xref: /linux/mm/mmap.c (revision 827634added7f38b7d724cab1dccdb2b004c13c3)
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
14 #include <linux/mm.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
24 #include <linux/fs.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 
45 #include <asm/uaccess.h>
46 #include <asm/cacheflush.h>
47 #include <asm/tlb.h>
48 #include <asm/mmu_context.h>
49 
50 #include "internal.h"
51 
52 #ifndef arch_mmap_check
53 #define arch_mmap_check(addr, len, flags)	(0)
54 #endif
55 
56 #ifndef arch_rebalance_pgtables
57 #define arch_rebalance_pgtables(addr, len)		(addr)
58 #endif
59 
60 static void unmap_region(struct mm_struct *mm,
61 		struct vm_area_struct *vma, struct vm_area_struct *prev,
62 		unsigned long start, unsigned long end);
63 
64 /* description of effects of mapping type and prot in current implementation.
65  * this is due to the limited x86 page protection hardware.  The expected
66  * behavior is in parens:
67  *
68  * map_type	prot
69  *		PROT_NONE	PROT_READ	PROT_WRITE	PROT_EXEC
70  * MAP_SHARED	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
71  *		w: (no) no	w: (no) no	w: (yes) yes	w: (no) no
72  *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
73  *
74  * MAP_PRIVATE	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
75  *		w: (no) no	w: (no) no	w: (copy) copy	w: (no) no
76  *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
77  *
78  */
79 pgprot_t protection_map[16] = {
80 	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
81 	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
82 };
83 
84 pgprot_t vm_get_page_prot(unsigned long vm_flags)
85 {
86 	return __pgprot(pgprot_val(protection_map[vm_flags &
87 				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
88 			pgprot_val(arch_vm_get_page_prot(vm_flags)));
89 }
90 EXPORT_SYMBOL(vm_get_page_prot);
91 
92 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
93 {
94 	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
95 }
96 
97 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
98 void vma_set_page_prot(struct vm_area_struct *vma)
99 {
100 	unsigned long vm_flags = vma->vm_flags;
101 
102 	vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
103 	if (vma_wants_writenotify(vma)) {
104 		vm_flags &= ~VM_SHARED;
105 		vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
106 						     vm_flags);
107 	}
108 }
109 
110 
111 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
112 int sysctl_overcommit_ratio __read_mostly = 50;	/* default is 50% */
113 unsigned long sysctl_overcommit_kbytes __read_mostly;
114 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
115 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
116 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
117 /*
118  * Make sure vm_committed_as in one cacheline and not cacheline shared with
119  * other variables. It can be updated by several CPUs frequently.
120  */
121 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
122 
123 /*
124  * The global memory commitment made in the system can be a metric
125  * that can be used to drive ballooning decisions when Linux is hosted
126  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
127  * balancing memory across competing virtual machines that are hosted.
128  * Several metrics drive this policy engine including the guest reported
129  * memory commitment.
130  */
131 unsigned long vm_memory_committed(void)
132 {
133 	return percpu_counter_read_positive(&vm_committed_as);
134 }
135 EXPORT_SYMBOL_GPL(vm_memory_committed);
136 
137 /*
138  * Check that a process has enough memory to allocate a new virtual
139  * mapping. 0 means there is enough memory for the allocation to
140  * succeed and -ENOMEM implies there is not.
141  *
142  * We currently support three overcommit policies, which are set via the
143  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
144  *
145  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
146  * Additional code 2002 Jul 20 by Robert Love.
147  *
148  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
149  *
150  * Note this is a helper function intended to be used by LSMs which
151  * wish to use this logic.
152  */
153 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
154 {
155 	long free, allowed, reserve;
156 
157 	VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
158 			-(s64)vm_committed_as_batch * num_online_cpus(),
159 			"memory commitment underflow");
160 
161 	vm_acct_memory(pages);
162 
163 	/*
164 	 * Sometimes we want to use more memory than we have
165 	 */
166 	if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
167 		return 0;
168 
169 	if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
170 		free = global_page_state(NR_FREE_PAGES);
171 		free += global_page_state(NR_FILE_PAGES);
172 
173 		/*
174 		 * shmem pages shouldn't be counted as free in this
175 		 * case, they can't be purged, only swapped out, and
176 		 * that won't affect the overall amount of available
177 		 * memory in the system.
178 		 */
179 		free -= global_page_state(NR_SHMEM);
180 
181 		free += get_nr_swap_pages();
182 
183 		/*
184 		 * Any slabs which are created with the
185 		 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
186 		 * which are reclaimable, under pressure.  The dentry
187 		 * cache and most inode caches should fall into this
188 		 */
189 		free += global_page_state(NR_SLAB_RECLAIMABLE);
190 
191 		/*
192 		 * Leave reserved pages. The pages are not for anonymous pages.
193 		 */
194 		if (free <= totalreserve_pages)
195 			goto error;
196 		else
197 			free -= totalreserve_pages;
198 
199 		/*
200 		 * Reserve some for root
201 		 */
202 		if (!cap_sys_admin)
203 			free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
204 
205 		if (free > pages)
206 			return 0;
207 
208 		goto error;
209 	}
210 
211 	allowed = vm_commit_limit();
212 	/*
213 	 * Reserve some for root
214 	 */
215 	if (!cap_sys_admin)
216 		allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
217 
218 	/*
219 	 * Don't let a single process grow so big a user can't recover
220 	 */
221 	if (mm) {
222 		reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
223 		allowed -= min_t(long, mm->total_vm / 32, reserve);
224 	}
225 
226 	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
227 		return 0;
228 error:
229 	vm_unacct_memory(pages);
230 
231 	return -ENOMEM;
232 }
233 
234 /*
235  * Requires inode->i_mapping->i_mmap_rwsem
236  */
237 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
238 		struct file *file, struct address_space *mapping)
239 {
240 	if (vma->vm_flags & VM_DENYWRITE)
241 		atomic_inc(&file_inode(file)->i_writecount);
242 	if (vma->vm_flags & VM_SHARED)
243 		mapping_unmap_writable(mapping);
244 
245 	flush_dcache_mmap_lock(mapping);
246 	vma_interval_tree_remove(vma, &mapping->i_mmap);
247 	flush_dcache_mmap_unlock(mapping);
248 }
249 
250 /*
251  * Unlink a file-based vm structure from its interval tree, to hide
252  * vma from rmap and vmtruncate before freeing its page tables.
253  */
254 void unlink_file_vma(struct vm_area_struct *vma)
255 {
256 	struct file *file = vma->vm_file;
257 
258 	if (file) {
259 		struct address_space *mapping = file->f_mapping;
260 		i_mmap_lock_write(mapping);
261 		__remove_shared_vm_struct(vma, file, mapping);
262 		i_mmap_unlock_write(mapping);
263 	}
264 }
265 
266 /*
267  * Close a vm structure and free it, returning the next.
268  */
269 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
270 {
271 	struct vm_area_struct *next = vma->vm_next;
272 
273 	might_sleep();
274 	if (vma->vm_ops && vma->vm_ops->close)
275 		vma->vm_ops->close(vma);
276 	if (vma->vm_file)
277 		fput(vma->vm_file);
278 	mpol_put(vma_policy(vma));
279 	kmem_cache_free(vm_area_cachep, vma);
280 	return next;
281 }
282 
283 static unsigned long do_brk(unsigned long addr, unsigned long len);
284 
285 SYSCALL_DEFINE1(brk, unsigned long, brk)
286 {
287 	unsigned long retval;
288 	unsigned long newbrk, oldbrk;
289 	struct mm_struct *mm = current->mm;
290 	unsigned long min_brk;
291 	bool populate;
292 
293 	down_write(&mm->mmap_sem);
294 
295 #ifdef CONFIG_COMPAT_BRK
296 	/*
297 	 * CONFIG_COMPAT_BRK can still be overridden by setting
298 	 * randomize_va_space to 2, which will still cause mm->start_brk
299 	 * to be arbitrarily shifted
300 	 */
301 	if (current->brk_randomized)
302 		min_brk = mm->start_brk;
303 	else
304 		min_brk = mm->end_data;
305 #else
306 	min_brk = mm->start_brk;
307 #endif
308 	if (brk < min_brk)
309 		goto out;
310 
311 	/*
312 	 * Check against rlimit here. If this check is done later after the test
313 	 * of oldbrk with newbrk then it can escape the test and let the data
314 	 * segment grow beyond its set limit the in case where the limit is
315 	 * not page aligned -Ram Gupta
316 	 */
317 	if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
318 			      mm->end_data, mm->start_data))
319 		goto out;
320 
321 	newbrk = PAGE_ALIGN(brk);
322 	oldbrk = PAGE_ALIGN(mm->brk);
323 	if (oldbrk == newbrk)
324 		goto set_brk;
325 
326 	/* Always allow shrinking brk. */
327 	if (brk <= mm->brk) {
328 		if (!do_munmap(mm, newbrk, oldbrk-newbrk))
329 			goto set_brk;
330 		goto out;
331 	}
332 
333 	/* Check against existing mmap mappings. */
334 	if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
335 		goto out;
336 
337 	/* Ok, looks good - let it rip. */
338 	if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
339 		goto out;
340 
341 set_brk:
342 	mm->brk = brk;
343 	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
344 	up_write(&mm->mmap_sem);
345 	if (populate)
346 		mm_populate(oldbrk, newbrk - oldbrk);
347 	return brk;
348 
349 out:
350 	retval = mm->brk;
351 	up_write(&mm->mmap_sem);
352 	return retval;
353 }
354 
355 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
356 {
357 	unsigned long max, subtree_gap;
358 	max = vma->vm_start;
359 	if (vma->vm_prev)
360 		max -= vma->vm_prev->vm_end;
361 	if (vma->vm_rb.rb_left) {
362 		subtree_gap = rb_entry(vma->vm_rb.rb_left,
363 				struct vm_area_struct, vm_rb)->rb_subtree_gap;
364 		if (subtree_gap > max)
365 			max = subtree_gap;
366 	}
367 	if (vma->vm_rb.rb_right) {
368 		subtree_gap = rb_entry(vma->vm_rb.rb_right,
369 				struct vm_area_struct, vm_rb)->rb_subtree_gap;
370 		if (subtree_gap > max)
371 			max = subtree_gap;
372 	}
373 	return max;
374 }
375 
376 #ifdef CONFIG_DEBUG_VM_RB
377 static int browse_rb(struct rb_root *root)
378 {
379 	int i = 0, j, bug = 0;
380 	struct rb_node *nd, *pn = NULL;
381 	unsigned long prev = 0, pend = 0;
382 
383 	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
384 		struct vm_area_struct *vma;
385 		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
386 		if (vma->vm_start < prev) {
387 			pr_emerg("vm_start %lx < prev %lx\n",
388 				  vma->vm_start, prev);
389 			bug = 1;
390 		}
391 		if (vma->vm_start < pend) {
392 			pr_emerg("vm_start %lx < pend %lx\n",
393 				  vma->vm_start, pend);
394 			bug = 1;
395 		}
396 		if (vma->vm_start > vma->vm_end) {
397 			pr_emerg("vm_start %lx > vm_end %lx\n",
398 				  vma->vm_start, vma->vm_end);
399 			bug = 1;
400 		}
401 		if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
402 			pr_emerg("free gap %lx, correct %lx\n",
403 			       vma->rb_subtree_gap,
404 			       vma_compute_subtree_gap(vma));
405 			bug = 1;
406 		}
407 		i++;
408 		pn = nd;
409 		prev = vma->vm_start;
410 		pend = vma->vm_end;
411 	}
412 	j = 0;
413 	for (nd = pn; nd; nd = rb_prev(nd))
414 		j++;
415 	if (i != j) {
416 		pr_emerg("backwards %d, forwards %d\n", j, i);
417 		bug = 1;
418 	}
419 	return bug ? -1 : i;
420 }
421 
422 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
423 {
424 	struct rb_node *nd;
425 
426 	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
427 		struct vm_area_struct *vma;
428 		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
429 		VM_BUG_ON_VMA(vma != ignore &&
430 			vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
431 			vma);
432 	}
433 }
434 
435 static void validate_mm(struct mm_struct *mm)
436 {
437 	int bug = 0;
438 	int i = 0;
439 	unsigned long highest_address = 0;
440 	struct vm_area_struct *vma = mm->mmap;
441 
442 	while (vma) {
443 		struct anon_vma_chain *avc;
444 
445 		vma_lock_anon_vma(vma);
446 		list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
447 			anon_vma_interval_tree_verify(avc);
448 		vma_unlock_anon_vma(vma);
449 		highest_address = vma->vm_end;
450 		vma = vma->vm_next;
451 		i++;
452 	}
453 	if (i != mm->map_count) {
454 		pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
455 		bug = 1;
456 	}
457 	if (highest_address != mm->highest_vm_end) {
458 		pr_emerg("mm->highest_vm_end %lx, found %lx\n",
459 			  mm->highest_vm_end, highest_address);
460 		bug = 1;
461 	}
462 	i = browse_rb(&mm->mm_rb);
463 	if (i != mm->map_count) {
464 		if (i != -1)
465 			pr_emerg("map_count %d rb %d\n", mm->map_count, i);
466 		bug = 1;
467 	}
468 	VM_BUG_ON_MM(bug, mm);
469 }
470 #else
471 #define validate_mm_rb(root, ignore) do { } while (0)
472 #define validate_mm(mm) do { } while (0)
473 #endif
474 
475 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
476 		     unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
477 
478 /*
479  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
480  * vma->vm_prev->vm_end values changed, without modifying the vma's position
481  * in the rbtree.
482  */
483 static void vma_gap_update(struct vm_area_struct *vma)
484 {
485 	/*
486 	 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
487 	 * function that does exacltly what we want.
