xref: /linux/mm/mmap.c (revision 26b0d14106954ae46d2f4f7eec3481828a210f7d)
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/export.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 #include <linux/uprobes.h>
34 
35 #include <asm/uaccess.h>
36 #include <asm/cacheflush.h>
37 #include <asm/tlb.h>
38 #include <asm/mmu_context.h>
39 
40 #include "internal.h"
41 
42 #ifndef arch_mmap_check
43 #define arch_mmap_check(addr, len, flags)	(0)
44 #endif
45 
46 #ifndef arch_rebalance_pgtables
47 #define arch_rebalance_pgtables(addr, len)		(addr)
48 #endif
49 
50 static void unmap_region(struct mm_struct *mm,
51 		struct vm_area_struct *vma, struct vm_area_struct *prev,
52 		unsigned long start, unsigned long end);
53 
54 /*
55  * WARNING: the debugging will use recursive algorithms so never enable this
56  * unless you know what you are doing.
57  */
58 #undef DEBUG_MM_RB
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  */
75 pgprot_t protection_map[16] = {
76 	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
77 	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
78 };
79 
80 pgprot_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 }
86 EXPORT_SYMBOL(vm_get_page_prot);
87 
88 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
89 int sysctl_overcommit_ratio __read_mostly = 50;	/* default is 50% */
90 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
91 /*
92  * Make sure vm_committed_as in one cacheline and not cacheline shared with
93  * other variables. It can be updated by several CPUs frequently.
94  */
95 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
96 
97 /*
98  * Check that a process has enough memory to allocate a new virtual
99  * mapping. 0 means there is enough memory for the allocation to
100  * succeed and -ENOMEM implies there is not.
101  *
102  * We currently support three overcommit policies, which are set via the
103  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
104  *
105  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
106  * Additional code 2002 Jul 20 by Robert Love.
107  *
108  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
109  *
110  * Note this is a helper function intended to be used by LSMs which
111  * wish to use this logic.
112  */
113 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
114 {
115 	unsigned long free, allowed;
116 
117 	vm_acct_memory(pages);
118 
119 	/*
120 	 * Sometimes we want to use more memory than we have
121 	 */
122 	if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
123 		return 0;
124 
125 	if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
126 		free = global_page_state(NR_FREE_PAGES);
127 		free += global_page_state(NR_FILE_PAGES);
128 
129 		/*
130 		 * shmem pages shouldn't be counted as free in this
131 		 * case, they can't be purged, only swapped out, and
132 		 * that won't affect the overall amount of available
133 		 * memory in the system.
134 		 */
135 		free -= global_page_state(NR_SHMEM);
136 
137 		free += nr_swap_pages;
138 
139 		/*
140 		 * Any slabs which are created with the
141 		 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
142 		 * which are reclaimable, under pressure.  The dentry
143 		 * cache and most inode caches should fall into this
144 		 */
145 		free += global_page_state(NR_SLAB_RECLAIMABLE);
146 
147 		/*
148 		 * Leave reserved pages. The pages are not for anonymous pages.
149 		 */
150 		if (free <= totalreserve_pages)
151 			goto error;
152 		else
153 			free -= totalreserve_pages;
154 
155 		/*
156 		 * Leave the last 3% for root
157 		 */
158 		if (!cap_sys_admin)
159 			free -= free / 32;
160 
161 		if (free > pages)
162 			return 0;
163 
164 		goto error;
165 	}
166 
167 	allowed = (totalram_pages - hugetlb_total_pages())
168 	       	* sysctl_overcommit_ratio / 100;
169 	/*
170 	 * Leave the last 3% for root
171 	 */
172 	if (!cap_sys_admin)
173 		allowed -= allowed / 32;
174 	allowed += total_swap_pages;
175 
176 	/* Don't let a single process grow too big:
177 	   leave 3% of the size of this process for other processes */
178 	if (mm)
179 		allowed -= mm->total_vm / 32;
180 
181 	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
182 		return 0;
183 error:
184 	vm_unacct_memory(pages);
185 
186 	return -ENOMEM;
187 }
188 
189 /*
190  * Requires inode->i_mapping->i_mmap_mutex
191  */
192 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
193 		struct file *file, struct address_space *mapping)
194 {
195 	if (vma->vm_flags & VM_DENYWRITE)
196 		atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
197 	if (vma->vm_flags & VM_SHARED)
198 		mapping->i_mmap_writable--;
199 
200 	flush_dcache_mmap_lock(mapping);
201 	if (unlikely(vma->vm_flags & VM_NONLINEAR))
202 		list_del_init(&vma->shared.vm_set.list);
203 	else
204 		vma_prio_tree_remove(vma, &mapping->i_mmap);
205 	flush_dcache_mmap_unlock(mapping);
206 }
207 
208 /*
209  * Unlink a file-based vm structure from its prio_tree, to hide
210  * vma from rmap and vmtruncate before freeing its page tables.
211  */
212 void unlink_file_vma(struct vm_area_struct *vma)
213 {
214 	struct file *file = vma->vm_file;
215 
216 	if (file) {
217 		struct address_space *mapping = file->f_mapping;
218 		mutex_lock(&mapping->i_mmap_mutex);
219 		__remove_shared_vm_struct(vma, file, mapping);
220 		mutex_unlock(&mapping->i_mmap_mutex);
221 	}
222 }
223 
224 /*
225  * Close a vm structure and free it, returning the next.
226  */
227 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
228 {
229 	struct vm_area_struct *next = vma->vm_next;
230 
231 	might_sleep();
232 	if (vma->vm_ops && vma->vm_ops->close)
233 		vma->vm_ops->close(vma);
234 	if (vma->vm_file) {
235 		fput(vma->vm_file);
236 		if (vma->vm_flags & VM_EXECUTABLE)
237 			removed_exe_file_vma(vma->vm_mm);
238 	}
239 	mpol_put(vma_policy(vma));
240 	kmem_cache_free(vm_area_cachep, vma);
241 	return next;
242 }
243 
244 static unsigned long do_brk(unsigned long addr, unsigned long len);
245 
246 SYSCALL_DEFINE1(brk, unsigned long, brk)
247 {
248 	unsigned long rlim, retval;
249 	unsigned long newbrk, oldbrk;
250 	struct mm_struct *mm = current->mm;
251 	unsigned long min_brk;
252 
253 	down_write(&mm->mmap_sem);
254 
255 #ifdef CONFIG_COMPAT_BRK
256 	/*
257 	 * CONFIG_COMPAT_BRK can still be overridden by setting
258 	 * randomize_va_space to 2, which will still cause mm->start_brk
259 	 * to be arbitrarily shifted
260 	 */
261 	if (current->brk_randomized)
262 		min_brk = mm->start_brk;
263 	else
264 		min_brk = mm->end_data;
265 #else
266 	min_brk = mm->start_brk;
267 #endif
268 	if (brk < min_brk)
269 		goto out;
270 
271 	/*
272 	 * Check against rlimit here. If this check is done later after the test
273 	 * of oldbrk with newbrk then it can escape the test and let the data
274 	 * segment grow beyond its set limit the in case where the limit is
275 	 * not page aligned -Ram Gupta
276 	 */
277 	rlim = rlimit(RLIMIT_DATA);
278 	if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
279 			(mm->end_data - mm->start_data) > rlim)
280 		goto out;
281 
282 	newbrk = PAGE_ALIGN(brk);
283 	oldbrk = PAGE_ALIGN(mm->brk);
284 	if (oldbrk == newbrk)
285 		goto set_brk;
286 
287 	/* Always allow shrinking brk. */
288 	if (brk <= mm->brk) {
289 		if (!do_munmap(mm, newbrk, oldbrk-newbrk))
290 			goto set_brk;
291 		goto out;
292 	}
293 
294 	/* Check against existing mmap mappings. */
295 	if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
296 		goto out;
297 
298 	/* Ok, looks good - let it rip. */
299 	if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
300 		goto out;
301 set_brk:
302 	mm->brk = brk;
303 out:
304 	retval = mm->brk;
305 	up_write(&mm->mmap_sem);
306 	return retval;
307 }
308 
309 #ifdef DEBUG_MM_RB
310 static int browse_rb(struct rb_root *root)
311 {
312 	int i = 0, j;
313 	struct rb_node *nd, *pn = NULL;
314 	unsigned long prev = 0, pend = 0;
315 
316 	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
317 		struct vm_area_struct *vma;
318 		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
319 		if (vma->vm_start < prev)
320 			printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
321 		if (vma->vm_start < pend)
322 			printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
323 		if (vma->vm_start > vma->vm_end)
324 			printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
325 		i++;
326 		pn = nd;
327 		prev = vma->vm_start;
328 		pend = vma->vm_end;
329 	}
330 	j = 0;
331 	for (nd = pn; nd; nd = rb_prev(nd)) {
332 		j++;
333 	}
334 	if (i != j)
335 		printk("backwards %d, forwards %d\n", j, i), i = 0;
336 	return i;
337 }
338 
339 void validate_mm(struct mm_struct *mm)
340 {
341 	int bug = 0;
342 	int i = 0;
343 	struct vm_area_struct *tmp = mm->mmap;
344 	while (tmp) {
345 		tmp = tmp->vm_next;
346 		i++;
347 	}
348 	if (i != mm->map_count)
349 		printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
350 	i = browse_rb(&mm->mm_rb);
351 	if (i != mm->map_count)
352 		printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
353 	BUG_ON(bug);
354 }
355 #else
356 #define validate_mm(mm) do { } while (0)
357 #endif
358 
359 static struct vm_area_struct *
360 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
361 		struct vm_area_struct **pprev, struct rb_node ***rb_link,
362 		struct rb_node ** rb_parent)
363 {
364 	struct vm_area_struct * vma;
365 	struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
366 
367 	__rb_link = &mm->mm_rb.rb_node;
368 	rb_prev = __rb_parent = NULL;
369 	vma = NULL;
370 
371 	while (*__rb_link) {
372 		struct vm_area_struct *vma_tmp;
373 
374 		__rb_parent = *__rb_link;
375 		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
376 
377 		if (vma_tmp->vm_end > addr) {
378 			vma = vma_tmp;
379 			if (vma_tmp->vm_start <= addr)
380 				break;
381 			__rb_link = &__rb_parent->rb_left;
382 		} else {
383 			rb_prev = __rb_parent;
384 			__rb_link = &__rb_parent->rb_right;
385 		}
386 	}
387 
388 	*pprev = NULL;
389 	if (rb_prev)
390 		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
391 	*rb_link = __rb_link;
392 	*rb_parent = __rb_parent;
393 	return vma;
394 }
395 
396 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
397 		struct rb_node **rb_link, struct rb_node *rb_parent)
398 {
399 	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
400 	rb_insert_color(&vma->vm_rb, &mm->mm_rb);
401 }
402 
403 static void __vma_link_file(struct vm_area_struct *vma)
404 {
405 	struct file *file;
406 
407 	file = vma->vm_file;
408 	if (file) {
409 		struct address_space *mapping = file->f_mapping;
410 
411 		if (vma->vm_flags & VM_DENYWRITE)
412 			atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
413 		if (vma->vm_flags & VM_SHARED)
414 			mapping->i_mmap_writable++;
415 
416 		flush_dcache_mmap_lock(mapping);
417 		if (unlikely(vma->vm_flags & VM_NONLINEAR))
418 			vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
419 		else
420 			vma_prio_tree_insert(vma, &mapping->i_mmap);
421 		flush_dcache_mmap_unlock(mapping);
422 	}
423 }
424 
425 static void
426 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
427 	struct vm_area_struct *prev, struct rb_node **rb_link,
428 	struct rb_node *rb_parent)
429 {
430 	__vma_link_list(mm, vma, prev, rb_parent);
431 	__vma_link_rb(mm, vma, rb_link, rb_parent);
432 }
433 
434 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435 			struct vm_area_struct *prev, struct rb_node **rb_link,
436 			struct rb_node *rb_parent)
437 {
438 	struct address_space *mapping = NULL;
439 
440 	if (vma->vm_file)
441 		mapping = vma->vm_file->f_mapping;
442 
443 	if (mapping)
444 		mutex_lock(&mapping->i_mmap_mutex);
445 
446 	__vma_link(mm, vma, prev, rb_link, rb_parent);
447 	__vma_link_file(vma);
448 
449 	if (mapping)
450 		mutex_unlock(&mapping->i_mmap_mutex);
451 
452 	mm->map_count++;
453 	validate_mm(mm);
454 }
455 
456 /*
457  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
458  * mm's list and rbtree.  It has already been inserted into the prio_tree.
