xref: /linux/mm/mmap.c (revision 62b5f7d013fc455b8db26cf01e421f4c0d264b92)
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 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
47 
48 #include <asm/uaccess.h>
49 #include <asm/cacheflush.h>
50 #include <asm/tlb.h>
51 #include <asm/mmu_context.h>
52 
53 #include "internal.h"
54 
55 #ifndef arch_mmap_check
56 #define arch_mmap_check(addr, len, flags)	(0)
57 #endif
58 
59 #ifndef arch_rebalance_pgtables
60 #define arch_rebalance_pgtables(addr, len)		(addr)
61 #endif
62 
63 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
64 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
65 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
66 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
67 #endif
68 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
69 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
70 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
71 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
72 #endif
73 
74 static bool ignore_rlimit_data = true;
75 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
76 
77 static void unmap_region(struct mm_struct *mm,
78 		struct vm_area_struct *vma, struct vm_area_struct *prev,
79 		unsigned long start, unsigned long end);
80 
81 /* description of effects of mapping type and prot in current implementation.
82  * this is due to the limited x86 page protection hardware.  The expected
83  * behavior is in parens:
84  *
85  * map_type	prot
86  *		PROT_NONE	PROT_READ	PROT_WRITE	PROT_EXEC
87  * MAP_SHARED	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
88  *		w: (no) no	w: (no) no	w: (yes) yes	w: (no) no
89  *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
90  *
91  * MAP_PRIVATE	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
92  *		w: (no) no	w: (no) no	w: (copy) copy	w: (no) no
93  *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
94  *
95  */
96 pgprot_t protection_map[16] = {
97 	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
98 	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
99 };
100 
101 pgprot_t vm_get_page_prot(unsigned long vm_flags)
102 {
103 	return __pgprot(pgprot_val(protection_map[vm_flags &
104 				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
105 			pgprot_val(arch_vm_get_page_prot(vm_flags)));
106 }
107 EXPORT_SYMBOL(vm_get_page_prot);
108 
109 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
110 {
111 	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
112 }
113 
114 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
115 void vma_set_page_prot(struct vm_area_struct *vma)
116 {
117 	unsigned long vm_flags = vma->vm_flags;
118 
119 	vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
120 	if (vma_wants_writenotify(vma)) {
121 		vm_flags &= ~VM_SHARED;
122 		vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
123 						     vm_flags);
124 	}
125 }
126 
127 
128 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
129 int sysctl_overcommit_ratio __read_mostly = 50;	/* default is 50% */
130 unsigned long sysctl_overcommit_kbytes __read_mostly;
131 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
132 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
133 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
134 /*
135  * Make sure vm_committed_as in one cacheline and not cacheline shared with
136  * other variables. It can be updated by several CPUs frequently.
137  */
138 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
139 
140 /*
141  * The global memory commitment made in the system can be a metric
142  * that can be used to drive ballooning decisions when Linux is hosted
143  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
144  * balancing memory across competing virtual machines that are hosted.
145  * Several metrics drive this policy engine including the guest reported
146  * memory commitment.
147  */
148 unsigned long vm_memory_committed(void)
149 {
150 	return percpu_counter_read_positive(&vm_committed_as);
151 }
152 EXPORT_SYMBOL_GPL(vm_memory_committed);
153 
154 /*
155  * Check that a process has enough memory to allocate a new virtual
156  * mapping. 0 means there is enough memory for the allocation to
157  * succeed and -ENOMEM implies there is not.
158  *
159  * We currently support three overcommit policies, which are set via the
160  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
161  *
162  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
163  * Additional code 2002 Jul 20 by Robert Love.
164  *
165  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
166  *
167  * Note this is a helper function intended to be used by LSMs which
168  * wish to use this logic.
169  */
170 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
171 {
172 	long free, allowed, reserve;
173 
174 	VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
175 			-(s64)vm_committed_as_batch * num_online_cpus(),
176 			"memory commitment underflow");
177 
178 	vm_acct_memory(pages);
179 
180 	/*
181 	 * Sometimes we want to use more memory than we have
182 	 */
183 	if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
184 		return 0;
185 
186 	if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
187 		free = global_page_state(NR_FREE_PAGES);
188 		free += global_page_state(NR_FILE_PAGES);
189 
190 		/*
191 		 * shmem pages shouldn't be counted as free in this
192 		 * case, they can't be purged, only swapped out, and
193 		 * that won't affect the overall amount of available
194 		 * memory in the system.
195 		 */
196 		free -= global_page_state(NR_SHMEM);
197 
198 		free += get_nr_swap_pages();
199 
200 		/*
201 		 * Any slabs which are created with the
202 		 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
203 		 * which are reclaimable, under pressure.  The dentry
204 		 * cache and most inode caches should fall into this
205 		 */
206 		free += global_page_state(NR_SLAB_RECLAIMABLE);
207 
208 		/*
209 		 * Leave reserved pages. The pages are not for anonymous pages.
210 		 */
211 		if (free <= totalreserve_pages)
212 			goto error;
213 		else
214 			free -= totalreserve_pages;
215 
216 		/*
217 		 * Reserve some for root
218 		 */
219 		if (!cap_sys_admin)
220 			free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
221 
222 		if (free > pages)
223 			return 0;
224 
225 		goto error;
226 	}
227 
228 	allowed = vm_commit_limit();
229 	/*
230 	 * Reserve some for root
231 	 */
232 	if (!cap_sys_admin)
233 		allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
234 
235 	/*
236 	 * Don't let a single process grow so big a user can't recover
237 	 */
238 	if (mm) {
239 		reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
240 		allowed -= min_t(long, mm->total_vm / 32, reserve);
241 	}
242 
243 	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
244 		return 0;
245 error:
246 	vm_unacct_memory(pages);
247 
248 	return -ENOMEM;
249 }
250 
251 /*
252  * Requires inode->i_mapping->i_mmap_rwsem
253  */
254 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
255 		struct file *file, struct address_space *mapping)
256 {
257 	if (vma->vm_flags & VM_DENYWRITE)
258 		atomic_inc(&file_inode(file)->i_writecount);
259 	if (vma->vm_flags & VM_SHARED)
260 		mapping_unmap_writable(mapping);
261 
262 	flush_dcache_mmap_lock(mapping);
263 	vma_interval_tree_remove(vma, &mapping->i_mmap);
264 	flush_dcache_mmap_unlock(mapping);
265 }
266 
267 /*
268  * Unlink a file-based vm structure from its interval tree, to hide
269  * vma from rmap and vmtruncate before freeing its page tables.
270  */
271 void unlink_file_vma(struct vm_area_struct *vma)
272 {
273 	struct file *file = vma->vm_file;
274 
275 	if (file) {
276 		struct address_space *mapping = file->f_mapping;
277 		i_mmap_lock_write(mapping);
278 		__remove_shared_vm_struct(vma, file, mapping);
279 		i_mmap_unlock_write(mapping);
280 	}
281 }
282 
283 /*
284  * Close a vm structure and free it, returning the next.
285  */
286 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
287 {
288 	struct vm_area_struct *next = vma->vm_next;
289 
290 	might_sleep();
291 	if (vma->vm_ops && vma->vm_ops->close)
292 		vma->vm_ops->close(vma);
293 	if (vma->vm_file)
294 		fput(vma->vm_file);
295 	mpol_put(vma_policy(vma));
296 	kmem_cache_free(vm_area_cachep, vma);
297 	return next;
298 }
299 
300 static unsigned long do_brk(unsigned long addr, unsigned long len);
301 
302 SYSCALL_DEFINE1(brk, unsigned long, brk)
303 {
304 	unsigned long retval;
305 	unsigned long newbrk, oldbrk;
306 	struct mm_struct *mm = current->mm;
307 	unsigned long min_brk;
308 	bool populate;
309 
310 	down_write(&mm->mmap_sem);
311 
312 #ifdef CONFIG_COMPAT_BRK
313 	/*
314 	 * CONFIG_COMPAT_BRK can still be overridden by setting
315 	 * randomize_va_space to 2, which will still cause mm->start_brk
316 	 * to be arbitrarily shifted
317 	 */
318 	if (current->brk_randomized)
319 		min_brk = mm->start_brk;
320 	else
321 		min_brk = mm->end_data;
322 #else
323 	min_brk = mm->start_brk;
324 #endif
325 	if (brk < min_brk)
326 		goto out;
327 
328 	/*
329 	 * Check against rlimit here. If this check is done later after the test
330 	 * of oldbrk with newbrk then it can escape the test and let the data
331 	 * segment grow beyond its set limit the in case where the limit is
332 	 * not page aligned -Ram Gupta
333 	 */
334 	if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
335 			      mm->end_data, mm->start_data))
336 		goto out;
337 
338 	newbrk = PAGE_ALIGN(brk);
339 	oldbrk = PAGE_ALIGN(mm->brk);
340 	if (oldbrk == newbrk)
341 		goto set_brk;
342 
343 	/* Always allow shrinking brk. */
344 	if (brk <= mm->brk) {
345 		if (!do_munmap(mm, newbrk, oldbrk-newbrk))
346 			goto set_brk;
347 		goto out;
348 	}
349 
350 	/* Check against existing mmap mappings. */
351 	if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
352 		goto out;
353 
354 	/* Ok, looks good - let it rip. */
355 	if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
356 		goto out;
357 
358 set_brk:
359 	mm->brk = brk;
360 	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
361 	up_write(&mm->mmap_sem);
362 	if (populate)
363 		mm_populate(oldbrk, newbrk - oldbrk);
364 	return brk;
365 
366 out:
367 	retval = mm->brk;
368 	up_write(&mm->mmap_sem);
369 	return retval;
370 }
371 
372 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
373 {
374 	unsigned long max, subtree_gap;
375 	max = vma->vm_start;
376 	if (vma->vm_prev)
377 		max -= vma->vm_prev->vm_end;
378 	if (vma->vm_rb.rb_left) {
379 		subtree_gap = rb_entry(vma->vm_rb.rb_left,
380 				struct vm_area_struct, vm_rb)->rb_subtree_gap;
381 		if (subtree_gap > max)
382 			max = subtree_gap;
383 	}
384 	if (vma->vm_rb.rb_right) {
385 		subtree_gap = rb_entry(vma->vm_rb.rb_right,
386 				struct vm_area_struct, vm_rb)->rb_subtree_gap;
387 		if (subtree_gap > max)
388 			max = subtree_gap;
389 	}
390 	return max;
391 }
392 
393 #ifdef CONFIG_DEBUG_VM_RB
394 static int browse_rb(struct mm_struct *mm)
395 {
396 	struct rb_root *root = &mm->mm_rb;
397 	int i = 0, j, bug = 0;
398 	struct rb_node *nd, *pn = NULL;
399 	unsigned long prev = 0, pend = 0;
400 
401 	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
402 		struct vm_area_struct *vma;
403 		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
404 		if (vma->vm_start < prev) {
405 			pr_emerg("vm_start %lx < prev %lx\n",
406 				  vma->vm_start, prev);
407 			bug = 1;
408 		}
409 		if (vma->vm_start < pend) {
410 			pr_emerg("vm_start %lx < pend %lx\n",
411 				  vma->vm_start, pend);
412 			bug = 1;
413 		}
414 		if (vma->vm_start > vma->vm_end) {
415 			pr_emerg("vm_start %lx > vm_end %lx\n",
416 				  vma->vm_start, vma->vm_end);
417 			bug = 1;
418 		}
419 		spin_lock(&mm->page_table_lock);
420 		if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
421 			pr_emerg("free gap %lx, correct %lx\n",
422 			       vma->rb_subtree_gap,
423 			       vma_compute_subtree_gap(vma));
424 			bug = 1;
425 		}
426 		spin_unlock(&mm->page_table_lock);
427 		i++;
428 		pn = nd;
429 		prev = vma->vm_start;
430 		pend = vma->vm_end;
431 	}
432 	j = 0;
433 	for (nd = pn; nd; nd = rb_prev(nd))
434 		j++;
435 	if (i != j) {
436 		pr_emerg("backwards %d, forwards %d\n", j, i);
437 		bug = 1;
438 	}
439 	return bug ? -1 : i;
440 }
441 
442 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
443 {
444 	struct rb_node *nd;
445 
446 	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
447 		struct vm_area_struct *vma;
448 		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
449 		VM_BUG_ON_VMA(vma != ignore &&
450 			vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
451 			vma);
452 	}
453 }
454 
455 static void validate_mm(struct mm_struct *mm)
456 {
457 	int bug = 0;
458 	int i = 0;
459 	unsigned long highest_address = 0;
460 	struct vm_area_struct *vma = mm->mmap;
461 
462 	while (vma) {
463 		struct anon_vma *anon_vma = vma->anon_vma;
464 		struct anon_vma_chain *avc;
465 
466 		if (anon_vma) {
467 			anon_vma_lock_read(anon_vma);
468 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
469 				anon_vma_interval_tree_verify(avc);
470 			anon_vma_unlock_read(anon_vma);
471 		}
472 
473 		highest_address = vma->vm_end;
474 		vma = vma->vm_next;
475 		i++;
476 	}
477 	if (i != mm->map_count) {
478 		pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
479 		bug = 1;
480 	}
481 	if (highest_address != mm->highest_vm_end) {
482 		pr_emerg("mm->highest_vm_end %lx, found %lx\n",
483 			  mm->highest_vm_end, highest_address);
484 		bug = 1;
485 	}
486 	i = browse_rb(mm);
487 	if (i != mm->map_count) {
488 		if (i != -1)
489 			pr_emerg("map_count %d rb %d\n", mm->map_count, i);
490 		bug = 1;
491 	}
492 	VM_BUG_ON_MM(bug, mm);
493 }
494 #else
495 #define validate_mm_rb(root, ignore) do { } while (0)
496 #define validate_mm(mm) do { } while (0)
497 #endif
498 
499 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
500 		     unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
501 
502 /*
503  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
504  * vma->vm_prev->vm_end values changed, without modifying the vma's position
505  * in the rbtree.