488 	 */
489 	vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
490 }
491 
492 static inline void vma_rb_insert(struct vm_area_struct *vma,
493 				 struct rb_root *root)
494 {
495 	/* All rb_subtree_gap values must be consistent prior to insertion */
496 	validate_mm_rb(root, NULL);
497 
498 	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
499 }
500 
501 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
502 {
503 	/*
504 	 * All rb_subtree_gap values must be consistent prior to erase,
505 	 * with the possible exception of the vma being erased.
506 	 */
507 	validate_mm_rb(root, vma);
508 
509 	/*
510 	 * Note rb_erase_augmented is a fairly large inline function,
511 	 * so make sure we instantiate it only once with our desired
512 	 * augmented rbtree callbacks.
513 	 */
514 	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
515 }
516 
517 /*
518  * vma has some anon_vma assigned, and is already inserted on that
519  * anon_vma's interval trees.
520  *
521  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
522  * vma must be removed from the anon_vma's interval trees using
523  * anon_vma_interval_tree_pre_update_vma().
524  *
525  * After the update, the vma will be reinserted using
526  * anon_vma_interval_tree_post_update_vma().
527  *
528  * The entire update must be protected by exclusive mmap_sem and by
529  * the root anon_vma's mutex.
530  */
531 static inline void
532 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
533 {
534 	struct anon_vma_chain *avc;
535 
536 	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
537 		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
538 }
539 
540 static inline void
541 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
542 {
543 	struct anon_vma_chain *avc;
544 
545 	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
546 		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
547 }
548 
549 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
550 		unsigned long end, struct vm_area_struct **pprev,
551 		struct rb_node ***rb_link, struct rb_node **rb_parent)
552 {
553 	struct rb_node **__rb_link, *__rb_parent, *rb_prev;
554 
555 	__rb_link = &mm->mm_rb.rb_node;
556 	rb_prev = __rb_parent = NULL;
557 
558 	while (*__rb_link) {
559 		struct vm_area_struct *vma_tmp;
560 
561 		__rb_parent = *__rb_link;
562 		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
563 
564 		if (vma_tmp->vm_end > addr) {
565 			/* Fail if an existing vma overlaps the area */
566 			if (vma_tmp->vm_start < end)
567 				return -ENOMEM;
568 			__rb_link = &__rb_parent->rb_left;
569 		} else {
570 			rb_prev = __rb_parent;
571 			__rb_link = &__rb_parent->rb_right;
572 		}
573 	}
574 
575 	*pprev = NULL;
576 	if (rb_prev)
577 		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
578 	*rb_link = __rb_link;
579 	*rb_parent = __rb_parent;
580 	return 0;
581 }
582 
583 static unsigned long count_vma_pages_range(struct mm_struct *mm,
584 		unsigned long addr, unsigned long end)
585 {
586 	unsigned long nr_pages = 0;
587 	struct vm_area_struct *vma;
588 
589 	/* Find first overlaping mapping */
590 	vma = find_vma_intersection(mm, addr, end);
591 	if (!vma)
592 		return 0;
593 
594 	nr_pages = (min(end, vma->vm_end) -
595 		max(addr, vma->vm_start)) >> PAGE_SHIFT;
596 
597 	/* Iterate over the rest of the overlaps */
598 	for (vma = vma->vm_next; vma; vma = vma->vm_next) {
599 		unsigned long overlap_len;
600 
601 		if (vma->vm_start > end)
602 			break;
603 
604 		overlap_len = min(end, vma->vm_end) - vma->vm_start;
605 		nr_pages += overlap_len >> PAGE_SHIFT;
606 	}
607 
608 	return nr_pages;
609 }
610 
611 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
612 		struct rb_node **rb_link, struct rb_node *rb_parent)
613 {
614 	/* Update tracking information for the gap following the new vma. */
615 	if (vma->vm_next)
616 		vma_gap_update(vma->vm_next);
617 	else
618 		mm->highest_vm_end = vma->vm_end;
619 
620 	/*
621 	 * vma->vm_prev wasn't known when we followed the rbtree to find the
622 	 * correct insertion point for that vma. As a result, we could not
623 	 * update the vma vm_rb parents rb_subtree_gap values on the way down.
624 	 * So, we first insert the vma with a zero rb_subtree_gap value
625 	 * (to be consistent with what we did on the way down), and then
626 	 * immediately update the gap to the correct value. Finally we
627 	 * rebalance the rbtree after all augmented values have been set.
628 	 */
629 	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
630 	vma->rb_subtree_gap = 0;
631 	vma_gap_update(vma);
632 	vma_rb_insert(vma, &mm->mm_rb);
633 }
634 
635 static void __vma_link_file(struct vm_area_struct *vma)
636 {
637 	struct file *file;
638 
639 	file = vma->vm_file;
640 	if (file) {
641 		struct address_space *mapping = file->f_mapping;
642 
643 		if (vma->vm_flags & VM_DENYWRITE)
644 			atomic_dec(&file_inode(file)->i_writecount);
645 		if (vma->vm_flags & VM_SHARED)
646 			atomic_inc(&mapping->i_mmap_writable);
647 
648 		flush_dcache_mmap_lock(mapping);
649 		vma_interval_tree_insert(vma, &mapping->i_mmap);
650 		flush_dcache_mmap_unlock(mapping);
651 	}
652 }
653 
654 static void
655 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
656 	struct vm_area_struct *prev, struct rb_node **rb_link,
657 	struct rb_node *rb_parent)
658 {
659 	__vma_link_list(mm, vma, prev, rb_parent);
660 	__vma_link_rb(mm, vma, rb_link, rb_parent);
661 }
662 
663 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
664 			struct vm_area_struct *prev, struct rb_node **rb_link,
665 			struct rb_node *rb_parent)
666 {
667 	struct address_space *mapping = NULL;
668 
669 	if (vma->vm_file) {
670 		mapping = vma->vm_file->f_mapping;
671 		i_mmap_lock_write(mapping);
672 	}
673 
674 	__vma_link(mm, vma, prev, rb_link, rb_parent);
675 	__vma_link_file(vma);
676 
677 	if (mapping)
678 		i_mmap_unlock_write(mapping);
679 
680 	mm->map_count++;
681 	validate_mm(mm);
682 }
683 
684 /*
685  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
686  * mm's list and rbtree.  It has already been inserted into the interval tree.
687  */
688 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
689 {
690 	struct vm_area_struct *prev;
691 	struct rb_node **rb_link, *rb_parent;
692 
693 	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
694 			   &prev, &rb_link, &rb_parent))
695 		BUG();
696 	__vma_link(mm, vma, prev, rb_link, rb_parent);
697 	mm->map_count++;
698 }
699 
700 static inline void
701 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
702 		struct vm_area_struct *prev)
703 {
704 	struct vm_area_struct *next;
705 
706 	vma_rb_erase(vma, &mm->mm_rb);
707 	prev->vm_next = next = vma->vm_next;
708 	if (next)
709 		next->vm_prev = prev;
710 
711 	/* Kill the cache */
712 	vmacache_invalidate(mm);
713 }
714 
715 /*
716  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
717  * is already present in an i_mmap tree without adjusting the tree.
718  * The following helper function should be used when such adjustments
719  * are necessary.  The "insert" vma (if any) is to be inserted
720  * before we drop the necessary locks.
721  */
722 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
723 	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
724 {
725 	struct mm_struct *mm = vma->vm_mm;
726 	struct vm_area_struct *next = vma->vm_next;
727 	struct vm_area_struct *importer = NULL;
728 	struct address_space *mapping = NULL;
729 	struct rb_root *root = NULL;
730 	struct anon_vma *anon_vma = NULL;
731 	struct file *file = vma->vm_file;
732 	bool start_changed = false, end_changed = false;
733 	long adjust_next = 0;
734 	int remove_next = 0;
735 
736 	if (next && !insert) {
737 		struct vm_area_struct *exporter = NULL;
738 
739 		if (end >= next->vm_end) {
740 			/*
741 			 * vma expands, overlapping all the next, and
742 			 * perhaps the one after too (mprotect case 6).
743 			 */
744 again:			remove_next = 1 + (end > next->vm_end);
745 			end = next->vm_end;
746 			exporter = next;
747 			importer = vma;
748 		} else if (end > next->vm_start) {
749 			/*
750 			 * vma expands, overlapping part of the next:
751 			 * mprotect case 5 shifting the boundary up.
752 			 */
753 			adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
754 			exporter = next;
755 			importer = vma;
756 		} else if (end < vma->vm_end) {
757 			/*
758 			 * vma shrinks, and !insert tells it's not
759 			 * split_vma inserting another: so it must be
760 			 * mprotect case 4 shifting the boundary down.
761 			 */
762 			adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
763 			exporter = vma;
764 			importer = next;
765 		}
766 
767 		/*
768 		 * Easily overlooked: when mprotect shifts the boundary,
769 		 * make sure the expanding vma has anon_vma set if the
770 		 * shrinking vma had, to cover any anon pages imported.
771 		 */
772 		if (exporter && exporter->anon_vma && !importer->anon_vma) {
773 			int error;
774 
775 			importer->anon_vma = exporter->anon_vma;
776 			error = anon_vma_clone(importer, exporter);
777 			if (error)
778 				return error;
779 		}
780 	}
781 
782 	if (file) {
783 		mapping = file->f_mapping;
784 		root = &mapping->i_mmap;
785 		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
786 
787 		if (adjust_next)
788 			uprobe_munmap(next, next->vm_start, next->vm_end);
789 
790 		i_mmap_lock_write(mapping);
791 		if (insert) {
792 			/*
793 			 * Put into interval tree now, so instantiated pages
794 			 * are visible to arm/parisc __flush_dcache_page
795 			 * throughout; but we cannot insert into address
796 			 * space until vma start or end is updated.
797 			 */
798 			__vma_link_file(insert);
799 		}
800 	}
801 
802 	vma_adjust_trans_huge(vma, start, end, adjust_next);
803 
804 	anon_vma = vma->anon_vma;
805 	if (!anon_vma && adjust_next)
806 		anon_vma = next->anon_vma;
807 	if (anon_vma) {
808 		VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
809 			  anon_vma != next->anon_vma, next);
810 		anon_vma_lock_write(anon_vma);
811 		anon_vma_interval_tree_pre_update_vma(vma);
812 		if (adjust_next)
813 			anon_vma_interval_tree_pre_update_vma(next);
814 	}
815 
816 	if (root) {
817 		flush_dcache_mmap_lock(mapping);
818 		vma_interval_tree_remove(vma, root);
819 		if (adjust_next)
820 			vma_interval_tree_remove(next, root);
821 	}
822 
823 	if (start != vma->vm_start) {
824 		vma->vm_start = start;
825 		start_changed = true;
826 	}
827 	if (end != vma->vm_end) {
828 		vma->vm_end = end;
829 		end_changed = true;
830 	}
831 	vma->vm_pgoff = pgoff;
832 	if (adjust_next) {
833 		next->vm_start += adjust_next << PAGE_SHIFT;
834 		next->vm_pgoff += adjust_next;
835 	}
836 
837 	if (root) {
838 		if (adjust_next)
839 			vma_interval_tree_insert(next, root);
840 		vma_interval_tree_insert(vma, root);
841 		flush_dcache_mmap_unlock(mapping);
842 	}
843 
844 	if (remove_next) {
845 		/*
846 		 * vma_merge has merged next into vma, and needs
847 		 * us to remove next before dropping the locks.
848 		 */
849 		__vma_unlink(mm, next, vma);
850 		if (file)
851 			__remove_shared_vm_struct(next, file, mapping);
852 	} else if (insert) {
853 		/*
854 		 * split_vma has split insert from vma, and needs
855 		 * us to insert it before dropping the locks
856 		 * (it may either follow vma or precede it).
857 		 */
858 		__insert_vm_struct(mm, insert);
859 	} else {
860 		if (start_changed)
861 			vma_gap_update(vma);
862 		if (end_changed) {
863 			if (!next)
864 				mm->highest_vm_end = end;
865 			else if (!adjust_next)
866 				vma_gap_update(next);
867 		}
868 	}
869 
870 	if (anon_vma) {
871 		anon_vma_interval_tree_post_update_vma(vma);
872 		if (adjust_next)
873 			anon_vma_interval_tree_post_update_vma(next);
874 		anon_vma_unlock_write(anon_vma);
875 	}
876 	if (mapping)
877 		i_mmap_unlock_write(mapping);
878 
879 	if (root) {
880 		uprobe_mmap(vma);
881 
882 		if (adjust_next)
883 			uprobe_mmap(next);
884 	}
885 
886 	if (remove_next) {
887 		if (file) {
888 			uprobe_munmap(next, next->vm_start, next->vm_end);
889 			fput(file);
890 		}
891 		if (next->anon_vma)
892 			anon_vma_merge(vma, next);
893 		mm->map_count--;
894 		mpol_put(vma_policy(next));
895 		kmem_cache_free(vm_area_cachep, next);
896 		/*
897 		 * In mprotect's case 6 (see comments on vma_merge),
898 		 * we must remove another next too. It would clutter
899 		 * up the code too much to do both in one go.
900 		 */
901 		next = vma->vm_next;
902 		if (remove_next == 2)
903 			goto again;
904 		else if (next)
905 			vma_gap_update(next);
906 		else
907 			mm->highest_vm_end = end;
908 	}
909 	if (insert && file)
910 		uprobe_mmap(insert);
911 
912 	validate_mm(mm);
913 
914 	return 0;
915 }
916 
917 /*
918  * If the vma has a ->close operation then the driver probably needs to release
919  * per-vma resources, so we don't attempt to merge those.
