xref: /openbmc/linux/mm/mmap.c (revision 3c881e05c814c970e4f9577446a9d3461d134607)
1  // SPDX-License-Identifier: GPL-2.0-only
2  /*
3   * mm/mmap.c
4   *
5   * Written by obz.
6   *
7   * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
8   */
9  
10  #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11  
12  #include <linux/kernel.h>
13  #include <linux/slab.h>
14  #include <linux/backing-dev.h>
15  #include <linux/mm.h>
16  #include <linux/vmacache.h>
17  #include <linux/shm.h>
18  #include <linux/mman.h>
19  #include <linux/pagemap.h>
20  #include <linux/swap.h>
21  #include <linux/syscalls.h>
22  #include <linux/capability.h>
23  #include <linux/init.h>
24  #include <linux/file.h>
25  #include <linux/fs.h>
26  #include <linux/personality.h>
27  #include <linux/security.h>
28  #include <linux/hugetlb.h>
29  #include <linux/shmem_fs.h>
30  #include <linux/profile.h>
31  #include <linux/export.h>
32  #include <linux/mount.h>
33  #include <linux/mempolicy.h>
34  #include <linux/rmap.h>
35  #include <linux/mmu_notifier.h>
36  #include <linux/mmdebug.h>
37  #include <linux/perf_event.h>
38  #include <linux/audit.h>
39  #include <linux/khugepaged.h>
40  #include <linux/uprobes.h>
41  #include <linux/rbtree_augmented.h>
42  #include <linux/notifier.h>
43  #include <linux/memory.h>
44  #include <linux/printk.h>
45  #include <linux/userfaultfd_k.h>
46  #include <linux/moduleparam.h>
47  #include <linux/pkeys.h>
48  #include <linux/oom.h>
49  #include <linux/sched/mm.h>
50  
51  #include <linux/uaccess.h>
52  #include <asm/cacheflush.h>
53  #include <asm/tlb.h>
54  #include <asm/mmu_context.h>
55  
56  #define CREATE_TRACE_POINTS
57  #include <trace/events/mmap.h>
58  
59  #include "internal.h"
60  
61  #ifndef arch_mmap_check
62  #define arch_mmap_check(addr, len, flags)	(0)
63  #endif
64  
65  #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66  const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67  const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68  int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69  #endif
70  #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71  const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72  const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73  int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
74  #endif
75  
76  static bool ignore_rlimit_data;
77  core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78  
79  static void unmap_region(struct mm_struct *mm,
80  		struct vm_area_struct *vma, struct vm_area_struct *prev,
81  		unsigned long start, unsigned long end);
82  
83  /* description of effects of mapping type and prot in current implementation.
84   * this is due to the limited x86 page protection hardware.  The expected
85   * behavior is in parens:
86   *
87   * map_type	prot
88   *		PROT_NONE	PROT_READ	PROT_WRITE	PROT_EXEC
89   * MAP_SHARED	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
90   *		w: (no) no	w: (no) no	w: (yes) yes	w: (no) no
91   *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
92   *
93   * MAP_PRIVATE	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
94   *		w: (no) no	w: (no) no	w: (copy) copy	w: (no) no
95   *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
96   *
97   * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
98   * MAP_PRIVATE (with Enhanced PAN supported):
99   *								r: (no) no
100   *								w: (no) no
101   *								x: (yes) yes
102   */
103  pgprot_t protection_map[16] __ro_after_init = {
104  	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
105  	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
106  };
107  
108  #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
109  static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
110  {
111  	return prot;
112  }
113  #endif
114  
115  pgprot_t vm_get_page_prot(unsigned long vm_flags)
116  {
117  	pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
118  				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
119  			pgprot_val(arch_vm_get_page_prot(vm_flags)));
120  
121  	return arch_filter_pgprot(ret);
122  }
123  EXPORT_SYMBOL(vm_get_page_prot);
124  
125  static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
126  {
127  	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
128  }
129  
130  /* Update vma->vm_page_prot to reflect vma->vm_flags. */
131  void vma_set_page_prot(struct vm_area_struct *vma)
132  {
133  	unsigned long vm_flags = vma->vm_flags;
134  	pgprot_t vm_page_prot;
135  
136  	vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
137  	if (vma_wants_writenotify(vma, vm_page_prot)) {
138  		vm_flags &= ~VM_SHARED;
139  		vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
140  	}
141  	/* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
142  	WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
143  }
144  
145  /*
146   * Requires inode->i_mapping->i_mmap_rwsem
147   */
148  static void __remove_shared_vm_struct(struct vm_area_struct *vma,
149  		struct file *file, struct address_space *mapping)
150  {
151  	if (vma->vm_flags & VM_DENYWRITE)
152  		allow_write_access(file);
153  	if (vma->vm_flags & VM_SHARED)
154  		mapping_unmap_writable(mapping);
155  
156  	flush_dcache_mmap_lock(mapping);
157  	vma_interval_tree_remove(vma, &mapping->i_mmap);
158  	flush_dcache_mmap_unlock(mapping);
159  }
160  
161  /*
162   * Unlink a file-based vm structure from its interval tree, to hide
163   * vma from rmap and vmtruncate before freeing its page tables.
164   */
165  void unlink_file_vma(struct vm_area_struct *vma)
166  {
167  	struct file *file = vma->vm_file;
168  
169  	if (file) {
170  		struct address_space *mapping = file->f_mapping;
171  		i_mmap_lock_write(mapping);
172  		__remove_shared_vm_struct(vma, file, mapping);
173  		i_mmap_unlock_write(mapping);
174  	}
175  }
176  
177  /*
178   * Close a vm structure and free it, returning the next.
179   */
180  static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
181  {
182  	struct vm_area_struct *next = vma->vm_next;
183  
184  	might_sleep();
185  	if (vma->vm_ops && vma->vm_ops->close)
186  		vma->vm_ops->close(vma);
187  	if (vma->vm_file)
188  		fput(vma->vm_file);
189  	mpol_put(vma_policy(vma));
190  	vm_area_free(vma);
191  	return next;
192  }
193  
194  static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
195  		struct list_head *uf);
196  SYSCALL_DEFINE1(brk, unsigned long, brk)
197  {
198  	unsigned long newbrk, oldbrk, origbrk;
199  	struct mm_struct *mm = current->mm;
200  	struct vm_area_struct *next;
201  	unsigned long min_brk;
202  	bool populate;
203  	bool downgraded = false;
204  	LIST_HEAD(uf);
205  
206  	if (mmap_write_lock_killable(mm))
207  		return -EINTR;
208  
209  	origbrk = mm->brk;
210  
211  #ifdef CONFIG_COMPAT_BRK
212  	/*
213  	 * CONFIG_COMPAT_BRK can still be overridden by setting
214  	 * randomize_va_space to 2, which will still cause mm->start_brk
215  	 * to be arbitrarily shifted
216  	 */
217  	if (current->brk_randomized)
218  		min_brk = mm->start_brk;
219  	else
220  		min_brk = mm->end_data;
221  #else
222  	min_brk = mm->start_brk;
223  #endif
224  	if (brk < min_brk)
225  		goto out;
226  
227  	/*
228  	 * Check against rlimit here. If this check is done later after the test
229  	 * of oldbrk with newbrk then it can escape the test and let the data
230  	 * segment grow beyond its set limit the in case where the limit is
231  	 * not page aligned -Ram Gupta
232  	 */
233  	if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
234  			      mm->end_data, mm->start_data))
235  		goto out;
236  
237  	newbrk = PAGE_ALIGN(brk);
238  	oldbrk = PAGE_ALIGN(mm->brk);
239  	if (oldbrk == newbrk) {
240  		mm->brk = brk;
241  		goto success;
242  	}
243  
244  	/*
245  	 * Always allow shrinking brk.
246  	 * __do_munmap() may downgrade mmap_lock to read.
247  	 */
248  	if (brk <= mm->brk) {
249  		int ret;
250  
251  		/*
252  		 * mm->brk must to be protected by write mmap_lock so update it
253  		 * before downgrading mmap_lock. When __do_munmap() fails,
254  		 * mm->brk will be restored from origbrk.
255  		 */
256  		mm->brk = brk;
257  		ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
258  		if (ret < 0) {
259  			mm->brk = origbrk;
260  			goto out;
261  		} else if (ret == 1) {
262  			downgraded = true;
263  		}
264  		goto success;
265  	}
266  
267  	/* Check against existing mmap mappings. */
268  	next = find_vma(mm, oldbrk);
269  	if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
270  		goto out;
271  
272  	/* Ok, looks good - let it rip. */
273  	if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
274  		goto out;
275  	mm->brk = brk;
276  
277  success:
278  	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
279  	if (downgraded)
280  		mmap_read_unlock(mm);
281  	else
282  		mmap_write_unlock(mm);
283  	userfaultfd_unmap_complete(mm, &uf);
284  	if (populate)
285  		mm_populate(oldbrk, newbrk - oldbrk);
286  	return brk;
287  
288  out:
289  	mmap_write_unlock(mm);
290  	return origbrk;
291  }
292  
293  static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
294  {
295  	unsigned long gap, prev_end;
296  
297  	/*
298  	 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
299  	 * allow two stack_guard_gaps between them here, and when choosing
300  	 * an unmapped area; whereas when expanding we only require one.
301  	 * That's a little inconsistent, but keeps the code here simpler.
302  	 */
303  	gap = vm_start_gap(vma);
304  	if (vma->vm_prev) {
305  		prev_end = vm_end_gap(vma->vm_prev);
306  		if (gap > prev_end)
307  			gap -= prev_end;
308  		else
309  			gap = 0;
310  	}
311  	return gap;
312  }
313  
314  #ifdef CONFIG_DEBUG_VM_RB
315  static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
316  {
317  	unsigned long max = vma_compute_gap(vma), subtree_gap;
318  	if (vma->vm_rb.rb_left) {
319  		subtree_gap = rb_entry(vma->vm_rb.rb_left,
320  				struct vm_area_struct, vm_rb)->rb_subtree_gap;
321  		if (subtree_gap > max)
322  			max = subtree_gap;
323  	}
324  	if (vma->vm_rb.rb_right) {
325  		subtree_gap = rb_entry(vma->vm_rb.rb_right,
326  				struct vm_area_struct, vm_rb)->rb_subtree_gap;
327  		if (subtree_gap > max)
328  			max = subtree_gap;
329  	}
330  	return max;
331  }
332  
333  static int browse_rb(struct mm_struct *mm)
334  {
335  	struct rb_root *root = &mm->mm_rb;
336  	int i = 0, j, bug = 0;
337  	struct rb_node *nd, *pn = NULL;
338  	unsigned long prev = 0, pend = 0;
339  
340  	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
341  		struct vm_area_struct *vma;
342  		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
343  		if (vma->vm_start < prev) {
344  			pr_emerg("vm_start %lx < prev %lx\n",
345  				  vma->vm_start, prev);
346  			bug = 1;
347  		}
348  		if (vma->vm_start < pend) {
349  			pr_emerg("vm_start %lx < pend %lx\n",
350  				  vma->vm_start, pend);
351  			bug = 1;
352  		}
353  		if (vma->vm_start > vma->vm_end) {
354  			pr_emerg("vm_start %lx > vm_end %lx\n",
355  				  vma->vm_start, vma->vm_end);
356  			bug = 1;
357  		}
358  		spin_lock(&mm->page_table_lock);
359  		if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
360  			pr_emerg("free gap %lx, correct %lx\n",
361  			       vma->rb_subtree_gap,
362  			       vma_compute_subtree_gap(vma));
363  			bug = 1;
364  		}
365  		spin_unlock(&mm->page_table_lock);
366  		i++;
367  		pn = nd;
368  		prev = vma->vm_start;
369  		pend = vma->vm_end;
370  	}
371  	j = 0;
372  	for (nd = pn; nd; nd = rb_prev(nd))
373  		j++;
374  	if (i != j) {
375  		pr_emerg("backwards %d, forwards %d\n", j, i);
376  		bug = 1;
377  	}
378  	return bug ? -1 : i;
379  }
380  
381  static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
382  {
383  	struct rb_node *nd;
384  
385  	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
386  		struct vm_area_struct *vma;
387  		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
388  		VM_BUG_ON_VMA(vma != ignore &&
389  			vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
390  			vma);
391  	}
392  }
393  
394  static void validate_mm(struct mm_struct *mm)
395  {
396  	int bug = 0;
397  	int i = 0;
398  	unsigned long highest_address = 0;
399  	struct vm_area_struct *vma = mm->mmap;
400  
401  	while (vma) {
402  		struct anon_vma *anon_vma = vma->anon_vma;
403  		struct anon_vma_chain *avc;
404  
405  		if (anon_vma) {
406  			anon_vma_lock_read(anon_vma);
407  			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
408  				anon_vma_interval_tree_verify(avc);
409  			anon_vma_unlock_read(anon_vma);
410  		}
411  
412  		highest_address = vm_end_gap(vma);
413  		vma = vma->vm_next;
414  		i++;
415  	}
416  	if (i != mm->map_count) {
417  		pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
418  		bug = 1;
419  	}
420  	if (highest_address != mm->highest_vm_end) {
421  		pr_emerg("mm->highest_vm_end %lx, found %lx\n",
422  			  mm->highest_vm_end, highest_address);
423  		bug = 1;
424  	}
425  	i = browse_rb(mm);
426  	if (i != mm->map_count) {
427  		if (i != -1)
428  			pr_emerg("map_count %d rb %d\n", mm->map_count, i);
429  		bug = 1;
430  	}
431  	VM_BUG_ON_MM(bug, mm);
432  }
433  #else
434  #define validate_mm_rb(root, ignore) do { } while (0)
435  #define validate_mm(mm) do { } while (0)
436  #endif
437  
438  RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
439  			 struct vm_area_struct, vm_rb,
440  			 unsigned long, rb_subtree_gap, vma_compute_gap)
441  
442  /*
443   * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
444   * vma->vm_prev->vm_end values changed, without modifying the vma's position
445   * in the rbtree.
446   */
447  static void vma_gap_update(struct vm_area_struct *vma)
448  {
449  	/*
450  	 * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
451  	 * a callback function that does exactly what we want.
452  	 */
453  	vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
454  }
455  
456  static inline void vma_rb_insert(struct vm_area_struct *vma,
457  				 struct rb_root *root)
458  {
459  	/* All rb_subtree_gap values must be consistent prior to insertion */
460  	validate_mm_rb(root, NULL);
461  
462  	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
463  }
464  
465  static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
466  {
467  	/*
468  	 * Note rb_erase_augmented is a fairly large inline function,
469  	 * so make sure we instantiate it only once with our desired
470  	 * augmented rbtree callbacks.