459  */
460 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
461 {
462 	struct vm_area_struct *__vma, *prev;
463 	struct rb_node **rb_link, *rb_parent;
464 
465 	__vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466 	BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467 	__vma_link(mm, vma, prev, rb_link, rb_parent);
468 	mm->map_count++;
469 }
470 
471 static inline void
472 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473 		struct vm_area_struct *prev)
474 {
475 	struct vm_area_struct *next = vma->vm_next;
476 
477 	prev->vm_next = next;
478 	if (next)
479 		next->vm_prev = prev;
480 	rb_erase(&vma->vm_rb, &mm->mm_rb);
481 	if (mm->mmap_cache == vma)
482 		mm->mmap_cache = prev;
483 }
484 
485 /*
486  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
487  * is already present in an i_mmap tree without adjusting the tree.
488  * The following helper function should be used when such adjustments
489  * are necessary.  The "insert" vma (if any) is to be inserted
490  * before we drop the necessary locks.
491  */
492 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
493 	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
494 {
495 	struct mm_struct *mm = vma->vm_mm;
496 	struct vm_area_struct *next = vma->vm_next;
497 	struct vm_area_struct *importer = NULL;
498 	struct address_space *mapping = NULL;
499 	struct prio_tree_root *root = NULL;
500 	struct anon_vma *anon_vma = NULL;
501 	struct file *file = vma->vm_file;
502 	long adjust_next = 0;
503 	int remove_next = 0;
504 
505 	if (next && !insert) {
506 		struct vm_area_struct *exporter = NULL;
507 
508 		if (end >= next->vm_end) {
509 			/*
510 			 * vma expands, overlapping all the next, and
511 			 * perhaps the one after too (mprotect case 6).
512 			 */
513 again:			remove_next = 1 + (end > next->vm_end);
514 			end = next->vm_end;
515 			exporter = next;
516 			importer = vma;
517 		} else if (end > next->vm_start) {
518 			/*
519 			 * vma expands, overlapping part of the next:
520 			 * mprotect case 5 shifting the boundary up.
521 			 */
522 			adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
523 			exporter = next;
524 			importer = vma;
525 		} else if (end < vma->vm_end) {
526 			/*
527 			 * vma shrinks, and !insert tells it's not
528 			 * split_vma inserting another: so it must be
529 			 * mprotect case 4 shifting the boundary down.
530 			 */
531 			adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
532 			exporter = vma;
533 			importer = next;
534 		}
535 
536 		/*
537 		 * Easily overlooked: when mprotect shifts the boundary,
538 		 * make sure the expanding vma has anon_vma set if the
539 		 * shrinking vma had, to cover any anon pages imported.
540 		 */
541 		if (exporter && exporter->anon_vma && !importer->anon_vma) {
542 			if (anon_vma_clone(importer, exporter))
543 				return -ENOMEM;
544 			importer->anon_vma = exporter->anon_vma;
545 		}
546 	}
547 
548 	if (file) {
549 		mapping = file->f_mapping;
550 		if (!(vma->vm_flags & VM_NONLINEAR)) {
551 			root = &mapping->i_mmap;
552 			uprobe_munmap(vma, vma->vm_start, vma->vm_end);
553 
554 			if (adjust_next)
555 				uprobe_munmap(next, next->vm_start,
556 							next->vm_end);
557 		}
558 
559 		mutex_lock(&mapping->i_mmap_mutex);
560 		if (insert) {
561 			/*
562 			 * Put into prio_tree now, so instantiated pages
563 			 * are visible to arm/parisc __flush_dcache_page
564 			 * throughout; but we cannot insert into address
565 			 * space until vma start or end is updated.
566 			 */
567 			__vma_link_file(insert);
568 		}
569 	}
570 
571 	vma_adjust_trans_huge(vma, start, end, adjust_next);
572 
573 	/*
574 	 * When changing only vma->vm_end, we don't really need anon_vma
575 	 * lock. This is a fairly rare case by itself, but the anon_vma
576 	 * lock may be shared between many sibling processes.  Skipping
577 	 * the lock for brk adjustments makes a difference sometimes.
578 	 */
579 	if (vma->anon_vma && (importer || start != vma->vm_start)) {
580 		anon_vma = vma->anon_vma;
581 		anon_vma_lock(anon_vma);
582 	}
583 
584 	if (root) {
585 		flush_dcache_mmap_lock(mapping);
586 		vma_prio_tree_remove(vma, root);
587 		if (adjust_next)
588 			vma_prio_tree_remove(next, root);
589 	}
590 
591 	vma->vm_start = start;
592 	vma->vm_end = end;
593 	vma->vm_pgoff = pgoff;
594 	if (adjust_next) {
595 		next->vm_start += adjust_next << PAGE_SHIFT;
596 		next->vm_pgoff += adjust_next;
597 	}
598 
599 	if (root) {
600 		if (adjust_next)
601 			vma_prio_tree_insert(next, root);
602 		vma_prio_tree_insert(vma, root);
603 		flush_dcache_mmap_unlock(mapping);
604 	}
605 
606 	if (remove_next) {
607 		/*
608 		 * vma_merge has merged next into vma, and needs
609 		 * us to remove next before dropping the locks.
610 		 */
611 		__vma_unlink(mm, next, vma);
612 		if (file)
613 			__remove_shared_vm_struct(next, file, mapping);
614 	} else if (insert) {
615 		/*
616 		 * split_vma has split insert from vma, and needs
617 		 * us to insert it before dropping the locks
618 		 * (it may either follow vma or precede it).
619 		 */
620 		__insert_vm_struct(mm, insert);
621 	}
622 
623 	if (anon_vma)
624 		anon_vma_unlock(anon_vma);
625 	if (mapping)
626 		mutex_unlock(&mapping->i_mmap_mutex);
627 
628 	if (root) {
629 		uprobe_mmap(vma);
630 
631 		if (adjust_next)
632 			uprobe_mmap(next);
633 	}
634 
635 	if (remove_next) {
636 		if (file) {
637 			uprobe_munmap(next, next->vm_start, next->vm_end);
638 			fput(file);
639 			if (next->vm_flags & VM_EXECUTABLE)
640 				removed_exe_file_vma(mm);
641 		}
642 		if (next->anon_vma)
643 			anon_vma_merge(vma, next);
644 		mm->map_count--;
645 		mpol_put(vma_policy(next));
646 		kmem_cache_free(vm_area_cachep, next);
647 		/*
648 		 * In mprotect's case 6 (see comments on vma_merge),
649 		 * we must remove another next too. It would clutter
650 		 * up the code too much to do both in one go.
651 		 */
652 		if (remove_next == 2) {
653 			next = vma->vm_next;
654 			goto again;
655 		}
656 	}
657 	if (insert && file)
658 		uprobe_mmap(insert);
659 
660 	validate_mm(mm);
661 
662 	return 0;
663 }
664 
665 /*
666  * If the vma has a ->close operation then the driver probably needs to release
667  * per-vma resources, so we don't attempt to merge those.
668  */
669 static inline int is_mergeable_vma(struct vm_area_struct *vma,
670 			struct file *file, unsigned long vm_flags)
671 {
672 	/* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
673 	if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
674 		return 0;
675 	if (vma->vm_file != file)
676 		return 0;
677 	if (vma->vm_ops && vma->vm_ops->close)
678 		return 0;
679 	return 1;
680 }
681 
682 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
683 					struct anon_vma *anon_vma2,
684 					struct vm_area_struct *vma)
685 {
686 	/*
687 	 * The list_is_singular() test is to avoid merging VMA cloned from
688 	 * parents. This can improve scalability caused by anon_vma lock.
689 	 */
690 	if ((!anon_vma1 || !anon_vma2) && (!vma ||
691 		list_is_singular(&vma->anon_vma_chain)))
692 		return 1;
693 	return anon_vma1 == anon_vma2;
694 }
695 
696 /*
697  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
698  * in front of (at a lower virtual address and file offset than) the vma.
699  *
700  * We cannot merge two vmas if they have differently assigned (non-NULL)
701  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
702  *
703  * We don't check here for the merged mmap wrapping around the end of pagecache
704  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
705  * wrap, nor mmaps which cover the final page at index -1UL.
706  */
707 static int
708 can_vma_merge_before(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 		if (vma->vm_pgoff == vm_pgoff)
714 			return 1;
715 	}
716 	return 0;
717 }
718 
719 /*
720  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
721  * beyond (at a higher virtual address and file offset than) the vma.
722  *
723  * We cannot merge two vmas if they have differently assigned (non-NULL)
724  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
725  */
726 static int
727 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
728 	struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
729 {
730 	if (is_mergeable_vma(vma, file, vm_flags) &&
731 	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
732 		pgoff_t vm_pglen;
733 		vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
734 		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
735 			return 1;
736 	}
737 	return 0;
738 }
739 
740 /*
741  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
742  * whether that can be merged with its predecessor or its successor.
743  * Or both (it neatly fills a hole).
744  *
745  * In most cases - when called for mmap, brk or mremap - [addr,end) is
746  * certain not to be mapped by the time vma_merge is called; but when
747  * called for mprotect, it is certain to be already mapped (either at
748  * an offset within prev, or at the start of next), and the flags of
749  * this area are about to be changed to vm_flags - and the no-change
750  * case has already been eliminated.