920  */
921 static inline int is_mergeable_vma(struct vm_area_struct *vma,
922 			struct file *file, unsigned long vm_flags)
923 {
924 	/*
925 	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
926 	 * match the flags but dirty bit -- the caller should mark
927 	 * merged VMA as dirty. If dirty bit won't be excluded from
928 	 * comparison, we increase pressue on the memory system forcing
929 	 * the kernel to generate new VMAs when old one could be
930 	 * extended instead.
931 	 */
932 	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
933 		return 0;
934 	if (vma->vm_file != file)
935 		return 0;
936 	if (vma->vm_ops && vma->vm_ops->close)
937 		return 0;
938 	return 1;
939 }
940 
941 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
942 					struct anon_vma *anon_vma2,
943 					struct vm_area_struct *vma)
944 {
945 	/*
946 	 * The list_is_singular() test is to avoid merging VMA cloned from
947 	 * parents. This can improve scalability caused by anon_vma lock.
948 	 */
949 	if ((!anon_vma1 || !anon_vma2) && (!vma ||
950 		list_is_singular(&vma->anon_vma_chain)))
951 		return 1;
952 	return anon_vma1 == anon_vma2;
953 }
954 
955 /*
956  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
957  * in front of (at a lower virtual address and file offset than) the vma.
958  *
959  * We cannot merge two vmas if they have differently assigned (non-NULL)
960  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
961  *
962  * We don't check here for the merged mmap wrapping around the end of pagecache
963  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
964  * wrap, nor mmaps which cover the final page at index -1UL.
965  */
966 static int
967 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
968 	struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
969 {
970 	if (is_mergeable_vma(vma, file, vm_flags) &&
971 	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
972 		if (vma->vm_pgoff == vm_pgoff)
973 			return 1;
974 	}
975 	return 0;
976 }
977 
978 /*
979  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
980  * beyond (at a higher virtual address and file offset than) the vma.
981  *
982  * We cannot merge two vmas if they have differently assigned (non-NULL)
983  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
984  */
985 static int
986 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
987 	struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
988 {
989 	if (is_mergeable_vma(vma, file, vm_flags) &&
990 	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
991 		pgoff_t vm_pglen;
992 		vm_pglen = vma_pages(vma);
993 		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
994 			return 1;
995 	}
996 	return 0;
997 }
998 
999 /*
1000  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1001  * whether that can be merged with its predecessor or its successor.
1002  * Or both (it neatly fills a hole).
1003  *
1004  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1005  * certain not to be mapped by the time vma_merge is called; but when
1006  * called for mprotect, it is certain to be already mapped (either at
1007  * an offset within prev, or at the start of next), and the flags of
1008  * this area are about to be changed to vm_flags - and the no-change
1009  * case has already been eliminated.
1010  *
1011  * The following mprotect cases have to be considered, where AAAA is
1012  * the area passed down from mprotect_fixup, never extending beyond one
1013  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1014  *
1015  *     AAAA             AAAA                AAAA          AAAA
1016  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1017  *    cannot merge    might become    might become    might become
1018  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1019  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1020  *    mremap move:                                    PPPPNNNNNNNN 8
1021  *        AAAA
1022  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1023  *    might become    case 1 below    case 2 below    case 3 below
1024  *
1025  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1026  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1027  */
1028 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1029 			struct vm_area_struct *prev, unsigned long addr,
1030 			unsigned long end, unsigned long vm_flags,
1031 			struct anon_vma *anon_vma, struct file *file,
1032 			pgoff_t pgoff, struct mempolicy *policy)
1033 {
1034 	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1035 	struct vm_area_struct *area, *next;
1036 	int err;
1037 
1038 	/*
1039 	 * We later require that vma->vm_flags == vm_flags,
1040 	 * so this tests vma->vm_flags & VM_SPECIAL, too.
1041 	 */
1042 	if (vm_flags & VM_SPECIAL)
1043 		return NULL;
1044 
1045 	if (prev)
1046 		next = prev->vm_next;
1047 	else
1048 		next = mm->mmap;
1049 	area = next;
1050 	if (next && next->vm_end == end)		/* cases 6, 7, 8 */
1051 		next = next->vm_next;
1052 
1053 	/*
1054 	 * Can it merge with the predecessor?
1055 	 */
1056 	if (prev && prev->vm_end == addr &&
1057 			mpol_equal(vma_policy(prev), policy) &&
1058 			can_vma_merge_after(prev, vm_flags,
1059 						anon_vma, file, pgoff)) {
1060 		/*
1061 		 * OK, it can.  Can we now merge in the successor as well?
1062 		 */
1063 		if (next && end == next->vm_start &&
1064 				mpol_equal(policy, vma_policy(next)) &&
1065 				can_vma_merge_before(next, vm_flags,
1066 					anon_vma, file, pgoff+pglen) &&
1067 				is_mergeable_anon_vma(prev->anon_vma,
1068 						      next->anon_vma, NULL)) {
1069 							/* cases 1, 6 */
1070 			err = vma_adjust(prev, prev->vm_start,
1071 				next->vm_end, prev->vm_pgoff, NULL);
1072 		} else					/* cases 2, 5, 7 */
1073 			err = vma_adjust(prev, prev->vm_start,
1074 				end, prev->vm_pgoff, NULL);
1075 		if (err)
1076 			return NULL;
1077 		khugepaged_enter_vma_merge(prev, vm_flags);
1078 		return prev;
1079 	}
1080 
1081 	/*
1082 	 * Can this new request be merged in front of next?
1083 	 */
1084 	if (next && end == next->vm_start &&
1085 			mpol_equal(policy, vma_policy(next)) &&
1086 			can_vma_merge_before(next, vm_flags,
1087 					anon_vma, file, pgoff+pglen)) {
1088 		if (prev && addr < prev->vm_end)	/* case 4 */
1089 			err = vma_adjust(prev, prev->vm_start,
1090 				addr, prev->vm_pgoff, NULL);
1091 		else					/* cases 3, 8 */
1092 			err = vma_adjust(area, addr, next->vm_end,
1093 				next->vm_pgoff - pglen, NULL);
1094 		if (err)
1095 			return NULL;
1096 		khugepaged_enter_vma_merge(area, vm_flags);
1097 		return area;
1098 	}
1099 
1100 	return NULL;
1101 }
1102 
1103 /*
1104  * Rough compatbility check to quickly see if it's even worth looking
1105  * at sharing an anon_vma.
1106  *
1107  * They need to have the same vm_file, and the flags can only differ
1108  * in things that mprotect may change.
1109  *
1110  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1111  * we can merge the two vma's. For example, we refuse to merge a vma if
1112  * there is a vm_ops->close() function, because that indicates that the
1113  * driver is doing some kind of reference counting. But that doesn't
1114  * really matter for the anon_vma sharing case.
1115  */
1116 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1117 {
1118 	return a->vm_end == b->vm_start &&
1119 		mpol_equal(vma_policy(a), vma_policy(b)) &&
1120 		a->vm_file == b->vm_file &&
1121 		!((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1122 		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1123 }
1124 
1125 /*
1126  * Do some basic sanity checking to see if we can re-use the anon_vma
1127  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1128  * the same as 'old', the other will be the new one that is trying
1129  * to share the anon_vma.
1130  *
1131  * NOTE! This runs with mm_sem held for reading, so it is possible that
1132  * the anon_vma of 'old' is concurrently in the process of being set up
1133  * by another page fault trying to merge _that_. But that's ok: if it
1134  * is being set up, that automatically means that it will be a singleton
1135  * acceptable for merging, so we can do all of this optimistically. But
1136  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1137  *
1138  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1139  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1140  * is to return an anon_vma that is "complex" due to having gone through
1141  * a fork).
1142  *
1143  * We also make sure that the two vma's are compatible (adjacent,
1144  * and with the same memory policies). That's all stable, even with just
1145  * a read lock on the mm_sem.
1146  */
1147 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1148 {
1149 	if (anon_vma_compatible(a, b)) {
1150 		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1151 
1152 		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1153 			return anon_vma;
1154 	}
1155 	return NULL;
1156 }
1157 
1158 /*
1159  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1160  * neighbouring vmas for a suitable anon_vma, before it goes off
1161  * to allocate a new anon_vma.  It checks because a repetitive
1162  * sequence of mprotects and faults may otherwise lead to distinct
1163  * anon_vmas being allocated, preventing vma merge in subsequent
1164  * mprotect.
1165  */
1166 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1167 {
1168 	struct anon_vma *anon_vma;
1169 	struct vm_area_struct *near;
1170 
1171 	near = vma->vm_next;
1172 	if (!near)
1173 		goto try_prev;
1174 
1175 	anon_vma = reusable_anon_vma(near, vma, near);
1176 	if (anon_vma)
1177 		return anon_vma;
1178 try_prev:
1179 	near = vma->vm_prev;
1180 	if (!near)
1181 		goto none;
1182 
1183 	anon_vma = reusable_anon_vma(near, near, vma);
1184 	if (anon_vma)
1185 		return anon_vma;
1186 none:
1187 	/*
1188 	 * There's no absolute need to look only at touching neighbours:
1189 	 * we could search further afield for "compatible" anon_vmas.
1190 	 * But it would probably just be a waste of time searching,
1191 	 * or lead to too many vmas hanging off the same anon_vma.
1192 	 * We're trying to allow mprotect remerging later on,
1193 	 * not trying to minimize memory used for anon_vmas.
1194 	 */
1195 	return NULL;
1196 }
1197 
1198 #ifdef CONFIG_PROC_FS
1199 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1200 						struct file *file, long pages)
1201 {
1202 	const unsigned long stack_flags
1203 		= VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1204 
1205 	mm->total_vm += pages;
1206 
1207 	if (file) {
1208 		mm->shared_vm += pages;
1209 		if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1210 			mm->exec_vm += pages;
1211 	} else if (flags & stack_flags)
1212 		mm->stack_vm += pages;
1213 }
1214 #endif /* CONFIG_PROC_FS */
1215 
1216 /*
1217  * If a hint addr is less than mmap_min_addr change hint to be as
1218  * low as possible but still greater than mmap_min_addr
1219  */
1220 static inline unsigned long round_hint_to_min(unsigned long hint)
1221 {
1222 	hint &= PAGE_MASK;
1223 	if (((void *)hint != NULL) &&
1224 	    (hint < mmap_min_addr))
1225 		return PAGE_ALIGN(mmap_min_addr);
1226 	return hint;
1227 }
1228 
1229 static inline int mlock_future_check(struct mm_struct *mm,
1230 				     unsigned long flags,
1231 				     unsigned long len)
1232 {
1233 	unsigned long locked, lock_limit;
1234 
1235 	/*  mlock MCL_FUTURE? */
1236 	if (flags & VM_LOCKED) {
1237 		locked = len >> PAGE_SHIFT;
1238 		locked += mm->locked_vm;
1239 		lock_limit = rlimit(RLIMIT_MEMLOCK);
1240 		lock_limit >>= PAGE_SHIFT;
1241 		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1242 			return -EAGAIN;
1243 	}
1244 	return 0;
1245 }
1246 
1247 /*
1248  * The caller must hold down_write(&current->mm->mmap_sem).
1249  */
1250 
1251 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1252 			unsigned long len, unsigned long prot,
1253 			unsigned long flags, unsigned long pgoff,
1254 			unsigned long *populate)
1255 {
1256 	struct mm_struct *mm = current->mm;
1257 	vm_flags_t vm_flags;
1258 
1259 	*populate = 0;
1260 
1261 	/*
1262 	 * Does the application expect PROT_READ to imply PROT_EXEC?
1263 	 *
1264 	 * (the exception is when the underlying filesystem is noexec
1265 	 *  mounted, in which case we dont add PROT_EXEC.)
1266 	 */
1267 	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1268 		if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1269 			prot |= PROT_EXEC;
1270 
1271 	if (!len)
1272 		return -EINVAL;
1273 
1274 	if (!(flags & MAP_FIXED))
1275 		addr = round_hint_to_min(addr);
1276 
1277 	/* Careful about overflows.. */
1278 	len = PAGE_ALIGN(len);
1279 	if (!len)
1280 		return -ENOMEM;
1281 
1282 	/* offset overflow? */
1283 	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1284 		return -EOVERFLOW;
1285 
1286 	/* Too many mappings? */
1287 	if (mm->map_count > sysctl_max_map_count)
1288 		return -ENOMEM;
1289 
1290 	/* Obtain the address to map to. we verify (or select) it and ensure
1291 	 * that it represents a valid section of the address space.
1292 	 */
1293 	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1294 	if (addr & ~PAGE_MASK)
1295 		return addr;
1296 
1297 	/* Do simple checking here so the lower-level routines won't have
1298 	 * to. we assume access permissions have been handled by the open
1299 	 * of the memory object, so we don't do any here.
1300 	 */
1301 	vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1302 			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1303 
1304 	if (flags & MAP_LOCKED)
1305 		if (!can_do_mlock())
1306 			return -EPERM;
1307 
1308 	if (mlock_future_check(mm, vm_flags, len))
1309 		return -EAGAIN;
1310 
1311 	if (file) {
1312 		struct inode *inode = file_inode(file);
1313 
1314 		switch (flags & MAP_TYPE) {
1315 		case MAP_SHARED:
1316 			if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1317 				return -EACCES;
1318 
1319 			/*
1320 			 * Make sure we don't allow writing to an append-only
1321 			 * file..
1322 			 */
1323 			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1324 				return -EACCES;
1325 
1326 			/*
1327 			 * Make sure there are no mandatory locks on the file.