471  	 */
472  	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
473  }
474  
475  static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
476  						struct rb_root *root,
477  						struct vm_area_struct *ignore)
478  {
479  	/*
480  	 * All rb_subtree_gap values must be consistent prior to erase,
481  	 * with the possible exception of
482  	 *
483  	 * a. the "next" vma being erased if next->vm_start was reduced in
484  	 *    __vma_adjust() -> __vma_unlink()
485  	 * b. the vma being erased in detach_vmas_to_be_unmapped() ->
486  	 *    vma_rb_erase()
487  	 */
488  	validate_mm_rb(root, ignore);
489  
490  	__vma_rb_erase(vma, root);
491  }
492  
493  static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
494  					 struct rb_root *root)
495  {
496  	vma_rb_erase_ignore(vma, root, vma);
497  }
498  
499  /*
500   * vma has some anon_vma assigned, and is already inserted on that
501   * anon_vma's interval trees.
502   *
503   * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
504   * vma must be removed from the anon_vma's interval trees using
505   * anon_vma_interval_tree_pre_update_vma().
506   *
507   * After the update, the vma will be reinserted using
508   * anon_vma_interval_tree_post_update_vma().
509   *
510   * The entire update must be protected by exclusive mmap_lock and by
511   * the root anon_vma's mutex.
512   */
513  static inline void
514  anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
515  {
516  	struct anon_vma_chain *avc;
517  
518  	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
519  		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
520  }
521  
522  static inline void
523  anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
524  {
525  	struct anon_vma_chain *avc;
526  
527  	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
528  		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
529  }
530  
531  static int find_vma_links(struct mm_struct *mm, unsigned long addr,
532  		unsigned long end, struct vm_area_struct **pprev,
533  		struct rb_node ***rb_link, struct rb_node **rb_parent)
534  {
535  	struct rb_node **__rb_link, *__rb_parent, *rb_prev;
536  
537  	__rb_link = &mm->mm_rb.rb_node;
538  	rb_prev = __rb_parent = NULL;
539  
540  	while (*__rb_link) {
541  		struct vm_area_struct *vma_tmp;
542  
543  		__rb_parent = *__rb_link;
544  		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
545  
546  		if (vma_tmp->vm_end > addr) {
547  			/* Fail if an existing vma overlaps the area */
548  			if (vma_tmp->vm_start < end)
549  				return -ENOMEM;
550  			__rb_link = &__rb_parent->rb_left;
551  		} else {
552  			rb_prev = __rb_parent;
553  			__rb_link = &__rb_parent->rb_right;
554  		}
555  	}
556  
557  	*pprev = NULL;
558  	if (rb_prev)
559  		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
560  	*rb_link = __rb_link;
561  	*rb_parent = __rb_parent;
562  	return 0;
563  }
564  
565  /*
566   * vma_next() - Get the next VMA.
567   * @mm: The mm_struct.
568   * @vma: The current vma.
569   *
570   * If @vma is NULL, return the first vma in the mm.
571   *
572   * Returns: The next VMA after @vma.
573   */
574  static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
575  					 struct vm_area_struct *vma)
576  {
577  	if (!vma)
578  		return mm->mmap;
579  
580  	return vma->vm_next;
581  }
582  
583  /*
584   * munmap_vma_range() - munmap VMAs that overlap a range.
585   * @mm: The mm struct
586   * @start: The start of the range.
587   * @len: The length of the range.
588   * @pprev: pointer to the pointer that will be set to previous vm_area_struct
589   * @rb_link: the rb_node
590   * @rb_parent: the parent rb_node
591   *
592   * Find all the vm_area_struct that overlap from @start to
593   * @end and munmap them.  Set @pprev to the previous vm_area_struct.
594   *
595   * Returns: -ENOMEM on munmap failure or 0 on success.
596   */
597  static inline int
598  munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
599  		 struct vm_area_struct **pprev, struct rb_node ***link,
600  		 struct rb_node **parent, struct list_head *uf)
601  {
602  
603  	while (find_vma_links(mm, start, start + len, pprev, link, parent))
604  		if (do_munmap(mm, start, len, uf))
605  			return -ENOMEM;
606  
607  	return 0;
608  }
609  static unsigned long count_vma_pages_range(struct mm_struct *mm,
610  		unsigned long addr, unsigned long end)
611  {
612  	unsigned long nr_pages = 0;
613  	struct vm_area_struct *vma;
614  
615  	/* Find first overlapping mapping */
616  	vma = find_vma_intersection(mm, addr, end);
617  	if (!vma)
618  		return 0;
619  
620  	nr_pages = (min(end, vma->vm_end) -
621  		max(addr, vma->vm_start)) >> PAGE_SHIFT;
622  
623  	/* Iterate over the rest of the overlaps */
624  	for (vma = vma->vm_next; vma; vma = vma->vm_next) {
625  		unsigned long overlap_len;
626  
627  		if (vma->vm_start > end)
628  			break;
629  
630  		overlap_len = min(end, vma->vm_end) - vma->vm_start;
631  		nr_pages += overlap_len >> PAGE_SHIFT;
632  	}
633  
634  	return nr_pages;
635  }
636  
637  void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
638  		struct rb_node **rb_link, struct rb_node *rb_parent)
639  {
640  	/* Update tracking information for the gap following the new vma. */
641  	if (vma->vm_next)
642  		vma_gap_update(vma->vm_next);
643  	else
644  		mm->highest_vm_end = vm_end_gap(vma);
645  
646  	/*
647  	 * vma->vm_prev wasn't known when we followed the rbtree to find the
648  	 * correct insertion point for that vma. As a result, we could not
649  	 * update the vma vm_rb parents rb_subtree_gap values on the way down.
650  	 * So, we first insert the vma with a zero rb_subtree_gap value
651  	 * (to be consistent with what we did on the way down), and then
652  	 * immediately update the gap to the correct value. Finally we
653  	 * rebalance the rbtree after all augmented values have been set.
654  	 */
655  	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
656  	vma->rb_subtree_gap = 0;
657  	vma_gap_update(vma);
658  	vma_rb_insert(vma, &mm->mm_rb);
659  }
660  
661  static void __vma_link_file(struct vm_area_struct *vma)
662  {
663  	struct file *file;
664  
665  	file = vma->vm_file;
666  	if (file) {
667  		struct address_space *mapping = file->f_mapping;
668  
669  		if (vma->vm_flags & VM_DENYWRITE)
670  			put_write_access(file_inode(file));
671  		if (vma->vm_flags & VM_SHARED)
672  			mapping_allow_writable(mapping);
673  
674  		flush_dcache_mmap_lock(mapping);
675  		vma_interval_tree_insert(vma, &mapping->i_mmap);
676  		flush_dcache_mmap_unlock(mapping);
677  	}
678  }
679  
680  static void
681  __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
682  	struct vm_area_struct *prev, struct rb_node **rb_link,
683  	struct rb_node *rb_parent)
684  {
685  	__vma_link_list(mm, vma, prev);
686  	__vma_link_rb(mm, vma, rb_link, rb_parent);
687  }
688  
689  static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
690  			struct vm_area_struct *prev, struct rb_node **rb_link,
691  			struct rb_node *rb_parent)
692  {
693  	struct address_space *mapping = NULL;
694  
695  	if (vma->vm_file) {
696  		mapping = vma->vm_file->f_mapping;
697  		i_mmap_lock_write(mapping);
698  	}
699  
700  	__vma_link(mm, vma, prev, rb_link, rb_parent);
701  	__vma_link_file(vma);
702  
703  	if (mapping)
704  		i_mmap_unlock_write(mapping);
705  
706  	mm->map_count++;
707  	validate_mm(mm);
708  }
709  
710  /*
711   * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
712   * mm's list and rbtree.  It has already been inserted into the interval tree.
713   */
714  static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
715  {
716  	struct vm_area_struct *prev;
717  	struct rb_node **rb_link, *rb_parent;
718  
719  	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
720  			   &prev, &rb_link, &rb_parent))
721  		BUG();
722  	__vma_link(mm, vma, prev, rb_link, rb_parent);
723  	mm->map_count++;
724  }
725  
726  static __always_inline void __vma_unlink(struct mm_struct *mm,
727  						struct vm_area_struct *vma,
728  						struct vm_area_struct *ignore)
729  {
730  	vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
731  	__vma_unlink_list(mm, vma);
732  	/* Kill the cache */
733  	vmacache_invalidate(mm);
734  }
735  
736  /*
737   * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
738   * is already present in an i_mmap tree without adjusting the tree.
739   * The following helper function should be used when such adjustments
740   * are necessary.  The "insert" vma (if any) is to be inserted
741   * before we drop the necessary locks.
742   */
743  int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
744  	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
745  	struct vm_area_struct *expand)
746  {
747  	struct mm_struct *mm = vma->vm_mm;
748  	struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
749  	struct address_space *mapping = NULL;
750  	struct rb_root_cached *root = NULL;
751  	struct anon_vma *anon_vma = NULL;
752  	struct file *file = vma->vm_file;
753  	bool start_changed = false, end_changed = false;
754  	long adjust_next = 0;
755  	int remove_next = 0;
756  
757  	if (next && !insert) {
758  		struct vm_area_struct *exporter = NULL, *importer = NULL;
759  
760  		if (end >= next->vm_end) {
761  			/*
762  			 * vma expands, overlapping all the next, and
763  			 * perhaps the one after too (mprotect case 6).
764  			 * The only other cases that gets here are
765  			 * case 1, case 7 and case 8.
766  			 */
767  			if (next == expand) {
768  				/*
769  				 * The only case where we don't expand "vma"
770  				 * and we expand "next" instead is case 8.
771  				 */
772  				VM_WARN_ON(end != next->vm_end);
773  				/*
774  				 * remove_next == 3 means we're
775  				 * removing "vma" and that to do so we
776  				 * swapped "vma" and "next".
777  				 */
778  				remove_next = 3;
779  				VM_WARN_ON(file != next->vm_file);
780  				swap(vma, next);
781  			} else {
782  				VM_WARN_ON(expand != vma);
783  				/*
784  				 * case 1, 6, 7, remove_next == 2 is case 6,
785  				 * remove_next == 1 is case 1 or 7.
786  				 */
787  				remove_next = 1 + (end > next->vm_end);
788  				VM_WARN_ON(remove_next == 2 &&
789  					   end != next->vm_next->vm_end);
790  				/* trim end to next, for case 6 first pass */
791  				end = next->vm_end;
792  			}
793  
794  			exporter = next;
795  			importer = vma;
796  
797  			/*
798  			 * If next doesn't have anon_vma, import from vma after
799  			 * next, if the vma overlaps with it.
800  			 */
801  			if (remove_next == 2 && !next->anon_vma)
802  				exporter = next->vm_next;
803  
804  		} else if (end > next->vm_start) {
805  			/*
806  			 * vma expands, overlapping part of the next:
807  			 * mprotect case 5 shifting the boundary up.
808  			 */
809  			adjust_next = (end - next->vm_start);
810  			exporter = next;
811  			importer = vma;
812  			VM_WARN_ON(expand != importer);
813  		} else if (end < vma->vm_end) {
814  			/*
815  			 * vma shrinks, and !insert tells it's not
816  			 * split_vma inserting another: so it must be
817  			 * mprotect case 4 shifting the boundary down.
818  			 */
819  			adjust_next = -(vma->vm_end - end);
820  			exporter = vma;
821  			importer = next;
822  			VM_WARN_ON(expand != importer);
823  		}
824  
825  		/*
826  		 * Easily overlooked: when mprotect shifts the boundary,
827  		 * make sure the expanding vma has anon_vma set if the
828  		 * shrinking vma had, to cover any anon pages imported.
829  		 */
830  		if (exporter && exporter->anon_vma && !importer->anon_vma) {
831  			int error;
832  
833  			importer->anon_vma = exporter->anon_vma;
834  			error = anon_vma_clone(importer, exporter);
835  			if (error)
836  				return error;
837  		}
838  	}
839  again:
840  	vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
841  
842  	if (file) {
843  		mapping = file->f_mapping;
844  		root = &mapping->i_mmap;
845  		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
846  
847  		if (adjust_next)
848  			uprobe_munmap(next, next->vm_start, next->vm_end);
849  
850  		i_mmap_lock_write(mapping);
851  		if (insert) {
852  			/*
853  			 * Put into interval tree now, so instantiated pages
854  			 * are visible to arm/parisc __flush_dcache_page
855  			 * throughout; but we cannot insert into address
856  			 * space until vma start or end is updated.
857  			 */
858  			__vma_link_file(insert);
859  		}
860  	}
861  
862  	anon_vma = vma->anon_vma;
863  	if (!anon_vma && adjust_next)
864  		anon_vma = next->anon_vma;
865  	if (anon_vma) {
866  		VM_WARN_ON(adjust_next && next->anon_vma &&
867  			   anon_vma != next->anon_vma);
868  		anon_vma_lock_write(anon_vma);
869  		anon_vma_interval_tree_pre_update_vma(vma);
870  		if (adjust_next)
871  			anon_vma_interval_tree_pre_update_vma(next);
872  	}
873  
874  	if (file) {
875  		flush_dcache_mmap_lock(mapping);
876  		vma_interval_tree_remove(vma, root);
877  		if (adjust_next)
878  			vma_interval_tree_remove(next, root);
879  	}
880  
881  	if (start != vma->vm_start) {
882  		vma->vm_start = start;
883  		start_changed = true;
884  	}
885  	if (end != vma->vm_end) {
886  		vma->vm_end = end;
887  		end_changed = true;
888  	}
889  	vma->vm_pgoff = pgoff;
890  	if (adjust_next) {
891  		next->vm_start += adjust_next;
892  		next->vm_pgoff += adjust_next >> PAGE_SHIFT;
893  	}
894  
895  	if (file) {
896  		if (adjust_next)
897  			vma_interval_tree_insert(next, root);
898  		vma_interval_tree_insert(vma, root);
899  		flush_dcache_mmap_unlock(mapping);
900  	}
901  
902  	if (remove_next) {
903  		/*
904  		 * vma_merge has merged next into vma, and needs
905  		 * us to remove next before dropping the locks.
906  		 */
907  		if (remove_next != 3)
908  			__vma_unlink(mm, next, next);
909  		else
910  			/*
911  			 * vma is not before next if they've been
912  			 * swapped.
913  			 *
914  			 * pre-swap() next->vm_start was reduced so
915  			 * tell validate_mm_rb to ignore pre-swap()
916  			 * "next" (which is stored in post-swap()
917  			 * "vma").
918  			 */
919  			__vma_unlink(mm, next, vma);
920  		if (file)
921  			__remove_shared_vm_struct(next, file, mapping);
922  	} else if (insert) {
923  		/*
924  		 * split_vma has split insert from vma, and needs
925  		 * us to insert it before dropping the locks
926  		 * (it may either follow vma or precede it).