751  *
752  * The following mprotect cases have to be considered, where AAAA is
753  * the area passed down from mprotect_fixup, never extending beyond one
754  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
755  *
756  *     AAAA             AAAA                AAAA          AAAA
757  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
758  *    cannot merge    might become    might become    might become
759  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
760  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
761  *    mremap move:                                    PPPPNNNNNNNN 8
762  *        AAAA
763  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
764  *    might become    case 1 below    case 2 below    case 3 below
765  *
766  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
767  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
768  */
769 struct vm_area_struct *vma_merge(struct mm_struct *mm,
770 			struct vm_area_struct *prev, unsigned long addr,
771 			unsigned long end, unsigned long vm_flags,
772 		     	struct anon_vma *anon_vma, struct file *file,
773 			pgoff_t pgoff, struct mempolicy *policy)
774 {
775 	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
776 	struct vm_area_struct *area, *next;
777 	int err;
778 
779 	/*
780 	 * We later require that vma->vm_flags == vm_flags,
781 	 * so this tests vma->vm_flags & VM_SPECIAL, too.
782 	 */
783 	if (vm_flags & VM_SPECIAL)
784 		return NULL;
785 
786 	if (prev)
787 		next = prev->vm_next;
788 	else
789 		next = mm->mmap;
790 	area = next;
791 	if (next && next->vm_end == end)		/* cases 6, 7, 8 */
792 		next = next->vm_next;
793 
794 	/*
795 	 * Can it merge with the predecessor?
796 	 */
797 	if (prev && prev->vm_end == addr &&
798   			mpol_equal(vma_policy(prev), policy) &&
799 			can_vma_merge_after(prev, vm_flags,
800 						anon_vma, file, pgoff)) {
801 		/*
802 		 * OK, it can.  Can we now merge in the successor as well?
803 		 */
804 		if (next && end == next->vm_start &&
805 				mpol_equal(policy, vma_policy(next)) &&
806 				can_vma_merge_before(next, vm_flags,
807 					anon_vma, file, pgoff+pglen) &&
808 				is_mergeable_anon_vma(prev->anon_vma,
809 						      next->anon_vma, NULL)) {
810 							/* cases 1, 6 */
811 			err = vma_adjust(prev, prev->vm_start,
812 				next->vm_end, prev->vm_pgoff, NULL);
813 		} else					/* cases 2, 5, 7 */
814 			err = vma_adjust(prev, prev->vm_start,
815 				end, prev->vm_pgoff, NULL);
816 		if (err)
817 			return NULL;
818 		khugepaged_enter_vma_merge(prev);
819 		return prev;
820 	}
821 
822 	/*
823 	 * Can this new request be merged in front of next?
824 	 */
825 	if (next && end == next->vm_start &&
826  			mpol_equal(policy, vma_policy(next)) &&
827 			can_vma_merge_before(next, vm_flags,
828 					anon_vma, file, pgoff+pglen)) {
829 		if (prev && addr < prev->vm_end)	/* case 4 */
830 			err = vma_adjust(prev, prev->vm_start,
831 				addr, prev->vm_pgoff, NULL);
832 		else					/* cases 3, 8 */
833 			err = vma_adjust(area, addr, next->vm_end,
834 				next->vm_pgoff - pglen, NULL);
835 		if (err)
836 			return NULL;
837 		khugepaged_enter_vma_merge(area);
838 		return area;
839 	}
840 
841 	return NULL;
842 }
843 
844 /*
845  * Rough compatbility check to quickly see if it's even worth looking
846  * at sharing an anon_vma.
847  *
848  * They need to have the same vm_file, and the flags can only differ
849  * in things that mprotect may change.
850  *
851  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
852  * we can merge the two vma's. For example, we refuse to merge a vma if
853  * there is a vm_ops->close() function, because that indicates that the
854  * driver is doing some kind of reference counting. But that doesn't
855  * really matter for the anon_vma sharing case.
856  */
857 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
858 {
859 	return a->vm_end == b->vm_start &&
860 		mpol_equal(vma_policy(a), vma_policy(b)) &&
861 		a->vm_file == b->vm_file &&
862 		!((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
863 		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
864 }
865 
866 /*
867  * Do some basic sanity checking to see if we can re-use the anon_vma
868  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
869  * the same as 'old', the other will be the new one that is trying
870  * to share the anon_vma.
871  *
872  * NOTE! This runs with mm_sem held for reading, so it is possible that
873  * the anon_vma of 'old' is concurrently in the process of being set up
874  * by another page fault trying to merge _that_. But that's ok: if it
875  * is being set up, that automatically means that it will be a singleton
876  * acceptable for merging, so we can do all of this optimistically. But
877  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
878  *
879  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
880  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
881  * is to return an anon_vma that is "complex" due to having gone through
882  * a fork).
883  *
884  * We also make sure that the two vma's are compatible (adjacent,
885  * and with the same memory policies). That's all stable, even with just
886  * a read lock on the mm_sem.
887  */
888 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
889 {
890 	if (anon_vma_compatible(a, b)) {
891 		struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
892 
893 		if (anon_vma && list_is_singular(&old->anon_vma_chain))
894 			return anon_vma;
895 	}
896 	return NULL;
897 }
898 
899 /*
900  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
901  * neighbouring vmas for a suitable anon_vma, before it goes off
902  * to allocate a new anon_vma.  It checks because a repetitive
903  * sequence of mprotects and faults may otherwise lead to distinct
904  * anon_vmas being allocated, preventing vma merge in subsequent
905  * mprotect.
906  */
907 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
908 {
909 	struct anon_vma *anon_vma;
910 	struct vm_area_struct *near;
911 
912 	near = vma->vm_next;
913 	if (!near)
914 		goto try_prev;
915 
916 	anon_vma = reusable_anon_vma(near, vma, near);
917 	if (anon_vma)
918 		return anon_vma;
919 try_prev:
920 	near = vma->vm_prev;
921 	if (!near)
922 		goto none;
923 
924 	anon_vma = reusable_anon_vma(near, near, vma);
925 	if (anon_vma)
926 		return anon_vma;
927 none:
928 	/*
929 	 * There's no absolute need to look only at touching neighbours:
930 	 * we could search further afield for "compatible" anon_vmas.
931 	 * But it would probably just be a waste of time searching,
932 	 * or lead to too many vmas hanging off the same anon_vma.
933 	 * We're trying to allow mprotect remerging later on,
934 	 * not trying to minimize memory used for anon_vmas.
935 	 */
936 	return NULL;
937 }
938 
939 #ifdef CONFIG_PROC_FS
940 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
941 						struct file *file, long pages)
942 {
943 	const unsigned long stack_flags
944 		= VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
945 
946 	if (file) {
947 		mm->shared_vm += pages;
948 		if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
949 			mm->exec_vm += pages;
950 	} else if (flags & stack_flags)
951 		mm->stack_vm += pages;
952 	if (flags & (VM_RESERVED|VM_IO))
953 		mm->reserved_vm += pages;
954 }
955 #endif /* CONFIG_PROC_FS */
956 
957 /*
958  * If a hint addr is less than mmap_min_addr change hint to be as
959  * low as possible but still greater than mmap_min_addr
960  */
961 static inline unsigned long round_hint_to_min(unsigned long hint)
962 {
963 	hint &= PAGE_MASK;
964 	if (((void *)hint != NULL) &&
965 	    (hint < mmap_min_addr))
966 		return PAGE_ALIGN(mmap_min_addr);
967 	return hint;
968 }
969 
970 /*
971  * The caller must hold down_write(&current->mm->mmap_sem).
972  */
973 
974 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
975 			unsigned long len, unsigned long prot,
976 			unsigned long flags, unsigned long pgoff)
977 {
978 	struct mm_struct * mm = current->mm;
979 	struct inode *inode;
980 	vm_flags_t vm_flags;
981 
982 	/*
983 	 * Does the application expect PROT_READ to imply PROT_EXEC?
984 	 *
985 	 * (the exception is when the underlying filesystem is noexec
986 	 *  mounted, in which case we dont add PROT_EXEC.)
987 	 */
988 	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
989 		if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
990 			prot |= PROT_EXEC;
991 
992 	if (!len)
993 		return -EINVAL;
994 
995 	if (!(flags & MAP_FIXED))
996 		addr = round_hint_to_min(addr);
997 
998 	/* Careful about overflows.. */
999 	len = PAGE_ALIGN(len);
1000 	if (!len)
1001 		return -ENOMEM;
1002 
1003 	/* offset overflow? */
1004 	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1005                return -EOVERFLOW;
1006 
1007 	/* Too many mappings? */
1008 	if (mm->map_count > sysctl_max_map_count)
1009 		return -ENOMEM;
1010 
1011 	/* Obtain the address to map to. we verify (or select) it and ensure
1012 	 * that it represents a valid section of the address space.
1013 	 */
1014 	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1015 	if (addr & ~PAGE_MASK)
1016 		return addr;
1017 
1018 	/* Do simple checking here so the lower-level routines won't have
1019 	 * to. we assume access permissions have been handled by the open
1020 	 * of the memory object, so we don't do any here.
1021 	 */
1022 	vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1023 			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1024 
1025 	if (flags & MAP_LOCKED)
1026 		if (!can_do_mlock())
1027 			return -EPERM;
1028 
1029 	/* mlock MCL_FUTURE? */
1030 	if (vm_flags & VM_LOCKED) {
1031 		unsigned long locked, lock_limit;
1032 		locked = len >> PAGE_SHIFT;
1033 		locked += mm->locked_vm;
1034 		lock_limit = rlimit(RLIMIT_MEMLOCK);
1035 		lock_limit >>= PAGE_SHIFT;
1036 		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1037 			return -EAGAIN;
1038 	}
1039 
1040 	inode = file ? file->f_path.dentry->d_inode : NULL;
1041 
1042 	if (file) {
1043 		switch (flags & MAP_TYPE) {
1044 		case MAP_SHARED:
1045 			if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1046 				return -EACCES;
1047 
1048 			/*
1049 			 * Make sure we don't allow writing to an append-only
1050 			 * file..
1051 			 */
1052 			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1053 				return -EACCES;
1054 
1055 			/*
1056 			 * Make sure there are no mandatory locks on the file.
1057 			 */
1058 			if (locks_verify_locked(inode))
1059 				return -EAGAIN;
1060 
1061 			vm_flags |= VM_SHARED | VM_MAYSHARE;
1062 			if (!(file->f_mode & FMODE_WRITE))
1063 				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1064 
1065 			/* fall through */
1066 		case MAP_PRIVATE:
1067 			if (!(file->f_mode & FMODE_READ))
1068 				return -EACCES;
1069 			if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1070 				if (vm_flags & VM_EXEC)
1071 					return -EPERM;
1072 				vm_flags &= ~VM_MAYEXEC;
1073 			}
1074 
1075 			if (!file->f_op || !file->f_op->mmap)
1076 				return -ENODEV;
1077 			break;
1078 
1079 		default:
1080 			return -EINVAL;
1081 		}
1082 	} else {
1083 		switch (flags & MAP_TYPE) {
1084 		case MAP_SHARED:
1085 			/*
1086 			 * Ignore pgoff.