506  */
507 static void vma_gap_update(struct vm_area_struct *vma)
508 {
509 	/*
510 	 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
511 	 * function that does exacltly what we want.
512 	 */
513 	vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
514 }
515 
516 static inline void vma_rb_insert(struct vm_area_struct *vma,
517 				 struct rb_root *root)
518 {
519 	/* All rb_subtree_gap values must be consistent prior to insertion */
520 	validate_mm_rb(root, NULL);
521 
522 	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
523 }
524 
525 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
526 {
527 	/*
528 	 * All rb_subtree_gap values must be consistent prior to erase,
529 	 * with the possible exception of the vma being erased.
530 	 */
531 	validate_mm_rb(root, vma);
532 
533 	/*
534 	 * Note rb_erase_augmented is a fairly large inline function,
535 	 * so make sure we instantiate it only once with our desired
536 	 * augmented rbtree callbacks.
537 	 */
538 	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
539 }
540 
541 /*
542  * vma has some anon_vma assigned, and is already inserted on that
543  * anon_vma's interval trees.
544  *
545  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
546  * vma must be removed from the anon_vma's interval trees using
547  * anon_vma_interval_tree_pre_update_vma().
548  *
549  * After the update, the vma will be reinserted using
550  * anon_vma_interval_tree_post_update_vma().
551  *
552  * The entire update must be protected by exclusive mmap_sem and by
553  * the root anon_vma's mutex.
554  */
555 static inline void
556 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
557 {
558 	struct anon_vma_chain *avc;
559 
560 	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
561 		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
562 }
563 
564 static inline void
565 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
566 {
567 	struct anon_vma_chain *avc;
568 
569 	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
570 		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
571 }
572 
573 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
574 		unsigned long end, struct vm_area_struct **pprev,
575 		struct rb_node ***rb_link, struct rb_node **rb_parent)
576 {
577 	struct rb_node **__rb_link, *__rb_parent, *rb_prev;
578 
579 	__rb_link = &mm->mm_rb.rb_node;
580 	rb_prev = __rb_parent = NULL;
581 
582 	while (*__rb_link) {
583 		struct vm_area_struct *vma_tmp;
584 
585 		__rb_parent = *__rb_link;
586 		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
587 
588 		if (vma_tmp->vm_end > addr) {
589 			/* Fail if an existing vma overlaps the area */
590 			if (vma_tmp->vm_start < end)
591 				return -ENOMEM;
592 			__rb_link = &__rb_parent->rb_left;
593 		} else {
594 			rb_prev = __rb_parent;
595 			__rb_link = &__rb_parent->rb_right;
596 		}
597 	}
598 
599 	*pprev = NULL;
600 	if (rb_prev)
601 		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
602 	*rb_link = __rb_link;
603 	*rb_parent = __rb_parent;
604 	return 0;
605 }
606 
607 static unsigned long count_vma_pages_range(struct mm_struct *mm,
608 		unsigned long addr, unsigned long end)
609 {
610 	unsigned long nr_pages = 0;
611 	struct vm_area_struct *vma;
612 
613 	/* Find first overlaping mapping */
614 	vma = find_vma_intersection(mm, addr, end);
615 	if (!vma)
616 		return 0;
617 
618 	nr_pages = (min(end, vma->vm_end) -
619 		max(addr, vma->vm_start)) >> PAGE_SHIFT;
620 
621 	/* Iterate over the rest of the overlaps */
622 	for (vma = vma->vm_next; vma; vma = vma->vm_next) {
623 		unsigned long overlap_len;
624 
625 		if (vma->vm_start > end)
626 			break;
627 
628 		overlap_len = min(end, vma->vm_end) - vma->vm_start;
629 		nr_pages += overlap_len >> PAGE_SHIFT;
630 	}
631 
632 	return nr_pages;
633 }
634 
635 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
636 		struct rb_node **rb_link, struct rb_node *rb_parent)
637 {
638 	/* Update tracking information for the gap following the new vma. */
639 	if (vma->vm_next)
640 		vma_gap_update(vma->vm_next);
641 	else
642 		mm->highest_vm_end = vma->vm_end;
643 
644 	/*
645 	 * vma->vm_prev wasn't known when we followed the rbtree to find the
646 	 * correct insertion point for that vma. As a result, we could not
647 	 * update the vma vm_rb parents rb_subtree_gap values on the way down.
648 	 * So, we first insert the vma with a zero rb_subtree_gap value
649 	 * (to be consistent with what we did on the way down), and then
650 	 * immediately update the gap to the correct value. Finally we
651 	 * rebalance the rbtree after all augmented values have been set.
652 	 */
653 	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
654 	vma->rb_subtree_gap = 0;
655 	vma_gap_update(vma);
656 	vma_rb_insert(vma, &mm->mm_rb);
657 }
658 
659 static void __vma_link_file(struct vm_area_struct *vma)
660 {
661 	struct file *file;
662 
663 	file = vma->vm_file;
664 	if (file) {
665 		struct address_space *mapping = file->f_mapping;
666 
667 		if (vma->vm_flags & VM_DENYWRITE)
668 			atomic_dec(&file_inode(file)->i_writecount);
669 		if (vma->vm_flags & VM_SHARED)
670 			atomic_inc(&mapping->i_mmap_writable);
671 
672 		flush_dcache_mmap_lock(mapping);
673 		vma_interval_tree_insert(vma, &mapping->i_mmap);
674 		flush_dcache_mmap_unlock(mapping);
675 	}
676 }
677 
678 static void
679 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
680 	struct vm_area_struct *prev, struct rb_node **rb_link,
681 	struct rb_node *rb_parent)
682 {
683 	__vma_link_list(mm, vma, prev, rb_parent);
684 	__vma_link_rb(mm, vma, rb_link, rb_parent);
685 }
686 
687 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
688 			struct vm_area_struct *prev, struct rb_node **rb_link,
689 			struct rb_node *rb_parent)
690 {
691 	struct address_space *mapping = NULL;
692 
693 	if (vma->vm_file) {
694 		mapping = vma->vm_file->f_mapping;
695 		i_mmap_lock_write(mapping);
696 	}
697 
698 	__vma_link(mm, vma, prev, rb_link, rb_parent);
699 	__vma_link_file(vma);
700 
701 	if (mapping)
702 		i_mmap_unlock_write(mapping);
703 
704 	mm->map_count++;
705 	validate_mm(mm);
706 }
707 
708 /*
709  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
710  * mm's list and rbtree.  It has already been inserted into the interval tree.
711  */
712 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
713 {
714 	struct vm_area_struct *prev;
715 	struct rb_node **rb_link, *rb_parent;
716 
717 	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
718 			   &prev, &rb_link, &rb_parent))
719 		BUG();
720 	__vma_link(mm, vma, prev, rb_link, rb_parent);
721 	mm->map_count++;
722 }
723 
724 static inline void
725 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
726 		struct vm_area_struct *prev)
727 {
728 	struct vm_area_struct *next;
729 
730 	vma_rb_erase(vma, &mm->mm_rb);
731 	prev->vm_next = next = vma->vm_next;
732 	if (next)
733 		next->vm_prev = prev;
734 
735 	/* Kill the cache */
736 	vmacache_invalidate(mm);
737 }
738 
739 /*
740  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
741  * is already present in an i_mmap tree without adjusting the tree.
742  * The following helper function should be used when such adjustments
743  * are necessary.  The "insert" vma (if any) is to be inserted
744  * before we drop the necessary locks.
745  */
746 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
747 	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
748 {
749 	struct mm_struct *mm = vma->vm_mm;
750 	struct vm_area_struct *next = vma->vm_next;
751 	struct vm_area_struct *importer = NULL;
752 	struct address_space *mapping = NULL;
753 	struct rb_root *root = NULL;
754 	struct anon_vma *anon_vma = NULL;
755 	struct file *file = vma->vm_file;
756 	bool start_changed = false, end_changed = false;
757 	long adjust_next = 0;
758 	int remove_next = 0;
759 
760 	if (next && !insert) {
761 		struct vm_area_struct *exporter = NULL;
762 
763 		if (end >= next->vm_end) {
764 			/*
765 			 * vma expands, overlapping all the next, and
766 			 * perhaps the one after too (mprotect case 6).
767 			 */
768 again:			remove_next = 1 + (end > next->vm_end);
769 			end = next->vm_end;
770 			exporter = next;
771 			importer = vma;
772 		} else if (end > next->vm_start) {
773 			/*
774 			 * vma expands, overlapping part of the next:
775 			 * mprotect case 5 shifting the boundary up.
776 			 */
777 			adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
778 			exporter = next;
779 			importer = vma;
780 		} else if (end < vma->vm_end) {
781 			/*
782 			 * vma shrinks, and !insert tells it's not
783 			 * split_vma inserting another: so it must be
784 			 * mprotect case 4 shifting the boundary down.
785 			 */
786 			adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
787 			exporter = vma;
788 			importer = next;
789 		}
790 
791 		/*
792 		 * Easily overlooked: when mprotect shifts the boundary,
793 		 * make sure the expanding vma has anon_vma set if the
794 		 * shrinking vma had, to cover any anon pages imported.
795 		 */
796 		if (exporter && exporter->anon_vma && !importer->anon_vma) {
797 			int error;
798 
799 			importer->anon_vma = exporter->anon_vma;
800 			error = anon_vma_clone(importer, exporter);
801 			if (error)
802 				return error;
803 		}
804 	}
805 
806 	if (file) {
807 		mapping = file->f_mapping;
808 		root = &mapping->i_mmap;
809 		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
810 
811 		if (adjust_next)
812 			uprobe_munmap(next, next->vm_start, next->vm_end);
813 
814 		i_mmap_lock_write(mapping);
815 		if (insert) {
816 			/*
817 			 * Put into interval tree now, so instantiated pages
818 			 * are visible to arm/parisc __flush_dcache_page
819 			 * throughout; but we cannot insert into address
820 			 * space until vma start or end is updated.
821 			 */
822 			__vma_link_file(insert);
823 		}
824 	}
825 
826 	vma_adjust_trans_huge(vma, start, end, adjust_next);
827 
828 	anon_vma = vma->anon_vma;
829 	if (!anon_vma && adjust_next)
830 		anon_vma = next->anon_vma;
831 	if (anon_vma) {
832 		VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
833 			  anon_vma != next->anon_vma, next);
834 		anon_vma_lock_write(anon_vma);
835 		anon_vma_interval_tree_pre_update_vma(vma);
836 		if (adjust_next)
837 			anon_vma_interval_tree_pre_update_vma(next);
838 	}
839 
840 	if (root) {
841 		flush_dcache_mmap_lock(mapping);
842 		vma_interval_tree_remove(vma, root);
843 		if (adjust_next)
844 			vma_interval_tree_remove(next, root);
845 	}
846 
847 	if (start != vma->vm_start) {
848 		vma->vm_start = start;
849 		start_changed = true;
850 	}
851 	if (end != vma->vm_end) {
852 		vma->vm_end = end;
853 		end_changed = true;
854 	}
855 	vma->vm_pgoff = pgoff;
856 	if (adjust_next) {
857 		next->vm_start += adjust_next << PAGE_SHIFT;
858 		next->vm_pgoff += adjust_next;
859 	}
860 
861 	if (root) {
862 		if (adjust_next)
863 			vma_interval_tree_insert(next, root);
864 		vma_interval_tree_insert(vma, root);
865 		flush_dcache_mmap_unlock(mapping);
866 	}
867 
868 	if (remove_next) {
869 		/*
870 		 * vma_merge has merged next into vma, and needs
871 		 * us to remove next before dropping the locks.
872 		 */
873 		__vma_unlink(mm, next, vma);
874 		if (file)
875 			__remove_shared_vm_struct(next, file, mapping);
876 	} else if (insert) {
877 		/*
878 		 * split_vma has split insert from vma, and needs
879 		 * us to insert it before dropping the locks
880 		 * (it may either follow vma or precede it).
881 		 */
882 		__insert_vm_struct(mm, insert);
883 	} else {
884 		if (start_changed)
885 			vma_gap_update(vma);
886 		if (end_changed) {
887 			if (!next)
888 				mm->highest_vm_end = end;
889 			else if (!adjust_next)
890 				vma_gap_update(next);
891 		}
892 	}
893 
894 	if (anon_vma) {
895 		anon_vma_interval_tree_post_update_vma(vma);
896 		if (adjust_next)
897 			anon_vma_interval_tree_post_update_vma(next);
898 		anon_vma_unlock_write(anon_vma);
899 	}
900 	if (mapping)
901 		i_mmap_unlock_write(mapping);
902 
903 	if (root) {
904 		uprobe_mmap(vma);
905 
906 		if (adjust_next)
907 			uprobe_mmap(next);
908 	}
909 
910 	if (remove_next) {
911 		if (file) {
912 			uprobe_munmap(next, next->vm_start, next->vm_end);
913 			fput(file);
914 		}
915 		if (next->anon_vma)
916 			anon_vma_merge(vma, next);
917 		mm->map_count--;
918 		mpol_put(vma_policy(next));
919 		kmem_cache_free(vm_area_cachep, next);
920 		/*
921 		 * In mprotect's case 6 (see comments on vma_merge),
922 		 * we must remove another next too. It would clutter
923 		 * up the code too much to do both in one go.
924 		 */
925 		next = vma->vm_next;
926 		if (remove_next == 2)
927 			goto again;
928 		else if (next)
929 			vma_gap_update(next);
930 		else
931 			mm->highest_vm_end = end;
932 	}
933 	if (insert && file)
934 		uprobe_mmap(insert);
935 
936 	validate_mm(mm);
937 
938 	return 0;
939 }
940 
941 /*
942  * If the vma has a ->close operation then the driver probably needs to release
943  * per-vma resources, so we don't attempt to merge those.
