * Called with the mm semaphore held.
2326 */
2327static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2328{
2329	unsigned long nr_accounted = 0;
2330
2331	/* Update high watermark before we lower total_vm */
2332	update_hiwater_vm(mm);
2333	do {
2334		long nrpages = vma_pages(vma);
2335
2336		if (vma->vm_flags & VM_ACCOUNT)
2337			nr_accounted += nrpages;
2338		vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2339		vma = remove_vma(vma);
2340	} while (vma);
2341	vm_unacct_memory(nr_accounted);
2342	validate_mm(mm);
2343}
2344
2345/*
2346 * Get rid of page table information in the indicated region.
2347 *
2348 * Called with the mm semaphore held.
2349 */
2350static void unmap_region(struct mm_struct *mm,
2351		struct vm_area_struct *vma, struct vm_area_struct *prev,
2352		unsigned long start, unsigned long end)
2353{
2354	struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2355	struct mmu_gather tlb;
 
2356
2357	lru_add_drain();
2358	tlb_gather_mmu(&tlb, mm, start, end);
2359	update_hiwater_rss(mm);
2360	unmap_vmas(&tlb, vma, start, end);
 
2361	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2362				 next ? next->vm_start : USER_PGTABLES_CEILING);
2363	tlb_finish_mmu(&tlb, start, end);
2364}
2365
2366/*
2367 * Create a list of vma's touched by the unmap, removing them from the mm's
2368 * vma list as we go..
2369 */
2370static void
2371detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2372	struct vm_area_struct *prev, unsigned long end)
2373{
2374	struct vm_area_struct **insertion_point;
2375	struct vm_area_struct *tail_vma = NULL;
 
2376
2377	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2378	vma->vm_prev = NULL;
2379	do {
2380		vma_rb_erase(vma, &mm->mm_rb);
2381		mm->map_count--;
2382		tail_vma = vma;
2383		vma = vma->vm_next;
2384	} while (vma && vma->vm_start < end);
2385	*insertion_point = vma;
2386	if (vma) {
2387		vma->vm_prev = prev;
2388		vma_gap_update(vma);
2389	} else
2390		mm->highest_vm_end = prev ? prev->vm_end : 0;
2391	tail_vma->vm_next = NULL;
2392
2393	/* Kill the cache */
2394	vmacache_invalidate(mm);
 
 
 
2395}
2396
2397/*
2398 * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2399 * munmap path where it doesn't make sense to fail.
2400 */
2401static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2402	      unsigned long addr, int new_below)
2403{
 
2404	struct vm_area_struct *new;
2405	int err = -ENOMEM;
2406
2407	if (is_vm_hugetlb_page(vma) && (addr &
2408					~(huge_page_mask(hstate_vma(vma)))))
2409		return -EINVAL;
2410
2411	new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2412	if (!new)
2413		goto out_err;
2414
2415	/* most fields are the same, copy all, and then fixup */
2416	*new = *vma;
2417
2418	INIT_LIST_HEAD(&new->anon_vma_chain);
2419
2420	if (new_below)
2421		new->vm_end = addr;
2422	else {
2423		new->vm_start = addr;
2424		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2425	}
2426
2427	err = vma_dup_policy(vma, new);
2428	if (err)
 
2429		goto out_free_vma;
 
 
2430
2431	if (anon_vma_clone(new, vma))
2432		goto out_free_mpol;
2433
2434	if (new->vm_file)
2435		get_file(new->vm_file);
 
 
 
2436
2437	if (new->vm_ops && new->vm_ops->open)
2438		new->vm_ops->open(new);
2439
2440	if (new_below)
2441		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2442			((addr - new->vm_start) >> PAGE_SHIFT), new);
2443	else
2444		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2445
2446	/* Success. */
2447	if (!err)
2448		return 0;
2449
2450	/* Clean everything up if vma_adjust failed. */
2451	if (new->vm_ops && new->vm_ops->close)
2452		new->vm_ops->close(new);
2453	if (new->vm_file)
 
 
2454		fput(new->vm_file);
 