1087 			 */
1088 			pgoff = 0;
1089 			vm_flags |= VM_SHARED | VM_MAYSHARE;
1090 			break;
1091 		case MAP_PRIVATE:
1092 			/*
1093 			 * Set pgoff according to addr for anon_vma.
1094 			 */
1095 			pgoff = addr >> PAGE_SHIFT;
1096 			break;
1097 		default:
1098 			return -EINVAL;
1099 		}
1100 	}
1101 
1102 	return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1103 }
1104 
1105 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1106 		unsigned long, prot, unsigned long, flags,
1107 		unsigned long, fd, unsigned long, pgoff)
1108 {
1109 	struct file *file = NULL;
1110 	unsigned long retval = -EBADF;
1111 
1112 	if (!(flags & MAP_ANONYMOUS)) {
1113 		audit_mmap_fd(fd, flags);
1114 		if (unlikely(flags & MAP_HUGETLB))
1115 			return -EINVAL;
1116 		file = fget(fd);
1117 		if (!file)
1118 			goto out;
1119 	} else if (flags & MAP_HUGETLB) {
1120 		struct user_struct *user = NULL;
1121 		/*
1122 		 * VM_NORESERVE is used because the reservations will be
1123 		 * taken when vm_ops->mmap() is called
1124 		 * A dummy user value is used because we are not locking
1125 		 * memory so no accounting is necessary
1126 		 */
1127 		file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1128 						VM_NORESERVE, &user,
1129 						HUGETLB_ANONHUGE_INODE);
1130 		if (IS_ERR(file))
1131 			return PTR_ERR(file);
1132 	}
1133 
1134 	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1135 
1136 	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1137 	if (file)
1138 		fput(file);
1139 out:
1140 	return retval;
1141 }
1142 
1143 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1144 struct mmap_arg_struct {
1145 	unsigned long addr;
1146 	unsigned long len;
1147 	unsigned long prot;
1148 	unsigned long flags;
1149 	unsigned long fd;
1150 	unsigned long offset;
1151 };
1152 
1153 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1154 {
1155 	struct mmap_arg_struct a;
1156 
1157 	if (copy_from_user(&a, arg, sizeof(a)))
1158 		return -EFAULT;
1159 	if (a.offset & ~PAGE_MASK)
1160 		return -EINVAL;
1161 
1162 	return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1163 			      a.offset >> PAGE_SHIFT);
1164 }
1165 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1166 
1167 /*
1168  * Some shared mappigns will want the pages marked read-only
1169  * to track write events. If so, we'll downgrade vm_page_prot
1170  * to the private version (using protection_map[] without the
1171  * VM_SHARED bit).
1172  */
1173 int vma_wants_writenotify(struct vm_area_struct *vma)
1174 {
1175 	vm_flags_t vm_flags = vma->vm_flags;
1176 
1177 	/* If it was private or non-writable, the write bit is already clear */
1178 	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1179 		return 0;
1180 
1181 	/* The backer wishes to know when pages are first written to? */
1182 	if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1183 		return 1;
1184 
1185 	/* The open routine did something to the protections already? */
1186 	if (pgprot_val(vma->vm_page_prot) !=
1187 	    pgprot_val(vm_get_page_prot(vm_flags)))
1188 		return 0;
1189 
1190 	/* Specialty mapping? */
1191 	if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1192 		return 0;
1193 
1194 	/* Can the mapping track the dirty pages? */
1195 	return vma->vm_file && vma->vm_file->f_mapping &&
1196 		mapping_cap_account_dirty(vma->vm_file->f_mapping);
1197 }
1198 
1199 /*
1200  * We account for memory if it's a private writeable mapping,
1201  * not hugepages and VM_NORESERVE wasn't set.
1202  */
1203 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1204 {
1205 	/*
1206 	 * hugetlb has its own accounting separate from the core VM
1207 	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1208 	 */
1209 	if (file && is_file_hugepages(file))
1210 		return 0;
1211 
1212 	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1213 }
1214 
1215 unsigned long mmap_region(struct file *file, unsigned long addr,
1216 			  unsigned long len, unsigned long flags,
1217 			  vm_flags_t vm_flags, unsigned long pgoff)
1218 {
1219 	struct mm_struct *mm = current->mm;
1220 	struct vm_area_struct *vma, *prev;
1221 	int correct_wcount = 0;
1222 	int error;
1223 	struct rb_node **rb_link, *rb_parent;
1224 	unsigned long charged = 0;
1225 	struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1226 
1227 	/* Clear old maps */
1228 	error = -ENOMEM;
1229 munmap_back:
1230 	vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1231 	if (vma && vma->vm_start < addr + len) {
1232 		if (do_munmap(mm, addr, len))
1233 			return -ENOMEM;
1234 		goto munmap_back;
1235 	}
1236 
1237 	/* Check against address space limit. */
1238 	if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1239 		return -ENOMEM;
1240 
1241 	/*
1242 	 * Set 'VM_NORESERVE' if we should not account for the
1243 	 * memory use of this mapping.
1244 	 */
1245 	if ((flags & MAP_NORESERVE)) {
1246 		/* We honor MAP_NORESERVE if allowed to overcommit */
1247 		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1248 			vm_flags |= VM_NORESERVE;
1249 
1250 		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1251 		if (file && is_file_hugepages(file))
1252 			vm_flags |= VM_NORESERVE;
1253 	}
1254 
1255 	/*
1256 	 * Private writable mapping: check memory availability
1257 	 */
1258 	if (accountable_mapping(file, vm_flags)) {
1259 		charged = len >> PAGE_SHIFT;
1260 		if (security_vm_enough_memory_mm(mm, charged))
1261 			return -ENOMEM;
1262 		vm_flags |= VM_ACCOUNT;
1263 	}
1264 
1265 	/*
1266 	 * Can we just expand an old mapping?
1267 	 */
1268 	vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1269 	if (vma)
1270 		goto out;
1271 
1272 	/*
1273 	 * Determine the object being mapped and call the appropriate
1274 	 * specific mapper. the address has already been validated, but
1275 	 * not unmapped, but the maps are removed from the list.
1276 	 */
1277 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1278 	if (!vma) {
1279 		error = -ENOMEM;
1280 		goto unacct_error;
1281 	}
1282 
1283 	vma->vm_mm = mm;
1284 	vma->vm_start = addr;
1285 	vma->vm_end = addr + len;
1286 	vma->vm_flags = vm_flags;
1287 	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1288 	vma->vm_pgoff = pgoff;
1289 	INIT_LIST_HEAD(&vma->anon_vma_chain);
1290 
1291 	error = -EINVAL;	/* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1292 
1293 	if (file) {
1294 		if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1295 			goto free_vma;
1296 		if (vm_flags & VM_DENYWRITE) {
1297 			error = deny_write_access(file);
1298 			if (error)
1299 				goto free_vma;
1300 			correct_wcount = 1;
1301 		}
1302 		vma->vm_file = file;
1303 		get_file(file);
1304 		error = file->f_op->mmap(file, vma);
1305 		if (error)
1306 			goto unmap_and_free_vma;
1307 		if (vm_flags & VM_EXECUTABLE)
1308 			added_exe_file_vma(mm);
1309 
1310 		/* Can addr have changed??
1311 		 *
1312 		 * Answer: Yes, several device drivers can do it in their
1313 		 *         f_op->mmap method. -DaveM
1314 		 */
1315 		addr = vma->vm_start;
1316 		pgoff = vma->vm_pgoff;
1317 		vm_flags = vma->vm_flags;
1318 	} else if (vm_flags & VM_SHARED) {
1319 		if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1320 			goto free_vma;
1321 		error = shmem_zero_setup(vma);
1322 		if (error)
1323 			goto free_vma;
1324 	}
1325 
1326 	if (vma_wants_writenotify(vma)) {
1327 		pgprot_t pprot = vma->vm_page_prot;
1328 
1329 		/* Can vma->vm_page_prot have changed??
1330 		 *
1331 		 * Answer: Yes, drivers may have changed it in their
1332 		 *         f_op->mmap method.
1333 		 *
1334 		 * Ensures that vmas marked as uncached stay that way.
1335 		 */
1336 		vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1337 		if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1338 			vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1339 	}
1340 
1341 	vma_link(mm, vma, prev, rb_link, rb_parent);
1342 	file = vma->vm_file;
1343 
1344 	/* Once vma denies write, undo our temporary denial count */
1345 	if (correct_wcount)
1346 		atomic_inc(&inode->i_writecount);
1347 out:
1348 	perf_event_mmap(vma);
1349 
1350 	mm->total_vm += len >> PAGE_SHIFT;
1351 	vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1352 	if (vm_flags & VM_LOCKED) {
1353 		if (!mlock_vma_pages_range(vma, addr, addr + len))
1354 			mm->locked_vm += (len >> PAGE_SHIFT);
1355 	} else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1356 		make_pages_present(addr, addr + len);
1357 
1358 	if (file && uprobe_mmap(vma))
1359 		/* matching probes but cannot insert */
1360 		goto unmap_and_free_vma;
1361 
1362 	return addr;
1363 
1364 unmap_and_free_vma:
1365 	if (correct_wcount)
1366 		atomic_inc(&inode->i_writecount);
1367 	vma->vm_file = NULL;
1368 	fput(file);
1369 
1370 	/* Undo any partial mapping done by a device driver. */
1371 	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1372 	charged = 0;
1373 free_vma:
1374 	kmem_cache_free(vm_area_cachep, vma);
1375 unacct_error:
1376 	if (charged)
1377 		vm_unacct_memory(charged);
1378 	return error;
1379 }
1380 
1381 /* Get an address range which is currently unmapped.
1382  * For shmat() with addr=0.
1383  *
1384  * Ugly calling convention alert:
1385  * Return value with the low bits set means error value,
1386  * ie
1387  *	if (ret & ~PAGE_MASK)
1388  *		error = ret;
1389  *
1390  * This function "knows" that -ENOMEM has the bits set.
1391  */
1392 #ifndef HAVE_ARCH_UNMAPPED_AREA
1393 unsigned long
1394 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1395 		unsigned long len, unsigned long pgoff, unsigned long flags)
1396 {
1397 	struct mm_struct *mm = current->mm;
1398 	struct vm_area_struct *vma;
1399 	unsigned long start_addr;
1400 
1401 	if (len > TASK_SIZE)
1402 		return -ENOMEM;
1403 
1404 	if (flags & MAP_FIXED)
1405 		return addr;
1406 
1407 	if (addr) {
1408 		addr = PAGE_ALIGN(addr);
1409 		vma = find_vma(mm, addr);
1410 		if (TASK_SIZE - len >= addr &&
1411 		    (!vma || addr + len <= vma->vm_start))
1412 			return addr;
1413 	}
1414 	if (len > mm->cached_hole_size) {
1415 	        start_addr = addr = mm->free_area_cache;
1416 	} else {
1417 	        start_addr = addr = TASK_UNMAPPED_BASE;
1418 	        mm->cached_hole_size = 0;
1419 	}
1420 
1421 full_search:
1422 	for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1423 		/* At this point:  (!vma || addr < vma->vm_end). */
1424 		if (TASK_SIZE - len < addr) {
1425 			/*
1426 			 * Start a new search - just in case we missed
1427 			 * some holes.