1328 			 */
1329 			if (locks_verify_locked(file))
1330 				return -EAGAIN;
1331 
1332 			vm_flags |= VM_SHARED | VM_MAYSHARE;
1333 			if (!(file->f_mode & FMODE_WRITE))
1334 				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1335 
1336 			/* fall through */
1337 		case MAP_PRIVATE:
1338 			if (!(file->f_mode & FMODE_READ))
1339 				return -EACCES;
1340 			if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1341 				if (vm_flags & VM_EXEC)
1342 					return -EPERM;
1343 				vm_flags &= ~VM_MAYEXEC;
1344 			}
1345 
1346 			if (!file->f_op->mmap)
1347 				return -ENODEV;
1348 			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1349 				return -EINVAL;
1350 			break;
1351 
1352 		default:
1353 			return -EINVAL;
1354 		}
1355 	} else {
1356 		switch (flags & MAP_TYPE) {
1357 		case MAP_SHARED:
1358 			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1359 				return -EINVAL;
1360 			/*
1361 			 * Ignore pgoff.
1362 			 */
1363 			pgoff = 0;
1364 			vm_flags |= VM_SHARED | VM_MAYSHARE;
1365 			break;
1366 		case MAP_PRIVATE:
1367 			/*
1368 			 * Set pgoff according to addr for anon_vma.
1369 			 */
1370 			pgoff = addr >> PAGE_SHIFT;
1371 			break;
1372 		default:
1373 			return -EINVAL;
1374 		}
1375 	}
1376 
1377 	/*
1378 	 * Set 'VM_NORESERVE' if we should not account for the
1379 	 * memory use of this mapping.
1380 	 */
1381 	if (flags & MAP_NORESERVE) {
1382 		/* We honor MAP_NORESERVE if allowed to overcommit */
1383 		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1384 			vm_flags |= VM_NORESERVE;
1385 
1386 		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1387 		if (file && is_file_hugepages(file))
1388 			vm_flags |= VM_NORESERVE;
1389 	}
1390 
1391 	addr = mmap_region(file, addr, len, vm_flags, pgoff);
1392 	if (!IS_ERR_VALUE(addr) &&
1393 	    ((vm_flags & VM_LOCKED) ||
1394 	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1395 		*populate = len;
1396 	return addr;
1397 }
1398 
1399 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1400 		unsigned long, prot, unsigned long, flags,
1401 		unsigned long, fd, unsigned long, pgoff)
1402 {
1403 	struct file *file = NULL;
1404 	unsigned long retval = -EBADF;
1405 
1406 	if (!(flags & MAP_ANONYMOUS)) {
1407 		audit_mmap_fd(fd, flags);
1408 		file = fget(fd);
1409 		if (!file)
1410 			goto out;
1411 		if (is_file_hugepages(file))
1412 			len = ALIGN(len, huge_page_size(hstate_file(file)));
1413 		retval = -EINVAL;
1414 		if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1415 			goto out_fput;
1416 	} else if (flags & MAP_HUGETLB) {
1417 		struct user_struct *user = NULL;
1418 		struct hstate *hs;
1419 
1420 		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1421 		if (!hs)
1422 			return -EINVAL;
1423 
1424 		len = ALIGN(len, huge_page_size(hs));
1425 		/*
1426 		 * VM_NORESERVE is used because the reservations will be
1427 		 * taken when vm_ops->mmap() is called
1428 		 * A dummy user value is used because we are not locking
1429 		 * memory so no accounting is necessary
1430 		 */
1431 		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1432 				VM_NORESERVE,
1433 				&user, HUGETLB_ANONHUGE_INODE,
1434 				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1435 		if (IS_ERR(file))
1436 			return PTR_ERR(file);
1437 	}
1438 
1439 	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1440 
1441 	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1442 out_fput:
1443 	if (file)
1444 		fput(file);
1445 out:
1446 	return retval;
1447 }
1448 
1449 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1450 struct mmap_arg_struct {
1451 	unsigned long addr;
1452 	unsigned long len;
1453 	unsigned long prot;
1454 	unsigned long flags;
1455 	unsigned long fd;
1456 	unsigned long offset;
1457 };
1458 
1459 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1460 {
1461 	struct mmap_arg_struct a;
1462 
1463 	if (copy_from_user(&a, arg, sizeof(a)))
1464 		return -EFAULT;
1465 	if (a.offset & ~PAGE_MASK)
1466 		return -EINVAL;
1467 
1468 	return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1469 			      a.offset >> PAGE_SHIFT);
1470 }
1471 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1472 
1473 /*
1474  * Some shared mappigns will want the pages marked read-only
1475  * to track write events. If so, we'll downgrade vm_page_prot
1476  * to the private version (using protection_map[] without the
1477  * VM_SHARED bit).
1478  */
1479 int vma_wants_writenotify(struct vm_area_struct *vma)
1480 {
1481 	vm_flags_t vm_flags = vma->vm_flags;
1482 
1483 	/* If it was private or non-writable, the write bit is already clear */
1484 	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1485 		return 0;
1486 
1487 	/* The backer wishes to know when pages are first written to? */
1488 	if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1489 		return 1;
1490 
1491 	/* The open routine did something to the protections that pgprot_modify
1492 	 * won't preserve? */
1493 	if (pgprot_val(vma->vm_page_prot) !=
1494 	    pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1495 		return 0;
1496 
1497 	/* Do we need to track softdirty? */
1498 	if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1499 		return 1;
1500 
1501 	/* Specialty mapping? */
1502 	if (vm_flags & VM_PFNMAP)
1503 		return 0;
1504 
1505 	/* Can the mapping track the dirty pages? */
1506 	return vma->vm_file && vma->vm_file->f_mapping &&
1507 		mapping_cap_account_dirty(vma->vm_file->f_mapping);
1508 }
1509 
1510 /*
1511  * We account for memory if it's a private writeable mapping,
1512  * not hugepages and VM_NORESERVE wasn't set.
1513  */
1514 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1515 {
1516 	/*
1517 	 * hugetlb has its own accounting separate from the core VM
1518 	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1519 	 */
1520 	if (file && is_file_hugepages(file))
1521 		return 0;
1522 
1523 	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1524 }
1525 
1526 unsigned long mmap_region(struct file *file, unsigned long addr,
1527 		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1528 {
1529 	struct mm_struct *mm = current->mm;
1530 	struct vm_area_struct *vma, *prev;
1531 	int error;
1532 	struct rb_node **rb_link, *rb_parent;
1533 	unsigned long charged = 0;
1534 
1535 	/* Check against address space limit. */
1536 	if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1537 		unsigned long nr_pages;
1538 
1539 		/*
1540 		 * MAP_FIXED may remove pages of mappings that intersects with
1541 		 * requested mapping. Account for the pages it would unmap.
1542 		 */
1543 		if (!(vm_flags & MAP_FIXED))
1544 			return -ENOMEM;
1545 
1546 		nr_pages = count_vma_pages_range(mm, addr, addr + len);
1547 
1548 		if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1549 			return -ENOMEM;
1550 	}
1551 
1552 	/* Clear old maps */
1553 	error = -ENOMEM;
1554 	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1555 			      &rb_parent)) {
1556 		if (do_munmap(mm, addr, len))
1557 			return -ENOMEM;
1558 	}
1559 
1560 	/*
1561 	 * Private writable mapping: check memory availability
1562 	 */
1563 	if (accountable_mapping(file, vm_flags)) {
1564 		charged = len >> PAGE_SHIFT;
1565 		if (security_vm_enough_memory_mm(mm, charged))
1566 			return -ENOMEM;
1567 		vm_flags |= VM_ACCOUNT;
1568 	}
1569 
1570 	/*
1571 	 * Can we just expand an old mapping?
1572 	 */
1573 	vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff,
1574 			NULL);
1575 	if (vma)
1576 		goto out;
1577 
1578 	/*
1579 	 * Determine the object being mapped and call the appropriate
1580 	 * specific mapper. the address has already been validated, but
1581 	 * not unmapped, but the maps are removed from the list.
1582 	 */
1583 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1584 	if (!vma) {
1585 		error = -ENOMEM;
1586 		goto unacct_error;
1587 	}
1588 
1589 	vma->vm_mm = mm;
1590 	vma->vm_start = addr;
1591 	vma->vm_end = addr + len;
1592 	vma->vm_flags = vm_flags;
1593 	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1594 	vma->vm_pgoff = pgoff;
1595 	INIT_LIST_HEAD(&vma->anon_vma_chain);
1596 
1597 	if (file) {
1598 		if (vm_flags & VM_DENYWRITE) {
1599 			error = deny_write_access(file);
1600 			if (error)
1601 				goto free_vma;
1602 		}
1603 		if (vm_flags & VM_SHARED) {
1604 			error = mapping_map_writable(file->f_mapping);
1605 			if (error)
1606 				goto allow_write_and_free_vma;
1607 		}
1608 
1609 		/* ->mmap() can change vma->vm_file, but must guarantee that
1610 		 * vma_link() below can deny write-access if VM_DENYWRITE is set
1611 		 * and map writably if VM_SHARED is set. This usually means the
1612 		 * new file must not have been exposed to user-space, yet.
1613 		 */
1614 		vma->vm_file = get_file(file);
1615 		error = file->f_op->mmap(file, vma);
1616 		if (error)
1617 			goto unmap_and_free_vma;
1618 
1619 		/* Can addr have changed??
1620 		 *
1621 		 * Answer: Yes, several device drivers can do it in their
1622 		 *         f_op->mmap method. -DaveM
1623 		 * Bug: If addr is changed, prev, rb_link, rb_parent should
1624 		 *      be updated for vma_link()
1625 		 */
1626 		WARN_ON_ONCE(addr != vma->vm_start);
1627 
1628 		addr = vma->vm_start;
1629 		vm_flags = vma->vm_flags;
1630 	} else if (vm_flags & VM_SHARED) {
1631 		error = shmem_zero_setup(vma);
1632 		if (error)
1633 			goto free_vma;
1634 	}
1635 
1636 	vma_link(mm, vma, prev, rb_link, rb_parent);
1637 	/* Once vma denies write, undo our temporary denial count */
1638 	if (file) {
1639 		if (vm_flags & VM_SHARED)
1640 			mapping_unmap_writable(file->f_mapping);
1641 		if (vm_flags & VM_DENYWRITE)
1642 			allow_write_access(file);
1643 	}
1644 	file = vma->vm_file;
1645 out:
1646 	perf_event_mmap(vma);
1647 
1648 	vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1649 	if (vm_flags & VM_LOCKED) {
1650 		if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1651 					vma == get_gate_vma(current->mm)))
1652 			mm->locked_vm += (len >> PAGE_SHIFT);
1653 		else
1654 			vma->vm_flags &= ~VM_LOCKED;
1655 	}
1656 
1657 	if (file)
1658 		uprobe_mmap(vma);
1659 
1660 	/*
1661 	 * New (or expanded) vma always get soft dirty status.
1662 	 * Otherwise user-space soft-dirty page tracker won't
1663 	 * be able to distinguish situation when vma area unmapped,
1664 	 * then new mapped in-place (which must be aimed as
1665 	 * a completely new data area).
1666 	 */
1667 	vma->vm_flags |= VM_SOFTDIRTY;
1668 
1669 	vma_set_page_prot(vma);
1670 
1671 	return addr;
1672 
1673 unmap_and_free_vma:
1674 	vma->vm_file = NULL;
1675 	fput(file);
1676 
1677 	/* Undo any partial mapping done by a device driver. */
1678 	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1679 	charged = 0;
1680 	if (vm_flags & VM_SHARED)
1681 		mapping_unmap_writable(file->f_mapping);
1682 allow_write_and_free_vma:
1683 	if (vm_flags & VM_DENYWRITE)
1684 		allow_write_access(file);
1685 free_vma:
1686 	kmem_cache_free(vm_area_cachep, vma);
1687 unacct_error:
1688 	if (charged)
1689 		vm_unacct_memory(charged);
1690 	return error;
1691 }
1692 
1693 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1694 {
1695 	/*
1696 	 * We implement the search by looking for an rbtree node that
1697 	 * immediately follows a suitable gap. That is,
1698 	 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1699 	 * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1700 	 * - gap_end - gap_start >= length
1701 	 */
1702 
1703 	struct mm_struct *mm = current->mm;
1704 	struct vm_area_struct *vma;
1705 	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1706 
1707 	/* Adjust search length to account for worst case alignment overhead */
1708 	length = info->length + info->align_mask;
1709 	if (length < info->length)
1710 		return -ENOMEM;
1711 
1712 	/* Adjust search limits by the desired length */
1713 	if (info->high_limit < length)
1714 		return -ENOMEM;
1715 	high_limit = info->high_limit - length;
1716 
1717 	if (info->low_limit > high_limit)
1718 		return -ENOMEM;
1719 	low_limit = info->low_limit + length;
1720 
1721 	/* Check if rbtree root looks promising */
1722 	if (RB_EMPTY_ROOT(&mm->mm_rb))
1723 		goto check_highest;
1724 	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1725 	if (vma->rb_subtree_gap < length)
1726 		goto check_highest;
1727 
1728 	while (true) {
1729 		/* Visit left subtree if it looks promising */
1730 		gap_end = vma->vm_start;
1731 		if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1732 			struct vm_area_struct *left =
1733 				rb_entry(vma->vm_rb.rb_left,
1734 					 struct vm_area_struct, vm_rb);
1735 			if (left->rb_subtree_gap >= length) {
1736 				vma = left;
1737 				continue;
1738 			}
1739 		}
1740 
1741 		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1742 check_current:
1743 		/* Check if current node has a suitable gap */
1744 		if (gap_start > high_limit)
1745 			return -ENOMEM;
1746 		if (gap_end >= low_limit && gap_end - gap_start >= length)
1747 			goto found;
1748 
1749 		/* Visit right subtree if it looks promising */
1750 		if (vma->vm_rb.rb_right) {
1751 			struct vm_area_struct *right =
1752 				rb_entry(vma->vm_rb.rb_right,
1753 					 struct vm_area_struct, vm_rb);
1754 			if (right->rb_subtree_gap >= length) {
1755 				vma = right;
1756 				continue;
1757 			}
1758 		}
1759 
1760 		/* Go back up the rbtree to find next candidate node */
1761 		while (true) {
1762 			struct rb_node *prev = &vma->vm_rb;
1763 			if (!rb_parent(prev))
1764 				goto check_highest;
1765 			vma = rb_entry(rb_parent(prev),
1766 				       struct vm_area_struct, vm_rb);
1767 			if (prev == vma->vm_rb.rb_left) {
1768 				gap_start = vma->vm_prev->vm_end;
1769 				gap_end = vma->vm_start;
1770 				goto check_current;
1771 			}
1772 		}
1773 	}
1774 
1775 check_highest:
1776 	/* Check highest gap, which does not precede any rbtree node */
1777 	gap_start = mm->highest_vm_end;
1778 	gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1779 	if (gap_start > high_limit)
1780 		return -ENOMEM;
1781 
1782 found:
1783 	/* We found a suitable gap. Clip it with the original low_limit. */
1784 	if (gap_start < info->low_limit)
1785 		gap_start = info->low_limit;
1786 
1787 	/* Adjust gap address to the desired alignment */
1788 	gap_start += (info->align_offset - gap_start) & info->align_mask;
1789 
1790 	VM_BUG_ON(gap_start + info->length > info->high_limit);
1791 	VM_BUG_ON(gap_start + info->length > gap_end);
1792 	return gap_start;
1793 }
1794 
1795 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1796 {
1797 	struct mm_struct *mm = current->mm;
1798 	struct vm_area_struct *vma;
1799 	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1800 
1801 	/* Adjust search length to account for worst case alignment overhead */
1802 	length = info->length + info->align_mask;
1803 	if (length < info->length)
1804 		return -ENOMEM;
1805 
1806 	/*
1807 	 * Adjust search limits by the desired length.