944  */
945 static inline int is_mergeable_vma(struct vm_area_struct *vma,
946 				struct file *file, unsigned long vm_flags,
947 				struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
948 {
949 	/*
950 	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
951 	 * match the flags but dirty bit -- the caller should mark
952 	 * merged VMA as dirty. If dirty bit won't be excluded from
953 	 * comparison, we increase pressue on the memory system forcing
954 	 * the kernel to generate new VMAs when old one could be
955 	 * extended instead.
956 	 */
957 	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
958 		return 0;
959 	if (vma->vm_file != file)
960 		return 0;
961 	if (vma->vm_ops && vma->vm_ops->close)
962 		return 0;
963 	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
964 		return 0;
965 	return 1;
966 }
967 
968 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
969 					struct anon_vma *anon_vma2,
970 					struct vm_area_struct *vma)
971 {
972 	/*
973 	 * The list_is_singular() test is to avoid merging VMA cloned from
974 	 * parents. This can improve scalability caused by anon_vma lock.
975 	 */
976 	if ((!anon_vma1 || !anon_vma2) && (!vma ||
977 		list_is_singular(&vma->anon_vma_chain)))
978 		return 1;
979 	return anon_vma1 == anon_vma2;
980 }
981 
982 /*
983  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
984  * in front of (at a lower virtual address and file offset than) the vma.
985  *
986  * We cannot merge two vmas if they have differently assigned (non-NULL)
987  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
988  *
989  * We don't check here for the merged mmap wrapping around the end of pagecache
990  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
991  * wrap, nor mmaps which cover the final page at index -1UL.
992  */
993 static int
994 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
995 		     struct anon_vma *anon_vma, struct file *file,
996 		     pgoff_t vm_pgoff,
997 		     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
998 {
999 	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1000 	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1001 		if (vma->vm_pgoff == vm_pgoff)
1002 			return 1;
1003 	}
1004 	return 0;
1005 }
1006 
1007 /*
1008  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1009  * beyond (at a higher virtual address and file offset than) the vma.
1010  *
1011  * We cannot merge two vmas if they have differently assigned (non-NULL)
1012  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1013  */
1014 static int
1015 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1016 		    struct anon_vma *anon_vma, struct file *file,
1017 		    pgoff_t vm_pgoff,
1018 		    struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1019 {
1020 	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1021 	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1022 		pgoff_t vm_pglen;
1023 		vm_pglen = vma_pages(vma);
1024 		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1025 			return 1;
1026 	}
1027 	return 0;
1028 }
1029 
1030 /*
1031  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1032  * whether that can be merged with its predecessor or its successor.
1033  * Or both (it neatly fills a hole).
1034  *
1035  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1036  * certain not to be mapped by the time vma_merge is called; but when
1037  * called for mprotect, it is certain to be already mapped (either at
1038  * an offset within prev, or at the start of next), and the flags of
1039  * this area are about to be changed to vm_flags - and the no-change
1040  * case has already been eliminated.
1041  *
1042  * The following mprotect cases have to be considered, where AAAA is
1043  * the area passed down from mprotect_fixup, never extending beyond one
1044  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1045  *
1046  *     AAAA             AAAA                AAAA          AAAA
1047  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1048  *    cannot merge    might become    might become    might become
1049  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1050  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1051  *    mremap move:                                    PPPPNNNNNNNN 8
1052  *        AAAA
1053  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1054  *    might become    case 1 below    case 2 below    case 3 below
1055  *
1056  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1057  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1058  */
1059 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1060 			struct vm_area_struct *prev, unsigned long addr,
1061 			unsigned long end, unsigned long vm_flags,
1062 			struct anon_vma *anon_vma, struct file *file,
1063 			pgoff_t pgoff, struct mempolicy *policy,
1064 			struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1065 {
1066 	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1067 	struct vm_area_struct *area, *next;
1068 	int err;
1069 
1070 	/*
1071 	 * We later require that vma->vm_flags == vm_flags,
1072 	 * so this tests vma->vm_flags & VM_SPECIAL, too.
1073 	 */
1074 	if (vm_flags & VM_SPECIAL)
1075 		return NULL;
1076 
1077 	if (prev)
1078 		next = prev->vm_next;
1079 	else
1080 		next = mm->mmap;
1081 	area = next;
1082 	if (next && next->vm_end == end)		/* cases 6, 7, 8 */
1083 		next = next->vm_next;
1084 
1085 	/*
1086 	 * Can it merge with the predecessor?
1087 	 */
1088 	if (prev && prev->vm_end == addr &&
1089 			mpol_equal(vma_policy(prev), policy) &&
1090 			can_vma_merge_after(prev, vm_flags,
1091 					    anon_vma, file, pgoff,
1092 					    vm_userfaultfd_ctx)) {
1093 		/*
1094 		 * OK, it can.  Can we now merge in the successor as well?
1095 		 */
1096 		if (next && end == next->vm_start &&
1097 				mpol_equal(policy, vma_policy(next)) &&
1098 				can_vma_merge_before(next, vm_flags,
1099 						     anon_vma, file,
1100 						     pgoff+pglen,
1101 						     vm_userfaultfd_ctx) &&
1102 				is_mergeable_anon_vma(prev->anon_vma,
1103 						      next->anon_vma, NULL)) {
1104 							/* cases 1, 6 */
1105 			err = vma_adjust(prev, prev->vm_start,
1106 				next->vm_end, prev->vm_pgoff, NULL);
1107 		} else					/* cases 2, 5, 7 */
1108 			err = vma_adjust(prev, prev->vm_start,
1109 				end, prev->vm_pgoff, NULL);
1110 		if (err)
1111 			return NULL;
1112 		khugepaged_enter_vma_merge(prev, vm_flags);
1113 		return prev;
1114 	}
1115 
1116 	/*
1117 	 * Can this new request be merged in front of next?
1118 	 */
1119 	if (next && end == next->vm_start &&
1120 			mpol_equal(policy, vma_policy(next)) &&
1121 			can_vma_merge_before(next, vm_flags,
1122 					     anon_vma, file, pgoff+pglen,
1123 					     vm_userfaultfd_ctx)) {
1124 		if (prev && addr < prev->vm_end)	/* case 4 */
1125 			err = vma_adjust(prev, prev->vm_start,
1126 				addr, prev->vm_pgoff, NULL);
1127 		else					/* cases 3, 8 */
1128 			err = vma_adjust(area, addr, next->vm_end,
1129 				next->vm_pgoff - pglen, NULL);
1130 		if (err)
1131 			return NULL;
1132 		khugepaged_enter_vma_merge(area, vm_flags);
1133 		return area;
1134 	}
1135 
1136 	return NULL;
1137 }
1138 
1139 /*
1140  * Rough compatbility check to quickly see if it's even worth looking
1141  * at sharing an anon_vma.
1142  *
1143  * They need to have the same vm_file, and the flags can only differ
1144  * in things that mprotect may change.
1145  *
1146  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1147  * we can merge the two vma's. For example, we refuse to merge a vma if
1148  * there is a vm_ops->close() function, because that indicates that the
1149  * driver is doing some kind of reference counting. But that doesn't
1150  * really matter for the anon_vma sharing case.
1151  */
1152 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1153 {
1154 	return a->vm_end == b->vm_start &&
1155 		mpol_equal(vma_policy(a), vma_policy(b)) &&
1156 		a->vm_file == b->vm_file &&
1157 		!((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1158 		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1159 }
1160 
1161 /*
1162  * Do some basic sanity checking to see if we can re-use the anon_vma
1163  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1164  * the same as 'old', the other will be the new one that is trying
1165  * to share the anon_vma.
1166  *
1167  * NOTE! This runs with mm_sem held for reading, so it is possible that
1168  * the anon_vma of 'old' is concurrently in the process of being set up
1169  * by another page fault trying to merge _that_. But that's ok: if it
1170  * is being set up, that automatically means that it will be a singleton
1171  * acceptable for merging, so we can do all of this optimistically. But
1172  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1173  *
1174  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1175  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1176  * is to return an anon_vma that is "complex" due to having gone through
1177  * a fork).
1178  *
1179  * We also make sure that the two vma's are compatible (adjacent,
1180  * and with the same memory policies). That's all stable, even with just
1181  * a read lock on the mm_sem.
1182  */
1183 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1184 {
1185 	if (anon_vma_compatible(a, b)) {
1186 		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1187 
1188 		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1189 			return anon_vma;
1190 	}
1191 	return NULL;
1192 }
1193 
1194 /*
1195  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1196  * neighbouring vmas for a suitable anon_vma, before it goes off
1197  * to allocate a new anon_vma.  It checks because a repetitive
1198  * sequence of mprotects and faults may otherwise lead to distinct
1199  * anon_vmas being allocated, preventing vma merge in subsequent
1200  * mprotect.
1201  */
1202 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1203 {
1204 	struct anon_vma *anon_vma;
1205 	struct vm_area_struct *near;
1206 
1207 	near = vma->vm_next;
1208 	if (!near)
1209 		goto try_prev;
1210 
1211 	anon_vma = reusable_anon_vma(near, vma, near);
1212 	if (anon_vma)
1213 		return anon_vma;
1214 try_prev:
1215 	near = vma->vm_prev;
1216 	if (!near)
1217 		goto none;
1218 
1219 	anon_vma = reusable_anon_vma(near, near, vma);
1220 	if (anon_vma)
1221 		return anon_vma;
1222 none:
1223 	/*
1224 	 * There's no absolute need to look only at touching neighbours:
1225 	 * we could search further afield for "compatible" anon_vmas.
1226 	 * But it would probably just be a waste of time searching,
1227 	 * or lead to too many vmas hanging off the same anon_vma.
1228 	 * We're trying to allow mprotect remerging later on,
1229 	 * not trying to minimize memory used for anon_vmas.
1230 	 */
1231 	return NULL;
1232 }
1233 
1234 /*
1235  * If a hint addr is less than mmap_min_addr change hint to be as
1236  * low as possible but still greater than mmap_min_addr
1237  */
1238 static inline unsigned long round_hint_to_min(unsigned long hint)
1239 {
1240 	hint &= PAGE_MASK;
1241 	if (((void *)hint != NULL) &&
1242 	    (hint < mmap_min_addr))
1243 		return PAGE_ALIGN(mmap_min_addr);
1244 	return hint;
1245 }
1246 
1247 static inline int mlock_future_check(struct mm_struct *mm,
1248 				     unsigned long flags,
1249 				     unsigned long len)
1250 {
1251 	unsigned long locked, lock_limit;
1252 
1253 	/*  mlock MCL_FUTURE? */
1254 	if (flags & VM_LOCKED) {
1255 		locked = len >> PAGE_SHIFT;
1256 		locked += mm->locked_vm;
1257 		lock_limit = rlimit(RLIMIT_MEMLOCK);
1258 		lock_limit >>= PAGE_SHIFT;
1259 		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1260 			return -EAGAIN;
1261 	}
1262 	return 0;
1263 }
1264 
1265 /*
1266  * The caller must hold down_write(&current->mm->mmap_sem).
1267  */
1268 unsigned long do_mmap(struct file *file, unsigned long addr,
1269 			unsigned long len, unsigned long prot,
1270 			unsigned long flags, vm_flags_t vm_flags,
1271 			unsigned long pgoff, unsigned long *populate)
1272 {
1273 	struct mm_struct *mm = current->mm;
1274 	int pkey = 0;
1275 
1276 	*populate = 0;
1277 
1278 	if (!len)
1279 		return -EINVAL;
1280 
1281 	/*
1282 	 * Does the application expect PROT_READ to imply PROT_EXEC?
1283 	 *
1284 	 * (the exception is when the underlying filesystem is noexec
1285 	 *  mounted, in which case we dont add PROT_EXEC.)
1286 	 */
1287 	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1288 		if (!(file && path_noexec(&file->f_path)))
1289 			prot |= PROT_EXEC;
1290 
1291 	if (!(flags & MAP_FIXED))
1292 		addr = round_hint_to_min(addr);
1293 
1294 	/* Careful about overflows.. */
1295 	len = PAGE_ALIGN(len);
1296 	if (!len)
1297 		return -ENOMEM;
1298 
1299 	/* offset overflow? */
1300 	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1301 		return -EOVERFLOW;
1302 
1303 	/* Too many mappings? */
1304 	if (mm->map_count > sysctl_max_map_count)
1305 		return -ENOMEM;
1306 
1307 	/* Obtain the address to map to. we verify (or select) it and ensure
1308 	 * that it represents a valid section of the address space.
1309 	 */
1310 	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1311 	if (offset_in_page(addr))
1312 		return addr;
1313 
1314 	if (prot == PROT_EXEC) {
1315 		pkey = execute_only_pkey(mm);
1316 		if (pkey < 0)
1317 			pkey = 0;
1318 	}
1319 
1320 	/* Do simple checking here so the lower-level routines won't have
1321 	 * to. we assume access permissions have been handled by the open
1322 	 * of the memory object, so we don't do any here.
1323 	 */
1324 	vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1325 			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1326 
1327 	if (flags & MAP_LOCKED)
1328 		if (!can_do_mlock())
1329 			return -EPERM;
1330 
1331 	if (mlock_future_check(mm, vm_flags, len))
1332 		return -EAGAIN;
1333 
1334 	if (file) {
1335 		struct inode *inode = file_inode(file);
1336 
1337 		switch (flags & MAP_TYPE) {
1338 		case MAP_SHARED:
1339 			if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1340 				return -EACCES;
1341 
1342 			/*
1343 			 * Make sure we don't allow writing to an append-only
1344 			 * file..
1345 			 */
1346 			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1347 				return -EACCES;
1348 
1349 			/*
1350 			 * Make sure there are no mandatory locks on the file.