1428 			 */
1429 			if (start_addr != TASK_UNMAPPED_BASE) {
1430 				addr = TASK_UNMAPPED_BASE;
1431 			        start_addr = addr;
1432 				mm->cached_hole_size = 0;
1433 				goto full_search;
1434 			}
1435 			return -ENOMEM;
1436 		}
1437 		if (!vma || addr + len <= vma->vm_start) {
1438 			/*
1439 			 * Remember the place where we stopped the search:
1440 			 */
1441 			mm->free_area_cache = addr + len;
1442 			return addr;
1443 		}
1444 		if (addr + mm->cached_hole_size < vma->vm_start)
1445 		        mm->cached_hole_size = vma->vm_start - addr;
1446 		addr = vma->vm_end;
1447 	}
1448 }
1449 #endif
1450 
1451 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1452 {
1453 	/*
1454 	 * Is this a new hole at the lowest possible address?
1455 	 */
1456 	if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1457 		mm->free_area_cache = addr;
1458 }
1459 
1460 /*
1461  * This mmap-allocator allocates new areas top-down from below the
1462  * stack's low limit (the base):
1463  */
1464 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1465 unsigned long
1466 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1467 			  const unsigned long len, const unsigned long pgoff,
1468 			  const unsigned long flags)
1469 {
1470 	struct vm_area_struct *vma;
1471 	struct mm_struct *mm = current->mm;
1472 	unsigned long addr = addr0, start_addr;
1473 
1474 	/* requested length too big for entire address space */
1475 	if (len > TASK_SIZE)
1476 		return -ENOMEM;
1477 
1478 	if (flags & MAP_FIXED)
1479 		return addr;
1480 
1481 	/* requesting a specific address */
1482 	if (addr) {
1483 		addr = PAGE_ALIGN(addr);
1484 		vma = find_vma(mm, addr);
1485 		if (TASK_SIZE - len >= addr &&
1486 				(!vma || addr + len <= vma->vm_start))
1487 			return addr;
1488 	}
1489 
1490 	/* check if free_area_cache is useful for us */
1491 	if (len <= mm->cached_hole_size) {
1492  	        mm->cached_hole_size = 0;
1493  		mm->free_area_cache = mm->mmap_base;
1494  	}
1495 
1496 try_again:
1497 	/* either no address requested or can't fit in requested address hole */
1498 	start_addr = addr = mm->free_area_cache;
1499 
1500 	if (addr < len)
1501 		goto fail;
1502 
1503 	addr -= len;
1504 	do {
1505 		/*
1506 		 * Lookup failure means no vma is above this address,
1507 		 * else if new region fits below vma->vm_start,
1508 		 * return with success:
1509 		 */
1510 		vma = find_vma(mm, addr);
1511 		if (!vma || addr+len <= vma->vm_start)
1512 			/* remember the address as a hint for next time */
1513 			return (mm->free_area_cache = addr);
1514 
1515  		/* remember the largest hole we saw so far */
1516  		if (addr + mm->cached_hole_size < vma->vm_start)
1517  		        mm->cached_hole_size = vma->vm_start - addr;
1518 
1519 		/* try just below the current vma->vm_start */
1520 		addr = vma->vm_start-len;
1521 	} while (len < vma->vm_start);
1522 
1523 fail:
1524 	/*
1525 	 * if hint left us with no space for the requested
1526 	 * mapping then try again:
1527 	 *
1528 	 * Note: this is different with the case of bottomup
1529 	 * which does the fully line-search, but we use find_vma
1530 	 * here that causes some holes skipped.
1531 	 */
1532 	if (start_addr != mm->mmap_base) {
1533 		mm->free_area_cache = mm->mmap_base;
1534 		mm->cached_hole_size = 0;
1535 		goto try_again;
1536 	}
1537 
1538 	/*
1539 	 * A failed mmap() very likely causes application failure,
1540 	 * so fall back to the bottom-up function here. This scenario
1541 	 * can happen with large stack limits and large mmap()
1542 	 * allocations.
1543 	 */
1544 	mm->cached_hole_size = ~0UL;
1545   	mm->free_area_cache = TASK_UNMAPPED_BASE;
1546 	addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1547 	/*
1548 	 * Restore the topdown base:
1549 	 */
1550 	mm->free_area_cache = mm->mmap_base;
1551 	mm->cached_hole_size = ~0UL;
1552 
1553 	return addr;
1554 }
1555 #endif
1556 
1557 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1558 {
1559 	/*
1560 	 * Is this a new hole at the highest possible address?
1561 	 */
1562 	if (addr > mm->free_area_cache)
1563 		mm->free_area_cache = addr;
1564 
1565 	/* dont allow allocations above current base */
1566 	if (mm->free_area_cache > mm->mmap_base)
1567 		mm->free_area_cache = mm->mmap_base;
1568 }
1569 
1570 unsigned long
1571 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1572 		unsigned long pgoff, unsigned long flags)
1573 {
1574 	unsigned long (*get_area)(struct file *, unsigned long,
1575 				  unsigned long, unsigned long, unsigned long);
1576 
1577 	unsigned long error = arch_mmap_check(addr, len, flags);
1578 	if (error)
1579 		return error;
1580 
1581 	/* Careful about overflows.. */
1582 	if (len > TASK_SIZE)
1583 		return -ENOMEM;
1584 
1585 	get_area = current->mm->get_unmapped_area;
1586 	if (file && file->f_op && file->f_op->get_unmapped_area)
1587 		get_area = file->f_op->get_unmapped_area;
1588 	addr = get_area(file, addr, len, pgoff, flags);
1589 	if (IS_ERR_VALUE(addr))
1590 		return addr;
1591 
1592 	if (addr > TASK_SIZE - len)
1593 		return -ENOMEM;
1594 	if (addr & ~PAGE_MASK)
1595 		return -EINVAL;
1596 
1597 	addr = arch_rebalance_pgtables(addr, len);
1598 	error = security_mmap_addr(addr);
1599 	return error ? error : addr;
1600 }
1601 
1602 EXPORT_SYMBOL(get_unmapped_area);
1603 
1604 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1605 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1606 {
1607 	struct vm_area_struct *vma = NULL;
1608 
1609 	if (WARN_ON_ONCE(!mm))		/* Remove this in linux-3.6 */
1610 		return NULL;
1611 
1612 	/* Check the cache first. */
1613 	/* (Cache hit rate is typically around 35%.) */
1614 	vma = mm->mmap_cache;
1615 	if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1616 		struct rb_node *rb_node;
1617 
1618 		rb_node = mm->mm_rb.rb_node;
1619 		vma = NULL;
1620 
1621 		while (rb_node) {
1622 			struct vm_area_struct *vma_tmp;
1623 
1624 			vma_tmp = rb_entry(rb_node,
1625 					   struct vm_area_struct, vm_rb);
1626 
1627 			if (vma_tmp->vm_end > addr) {
1628 				vma = vma_tmp;
1629 				if (vma_tmp->vm_start <= addr)
1630 					break;
1631 				rb_node = rb_node->rb_left;
1632 			} else
1633 				rb_node = rb_node->rb_right;
1634 		}
1635 		if (vma)
1636 			mm->mmap_cache = vma;
1637 	}
1638 	return vma;
1639 }
1640 
1641 EXPORT_SYMBOL(find_vma);
1642 
1643 /*
1644  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1645  */
1646 struct vm_area_struct *
1647 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1648 			struct vm_area_struct **pprev)
1649 {
1650 	struct vm_area_struct *vma;
1651 
1652 	vma = find_vma(mm, addr);
1653 	if (vma) {
1654 		*pprev = vma->vm_prev;
1655 	} else {
1656 		struct rb_node *rb_node = mm->mm_rb.rb_node;
1657 		*pprev = NULL;
1658 		while (rb_node) {
1659 			*pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1660 			rb_node = rb_node->rb_right;
1661 		}
1662 	}
1663 	return vma;
1664 }
1665 
1666 /*
1667  * Verify that the stack growth is acceptable and
1668  * update accounting. This is shared with both the
1669  * grow-up and grow-down cases.
1670  */
1671 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1672 {
1673 	struct mm_struct *mm = vma->vm_mm;
1674 	struct rlimit *rlim = current->signal->rlim;
1675 	unsigned long new_start;
1676 
1677 	/* address space limit tests */
1678 	if (!may_expand_vm(mm, grow))
1679 		return -ENOMEM;
1680 
1681 	/* Stack limit test */
1682 	if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1683 		return -ENOMEM;
1684 
1685 	/* mlock limit tests */
1686 	if (vma->vm_flags & VM_LOCKED) {
1687 		unsigned long locked;
1688 		unsigned long limit;
1689 		locked = mm->locked_vm + grow;
1690 		limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1691 		limit >>= PAGE_SHIFT;
1692 		if (locked > limit && !capable(CAP_IPC_LOCK))
1693 			return -ENOMEM;
1694 	}
1695 
1696 	/* Check to ensure the stack will not grow into a hugetlb-only region */
1697 	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1698 			vma->vm_end - size;
1699 	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1700 		return -EFAULT;
1701 
1702 	/*
1703 	 * Overcommit..  This must be the final test, as it will
1704 	 * update security statistics.
1705 	 */
1706 	if (security_vm_enough_memory_mm(mm, grow))
1707 		return -ENOMEM;
1708 
1709 	/* Ok, everything looks good - let it rip */
1710 	mm->total_vm += grow;
1711 	if (vma->vm_flags & VM_LOCKED)
1712 		mm->locked_vm += grow;
1713 	vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1714 	return 0;
1715 }
1716 
1717 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1718 /*
1719  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1720  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1721  */
1722 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1723 {
1724 	int error;
1725 
1726 	if (!(vma->vm_flags & VM_GROWSUP))
1727 		return -EFAULT;
1728 
1729 	/*
1730 	 * We must make sure the anon_vma is allocated
1731 	 * so that the anon_vma locking is not a noop.
1732 	 */
1733 	if (unlikely(anon_vma_prepare(vma)))
1734 		return -ENOMEM;
1735 	vma_lock_anon_vma(vma);
1736 
1737 	/*
1738 	 * vma->vm_start/vm_end cannot change under us because the caller
1739 	 * is required to hold the mmap_sem in read mode.  We need the
1740 	 * anon_vma lock to serialize against concurrent expand_stacks.
1741 	 * Also guard against wrapping around to address 0.