927  		 */
928  		__insert_vm_struct(mm, insert);
929  	} else {
930  		if (start_changed)
931  			vma_gap_update(vma);
932  		if (end_changed) {
933  			if (!next)
934  				mm->highest_vm_end = vm_end_gap(vma);
935  			else if (!adjust_next)
936  				vma_gap_update(next);
937  		}
938  	}
939  
940  	if (anon_vma) {
941  		anon_vma_interval_tree_post_update_vma(vma);
942  		if (adjust_next)
943  			anon_vma_interval_tree_post_update_vma(next);
944  		anon_vma_unlock_write(anon_vma);
945  	}
946  
947  	if (file) {
948  		i_mmap_unlock_write(mapping);
949  		uprobe_mmap(vma);
950  
951  		if (adjust_next)
952  			uprobe_mmap(next);
953  	}
954  
955  	if (remove_next) {
956  		if (file) {
957  			uprobe_munmap(next, next->vm_start, next->vm_end);
958  			fput(file);
959  		}
960  		if (next->anon_vma)
961  			anon_vma_merge(vma, next);
962  		mm->map_count--;
963  		mpol_put(vma_policy(next));
964  		vm_area_free(next);
965  		/*
966  		 * In mprotect's case 6 (see comments on vma_merge),
967  		 * we must remove another next too. It would clutter
968  		 * up the code too much to do both in one go.
969  		 */
970  		if (remove_next != 3) {
971  			/*
972  			 * If "next" was removed and vma->vm_end was
973  			 * expanded (up) over it, in turn
974  			 * "next->vm_prev->vm_end" changed and the
975  			 * "vma->vm_next" gap must be updated.
976  			 */
977  			next = vma->vm_next;
978  		} else {
979  			/*
980  			 * For the scope of the comment "next" and
981  			 * "vma" considered pre-swap(): if "vma" was
982  			 * removed, next->vm_start was expanded (down)
983  			 * over it and the "next" gap must be updated.
984  			 * Because of the swap() the post-swap() "vma"
985  			 * actually points to pre-swap() "next"
986  			 * (post-swap() "next" as opposed is now a
987  			 * dangling pointer).
988  			 */
989  			next = vma;
990  		}
991  		if (remove_next == 2) {
992  			remove_next = 1;
993  			end = next->vm_end;
994  			goto again;
995  		}
996  		else if (next)
997  			vma_gap_update(next);
998  		else {
999  			/*
1000  			 * If remove_next == 2 we obviously can't
1001  			 * reach this path.
1002  			 *
1003  			 * If remove_next == 3 we can't reach this
1004  			 * path because pre-swap() next is always not
1005  			 * NULL. pre-swap() "next" is not being
1006  			 * removed and its next->vm_end is not altered
1007  			 * (and furthermore "end" already matches
1008  			 * next->vm_end in remove_next == 3).
1009  			 *
1010  			 * We reach this only in the remove_next == 1
1011  			 * case if the "next" vma that was removed was
1012  			 * the highest vma of the mm. However in such
1013  			 * case next->vm_end == "end" and the extended
1014  			 * "vma" has vma->vm_end == next->vm_end so
1015  			 * mm->highest_vm_end doesn't need any update
1016  			 * in remove_next == 1 case.
1017  			 */
1018  			VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1019  		}
1020  	}
1021  	if (insert && file)
1022  		uprobe_mmap(insert);
1023  
1024  	validate_mm(mm);
1025  
1026  	return 0;
1027  }
1028  
1029  /*
1030   * If the vma has a ->close operation then the driver probably needs to release
1031   * per-vma resources, so we don't attempt to merge those.
1032   */
1033  static inline int is_mergeable_vma(struct vm_area_struct *vma,
1034  				struct file *file, unsigned long vm_flags,
1035  				struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1036  {
1037  	/*
1038  	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
1039  	 * match the flags but dirty bit -- the caller should mark
1040  	 * merged VMA as dirty. If dirty bit won't be excluded from
1041  	 * comparison, we increase pressure on the memory system forcing
1042  	 * the kernel to generate new VMAs when old one could be
1043  	 * extended instead.
1044  	 */
1045  	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1046  		return 0;
1047  	if (vma->vm_file != file)
1048  		return 0;
1049  	if (vma->vm_ops && vma->vm_ops->close)
1050  		return 0;
1051  	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1052  		return 0;
1053  	return 1;
1054  }
1055  
1056  static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1057  					struct anon_vma *anon_vma2,
1058  					struct vm_area_struct *vma)
1059  {
1060  	/*
1061  	 * The list_is_singular() test is to avoid merging VMA cloned from
1062  	 * parents. This can improve scalability caused by anon_vma lock.
1063  	 */
1064  	if ((!anon_vma1 || !anon_vma2) && (!vma ||
1065  		list_is_singular(&vma->anon_vma_chain)))
1066  		return 1;
1067  	return anon_vma1 == anon_vma2;
1068  }
1069  
1070  /*
1071   * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1072   * in front of (at a lower virtual address and file offset than) the vma.
1073   *
1074   * We cannot merge two vmas if they have differently assigned (non-NULL)
1075   * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1076   *
1077   * We don't check here for the merged mmap wrapping around the end of pagecache
1078   * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1079   * wrap, nor mmaps which cover the final page at index -1UL.
1080   */
1081  static int
1082  can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1083  		     struct anon_vma *anon_vma, struct file *file,
1084  		     pgoff_t vm_pgoff,
1085  		     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1086  {
1087  	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1088  	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1089  		if (vma->vm_pgoff == vm_pgoff)
1090  			return 1;
1091  	}
1092  	return 0;
1093  }
1094  
1095  /*
1096   * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1097   * beyond (at a higher virtual address and file offset than) the vma.
1098   *
1099   * We cannot merge two vmas if they have differently assigned (non-NULL)
1100   * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1101   */
1102  static int
1103  can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1104  		    struct anon_vma *anon_vma, struct file *file,
1105  		    pgoff_t vm_pgoff,
1106  		    struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1107  {
1108  	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1109  	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1110  		pgoff_t vm_pglen;
1111  		vm_pglen = vma_pages(vma);
1112  		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1113  			return 1;
1114  	}
1115  	return 0;
1116  }
1117  
1118  /*
1119   * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1120   * whether that can be merged with its predecessor or its successor.
1121   * Or both (it neatly fills a hole).
1122   *
1123   * In most cases - when called for mmap, brk or mremap - [addr,end) is
1124   * certain not to be mapped by the time vma_merge is called; but when
1125   * called for mprotect, it is certain to be already mapped (either at
1126   * an offset within prev, or at the start of next), and the flags of
1127   * this area are about to be changed to vm_flags - and the no-change
1128   * case has already been eliminated.
1129   *
1130   * The following mprotect cases have to be considered, where AAAA is
1131   * the area passed down from mprotect_fixup, never extending beyond one
1132   * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1133   *
1134   *     AAAA             AAAA                   AAAA
1135   *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
1136   *    cannot merge    might become       might become
1137   *                    PPNNNNNNNNNN       PPPPPPPPPPNN
1138   *    mmap, brk or    case 4 below       case 5 below
1139   *    mremap move:
1140   *                        AAAA               AAAA
1141   *                    PPPP    NNNN       PPPPNNNNXXXX
1142   *                    might become       might become
1143   *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
1144   *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
1145   *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
1146   *
1147   * It is important for case 8 that the vma NNNN overlapping the
1148   * region AAAA is never going to extended over XXXX. Instead XXXX must
1149   * be extended in region AAAA and NNNN must be removed. This way in
1150   * all cases where vma_merge succeeds, the moment vma_adjust drops the
1151   * rmap_locks, the properties of the merged vma will be already
1152   * correct for the whole merged range. Some of those properties like
1153   * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1154   * be correct for the whole merged range immediately after the
1155   * rmap_locks are released. Otherwise if XXXX would be removed and
1156   * NNNN would be extended over the XXXX range, remove_migration_ptes
1157   * or other rmap walkers (if working on addresses beyond the "end"
1158   * parameter) may establish ptes with the wrong permissions of NNNN
1159   * instead of the right permissions of XXXX.
1160   */
1161  struct vm_area_struct *vma_merge(struct mm_struct *mm,
1162  			struct vm_area_struct *prev, unsigned long addr,
1163  			unsigned long end, unsigned long vm_flags,
1164  			struct anon_vma *anon_vma, struct file *file,
1165  			pgoff_t pgoff, struct mempolicy *policy,
1166  			struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1167  {
1168  	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1169  	struct vm_area_struct *area, *next;
1170  	int err;
1171  
1172  	/*
1173  	 * We later require that vma->vm_flags == vm_flags,
1174  	 * so this tests vma->vm_flags & VM_SPECIAL, too.
1175  	 */
1176  	if (vm_flags & VM_SPECIAL)
1177  		return NULL;
1178  
1179  	next = vma_next(mm, prev);
1180  	area = next;
1181  	if (area && area->vm_end == end)		/* cases 6, 7, 8 */
1182  		next = next->vm_next;
1183  
1184  	/* verify some invariant that must be enforced by the caller */
1185  	VM_WARN_ON(prev && addr <= prev->vm_start);
1186  	VM_WARN_ON(area && end > area->vm_end);
1187  	VM_WARN_ON(addr >= end);
1188  
1189  	/*
1190  	 * Can it merge with the predecessor?
1191  	 */
1192  	if (prev && prev->vm_end == addr &&
1193  			mpol_equal(vma_policy(prev), policy) &&
1194  			can_vma_merge_after(prev, vm_flags,
1195  					    anon_vma, file, pgoff,
1196  					    vm_userfaultfd_ctx)) {
1197  		/*
1198  		 * OK, it can.  Can we now merge in the successor as well?
1199  		 */
1200  		if (next && end == next->vm_start &&
1201  				mpol_equal(policy, vma_policy(next)) &&
1202  				can_vma_merge_before(next, vm_flags,
1203  						     anon_vma, file,
1204  						     pgoff+pglen,
1205  						     vm_userfaultfd_ctx) &&
1206  				is_mergeable_anon_vma(prev->anon_vma,
1207  						      next->anon_vma, NULL)) {
1208  							/* cases 1, 6 */
1209  			err = __vma_adjust(prev, prev->vm_start,
1210  					 next->vm_end, prev->vm_pgoff, NULL,
1211  					 prev);
1212  		} else					/* cases 2, 5, 7 */
1213  			err = __vma_adjust(prev, prev->vm_start,
1214  					 end, prev->vm_pgoff, NULL, prev);
1215  		if (err)
1216  			return NULL;
1217  		khugepaged_enter_vma_merge(prev, vm_flags);
1218  		return prev;
1219  	}
1220  
1221  	/*
1222  	 * Can this new request be merged in front of next?
1223  	 */
1224  	if (next && end == next->vm_start &&
1225  			mpol_equal(policy, vma_policy(next)) &&
1226  			can_vma_merge_before(next, vm_flags,
1227  					     anon_vma, file, pgoff+pglen,
1228  					     vm_userfaultfd_ctx)) {
1229  		if (prev && addr < prev->vm_end)	/* case 4 */
1230  			err = __vma_adjust(prev, prev->vm_start,
1231  					 addr, prev->vm_pgoff, NULL, next);
1232  		else {					/* cases 3, 8 */
1233  			err = __vma_adjust(area, addr, next->vm_end,
1234  					 next->vm_pgoff - pglen, NULL, next);
1235  			/*
1236  			 * In case 3 area is already equal to next and
1237  			 * this is a noop, but in case 8 "area" has
1238  			 * been removed and next was expanded over it.
1239  			 */
1240  			area = next;
1241  		}
1242  		if (err)
1243  			return NULL;
1244  		khugepaged_enter_vma_merge(area, vm_flags);
1245  		return area;
1246  	}
1247  
1248  	return NULL;
1249  }
1250  
1251  /*
1252   * Rough compatibility check to quickly see if it's even worth looking
1253   * at sharing an anon_vma.
1254   *
1255   * They need to have the same vm_file, and the flags can only differ
1256   * in things that mprotect may change.
1257   *
1258   * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1259   * we can merge the two vma's. For example, we refuse to merge a vma if
1260   * there is a vm_ops->close() function, because that indicates that the
1261   * driver is doing some kind of reference counting. But that doesn't
1262   * really matter for the anon_vma sharing case.
1263   */
1264  static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1265  {
1266  	return a->vm_end == b->vm_start &&
1267  		mpol_equal(vma_policy(a), vma_policy(b)) &&
1268  		a->vm_file == b->vm_file &&
1269  		!((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1270  		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1271  }
1272  
1273  /*
1274   * Do some basic sanity checking to see if we can re-use the anon_vma
1275   * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1276   * the same as 'old', the other will be the new one that is trying
1277   * to share the anon_vma.
1278   *
1279   * NOTE! This runs with mm_sem held for reading, so it is possible that
1280   * the anon_vma of 'old' is concurrently in the process of being set up
1281   * by another page fault trying to merge _that_. But that's ok: if it
1282   * is being set up, that automatically means that it will be a singleton
1283   * acceptable for merging, so we can do all of this optimistically. But
1284   * we do that READ_ONCE() to make sure that we never re-load the pointer.
1285   *
1286   * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1287   * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1288   * is to return an anon_vma that is "complex" due to having gone through
1289   * a fork).
1290   *
1291   * We also make sure that the two vma's are compatible (adjacent,
1292   * and with the same memory policies). That's all stable, even with just
1293   * a read lock on the mm_sem.
1294   */
1295  static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1296  {
1297  	if (anon_vma_compatible(a, b)) {
1298  		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1299  
1300  		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1301  			return anon_vma;
1302  	}
1303  	return NULL;
1304  }
1305  
1306  /*
1307   * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1308   * neighbouring vmas for a suitable anon_vma, before it goes off
1309   * to allocate a new anon_vma.  It checks because a repetitive
1310   * sequence of mprotects and faults may otherwise lead to distinct
1311   * anon_vmas being allocated, preventing vma merge in subsequent
1312   * mprotect.
1313   */
1314  struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1315  {
1316  	struct anon_vma *anon_vma = NULL;
1317  
1318  	/* Try next first. */
1319  	if (vma->vm_next) {
1320  		anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1321  		if (anon_vma)
1322  			return anon_vma;
1323  	}
1324  
1325  	/* Try prev next. */
1326  	if (vma->vm_prev)
1327  		anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1328  
1329  	/*
1330  	 * We might reach here with anon_vma == NULL if we can't find
1331  	 * any reusable anon_vma.
1332  	 * There's no absolute need to look only at touching neighbours:
1333  	 * we could search further afield for "compatible" anon_vmas.
1334  	 * But it would probably just be a waste of time searching,
1335  	 * or lead to too many vmas hanging off the same anon_vma.
1336  	 * We're trying to allow mprotect remerging later on,
1337  	 * not trying to minimize memory used for anon_vmas.