1808 	 * See implementation comment at top of unmapped_area().
1809 	 */
1810 	gap_end = info->high_limit;
1811 	if (gap_end < length)
1812 		return -ENOMEM;
1813 	high_limit = gap_end - length;
1814 
1815 	if (info->low_limit > high_limit)
1816 		return -ENOMEM;
1817 	low_limit = info->low_limit + length;
1818 
1819 	/* Check highest gap, which does not precede any rbtree node */
1820 	gap_start = mm->highest_vm_end;
1821 	if (gap_start <= high_limit)
1822 		goto found_highest;
1823 
1824 	/* Check if rbtree root looks promising */
1825 	if (RB_EMPTY_ROOT(&mm->mm_rb))
1826 		return -ENOMEM;
1827 	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1828 	if (vma->rb_subtree_gap < length)
1829 		return -ENOMEM;
1830 
1831 	while (true) {
1832 		/* Visit right subtree if it looks promising */
1833 		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1834 		if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1835 			struct vm_area_struct *right =
1836 				rb_entry(vma->vm_rb.rb_right,
1837 					 struct vm_area_struct, vm_rb);
1838 			if (right->rb_subtree_gap >= length) {
1839 				vma = right;
1840 				continue;
1841 			}
1842 		}
1843 
1844 check_current:
1845 		/* Check if current node has a suitable gap */
1846 		gap_end = vma->vm_start;
1847 		if (gap_end < low_limit)
1848 			return -ENOMEM;
1849 		if (gap_start <= high_limit && gap_end - gap_start >= length)
1850 			goto found;
1851 
1852 		/* Visit left subtree if it looks promising */
1853 		if (vma->vm_rb.rb_left) {
1854 			struct vm_area_struct *left =
1855 				rb_entry(vma->vm_rb.rb_left,
1856 					 struct vm_area_struct, vm_rb);
1857 			if (left->rb_subtree_gap >= length) {
1858 				vma = left;
1859 				continue;
1860 			}
1861 		}
1862 
1863 		/* Go back up the rbtree to find next candidate node */
1864 		while (true) {
1865 			struct rb_node *prev = &vma->vm_rb;
1866 			if (!rb_parent(prev))
1867 				return -ENOMEM;
1868 			vma = rb_entry(rb_parent(prev),
1869 				       struct vm_area_struct, vm_rb);
1870 			if (prev == vma->vm_rb.rb_right) {
1871 				gap_start = vma->vm_prev ?
1872 					vma->vm_prev->vm_end : 0;
1873 				goto check_current;
1874 			}
1875 		}
1876 	}
1877 
1878 found:
1879 	/* We found a suitable gap. Clip it with the original high_limit. */
1880 	if (gap_end > info->high_limit)
1881 		gap_end = info->high_limit;
1882 
1883 found_highest:
1884 	/* Compute highest gap address at the desired alignment */
1885 	gap_end -= info->length;
1886 	gap_end -= (gap_end - info->align_offset) & info->align_mask;
1887 
1888 	VM_BUG_ON(gap_end < info->low_limit);
1889 	VM_BUG_ON(gap_end < gap_start);
1890 	return gap_end;
1891 }
1892 
1893 /* Get an address range which is currently unmapped.
1894  * For shmat() with addr=0.
1895  *
1896  * Ugly calling convention alert:
1897  * Return value with the low bits set means error value,
1898  * ie
1899  *	if (ret & ~PAGE_MASK)
1900  *		error = ret;
1901  *
1902  * This function "knows" that -ENOMEM has the bits set.
1903  */
1904 #ifndef HAVE_ARCH_UNMAPPED_AREA
1905 unsigned long
1906 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1907 		unsigned long len, unsigned long pgoff, unsigned long flags)
1908 {
1909 	struct mm_struct *mm = current->mm;
1910 	struct vm_area_struct *vma;
1911 	struct vm_unmapped_area_info info;
1912 
1913 	if (len > TASK_SIZE - mmap_min_addr)
1914 		return -ENOMEM;
1915 
1916 	if (flags & MAP_FIXED)
1917 		return addr;
1918 
1919 	if (addr) {
1920 		addr = PAGE_ALIGN(addr);
1921 		vma = find_vma(mm, addr);
1922 		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1923 		    (!vma || addr + len <= vma->vm_start))
1924 			return addr;
1925 	}
1926 
1927 	info.flags = 0;
1928 	info.length = len;
1929 	info.low_limit = mm->mmap_base;
1930 	info.high_limit = TASK_SIZE;
1931 	info.align_mask = 0;
1932 	return vm_unmapped_area(&info);
1933 }
1934 #endif
1935 
1936 /*
1937  * This mmap-allocator allocates new areas top-down from below the
1938  * stack's low limit (the base):
1939  */
1940 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1941 unsigned long
1942 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1943 			  const unsigned long len, const unsigned long pgoff,
1944 			  const unsigned long flags)
1945 {
1946 	struct vm_area_struct *vma;
1947 	struct mm_struct *mm = current->mm;
1948 	unsigned long addr = addr0;
1949 	struct vm_unmapped_area_info info;
1950 
1951 	/* requested length too big for entire address space */
1952 	if (len > TASK_SIZE - mmap_min_addr)
1953 		return -ENOMEM;
1954 
1955 	if (flags & MAP_FIXED)
1956 		return addr;
1957 
1958 	/* requesting a specific address */
1959 	if (addr) {
1960 		addr = PAGE_ALIGN(addr);
1961 		vma = find_vma(mm, addr);
1962 		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1963 				(!vma || addr + len <= vma->vm_start))
1964 			return addr;
1965 	}
1966 
1967 	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1968 	info.length = len;
1969 	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1970 	info.high_limit = mm->mmap_base;
1971 	info.align_mask = 0;
1972 	addr = vm_unmapped_area(&info);
1973 
1974 	/*
1975 	 * A failed mmap() very likely causes application failure,
1976 	 * so fall back to the bottom-up function here. This scenario
1977 	 * can happen with large stack limits and large mmap()
1978 	 * allocations.
1979 	 */
1980 	if (addr & ~PAGE_MASK) {
1981 		VM_BUG_ON(addr != -ENOMEM);
1982 		info.flags = 0;
1983 		info.low_limit = TASK_UNMAPPED_BASE;
1984 		info.high_limit = TASK_SIZE;
1985 		addr = vm_unmapped_area(&info);
1986 	}
1987 
1988 	return addr;
1989 }
1990 #endif
1991 
1992 unsigned long
1993 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1994 		unsigned long pgoff, unsigned long flags)
1995 {
1996 	unsigned long (*get_area)(struct file *, unsigned long,
1997 				  unsigned long, unsigned long, unsigned long);
1998 
1999 	unsigned long error = arch_mmap_check(addr, len, flags);
2000 	if (error)
2001 		return error;
2002 
2003 	/* Careful about overflows.. */
2004 	if (len > TASK_SIZE)
2005 		return -ENOMEM;
2006 
2007 	get_area = current->mm->get_unmapped_area;
2008 	if (file && file->f_op->get_unmapped_area)
2009 		get_area = file->f_op->get_unmapped_area;
2010 	addr = get_area(file, addr, len, pgoff, flags);
2011 	if (IS_ERR_VALUE(addr))
2012 		return addr;
2013 
2014 	if (addr > TASK_SIZE - len)
2015 		return -ENOMEM;
2016 	if (addr & ~PAGE_MASK)
2017 		return -EINVAL;
2018 
2019 	addr = arch_rebalance_pgtables(addr, len);
2020 	error = security_mmap_addr(addr);
2021 	return error ? error : addr;
2022 }
2023 
2024 EXPORT_SYMBOL(get_unmapped_area);
2025 
2026 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2027 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2028 {
2029 	struct rb_node *rb_node;
2030 	struct vm_area_struct *vma;
2031 
2032 	/* Check the cache first. */
2033 	vma = vmacache_find(mm, addr);
2034 	if (likely(vma))
2035 		return vma;
2036 
2037 	rb_node = mm->mm_rb.rb_node;
2038 	vma = NULL;
2039 
2040 	while (rb_node) {
2041 		struct vm_area_struct *tmp;
2042 
2043 		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2044 
2045 		if (tmp->vm_end > addr) {
2046 			vma = tmp;
2047 			if (tmp->vm_start <= addr)
2048 				break;
2049 			rb_node = rb_node->rb_left;
2050 		} else
2051 			rb_node = rb_node->rb_right;
2052 	}
2053 
2054 	if (vma)
2055 		vmacache_update(addr, vma);
2056 	return vma;
2057 }
2058 
2059 EXPORT_SYMBOL(find_vma);
2060 
2061 /*
2062  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2063  */
2064 struct vm_area_struct *
2065 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2066 			struct vm_area_struct **pprev)
2067 {
2068 	struct vm_area_struct *vma;
2069 
2070 	vma = find_vma(mm, addr);
2071 	if (vma) {
2072 		*pprev = vma->vm_prev;
2073 	} else {
2074 		struct rb_node *rb_node = mm->mm_rb.rb_node;
2075 		*pprev = NULL;
2076 		while (rb_node) {
2077 			*pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2078 			rb_node = rb_node->rb_right;
2079 		}
2080 	}
2081 	return vma;
2082 }
2083 
2084 /*
2085  * Verify that the stack growth is acceptable and
2086  * update accounting. This is shared with both the
2087  * grow-up and grow-down cases.
2088  */
2089 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2090 {
2091 	struct mm_struct *mm = vma->vm_mm;
2092 	struct rlimit *rlim = current->signal->rlim;
2093 	unsigned long new_start, actual_size;
2094 
2095 	/* address space limit tests */
2096 	if (!may_expand_vm(mm, grow))
2097 		return -ENOMEM;
2098 
2099 	/* Stack limit test */
2100 	actual_size = size;
2101 	if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
2102 		actual_size -= PAGE_SIZE;
2103 	if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2104 		return -ENOMEM;
2105 
2106 	/* mlock limit tests */
2107 	if (vma->vm_flags & VM_LOCKED) {
2108 		unsigned long locked;
2109 		unsigned long limit;
2110 		locked = mm->locked_vm + grow;
2111 		limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2112 		limit >>= PAGE_SHIFT;
2113 		if (locked > limit && !capable(CAP_IPC_LOCK))
2114 			return -ENOMEM;
2115 	}
2116 
2117 	/* Check to ensure the stack will not grow into a hugetlb-only region */
2118 	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2119 			vma->vm_end - size;
2120 	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2121 		return -EFAULT;
2122 
2123 	/*
2124 	 * Overcommit..  This must be the final test, as it will
2125 	 * update security statistics.
2126 	 */
2127 	if (security_vm_enough_memory_mm(mm, grow))
2128 		return -ENOMEM;
2129 
2130 	/* Ok, everything looks good - let it rip */
2131 	if (vma->vm_flags & VM_LOCKED)
2132 		mm->locked_vm += grow;
2133 	vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2134 	return 0;
2135 }
2136 
2137 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2138 /*
2139  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2140  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2141  */
2142 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2143 {
2144 	int error;
2145 
2146 	if (!(vma->vm_flags & VM_GROWSUP))
2147 		return -EFAULT;
2148 
2149 	/*
2150 	 * We must make sure the anon_vma is allocated
2151 	 * so that the anon_vma locking is not a noop.
2152 	 */
2153 	if (unlikely(anon_vma_prepare(vma)))
2154 		return -ENOMEM;
2155 	vma_lock_anon_vma(vma);
2156 
2157 	/*
2158 	 * vma->vm_start/vm_end cannot change under us because the caller
2159 	 * is required to hold the mmap_sem in read mode.  We need the
2160 	 * anon_vma lock to serialize against concurrent expand_stacks.
2161 	 * Also guard against wrapping around to address 0.