2455	unlink_anon_vmas(new);
2456 out_free_mpol:
2457	mpol_put(vma_policy(new));
2458 out_free_vma:
2459	kmem_cache_free(vm_area_cachep, new);
2460 out_err:
2461	return err;
2462}
2463
2464/*
2465 * Split a vma into two pieces at address 'addr', a new vma is allocated
2466 * either for the first part or the tail.
2467 */
2468int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2469	      unsigned long addr, int new_below)
2470{
2471	if (mm->map_count >= sysctl_max_map_count)
2472		return -ENOMEM;
2473
2474	return __split_vma(mm, vma, addr, new_below);
2475}
2476
2477/* Munmap is split into 2 main parts -- this part which finds
2478 * what needs doing, and the areas themselves, which do the
2479 * work.  This now handles partial unmappings.
2480 * Jeremy Fitzhardinge <jeremy@goop.org>
2481 */
2482int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2483{
2484	unsigned long end;
2485	struct vm_area_struct *vma, *prev, *last;
2486
2487	if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2488		return -EINVAL;
2489
2490	if ((len = PAGE_ALIGN(len)) == 0)
2491		return -EINVAL;
2492
2493	/* Find the first overlapping VMA */
2494	vma = find_vma(mm, start);
2495	if (!vma)
2496		return 0;
2497	prev = vma->vm_prev;
2498	/* we have  start < vma->vm_end  */
2499
2500	/* if it doesn't overlap, we have nothing.. */
2501	end = start + len;
2502	if (vma->vm_start >= end)
2503		return 0;
2504
2505	/*
2506	 * If we need to split any vma, do it now to save pain later.
2507	 *
2508	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2509	 * unmapped vm_area_struct will remain in use: so lower split_vma
2510	 * places tmp vma above, and higher split_vma places tmp vma below.
2511	 */
2512	if (start > vma->vm_start) {
2513		int error;
2514
2515		/*
2516		 * Make sure that map_count on return from munmap() will
2517		 * not exceed its limit; but let map_count go just above
2518		 * its limit temporarily, to help free resources as expected.
2519		 */
2520		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2521			return -ENOMEM;
2522
2523		error = __split_vma(mm, vma, start, 0);
2524		if (error)
2525			return error;
2526		prev = vma;
2527	}
2528
2529	/* Does it split the last one? */
2530	last = find_vma(mm, end);
2531	if (last && end > last->vm_start) {
2532		int error = __split_vma(mm, last, end, 1);
2533		if (error)
2534			return error;
2535	}
2536	vma = prev? prev->vm_next: mm->mmap;
2537
2538	/*
2539	 * unlock any mlock()ed ranges before detaching vmas
2540	 */
2541	if (mm->locked_vm) {
2542		struct vm_area_struct *tmp = vma;
2543		while (tmp && tmp->vm_start < end) {
2544			if (tmp->vm_flags & VM_LOCKED) {
2545				mm->locked_vm -= vma_pages(tmp);
2546				munlock_vma_pages_all(tmp);
2547			}
2548			tmp = tmp->vm_next;
2549		}
2550	}
2551
2552	/*
2553	 * Remove the vma's, and unmap the actual pages
2554	 */
2555	detach_vmas_to_be_unmapped(mm, vma, prev, end);
2556	unmap_region(mm, vma, prev, start, end);
2557
2558	/* Fix up all other VM information */
2559	remove_vma_list(mm, vma);
2560
2561	return 0;
2562}
2563
2564int vm_munmap(unsigned long start, size_t len)
 
 
2565{
2566	int ret;
2567	struct mm_struct *mm = current->mm;
2568
 
 
2569	down_write(&mm->mmap_sem);
2570	ret = do_munmap(mm, start, len);
2571	up_write(&mm->mmap_sem);
2572	return ret;
2573}
2574EXPORT_SYMBOL(vm_munmap);
2575
2576SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2577{
2578	profile_munmap(addr);
2579	return vm_munmap(addr, len);
2580}
2581
2582static inline void verify_mm_writelocked(struct mm_struct *mm)
2583{
2584#ifdef CONFIG_DEBUG_VM
2585	if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2586		WARN_ON(1);
2587		up_read(&mm->mmap_sem);
2588	}
2589#endif
2590}
2591
2592/*
2593 *  this is really a simplified "do_mmap".  it only handles
2594 *  anonymous maps.  eventually we may be able to do some
2595 *  brk-specific accounting here.
2596 */
2597static unsigned long do_brk(unsigned long addr, unsigned long len)
2598{
2599	struct mm_struct * mm = current->mm;
2600	struct vm_area_struct * vma, * prev;
2601	unsigned long flags;
2602	struct rb_node ** rb_link, * rb_parent;
2603	pgoff_t pgoff = addr >> PAGE_SHIFT;
2604	int error;
2605
2606	len = PAGE_ALIGN(len);
2607	if (!len)
2608		return addr;
2609
 