1338  	 */
1339  	return anon_vma;
1340  }
1341  
1342  /*
1343   * If a hint addr is less than mmap_min_addr change hint to be as
1344   * low as possible but still greater than mmap_min_addr
1345   */
1346  static inline unsigned long round_hint_to_min(unsigned long hint)
1347  {
1348  	hint &= PAGE_MASK;
1349  	if (((void *)hint != NULL) &&
1350  	    (hint < mmap_min_addr))
1351  		return PAGE_ALIGN(mmap_min_addr);
1352  	return hint;
1353  }
1354  
1355  static inline int mlock_future_check(struct mm_struct *mm,
1356  				     unsigned long flags,
1357  				     unsigned long len)
1358  {
1359  	unsigned long locked, lock_limit;
1360  
1361  	/*  mlock MCL_FUTURE? */
1362  	if (flags & VM_LOCKED) {
1363  		locked = len >> PAGE_SHIFT;
1364  		locked += mm->locked_vm;
1365  		lock_limit = rlimit(RLIMIT_MEMLOCK);
1366  		lock_limit >>= PAGE_SHIFT;
1367  		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1368  			return -EAGAIN;
1369  	}
1370  	return 0;
1371  }
1372  
1373  static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1374  {
1375  	if (S_ISREG(inode->i_mode))
1376  		return MAX_LFS_FILESIZE;
1377  
1378  	if (S_ISBLK(inode->i_mode))
1379  		return MAX_LFS_FILESIZE;
1380  
1381  	if (S_ISSOCK(inode->i_mode))
1382  		return MAX_LFS_FILESIZE;
1383  
1384  	/* Special "we do even unsigned file positions" case */
1385  	if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1386  		return 0;
1387  
1388  	/* Yes, random drivers might want more. But I'm tired of buggy drivers */
1389  	return ULONG_MAX;
1390  }
1391  
1392  static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1393  				unsigned long pgoff, unsigned long len)
1394  {
1395  	u64 maxsize = file_mmap_size_max(file, inode);
1396  
1397  	if (maxsize && len > maxsize)
1398  		return false;
1399  	maxsize -= len;
1400  	if (pgoff > maxsize >> PAGE_SHIFT)
1401  		return false;
1402  	return true;
1403  }
1404  
1405  /*
1406   * The caller must write-lock current->mm->mmap_lock.
1407   */
1408  unsigned long do_mmap(struct file *file, unsigned long addr,
1409  			unsigned long len, unsigned long prot,
1410  			unsigned long flags, unsigned long pgoff,
1411  			unsigned long *populate, struct list_head *uf)
1412  {
1413  	struct mm_struct *mm = current->mm;
1414  	vm_flags_t vm_flags;
1415  	int pkey = 0;
1416  
1417  	*populate = 0;
1418  
1419  	if (!len)
1420  		return -EINVAL;
1421  
1422  	/*
1423  	 * Does the application expect PROT_READ to imply PROT_EXEC?
1424  	 *
1425  	 * (the exception is when the underlying filesystem is noexec
1426  	 *  mounted, in which case we dont add PROT_EXEC.)
1427  	 */
1428  	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1429  		if (!(file && path_noexec(&file->f_path)))
1430  			prot |= PROT_EXEC;
1431  
1432  	/* force arch specific MAP_FIXED handling in get_unmapped_area */
1433  	if (flags & MAP_FIXED_NOREPLACE)
1434  		flags |= MAP_FIXED;
1435  
1436  	if (!(flags & MAP_FIXED))
1437  		addr = round_hint_to_min(addr);
1438  
1439  	/* Careful about overflows.. */
1440  	len = PAGE_ALIGN(len);
1441  	if (!len)
1442  		return -ENOMEM;
1443  
1444  	/* offset overflow? */
1445  	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1446  		return -EOVERFLOW;
1447  
1448  	/* Too many mappings? */
1449  	if (mm->map_count > sysctl_max_map_count)
1450  		return -ENOMEM;
1451  
1452  	/* Obtain the address to map to. we verify (or select) it and ensure
1453  	 * that it represents a valid section of the address space.
1454  	 */
1455  	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1456  	if (IS_ERR_VALUE(addr))
1457  		return addr;
1458  
1459  	if (flags & MAP_FIXED_NOREPLACE) {
1460  		struct vm_area_struct *vma = find_vma(mm, addr);
1461  
1462  		if (vma && vma->vm_start < addr + len)
1463  			return -EEXIST;
1464  	}
1465  
1466  	if (prot == PROT_EXEC) {
1467  		pkey = execute_only_pkey(mm);
1468  		if (pkey < 0)
1469  			pkey = 0;
1470  	}
1471  
1472  	/* Do simple checking here so the lower-level routines won't have
1473  	 * to. we assume access permissions have been handled by the open
1474  	 * of the memory object, so we don't do any here.
1475  	 */
1476  	vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1477  			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1478  
1479  	if (flags & MAP_LOCKED)
1480  		if (!can_do_mlock())
1481  			return -EPERM;
1482  
1483  	if (mlock_future_check(mm, vm_flags, len))
1484  		return -EAGAIN;
1485  
1486  	if (file) {
1487  		struct inode *inode = file_inode(file);
1488  		unsigned long flags_mask;
1489  
1490  		if (!file_mmap_ok(file, inode, pgoff, len))
1491  			return -EOVERFLOW;
1492  
1493  		flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1494  
1495  		switch (flags & MAP_TYPE) {
1496  		case MAP_SHARED:
1497  			/*
1498  			 * Force use of MAP_SHARED_VALIDATE with non-legacy
1499  			 * flags. E.g. MAP_SYNC is dangerous to use with
1500  			 * MAP_SHARED as you don't know which consistency model
1501  			 * you will get. We silently ignore unsupported flags
1502  			 * with MAP_SHARED to preserve backward compatibility.
1503  			 */
1504  			flags &= LEGACY_MAP_MASK;
1505  			fallthrough;
1506  		case MAP_SHARED_VALIDATE:
1507  			if (flags & ~flags_mask)
1508  				return -EOPNOTSUPP;
1509  			if (prot & PROT_WRITE) {
1510  				if (!(file->f_mode & FMODE_WRITE))
1511  					return -EACCES;
1512  				if (IS_SWAPFILE(file->f_mapping->host))
1513  					return -ETXTBSY;
1514  			}
1515  
1516  			/*
1517  			 * Make sure we don't allow writing to an append-only
1518  			 * file..
1519  			 */
1520  			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1521  				return -EACCES;
1522  
1523  			/*
1524  			 * Make sure there are no mandatory locks on the file.
1525  			 */
1526  			if (locks_verify_locked(file))
1527  				return -EAGAIN;
1528  
1529  			vm_flags |= VM_SHARED | VM_MAYSHARE;
1530  			if (!(file->f_mode & FMODE_WRITE))
1531  				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1532  			fallthrough;
1533  		case MAP_PRIVATE:
1534  			if (!(file->f_mode & FMODE_READ))
1535  				return -EACCES;
1536  			if (path_noexec(&file->f_path)) {
1537  				if (vm_flags & VM_EXEC)
1538  					return -EPERM;
1539  				vm_flags &= ~VM_MAYEXEC;
1540  			}
1541  
1542  			if (!file->f_op->mmap)
1543  				return -ENODEV;
1544  			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1545  				return -EINVAL;
1546  			break;
1547  
1548  		default:
1549  			return -EINVAL;
1550  		}
1551  	} else {
1552  		switch (flags & MAP_TYPE) {
1553  		case MAP_SHARED:
1554  			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1555  				return -EINVAL;
1556  			/*
1557  			 * Ignore pgoff.
1558  			 */
1559  			pgoff = 0;
1560  			vm_flags |= VM_SHARED | VM_MAYSHARE;
1561  			break;
1562  		case MAP_PRIVATE:
1563  			/*
1564  			 * Set pgoff according to addr for anon_vma.
1565  			 */
1566  			pgoff = addr >> PAGE_SHIFT;
1567  			break;
1568  		default:
1569  			return -EINVAL;
1570  		}
1571  	}
1572  
1573  	/*
1574  	 * Set 'VM_NORESERVE' if we should not account for the
1575  	 * memory use of this mapping.
1576  	 */
1577  	if (flags & MAP_NORESERVE) {
1578  		/* We honor MAP_NORESERVE if allowed to overcommit */
1579  		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1580  			vm_flags |= VM_NORESERVE;
1581  
1582  		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1583  		if (file && is_file_hugepages(file))
1584  			vm_flags |= VM_NORESERVE;
1585  	}
1586  
1587  	addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1588  	if (!IS_ERR_VALUE(addr) &&
1589  	    ((vm_flags & VM_LOCKED) ||
1590  	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1591  		*populate = len;
1592  	return addr;
1593  }
1594  
1595  unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1596  			      unsigned long prot, unsigned long flags,
1597  			      unsigned long fd, unsigned long pgoff)
1598  {
1599  	struct file *file = NULL;
1600  	unsigned long retval;
1601  
1602  	if (!(flags & MAP_ANONYMOUS)) {
1603  		audit_mmap_fd(fd, flags);
1604  		file = fget(fd);
1605  		if (!file)
1606  			return -EBADF;
1607  		if (is_file_hugepages(file)) {
1608  			len = ALIGN(len, huge_page_size(hstate_file(file)));
1609  		} else if (unlikely(flags & MAP_HUGETLB)) {
1610  			retval = -EINVAL;
1611  			goto out_fput;
1612  		}
1613  	} else if (flags & MAP_HUGETLB) {
1614  		struct user_struct *user = NULL;
1615  		struct hstate *hs;
1616  
1617  		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1618  		if (!hs)
1619  			return -EINVAL;
1620  
1621  		len = ALIGN(len, huge_page_size(hs));
1622  		/*
1623  		 * VM_NORESERVE is used because the reservations will be
1624  		 * taken when vm_ops->mmap() is called
1625  		 * A dummy user value is used because we are not locking
1626  		 * memory so no accounting is necessary
1627  		 */
1628  		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1629  				VM_NORESERVE,
1630  				&user, HUGETLB_ANONHUGE_INODE,
1631  				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1632  		if (IS_ERR(file))
1633  			return PTR_ERR(file);
1634  	}
1635  
1636  	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1637  
1638  	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1639  out_fput:
1640  	if (file)
1641  		fput(file);
1642  	return retval;
1643  }
1644  
1645  SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1646  		unsigned long, prot, unsigned long, flags,
1647  		unsigned long, fd, unsigned long, pgoff)
1648  {
1649  	return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1650  }
1651  
1652  #ifdef __ARCH_WANT_SYS_OLD_MMAP
1653  struct mmap_arg_struct {
1654  	unsigned long addr;
1655  	unsigned long len;
1656  	unsigned long prot;
1657  	unsigned long flags;
1658  	unsigned long fd;
1659  	unsigned long offset;
1660  };
1661  
1662  SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1663  {
1664  	struct mmap_arg_struct a;
1665  
1666  	if (copy_from_user(&a, arg, sizeof(a)))
1667  		return -EFAULT;
1668  	if (offset_in_page(a.offset))
1669  		return -EINVAL;
1670  
1671  	return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1672  			       a.offset >> PAGE_SHIFT);
1673  }
1674  #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1675  
1676  /*
1677   * Some shared mappings will want the pages marked read-only
1678   * to track write events. If so, we'll downgrade vm_page_prot
1679   * to the private version (using protection_map[] without the
1680   * VM_SHARED bit).
1681   */
1682  int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1683  {
1684  	vm_flags_t vm_flags = vma->vm_flags;
1685  	const struct vm_operations_struct *vm_ops = vma->vm_ops;
1686  
1687  	/* If it was private or non-writable, the write bit is already clear */
1688  	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1689  		return 0;
1690  
1691  	/* The backer wishes to know when pages are first written to? */
1692  	if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1693  		return 1;
1694  
1695  	/* The open routine did something to the protections that pgprot_modify
1696  	 * won't preserve? */
1697  	if (pgprot_val(vm_page_prot) !=
1698  	    pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1699  		return 0;
1700  
1701  	/* Do we need to track softdirty? */
1702  	if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1703  		return 1;
1704  
1705  	/* Specialty mapping? */
1706  	if (vm_flags & VM_PFNMAP)
1707  		return 0;
1708  
1709  	/* Can the mapping track the dirty pages? */
1710  	return vma->vm_file && vma->vm_file->f_mapping &&
1711  		mapping_can_writeback(vma->vm_file->f_mapping);
1712  }
1713  
1714  /*
1715   * We account for memory if it's a private writeable mapping,
1716   * not hugepages and VM_NORESERVE wasn't set.
1717   */
1718  static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1719  {
1720  	/*
1721  	 * hugetlb has its own accounting separate from the core VM
1722  	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1723  	 */
1724  	if (file && is_file_hugepages(file))
1725  		return 0;
1726  
1727  	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1728  }
1729  
1730  unsigned long mmap_region(struct file *file, unsigned long addr,
1731  		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1732  		struct list_head *uf)
1733  {
1734  	struct mm_struct *mm = current->mm;
1735  	struct vm_area_struct *vma, *prev, *merge;
1736  	int error;
1737  	struct rb_node **rb_link, *rb_parent;
1738  	unsigned long charged = 0;
1739  
1740  	/* Check against address space limit. */
1741  	if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1742  		unsigned long nr_pages;
1743  
1744  		/*
1745  		 * MAP_FIXED may remove pages of mappings that intersects with
1746  		 * requested mapping. Account for the pages it would unmap.
1747  		 */
1748  		nr_pages = count_vma_pages_range(mm, addr, addr + len);
1749  
1750  		if (!may_expand_vm(mm, vm_flags,
1751  					(len >> PAGE_SHIFT) - nr_pages))
1752  			return -ENOMEM;
1753  	}
1754  
1755  	/* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1756  	if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1757  		return -ENOMEM;
1758  	/*
1759  	 * Private writable mapping: check memory availability
1760  	 */
1761  	if (accountable_mapping(file, vm_flags)) {
1762  		charged = len >> PAGE_SHIFT;
1763  		if (security_vm_enough_memory_mm(mm, charged))
1764  			return -ENOMEM;
1765  		vm_flags |= VM_ACCOUNT;
1766  	}
1767  
1768  	/*
1769  	 * Can we just expand an old mapping?
1770  	 */
1771  	vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1772  			NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1773  	if (vma)
1774  		goto out;
1775  
1776  	/*
1777  	 * Determine the object being mapped and call the appropriate
1778  	 * specific mapper. the address has already been validated, but
1779  	 * not unmapped, but the maps are removed from the list.
1780  	 */
1781  	vma = vm_area_alloc(mm);
1782  	if (!vma) {
1783  		error = -ENOMEM;
1784  		goto unacct_error;
1785  	}
1786  
1787  	vma->vm_start = addr;
1788  	vma->vm_end = addr + len;
1789  	vma->vm_flags = vm_flags;
1790  	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1791  	vma->vm_pgoff = pgoff;
1792  
1793  	if (file) {
1794  		if (vm_flags & VM_DENYWRITE) {
1795  			error = deny_write_access(file);
1796  			if (error)
1797  				goto free_vma;
1798  		}
1799  		if (vm_flags & VM_SHARED) {
1800  			error = mapping_map_writable(file->f_mapping);
1801  			if (error)
1802  				goto allow_write_and_free_vma;
1803  		}
1804  
1805  		/* ->mmap() can change vma->vm_file, but must guarantee that
1806  		 * vma_link() below can deny write-access if VM_DENYWRITE is set
1807  		 * and map writably if VM_SHARED is set. This usually means the
1808  		 * new file must not have been exposed to user-space, yet.
1809  		 */
1810  		vma->vm_file = get_file(file);
1811  		error = call_mmap(file, vma);
1812  		if (error)
1813  			goto unmap_and_free_vma;
1814  
1815  		/* Can addr have changed??