2162 	 */
2163 	if (address < PAGE_ALIGN(address+4))
2164 		address = PAGE_ALIGN(address+4);
2165 	else {
2166 		vma_unlock_anon_vma(vma);
2167 		return -ENOMEM;
2168 	}
2169 	error = 0;
2170 
2171 	/* Somebody else might have raced and expanded it already */
2172 	if (address > vma->vm_end) {
2173 		unsigned long size, grow;
2174 
2175 		size = address - vma->vm_start;
2176 		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2177 
2178 		error = -ENOMEM;
2179 		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2180 			error = acct_stack_growth(vma, size, grow);
2181 			if (!error) {
2182 				/*
2183 				 * vma_gap_update() doesn't support concurrent
2184 				 * updates, but we only hold a shared mmap_sem
2185 				 * lock here, so we need to protect against
2186 				 * concurrent vma expansions.
2187 				 * vma_lock_anon_vma() doesn't help here, as
2188 				 * we don't guarantee that all growable vmas
2189 				 * in a mm share the same root anon vma.
2190 				 * So, we reuse mm->page_table_lock to guard
2191 				 * against concurrent vma expansions.
2192 				 */
2193 				spin_lock(&vma->vm_mm->page_table_lock);
2194 				anon_vma_interval_tree_pre_update_vma(vma);
2195 				vma->vm_end = address;
2196 				anon_vma_interval_tree_post_update_vma(vma);
2197 				if (vma->vm_next)
2198 					vma_gap_update(vma->vm_next);
2199 				else
2200 					vma->vm_mm->highest_vm_end = address;
2201 				spin_unlock(&vma->vm_mm->page_table_lock);
2202 
2203 				perf_event_mmap(vma);
2204 			}
2205 		}
2206 	}
2207 	vma_unlock_anon_vma(vma);
2208 	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2209 	validate_mm(vma->vm_mm);
2210 	return error;
2211 }
2212 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2213 
2214 /*
2215  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2216  */
2217 int expand_downwards(struct vm_area_struct *vma,
2218 				   unsigned long address)
2219 {
2220 	int error;
2221 
2222 	/*
2223 	 * We must make sure the anon_vma is allocated
2224 	 * so that the anon_vma locking is not a noop.
2225 	 */
2226 	if (unlikely(anon_vma_prepare(vma)))
2227 		return -ENOMEM;
2228 
2229 	address &= PAGE_MASK;
2230 	error = security_mmap_addr(address);
2231 	if (error)
2232 		return error;
2233 
2234 	vma_lock_anon_vma(vma);
2235 
2236 	/*
2237 	 * vma->vm_start/vm_end cannot change under us because the caller
2238 	 * is required to hold the mmap_sem in read mode.  We need the
2239 	 * anon_vma lock to serialize against concurrent expand_stacks.
2240 	 */
2241 
2242 	/* Somebody else might have raced and expanded it already */
2243 	if (address < vma->vm_start) {
2244 		unsigned long size, grow;
2245 
2246 		size = vma->vm_end - address;
2247 		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2248 
2249 		error = -ENOMEM;
2250 		if (grow <= vma->vm_pgoff) {
2251 			error = acct_stack_growth(vma, size, grow);
2252 			if (!error) {
2253 				/*
2254 				 * vma_gap_update() doesn't support concurrent
2255 				 * updates, but we only hold a shared mmap_sem
2256 				 * lock here, so we need to protect against
2257 				 * concurrent vma expansions.
2258 				 * vma_lock_anon_vma() doesn't help here, as
2259 				 * we don't guarantee that all growable vmas
2260 				 * in a mm share the same root anon vma.
2261 				 * So, we reuse mm->page_table_lock to guard
2262 				 * against concurrent vma expansions.
2263 				 */
2264 				spin_lock(&vma->vm_mm->page_table_lock);
2265 				anon_vma_interval_tree_pre_update_vma(vma);
2266 				vma->vm_start = address;
2267 				vma->vm_pgoff -= grow;
2268 				anon_vma_interval_tree_post_update_vma(vma);
2269 				vma_gap_update(vma);
2270 				spin_unlock(&vma->vm_mm->page_table_lock);
2271 
2272 				perf_event_mmap(vma);
2273 			}
2274 		}
2275 	}
2276 	vma_unlock_anon_vma(vma);
2277 	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2278 	validate_mm(vma->vm_mm);
2279 	return error;
2280 }
2281 
2282 /*
2283  * Note how expand_stack() refuses to expand the stack all the way to
2284  * abut the next virtual mapping, *unless* that mapping itself is also
2285  * a stack mapping. We want to leave room for a guard page, after all
2286  * (the guard page itself is not added here, that is done by the
2287  * actual page faulting logic)
2288  *
2289  * This matches the behavior of the guard page logic (see mm/memory.c:
2290  * check_stack_guard_page()), which only allows the guard page to be
2291  * removed under these circumstances.
2292  */
2293 #ifdef CONFIG_STACK_GROWSUP
2294 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2295 {
2296 	struct vm_area_struct *next;
2297 
2298 	address &= PAGE_MASK;
2299 	next = vma->vm_next;
2300 	if (next && next->vm_start == address + PAGE_SIZE) {
2301 		if (!(next->vm_flags & VM_GROWSUP))
2302 			return -ENOMEM;
2303 	}
2304 	return expand_upwards(vma, address);
2305 }
2306 
2307 struct vm_area_struct *
2308 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2309 {
2310 	struct vm_area_struct *vma, *prev;
2311 
2312 	addr &= PAGE_MASK;
2313 	vma = find_vma_prev(mm, addr, &prev);
2314 	if (vma && (vma->vm_start <= addr))
2315 		return vma;
2316 	if (!prev || expand_stack(prev, addr))
2317 		return NULL;
2318 	if (prev->vm_flags & VM_LOCKED)
2319 		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2320 	return prev;
2321 }
2322 #else
2323 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2324 {
2325 	struct vm_area_struct *prev;
2326 
2327 	address &= PAGE_MASK;
2328 	prev = vma->vm_prev;
2329 	if (prev && prev->vm_end == address) {
2330 		if (!(prev->vm_flags & VM_GROWSDOWN))
2331 			return -ENOMEM;
2332 	}
2333 	return expand_downwards(vma, address);
2334 }
2335 
2336 struct vm_area_struct *
2337 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2338 {
2339 	struct vm_area_struct *vma;
2340 	unsigned long start;
2341 
2342 	addr &= PAGE_MASK;
2343 	vma = find_vma(mm, addr);
2344 	if (!vma)
2345 		return NULL;
2346 	if (vma->vm_start <= addr)
2347 		return vma;
2348 	if (!(vma->vm_flags & VM_GROWSDOWN))
2349 		return NULL;
2350 	start = vma->vm_start;
2351 	if (expand_stack(vma, addr))
2352 		return NULL;
2353 	if (vma->vm_flags & VM_LOCKED)
2354 		populate_vma_page_range(vma, addr, start, NULL);
2355 	return vma;
2356 }
2357 #endif
2358 
2359 EXPORT_SYMBOL_GPL(find_extend_vma);
2360 
2361 /*
2362  * Ok - we have the memory areas we should free on the vma list,
2363  * so release them, and do the vma updates.
2364  *
2365  * Called with the mm semaphore held.
2366  */
2367 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2368 {
2369 	unsigned long nr_accounted = 0;
2370 
2371 	/* Update high watermark before we lower total_vm */
2372 	update_hiwater_vm(mm);
2373 	do {
2374 		long nrpages = vma_pages(vma);
2375 
2376 		if (vma->vm_flags & VM_ACCOUNT)
2377 			nr_accounted += nrpages;
2378 		vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2379 		vma = remove_vma(vma);
2380 	} while (vma);
2381 	vm_unacct_memory(nr_accounted);
2382 	validate_mm(mm);
2383 }
2384 
2385 /*
2386  * Get rid of page table information in the indicated region.
2387  *
2388  * Called with the mm semaphore held.
2389  */
2390 static void unmap_region(struct mm_struct *mm,
2391 		struct vm_area_struct *vma, struct vm_area_struct *prev,
2392 		unsigned long start, unsigned long end)
2393 {
2394 	struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2395 	struct mmu_gather tlb;
2396 
2397 	lru_add_drain();
2398 	tlb_gather_mmu(&tlb, mm, start, end);
2399 	update_hiwater_rss(mm);
2400 	unmap_vmas(&tlb, vma, start, end);
2401 	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2402 				 next ? next->vm_start : USER_PGTABLES_CEILING);
2403 	tlb_finish_mmu(&tlb, start, end);
2404 }
2405 
2406 /*
2407  * Create a list of vma's touched by the unmap, removing them from the mm's
2408  * vma list as we go..
2409  */
2410 static void
2411 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2412 	struct vm_area_struct *prev, unsigned long end)
2413 {
2414 	struct vm_area_struct **insertion_point;
2415 	struct vm_area_struct *tail_vma = NULL;
2416 
2417 	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2418 	vma->vm_prev = NULL;
2419 	do {
2420 		vma_rb_erase(vma, &mm->mm_rb);
2421 		mm->map_count--;
2422 		tail_vma = vma;
2423 		vma = vma->vm_next;
2424 	} while (vma && vma->vm_start < end);
2425 	*insertion_point = vma;
2426 	if (vma) {
2427 		vma->vm_prev = prev;
2428 		vma_gap_update(vma);
2429 	} else
2430 		mm->highest_vm_end = prev ? prev->vm_end : 0;
2431 	tail_vma->vm_next = NULL;
2432 
2433 	/* Kill the cache */
2434 	vmacache_invalidate(mm);
2435 }
2436 
2437 /*
2438  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2439  * munmap path where it doesn't make sense to fail.
2440  */
2441 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2442 	      unsigned long addr, int new_below)
2443 {
2444 	struct vm_area_struct *new;
2445 	int err = -ENOMEM;
2446 
2447 	if (is_vm_hugetlb_page(vma) && (addr &
2448 					~(huge_page_mask(hstate_vma(vma)))))
2449 		return -EINVAL;
2450 
2451 	new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2452 	if (!new)
2453 		goto out_err;
2454 
2455 	/* most fields are the same, copy all, and then fixup */
2456 	*new = *vma;
2457 
2458 	INIT_LIST_HEAD(&new->anon_vma_chain);
2459 
2460 	if (new_below)
2461 		new->vm_end = addr;
2462 	else {
2463 		new->vm_start = addr;
2464 		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2465 	}
2466 
2467 	err = vma_dup_policy(vma, new);
2468 	if (err)
2469 		goto out_free_vma;
2470 
2471 	err = anon_vma_clone(new, vma);
2472 	if (err)
2473 		goto out_free_mpol;
2474 
2475 	if (new->vm_file)
2476 		get_file(new->vm_file);
2477 
2478 	if (new->vm_ops && new->vm_ops->open)
2479 		new->vm_ops->open(new);
2480 
2481 	if (new_below)
2482 		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2483 			((addr - new->vm_start) >> PAGE_SHIFT), new);
2484 	else
2485 		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2486 
2487 	/* Success. */
2488 	if (!err)
2489 		return 0;
2490 
2491 	/* Clean everything up if vma_adjust failed. */
2492 	if (new->vm_ops && new->vm_ops->close)
2493 		new->vm_ops->close(new);
2494 	if (new->vm_file)
2495 		fput(new->vm_file);
2496 	unlink_anon_vmas(new);
2497  out_free_mpol:
2498 	mpol_put(vma_policy(new));
2499  out_free_vma:
2500 	kmem_cache_free(vm_area_cachep, new);
2501  out_err:
2502 	return err;
2503 }
2504 
2505 /*
2506  * Split a vma into two pieces at address 'addr', a new vma is allocated
2507  * either for the first part or the tail.
2508  */
2509 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2510 	      unsigned long addr, int new_below)
2511 {
2512 	if (mm->map_count >= sysctl_max_map_count)
2513 		return -ENOMEM;
2514 
2515 	return __split_vma(mm, vma, addr, new_below);
2516 }
2517 
2518 /* Munmap is split into 2 main parts -- this part which finds
2519  * what needs doing, and the areas themselves, which do the
2520  * work.  This now handles partial unmappings.
2521  * Jeremy Fitzhardinge <jeremy@goop.org>
2522  */
2523 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2524 {
2525 	unsigned long end;
2526 	struct vm_area_struct *vma, *prev, *last;
2527 
2528 	if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2529 		return -EINVAL;
2530 
2531 	len = PAGE_ALIGN(len);
2532 	if (len == 0)
2533 		return -EINVAL;
2534 
2535 	/* Find the first overlapping VMA */
2536 	vma = find_vma(mm, start);
2537 	if (!vma)
2538 		return 0;
2539 	prev = vma->vm_prev;
2540 	/* we have  start < vma->vm_end  */
2541 
2542 	/* if it doesn't overlap, we have nothing.. */
2543 	end = start + len;
2544 	if (vma->vm_start >= end)
2545 		return 0;
2546 
2547 	/*
2548 	 * If we need to split any vma, do it now to save pain later.
2549 	 *
2550 	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2551 	 * unmapped vm_area_struct will remain in use: so lower split_vma
2552 	 * places tmp vma above, and higher split_vma places tmp vma below.
2553 	 */
2554 	if (start > vma->vm_start) {
2555 		int error;
2556 
2557 		/*
2558 		 * Make sure that map_count on return from munmap() will
2559 		 * not exceed its limit; but let map_count go just above
2560 		 * its limit temporarily, to help free resources as expected.