 
 
 
2610	flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2611
2612	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2613	if (error & ~PAGE_MASK)
2614		return error;
2615
2616	error = mlock_future_check(mm, mm->def_flags, len);
2617	if (error)
2618		return error;
 
 
 
 
 
 
 
 
 
2619
2620	/*
2621	 * mm->mmap_sem is required to protect against another thread
2622	 * changing the mappings in case we sleep.
2623	 */
2624	verify_mm_writelocked(mm);
2625
2626	/*
2627	 * Clear old maps.  this also does some error checking for us
2628	 */
2629 munmap_back:
2630	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
 
2631		if (do_munmap(mm, addr, len))
2632			return -ENOMEM;
2633		goto munmap_back;
2634	}
2635
2636	/* Check against address space limits *after* clearing old maps... */
2637	if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2638		return -ENOMEM;
2639
2640	if (mm->map_count > sysctl_max_map_count)
2641		return -ENOMEM;
2642
2643	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2644		return -ENOMEM;
2645
2646	/* Can we just expand an old private anonymous mapping? */
2647	vma = vma_merge(mm, prev, addr, addr + len, flags,
2648					NULL, NULL, pgoff, NULL);
2649	if (vma)
2650		goto out;
2651
2652	/*
2653	 * create a vma struct for an anonymous mapping
2654	 */
2655	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2656	if (!vma) {
2657		vm_unacct_memory(len >> PAGE_SHIFT);
2658		return -ENOMEM;
2659	}
2660
2661	INIT_LIST_HEAD(&vma->anon_vma_chain);
2662	vma->vm_mm = mm;
2663	vma->vm_start = addr;
2664	vma->vm_end = addr + len;
2665	vma->vm_pgoff = pgoff;
2666	vma->vm_flags = flags;
2667	vma->vm_page_prot = vm_get_page_prot(flags);
2668	vma_link(mm, vma, prev, rb_link, rb_parent);
2669out:
2670	perf_event_mmap(vma);
2671	mm->total_vm += len >> PAGE_SHIFT;
2672	if (flags & VM_LOCKED)
2673		mm->locked_vm += (len >> PAGE_SHIFT);
2674	vma->vm_flags |= VM_SOFTDIRTY;
 
2675	return addr;
2676}
2677
2678unsigned long vm_brk(unsigned long addr, unsigned long len)
2679{
2680	struct mm_struct *mm = current->mm;
2681	unsigned long ret;
2682	bool populate;
2683
2684	down_write(&mm->mmap_sem);
2685	ret = do_brk(addr, len);
2686	populate = ((mm->def_flags & VM_LOCKED) != 0);
2687	up_write(&mm->mmap_sem);
2688	if (populate)
2689		mm_populate(addr, len);
2690	return ret;
2691}
2692EXPORT_SYMBOL(vm_brk);
2693
2694/* Release all mmaps. */
2695void exit_mmap(struct mm_struct *mm)
2696{
2697	struct mmu_gather tlb;
2698	struct vm_area_struct *vma;
2699	unsigned long nr_accounted = 0;
 
2700
2701	/* mm's last user has gone, and its about to be pulled down */
2702	mmu_notifier_release(mm);
2703
2704	if (mm->locked_vm) {
2705		vma = mm->mmap;
2706		while (vma) {
2707			if (vma->vm_flags & VM_LOCKED)
2708				munlock_vma_pages_all(vma);
2709			vma = vma->vm_next;
2710		}
2711	}
2712
2713	arch_exit_mmap(mm);
2714
2715	vma = mm->mmap;
2716	if (!vma)	/* Can happen if dup_mmap() received an OOM */
2717		return;
2718
2719	lru_add_drain();
2720	flush_cache_mm(mm);
2721	tlb_gather_mmu(&tlb, mm, 0, -1);
2722	/* update_hiwater_rss(mm) here? but nobody should be looking */
2723	/* Use -1 here to ensure all VMAs in the mm are unmapped */
2724	unmap_vmas(&tlb, vma, 0, -1);
 