1351 			 */
1352 			if (locks_verify_locked(file))
1353 				return -EAGAIN;
1354 
1355 			vm_flags |= VM_SHARED | VM_MAYSHARE;
1356 			if (!(file->f_mode & FMODE_WRITE))
1357 				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1358 
1359 			/* fall through */
1360 		case MAP_PRIVATE:
1361 			if (!(file->f_mode & FMODE_READ))
1362 				return -EACCES;
1363 			if (path_noexec(&file->f_path)) {
1364 				if (vm_flags & VM_EXEC)
1365 					return -EPERM;
1366 				vm_flags &= ~VM_MAYEXEC;
1367 			}
1368 
1369 			if (!file->f_op->mmap)
1370 				return -ENODEV;
1371 			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1372 				return -EINVAL;
1373 			break;
1374 
1375 		default:
1376 			return -EINVAL;
1377 		}
1378 	} else {
1379 		switch (flags & MAP_TYPE) {
1380 		case MAP_SHARED:
1381 			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1382 				return -EINVAL;
1383 			/*
1384 			 * Ignore pgoff.
1385 			 */
1386 			pgoff = 0;
1387 			vm_flags |= VM_SHARED | VM_MAYSHARE;
1388 			break;
1389 		case MAP_PRIVATE:
1390 			/*
1391 			 * Set pgoff according to addr for anon_vma.
1392 			 */
1393 			pgoff = addr >> PAGE_SHIFT;
1394 			break;
1395 		default:
1396 			return -EINVAL;
1397 		}
1398 	}
1399 
1400 	/*
1401 	 * Set 'VM_NORESERVE' if we should not account for the
1402 	 * memory use of this mapping.
1403 	 */
1404 	if (flags & MAP_NORESERVE) {
1405 		/* We honor MAP_NORESERVE if allowed to overcommit */
1406 		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1407 			vm_flags |= VM_NORESERVE;
1408 
1409 		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1410 		if (file && is_file_hugepages(file))
1411 			vm_flags |= VM_NORESERVE;
1412 	}
1413 
1414 	addr = mmap_region(file, addr, len, vm_flags, pgoff);
1415 	if (!IS_ERR_VALUE(addr) &&
1416 	    ((vm_flags & VM_LOCKED) ||
1417 	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1418 		*populate = len;
1419 	return addr;
1420 }
1421 
1422 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1423 		unsigned long, prot, unsigned long, flags,
1424 		unsigned long, fd, unsigned long, pgoff)
1425 {
1426 	struct file *file = NULL;
1427 	unsigned long retval;
1428 
1429 	if (!(flags & MAP_ANONYMOUS)) {
1430 		audit_mmap_fd(fd, flags);
1431 		file = fget(fd);
1432 		if (!file)
1433 			return -EBADF;
1434 		if (is_file_hugepages(file))
1435 			len = ALIGN(len, huge_page_size(hstate_file(file)));
1436 		retval = -EINVAL;
1437 		if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1438 			goto out_fput;
1439 	} else if (flags & MAP_HUGETLB) {
1440 		struct user_struct *user = NULL;
1441 		struct hstate *hs;
1442 
1443 		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1444 		if (!hs)
1445 			return -EINVAL;
1446 
1447 		len = ALIGN(len, huge_page_size(hs));
1448 		/*
1449 		 * VM_NORESERVE is used because the reservations will be
1450 		 * taken when vm_ops->mmap() is called
1451 		 * A dummy user value is used because we are not locking
1452 		 * memory so no accounting is necessary
1453 		 */
1454 		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1455 				VM_NORESERVE,
1456 				&user, HUGETLB_ANONHUGE_INODE,
1457 				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1458 		if (IS_ERR(file))
1459 			return PTR_ERR(file);
1460 	}
1461 
1462 	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1463 
1464 	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1465 out_fput:
1466 	if (file)
1467 		fput(file);
1468 	return retval;
1469 }
1470 
1471 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1472 struct mmap_arg_struct {
1473 	unsigned long addr;
1474 	unsigned long len;
1475 	unsigned long prot;
1476 	unsigned long flags;
1477 	unsigned long fd;
1478 	unsigned long offset;
1479 };
1480 
1481 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1482 {
1483 	struct mmap_arg_struct a;
1484 
1485 	if (copy_from_user(&a, arg, sizeof(a)))
1486 		return -EFAULT;
1487 	if (offset_in_page(a.offset))
1488 		return -EINVAL;
1489 
1490 	return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1491 			      a.offset >> PAGE_SHIFT);
1492 }
1493 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1494 
1495 /*
1496  * Some shared mappigns will want the pages marked read-only
1497  * to track write events. If so, we'll downgrade vm_page_prot
1498  * to the private version (using protection_map[] without the
1499  * VM_SHARED bit).
1500  */
1501 int vma_wants_writenotify(struct vm_area_struct *vma)
1502 {
1503 	vm_flags_t vm_flags = vma->vm_flags;
1504 	const struct vm_operations_struct *vm_ops = vma->vm_ops;
1505 
1506 	/* If it was private or non-writable, the write bit is already clear */
1507 	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1508 		return 0;
1509 
1510 	/* The backer wishes to know when pages are first written to? */
1511 	if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1512 		return 1;
1513 
1514 	/* The open routine did something to the protections that pgprot_modify
1515 	 * won't preserve? */
1516 	if (pgprot_val(vma->vm_page_prot) !=
1517 	    pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1518 		return 0;
1519 
1520 	/* Do we need to track softdirty? */
1521 	if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1522 		return 1;
1523 
1524 	/* Specialty mapping? */
1525 	if (vm_flags & VM_PFNMAP)
1526 		return 0;
1527 
1528 	/* Can the mapping track the dirty pages? */
1529 	return vma->vm_file && vma->vm_file->f_mapping &&
1530 		mapping_cap_account_dirty(vma->vm_file->f_mapping);
1531 }
1532 
1533 /*
1534  * We account for memory if it's a private writeable mapping,
1535  * not hugepages and VM_NORESERVE wasn't set.
1536  */
1537 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1538 {
1539 	/*
1540 	 * hugetlb has its own accounting separate from the core VM
1541 	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1542 	 */
1543 	if (file && is_file_hugepages(file))
1544 		return 0;
1545 
1546 	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1547 }
1548 
1549 unsigned long mmap_region(struct file *file, unsigned long addr,
1550 		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1551 {
1552 	struct mm_struct *mm = current->mm;
1553 	struct vm_area_struct *vma, *prev;
1554 	int error;
1555 	struct rb_node **rb_link, *rb_parent;
1556 	unsigned long charged = 0;
1557 
1558 	/* Check against address space limit. */
1559 	if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1560 		unsigned long nr_pages;
1561 
1562 		/*
1563 		 * MAP_FIXED may remove pages of mappings that intersects with
1564 		 * requested mapping. Account for the pages it would unmap.
1565 		 */
1566 		nr_pages = count_vma_pages_range(mm, addr, addr + len);
1567 
1568 		if (!may_expand_vm(mm, vm_flags,
1569 					(len >> PAGE_SHIFT) - nr_pages))
1570 			return -ENOMEM;
1571 	}
1572 
1573 	/* Clear old maps */
1574 	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1575 			      &rb_parent)) {
1576 		if (do_munmap(mm, addr, len))
1577 			return -ENOMEM;
1578 	}
1579 
1580 	/*
1581 	 * Private writable mapping: check memory availability
1582 	 */
1583 	if (accountable_mapping(file, vm_flags)) {
1584 		charged = len >> PAGE_SHIFT;
1585 		if (security_vm_enough_memory_mm(mm, charged))
1586 			return -ENOMEM;
1587 		vm_flags |= VM_ACCOUNT;
1588 	}
1589 
1590 	/*
1591 	 * Can we just expand an old mapping?
1592 	 */
1593 	vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1594 			NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1595 	if (vma)
1596 		goto out;
1597 
1598 	/*
1599 	 * Determine the object being mapped and call the appropriate
1600 	 * specific mapper. the address has already been validated, but
1601 	 * not unmapped, but the maps are removed from the list.
1602 	 */
1603 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1604 	if (!vma) {
1605 		error = -ENOMEM;
1606 		goto unacct_error;
1607 	}
1608 
1609 	vma->vm_mm = mm;
1610 	vma->vm_start = addr;
1611 	vma->vm_end = addr + len;
1612 	vma->vm_flags = vm_flags;
1613 	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1614 	vma->vm_pgoff = pgoff;
1615 	INIT_LIST_HEAD(&vma->anon_vma_chain);
1616 
1617 	if (file) {
1618 		if (vm_flags & VM_DENYWRITE) {
1619 			error = deny_write_access(file);
1620 			if (error)
1621 				goto free_vma;
1622 		}
1623 		if (vm_flags & VM_SHARED) {
1624 			error = mapping_map_writable(file->f_mapping);
1625 			if (error)
1626 				goto allow_write_and_free_vma;
1627 		}
1628 
1629 		/* ->mmap() can change vma->vm_file, but must guarantee that
1630 		 * vma_link() below can deny write-access if VM_DENYWRITE is set
1631 		 * and map writably if VM_SHARED is set. This usually means the
1632 		 * new file must not have been exposed to user-space, yet.
1633 		 */
1634 		vma->vm_file = get_file(file);
1635 		error = file->f_op->mmap(file, vma);
1636 		if (error)
1637 			goto unmap_and_free_vma;
1638 
1639 		/* Can addr have changed??
1640 		 *
1641 		 * Answer: Yes, several device drivers can do it in their
1642 		 *         f_op->mmap method. -DaveM
1643 		 * Bug: If addr is changed, prev, rb_link, rb_parent should
1644 		 *      be updated for vma_link()
1645 		 */
1646 		WARN_ON_ONCE(addr != vma->vm_start);
1647 
1648 		addr = vma->vm_start;
1649 		vm_flags = vma->vm_flags;
1650 	} else if (vm_flags & VM_SHARED) {
1651 		error = shmem_zero_setup(vma);
1652 		if (error)
1653 			goto free_vma;
1654 	}
1655 
1656 	vma_link(mm, vma, prev, rb_link, rb_parent);
1657 	/* Once vma denies write, undo our temporary denial count */
1658 	if (file) {
1659 		if (vm_flags & VM_SHARED)
1660 			mapping_unmap_writable(file->f_mapping);
1661 		if (vm_flags & VM_DENYWRITE)
1662 			allow_write_access(file);
1663 	}
1664 	file = vma->vm_file;
1665 out:
1666 	perf_event_mmap(vma);
1667 
1668 	vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1669 	if (vm_flags & VM_LOCKED) {
1670 		if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1671 					vma == get_gate_vma(current->mm)))
1672 			mm->locked_vm += (len >> PAGE_SHIFT);
1673 		else
1674 			vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1675 	}
1676 
1677 	if (file)
1678 		uprobe_mmap(vma);
1679 
1680 	/*
1681 	 * New (or expanded) vma always get soft dirty status.
1682 	 * Otherwise user-space soft-dirty page tracker won't
1683 	 * be able to distinguish situation when vma area unmapped,
1684 	 * then new mapped in-place (which must be aimed as
1685 	 * a completely new data area).
1686 	 */
1687 	vma->vm_flags |= VM_SOFTDIRTY;
1688 
1689 	vma_set_page_prot(vma);
1690 
1691 	return addr;
1692 
1693 unmap_and_free_vma:
1694 	vma->vm_file = NULL;
1695 	fput(file);
1696 
1697 	/* Undo any partial mapping done by a device driver. */
1698 	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1699 	charged = 0;
1700 	if (vm_flags & VM_SHARED)
1701 		mapping_unmap_writable(file->f_mapping);
1702 allow_write_and_free_vma:
1703 	if (vm_flags & VM_DENYWRITE)
1704 		allow_write_access(file);
1705 free_vma:
1706 	kmem_cache_free(vm_area_cachep, vma);
1707 unacct_error:
1708 	if (charged)
1709 		vm_unacct_memory(charged);
1710 	return error;
1711 }
1712 
1713 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1714 {
1715 	/*
1716 	 * We implement the search by looking for an rbtree node that
1717 	 * immediately follows a suitable gap. That is,
1718 	 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1719 	 * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1720 	 * - gap_end - gap_start >= length
1721 	 */
1722 
1723 	struct mm_struct *mm = current->mm;
1724 	struct vm_area_struct *vma;
1725 	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1726 
1727 	/* Adjust search length to account for worst case alignment overhead */
1728 	length = info->length + info->align_mask;
1729 	if (length < info->length)
1730 		return -ENOMEM;
1731 
1732 	/* Adjust search limits by the desired length */
1733 	if (info->high_limit < length)
1734 		return -ENOMEM;
1735 	high_limit = info->high_limit - length;
1736 
1737 	if (info->low_limit > high_limit)
1738 		return -ENOMEM;
1739 	low_limit = info->low_limit + length;
1740 
1741 	/* Check if rbtree root looks promising */
1742 	if (RB_EMPTY_ROOT(&mm->mm_rb))
1743 		goto check_highest;
1744 	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1745 	if (vma->rb_subtree_gap < length)
1746 		goto check_highest;
1747 
1748 	while (true) {
1749 		/* Visit left subtree if it looks promising */
1750 		gap_end = vma->vm_start;
1751 		if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1752 			struct vm_area_struct *left =
1753 				rb_entry(vma->vm_rb.rb_left,
1754 					 struct vm_area_struct, vm_rb);
1755 			if (left->rb_subtree_gap >= length) {
1756 				vma = left;
1757 				continue;
1758 			}
1759 		}
1760 
1761 		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1762 check_current:
1763 		/* Check if current node has a suitable gap */
1764 		if (gap_start > high_limit)
1765 			return -ENOMEM;
1766 		if (gap_end >= low_limit && gap_end - gap_start >= length)
1767 			goto found;
1768 
1769 		/* Visit right subtree if it looks promising */
1770 		if (vma->vm_rb.rb_right) {
1771 			struct vm_area_struct *right =
1772 				rb_entry(vma->vm_rb.rb_right,
1773 					 struct vm_area_struct, vm_rb);
1774 			if (right->rb_subtree_gap >= length) {
1775 				vma = right;
1776 				continue;
1777 			}
1778 		}
1779 
1780 		/* Go back up the rbtree to find next candidate node */
1781 		while (true) {
1782 			struct rb_node *prev = &vma->vm_rb;
1783 			if (!rb_parent(prev))
1784 				goto check_highest;
1785 			vma = rb_entry(rb_parent(prev),
1786 				       struct vm_area_struct, vm_rb);
1787 			if (prev == vma->vm_rb.rb_left) {
1788 				gap_start = vma->vm_prev->vm_end;
1789 				gap_end = vma->vm_start;
1790 				goto check_current;
1791 			}
1792 		}
1793 	}
1794 
1795 check_highest:
1796 	/* Check highest gap, which does not precede any rbtree node */
1797 	gap_start = mm->highest_vm_end;
1798 	gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1799 	if (gap_start > high_limit)
1800 		return -ENOMEM;
1801 
1802 found:
1803 	/* We found a suitable gap. Clip it with the original low_limit. */
1804 	if (gap_start < info->low_limit)
1805 		gap_start = info->low_limit;
1806 
1807 	/* Adjust gap address to the desired alignment */
1808 	gap_start += (info->align_offset - gap_start) & info->align_mask;
1809 
1810 	VM_BUG_ON(gap_start + info->length > info->high_limit);
1811 	VM_BUG_ON(gap_start + info->length > gap_end);
1812 	return gap_start;
1813 }
1814 
1815 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1816 {
1817 	struct mm_struct *mm = current->mm;
1818 	struct vm_area_struct *vma;
1819 	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1820 
1821 	/* Adjust search length to account for worst case alignment overhead */
1822 	length = info->length + info->align_mask;
1823 	if (length < info->length)
1824 		return -ENOMEM;
1825 
1826 	/*
1827 	 * Adjust search limits by the desired length.