1742 	 */
1743 	if (address < PAGE_ALIGN(address+4))
1744 		address = PAGE_ALIGN(address+4);
1745 	else {
1746 		vma_unlock_anon_vma(vma);
1747 		return -ENOMEM;
1748 	}
1749 	error = 0;
1750 
1751 	/* Somebody else might have raced and expanded it already */
1752 	if (address > vma->vm_end) {
1753 		unsigned long size, grow;
1754 
1755 		size = address - vma->vm_start;
1756 		grow = (address - vma->vm_end) >> PAGE_SHIFT;
1757 
1758 		error = -ENOMEM;
1759 		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1760 			error = acct_stack_growth(vma, size, grow);
1761 			if (!error) {
1762 				vma->vm_end = address;
1763 				perf_event_mmap(vma);
1764 			}
1765 		}
1766 	}
1767 	vma_unlock_anon_vma(vma);
1768 	khugepaged_enter_vma_merge(vma);
1769 	return error;
1770 }
1771 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1772 
1773 /*
1774  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1775  */
1776 int expand_downwards(struct vm_area_struct *vma,
1777 				   unsigned long address)
1778 {
1779 	int error;
1780 
1781 	/*
1782 	 * We must make sure the anon_vma is allocated
1783 	 * so that the anon_vma locking is not a noop.
1784 	 */
1785 	if (unlikely(anon_vma_prepare(vma)))
1786 		return -ENOMEM;
1787 
1788 	address &= PAGE_MASK;
1789 	error = security_mmap_addr(address);
1790 	if (error)
1791 		return error;
1792 
1793 	vma_lock_anon_vma(vma);
1794 
1795 	/*
1796 	 * vma->vm_start/vm_end cannot change under us because the caller
1797 	 * is required to hold the mmap_sem in read mode.  We need the
1798 	 * anon_vma lock to serialize against concurrent expand_stacks.
1799 	 */
1800 
1801 	/* Somebody else might have raced and expanded it already */
1802 	if (address < vma->vm_start) {
1803 		unsigned long size, grow;
1804 
1805 		size = vma->vm_end - address;
1806 		grow = (vma->vm_start - address) >> PAGE_SHIFT;
1807 
1808 		error = -ENOMEM;
1809 		if (grow <= vma->vm_pgoff) {
1810 			error = acct_stack_growth(vma, size, grow);
1811 			if (!error) {
1812 				vma->vm_start = address;
1813 				vma->vm_pgoff -= grow;
1814 				perf_event_mmap(vma);
1815 			}
1816 		}
1817 	}
1818 	vma_unlock_anon_vma(vma);
1819 	khugepaged_enter_vma_merge(vma);
1820 	return error;
1821 }
1822 
1823 #ifdef CONFIG_STACK_GROWSUP
1824 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1825 {
1826 	return expand_upwards(vma, address);
1827 }
1828 
1829 struct vm_area_struct *
1830 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1831 {
1832 	struct vm_area_struct *vma, *prev;
1833 
1834 	addr &= PAGE_MASK;
1835 	vma = find_vma_prev(mm, addr, &prev);
1836 	if (vma && (vma->vm_start <= addr))
1837 		return vma;
1838 	if (!prev || expand_stack(prev, addr))
1839 		return NULL;
1840 	if (prev->vm_flags & VM_LOCKED) {
1841 		mlock_vma_pages_range(prev, addr, prev->vm_end);
1842 	}
1843 	return prev;
1844 }
1845 #else
1846 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1847 {
1848 	return expand_downwards(vma, address);
1849 }
1850 
1851 struct vm_area_struct *
1852 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1853 {
1854 	struct vm_area_struct * vma;
1855 	unsigned long start;
1856 
1857 	addr &= PAGE_MASK;
1858 	vma = find_vma(mm,addr);
1859 	if (!vma)
1860 		return NULL;
1861 	if (vma->vm_start <= addr)
1862 		return vma;
1863 	if (!(vma->vm_flags & VM_GROWSDOWN))
1864 		return NULL;
1865 	start = vma->vm_start;
1866 	if (expand_stack(vma, addr))
1867 		return NULL;
1868 	if (vma->vm_flags & VM_LOCKED) {
1869 		mlock_vma_pages_range(vma, addr, start);
1870 	}
1871 	return vma;
1872 }
1873 #endif
1874 
1875 /*
1876  * Ok - we have the memory areas we should free on the vma list,
1877  * so release them, and do the vma updates.
1878  *
1879  * Called with the mm semaphore held.
1880  */
1881 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1882 {
1883 	unsigned long nr_accounted = 0;
1884 
1885 	/* Update high watermark before we lower total_vm */
1886 	update_hiwater_vm(mm);
1887 	do {
1888 		long nrpages = vma_pages(vma);
1889 
1890 		if (vma->vm_flags & VM_ACCOUNT)
1891 			nr_accounted += nrpages;
1892 		mm->total_vm -= nrpages;
1893 		vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1894 		vma = remove_vma(vma);
1895 	} while (vma);
1896 	vm_unacct_memory(nr_accounted);
1897 	validate_mm(mm);
1898 }
1899 
1900 /*
1901  * Get rid of page table information in the indicated region.
1902  *
1903  * Called with the mm semaphore held.
1904  */
1905 static void unmap_region(struct mm_struct *mm,
1906 		struct vm_area_struct *vma, struct vm_area_struct *prev,
1907 		unsigned long start, unsigned long end)
1908 {
1909 	struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1910 	struct mmu_gather tlb;
1911 
1912 	lru_add_drain();
1913 	tlb_gather_mmu(&tlb, mm, 0);
1914 	update_hiwater_rss(mm);
1915 	unmap_vmas(&tlb, vma, start, end);
1916 	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1917 				 next ? next->vm_start : 0);
1918 	tlb_finish_mmu(&tlb, start, end);
1919 }
1920 
1921 /*
1922  * Create a list of vma's touched by the unmap, removing them from the mm's
1923  * vma list as we go..
1924  */
1925 static void
1926 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1927 	struct vm_area_struct *prev, unsigned long end)
1928 {
1929 	struct vm_area_struct **insertion_point;
1930 	struct vm_area_struct *tail_vma = NULL;
1931 	unsigned long addr;
1932 
1933 	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1934 	vma->vm_prev = NULL;
1935 	do {
1936 		rb_erase(&vma->vm_rb, &mm->mm_rb);
1937 		mm->map_count--;
1938 		tail_vma = vma;
1939 		vma = vma->vm_next;
1940 	} while (vma && vma->vm_start < end);
1941 	*insertion_point = vma;
1942 	if (vma)
1943 		vma->vm_prev = prev;
1944 	tail_vma->vm_next = NULL;
1945 	if (mm->unmap_area == arch_unmap_area)
1946 		addr = prev ? prev->vm_end : mm->mmap_base;
1947 	else
1948 		addr = vma ?  vma->vm_start : mm->mmap_base;
1949 	mm->unmap_area(mm, addr);
1950 	mm->mmap_cache = NULL;		/* Kill the cache. */
1951 }
1952 
1953 /*
1954  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1955  * munmap path where it doesn't make sense to fail.
1956  */
1957 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1958 	      unsigned long addr, int new_below)
1959 {
1960 	struct mempolicy *pol;
1961 	struct vm_area_struct *new;
1962 	int err = -ENOMEM;
1963 
1964 	if (is_vm_hugetlb_page(vma) && (addr &
1965 					~(huge_page_mask(hstate_vma(vma)))))
1966 		return -EINVAL;
1967 
1968 	new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1969 	if (!new)
1970 		goto out_err;
1971 
1972 	/* most fields are the same, copy all, and then fixup */
1973 	*new = *vma;
1974 
1975 	INIT_LIST_HEAD(&new->anon_vma_chain);
1976 
1977 	if (new_below)
1978 		new->vm_end = addr;
1979 	else {
1980 		new->vm_start = addr;
1981 		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1982 	}
1983 
1984 	pol = mpol_dup(vma_policy(vma));
1985 	if (IS_ERR(pol)) {
1986 		err = PTR_ERR(pol);
1987 		goto out_free_vma;
1988 	}
1989 	vma_set_policy(new, pol);
1990 
1991 	if (anon_vma_clone(new, vma))
1992 		goto out_free_mpol;
1993 
1994 	if (new->vm_file) {
1995 		get_file(new->vm_file);
1996 		if (vma->vm_flags & VM_EXECUTABLE)
1997 			added_exe_file_vma(mm);
1998 	}
1999 
2000 	if (new->vm_ops && new->vm_ops->open)
2001 		new->vm_ops->open(new);
2002 
2003 	if (new_below)
2004 		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2005 			((addr - new->vm_start) >> PAGE_SHIFT), new);
2006 	else
2007 		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2008 
2009 	/* Success. */
2010 	if (!err)
2011 		return 0;
2012 
2013 	/* Clean everything up if vma_adjust failed. */
2014 	if (new->vm_ops && new->vm_ops->close)
2015 		new->vm_ops->close(new);
2016 	if (new->vm_file) {
2017 		if (vma->vm_flags & VM_EXECUTABLE)
2018 			removed_exe_file_vma(mm);
2019 		fput(new->vm_file);
2020 	}
2021 	unlink_anon_vmas(new);
2022  out_free_mpol:
2023 	mpol_put(pol);
2024  out_free_vma:
2025 	kmem_cache_free(vm_area_cachep, new);
2026  out_err:
2027 	return err;
2028 }
2029 
2030 /*
2031  * Split a vma into two pieces at address 'addr', a new vma is allocated
2032  * either for the first part or the tail.
2033  */
2034 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2035 	      unsigned long addr, int new_below)
2036 {
2037 	if (mm->map_count >= sysctl_max_map_count)
2038 		return -ENOMEM;
2039 
2040 	return __split_vma(mm, vma, addr, new_below);
2041 }
2042 
2043 /* Munmap is split into 2 main parts -- this part which finds
2044  * what needs doing, and the areas themselves, which do the
2045  * work.  This now handles partial unmappings.
2046  * Jeremy Fitzhardinge <jeremy@goop.org>
2047  */
2048 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2049 {
2050 	unsigned long end;
2051 	struct vm_area_struct *vma, *prev, *last;
2052 
2053 	if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2054 		return -EINVAL;
2055 
2056 	if ((len = PAGE_ALIGN(len)) == 0)
2057 		return -EINVAL;
2058 
2059 	/* Find the first overlapping VMA */
2060 	vma = find_vma(mm, start);
2061 	if (!vma)
2062 		return 0;
2063 	prev = vma->vm_prev;
2064 	/* we have  start < vma->vm_end  */
2065 
2066 	/* if it doesn't overlap, we have nothing.. */
2067 	end = start + len;
2068 	if (vma->vm_start >= end)
2069 		return 0;
2070 
2071 	/*
2072 	 * If we need to split any vma, do it now to save pain later.
2073 	 *
2074 	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2075 	 * unmapped vm_area_struct will remain in use: so lower split_vma
2076 	 * places tmp vma above, and higher split_vma places tmp vma below.
2077 	 */
2078 	if (start > vma->vm_start) {
2079 		int error;
2080 
2081 		/*
2082 		 * Make sure that map_count on return from munmap() will
2083 		 * not exceed its limit; but let map_count go just above
2084 		 * its limit temporarily, to help free resources as expected.