2725
2726	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2727	tlb_finish_mmu(&tlb, 0, -1);
2728
2729	/*
2730	 * Walk the list again, actually closing and freeing it,
2731	 * with preemption enabled, without holding any MM locks.
2732	 */
2733	while (vma) {
2734		if (vma->vm_flags & VM_ACCOUNT)
2735			nr_accounted += vma_pages(vma);
2736		vma = remove_vma(vma);
2737	}
2738	vm_unacct_memory(nr_accounted);
2739
2740	WARN_ON(atomic_long_read(&mm->nr_ptes) >
2741			(FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2742}
2743
2744/* Insert vm structure into process list sorted by address
2745 * and into the inode's i_mmap tree.  If vm_file is non-NULL
2746 * then i_mmap_mutex is taken here.
2747 */
2748int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2749{
2750	struct vm_area_struct *prev;
2751	struct rb_node **rb_link, *rb_parent;
2752
2753	/*
2754	 * The vm_pgoff of a purely anonymous vma should be irrelevant
2755	 * until its first write fault, when page's anon_vma and index
2756	 * are set.  But now set the vm_pgoff it will almost certainly
2757	 * end up with (unless mremap moves it elsewhere before that
2758	 * first wfault), so /proc/pid/maps tells a consistent story.
2759	 *
2760	 * By setting it to reflect the virtual start address of the
2761	 * vma, merges and splits can happen in a seamless way, just
2762	 * using the existing file pgoff checks and manipulations.
2763	 * Similarly in do_mmap_pgoff and in do_brk.
2764	 */
2765	if (!vma->vm_file) {
2766		BUG_ON(vma->anon_vma);
2767		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2768	}
2769	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2770			   &prev, &rb_link, &rb_parent))
2771		return -ENOMEM;
2772	if ((vma->vm_flags & VM_ACCOUNT) &&
2773	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
2774		return -ENOMEM;
2775
2776	vma_link(mm, vma, prev, rb_link, rb_parent);
2777	return 0;
2778}
2779
2780/*
2781 * Copy the vma structure to a new location in the same mm,
2782 * prior to moving page table entries, to effect an mremap move.
2783 */
2784struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2785	unsigned long addr, unsigned long len, pgoff_t pgoff,
2786	bool *need_rmap_locks)
2787{
2788	struct vm_area_struct *vma = *vmap;
2789	unsigned long vma_start = vma->vm_start;
2790	struct mm_struct *mm = vma->vm_mm;
2791	struct vm_area_struct *new_vma, *prev;
2792	struct rb_node **rb_link, *rb_parent;
2793	bool faulted_in_anon_vma = true;
2794
2795	/*
2796	 * If anonymous vma has not yet been faulted, update new pgoff
2797	 * to match new location, to increase its chance of merging.
2798	 */
2799	if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2800		pgoff = addr >> PAGE_SHIFT;
2801		faulted_in_anon_vma = false;
2802	}
2803
2804	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2805		return NULL;	/* should never get here */
2806	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2807			vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2808	if (new_vma) {
2809		/*
2810		 * Source vma may have been merged into new_vma
2811		 */
2812		if (unlikely(vma_start >= new_vma->vm_start &&
2813			     vma_start < new_vma->vm_end)) {
2814			/*
2815			 * The only way we can get a vma_merge with
2816			 * self during an mremap is if the vma hasn't
2817			 * been faulted in yet and we were allowed to
2818			 * reset the dst vma->vm_pgoff to the
2819			 * destination address of the mremap to allow
2820			 * the merge to happen. mremap must change the
2821			 * vm_pgoff linearity between src and dst vmas
2822			 * (in turn preventing a vma_merge) to be
2823			 * safe. It is only safe to keep the vm_pgoff
2824			 * linear if there are no pages mapped yet.
2825			 */
2826			VM_BUG_ON(faulted_in_anon_vma);
2827			*vmap = vma = new_vma;
2828		}
2829		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2830	} else {
2831		new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2832		if (new_vma) {
2833			*new_vma = *vma;
2834			new_vma->vm_start = addr;
2835			new_vma->vm_end = addr + len;
2836			new_vma->vm_pgoff = pgoff;
2837			if (vma_dup_policy(vma, new_vma))
2838				goto out_free_vma;
2839			INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2840			if (anon_vma_clone(new_vma, vma))
2841				goto out_free_mempol;
2842			if (new_vma->vm_file)
 