1828 	 * See implementation comment at top of unmapped_area().
1829 	 */
1830 	gap_end = info->high_limit;
1831 	if (gap_end < length)
1832 		return -ENOMEM;
1833 	high_limit = gap_end - length;
1834 
1835 	if (info->low_limit > high_limit)
1836 		return -ENOMEM;
1837 	low_limit = info->low_limit + length;
1838 
1839 	/* Check highest gap, which does not precede any rbtree node */
1840 	gap_start = mm->highest_vm_end;
1841 	if (gap_start <= high_limit)
1842 		goto found_highest;
1843 
1844 	/* Check if rbtree root looks promising */
1845 	if (RB_EMPTY_ROOT(&mm->mm_rb))
1846 		return -ENOMEM;
1847 	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1848 	if (vma->rb_subtree_gap < length)
1849 		return -ENOMEM;
1850 
1851 	while (true) {
1852 		/* Visit right subtree if it looks promising */
1853 		gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1854 		if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1855 			struct vm_area_struct *right =
1856 				rb_entry(vma->vm_rb.rb_right,
1857 					 struct vm_area_struct, vm_rb);
1858 			if (right->rb_subtree_gap >= length) {
1859 				vma = right;
1860 				continue;
1861 			}
1862 		}
1863 
1864 check_current:
1865 		/* Check if current node has a suitable gap */
1866 		gap_end = vma->vm_start;
1867 		if (gap_end < low_limit)
1868 			return -ENOMEM;
1869 		if (gap_start <= high_limit && gap_end - gap_start >= length)
1870 			goto found;
1871 
1872 		/* Visit left subtree if it looks promising */
1873 		if (vma->vm_rb.rb_left) {
1874 			struct vm_area_struct *left =
1875 				rb_entry(vma->vm_rb.rb_left,
1876 					 struct vm_area_struct, vm_rb);
1877 			if (left->rb_subtree_gap >= length) {
1878 				vma = left;
1879 				continue;
1880 			}
1881 		}
1882 
1883 		/* Go back up the rbtree to find next candidate node */
1884 		while (true) {
1885 			struct rb_node *prev = &vma->vm_rb;
1886 			if (!rb_parent(prev))
1887 				return -ENOMEM;
1888 			vma = rb_entry(rb_parent(prev),
1889 				       struct vm_area_struct, vm_rb);
1890 			if (prev == vma->vm_rb.rb_right) {
1891 				gap_start = vma->vm_prev ?
1892 					vma->vm_prev->vm_end : 0;
1893 				goto check_current;
1894 			}
1895 		}
1896 	}
1897 
1898 found:
1899 	/* We found a suitable gap. Clip it with the original high_limit. */
1900 	if (gap_end > info->high_limit)
1901 		gap_end = info->high_limit;
1902 
1903 found_highest:
1904 	/* Compute highest gap address at the desired alignment */
1905 	gap_end -= info->length;
1906 	gap_end -= (gap_end - info->align_offset) & info->align_mask;
1907 
1908 	VM_BUG_ON(gap_end < info->low_limit);
1909 	VM_BUG_ON(gap_end < gap_start);
1910 	return gap_end;
1911 }
1912 
1913 /* Get an address range which is currently unmapped.
1914  * For shmat() with addr=0.
1915  *
1916  * Ugly calling convention alert:
1917  * Return value with the low bits set means error value,
1918  * ie
1919  *	if (ret & ~PAGE_MASK)
1920  *		error = ret;
1921  *
1922  * This function "knows" that -ENOMEM has the bits set.
1923  */
1924 #ifndef HAVE_ARCH_UNMAPPED_AREA
1925 unsigned long
1926 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1927 		unsigned long len, unsigned long pgoff, unsigned long flags)
1928 {
1929 	struct mm_struct *mm = current->mm;
1930 	struct vm_area_struct *vma;
1931 	struct vm_unmapped_area_info info;
1932 
1933 	if (len > TASK_SIZE - mmap_min_addr)
1934 		return -ENOMEM;
1935 
1936 	if (flags & MAP_FIXED)
1937 		return addr;
1938 
1939 	if (addr) {
1940 		addr = PAGE_ALIGN(addr);
1941 		vma = find_vma(mm, addr);
1942 		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1943 		    (!vma || addr + len <= vma->vm_start))
1944 			return addr;
1945 	}
1946 
1947 	info.flags = 0;
1948 	info.length = len;
1949 	info.low_limit = mm->mmap_base;
1950 	info.high_limit = TASK_SIZE;
1951 	info.align_mask = 0;
1952 	return vm_unmapped_area(&info);
1953 }
1954 #endif
1955 
1956 /*
1957  * This mmap-allocator allocates new areas top-down from below the
1958  * stack's low limit (the base):
1959  */
1960 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1961 unsigned long
1962 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1963 			  const unsigned long len, const unsigned long pgoff,
1964 			  const unsigned long flags)
1965 {
1966 	struct vm_area_struct *vma;
1967 	struct mm_struct *mm = current->mm;
1968 	unsigned long addr = addr0;
1969 	struct vm_unmapped_area_info info;
1970 
1971 	/* requested length too big for entire address space */
1972 	if (len > TASK_SIZE - mmap_min_addr)
1973 		return -ENOMEM;
1974 
1975 	if (flags & MAP_FIXED)
1976 		return addr;
1977 
1978 	/* requesting a specific address */
1979 	if (addr) {
1980 		addr = PAGE_ALIGN(addr);
1981 		vma = find_vma(mm, addr);
1982 		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1983 				(!vma || addr + len <= vma->vm_start))
1984 			return addr;
1985 	}
1986 
1987 	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1988 	info.length = len;
1989 	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1990 	info.high_limit = mm->mmap_base;
1991 	info.align_mask = 0;
1992 	addr = vm_unmapped_area(&info);
1993 
1994 	/*
1995 	 * A failed mmap() very likely causes application failure,
1996 	 * so fall back to the bottom-up function here. This scenario
1997 	 * can happen with large stack limits and large mmap()
1998 	 * allocations.
1999 	 */
2000 	if (offset_in_page(addr)) {
2001 		VM_BUG_ON(addr != -ENOMEM);
2002 		info.flags = 0;
2003 		info.low_limit = TASK_UNMAPPED_BASE;
2004 		info.high_limit = TASK_SIZE;
2005 		addr = vm_unmapped_area(&info);
2006 	}
2007 
2008 	return addr;
2009 }
2010 #endif
2011 
2012 unsigned long
2013 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2014 		unsigned long pgoff, unsigned long flags)
2015 {
2016 	unsigned long (*get_area)(struct file *, unsigned long,
2017 				  unsigned long, unsigned long, unsigned long);
2018 
2019 	unsigned long error = arch_mmap_check(addr, len, flags);
2020 	if (error)
2021 		return error;
2022 
2023 	/* Careful about overflows.. */
2024 	if (len > TASK_SIZE)
2025 		return -ENOMEM;
2026 
2027 	get_area = current->mm->get_unmapped_area;
2028 	if (file && file->f_op->get_unmapped_area)
2029 		get_area = file->f_op->get_unmapped_area;
2030 	addr = get_area(file, addr, len, pgoff, flags);
2031 	if (IS_ERR_VALUE(addr))
2032 		return addr;
2033 
2034 	if (addr > TASK_SIZE - len)
2035 		return -ENOMEM;
2036 	if (offset_in_page(addr))
2037 		return -EINVAL;
2038 
2039 	addr = arch_rebalance_pgtables(addr, len);
2040 	error = security_mmap_addr(addr);
2041 	return error ? error : addr;
2042 }
2043 
2044 EXPORT_SYMBOL(get_unmapped_area);
2045 
2046 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2047 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2048 {
2049 	struct rb_node *rb_node;
2050 	struct vm_area_struct *vma;
2051 
2052 	/* Check the cache first. */
2053 	vma = vmacache_find(mm, addr);
2054 	if (likely(vma))
2055 		return vma;
2056 
2057 	rb_node = mm->mm_rb.rb_node;
2058 
2059 	while (rb_node) {
2060 		struct vm_area_struct *tmp;
2061 
2062 		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2063 
2064 		if (tmp->vm_end > addr) {
2065 			vma = tmp;
2066 			if (tmp->vm_start <= addr)
2067 				break;
2068 			rb_node = rb_node->rb_left;
2069 		} else
2070 			rb_node = rb_node->rb_right;
2071 	}
2072 
2073 	if (vma)
2074 		vmacache_update(addr, vma);
2075 	return vma;
2076 }
2077 
2078 EXPORT_SYMBOL(find_vma);
2079 
2080 /*
2081  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2082  */
2083 struct vm_area_struct *
2084 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2085 			struct vm_area_struct **pprev)
2086 {
2087 	struct vm_area_struct *vma;
2088 
2089 	vma = find_vma(mm, addr);
2090 	if (vma) {
2091 		*pprev = vma->vm_prev;
2092 	} else {
2093 		struct rb_node *rb_node = mm->mm_rb.rb_node;
2094 		*pprev = NULL;
2095 		while (rb_node) {
2096 			*pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2097 			rb_node = rb_node->rb_right;
2098 		}
2099 	}
2100 	return vma;
2101 }
2102 
2103 /*
2104  * Verify that the stack growth is acceptable and
2105  * update accounting. This is shared with both the
2106  * grow-up and grow-down cases.
2107  */
2108 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2109 {
2110 	struct mm_struct *mm = vma->vm_mm;
2111 	struct rlimit *rlim = current->signal->rlim;
2112 	unsigned long new_start, actual_size;
2113 
2114 	/* address space limit tests */
2115 	if (!may_expand_vm(mm, vma->vm_flags, grow))
2116 		return -ENOMEM;
2117 
2118 	/* Stack limit test */
2119 	actual_size = size;
2120 	if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
2121 		actual_size -= PAGE_SIZE;
2122 	if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2123 		return -ENOMEM;
2124 
2125 	/* mlock limit tests */
2126 	if (vma->vm_flags & VM_LOCKED) {
2127 		unsigned long locked;
2128 		unsigned long limit;
2129 		locked = mm->locked_vm + grow;
2130 		limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2131 		limit >>= PAGE_SHIFT;
2132 		if (locked > limit && !capable(CAP_IPC_LOCK))
2133 			return -ENOMEM;
2134 	}
2135 
2136 	/* Check to ensure the stack will not grow into a hugetlb-only region */
2137 	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2138 			vma->vm_end - size;
2139 	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2140 		return -EFAULT;
2141 
2142 	/*
2143 	 * Overcommit..  This must be the final test, as it will
2144 	 * update security statistics.
2145 	 */
2146 	if (security_vm_enough_memory_mm(mm, grow))
2147 		return -ENOMEM;
2148 
2149 	return 0;
2150 }
2151 
2152 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2153 /*
2154  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2155  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2156  */
2157 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2158 {
2159 	struct mm_struct *mm = vma->vm_mm;
2160 	int error = 0;
2161 
2162 	if (!(vma->vm_flags & VM_GROWSUP))
2163 		return -EFAULT;
2164 
2165 	/* Guard against wrapping around to address 0. */
2166 	if (address < PAGE_ALIGN(address+4))
2167 		address = PAGE_ALIGN(address+4);
2168 	else
2169 		return -ENOMEM;
2170 
2171 	/* We must make sure the anon_vma is allocated. */
2172 	if (unlikely(anon_vma_prepare(vma)))
2173 		return -ENOMEM;
2174 
2175 	/*
2176 	 * vma->vm_start/vm_end cannot change under us because the caller
2177 	 * is required to hold the mmap_sem in read mode.  We need the
2178 	 * anon_vma lock to serialize against concurrent expand_stacks.
2179 	 */
2180 	anon_vma_lock_write(vma->anon_vma);
2181 
2182 	/* Somebody else might have raced and expanded it already */
2183 	if (address > vma->vm_end) {
2184 		unsigned long size, grow;
2185 
2186 		size = address - vma->vm_start;
2187 		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2188 
2189 		error = -ENOMEM;
2190 		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2191 			error = acct_stack_growth(vma, size, grow);
2192 			if (!error) {
2193 				/*
2194 				 * vma_gap_update() doesn't support concurrent
2195 				 * updates, but we only hold a shared mmap_sem
2196 				 * lock here, so we need to protect against
2197 				 * concurrent vma expansions.