2085 		 */
2086 		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2087 			return -ENOMEM;
2088 
2089 		error = __split_vma(mm, vma, start, 0);
2090 		if (error)
2091 			return error;
2092 		prev = vma;
2093 	}
2094 
2095 	/* Does it split the last one? */
2096 	last = find_vma(mm, end);
2097 	if (last && end > last->vm_start) {
2098 		int error = __split_vma(mm, last, end, 1);
2099 		if (error)
2100 			return error;
2101 	}
2102 	vma = prev? prev->vm_next: mm->mmap;
2103 
2104 	/*
2105 	 * unlock any mlock()ed ranges before detaching vmas
2106 	 */
2107 	if (mm->locked_vm) {
2108 		struct vm_area_struct *tmp = vma;
2109 		while (tmp && tmp->vm_start < end) {
2110 			if (tmp->vm_flags & VM_LOCKED) {
2111 				mm->locked_vm -= vma_pages(tmp);
2112 				munlock_vma_pages_all(tmp);
2113 			}
2114 			tmp = tmp->vm_next;
2115 		}
2116 	}
2117 
2118 	/*
2119 	 * Remove the vma's, and unmap the actual pages
2120 	 */
2121 	detach_vmas_to_be_unmapped(mm, vma, prev, end);
2122 	unmap_region(mm, vma, prev, start, end);
2123 
2124 	/* Fix up all other VM information */
2125 	remove_vma_list(mm, vma);
2126 
2127 	return 0;
2128 }
2129 
2130 int vm_munmap(unsigned long start, size_t len)
2131 {
2132 	int ret;
2133 	struct mm_struct *mm = current->mm;
2134 
2135 	down_write(&mm->mmap_sem);
2136 	ret = do_munmap(mm, start, len);
2137 	up_write(&mm->mmap_sem);
2138 	return ret;
2139 }
2140 EXPORT_SYMBOL(vm_munmap);
2141 
2142 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2143 {
2144 	profile_munmap(addr);
2145 	return vm_munmap(addr, len);
2146 }
2147 
2148 static inline void verify_mm_writelocked(struct mm_struct *mm)
2149 {
2150 #ifdef CONFIG_DEBUG_VM
2151 	if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2152 		WARN_ON(1);
2153 		up_read(&mm->mmap_sem);
2154 	}
2155 #endif
2156 }
2157 
2158 /*
2159  *  this is really a simplified "do_mmap".  it only handles
2160  *  anonymous maps.  eventually we may be able to do some
2161  *  brk-specific accounting here.
2162  */
2163 static unsigned long do_brk(unsigned long addr, unsigned long len)
2164 {
2165 	struct mm_struct * mm = current->mm;
2166 	struct vm_area_struct * vma, * prev;
2167 	unsigned long flags;
2168 	struct rb_node ** rb_link, * rb_parent;
2169 	pgoff_t pgoff = addr >> PAGE_SHIFT;
2170 	int error;
2171 
2172 	len = PAGE_ALIGN(len);
2173 	if (!len)
2174 		return addr;
2175 
2176 	flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2177 
2178 	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2179 	if (error & ~PAGE_MASK)
2180 		return error;
2181 
2182 	/*
2183 	 * mlock MCL_FUTURE?
2184 	 */
2185 	if (mm->def_flags & VM_LOCKED) {
2186 		unsigned long locked, lock_limit;
2187 		locked = len >> PAGE_SHIFT;
2188 		locked += mm->locked_vm;
2189 		lock_limit = rlimit(RLIMIT_MEMLOCK);
2190 		lock_limit >>= PAGE_SHIFT;
2191 		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2192 			return -EAGAIN;
2193 	}
2194 
2195 	/*
2196 	 * mm->mmap_sem is required to protect against another thread
2197 	 * changing the mappings in case we sleep.
2198 	 */
2199 	verify_mm_writelocked(mm);
2200 
2201 	/*
2202 	 * Clear old maps.  this also does some error checking for us
2203 	 */
2204  munmap_back:
2205 	vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2206 	if (vma && vma->vm_start < addr + len) {
2207 		if (do_munmap(mm, addr, len))
2208 			return -ENOMEM;
2209 		goto munmap_back;
2210 	}
2211 
2212 	/* Check against address space limits *after* clearing old maps... */
2213 	if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2214 		return -ENOMEM;
2215 
2216 	if (mm->map_count > sysctl_max_map_count)
2217 		return -ENOMEM;
2218 
2219 	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2220 		return -ENOMEM;
2221 
2222 	/* Can we just expand an old private anonymous mapping? */
2223 	vma = vma_merge(mm, prev, addr, addr + len, flags,
2224 					NULL, NULL, pgoff, NULL);
2225 	if (vma)
2226 		goto out;
2227 
2228 	/*
2229 	 * create a vma struct for an anonymous mapping
2230 	 */
2231 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2232 	if (!vma) {
2233 		vm_unacct_memory(len >> PAGE_SHIFT);
2234 		return -ENOMEM;
2235 	}
2236 
2237 	INIT_LIST_HEAD(&vma->anon_vma_chain);
2238 	vma->vm_mm = mm;
2239 	vma->vm_start = addr;
2240 	vma->vm_end = addr + len;
2241 	vma->vm_pgoff = pgoff;
2242 	vma->vm_flags = flags;
2243 	vma->vm_page_prot = vm_get_page_prot(flags);
2244 	vma_link(mm, vma, prev, rb_link, rb_parent);
2245 out:
2246 	perf_event_mmap(vma);
2247 	mm->total_vm += len >> PAGE_SHIFT;
2248 	if (flags & VM_LOCKED) {
2249 		if (!mlock_vma_pages_range(vma, addr, addr + len))
2250 			mm->locked_vm += (len >> PAGE_SHIFT);
2251 	}
2252 	return addr;
2253 }
2254 
2255 unsigned long vm_brk(unsigned long addr, unsigned long len)
2256 {
2257 	struct mm_struct *mm = current->mm;
2258 	unsigned long ret;
2259 
2260 	down_write(&mm->mmap_sem);
2261 	ret = do_brk(addr, len);
2262 	up_write(&mm->mmap_sem);
2263 	return ret;
2264 }
2265 EXPORT_SYMBOL(vm_brk);
2266 
2267 /* Release all mmaps. */
2268 void exit_mmap(struct mm_struct *mm)
2269 {
2270 	struct mmu_gather tlb;
2271 	struct vm_area_struct *vma;
2272 	unsigned long nr_accounted = 0;
2273 
2274 	/* mm's last user has gone, and its about to be pulled down */
2275 	mmu_notifier_release(mm);
2276 
2277 	if (mm->locked_vm) {
2278 		vma = mm->mmap;
2279 		while (vma) {
2280 			if (vma->vm_flags & VM_LOCKED)
2281 				munlock_vma_pages_all(vma);
2282 			vma = vma->vm_next;
2283 		}
2284 	}
2285 
2286 	arch_exit_mmap(mm);
2287 
2288 	vma = mm->mmap;
2289 	if (!vma)	/* Can happen if dup_mmap() received an OOM */
2290 		return;
2291 
2292 	lru_add_drain();
2293 	flush_cache_mm(mm);
2294 	tlb_gather_mmu(&tlb, mm, 1);
2295 	/* update_hiwater_rss(mm) here? but nobody should be looking */
2296 	/* Use -1 here to ensure all VMAs in the mm are unmapped */
2297 	unmap_vmas(&tlb, vma, 0, -1);
2298 
2299 	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2300 	tlb_finish_mmu(&tlb, 0, -1);
2301 
2302 	/*
2303 	 * Walk the list again, actually closing and freeing it,
2304 	 * with preemption enabled, without holding any MM locks.
2305 	 */
2306 	while (vma) {
2307 		if (vma->vm_flags & VM_ACCOUNT)
2308 			nr_accounted += vma_pages(vma);
2309 		vma = remove_vma(vma);
2310 	}
2311 	vm_unacct_memory(nr_accounted);
2312 
2313 	BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2314 }
2315 
2316 /* Insert vm structure into process list sorted by address
2317  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2318  * then i_mmap_mutex is taken here.
2319  */
2320 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2321 {
2322 	struct vm_area_struct * __vma, * prev;
2323 	struct rb_node ** rb_link, * rb_parent;
2324 
2325 	/*
2326 	 * The vm_pgoff of a purely anonymous vma should be irrelevant
2327 	 * until its first write fault, when page's anon_vma and index
2328 	 * are set.  But now set the vm_pgoff it will almost certainly
2329 	 * end up with (unless mremap moves it elsewhere before that
2330 	 * first wfault), so /proc/pid/maps tells a consistent story.
2331 	 *
2332 	 * By setting it to reflect the virtual start address of the
2333 	 * vma, merges and splits can happen in a seamless way, just
2334 	 * using the existing file pgoff checks and manipulations.
2335 	 * Similarly in do_mmap_pgoff and in do_brk.
2336 	 */
2337 	if (!vma->vm_file) {
2338 		BUG_ON(vma->anon_vma);
2339 		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2340 	}
2341 	__vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2342 	if (__vma && __vma->vm_start < vma->vm_end)
2343 		return -ENOMEM;
2344 	if ((vma->vm_flags & VM_ACCOUNT) &&
2345 	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
2346 		return -ENOMEM;
2347 
2348 	if (vma->vm_file && uprobe_mmap(vma))
2349 		return -EINVAL;
2350 
2351 	vma_link(mm, vma, prev, rb_link, rb_parent);
2352 	return 0;
2353 }
2354 
2355 /*
2356  * Copy the vma structure to a new location in the same mm,
2357  * prior to moving page table entries, to effect an mremap move.
2358  */
2359 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2360 	unsigned long addr, unsigned long len, pgoff_t pgoff)
2361 {
2362 	struct vm_area_struct *vma = *vmap;
2363 	unsigned long vma_start = vma->vm_start;
2364 	struct mm_struct *mm = vma->vm_mm;
2365 	struct vm_area_struct *new_vma, *prev;
2366 	struct rb_node **rb_link, *rb_parent;
2367 	struct mempolicy *pol;
2368 	bool faulted_in_anon_vma = true;
2369 
2370 	/*
2371 	 * If anonymous vma has not yet been faulted, update new pgoff
2372 	 * to match new location, to increase its chance of merging.
2373 	 */
2374 	if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2375 		pgoff = addr >> PAGE_SHIFT;
2376 		faulted_in_anon_vma = false;
2377 	}
2378 
2379 	find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2380 	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2381 			vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2382 	if (new_vma) {
2383 		/*
2384 		 * Source vma may have been merged into new_vma
2385 		 */
2386 		if (unlikely(vma_start >= new_vma->vm_start &&
2387 			     vma_start < new_vma->vm_end)) {
2388 			/*
2389 			 * The only way we can get a vma_merge with
2390 			 * self during an mremap is if the vma hasn't
2391 			 * been faulted in yet and we were allowed to
2392 			 * reset the dst vma->vm_pgoff to the
2393 			 * destination address of the mremap to allow
2394 			 * the merge to happen. mremap must change the
2395 			 * vm_pgoff linearity between src and dst vmas
2396 			 * (in turn preventing a vma_merge) to be
2397 			 * safe. It is only safe to keep the vm_pgoff
2398 			 * linear if there are no pages mapped yet.