 
 
 
2843				get_file(new_vma->vm_file);
 
 
 
2844			if (new_vma->vm_ops && new_vma->vm_ops->open)
2845				new_vma->vm_ops->open(new_vma);
2846			vma_link(mm, new_vma, prev, rb_link, rb_parent);
2847			*need_rmap_locks = false;
2848		}
2849	}
2850	return new_vma;
2851
2852 out_free_mempol:
2853	mpol_put(vma_policy(new_vma));
2854 out_free_vma:
2855	kmem_cache_free(vm_area_cachep, new_vma);
2856	return NULL;
2857}
2858
2859/*
2860 * Return true if the calling process may expand its vm space by the passed
2861 * number of pages
2862 */
2863int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2864{
2865	unsigned long cur = mm->total_vm;	/* pages */
2866	unsigned long lim;
2867
2868	lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2869
2870	if (cur + npages > lim)
2871		return 0;
2872	return 1;
2873}
2874
2875
2876static int special_mapping_fault(struct vm_area_struct *vma,
2877				struct vm_fault *vmf)
2878{
2879	pgoff_t pgoff;
2880	struct page **pages;
2881
2882	/*
2883	 * special mappings have no vm_file, and in that case, the mm
2884	 * uses vm_pgoff internally. So we have to subtract it from here.
2885	 * We are allowed to do this because we are the mm; do not copy
2886	 * this code into drivers!
2887	 */
2888	pgoff = vmf->pgoff - vma->vm_pgoff;
2889
2890	for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2891		pgoff--;
2892
2893	if (*pages) {
2894		struct page *page = *pages;
2895		get_page(page);
2896		vmf->page = page;
2897		return 0;
2898	}
2899
2900	return VM_FAULT_SIGBUS;
2901}
2902
2903/*
2904 * Having a close hook prevents vma merging regardless of flags.
2905 */
2906static void special_mapping_close(struct vm_area_struct *vma)
2907{
2908}
2909
2910static const struct vm_operations_struct special_mapping_vmops = {
2911	.close = special_mapping_close,
2912	.fault = special_mapping_fault,
2913};
2914
2915/*
2916 * Called with mm->mmap_sem held for writing.
2917 * Insert a new vma covering the given region, with the given flags.
2918 * Its pages are supplied by the given array of struct page *.
2919 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2920 * The region past the last page supplied will always produce SIGBUS.
2921 * The array pointer and the pages it points to are assumed to stay alive
2922 * for as long as this mapping might exist.
2923 */
2924struct vm_area_struct *_install_special_mapping(struct mm_struct *mm,
2925			    unsigned long addr, unsigned long len,
2926			    unsigned long vm_flags, struct page **pages)
2927{
2928	int ret;
2929	struct vm_area_struct *vma;
2930
2931	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2932	if (unlikely(vma == NULL))
2933		return ERR_PTR(-ENOMEM);
2934
2935	INIT_LIST_HEAD(&vma->anon_vma_chain);
2936	vma->vm_mm = mm;
2937	vma->vm_start = addr;
2938	vma->vm_end = addr + len;
2939
2940	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
2941	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2942
2943	vma->vm_ops = &special_mapping_vmops;
2944	vma->vm_private_data = pages;
2945
 
 
 