1816  		 *
1817  		 * Answer: Yes, several device drivers can do it in their
1818  		 *         f_op->mmap method. -DaveM
1819  		 * Bug: If addr is changed, prev, rb_link, rb_parent should
1820  		 *      be updated for vma_link()
1821  		 */
1822  		WARN_ON_ONCE(addr != vma->vm_start);
1823  
1824  		addr = vma->vm_start;
1825  
1826  		/* If vm_flags changed after call_mmap(), we should try merge vma again
1827  		 * as we may succeed this time.
1828  		 */
1829  		if (unlikely(vm_flags != vma->vm_flags && prev)) {
1830  			merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1831  				NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1832  			if (merge) {
1833  				/* ->mmap() can change vma->vm_file and fput the original file. So
1834  				 * fput the vma->vm_file here or we would add an extra fput for file
1835  				 * and cause general protection fault ultimately.
1836  				 */
1837  				fput(vma->vm_file);
1838  				vm_area_free(vma);
1839  				vma = merge;
1840  				/* Update vm_flags to pick up the change. */
1841  				vm_flags = vma->vm_flags;
1842  				goto unmap_writable;
1843  			}
1844  		}
1845  
1846  		vm_flags = vma->vm_flags;
1847  	} else if (vm_flags & VM_SHARED) {
1848  		error = shmem_zero_setup(vma);
1849  		if (error)
1850  			goto free_vma;
1851  	} else {
1852  		vma_set_anonymous(vma);
1853  	}
1854  
1855  	/* Allow architectures to sanity-check the vm_flags */
1856  	if (!arch_validate_flags(vma->vm_flags)) {
1857  		error = -EINVAL;
1858  		if (file)
1859  			goto unmap_and_free_vma;
1860  		else
1861  			goto free_vma;
1862  	}
1863  
1864  	vma_link(mm, vma, prev, rb_link, rb_parent);
1865  	/* Once vma denies write, undo our temporary denial count */
1866  	if (file) {
1867  unmap_writable:
1868  		if (vm_flags & VM_SHARED)
1869  			mapping_unmap_writable(file->f_mapping);
1870  		if (vm_flags & VM_DENYWRITE)
1871  			allow_write_access(file);
1872  	}
1873  	file = vma->vm_file;
1874  out:
1875  	perf_event_mmap(vma);
1876  
1877  	vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1878  	if (vm_flags & VM_LOCKED) {
1879  		if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1880  					is_vm_hugetlb_page(vma) ||
1881  					vma == get_gate_vma(current->mm))
1882  			vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1883  		else
1884  			mm->locked_vm += (len >> PAGE_SHIFT);
1885  	}
1886  
1887  	if (file)
1888  		uprobe_mmap(vma);
1889  
1890  	/*
1891  	 * New (or expanded) vma always get soft dirty status.
1892  	 * Otherwise user-space soft-dirty page tracker won't
1893  	 * be able to distinguish situation when vma area unmapped,
1894  	 * then new mapped in-place (which must be aimed as
1895  	 * a completely new data area).
1896  	 */
1897  	vma->vm_flags |= VM_SOFTDIRTY;
1898  
1899  	vma_set_page_prot(vma);
1900  
1901  	return addr;
1902  
1903  unmap_and_free_vma:
1904  	fput(vma->vm_file);
1905  	vma->vm_file = NULL;
1906  
1907  	/* Undo any partial mapping done by a device driver. */
1908  	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1909  	charged = 0;
1910  	if (vm_flags & VM_SHARED)
1911  		mapping_unmap_writable(file->f_mapping);
1912  allow_write_and_free_vma:
1913  	if (vm_flags & VM_DENYWRITE)
1914  		allow_write_access(file);
1915  free_vma:
1916  	vm_area_free(vma);
1917  unacct_error:
1918  	if (charged)
1919  		vm_unacct_memory(charged);
1920  	return error;
1921  }
1922  
1923  static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1924  {
1925  	/*
1926  	 * We implement the search by looking for an rbtree node that
1927  	 * immediately follows a suitable gap. That is,
1928  	 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1929  	 * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1930  	 * - gap_end - gap_start >= length
1931  	 */
1932  
1933  	struct mm_struct *mm = current->mm;
1934  	struct vm_area_struct *vma;
1935  	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1936  
1937  	/* Adjust search length to account for worst case alignment overhead */
1938  	length = info->length + info->align_mask;
1939  	if (length < info->length)
1940  		return -ENOMEM;
1941  
1942  	/* Adjust search limits by the desired length */
1943  	if (info->high_limit < length)
1944  		return -ENOMEM;
1945  	high_limit = info->high_limit - length;
1946  
1947  	if (info->low_limit > high_limit)
1948  		return -ENOMEM;
1949  	low_limit = info->low_limit + length;
1950  
1951  	/* Check if rbtree root looks promising */
1952  	if (RB_EMPTY_ROOT(&mm->mm_rb))
1953  		goto check_highest;
1954  	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1955  	if (vma->rb_subtree_gap < length)
1956  		goto check_highest;
1957  
1958  	while (true) {
1959  		/* Visit left subtree if it looks promising */
1960  		gap_end = vm_start_gap(vma);
1961  		if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1962  			struct vm_area_struct *left =
1963  				rb_entry(vma->vm_rb.rb_left,
1964  					 struct vm_area_struct, vm_rb);
1965  			if (left->rb_subtree_gap >= length) {
1966  				vma = left;
1967  				continue;
1968  			}
1969  		}
1970  
1971  		gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1972  check_current:
1973  		/* Check if current node has a suitable gap */
1974  		if (gap_start > high_limit)
1975  			return -ENOMEM;
1976  		if (gap_end >= low_limit &&
1977  		    gap_end > gap_start && gap_end - gap_start >= length)
1978  			goto found;
1979  
1980  		/* Visit right subtree if it looks promising */
1981  		if (vma->vm_rb.rb_right) {
1982  			struct vm_area_struct *right =
1983  				rb_entry(vma->vm_rb.rb_right,
1984  					 struct vm_area_struct, vm_rb);
1985  			if (right->rb_subtree_gap >= length) {
1986  				vma = right;
1987  				continue;
1988  			}
1989  		}
1990  
1991  		/* Go back up the rbtree to find next candidate node */
1992  		while (true) {
1993  			struct rb_node *prev = &vma->vm_rb;
1994  			if (!rb_parent(prev))
1995  				goto check_highest;
1996  			vma = rb_entry(rb_parent(prev),
1997  				       struct vm_area_struct, vm_rb);
1998  			if (prev == vma->vm_rb.rb_left) {
1999  				gap_start = vm_end_gap(vma->vm_prev);
2000  				gap_end = vm_start_gap(vma);
2001  				goto check_current;
2002  			}
2003  		}
2004  	}
2005  
2006  check_highest:
2007  	/* Check highest gap, which does not precede any rbtree node */
2008  	gap_start = mm->highest_vm_end;
2009  	gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
2010  	if (gap_start > high_limit)
2011  		return -ENOMEM;
2012  
2013  found:
2014  	/* We found a suitable gap. Clip it with the original low_limit. */
2015  	if (gap_start < info->low_limit)
2016  		gap_start = info->low_limit;
2017  
2018  	/* Adjust gap address to the desired alignment */
2019  	gap_start += (info->align_offset - gap_start) & info->align_mask;
2020  
2021  	VM_BUG_ON(gap_start + info->length > info->high_limit);
2022  	VM_BUG_ON(gap_start + info->length > gap_end);
2023  	return gap_start;
2024  }
2025  
2026  static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
2027  {
2028  	struct mm_struct *mm = current->mm;
2029  	struct vm_area_struct *vma;
2030  	unsigned long length, low_limit, high_limit, gap_start, gap_end;
2031  
2032  	/* Adjust search length to account for worst case alignment overhead */
2033  	length = info->length + info->align_mask;
2034  	if (length < info->length)
2035  		return -ENOMEM;
2036  
2037  	/*
2038  	 * Adjust search limits by the desired length.
2039  	 * See implementation comment at top of unmapped_area().
2040  	 */
2041  	gap_end = info->high_limit;
2042  	if (gap_end < length)
2043  		return -ENOMEM;
2044  	high_limit = gap_end - length;
2045  
2046  	if (info->low_limit > high_limit)
2047  		return -ENOMEM;
2048  	low_limit = info->low_limit + length;
2049  
2050  	/* Check highest gap, which does not precede any rbtree node */
2051  	gap_start = mm->highest_vm_end;
2052  	if (gap_start <= high_limit)
2053  		goto found_highest;
2054  
2055  	/* Check if rbtree root looks promising */
2056  	if (RB_EMPTY_ROOT(&mm->mm_rb))
2057  		return -ENOMEM;
2058  	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2059  	if (vma->rb_subtree_gap < length)
2060  		return -ENOMEM;
2061  
2062  	while (true) {
2063  		/* Visit right subtree if it looks promising */
2064  		gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2065  		if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2066  			struct vm_area_struct *right =
2067  				rb_entry(vma->vm_rb.rb_right,
2068  					 struct vm_area_struct, vm_rb);
2069  			if (right->rb_subtree_gap >= length) {
2070  				vma = right;
2071  				continue;
2072  			}
2073  		}
2074  
2075  check_current:
2076  		/* Check if current node has a suitable gap */
2077  		gap_end = vm_start_gap(vma);
2078  		if (gap_end < low_limit)
2079  			return -ENOMEM;
2080  		if (gap_start <= high_limit &&
2081  		    gap_end > gap_start && gap_end - gap_start >= length)
2082  			goto found;
2083  
2084  		/* Visit left subtree if it looks promising */
2085  		if (vma->vm_rb.rb_left) {
2086  			struct vm_area_struct *left =
2087  				rb_entry(vma->vm_rb.rb_left,
2088  					 struct vm_area_struct, vm_rb);
2089  			if (left->rb_subtree_gap >= length) {
2090  				vma = left;
2091  				continue;
2092  			}
2093  		}
2094  
2095  		/* Go back up the rbtree to find next candidate node */
2096  		while (true) {
2097  			struct rb_node *prev = &vma->vm_rb;
2098  			if (!rb_parent(prev))
2099  				return -ENOMEM;
2100  			vma = rb_entry(rb_parent(prev),
2101  				       struct vm_area_struct, vm_rb);
2102  			if (prev == vma->vm_rb.rb_right) {
2103  				gap_start = vma->vm_prev ?
2104  					vm_end_gap(vma->vm_prev) : 0;
2105  				goto check_current;
2106  			}
2107  		}
2108  	}
2109  
2110  found:
2111  	/* We found a suitable gap. Clip it with the original high_limit. */
2112  	if (gap_end > info->high_limit)
2113  		gap_end = info->high_limit;
2114  
2115  found_highest:
2116  	/* Compute highest gap address at the desired alignment */
2117  	gap_end -= info->length;
2118  	gap_end -= (gap_end - info->align_offset) & info->align_mask;
2119  
2120  	VM_BUG_ON(gap_end < info->low_limit);
2121  	VM_BUG_ON(gap_end < gap_start);
2122  	return gap_end;
2123  }
2124  
2125  /*
2126   * Search for an unmapped address range.
2127   *
2128   * We are looking for a range that:
2129   * - does not intersect with any VMA;
2130   * - is contained within the [low_limit, high_limit) interval;
2131   * - is at least the desired size.
2132   * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2133   */
2134  unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2135  {
2136  	unsigned long addr;
2137  
2138  	if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2139  		addr = unmapped_area_topdown(info);
2140  	else
2141  		addr = unmapped_area(info);
2142  
2143  	trace_vm_unmapped_area(addr, info);
2144  	return addr;
2145  }
2146  
2147  #ifndef arch_get_mmap_end
2148  #define arch_get_mmap_end(addr)	(TASK_SIZE)
2149  #endif
2150  
2151  #ifndef arch_get_mmap_base
2152  #define arch_get_mmap_base(addr, base) (base)
2153  #endif
2154  
2155  /* Get an address range which is currently unmapped.
2156   * For shmat() with addr=0.
2157   *
2158   * Ugly calling convention alert:
2159   * Return value with the low bits set means error value,
2160   * ie
2161   *	if (ret & ~PAGE_MASK)
2162   *		error = ret;
2163   *
2164   * This function "knows" that -ENOMEM has the bits set.
2165   */
2166  #ifndef HAVE_ARCH_UNMAPPED_AREA
2167  unsigned long
2168  arch_get_unmapped_area(struct file *filp, unsigned long addr,
2169  		unsigned long len, unsigned long pgoff, unsigned long flags)
2170  {
2171  	struct mm_struct *mm = current->mm;
2172  	struct vm_area_struct *vma, *prev;
2173  	struct vm_unmapped_area_info info;
2174  	const unsigned long mmap_end = arch_get_mmap_end(addr);
2175  
2176  	if (len > mmap_end - mmap_min_addr)
2177  		return -ENOMEM;
2178  
2179  	if (flags & MAP_FIXED)
2180  		return addr;
2181  
2182  	if (addr) {
2183  		addr = PAGE_ALIGN(addr);
2184  		vma = find_vma_prev(mm, addr, &prev);
2185  		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2186  		    (!vma || addr + len <= vm_start_gap(vma)) &&
2187  		    (!prev || addr >= vm_end_gap(prev)))
2188  			return addr;
2189  	}
2190  
2191  	info.flags = 0;
2192  	info.length = len;
2193  	info.low_limit = mm->mmap_base;
2194  	info.high_limit = mmap_end;
2195  	info.align_mask = 0;
2196  	info.align_offset = 0;
2197  	return vm_unmapped_area(&info);
2198  }
2199  #endif
2200  
2201  /*
2202   * This mmap-allocator allocates new areas top-down from below the
2203   * stack's low limit (the base):
2204   */
2205  #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2206  unsigned long
2207  arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2208  			  unsigned long len, unsigned long pgoff,
2209  			  unsigned long flags)
2210  {
2211  	struct vm_area_struct *vma, *prev;
2212  	struct mm_struct *mm = current->mm;
2213  	struct vm_unmapped_area_info info;
2214  	const unsigned long mmap_end = arch_get_mmap_end(addr);
2215  
2216  	/* requested length too big for entire address space */
2217  	if (len > mmap_end - mmap_min_addr)
2218  		return -ENOMEM;
2219  
2220  	if (flags & MAP_FIXED)
2221  		return addr;
2222  
2223  	/* requesting a specific address */
2224  	if (addr) {
2225  		addr = PAGE_ALIGN(addr);
2226  		vma = find_vma_prev(mm, addr, &prev);
2227  		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2228  				(!vma || addr + len <= vm_start_gap(vma)) &&
2229  				(!prev || addr >= vm_end_gap(prev)))
2230  			return addr;
2231  	}
2232  
2233  	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2234  	info.length = len;
2235  	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2236  	info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2237  	info.align_mask = 0;
2238  	info.align_offset = 0;
2239  	addr = vm_unmapped_area(&info);
2240  
2241  	/*
2242  	 * A failed mmap() very likely causes application failure,
2243  	 * so fall back to the bottom-up function here. This scenario
2244  	 * can happen with large stack limits and large mmap()
2245  	 * allocations.