2561 		 */
2562 		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2563 			return -ENOMEM;
2564 
2565 		error = __split_vma(mm, vma, start, 0);
2566 		if (error)
2567 			return error;
2568 		prev = vma;
2569 	}
2570 
2571 	/* Does it split the last one? */
2572 	last = find_vma(mm, end);
2573 	if (last && end > last->vm_start) {
2574 		int error = __split_vma(mm, last, end, 1);
2575 		if (error)
2576 			return error;
2577 	}
2578 	vma = prev ? prev->vm_next : mm->mmap;
2579 
2580 	/*
2581 	 * unlock any mlock()ed ranges before detaching vmas
2582 	 */
2583 	if (mm->locked_vm) {
2584 		struct vm_area_struct *tmp = vma;
2585 		while (tmp && tmp->vm_start < end) {
2586 			if (tmp->vm_flags & VM_LOCKED) {
2587 				mm->locked_vm -= vma_pages(tmp);
2588 				munlock_vma_pages_all(tmp);
2589 			}
2590 			tmp = tmp->vm_next;
2591 		}
2592 	}
2593 
2594 	/*
2595 	 * Remove the vma's, and unmap the actual pages
2596 	 */
2597 	detach_vmas_to_be_unmapped(mm, vma, prev, end);
2598 	unmap_region(mm, vma, prev, start, end);
2599 
2600 	arch_unmap(mm, vma, start, end);
2601 
2602 	/* Fix up all other VM information */
2603 	remove_vma_list(mm, vma);
2604 
2605 	return 0;
2606 }
2607 
2608 int vm_munmap(unsigned long start, size_t len)
2609 {
2610 	int ret;
2611 	struct mm_struct *mm = current->mm;
2612 
2613 	down_write(&mm->mmap_sem);
2614 	ret = do_munmap(mm, start, len);
2615 	up_write(&mm->mmap_sem);
2616 	return ret;
2617 }
2618 EXPORT_SYMBOL(vm_munmap);
2619 
2620 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2621 {
2622 	profile_munmap(addr);
2623 	return vm_munmap(addr, len);
2624 }
2625 
2626 
2627 /*
2628  * Emulation of deprecated remap_file_pages() syscall.
2629  */
2630 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2631 		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2632 {
2633 
2634 	struct mm_struct *mm = current->mm;
2635 	struct vm_area_struct *vma;
2636 	unsigned long populate = 0;
2637 	unsigned long ret = -EINVAL;
2638 	struct file *file;
2639 
2640 	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2641 			"See Documentation/vm/remap_file_pages.txt.\n",
2642 			current->comm, current->pid);
2643 
2644 	if (prot)
2645 		return ret;
2646 	start = start & PAGE_MASK;
2647 	size = size & PAGE_MASK;
2648 
2649 	if (start + size <= start)
2650 		return ret;
2651 
2652 	/* Does pgoff wrap? */
2653 	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2654 		return ret;
2655 
2656 	down_write(&mm->mmap_sem);
2657 	vma = find_vma(mm, start);
2658 
2659 	if (!vma || !(vma->vm_flags & VM_SHARED))
2660 		goto out;
2661 
2662 	if (start < vma->vm_start || start + size > vma->vm_end)
2663 		goto out;
2664 
2665 	if (pgoff == linear_page_index(vma, start)) {
2666 		ret = 0;
2667 		goto out;
2668 	}
2669 
2670 	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2671 	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2672 	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2673 
2674 	flags &= MAP_NONBLOCK;
2675 	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2676 	if (vma->vm_flags & VM_LOCKED) {
2677 		flags |= MAP_LOCKED;
2678 		/* drop PG_Mlocked flag for over-mapped range */
2679 		munlock_vma_pages_range(vma, start, start + size);
2680 	}
2681 
2682 	file = get_file(vma->vm_file);
2683 	ret = do_mmap_pgoff(vma->vm_file, start, size,
2684 			prot, flags, pgoff, &populate);
2685 	fput(file);
2686 out:
2687 	up_write(&mm->mmap_sem);
2688 	if (populate)
2689 		mm_populate(ret, populate);
2690 	if (!IS_ERR_VALUE(ret))
2691 		ret = 0;
2692 	return ret;
2693 }
2694 
2695 static inline void verify_mm_writelocked(struct mm_struct *mm)
2696 {
2697 #ifdef CONFIG_DEBUG_VM
2698 	if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2699 		WARN_ON(1);
2700 		up_read(&mm->mmap_sem);
2701 	}
2702 #endif
2703 }
2704 
2705 /*
2706  *  this is really a simplified "do_mmap".  it only handles
2707  *  anonymous maps.  eventually we may be able to do some
2708  *  brk-specific accounting here.
2709  */
2710 static unsigned long do_brk(unsigned long addr, unsigned long len)
2711 {
2712 	struct mm_struct *mm = current->mm;
2713 	struct vm_area_struct *vma, *prev;
2714 	unsigned long flags;
2715 	struct rb_node **rb_link, *rb_parent;
2716 	pgoff_t pgoff = addr >> PAGE_SHIFT;
2717 	int error;
2718 
2719 	len = PAGE_ALIGN(len);
2720 	if (!len)
2721 		return addr;
2722 
2723 	flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2724 
2725 	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2726 	if (error & ~PAGE_MASK)
2727 		return error;
2728 
2729 	error = mlock_future_check(mm, mm->def_flags, len);
2730 	if (error)
2731 		return error;
2732 
2733 	/*
2734 	 * mm->mmap_sem is required to protect against another thread
2735 	 * changing the mappings in case we sleep.
2736 	 */
2737 	verify_mm_writelocked(mm);
2738 
2739 	/*
2740 	 * Clear old maps.  this also does some error checking for us
2741 	 */
2742 	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2743 			      &rb_parent)) {
2744 		if (do_munmap(mm, addr, len))
2745 			return -ENOMEM;
2746 	}
2747 
2748 	/* Check against address space limits *after* clearing old maps... */
2749 	if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2750 		return -ENOMEM;
2751 
2752 	if (mm->map_count > sysctl_max_map_count)
2753 		return -ENOMEM;
2754 
2755 	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2756 		return -ENOMEM;
2757 
2758 	/* Can we just expand an old private anonymous mapping? */
2759 	vma = vma_merge(mm, prev, addr, addr + len, flags,
2760 					NULL, NULL, pgoff, NULL);
2761 	if (vma)
2762 		goto out;
2763 
2764 	/*
2765 	 * create a vma struct for an anonymous mapping
2766 	 */
2767 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2768 	if (!vma) {
2769 		vm_unacct_memory(len >> PAGE_SHIFT);
2770 		return -ENOMEM;
2771 	}
2772 
2773 	INIT_LIST_HEAD(&vma->anon_vma_chain);
2774 	vma->vm_mm = mm;
2775 	vma->vm_start = addr;
2776 	vma->vm_end = addr + len;
2777 	vma->vm_pgoff = pgoff;
2778 	vma->vm_flags = flags;
2779 	vma->vm_page_prot = vm_get_page_prot(flags);
2780 	vma_link(mm, vma, prev, rb_link, rb_parent);
2781 out:
2782 	perf_event_mmap(vma);
2783 	mm->total_vm += len >> PAGE_SHIFT;
2784 	if (flags & VM_LOCKED)
2785 		mm->locked_vm += (len >> PAGE_SHIFT);
2786 	vma->vm_flags |= VM_SOFTDIRTY;
2787 	return addr;
2788 }
2789 
2790 unsigned long vm_brk(unsigned long addr, unsigned long len)
2791 {
2792 	struct mm_struct *mm = current->mm;
2793 	unsigned long ret;
2794 	bool populate;
2795 
2796 	down_write(&mm->mmap_sem);
2797 	ret = do_brk(addr, len);
2798 	populate = ((mm->def_flags & VM_LOCKED) != 0);
2799 	up_write(&mm->mmap_sem);
2800 	if (populate)
2801 		mm_populate(addr, len);
2802 	return ret;
2803 }
2804 EXPORT_SYMBOL(vm_brk);
2805 
2806 /* Release all mmaps. */
2807 void exit_mmap(struct mm_struct *mm)
2808 {
2809 	struct mmu_gather tlb;
2810 	struct vm_area_struct *vma;
2811 	unsigned long nr_accounted = 0;
2812 
2813 	/* mm's last user has gone, and its about to be pulled down */
2814 	mmu_notifier_release(mm);
2815 
2816 	if (mm->locked_vm) {
2817 		vma = mm->mmap;
2818 		while (vma) {
2819 			if (vma->vm_flags & VM_LOCKED)
2820 				munlock_vma_pages_all(vma);
2821 			vma = vma->vm_next;
2822 		}
2823 	}
2824 
2825 	arch_exit_mmap(mm);
2826 
2827 	vma = mm->mmap;
2828 	if (!vma)	/* Can happen if dup_mmap() received an OOM */
2829 		return;
2830 
2831 	lru_add_drain();
2832 	flush_cache_mm(mm);
2833 	tlb_gather_mmu(&tlb, mm, 0, -1);
2834 	/* update_hiwater_rss(mm) here? but nobody should be looking */
2835 	/* Use -1 here to ensure all VMAs in the mm are unmapped */
2836 	unmap_vmas(&tlb, vma, 0, -1);
2837 
2838 	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2839 	tlb_finish_mmu(&tlb, 0, -1);
2840 
2841 	/*
2842 	 * Walk the list again, actually closing and freeing it,
2843 	 * with preemption enabled, without holding any MM locks.
2844 	 */
2845 	while (vma) {
2846 		if (vma->vm_flags & VM_ACCOUNT)
2847 			nr_accounted += vma_pages(vma);
2848 		vma = remove_vma(vma);
2849 	}
2850 	vm_unacct_memory(nr_accounted);
2851 }
2852 
2853 /* Insert vm structure into process list sorted by address
2854  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2855  * then i_mmap_rwsem is taken here.
2856  */
2857 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2858 {
2859 	struct vm_area_struct *prev;
2860 	struct rb_node **rb_link, *rb_parent;
2861 
2862 	/*
2863 	 * The vm_pgoff of a purely anonymous vma should be irrelevant
2864 	 * until its first write fault, when page's anon_vma and index
2865 	 * are set.  But now set the vm_pgoff it will almost certainly
2866 	 * end up with (unless mremap moves it elsewhere before that
2867 	 * first wfault), so /proc/pid/maps tells a consistent story.
2868 	 *
2869 	 * By setting it to reflect the virtual start address of the
2870 	 * vma, merges and splits can happen in a seamless way, just
2871 	 * using the existing file pgoff checks and manipulations.
2872 	 * Similarly in do_mmap_pgoff and in do_brk.
2873 	 */
2874 	if (!vma->vm_file) {
2875 		BUG_ON(vma->anon_vma);
2876 		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2877 	}
2878 	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2879 			   &prev, &rb_link, &rb_parent))
2880 		return -ENOMEM;
2881 	if ((vma->vm_flags & VM_ACCOUNT) &&
2882 	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
2883 		return -ENOMEM;
2884 
2885 	vma_link(mm, vma, prev, rb_link, rb_parent);
2886 	return 0;
2887 }
2888 
2889 /*
2890  * Copy the vma structure to a new location in the same mm,
2891  * prior to moving page table entries, to effect an mremap move.
2892  */
2893 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2894 	unsigned long addr, unsigned long len, pgoff_t pgoff,
2895 	bool *need_rmap_locks)
2896 {
2897 	struct vm_area_struct *vma = *vmap;
2898 	unsigned long vma_start = vma->vm_start;
2899 	struct mm_struct *mm = vma->vm_mm;
2900 	struct vm_area_struct *new_vma, *prev;
2901 	struct rb_node **rb_link, *rb_parent;
2902 	bool faulted_in_anon_vma = true;
2903 
2904 	/*
2905 	 * If anonymous vma has not yet been faulted, update new pgoff
2906 	 * to match new location, to increase its chance of merging.
2907 	 */
2908 	if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2909 		pgoff = addr >> PAGE_SHIFT;
2910 		faulted_in_anon_vma = false;
2911 	}
2912 
2913 	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2914 		return NULL;	/* should never get here */
2915 	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2916 			vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2917 	if (new_vma) {
2918 		/*
2919 		 * Source vma may have been merged into new_vma
2920 		 */
2921 		if (unlikely(vma_start >= new_vma->vm_start &&
2922 			     vma_start < new_vma->vm_end)) {
2923 			/*
2924 			 * The only way we can get a vma_merge with
2925 			 * self during an mremap is if the vma hasn't
2926 			 * been faulted in yet and we were allowed to
2927 			 * reset the dst vma->vm_pgoff to the
2928 			 * destination address of the mremap to allow
2929 			 * the merge to happen. mremap must change the
2930 			 * vm_pgoff linearity between src and dst vmas
2931 			 * (in turn preventing a vma_merge) to be
2932 			 * safe. It is only safe to keep the vm_pgoff
2933 			 * linear if there are no pages mapped yet.
2934 			 */
2935 			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
2936 			*vmap = vma = new_vma;
2937 		}
2938 		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2939 	} else {
2940 		new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2941 		if (new_vma) {
2942 			*new_vma = *vma;
2943 			new_vma->vm_start = addr;
2944 			new_vma->vm_end = addr + len;
2945 			new_vma->vm_pgoff = pgoff;
2946 			if (vma_dup_policy(vma, new_vma))
2947 				goto out_free_vma;
2948 			INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2949 			if (anon_vma_clone(new_vma, vma))
2950 				goto out_free_mempol;
2951 			if (new_vma->vm_file)
2952 				get_file(new_vma->vm_file);
2953 			if (new_vma->vm_ops && new_vma->vm_ops->open)
2954 				new_vma->vm_ops->open(new_vma);
2955 			vma_link(mm, new_vma, prev, rb_link, rb_parent);
2956 			*need_rmap_locks = false;
2957 		}
2958 	}
2959 	return new_vma;
2960 
2961  out_free_mempol:
2962 	mpol_put(vma_policy(new_vma));
2963  out_free_vma:
2964 	kmem_cache_free(vm_area_cachep, new_vma);
2965 	return NULL;
2966 }
2967 
2968 /*
2969  * Return true if the calling process may expand its vm space by the passed
2970  * number of pages
2971  */
2972 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2973 {
2974 	unsigned long cur = mm->total_vm;	/* pages */
2975 	unsigned long lim;
2976 
2977 	lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2978 
2979 	if (cur + npages > lim)
2980 		return 0;
2981 	return 1;
2982 }
2983 
2984 static int special_mapping_fault(struct vm_area_struct *vma,
2985 				 struct vm_fault *vmf);
2986 
2987 /*
2988  * Having a close hook prevents vma merging regardless of flags.