 
2946	ret = insert_vm_struct(mm, vma);
2947	if (ret)
2948		goto out;
2949
2950	mm->total_vm += len >> PAGE_SHIFT;
2951
2952	perf_event_mmap(vma);
2953
2954	return vma;
2955
2956out:
2957	kmem_cache_free(vm_area_cachep, vma);
2958	return ERR_PTR(ret);
2959}
2960
2961int install_special_mapping(struct mm_struct *mm,
2962			    unsigned long addr, unsigned long len,
2963			    unsigned long vm_flags, struct page **pages)
2964{
2965	struct vm_area_struct *vma = _install_special_mapping(mm,
2966			    addr, len, vm_flags, pages);
2967
2968	if (IS_ERR(vma))
2969		return PTR_ERR(vma);
2970	return 0;
2971}
2972
2973static DEFINE_MUTEX(mm_all_locks_mutex);
2974
2975static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2976{
2977	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2978		/*
2979		 * The LSB of head.next can't change from under us
2980		 * because we hold the mm_all_locks_mutex.
2981		 */
2982		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
2983		/*
2984		 * We can safely modify head.next after taking the
2985		 * anon_vma->root->rwsem. If some other vma in this mm shares
2986		 * the same anon_vma we won't take it again.
2987		 *
2988		 * No need of atomic instructions here, head.next
2989		 * can't change from under us thanks to the
2990		 * anon_vma->root->rwsem.
2991		 */
2992		if (__test_and_set_bit(0, (unsigned long *)
2993				       &anon_vma->root->rb_root.rb_node))
2994			BUG();
2995	}
2996}
2997
2998static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2999{
3000	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3001		/*
3002		 * AS_MM_ALL_LOCKS can't change from under us because
3003		 * we hold the mm_all_locks_mutex.
3004		 *
3005		 * Operations on ->flags have to be atomic because
3006		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3007		 * mm_all_locks_mutex, there may be other cpus
3008		 * changing other bitflags in parallel to us.
3009		 */
3010		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3011			BUG();
3012		mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
3013	}
3014}
3015
3016/*
3017 * This operation locks against the VM for all pte/vma/mm related
3018 * operations that could ever happen on a certain mm. This includes
3019 * vmtruncate, try_to_unmap, and all page faults.
3020 *
3021 * The caller must take the mmap_sem in write mode before calling
3022 * mm_take_all_locks(). The caller isn't allowed to release the
3023 * mmap_sem until mm_drop_all_locks() returns.
3024 *
3025 * mmap_sem in write mode is required in order to block all operations
3026 * that could modify pagetables and free pages without need of
3027 * altering the vma layout (for example populate_range() with
3028 * nonlinear vmas). It's also needed in write mode to avoid new
3029 * anon_vmas to be associated with existing vmas.
3030 *
3031 * A single task can't take more than one mm_take_all_locks() in a row
3032 * or it would deadlock.
3033 *
3034 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3035 * mapping->flags avoid to take the same lock twice, if more than one
3036 * vma in this mm is backed by the same anon_vma or address_space.
3037 *
3038 * We can take all the locks in random order because the VM code
3039 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3040 * takes more than one of them in a row. Secondly we're protected
3041 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3042 *
3043 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3044 * that may have to take thousand of locks.
3045 *
3046 * mm_take_all_locks() can fail if it's interrupted by signals.
3047 */
3048int mm_take_all_locks(struct mm_struct *mm)
3049{
3050	struct vm_area_struct *vma;
3051	struct anon_vma_chain *avc;
 
3052
3053	BUG_ON(down_read_trylock(&mm->mmap_sem));
3054
3055	mutex_lock(&mm_all_locks_mutex);
3056
3057	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3058		if (signal_pending(current))
3059			goto out_unlock;
3060		if (vma->vm_file && vma->vm_file->f_mapping)
3061			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3062	}
3063
3064	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3065		if (signal_pending(current))
3066			goto out_unlock;
3067		if (vma->anon_vma)
3068			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3069				vm_lock_anon_vma(mm, avc->anon_vma);
3070	}
3071
3072	return 0;
3073
3074out_unlock:
3075	mm_drop_all_locks(mm);
3076	return -EINTR;
 