2246  	 */
2247  	if (offset_in_page(addr)) {
2248  		VM_BUG_ON(addr != -ENOMEM);
2249  		info.flags = 0;
2250  		info.low_limit = TASK_UNMAPPED_BASE;
2251  		info.high_limit = mmap_end;
2252  		addr = vm_unmapped_area(&info);
2253  	}
2254  
2255  	return addr;
2256  }
2257  #endif
2258  
2259  unsigned long
2260  get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2261  		unsigned long pgoff, unsigned long flags)
2262  {
2263  	unsigned long (*get_area)(struct file *, unsigned long,
2264  				  unsigned long, unsigned long, unsigned long);
2265  
2266  	unsigned long error = arch_mmap_check(addr, len, flags);
2267  	if (error)
2268  		return error;
2269  
2270  	/* Careful about overflows.. */
2271  	if (len > TASK_SIZE)
2272  		return -ENOMEM;
2273  
2274  	get_area = current->mm->get_unmapped_area;
2275  	if (file) {
2276  		if (file->f_op->get_unmapped_area)
2277  			get_area = file->f_op->get_unmapped_area;
2278  	} else if (flags & MAP_SHARED) {
2279  		/*
2280  		 * mmap_region() will call shmem_zero_setup() to create a file,
2281  		 * so use shmem's get_unmapped_area in case it can be huge.
2282  		 * do_mmap() will clear pgoff, so match alignment.
2283  		 */
2284  		pgoff = 0;
2285  		get_area = shmem_get_unmapped_area;
2286  	}
2287  
2288  	addr = get_area(file, addr, len, pgoff, flags);
2289  	if (IS_ERR_VALUE(addr))
2290  		return addr;
2291  
2292  	if (addr > TASK_SIZE - len)
2293  		return -ENOMEM;
2294  	if (offset_in_page(addr))
2295  		return -EINVAL;
2296  
2297  	error = security_mmap_addr(addr);
2298  	return error ? error : addr;
2299  }
2300  
2301  EXPORT_SYMBOL(get_unmapped_area);
2302  
2303  /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2304  struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2305  {
2306  	struct rb_node *rb_node;
2307  	struct vm_area_struct *vma;
2308  
2309  	/* Check the cache first. */
2310  	vma = vmacache_find(mm, addr);
2311  	if (likely(vma))
2312  		return vma;
2313  
2314  	rb_node = mm->mm_rb.rb_node;
2315  
2316  	while (rb_node) {
2317  		struct vm_area_struct *tmp;
2318  
2319  		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2320  
2321  		if (tmp->vm_end > addr) {
2322  			vma = tmp;
2323  			if (tmp->vm_start <= addr)
2324  				break;
2325  			rb_node = rb_node->rb_left;
2326  		} else
2327  			rb_node = rb_node->rb_right;
2328  	}
2329  
2330  	if (vma)
2331  		vmacache_update(addr, vma);
2332  	return vma;
2333  }
2334  
2335  EXPORT_SYMBOL(find_vma);
2336  
2337  /*
2338   * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2339   */
2340  struct vm_area_struct *
2341  find_vma_prev(struct mm_struct *mm, unsigned long addr,
2342  			struct vm_area_struct **pprev)
2343  {
2344  	struct vm_area_struct *vma;
2345  
2346  	vma = find_vma(mm, addr);
2347  	if (vma) {
2348  		*pprev = vma->vm_prev;
2349  	} else {
2350  		struct rb_node *rb_node = rb_last(&mm->mm_rb);
2351  
2352  		*pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2353  	}
2354  	return vma;
2355  }
2356  
2357  /*
2358   * Verify that the stack growth is acceptable and
2359   * update accounting. This is shared with both the
2360   * grow-up and grow-down cases.
2361   */
2362  static int acct_stack_growth(struct vm_area_struct *vma,
2363  			     unsigned long size, unsigned long grow)
2364  {
2365  	struct mm_struct *mm = vma->vm_mm;
2366  	unsigned long new_start;
2367  
2368  	/* address space limit tests */
2369  	if (!may_expand_vm(mm, vma->vm_flags, grow))
2370  		return -ENOMEM;
2371  
2372  	/* Stack limit test */
2373  	if (size > rlimit(RLIMIT_STACK))
2374  		return -ENOMEM;
2375  
2376  	/* mlock limit tests */
2377  	if (vma->vm_flags & VM_LOCKED) {
2378  		unsigned long locked;
2379  		unsigned long limit;
2380  		locked = mm->locked_vm + grow;
2381  		limit = rlimit(RLIMIT_MEMLOCK);
2382  		limit >>= PAGE_SHIFT;
2383  		if (locked > limit && !capable(CAP_IPC_LOCK))
2384  			return -ENOMEM;
2385  	}
2386  
2387  	/* Check to ensure the stack will not grow into a hugetlb-only region */
2388  	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2389  			vma->vm_end - size;
2390  	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2391  		return -EFAULT;
2392  
2393  	/*
2394  	 * Overcommit..  This must be the final test, as it will
2395  	 * update security statistics.
2396  	 */
2397  	if (security_vm_enough_memory_mm(mm, grow))
2398  		return -ENOMEM;
2399  
2400  	return 0;
2401  }
2402  
2403  #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2404  /*
2405   * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2406   * vma is the last one with address > vma->vm_end.  Have to extend vma.
2407   */
2408  int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2409  {
2410  	struct mm_struct *mm = vma->vm_mm;
2411  	struct vm_area_struct *next;
2412  	unsigned long gap_addr;
2413  	int error = 0;
2414  
2415  	if (!(vma->vm_flags & VM_GROWSUP))
2416  		return -EFAULT;
2417  
2418  	/* Guard against exceeding limits of the address space. */
2419  	address &= PAGE_MASK;
2420  	if (address >= (TASK_SIZE & PAGE_MASK))
2421  		return -ENOMEM;
2422  	address += PAGE_SIZE;
2423  
2424  	/* Enforce stack_guard_gap */
2425  	gap_addr = address + stack_guard_gap;
2426  
2427  	/* Guard against overflow */
2428  	if (gap_addr < address || gap_addr > TASK_SIZE)
2429  		gap_addr = TASK_SIZE;
2430  
2431  	next = vma->vm_next;
2432  	if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2433  		if (!(next->vm_flags & VM_GROWSUP))
2434  			return -ENOMEM;
2435  		/* Check that both stack segments have the same anon_vma? */
2436  	}
2437  
2438  	/* We must make sure the anon_vma is allocated. */
2439  	if (unlikely(anon_vma_prepare(vma)))
2440  		return -ENOMEM;
2441  
2442  	/*
2443  	 * vma->vm_start/vm_end cannot change under us because the caller
2444  	 * is required to hold the mmap_lock in read mode.  We need the
2445  	 * anon_vma lock to serialize against concurrent expand_stacks.
2446  	 */
2447  	anon_vma_lock_write(vma->anon_vma);
2448  
2449  	/* Somebody else might have raced and expanded it already */
2450  	if (address > vma->vm_end) {
2451  		unsigned long size, grow;
2452  
2453  		size = address - vma->vm_start;
2454  		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2455  
2456  		error = -ENOMEM;
2457  		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2458  			error = acct_stack_growth(vma, size, grow);
2459  			if (!error) {
2460  				/*
2461  				 * vma_gap_update() doesn't support concurrent
2462  				 * updates, but we only hold a shared mmap_lock
2463  				 * lock here, so we need to protect against
2464  				 * concurrent vma expansions.
2465  				 * anon_vma_lock_write() doesn't help here, as
2466  				 * we don't guarantee that all growable vmas
2467  				 * in a mm share the same root anon vma.
2468  				 * So, we reuse mm->page_table_lock to guard
2469  				 * against concurrent vma expansions.
2470  				 */
2471  				spin_lock(&mm->page_table_lock);
2472  				if (vma->vm_flags & VM_LOCKED)
2473  					mm->locked_vm += grow;
2474  				vm_stat_account(mm, vma->vm_flags, grow);
2475  				anon_vma_interval_tree_pre_update_vma(vma);
2476  				vma->vm_end = address;
2477  				anon_vma_interval_tree_post_update_vma(vma);
2478  				if (vma->vm_next)
2479  					vma_gap_update(vma->vm_next);
2480  				else
2481  					mm->highest_vm_end = vm_end_gap(vma);
2482  				spin_unlock(&mm->page_table_lock);
2483  
2484  				perf_event_mmap(vma);
2485  			}
2486  		}
2487  	}
2488  	anon_vma_unlock_write(vma->anon_vma);
2489  	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2490  	validate_mm(mm);
2491  	return error;
2492  }
2493  #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2494  
2495  /*
2496   * vma is the first one with address < vma->vm_start.  Have to extend vma.
2497   */
2498  int expand_downwards(struct vm_area_struct *vma,
2499  				   unsigned long address)
2500  {
2501  	struct mm_struct *mm = vma->vm_mm;
2502  	struct vm_area_struct *prev;
2503  	int error = 0;
2504  
2505  	address &= PAGE_MASK;
2506  	if (address < mmap_min_addr)
2507  		return -EPERM;
2508  
2509  	/* Enforce stack_guard_gap */
2510  	prev = vma->vm_prev;
2511  	/* Check that both stack segments have the same anon_vma? */
2512  	if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2513  			vma_is_accessible(prev)) {
2514  		if (address - prev->vm_end < stack_guard_gap)
2515  			return -ENOMEM;
2516  	}
2517  
2518  	/* We must make sure the anon_vma is allocated. */
2519  	if (unlikely(anon_vma_prepare(vma)))
2520  		return -ENOMEM;
2521  
2522  	/*
2523  	 * vma->vm_start/vm_end cannot change under us because the caller
2524  	 * is required to hold the mmap_lock in read mode.  We need the
2525  	 * anon_vma lock to serialize against concurrent expand_stacks.
2526  	 */
2527  	anon_vma_lock_write(vma->anon_vma);
2528  
2529  	/* Somebody else might have raced and expanded it already */
2530  	if (address < vma->vm_start) {
2531  		unsigned long size, grow;
2532  
2533  		size = vma->vm_end - address;
2534  		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2535  
2536  		error = -ENOMEM;
2537  		if (grow <= vma->vm_pgoff) {
2538  			error = acct_stack_growth(vma, size, grow);
2539  			if (!error) {
2540  				/*
2541  				 * vma_gap_update() doesn't support concurrent
2542  				 * updates, but we only hold a shared mmap_lock
2543  				 * lock here, so we need to protect against
2544  				 * concurrent vma expansions.
2545  				 * anon_vma_lock_write() doesn't help here, as
2546  				 * we don't guarantee that all growable vmas
2547  				 * in a mm share the same root anon vma.
2548  				 * So, we reuse mm->page_table_lock to guard
2549  				 * against concurrent vma expansions.
2550  				 */
2551  				spin_lock(&mm->page_table_lock);
2552  				if (vma->vm_flags & VM_LOCKED)
2553  					mm->locked_vm += grow;
2554  				vm_stat_account(mm, vma->vm_flags, grow);
2555  				anon_vma_interval_tree_pre_update_vma(vma);
2556  				vma->vm_start = address;
2557  				vma->vm_pgoff -= grow;
2558  				anon_vma_interval_tree_post_update_vma(vma);
2559  				vma_gap_update(vma);
2560  				spin_unlock(&mm->page_table_lock);
2561  
2562  				perf_event_mmap(vma);
2563  			}
2564  		}
2565  	}
2566  	anon_vma_unlock_write(vma->anon_vma);
2567  	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2568  	validate_mm(mm);
2569  	return error;
2570  }
2571  
2572  /* enforced gap between the expanding stack and other mappings. */
2573  unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2574  
2575  static int __init cmdline_parse_stack_guard_gap(char *p)
2576  {
2577  	unsigned long val;
2578  	char *endptr;
2579  
2580  	val = simple_strtoul(p, &endptr, 10);
2581  	if (!*endptr)
2582  		stack_guard_gap = val << PAGE_SHIFT;
2583  
2584  	return 0;
2585  }
2586  __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2587  
2588  #ifdef CONFIG_STACK_GROWSUP
2589  int expand_stack(struct vm_area_struct *vma, unsigned long address)
2590  {
2591  	return expand_upwards(vma, address);
2592  }
2593  
2594  struct vm_area_struct *
2595  find_extend_vma(struct mm_struct *mm, unsigned long addr)
2596  {
2597  	struct vm_area_struct *vma, *prev;
2598  
2599  	addr &= PAGE_MASK;
2600  	vma = find_vma_prev(mm, addr, &prev);
2601  	if (vma && (vma->vm_start <= addr))
2602  		return vma;
2603  	/* don't alter vm_end if the coredump is running */
2604  	if (!prev || expand_stack(prev, addr))
2605  		return NULL;
2606  	if (prev->vm_flags & VM_LOCKED)
2607  		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2608  	return prev;
2609  }
2610  #else
2611  int expand_stack(struct vm_area_struct *vma, unsigned long address)
2612  {
2613  	return expand_downwards(vma, address);
2614  }
2615  
2616  struct vm_area_struct *
2617  find_extend_vma(struct mm_struct *mm, unsigned long addr)
2618  {
2619  	struct vm_area_struct *vma;
2620  	unsigned long start;
2621  
2622  	addr &= PAGE_MASK;
2623  	vma = find_vma(mm, addr);
2624  	if (!vma)
2625  		return NULL;
2626  	if (vma->vm_start <= addr)
2627  		return vma;
2628  	if (!(vma->vm_flags & VM_GROWSDOWN))
2629  		return NULL;
2630  	start = vma->vm_start;
2631  	if (expand_stack(vma, addr))
2632  		return NULL;
2633  	if (vma->vm_flags & VM_LOCKED)
2634  		populate_vma_page_range(vma, addr, start, NULL);
2635  	return vma;
2636  }
2637  #endif
2638  
2639  EXPORT_SYMBOL_GPL(find_extend_vma);
2640  
2641  /*
2642   * Ok - we have the memory areas we should free on the vma list,
2643   * so release them, and do the vma updates.
2644   *
2645   * Called with the mm semaphore held.
2646   */
2647  static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2648  {
2649  	unsigned long nr_accounted = 0;
2650  
2651  	/* Update high watermark before we lower total_vm */
2652  	update_hiwater_vm(mm);
2653  	do {
2654  		long nrpages = vma_pages(vma);
2655  
2656  		if (vma->vm_flags & VM_ACCOUNT)
2657  			nr_accounted += nrpages;
2658  		vm_stat_account(mm, vma->vm_flags, -nrpages);
2659  		vma = remove_vma(vma);
2660  	} while (vma);
2661  	vm_unacct_memory(nr_accounted);
2662  	validate_mm(mm);
2663  }
2664  
2665  /*
2666   * Get rid of page table information in the indicated region.
2667   *
2668   * Called with the mm semaphore held.