2198 				 * anon_vma_lock_write() doesn't help here, as
2199 				 * we don't guarantee that all growable vmas
2200 				 * in a mm share the same root anon vma.
2201 				 * So, we reuse mm->page_table_lock to guard
2202 				 * against concurrent vma expansions.
2203 				 */
2204 				spin_lock(&mm->page_table_lock);
2205 				if (vma->vm_flags & VM_LOCKED)
2206 					mm->locked_vm += grow;
2207 				vm_stat_account(mm, vma->vm_flags, grow);
2208 				anon_vma_interval_tree_pre_update_vma(vma);
2209 				vma->vm_end = address;
2210 				anon_vma_interval_tree_post_update_vma(vma);
2211 				if (vma->vm_next)
2212 					vma_gap_update(vma->vm_next);
2213 				else
2214 					mm->highest_vm_end = address;
2215 				spin_unlock(&mm->page_table_lock);
2216 
2217 				perf_event_mmap(vma);
2218 			}
2219 		}
2220 	}
2221 	anon_vma_unlock_write(vma->anon_vma);
2222 	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2223 	validate_mm(mm);
2224 	return error;
2225 }
2226 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2227 
2228 /*
2229  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2230  */
2231 int expand_downwards(struct vm_area_struct *vma,
2232 				   unsigned long address)
2233 {
2234 	struct mm_struct *mm = vma->vm_mm;
2235 	int error;
2236 
2237 	address &= PAGE_MASK;
2238 	error = security_mmap_addr(address);
2239 	if (error)
2240 		return error;
2241 
2242 	/* We must make sure the anon_vma is allocated. */
2243 	if (unlikely(anon_vma_prepare(vma)))
2244 		return -ENOMEM;
2245 
2246 	/*
2247 	 * vma->vm_start/vm_end cannot change under us because the caller
2248 	 * is required to hold the mmap_sem in read mode.  We need the
2249 	 * anon_vma lock to serialize against concurrent expand_stacks.
2250 	 */
2251 	anon_vma_lock_write(vma->anon_vma);
2252 
2253 	/* Somebody else might have raced and expanded it already */
2254 	if (address < vma->vm_start) {
2255 		unsigned long size, grow;
2256 
2257 		size = vma->vm_end - address;
2258 		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2259 
2260 		error = -ENOMEM;
2261 		if (grow <= vma->vm_pgoff) {
2262 			error = acct_stack_growth(vma, size, grow);
2263 			if (!error) {
2264 				/*
2265 				 * vma_gap_update() doesn't support concurrent
2266 				 * updates, but we only hold a shared mmap_sem
2267 				 * lock here, so we need to protect against
2268 				 * concurrent vma expansions.
2269 				 * anon_vma_lock_write() doesn't help here, as
2270 				 * we don't guarantee that all growable vmas
2271 				 * in a mm share the same root anon vma.
2272 				 * So, we reuse mm->page_table_lock to guard
2273 				 * against concurrent vma expansions.
2274 				 */
2275 				spin_lock(&mm->page_table_lock);
2276 				if (vma->vm_flags & VM_LOCKED)
2277 					mm->locked_vm += grow;
2278 				vm_stat_account(mm, vma->vm_flags, grow);
2279 				anon_vma_interval_tree_pre_update_vma(vma);
2280 				vma->vm_start = address;
2281 				vma->vm_pgoff -= grow;
2282 				anon_vma_interval_tree_post_update_vma(vma);
2283 				vma_gap_update(vma);
2284 				spin_unlock(&mm->page_table_lock);
2285 
2286 				perf_event_mmap(vma);
2287 			}
2288 		}
2289 	}
2290 	anon_vma_unlock_write(vma->anon_vma);
2291 	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2292 	validate_mm(mm);
2293 	return error;
2294 }
2295 
2296 /*
2297  * Note how expand_stack() refuses to expand the stack all the way to
2298  * abut the next virtual mapping, *unless* that mapping itself is also
2299  * a stack mapping. We want to leave room for a guard page, after all
2300  * (the guard page itself is not added here, that is done by the
2301  * actual page faulting logic)
2302  *
2303  * This matches the behavior of the guard page logic (see mm/memory.c:
2304  * check_stack_guard_page()), which only allows the guard page to be
2305  * removed under these circumstances.
2306  */
2307 #ifdef CONFIG_STACK_GROWSUP
2308 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2309 {
2310 	struct vm_area_struct *next;
2311 
2312 	address &= PAGE_MASK;
2313 	next = vma->vm_next;
2314 	if (next && next->vm_start == address + PAGE_SIZE) {
2315 		if (!(next->vm_flags & VM_GROWSUP))
2316 			return -ENOMEM;
2317 	}
2318 	return expand_upwards(vma, address);
2319 }
2320 
2321 struct vm_area_struct *
2322 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2323 {
2324 	struct vm_area_struct *vma, *prev;
2325 
2326 	addr &= PAGE_MASK;
2327 	vma = find_vma_prev(mm, addr, &prev);
2328 	if (vma && (vma->vm_start <= addr))
2329 		return vma;
2330 	if (!prev || expand_stack(prev, addr))
2331 		return NULL;
2332 	if (prev->vm_flags & VM_LOCKED)
2333 		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2334 	return prev;
2335 }
2336 #else
2337 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2338 {
2339 	struct vm_area_struct *prev;
2340 
2341 	address &= PAGE_MASK;
2342 	prev = vma->vm_prev;
2343 	if (prev && prev->vm_end == address) {
2344 		if (!(prev->vm_flags & VM_GROWSDOWN))
2345 			return -ENOMEM;
2346 	}
2347 	return expand_downwards(vma, address);
2348 }
2349 
2350 struct vm_area_struct *
2351 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2352 {
2353 	struct vm_area_struct *vma;
2354 	unsigned long start;
2355 
2356 	addr &= PAGE_MASK;
2357 	vma = find_vma(mm, addr);
2358 	if (!vma)
2359 		return NULL;
2360 	if (vma->vm_start <= addr)
2361 		return vma;
2362 	if (!(vma->vm_flags & VM_GROWSDOWN))
2363 		return NULL;
2364 	start = vma->vm_start;
2365 	if (expand_stack(vma, addr))
2366 		return NULL;
2367 	if (vma->vm_flags & VM_LOCKED)
2368 		populate_vma_page_range(vma, addr, start, NULL);
2369 	return vma;
2370 }
2371 #endif
2372 
2373 EXPORT_SYMBOL_GPL(find_extend_vma);
2374 
2375 /*
2376  * Ok - we have the memory areas we should free on the vma list,
2377  * so release them, and do the vma updates.
2378  *
2379  * Called with the mm semaphore held.
2380  */
2381 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2382 {
2383 	unsigned long nr_accounted = 0;
2384 
2385 	/* Update high watermark before we lower total_vm */
2386 	update_hiwater_vm(mm);
2387 	do {
2388 		long nrpages = vma_pages(vma);
2389 
2390 		if (vma->vm_flags & VM_ACCOUNT)
2391 			nr_accounted += nrpages;
2392 		vm_stat_account(mm, vma->vm_flags, -nrpages);
2393 		vma = remove_vma(vma);
2394 	} while (vma);
2395 	vm_unacct_memory(nr_accounted);
2396 	validate_mm(mm);
2397 }
2398 
2399 /*
2400  * Get rid of page table information in the indicated region.
2401  *
2402  * Called with the mm semaphore held.
2403  */
2404 static void unmap_region(struct mm_struct *mm,
2405 		struct vm_area_struct *vma, struct vm_area_struct *prev,
2406 		unsigned long start, unsigned long end)
2407 {
2408 	struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2409 	struct mmu_gather tlb;
2410 
2411 	lru_add_drain();
2412 	tlb_gather_mmu(&tlb, mm, start, end);
2413 	update_hiwater_rss(mm);
2414 	unmap_vmas(&tlb, vma, start, end);
2415 	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2416 				 next ? next->vm_start : USER_PGTABLES_CEILING);
2417 	tlb_finish_mmu(&tlb, start, end);
2418 }
2419 
2420 /*
2421  * Create a list of vma's touched by the unmap, removing them from the mm's
2422  * vma list as we go..
2423  */
2424 static void
2425 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2426 	struct vm_area_struct *prev, unsigned long end)
2427 {
2428 	struct vm_area_struct **insertion_point;
2429 	struct vm_area_struct *tail_vma = NULL;
2430 
2431 	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2432 	vma->vm_prev = NULL;
2433 	do {
2434 		vma_rb_erase(vma, &mm->mm_rb);
2435 		mm->map_count--;
2436 		tail_vma = vma;
2437 		vma = vma->vm_next;
2438 	} while (vma && vma->vm_start < end);
2439 	*insertion_point = vma;
2440 	if (vma) {
2441 		vma->vm_prev = prev;
2442 		vma_gap_update(vma);
2443 	} else
2444 		mm->highest_vm_end = prev ? prev->vm_end : 0;
2445 	tail_vma->vm_next = NULL;
2446 
2447 	/* Kill the cache */
2448 	vmacache_invalidate(mm);
2449 }
2450 
2451 /*
2452  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2453  * munmap path where it doesn't make sense to fail.
2454  */
2455 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2456 	      unsigned long addr, int new_below)
2457 {
2458 	struct vm_area_struct *new;
2459 	int err;
2460 
2461 	if (is_vm_hugetlb_page(vma) && (addr &
2462 					~(huge_page_mask(hstate_vma(vma)))))
2463 		return -EINVAL;
2464 
2465 	new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2466 	if (!new)
2467 		return -ENOMEM;
2468 
2469 	/* most fields are the same, copy all, and then fixup */
2470 	*new = *vma;
2471 
2472 	INIT_LIST_HEAD(&new->anon_vma_chain);
2473 
2474 	if (new_below)
2475 		new->vm_end = addr;
2476 	else {
2477 		new->vm_start = addr;
2478 		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2479 	}
2480 
2481 	err = vma_dup_policy(vma, new);
2482 	if (err)
2483 		goto out_free_vma;
2484 
2485 	err = anon_vma_clone(new, vma);
2486 	if (err)
2487 		goto out_free_mpol;
2488 
2489 	if (new->vm_file)
2490 		get_file(new->vm_file);
2491 
2492 	if (new->vm_ops && new->vm_ops->open)
2493 		new->vm_ops->open(new);
2494 
2495 	if (new_below)
2496 		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2497 			((addr - new->vm_start) >> PAGE_SHIFT), new);
2498 	else
2499 		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2500 
2501 	/* Success. */
2502 	if (!err)
2503 		return 0;
2504 
2505 	/* Clean everything up if vma_adjust failed. */
2506 	if (new->vm_ops && new->vm_ops->close)
2507 		new->vm_ops->close(new);
2508 	if (new->vm_file)
2509 		fput(new->vm_file);
2510 	unlink_anon_vmas(new);
2511  out_free_mpol:
2512 	mpol_put(vma_policy(new));
2513  out_free_vma:
2514 	kmem_cache_free(vm_area_cachep, new);
2515 	return err;
2516 }
2517 
2518 /*
2519  * Split a vma into two pieces at address 'addr', a new vma is allocated
2520  * either for the first part or the tail.
2521  */
2522 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2523 	      unsigned long addr, int new_below)
2524 {
2525 	if (mm->map_count >= sysctl_max_map_count)
2526 		return -ENOMEM;
2527 
2528 	return __split_vma(mm, vma, addr, new_below);
2529 }
2530 
2531 /* Munmap is split into 2 main parts -- this part which finds
2532  * what needs doing, and the areas themselves, which do the
2533  * work.  This now handles partial unmappings.
2534  * Jeremy Fitzhardinge <jeremy@goop.org>
2535  */
2536 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2537 {
2538 	unsigned long end;
2539 	struct vm_area_struct *vma, *prev, *last;
2540 
2541 	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2542 		return -EINVAL;
2543 
2544 	len = PAGE_ALIGN(len);
2545 	if (len == 0)
2546 		return -EINVAL;
2547 
2548 	/* Find the first overlapping VMA */
2549 	vma = find_vma(mm, start);
2550 	if (!vma)
2551 		return 0;
2552 	prev = vma->vm_prev;
2553 	/* we have  start < vma->vm_end  */
2554 
2555 	/* if it doesn't overlap, we have nothing.. */
2556 	end = start + len;
2557 	if (vma->vm_start >= end)
2558 		return 0;
2559 
2560 	/*
2561 	 * If we need to split any vma, do it now to save pain later.
2562 	 *
2563 	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2564 	 * unmapped vm_area_struct will remain in use: so lower split_vma
2565 	 * places tmp vma above, and higher split_vma places tmp vma below.
2566 	 */
2567 	if (start > vma->vm_start) {
2568 		int error;
2569 
2570 		/*
2571 		 * Make sure that map_count on return from munmap() will
2572 		 * not exceed its limit; but let map_count go just above
2573 		 * its limit temporarily, to help free resources as expected.