2399 			 */
2400 			VM_BUG_ON(faulted_in_anon_vma);
2401 			*vmap = new_vma;
2402 		} else
2403 			anon_vma_moveto_tail(new_vma);
2404 	} else {
2405 		new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2406 		if (new_vma) {
2407 			*new_vma = *vma;
2408 			pol = mpol_dup(vma_policy(vma));
2409 			if (IS_ERR(pol))
2410 				goto out_free_vma;
2411 			INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2412 			if (anon_vma_clone(new_vma, vma))
2413 				goto out_free_mempol;
2414 			vma_set_policy(new_vma, pol);
2415 			new_vma->vm_start = addr;
2416 			new_vma->vm_end = addr + len;
2417 			new_vma->vm_pgoff = pgoff;
2418 			if (new_vma->vm_file) {
2419 				get_file(new_vma->vm_file);
2420 
2421 				if (uprobe_mmap(new_vma))
2422 					goto out_free_mempol;
2423 
2424 				if (vma->vm_flags & VM_EXECUTABLE)
2425 					added_exe_file_vma(mm);
2426 			}
2427 			if (new_vma->vm_ops && new_vma->vm_ops->open)
2428 				new_vma->vm_ops->open(new_vma);
2429 			vma_link(mm, new_vma, prev, rb_link, rb_parent);
2430 		}
2431 	}
2432 	return new_vma;
2433 
2434  out_free_mempol:
2435 	mpol_put(pol);
2436  out_free_vma:
2437 	kmem_cache_free(vm_area_cachep, new_vma);
2438 	return NULL;
2439 }
2440 
2441 /*
2442  * Return true if the calling process may expand its vm space by the passed
2443  * number of pages
2444  */
2445 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2446 {
2447 	unsigned long cur = mm->total_vm;	/* pages */
2448 	unsigned long lim;
2449 
2450 	lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2451 
2452 	if (cur + npages > lim)
2453 		return 0;
2454 	return 1;
2455 }
2456 
2457 
2458 static int special_mapping_fault(struct vm_area_struct *vma,
2459 				struct vm_fault *vmf)
2460 {
2461 	pgoff_t pgoff;
2462 	struct page **pages;
2463 
2464 	/*
2465 	 * special mappings have no vm_file, and in that case, the mm
2466 	 * uses vm_pgoff internally. So we have to subtract it from here.
2467 	 * We are allowed to do this because we are the mm; do not copy
2468 	 * this code into drivers!
2469 	 */
2470 	pgoff = vmf->pgoff - vma->vm_pgoff;
2471 
2472 	for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2473 		pgoff--;
2474 
2475 	if (*pages) {
2476 		struct page *page = *pages;
2477 		get_page(page);
2478 		vmf->page = page;
2479 		return 0;
2480 	}
2481 
2482 	return VM_FAULT_SIGBUS;
2483 }
2484 
2485 /*
2486  * Having a close hook prevents vma merging regardless of flags.
2487  */
2488 static void special_mapping_close(struct vm_area_struct *vma)
2489 {
2490 }
2491 
2492 static const struct vm_operations_struct special_mapping_vmops = {
2493 	.close = special_mapping_close,
2494 	.fault = special_mapping_fault,
2495 };
2496 
2497 /*
2498  * Called with mm->mmap_sem held for writing.
2499  * Insert a new vma covering the given region, with the given flags.
2500  * Its pages are supplied by the given array of struct page *.
2501  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2502  * The region past the last page supplied will always produce SIGBUS.
2503  * The array pointer and the pages it points to are assumed to stay alive
2504  * for as long as this mapping might exist.
2505  */
2506 int install_special_mapping(struct mm_struct *mm,
2507 			    unsigned long addr, unsigned long len,
2508 			    unsigned long vm_flags, struct page **pages)
2509 {
2510 	int ret;
2511 	struct vm_area_struct *vma;
2512 
2513 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2514 	if (unlikely(vma == NULL))
2515 		return -ENOMEM;
2516 
2517 	INIT_LIST_HEAD(&vma->anon_vma_chain);
2518 	vma->vm_mm = mm;
2519 	vma->vm_start = addr;
2520 	vma->vm_end = addr + len;
2521 
2522 	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2523 	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2524 
2525 	vma->vm_ops = &special_mapping_vmops;
2526 	vma->vm_private_data = pages;
2527 
2528 	ret = insert_vm_struct(mm, vma);
2529 	if (ret)
2530 		goto out;
2531 
2532 	mm->total_vm += len >> PAGE_SHIFT;
2533 
2534 	perf_event_mmap(vma);
2535 
2536 	return 0;
2537 
2538 out:
2539 	kmem_cache_free(vm_area_cachep, vma);
2540 	return ret;
2541 }
2542 
2543 static DEFINE_MUTEX(mm_all_locks_mutex);
2544 
2545 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2546 {
2547 	if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2548 		/*
2549 		 * The LSB of head.next can't change from under us
2550 		 * because we hold the mm_all_locks_mutex.
2551 		 */
2552 		mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2553 		/*
2554 		 * We can safely modify head.next after taking the
2555 		 * anon_vma->root->mutex. If some other vma in this mm shares
2556 		 * the same anon_vma we won't take it again.
2557 		 *
2558 		 * No need of atomic instructions here, head.next
2559 		 * can't change from under us thanks to the
2560 		 * anon_vma->root->mutex.
2561 		 */
2562 		if (__test_and_set_bit(0, (unsigned long *)
2563 				       &anon_vma->root->head.next))
2564 			BUG();
2565 	}
2566 }
2567 
2568 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2569 {
2570 	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2571 		/*
2572 		 * AS_MM_ALL_LOCKS can't change from under us because
2573 		 * we hold the mm_all_locks_mutex.
2574 		 *
2575 		 * Operations on ->flags have to be atomic because
2576 		 * even if AS_MM_ALL_LOCKS is stable thanks to the
2577 		 * mm_all_locks_mutex, there may be other cpus
2578 		 * changing other bitflags in parallel to us.
2579 		 */
2580 		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2581 			BUG();
2582 		mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2583 	}
2584 }
2585 
2586 /*
2587  * This operation locks against the VM for all pte/vma/mm related
2588  * operations that could ever happen on a certain mm. This includes
2589  * vmtruncate, try_to_unmap, and all page faults.
2590  *
2591  * The caller must take the mmap_sem in write mode before calling
2592  * mm_take_all_locks(). The caller isn't allowed to release the
2593  * mmap_sem until mm_drop_all_locks() returns.
2594  *
2595  * mmap_sem in write mode is required in order to block all operations
2596  * that could modify pagetables and free pages without need of
2597  * altering the vma layout (for example populate_range() with
2598  * nonlinear vmas). It's also needed in write mode to avoid new
2599  * anon_vmas to be associated with existing vmas.
2600  *
2601  * A single task can't take more than one mm_take_all_locks() in a row
2602  * or it would deadlock.
2603  *
2604  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2605  * mapping->flags avoid to take the same lock twice, if more than one
2606  * vma in this mm is backed by the same anon_vma or address_space.
2607  *
2608  * We can take all the locks in random order because the VM code
2609  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2610  * takes more than one of them in a row. Secondly we're protected
2611  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2612  *
2613  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2614  * that may have to take thousand of locks.
2615  *
2616  * mm_take_all_locks() can fail if it's interrupted by signals.
2617  */
2618 int mm_take_all_locks(struct mm_struct *mm)
2619 {
2620 	struct vm_area_struct *vma;
2621 	struct anon_vma_chain *avc;
2622 
2623 	BUG_ON(down_read_trylock(&mm->mmap_sem));
2624 
2625 	mutex_lock(&mm_all_locks_mutex);
2626 
2627 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
2628 		if (signal_pending(current))
2629 			goto out_unlock;
2630 		if (vma->vm_file && vma->vm_file->f_mapping)
2631 			vm_lock_mapping(mm, vma->vm_file->f_mapping);
2632 	}
2633 
2634 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
2635 		if (signal_pending(current))
2636 			goto out_unlock;
2637 		if (vma->anon_vma)
2638 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2639 				vm_lock_anon_vma(mm, avc->anon_vma);
2640 	}
2641 
2642 	return 0;
2643 
2644 out_unlock:
2645 	mm_drop_all_locks(mm);
2646 	return -EINTR;
2647 }
2648 
2649 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2650 {
2651 	if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2652 		/*
2653 		 * The LSB of head.next can't change to 0 from under
2654 		 * us because we hold the mm_all_locks_mutex.
2655 		 *
2656 		 * We must however clear the bitflag before unlocking
2657 		 * the vma so the users using the anon_vma->head will
2658 		 * never see our bitflag.
2659 		 *
2660 		 * No need of atomic instructions here, head.next
2661 		 * can't change from under us until we release the
2662 		 * anon_vma->root->mutex.
2663 		 */
2664 		if (!__test_and_clear_bit(0, (unsigned long *)
2665 					  &anon_vma->root->head.next))
2666 			BUG();
2667 		anon_vma_unlock(anon_vma);
2668 	}
2669 }
2670 
2671 static void vm_unlock_mapping(struct address_space *mapping)
2672 {
2673 	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2674 		/*
2675 		 * AS_MM_ALL_LOCKS can't change to 0 from under us
2676 		 * because we hold the mm_all_locks_mutex.
2677 		 */
2678 		mutex_unlock(&mapping->i_mmap_mutex);
2679 		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2680 					&mapping->flags))
2681 			BUG();
2682 	}
2683 }
2684 
2685 /*
2686  * The mmap_sem cannot be released by the caller until
2687  * mm_drop_all_locks() returns.
2688  */
2689 void mm_drop_all_locks(struct mm_struct *mm)
2690 {
2691 	struct vm_area_struct *vma;
2692 	struct anon_vma_chain *avc;
2693 
2694 	BUG_ON(down_read_trylock(&mm->mmap_sem));
2695 	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2696 
2697 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
2698 		if (vma->anon_vma)
2699 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2700 				vm_unlock_anon_vma(avc->anon_vma);
2701 		if (vma->vm_file && vma->vm_file->f_mapping)
2702 			vm_unlock_mapping(vma->vm_file->f_mapping);
2703 	}
2704 
2705 	mutex_unlock(&mm_all_locks_mutex);
2706 }
2707 
2708 /*
2709  * initialise the VMA slab
2710  */
2711 void __init mmap_init(void)
2712 {
2713 	int ret;
2714 
2715 	ret = percpu_counter_init(&vm_committed_as, 0);
2716 	VM_BUG_ON(ret);
2717 }
2718