2669   */
2670  static void unmap_region(struct mm_struct *mm,
2671  		struct vm_area_struct *vma, struct vm_area_struct *prev,
2672  		unsigned long start, unsigned long end)
2673  {
2674  	struct vm_area_struct *next = vma_next(mm, prev);
2675  	struct mmu_gather tlb;
2676  
2677  	lru_add_drain();
2678  	tlb_gather_mmu(&tlb, mm);
2679  	update_hiwater_rss(mm);
2680  	unmap_vmas(&tlb, vma, start, end);
2681  	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2682  				 next ? next->vm_start : USER_PGTABLES_CEILING);
2683  	tlb_finish_mmu(&tlb);
2684  }
2685  
2686  /*
2687   * Create a list of vma's touched by the unmap, removing them from the mm's
2688   * vma list as we go..
2689   */
2690  static bool
2691  detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2692  	struct vm_area_struct *prev, unsigned long end)
2693  {
2694  	struct vm_area_struct **insertion_point;
2695  	struct vm_area_struct *tail_vma = NULL;
2696  
2697  	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2698  	vma->vm_prev = NULL;
2699  	do {
2700  		vma_rb_erase(vma, &mm->mm_rb);
2701  		mm->map_count--;
2702  		tail_vma = vma;
2703  		vma = vma->vm_next;
2704  	} while (vma && vma->vm_start < end);
2705  	*insertion_point = vma;
2706  	if (vma) {
2707  		vma->vm_prev = prev;
2708  		vma_gap_update(vma);
2709  	} else
2710  		mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2711  	tail_vma->vm_next = NULL;
2712  
2713  	/* Kill the cache */
2714  	vmacache_invalidate(mm);
2715  
2716  	/*
2717  	 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2718  	 * VM_GROWSUP VMA. Such VMAs can change their size under
2719  	 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2720  	 */
2721  	if (vma && (vma->vm_flags & VM_GROWSDOWN))
2722  		return false;
2723  	if (prev && (prev->vm_flags & VM_GROWSUP))
2724  		return false;
2725  	return true;
2726  }
2727  
2728  /*
2729   * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2730   * has already been checked or doesn't make sense to fail.
2731   */
2732  int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2733  		unsigned long addr, int new_below)
2734  {
2735  	struct vm_area_struct *new;
2736  	int err;
2737  
2738  	if (vma->vm_ops && vma->vm_ops->may_split) {
2739  		err = vma->vm_ops->may_split(vma, addr);
2740  		if (err)
2741  			return err;
2742  	}
2743  
2744  	new = vm_area_dup(vma);
2745  	if (!new)
2746  		return -ENOMEM;
2747  
2748  	if (new_below)
2749  		new->vm_end = addr;
2750  	else {
2751  		new->vm_start = addr;
2752  		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2753  	}
2754  
2755  	err = vma_dup_policy(vma, new);
2756  	if (err)
2757  		goto out_free_vma;
2758  
2759  	err = anon_vma_clone(new, vma);
2760  	if (err)
2761  		goto out_free_mpol;
2762  
2763  	if (new->vm_file)
2764  		get_file(new->vm_file);
2765  
2766  	if (new->vm_ops && new->vm_ops->open)
2767  		new->vm_ops->open(new);
2768  
2769  	if (new_below)
2770  		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2771  			((addr - new->vm_start) >> PAGE_SHIFT), new);
2772  	else
2773  		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2774  
2775  	/* Success. */
2776  	if (!err)
2777  		return 0;
2778  
2779  	/* Clean everything up if vma_adjust failed. */
2780  	if (new->vm_ops && new->vm_ops->close)
2781  		new->vm_ops->close(new);
2782  	if (new->vm_file)
2783  		fput(new->vm_file);
2784  	unlink_anon_vmas(new);
2785   out_free_mpol:
2786  	mpol_put(vma_policy(new));
2787   out_free_vma:
2788  	vm_area_free(new);
2789  	return err;
2790  }
2791  
2792  /*
2793   * Split a vma into two pieces at address 'addr', a new vma is allocated
2794   * either for the first part or the tail.
2795   */
2796  int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2797  	      unsigned long addr, int new_below)
2798  {
2799  	if (mm->map_count >= sysctl_max_map_count)
2800  		return -ENOMEM;
2801  
2802  	return __split_vma(mm, vma, addr, new_below);
2803  }
2804  
2805  /* Munmap is split into 2 main parts -- this part which finds
2806   * what needs doing, and the areas themselves, which do the
2807   * work.  This now handles partial unmappings.
2808   * Jeremy Fitzhardinge <jeremy@goop.org>
2809   */
2810  int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2811  		struct list_head *uf, bool downgrade)
2812  {
2813  	unsigned long end;
2814  	struct vm_area_struct *vma, *prev, *last;
2815  
2816  	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2817  		return -EINVAL;
2818  
2819  	len = PAGE_ALIGN(len);
2820  	end = start + len;
2821  	if (len == 0)
2822  		return -EINVAL;
2823  
2824  	/*
2825  	 * arch_unmap() might do unmaps itself.  It must be called
2826  	 * and finish any rbtree manipulation before this code
2827  	 * runs and also starts to manipulate the rbtree.
2828  	 */
2829  	arch_unmap(mm, start, end);
2830  
2831  	/* Find the first overlapping VMA */
2832  	vma = find_vma(mm, start);
2833  	if (!vma)
2834  		return 0;
2835  	prev = vma->vm_prev;
2836  	/* we have  start < vma->vm_end  */
2837  
2838  	/* if it doesn't overlap, we have nothing.. */
2839  	if (vma->vm_start >= end)
2840  		return 0;
2841  
2842  	/*
2843  	 * If we need to split any vma, do it now to save pain later.
2844  	 *
2845  	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2846  	 * unmapped vm_area_struct will remain in use: so lower split_vma
2847  	 * places tmp vma above, and higher split_vma places tmp vma below.
2848  	 */
2849  	if (start > vma->vm_start) {
2850  		int error;
2851  
2852  		/*
2853  		 * Make sure that map_count on return from munmap() will
2854  		 * not exceed its limit; but let map_count go just above
2855  		 * its limit temporarily, to help free resources as expected.
2856  		 */
2857  		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2858  			return -ENOMEM;
2859  
2860  		error = __split_vma(mm, vma, start, 0);
2861  		if (error)
2862  			return error;
2863  		prev = vma;
2864  	}
2865  
2866  	/* Does it split the last one? */
2867  	last = find_vma(mm, end);
2868  	if (last && end > last->vm_start) {
2869  		int error = __split_vma(mm, last, end, 1);
2870  		if (error)
2871  			return error;
2872  	}
2873  	vma = vma_next(mm, prev);
2874  
2875  	if (unlikely(uf)) {
2876  		/*
2877  		 * If userfaultfd_unmap_prep returns an error the vmas
2878  		 * will remain split, but userland will get a
2879  		 * highly unexpected error anyway. This is no
2880  		 * different than the case where the first of the two
2881  		 * __split_vma fails, but we don't undo the first
2882  		 * split, despite we could. This is unlikely enough
2883  		 * failure that it's not worth optimizing it for.
2884  		 */
2885  		int error = userfaultfd_unmap_prep(vma, start, end, uf);
2886  		if (error)
2887  			return error;
2888  	}
2889  
2890  	/*
2891  	 * unlock any mlock()ed ranges before detaching vmas
2892  	 */
2893  	if (mm->locked_vm) {
2894  		struct vm_area_struct *tmp = vma;
2895  		while (tmp && tmp->vm_start < end) {
2896  			if (tmp->vm_flags & VM_LOCKED) {
2897  				mm->locked_vm -= vma_pages(tmp);
2898  				munlock_vma_pages_all(tmp);
2899  			}
2900  
2901  			tmp = tmp->vm_next;
2902  		}
2903  	}
2904  
2905  	/* Detach vmas from rbtree */
2906  	if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2907  		downgrade = false;
2908  
2909  	if (downgrade)
2910  		mmap_write_downgrade(mm);
2911  
2912  	unmap_region(mm, vma, prev, start, end);
2913  
2914  	/* Fix up all other VM information */
2915  	remove_vma_list(mm, vma);
2916  
2917  	return downgrade ? 1 : 0;
2918  }
2919  
2920  int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2921  	      struct list_head *uf)
2922  {
2923  	return __do_munmap(mm, start, len, uf, false);
2924  }
2925  
2926  static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2927  {
2928  	int ret;
2929  	struct mm_struct *mm = current->mm;
2930  	LIST_HEAD(uf);
2931  
2932  	if (mmap_write_lock_killable(mm))
2933  		return -EINTR;
2934  
2935  	ret = __do_munmap(mm, start, len, &uf, downgrade);
2936  	/*
2937  	 * Returning 1 indicates mmap_lock is downgraded.
2938  	 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2939  	 * it to 0 before return.
2940  	 */
2941  	if (ret == 1) {
2942  		mmap_read_unlock(mm);
2943  		ret = 0;
2944  	} else
2945  		mmap_write_unlock(mm);
2946  
2947  	userfaultfd_unmap_complete(mm, &uf);
2948  	return ret;
2949  }
2950  
2951  int vm_munmap(unsigned long start, size_t len)
2952  {
2953  	return __vm_munmap(start, len, false);
2954  }
2955  EXPORT_SYMBOL(vm_munmap);
2956  
2957  SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2958  {
2959  	addr = untagged_addr(addr);
2960  	profile_munmap(addr);
2961  	return __vm_munmap(addr, len, true);
2962  }
2963  
2964  
2965  /*
2966   * Emulation of deprecated remap_file_pages() syscall.
2967   */
2968  SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2969  		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2970  {
2971  
2972  	struct mm_struct *mm = current->mm;
2973  	struct vm_area_struct *vma;
2974  	unsigned long populate = 0;
2975  	unsigned long ret = -EINVAL;
2976  	struct file *file;
2977  
2978  	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2979  		     current->comm, current->pid);
2980  
2981  	if (prot)
2982  		return ret;
2983  	start = start & PAGE_MASK;
2984  	size = size & PAGE_MASK;
2985  
2986  	if (start + size <= start)
2987  		return ret;
2988  
2989  	/* Does pgoff wrap? */
2990  	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2991  		return ret;
2992  
2993  	if (mmap_write_lock_killable(mm))
2994  		return -EINTR;
2995  
2996  	vma = find_vma(mm, start);
2997  
2998  	if (!vma || !(vma->vm_flags & VM_SHARED))
2999  		goto out;
3000  
3001  	if (start < vma->vm_start)
3002  		goto out;
3003  
3004  	if (start + size > vma->vm_end) {
3005  		struct vm_area_struct *next;
3006  
3007  		for (next = vma->vm_next; next; next = next->vm_next) {
3008  			/* hole between vmas ? */
3009  			if (next->vm_start != next->vm_prev->vm_end)
3010  				goto out;
3011  
3012  			if (next->vm_file != vma->vm_file)
3013  				goto out;
3014  
3015  			if (next->vm_flags != vma->vm_flags)
3016  				goto out;
3017  
3018  			if (start + size <= next->vm_end)
3019  				break;
3020  		}
3021  
3022  		if (!next)
3023  			goto out;
3024  	}
3025  
3026  	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3027  	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3028  	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3029  
3030  	flags &= MAP_NONBLOCK;
3031  	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3032  	if (vma->vm_flags & VM_LOCKED)
3033  		flags |= MAP_LOCKED;
3034  
3035  	file = get_file(vma->vm_file);
3036  	ret = do_mmap(vma->vm_file, start, size,
3037  			prot, flags, pgoff, &populate, NULL);
3038  	fput(file);
3039  out:
3040  	mmap_write_unlock(mm);
3041  	if (populate)
3042  		mm_populate(ret, populate);
3043  	if (!IS_ERR_VALUE(ret))
3044  		ret = 0;
3045  	return ret;
3046  }
3047  
3048  /*
3049   *  this is really a simplified "do_mmap".  it only handles
3050   *  anonymous maps.  eventually we may be able to do some
3051   *  brk-specific accounting here.
3052   */
3053  static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3054  {
3055  	struct mm_struct *mm = current->mm;
3056  	struct vm_area_struct *vma, *prev;
3057  	struct rb_node **rb_link, *rb_parent;
3058  	pgoff_t pgoff = addr >> PAGE_SHIFT;
3059  	int error;
3060  	unsigned long mapped_addr;
3061  
3062  	/* Until we need other flags, refuse anything except VM_EXEC. */
3063  	if ((flags & (~VM_EXEC)) != 0)
3064  		return -EINVAL;
3065  	flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3066  
3067  	mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3068  	if (IS_ERR_VALUE(mapped_addr))
3069  		return mapped_addr;
3070  
3071  	error = mlock_future_check(mm, mm->def_flags, len);
3072  	if (error)
3073  		return error;
3074  
3075  	/* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3076  	if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3077  		return -ENOMEM;
3078  
3079  	/* Check against address space limits *after* clearing old maps... */
3080  	if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3081  		return -ENOMEM;
3082  
3083  	if (mm->map_count > sysctl_max_map_count)
3084  		return -ENOMEM;
3085  
3086  	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3087  		return -ENOMEM;
3088  
3089  	/* Can we just expand an old private anonymous mapping? */
3090  	vma = vma_merge(mm, prev, addr, addr + len, flags,
3091  			NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3092  	if (vma)
3093  		goto out;
3094  
3095  	/*
3096  	 * create a vma struct for an anonymous mapping
3097  	 */
3098  	vma = vm_area_alloc(mm);
3099  	if (!vma) {
3100  		vm_unacct_memory(len >> PAGE_SHIFT);
3101  		return -ENOMEM;
3102  	}
3103  
3104  	vma_set_anonymous(vma);
3105  	vma->vm_start = addr;
3106  	vma->vm_end = addr + len;
3107  	vma->vm_pgoff = pgoff;
3108  	vma->vm_flags = flags;
3109  	vma->vm_page_prot = vm_get_page_prot(flags);
3110  	vma_link(mm, vma, prev, rb_link, rb_parent);
3111  out:
3112  	perf_event_mmap(vma);
3113  	mm->total_vm += len >> PAGE_SHIFT;
3114  	mm->data_vm += len >> PAGE_SHIFT;
3115  	if (flags & VM_LOCKED)
3116  		mm->locked_vm += (len >> PAGE_SHIFT);
3117  	vma->vm_flags |= VM_SOFTDIRTY;
3118  	return 0;
3119  }
3120  
3121  int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3122  {
3123  	struct mm_struct *mm = current->mm;
3124  	unsigned long len;
3125  	int ret;
3126  	bool populate;
3127  	LIST_HEAD(uf);
3128  
3129  	len = PAGE_ALIGN(request);
3130  	if (len < request)
3131  		return -ENOMEM;
3132  	if (!len)
3133  		return 0;
3134  
3135  	if (mmap_write_lock_killable(mm))
3136  		return -EINTR;
3137  
3138  	ret = do_brk_flags(addr, len, flags, &uf);
3139  	populate = ((mm->def_flags & VM_LOCKED) != 0);
3140  	mmap_write_unlock(mm);
3141  	userfaultfd_unmap_complete(mm, &uf);
3142  	if (populate && !ret)
3143  		mm_populate(addr, len);
3144  	return ret;
3145  }
3146  EXPORT_SYMBOL(vm_brk_flags);
3147  
3148  int vm_brk(unsigned long addr, unsigned long len)
3149  {
3150  	return vm_brk_flags(addr, len, 0);
3151  }
3152  EXPORT_SYMBOL(vm_brk);
3153  
3154  /* Release all mmaps. */
3155  void exit_mmap(struct mm_struct *mm)
3156  {
3157  	struct mmu_gather tlb;
3158  	struct vm_area_struct *vma;
3159  	unsigned long nr_accounted = 0;
3160  
3161  	/* mm's last user has gone, and its about to be pulled down */
3162  	mmu_notifier_release(mm);
3163  
3164  	if (unlikely(mm_is_oom_victim(mm))) {
3165  		/*
3166  		 * Manually reap the mm to free as much memory as possible.