2574 		 */
2575 		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2576 			return -ENOMEM;
2577 
2578 		error = __split_vma(mm, vma, start, 0);
2579 		if (error)
2580 			return error;
2581 		prev = vma;
2582 	}
2583 
2584 	/* Does it split the last one? */
2585 	last = find_vma(mm, end);
2586 	if (last && end > last->vm_start) {
2587 		int error = __split_vma(mm, last, end, 1);
2588 		if (error)
2589 			return error;
2590 	}
2591 	vma = prev ? prev->vm_next : mm->mmap;
2592 
2593 	/*
2594 	 * unlock any mlock()ed ranges before detaching vmas
2595 	 */
2596 	if (mm->locked_vm) {
2597 		struct vm_area_struct *tmp = vma;
2598 		while (tmp && tmp->vm_start < end) {
2599 			if (tmp->vm_flags & VM_LOCKED) {
2600 				mm->locked_vm -= vma_pages(tmp);
2601 				munlock_vma_pages_all(tmp);
2602 			}
2603 			tmp = tmp->vm_next;
2604 		}
2605 	}
2606 
2607 	/*
2608 	 * Remove the vma's, and unmap the actual pages
2609 	 */
2610 	detach_vmas_to_be_unmapped(mm, vma, prev, end);
2611 	unmap_region(mm, vma, prev, start, end);
2612 
2613 	arch_unmap(mm, vma, start, end);
2614 
2615 	/* Fix up all other VM information */
2616 	remove_vma_list(mm, vma);
2617 
2618 	return 0;
2619 }
2620 
2621 int vm_munmap(unsigned long start, size_t len)
2622 {
2623 	int ret;
2624 	struct mm_struct *mm = current->mm;
2625 
2626 	down_write(&mm->mmap_sem);
2627 	ret = do_munmap(mm, start, len);
2628 	up_write(&mm->mmap_sem);
2629 	return ret;
2630 }
2631 EXPORT_SYMBOL(vm_munmap);
2632 
2633 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2634 {
2635 	profile_munmap(addr);
2636 	return vm_munmap(addr, len);
2637 }
2638 
2639 
2640 /*
2641  * Emulation of deprecated remap_file_pages() syscall.
2642  */
2643 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2644 		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2645 {
2646 
2647 	struct mm_struct *mm = current->mm;
2648 	struct vm_area_struct *vma;
2649 	unsigned long populate = 0;
2650 	unsigned long ret = -EINVAL;
2651 	struct file *file;
2652 
2653 	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2654 			"See Documentation/vm/remap_file_pages.txt.\n",
2655 			current->comm, current->pid);
2656 
2657 	if (prot)
2658 		return ret;
2659 	start = start & PAGE_MASK;
2660 	size = size & PAGE_MASK;
2661 
2662 	if (start + size <= start)
2663 		return ret;
2664 
2665 	/* Does pgoff wrap? */
2666 	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2667 		return ret;
2668 
2669 	down_write(&mm->mmap_sem);
2670 	vma = find_vma(mm, start);
2671 
2672 	if (!vma || !(vma->vm_flags & VM_SHARED))
2673 		goto out;
2674 
2675 	if (start < vma->vm_start || start + size > vma->vm_end)
2676 		goto out;
2677 
2678 	if (pgoff == linear_page_index(vma, start)) {
2679 		ret = 0;
2680 		goto out;
2681 	}
2682 
2683 	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2684 	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2685 	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2686 
2687 	flags &= MAP_NONBLOCK;
2688 	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2689 	if (vma->vm_flags & VM_LOCKED) {
2690 		flags |= MAP_LOCKED;
2691 		/* drop PG_Mlocked flag for over-mapped range */
2692 		munlock_vma_pages_range(vma, start, start + size);
2693 	}
2694 
2695 	file = get_file(vma->vm_file);
2696 	ret = do_mmap_pgoff(vma->vm_file, start, size,
2697 			prot, flags, pgoff, &populate);
2698 	fput(file);
2699 out:
2700 	up_write(&mm->mmap_sem);
2701 	if (populate)
2702 		mm_populate(ret, populate);
2703 	if (!IS_ERR_VALUE(ret))
2704 		ret = 0;
2705 	return ret;
2706 }
2707 
2708 static inline void verify_mm_writelocked(struct mm_struct *mm)
2709 {
2710 #ifdef CONFIG_DEBUG_VM
2711 	if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2712 		WARN_ON(1);
2713 		up_read(&mm->mmap_sem);
2714 	}
2715 #endif
2716 }
2717 
2718 /*
2719  *  this is really a simplified "do_mmap".  it only handles
2720  *  anonymous maps.  eventually we may be able to do some
2721  *  brk-specific accounting here.
2722  */
2723 static unsigned long do_brk(unsigned long addr, unsigned long len)
2724 {
2725 	struct mm_struct *mm = current->mm;
2726 	struct vm_area_struct *vma, *prev;
2727 	unsigned long flags;
2728 	struct rb_node **rb_link, *rb_parent;
2729 	pgoff_t pgoff = addr >> PAGE_SHIFT;
2730 	int error;
2731 
2732 	len = PAGE_ALIGN(len);
2733 	if (!len)
2734 		return addr;
2735 
2736 	flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2737 
2738 	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2739 	if (offset_in_page(error))
2740 		return error;
2741 
2742 	error = mlock_future_check(mm, mm->def_flags, len);
2743 	if (error)
2744 		return error;
2745 
2746 	/*
2747 	 * mm->mmap_sem is required to protect against another thread
2748 	 * changing the mappings in case we sleep.
2749 	 */
2750 	verify_mm_writelocked(mm);
2751 
2752 	/*
2753 	 * Clear old maps.  this also does some error checking for us
2754 	 */
2755 	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2756 			      &rb_parent)) {
2757 		if (do_munmap(mm, addr, len))
2758 			return -ENOMEM;
2759 	}
2760 
2761 	/* Check against address space limits *after* clearing old maps... */
2762 	if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2763 		return -ENOMEM;
2764 
2765 	if (mm->map_count > sysctl_max_map_count)
2766 		return -ENOMEM;
2767 
2768 	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2769 		return -ENOMEM;
2770 
2771 	/* Can we just expand an old private anonymous mapping? */
2772 	vma = vma_merge(mm, prev, addr, addr + len, flags,
2773 			NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2774 	if (vma)
2775 		goto out;
2776 
2777 	/*
2778 	 * create a vma struct for an anonymous mapping
2779 	 */
2780 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2781 	if (!vma) {
2782 		vm_unacct_memory(len >> PAGE_SHIFT);
2783 		return -ENOMEM;
2784 	}
2785 
2786 	INIT_LIST_HEAD(&vma->anon_vma_chain);
2787 	vma->vm_mm = mm;
2788 	vma->vm_start = addr;
2789 	vma->vm_end = addr + len;
2790 	vma->vm_pgoff = pgoff;
2791 	vma->vm_flags = flags;
2792 	vma->vm_page_prot = vm_get_page_prot(flags);
2793 	vma_link(mm, vma, prev, rb_link, rb_parent);
2794 out:
2795 	perf_event_mmap(vma);
2796 	mm->total_vm += len >> PAGE_SHIFT;
2797 	mm->data_vm += len >> PAGE_SHIFT;
2798 	if (flags & VM_LOCKED)
2799 		mm->locked_vm += (len >> PAGE_SHIFT);
2800 	vma->vm_flags |= VM_SOFTDIRTY;
2801 	return addr;
2802 }
2803 
2804 unsigned long vm_brk(unsigned long addr, unsigned long len)
2805 {
2806 	struct mm_struct *mm = current->mm;
2807 	unsigned long ret;
2808 	bool populate;
2809 
2810 	down_write(&mm->mmap_sem);
2811 	ret = do_brk(addr, len);
2812 	populate = ((mm->def_flags & VM_LOCKED) != 0);
2813 	up_write(&mm->mmap_sem);
2814 	if (populate)
2815 		mm_populate(addr, len);
2816 	return ret;
2817 }
2818 EXPORT_SYMBOL(vm_brk);
2819 
2820 /* Release all mmaps. */
2821 void exit_mmap(struct mm_struct *mm)
2822 {
2823 	struct mmu_gather tlb;
2824 	struct vm_area_struct *vma;
2825 	unsigned long nr_accounted = 0;
2826 
2827 	/* mm's last user has gone, and its about to be pulled down */
2828 	mmu_notifier_release(mm);
2829 
2830 	if (mm->locked_vm) {
2831 		vma = mm->mmap;
2832 		while (vma) {
2833 			if (vma->vm_flags & VM_LOCKED)
2834 				munlock_vma_pages_all(vma);
2835 			vma = vma->vm_next;
2836 		}
2837 	}
2838 
2839 	arch_exit_mmap(mm);
2840 
2841 	vma = mm->mmap;
2842 	if (!vma)	/* Can happen if dup_mmap() received an OOM */
2843 		return;
2844 
2845 	lru_add_drain();
2846 	flush_cache_mm(mm);
2847 	tlb_gather_mmu(&tlb, mm, 0, -1);
2848 	/* update_hiwater_rss(mm) here? but nobody should be looking */
2849 	/* Use -1 here to ensure all VMAs in the mm are unmapped */
2850 	unmap_vmas(&tlb, vma, 0, -1);
2851 
2852 	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2853 	tlb_finish_mmu(&tlb, 0, -1);
2854 
2855 	/*
2856 	 * Walk the list again, actually closing and freeing it,
2857 	 * with preemption enabled, without holding any MM locks.
2858 	 */
2859 	while (vma) {
2860 		if (vma->vm_flags & VM_ACCOUNT)
2861 			nr_accounted += vma_pages(vma);
2862 		vma = remove_vma(vma);
2863 	}
2864 	vm_unacct_memory(nr_accounted);
2865 }
2866 
2867 /* Insert vm structure into process list sorted by address
2868  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2869  * then i_mmap_rwsem is taken here.
2870  */
2871 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2872 {
2873 	struct vm_area_struct *prev;
2874 	struct rb_node **rb_link, *rb_parent;
2875 
2876 	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2877 			   &prev, &rb_link, &rb_parent))
2878 		return -ENOMEM;
2879 	if ((vma->vm_flags & VM_ACCOUNT) &&
2880 	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
2881 		return -ENOMEM;
2882 
2883 	/*
2884 	 * The vm_pgoff of a purely anonymous vma should be irrelevant
2885 	 * until its first write fault, when page's anon_vma and index
2886 	 * are set.  But now set the vm_pgoff it will almost certainly
2887 	 * end up with (unless mremap moves it elsewhere before that
2888 	 * first wfault), so /proc/pid/maps tells a consistent story.
2889 	 *
2890 	 * By setting it to reflect the virtual start address of the
2891 	 * vma, merges and splits can happen in a seamless way, just
2892 	 * using the existing file pgoff checks and manipulations.
2893 	 * Similarly in do_mmap_pgoff and in do_brk.
2894 	 */
2895 	if (vma_is_anonymous(vma)) {
2896 		BUG_ON(vma->anon_vma);
2897 		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2898 	}
2899 
2900 	vma_link(mm, vma, prev, rb_link, rb_parent);
2901 	return 0;
2902 }
2903 
2904 /*
2905  * Copy the vma structure to a new location in the same mm,
2906  * prior to moving page table entries, to effect an mremap move.
2907  */
2908 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2909 	unsigned long addr, unsigned long len, pgoff_t pgoff,
2910 	bool *need_rmap_locks)
2911 {
2912 	struct vm_area_struct *vma = *vmap;
2913 	unsigned long vma_start = vma->vm_start;
2914 	struct mm_struct *mm = vma->vm_mm;
2915 	struct vm_area_struct *new_vma, *prev;
2916 	struct rb_node **rb_link, *rb_parent;
2917 	bool faulted_in_anon_vma = true;
2918 
2919 	/*
2920 	 * If anonymous vma has not yet been faulted, update new pgoff
2921 	 * to match new location, to increase its chance of merging.
2922 	 */
2923 	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2924 		pgoff = addr >> PAGE_SHIFT;
2925 		faulted_in_anon_vma = false;
2926 	}
2927 
2928 	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2929 		return NULL;	/* should never get here */
2930 	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2931 			    vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2932 			    vma->vm_userfaultfd_ctx);
2933 	if (new_vma) {
2934 		/*
2935 		 * Source vma may have been merged into new_vma
2936 		 */
2937 		if (unlikely(vma_start >= new_vma->vm_start &&
2938 			     vma_start < new_vma->vm_end)) {
2939 			/*
2940 			 * The only way we can get a vma_merge with
2941 			 * self during an mremap is if the vma hasn't
2942 			 * been faulted in yet and we were allowed to
2943 			 * reset the dst vma->vm_pgoff to the
2944 			 * destination address of the mremap to allow
2945 			 * the merge to happen. mremap must change the
2946 			 * vm_pgoff linearity between src and dst vmas
2947 			 * (in turn preventing a vma_merge) to be
2948 			 * safe. It is only safe to keep the vm_pgoff
2949 			 * linear if there are no pages mapped yet.
2950 			 */
2951 			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
2952 			*vmap = vma = new_vma;
2953 		}
2954 		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2955 	} else {
2956 		new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2957 		if (!new_vma)
2958 			goto out;
2959 		*new_vma = *vma;
2960 		new_vma->vm_start = addr;
2961 		new_vma->vm_end = addr + len;
2962 		new_vma->vm_pgoff = pgoff;
2963 		if (vma_dup_policy(vma, new_vma))
2964 			goto out_free_vma;
2965 		INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2966 		if (anon_vma_clone(new_vma, vma))
2967 			goto out_free_mempol;
2968 		if (new_vma->vm_file)
2969 			get_file(new_vma->vm_file);
2970 		if (new_vma->vm_ops && new_vma->vm_ops->open)
2971 			new_vma->vm_ops->open(new_vma);
2972 		vma_link(mm, new_vma, prev, rb_link, rb_parent);
2973 		*need_rmap_locks = false;
2974 	}
2975 	return new_vma;
2976 
2977 out_free_mempol:
2978 	mpol_put(vma_policy(new_vma));
2979 out_free_vma:
2980 	kmem_cache_free(vm_area_cachep, new_vma);
2981 out:
2982 	return NULL;
2983 }
2984 
2985 /*
2986  * Return true if the calling process may expand its vm space by the passed
2987  * number of pages
2988  */
2989 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
2990 {
2991 	if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
2992 		return false;
2993 
2994 	if (is_data_mapping(flags) &&
2995 	    mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
2996 		if (ignore_rlimit_data)
2997 			pr_warn_once("%s (%d): VmData %lu exceed data ulimit "
2998 				     "%lu. Will be forbidden soon.\n",
2999 				     current->comm, current->pid,
3000 				     (mm->data_vm + npages) << PAGE_SHIFT,
3001 				     rlimit(RLIMIT_DATA));
3002 		else
3003 			return false;
3004 	}
3005 
3006 	return true;
3007 }
3008 
3009 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3010 {
3011 	mm->total_vm += npages;
3012 
3013 	if (is_exec_mapping(flags))
3014 		mm->exec_vm += npages;
3015 	else if (is_stack_mapping(flags))
3016 		mm->stack_vm += npages;
3017 	else if (is_data_mapping(flags))
3018 		mm->data_vm += npages;
3019 }
3020 
3021 static int special_mapping_fault(struct vm_area_struct *vma,
3022 				 struct vm_fault *vmf);
3023 
3024 /*
3025  * Having a close hook prevents vma merging regardless of flags.