2989  */
2990 static void special_mapping_close(struct vm_area_struct *vma)
2991 {
2992 }
2993 
2994 static const char *special_mapping_name(struct vm_area_struct *vma)
2995 {
2996 	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
2997 }
2998 
2999 static const struct vm_operations_struct special_mapping_vmops = {
3000 	.close = special_mapping_close,
3001 	.fault = special_mapping_fault,
3002 	.name = special_mapping_name,
3003 };
3004 
3005 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3006 	.close = special_mapping_close,
3007 	.fault = special_mapping_fault,
3008 };
3009 
3010 static int special_mapping_fault(struct vm_area_struct *vma,
3011 				struct vm_fault *vmf)
3012 {
3013 	pgoff_t pgoff;
3014 	struct page **pages;
3015 
3016 	/*
3017 	 * special mappings have no vm_file, and in that case, the mm
3018 	 * uses vm_pgoff internally. So we have to subtract it from here.
3019 	 * We are allowed to do this because we are the mm; do not copy
3020 	 * this code into drivers!
3021 	 */
3022 	pgoff = vmf->pgoff - vma->vm_pgoff;
3023 
3024 	if (vma->vm_ops == &legacy_special_mapping_vmops)
3025 		pages = vma->vm_private_data;
3026 	else
3027 		pages = ((struct vm_special_mapping *)vma->vm_private_data)->
3028 			pages;
3029 
3030 	for (; pgoff && *pages; ++pages)
3031 		pgoff--;
3032 
3033 	if (*pages) {
3034 		struct page *page = *pages;
3035 		get_page(page);
3036 		vmf->page = page;
3037 		return 0;
3038 	}
3039 
3040 	return VM_FAULT_SIGBUS;
3041 }
3042 
3043 static struct vm_area_struct *__install_special_mapping(
3044 	struct mm_struct *mm,
3045 	unsigned long addr, unsigned long len,
3046 	unsigned long vm_flags, const struct vm_operations_struct *ops,
3047 	void *priv)
3048 {
3049 	int ret;
3050 	struct vm_area_struct *vma;
3051 
3052 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3053 	if (unlikely(vma == NULL))
3054 		return ERR_PTR(-ENOMEM);
3055 
3056 	INIT_LIST_HEAD(&vma->anon_vma_chain);
3057 	vma->vm_mm = mm;
3058 	vma->vm_start = addr;
3059 	vma->vm_end = addr + len;
3060 
3061 	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3062 	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3063 
3064 	vma->vm_ops = ops;
3065 	vma->vm_private_data = priv;
3066 
3067 	ret = insert_vm_struct(mm, vma);
3068 	if (ret)
3069 		goto out;
3070 
3071 	mm->total_vm += len >> PAGE_SHIFT;
3072 
3073 	perf_event_mmap(vma);
3074 
3075 	return vma;
3076 
3077 out:
3078 	kmem_cache_free(vm_area_cachep, vma);
3079 	return ERR_PTR(ret);
3080 }
3081 
3082 /*
3083  * Called with mm->mmap_sem held for writing.
3084  * Insert a new vma covering the given region, with the given flags.
3085  * Its pages are supplied by the given array of struct page *.
3086  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3087  * The region past the last page supplied will always produce SIGBUS.
3088  * The array pointer and the pages it points to are assumed to stay alive
3089  * for as long as this mapping might exist.
3090  */
3091 struct vm_area_struct *_install_special_mapping(
3092 	struct mm_struct *mm,
3093 	unsigned long addr, unsigned long len,
3094 	unsigned long vm_flags, const struct vm_special_mapping *spec)
3095 {
3096 	return __install_special_mapping(mm, addr, len, vm_flags,
3097 					 &special_mapping_vmops, (void *)spec);
3098 }
3099 
3100 int install_special_mapping(struct mm_struct *mm,
3101 			    unsigned long addr, unsigned long len,
3102 			    unsigned long vm_flags, struct page **pages)
3103 {
3104 	struct vm_area_struct *vma = __install_special_mapping(
3105 		mm, addr, len, vm_flags, &legacy_special_mapping_vmops,
3106 		(void *)pages);
3107 
3108 	return PTR_ERR_OR_ZERO(vma);
3109 }
3110 
3111 static DEFINE_MUTEX(mm_all_locks_mutex);
3112 
3113 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3114 {
3115 	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3116 		/*
3117 		 * The LSB of head.next can't change from under us
3118 		 * because we hold the mm_all_locks_mutex.
3119 		 */
3120 		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3121 		/*
3122 		 * We can safely modify head.next after taking the
3123 		 * anon_vma->root->rwsem. If some other vma in this mm shares
3124 		 * the same anon_vma we won't take it again.
3125 		 *
3126 		 * No need of atomic instructions here, head.next
3127 		 * can't change from under us thanks to the
3128 		 * anon_vma->root->rwsem.
3129 		 */
3130 		if (__test_and_set_bit(0, (unsigned long *)
3131 				       &anon_vma->root->rb_root.rb_node))
3132 			BUG();
3133 	}
3134 }
3135 
3136 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3137 {
3138 	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3139 		/*
3140 		 * AS_MM_ALL_LOCKS can't change from under us because
3141 		 * we hold the mm_all_locks_mutex.
3142 		 *
3143 		 * Operations on ->flags have to be atomic because
3144 		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3145 		 * mm_all_locks_mutex, there may be other cpus
3146 		 * changing other bitflags in parallel to us.
3147 		 */
3148 		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3149 			BUG();
3150 		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3151 	}
3152 }
3153 
3154 /*
3155  * This operation locks against the VM for all pte/vma/mm related
3156  * operations that could ever happen on a certain mm. This includes
3157  * vmtruncate, try_to_unmap, and all page faults.
3158  *
3159  * The caller must take the mmap_sem in write mode before calling
3160  * mm_take_all_locks(). The caller isn't allowed to release the
3161  * mmap_sem until mm_drop_all_locks() returns.
3162  *
3163  * mmap_sem in write mode is required in order to block all operations
3164  * that could modify pagetables and free pages without need of
3165  * altering the vma layout. It's also needed in write mode to avoid new
3166  * anon_vmas to be associated with existing vmas.
3167  *
3168  * A single task can't take more than one mm_take_all_locks() in a row
3169  * or it would deadlock.
3170  *
3171  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3172  * mapping->flags avoid to take the same lock twice, if more than one
3173  * vma in this mm is backed by the same anon_vma or address_space.
3174  *
3175  * We can take all the locks in random order because the VM code
3176  * taking i_mmap_rwsem or anon_vma->rwsem outside the mmap_sem never
3177  * takes more than one of them in a row. Secondly we're protected
3178  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3179  *
3180  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3181  * that may have to take thousand of locks.
3182  *
3183  * mm_take_all_locks() can fail if it's interrupted by signals.
3184  */
3185 int mm_take_all_locks(struct mm_struct *mm)
3186 {
3187 	struct vm_area_struct *vma;
3188 	struct anon_vma_chain *avc;
3189 
3190 	BUG_ON(down_read_trylock(&mm->mmap_sem));
3191 
3192 	mutex_lock(&mm_all_locks_mutex);
3193 
3194 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3195 		if (signal_pending(current))
3196 			goto out_unlock;
3197 		if (vma->vm_file && vma->vm_file->f_mapping)
3198 			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3199 	}
3200 
3201 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3202 		if (signal_pending(current))
3203 			goto out_unlock;
3204 		if (vma->anon_vma)
3205 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3206 				vm_lock_anon_vma(mm, avc->anon_vma);
3207 	}
3208 
3209 	return 0;
3210 
3211 out_unlock:
3212 	mm_drop_all_locks(mm);
3213 	return -EINTR;
3214 }
3215 
3216 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3217 {
3218 	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3219 		/*
3220 		 * The LSB of head.next can't change to 0 from under
3221 		 * us because we hold the mm_all_locks_mutex.
3222 		 *
3223 		 * We must however clear the bitflag before unlocking
3224 		 * the vma so the users using the anon_vma->rb_root will
3225 		 * never see our bitflag.
3226 		 *
3227 		 * No need of atomic instructions here, head.next
3228 		 * can't change from under us until we release the
3229 		 * anon_vma->root->rwsem.
3230 		 */
3231 		if (!__test_and_clear_bit(0, (unsigned long *)
3232 					  &anon_vma->root->rb_root.rb_node))
3233 			BUG();
3234 		anon_vma_unlock_write(anon_vma);
3235 	}
3236 }
3237 
3238 static void vm_unlock_mapping(struct address_space *mapping)
3239 {
3240 	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3241 		/*
3242 		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3243 		 * because we hold the mm_all_locks_mutex.
3244 		 */
3245 		i_mmap_unlock_write(mapping);
3246 		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3247 					&mapping->flags))
3248 			BUG();
3249 	}
3250 }
3251 
3252 /*
3253  * The mmap_sem cannot be released by the caller until
3254  * mm_drop_all_locks() returns.
3255  */
3256 void mm_drop_all_locks(struct mm_struct *mm)
3257 {
3258 	struct vm_area_struct *vma;
3259 	struct anon_vma_chain *avc;
3260 
3261 	BUG_ON(down_read_trylock(&mm->mmap_sem));
3262 	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3263 
3264 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3265 		if (vma->anon_vma)
3266 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3267 				vm_unlock_anon_vma(avc->anon_vma);
3268 		if (vma->vm_file && vma->vm_file->f_mapping)
3269 			vm_unlock_mapping(vma->vm_file->f_mapping);
3270 	}
3271 
3272 	mutex_unlock(&mm_all_locks_mutex);
3273 }
3274 
3275 /*
3276  * initialise the VMA slab
3277  */
3278 void __init mmap_init(void)
3279 {
3280 	int ret;
3281 
3282 	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3283 	VM_BUG_ON(ret);
3284 }
3285 
3286 /*
3287  * Initialise sysctl_user_reserve_kbytes.
3288  *
3289  * This is intended to prevent a user from starting a single memory hogging
3290  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3291  * mode.
3292  *
3293  * The default value is min(3% of free memory, 128MB)
3294  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3295  */
3296 static int init_user_reserve(void)
3297 {
3298 	unsigned long free_kbytes;
3299 
3300 	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3301 
3302 	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3303 	return 0;
3304 }
3305 subsys_initcall(init_user_reserve);
3306 
3307 /*
3308  * Initialise sysctl_admin_reserve_kbytes.
3309  *
3310  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3311  * to log in and kill a memory hogging process.
3312  *
3313  * Systems with more than 256MB will reserve 8MB, enough to recover
3314  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3315  * only reserve 3% of free pages by default.
3316  */
3317 static int init_admin_reserve(void)
3318 {
3319 	unsigned long free_kbytes;
3320 
3321 	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3322 
3323 	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3324 	return 0;
3325 }
3326 subsys_initcall(init_admin_reserve);
3327 
3328 /*
3329  * Reinititalise user and admin reserves if memory is added or removed.
3330  *
3331  * The default user reserve max is 128MB, and the default max for the
3332  * admin reserve is 8MB. These are usually, but not always, enough to
3333  * enable recovery from a memory hogging process using login/sshd, a shell,
3334  * and tools like top. It may make sense to increase or even disable the
3335  * reserve depending on the existence of swap or variations in the recovery
3336  * tools. So, the admin may have changed them.
3337  *
3338  * If memory is added and the reserves have been eliminated or increased above
3339  * the default max, then we'll trust the admin.
3340  *
3341  * If memory is removed and there isn't enough free memory, then we
3342  * need to reset the reserves.
3343  *
3344  * Otherwise keep the reserve set by the admin.
3345  */
3346 static int reserve_mem_notifier(struct notifier_block *nb,
3347 			     unsigned long action, void *data)
3348 {
3349 	unsigned long tmp, free_kbytes;
3350 
3351 	switch (action) {
3352 	case MEM_ONLINE:
3353 		/* Default max is 128MB. Leave alone if modified by operator. */
3354 		tmp = sysctl_user_reserve_kbytes;
3355 		if (0 < tmp && tmp < (1UL << 17))
3356 			init_user_reserve();
3357 
3358 		/* Default max is 8MB.  Leave alone if modified by operator. */
3359 		tmp = sysctl_admin_reserve_kbytes;
3360 		if (0 < tmp && tmp < (1UL << 13))
3361 			init_admin_reserve();
3362 
3363 		break;
3364 	case MEM_OFFLINE:
3365 		free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3366 
3367 		if (sysctl_user_reserve_kbytes > free_kbytes) {
3368 			init_user_reserve();
3369 			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3370 				sysctl_user_reserve_kbytes);
3371 		}
3372 
3373 		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3374 			init_admin_reserve();
3375 			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3376 				sysctl_admin_reserve_kbytes);
3377 		}
3378 		break;
3379 	default:
3380 		break;
3381 	}
3382 	return NOTIFY_OK;
3383 }
3384 
3385 static struct notifier_block reserve_mem_nb = {
3386 	.notifier_call = reserve_mem_notifier,
3387 };
3388 
3389 static int __meminit init_reserve_notifier(void)
3390 {
3391 	if (register_hotmemory_notifier(&reserve_mem_nb))
3392 		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3393 
3394 	return 0;
3395 }
3396 subsys_initcall(init_reserve_notifier);
3397