 
3077}
3078
3079static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3080{
3081	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3082		/*
3083		 * The LSB of head.next can't change to 0 from under
3084		 * us because we hold the mm_all_locks_mutex.
3085		 *
3086		 * We must however clear the bitflag before unlocking
3087		 * the vma so the users using the anon_vma->rb_root will
3088		 * never see our bitflag.
3089		 *
3090		 * No need of atomic instructions here, head.next
3091		 * can't change from under us until we release the
3092		 * anon_vma->root->rwsem.
3093		 */
3094		if (!__test_and_clear_bit(0, (unsigned long *)
3095					  &anon_vma->root->rb_root.rb_node))
3096			BUG();
3097		anon_vma_unlock_write(anon_vma);
3098	}
3099}
3100
3101static void vm_unlock_mapping(struct address_space *mapping)
3102{
3103	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3104		/*
3105		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3106		 * because we hold the mm_all_locks_mutex.
3107		 */
3108		mutex_unlock(&mapping->i_mmap_mutex);
3109		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3110					&mapping->flags))
3111			BUG();
3112	}
3113}
3114
3115/*
3116 * The mmap_sem cannot be released by the caller until
3117 * mm_drop_all_locks() returns.
3118 */
3119void mm_drop_all_locks(struct mm_struct *mm)
3120{
3121	struct vm_area_struct *vma;
3122	struct anon_vma_chain *avc;
3123
3124	BUG_ON(down_read_trylock(&mm->mmap_sem));
3125	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3126
3127	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3128		if (vma->anon_vma)
3129			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3130				vm_unlock_anon_vma(avc->anon_vma);
3131		if (vma->vm_file && vma->vm_file->f_mapping)
3132			vm_unlock_mapping(vma->vm_file->f_mapping);
3133	}
3134
3135	mutex_unlock(&mm_all_locks_mutex);
3136}
3137
3138/*
3139 * initialise the VMA slab
3140 */
3141void __init mmap_init(void)
3142{
3143	int ret;
3144
3145	ret = percpu_counter_init(&vm_committed_as, 0);
3146	VM_BUG_ON(ret);
3147}
3148
3149/*
3150 * Initialise sysctl_user_reserve_kbytes.
3151 *
3152 * This is intended to prevent a user from starting a single memory hogging
3153 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3154 * mode.
3155 *
3156 * The default value is min(3% of free memory, 128MB)
3157 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3158 */
3159static int init_user_reserve(void)
3160{
3161	unsigned long free_kbytes;
3162
3163	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3164
3165	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3166	return 0;
3167}
3168subsys_initcall(init_user_reserve);
3169
3170/*
3171 * Initialise sysctl_admin_reserve_kbytes.
3172 *
3173 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3174 * to log in and kill a memory hogging process.
3175 *
3176 * Systems with more than 256MB will reserve 8MB, enough to recover
3177 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3178 * only reserve 3% of free pages by default.
3179 */
3180static int init_admin_reserve(void)
3181{
3182	unsigned long free_kbytes;
3183
3184	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3185
3186	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3187	return 0;
3188}
3189subsys_initcall(init_admin_reserve);
3190
3191/*
3192 * Reinititalise user and admin reserves if memory is added or removed.
3193 *
3194 * The default user reserve max is 128MB, and the default max for the
3195 * admin reserve is 8MB. These are usually, but not always, enough to
3196 * enable recovery from a memory hogging process using login/sshd, a shell,
3197 * and tools like top. It may make sense to increase or even disable the
3198 * reserve depending on the existence of swap or variations in the recovery
3199 * tools. So, the admin may have changed them.
3200 *
3201 * If memory is added and the reserves have been eliminated or increased above
3202 * the default max, then we'll trust the admin.
3203 *
3204 * If memory is removed and there isn't enough free memory, then we
3205 * need to reset the reserves.
3206 *
3207 * Otherwise keep the reserve set by the admin.
3208 */
3209static int reserve_mem_notifier(struct notifier_block *nb,
3210			     unsigned long action, void *data)
3211{
3212	unsigned long tmp, free_kbytes;
3213
3214	switch (action) {
3215	case MEM_ONLINE:
3216		/* Default max is 128MB. Leave alone if modified by operator. */
3217		tmp = sysctl_user_reserve_kbytes;
3218		if (0 < tmp && tmp < (1UL << 17))
3219			init_user_reserve();
3220
3221		/* Default max is 8MB.  Leave alone if modified by operator. */
3222		tmp = sysctl_admin_reserve_kbytes;
3223		if (0 < tmp && tmp < (1UL << 13))
3224			init_admin_reserve();
3225
3226		break;
3227	case MEM_OFFLINE:
3228		free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3229
3230		if (sysctl_user_reserve_kbytes > free_kbytes) {
3231			init_user_reserve();
3232			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3233				sysctl_user_reserve_kbytes);
3234		}
3235
3236		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3237			init_admin_reserve();
3238			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3239				sysctl_admin_reserve_kbytes);
3240		}
3241		break;
3242	default:
3243		break;
3244	}
3245	return NOTIFY_OK;
3246}
3247
3248static struct notifier_block reserve_mem_nb = {
3249	.notifier_call = reserve_mem_notifier,
3250};
3251
3252static int __meminit init_reserve_notifier(void)
3253{
3254	if (register_hotmemory_notifier(&reserve_mem_nb))
3255		printk("Failed registering memory add/remove notifier for admin reserve");
3256
3257	return 0;
3258}
3259subsys_initcall(init_reserve_notifier);