3026  */
3027 static void special_mapping_close(struct vm_area_struct *vma)
3028 {
3029 }
3030 
3031 static const char *special_mapping_name(struct vm_area_struct *vma)
3032 {
3033 	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3034 }
3035 
3036 static const struct vm_operations_struct special_mapping_vmops = {
3037 	.close = special_mapping_close,
3038 	.fault = special_mapping_fault,
3039 	.name = special_mapping_name,
3040 };
3041 
3042 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3043 	.close = special_mapping_close,
3044 	.fault = special_mapping_fault,
3045 };
3046 
3047 static int special_mapping_fault(struct vm_area_struct *vma,
3048 				struct vm_fault *vmf)
3049 {
3050 	pgoff_t pgoff;
3051 	struct page **pages;
3052 
3053 	if (vma->vm_ops == &legacy_special_mapping_vmops) {
3054 		pages = vma->vm_private_data;
3055 	} else {
3056 		struct vm_special_mapping *sm = vma->vm_private_data;
3057 
3058 		if (sm->fault)
3059 			return sm->fault(sm, vma, vmf);
3060 
3061 		pages = sm->pages;
3062 	}
3063 
3064 	for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3065 		pgoff--;
3066 
3067 	if (*pages) {
3068 		struct page *page = *pages;
3069 		get_page(page);
3070 		vmf->page = page;
3071 		return 0;
3072 	}
3073 
3074 	return VM_FAULT_SIGBUS;
3075 }
3076 
3077 static struct vm_area_struct *__install_special_mapping(
3078 	struct mm_struct *mm,
3079 	unsigned long addr, unsigned long len,
3080 	unsigned long vm_flags, void *priv,
3081 	const struct vm_operations_struct *ops)
3082 {
3083 	int ret;
3084 	struct vm_area_struct *vma;
3085 
3086 	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3087 	if (unlikely(vma == NULL))
3088 		return ERR_PTR(-ENOMEM);
3089 
3090 	INIT_LIST_HEAD(&vma->anon_vma_chain);
3091 	vma->vm_mm = mm;
3092 	vma->vm_start = addr;
3093 	vma->vm_end = addr + len;
3094 
3095 	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3096 	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3097 
3098 	vma->vm_ops = ops;
3099 	vma->vm_private_data = priv;
3100 
3101 	ret = insert_vm_struct(mm, vma);
3102 	if (ret)
3103 		goto out;
3104 
3105 	vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3106 
3107 	perf_event_mmap(vma);
3108 
3109 	return vma;
3110 
3111 out:
3112 	kmem_cache_free(vm_area_cachep, vma);
3113 	return ERR_PTR(ret);
3114 }
3115 
3116 /*
3117  * Called with mm->mmap_sem held for writing.
3118  * Insert a new vma covering the given region, with the given flags.
3119  * Its pages are supplied by the given array of struct page *.
3120  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3121  * The region past the last page supplied will always produce SIGBUS.
3122  * The array pointer and the pages it points to are assumed to stay alive
3123  * for as long as this mapping might exist.
3124  */
3125 struct vm_area_struct *_install_special_mapping(
3126 	struct mm_struct *mm,
3127 	unsigned long addr, unsigned long len,
3128 	unsigned long vm_flags, const struct vm_special_mapping *spec)
3129 {
3130 	return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3131 					&special_mapping_vmops);
3132 }
3133 
3134 int install_special_mapping(struct mm_struct *mm,
3135 			    unsigned long addr, unsigned long len,
3136 			    unsigned long vm_flags, struct page **pages)
3137 {
3138 	struct vm_area_struct *vma = __install_special_mapping(
3139 		mm, addr, len, vm_flags, (void *)pages,
3140 		&legacy_special_mapping_vmops);
3141 
3142 	return PTR_ERR_OR_ZERO(vma);
3143 }
3144 
3145 static DEFINE_MUTEX(mm_all_locks_mutex);
3146 
3147 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3148 {
3149 	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3150 		/*
3151 		 * The LSB of head.next can't change from under us
3152 		 * because we hold the mm_all_locks_mutex.
3153 		 */
3154 		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3155 		/*
3156 		 * We can safely modify head.next after taking the
3157 		 * anon_vma->root->rwsem. If some other vma in this mm shares
3158 		 * the same anon_vma we won't take it again.
3159 		 *
3160 		 * No need of atomic instructions here, head.next
3161 		 * can't change from under us thanks to the
3162 		 * anon_vma->root->rwsem.
3163 		 */
3164 		if (__test_and_set_bit(0, (unsigned long *)
3165 				       &anon_vma->root->rb_root.rb_node))
3166 			BUG();
3167 	}
3168 }
3169 
3170 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3171 {
3172 	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3173 		/*
3174 		 * AS_MM_ALL_LOCKS can't change from under us because
3175 		 * we hold the mm_all_locks_mutex.
3176 		 *
3177 		 * Operations on ->flags have to be atomic because
3178 		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3179 		 * mm_all_locks_mutex, there may be other cpus
3180 		 * changing other bitflags in parallel to us.
3181 		 */
3182 		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3183 			BUG();
3184 		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3185 	}
3186 }
3187 
3188 /*
3189  * This operation locks against the VM for all pte/vma/mm related
3190  * operations that could ever happen on a certain mm. This includes
3191  * vmtruncate, try_to_unmap, and all page faults.
3192  *
3193  * The caller must take the mmap_sem in write mode before calling
3194  * mm_take_all_locks(). The caller isn't allowed to release the
3195  * mmap_sem until mm_drop_all_locks() returns.
3196  *
3197  * mmap_sem in write mode is required in order to block all operations
3198  * that could modify pagetables and free pages without need of
3199  * altering the vma layout. It's also needed in write mode to avoid new
3200  * anon_vmas to be associated with existing vmas.
3201  *
3202  * A single task can't take more than one mm_take_all_locks() in a row
3203  * or it would deadlock.
3204  *
3205  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3206  * mapping->flags avoid to take the same lock twice, if more than one
3207  * vma in this mm is backed by the same anon_vma or address_space.
3208  *
3209  * We take locks in following order, accordingly to comment at beginning
3210  * of mm/rmap.c:
3211  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3212  *     hugetlb mapping);
3213  *   - all i_mmap_rwsem locks;
3214  *   - all anon_vma->rwseml
3215  *
3216  * We can take all locks within these types randomly because the VM code
3217  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3218  * mm_all_locks_mutex.
3219  *
3220  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3221  * that may have to take thousand of locks.
3222  *
3223  * mm_take_all_locks() can fail if it's interrupted by signals.
3224  */
3225 int mm_take_all_locks(struct mm_struct *mm)
3226 {
3227 	struct vm_area_struct *vma;
3228 	struct anon_vma_chain *avc;
3229 
3230 	BUG_ON(down_read_trylock(&mm->mmap_sem));
3231 
3232 	mutex_lock(&mm_all_locks_mutex);
3233 
3234 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3235 		if (signal_pending(current))
3236 			goto out_unlock;
3237 		if (vma->vm_file && vma->vm_file->f_mapping &&
3238 				is_vm_hugetlb_page(vma))
3239 			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3240 	}
3241 
3242 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3243 		if (signal_pending(current))
3244 			goto out_unlock;
3245 		if (vma->vm_file && vma->vm_file->f_mapping &&
3246 				!is_vm_hugetlb_page(vma))
3247 			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3248 	}
3249 
3250 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3251 		if (signal_pending(current))
3252 			goto out_unlock;
3253 		if (vma->anon_vma)
3254 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3255 				vm_lock_anon_vma(mm, avc->anon_vma);
3256 	}
3257 
3258 	return 0;
3259 
3260 out_unlock:
3261 	mm_drop_all_locks(mm);
3262 	return -EINTR;
3263 }
3264 
3265 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3266 {
3267 	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3268 		/*
3269 		 * The LSB of head.next can't change to 0 from under
3270 		 * us because we hold the mm_all_locks_mutex.
3271 		 *
3272 		 * We must however clear the bitflag before unlocking
3273 		 * the vma so the users using the anon_vma->rb_root will
3274 		 * never see our bitflag.
3275 		 *
3276 		 * No need of atomic instructions here, head.next
3277 		 * can't change from under us until we release the
3278 		 * anon_vma->root->rwsem.
3279 		 */
3280 		if (!__test_and_clear_bit(0, (unsigned long *)
3281 					  &anon_vma->root->rb_root.rb_node))
3282 			BUG();
3283 		anon_vma_unlock_write(anon_vma);
3284 	}
3285 }
3286 
3287 static void vm_unlock_mapping(struct address_space *mapping)
3288 {
3289 	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3290 		/*
3291 		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3292 		 * because we hold the mm_all_locks_mutex.
3293 		 */
3294 		i_mmap_unlock_write(mapping);
3295 		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3296 					&mapping->flags))
3297 			BUG();
3298 	}
3299 }
3300 
3301 /*
3302  * The mmap_sem cannot be released by the caller until
3303  * mm_drop_all_locks() returns.
3304  */
3305 void mm_drop_all_locks(struct mm_struct *mm)
3306 {
3307 	struct vm_area_struct *vma;
3308 	struct anon_vma_chain *avc;
3309 
3310 	BUG_ON(down_read_trylock(&mm->mmap_sem));
3311 	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3312 
3313 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3314 		if (vma->anon_vma)
3315 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3316 				vm_unlock_anon_vma(avc->anon_vma);
3317 		if (vma->vm_file && vma->vm_file->f_mapping)
3318 			vm_unlock_mapping(vma->vm_file->f_mapping);
3319 	}
3320 
3321 	mutex_unlock(&mm_all_locks_mutex);
3322 }
3323 
3324 /*
3325  * initialise the VMA slab
3326  */
3327 void __init mmap_init(void)
3328 {
3329 	int ret;
3330 
3331 	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3332 	VM_BUG_ON(ret);
3333 }
3334 
3335 /*
3336  * Initialise sysctl_user_reserve_kbytes.
3337  *
3338  * This is intended to prevent a user from starting a single memory hogging
3339  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3340  * mode.
3341  *
3342  * The default value is min(3% of free memory, 128MB)
3343  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3344  */
3345 static int init_user_reserve(void)
3346 {
3347 	unsigned long free_kbytes;
3348 
3349 	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3350 
3351 	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3352 	return 0;
3353 }
3354 subsys_initcall(init_user_reserve);
3355 
3356 /*
3357  * Initialise sysctl_admin_reserve_kbytes.
3358  *
3359  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3360  * to log in and kill a memory hogging process.
3361  *
3362  * Systems with more than 256MB will reserve 8MB, enough to recover
3363  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3364  * only reserve 3% of free pages by default.
3365  */
3366 static int init_admin_reserve(void)
3367 {
3368 	unsigned long free_kbytes;
3369 
3370 	free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3371 
3372 	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3373 	return 0;
3374 }
3375 subsys_initcall(init_admin_reserve);
3376 
3377 /*
3378  * Reinititalise user and admin reserves if memory is added or removed.
3379  *
3380  * The default user reserve max is 128MB, and the default max for the
3381  * admin reserve is 8MB. These are usually, but not always, enough to
3382  * enable recovery from a memory hogging process using login/sshd, a shell,
3383  * and tools like top. It may make sense to increase or even disable the
3384  * reserve depending on the existence of swap or variations in the recovery
3385  * tools. So, the admin may have changed them.
3386  *
3387  * If memory is added and the reserves have been eliminated or increased above
3388  * the default max, then we'll trust the admin.
3389  *
3390  * If memory is removed and there isn't enough free memory, then we
3391  * need to reset the reserves.
3392  *
3393  * Otherwise keep the reserve set by the admin.
3394  */
3395 static int reserve_mem_notifier(struct notifier_block *nb,
3396 			     unsigned long action, void *data)
3397 {
3398 	unsigned long tmp, free_kbytes;
3399 
3400 	switch (action) {
3401 	case MEM_ONLINE:
3402 		/* Default max is 128MB. Leave alone if modified by operator. */
3403 		tmp = sysctl_user_reserve_kbytes;
3404 		if (0 < tmp && tmp < (1UL << 17))
3405 			init_user_reserve();
3406 
3407 		/* Default max is 8MB.  Leave alone if modified by operator. */
3408 		tmp = sysctl_admin_reserve_kbytes;
3409 		if (0 < tmp && tmp < (1UL << 13))
3410 			init_admin_reserve();
3411 
3412 		break;
3413 	case MEM_OFFLINE:
3414 		free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3415 
3416 		if (sysctl_user_reserve_kbytes > free_kbytes) {
3417 			init_user_reserve();
3418 			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3419 				sysctl_user_reserve_kbytes);
3420 		}
3421 
3422 		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3423 			init_admin_reserve();
3424 			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3425 				sysctl_admin_reserve_kbytes);
3426 		}
3427 		break;
3428 	default:
3429 		break;
3430 	}
3431 	return NOTIFY_OK;
3432 }
3433 
3434 static struct notifier_block reserve_mem_nb = {
3435 	.notifier_call = reserve_mem_notifier,
3436 };
3437 
3438 static int __meminit init_reserve_notifier(void)
3439 {
3440 	if (register_hotmemory_notifier(&reserve_mem_nb))
3441 		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3442 
3443 	return 0;
3444 }
3445 subsys_initcall(init_reserve_notifier);
3446