3167  		 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3168  		 * this mm from further consideration.  Taking mm->mmap_lock for
3169  		 * write after setting MMF_OOM_SKIP will guarantee that the oom
3170  		 * reaper will not run on this mm again after mmap_lock is
3171  		 * dropped.
3172  		 *
3173  		 * Nothing can be holding mm->mmap_lock here and the above call
3174  		 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3175  		 * __oom_reap_task_mm() will not block.
3176  		 *
3177  		 * This needs to be done before calling munlock_vma_pages_all(),
3178  		 * which clears VM_LOCKED, otherwise the oom reaper cannot
3179  		 * reliably test it.
3180  		 */
3181  		(void)__oom_reap_task_mm(mm);
3182  
3183  		set_bit(MMF_OOM_SKIP, &mm->flags);
3184  		mmap_write_lock(mm);
3185  		mmap_write_unlock(mm);
3186  	}
3187  
3188  	if (mm->locked_vm) {
3189  		vma = mm->mmap;
3190  		while (vma) {
3191  			if (vma->vm_flags & VM_LOCKED)
3192  				munlock_vma_pages_all(vma);
3193  			vma = vma->vm_next;
3194  		}
3195  	}
3196  
3197  	arch_exit_mmap(mm);
3198  
3199  	vma = mm->mmap;
3200  	if (!vma)	/* Can happen if dup_mmap() received an OOM */
3201  		return;
3202  
3203  	lru_add_drain();
3204  	flush_cache_mm(mm);
3205  	tlb_gather_mmu_fullmm(&tlb, mm);
3206  	/* update_hiwater_rss(mm) here? but nobody should be looking */
3207  	/* Use -1 here to ensure all VMAs in the mm are unmapped */
3208  	unmap_vmas(&tlb, vma, 0, -1);
3209  	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3210  	tlb_finish_mmu(&tlb);
3211  
3212  	/*
3213  	 * Walk the list again, actually closing and freeing it,
3214  	 * with preemption enabled, without holding any MM locks.
3215  	 */
3216  	while (vma) {
3217  		if (vma->vm_flags & VM_ACCOUNT)
3218  			nr_accounted += vma_pages(vma);
3219  		vma = remove_vma(vma);
3220  		cond_resched();
3221  	}
3222  	vm_unacct_memory(nr_accounted);
3223  }
3224  
3225  /* Insert vm structure into process list sorted by address
3226   * and into the inode's i_mmap tree.  If vm_file is non-NULL
3227   * then i_mmap_rwsem is taken here.
3228   */
3229  int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3230  {
3231  	struct vm_area_struct *prev;
3232  	struct rb_node **rb_link, *rb_parent;
3233  
3234  	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3235  			   &prev, &rb_link, &rb_parent))
3236  		return -ENOMEM;
3237  	if ((vma->vm_flags & VM_ACCOUNT) &&
3238  	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
3239  		return -ENOMEM;
3240  
3241  	/*
3242  	 * The vm_pgoff of a purely anonymous vma should be irrelevant
3243  	 * until its first write fault, when page's anon_vma and index
3244  	 * are set.  But now set the vm_pgoff it will almost certainly
3245  	 * end up with (unless mremap moves it elsewhere before that
3246  	 * first wfault), so /proc/pid/maps tells a consistent story.
3247  	 *
3248  	 * By setting it to reflect the virtual start address of the
3249  	 * vma, merges and splits can happen in a seamless way, just
3250  	 * using the existing file pgoff checks and manipulations.
3251  	 * Similarly in do_mmap and in do_brk_flags.
3252  	 */
3253  	if (vma_is_anonymous(vma)) {
3254  		BUG_ON(vma->anon_vma);
3255  		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3256  	}
3257  
3258  	vma_link(mm, vma, prev, rb_link, rb_parent);
3259  	return 0;
3260  }
3261  
3262  /*
3263   * Copy the vma structure to a new location in the same mm,
3264   * prior to moving page table entries, to effect an mremap move.
3265   */
3266  struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3267  	unsigned long addr, unsigned long len, pgoff_t pgoff,
3268  	bool *need_rmap_locks)
3269  {
3270  	struct vm_area_struct *vma = *vmap;
3271  	unsigned long vma_start = vma->vm_start;
3272  	struct mm_struct *mm = vma->vm_mm;
3273  	struct vm_area_struct *new_vma, *prev;
3274  	struct rb_node **rb_link, *rb_parent;
3275  	bool faulted_in_anon_vma = true;
3276  
3277  	/*
3278  	 * If anonymous vma has not yet been faulted, update new pgoff
3279  	 * to match new location, to increase its chance of merging.
3280  	 */
3281  	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3282  		pgoff = addr >> PAGE_SHIFT;
3283  		faulted_in_anon_vma = false;
3284  	}
3285  
3286  	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3287  		return NULL;	/* should never get here */
3288  	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3289  			    vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3290  			    vma->vm_userfaultfd_ctx);
3291  	if (new_vma) {
3292  		/*
3293  		 * Source vma may have been merged into new_vma
3294  		 */
3295  		if (unlikely(vma_start >= new_vma->vm_start &&
3296  			     vma_start < new_vma->vm_end)) {
3297  			/*
3298  			 * The only way we can get a vma_merge with
3299  			 * self during an mremap is if the vma hasn't
3300  			 * been faulted in yet and we were allowed to
3301  			 * reset the dst vma->vm_pgoff to the
3302  			 * destination address of the mremap to allow
3303  			 * the merge to happen. mremap must change the
3304  			 * vm_pgoff linearity between src and dst vmas
3305  			 * (in turn preventing a vma_merge) to be
3306  			 * safe. It is only safe to keep the vm_pgoff
3307  			 * linear if there are no pages mapped yet.
3308  			 */
3309  			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3310  			*vmap = vma = new_vma;
3311  		}
3312  		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3313  	} else {
3314  		new_vma = vm_area_dup(vma);
3315  		if (!new_vma)
3316  			goto out;
3317  		new_vma->vm_start = addr;
3318  		new_vma->vm_end = addr + len;
3319  		new_vma->vm_pgoff = pgoff;
3320  		if (vma_dup_policy(vma, new_vma))
3321  			goto out_free_vma;
3322  		if (anon_vma_clone(new_vma, vma))
3323  			goto out_free_mempol;
3324  		if (new_vma->vm_file)
3325  			get_file(new_vma->vm_file);
3326  		if (new_vma->vm_ops && new_vma->vm_ops->open)
3327  			new_vma->vm_ops->open(new_vma);
3328  		vma_link(mm, new_vma, prev, rb_link, rb_parent);
3329  		*need_rmap_locks = false;
3330  	}
3331  	return new_vma;
3332  
3333  out_free_mempol:
3334  	mpol_put(vma_policy(new_vma));
3335  out_free_vma:
3336  	vm_area_free(new_vma);
3337  out:
3338  	return NULL;
3339  }
3340  
3341  /*
3342   * Return true if the calling process may expand its vm space by the passed
3343   * number of pages
3344   */
3345  bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3346  {
3347  	if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3348  		return false;
3349  
3350  	if (is_data_mapping(flags) &&
3351  	    mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3352  		/* Workaround for Valgrind */
3353  		if (rlimit(RLIMIT_DATA) == 0 &&
3354  		    mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3355  			return true;
3356  
3357  		pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3358  			     current->comm, current->pid,
3359  			     (mm->data_vm + npages) << PAGE_SHIFT,
3360  			     rlimit(RLIMIT_DATA),
3361  			     ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3362  
3363  		if (!ignore_rlimit_data)
3364  			return false;
3365  	}
3366  
3367  	return true;
3368  }
3369  
3370  void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3371  {
3372  	mm->total_vm += npages;
3373  
3374  	if (is_exec_mapping(flags))
3375  		mm->exec_vm += npages;
3376  	else if (is_stack_mapping(flags))
3377  		mm->stack_vm += npages;
3378  	else if (is_data_mapping(flags))
3379  		mm->data_vm += npages;
3380  }
3381  
3382  static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3383  
3384  /*
3385   * Having a close hook prevents vma merging regardless of flags.
3386   */
3387  static void special_mapping_close(struct vm_area_struct *vma)
3388  {
3389  }
3390  
3391  static const char *special_mapping_name(struct vm_area_struct *vma)
3392  {
3393  	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3394  }
3395  
3396  static int special_mapping_mremap(struct vm_area_struct *new_vma)
3397  {
3398  	struct vm_special_mapping *sm = new_vma->vm_private_data;
3399  
3400  	if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3401  		return -EFAULT;
3402  
3403  	if (sm->mremap)
3404  		return sm->mremap(sm, new_vma);
3405  
3406  	return 0;
3407  }
3408  
3409  static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3410  {
3411  	/*
3412  	 * Forbid splitting special mappings - kernel has expectations over
3413  	 * the number of pages in mapping. Together with VM_DONTEXPAND
3414  	 * the size of vma should stay the same over the special mapping's
3415  	 * lifetime.
3416  	 */
3417  	return -EINVAL;
3418  }
3419  
3420  static const struct vm_operations_struct special_mapping_vmops = {
3421  	.close = special_mapping_close,
3422  	.fault = special_mapping_fault,
3423  	.mremap = special_mapping_mremap,
3424  	.name = special_mapping_name,
3425  	/* vDSO code relies that VVAR can't be accessed remotely */
3426  	.access = NULL,
3427  	.may_split = special_mapping_split,
3428  };
3429  
3430  static const struct vm_operations_struct legacy_special_mapping_vmops = {
3431  	.close = special_mapping_close,
3432  	.fault = special_mapping_fault,
3433  };
3434  
3435  static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3436  {
3437  	struct vm_area_struct *vma = vmf->vma;
3438  	pgoff_t pgoff;
3439  	struct page **pages;
3440  
3441  	if (vma->vm_ops == &legacy_special_mapping_vmops) {
3442  		pages = vma->vm_private_data;
3443  	} else {
3444  		struct vm_special_mapping *sm = vma->vm_private_data;
3445  
3446  		if (sm->fault)
3447  			return sm->fault(sm, vmf->vma, vmf);
3448  
3449  		pages = sm->pages;
3450  	}
3451  
3452  	for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3453  		pgoff--;
3454  
3455  	if (*pages) {
3456  		struct page *page = *pages;
3457  		get_page(page);
3458  		vmf->page = page;
3459  		return 0;
3460  	}
3461  
3462  	return VM_FAULT_SIGBUS;
3463  }
3464  
3465  static struct vm_area_struct *__install_special_mapping(
3466  	struct mm_struct *mm,
3467  	unsigned long addr, unsigned long len,
3468  	unsigned long vm_flags, void *priv,
3469  	const struct vm_operations_struct *ops)
3470  {
3471  	int ret;
3472  	struct vm_area_struct *vma;
3473  
3474  	vma = vm_area_alloc(mm);
3475  	if (unlikely(vma == NULL))
3476  		return ERR_PTR(-ENOMEM);
3477  
3478  	vma->vm_start = addr;
3479  	vma->vm_end = addr + len;
3480  
3481  	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3482  	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3483  
3484  	vma->vm_ops = ops;
3485  	vma->vm_private_data = priv;
3486  
3487  	ret = insert_vm_struct(mm, vma);
3488  	if (ret)
3489  		goto out;
3490  
3491  	vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3492  
3493  	perf_event_mmap(vma);
3494  
3495  	return vma;
3496  
3497  out:
3498  	vm_area_free(vma);
3499  	return ERR_PTR(ret);
3500  }
3501  
3502  bool vma_is_special_mapping(const struct vm_area_struct *vma,
3503  	const struct vm_special_mapping *sm)
3504  {
3505  	return vma->vm_private_data == sm &&
3506  		(vma->vm_ops == &special_mapping_vmops ||
3507  		 vma->vm_ops == &legacy_special_mapping_vmops);
3508  }
3509  
3510  /*
3511   * Called with mm->mmap_lock held for writing.
3512   * Insert a new vma covering the given region, with the given flags.
3513   * Its pages are supplied by the given array of struct page *.
3514   * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3515   * The region past the last page supplied will always produce SIGBUS.
3516   * The array pointer and the pages it points to are assumed to stay alive
3517   * for as long as this mapping might exist.
3518   */
3519  struct vm_area_struct *_install_special_mapping(
3520  	struct mm_struct *mm,
3521  	unsigned long addr, unsigned long len,
3522  	unsigned long vm_flags, const struct vm_special_mapping *spec)
3523  {
3524  	return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3525  					&special_mapping_vmops);
3526  }
3527  
3528  int install_special_mapping(struct mm_struct *mm,
3529  			    unsigned long addr, unsigned long len,
3530  			    unsigned long vm_flags, struct page **pages)
3531  {
3532  	struct vm_area_struct *vma = __install_special_mapping(
3533  		mm, addr, len, vm_flags, (void *)pages,
3534  		&legacy_special_mapping_vmops);
3535  
3536  	return PTR_ERR_OR_ZERO(vma);
3537  }
3538  
3539  static DEFINE_MUTEX(mm_all_locks_mutex);
3540  
3541  static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3542  {
3543  	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3544  		/*
3545  		 * The LSB of head.next can't change from under us
3546  		 * because we hold the mm_all_locks_mutex.
3547  		 */
3548  		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3549  		/*
3550  		 * We can safely modify head.next after taking the
3551  		 * anon_vma->root->rwsem. If some other vma in this mm shares
3552  		 * the same anon_vma we won't take it again.
3553  		 *
3554  		 * No need of atomic instructions here, head.next
3555  		 * can't change from under us thanks to the
3556  		 * anon_vma->root->rwsem.
3557  		 */
3558  		if (__test_and_set_bit(0, (unsigned long *)
3559  				       &anon_vma->root->rb_root.rb_root.rb_node))
3560  			BUG();
3561  	}
3562  }
3563  
3564  static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3565  {
3566  	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3567  		/*
3568  		 * AS_MM_ALL_LOCKS can't change from under us because
3569  		 * we hold the mm_all_locks_mutex.
3570  		 *
3571  		 * Operations on ->flags have to be atomic because
3572  		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3573  		 * mm_all_locks_mutex, there may be other cpus
3574  		 * changing other bitflags in parallel to us.
3575  		 */
3576  		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3577  			BUG();
3578  		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3579  	}
3580  }
3581  
3582  /*
3583   * This operation locks against the VM for all pte/vma/mm related
3584   * operations that could ever happen on a certain mm. This includes
3585   * vmtruncate, try_to_unmap, and all page faults.
3586   *
3587   * The caller must take the mmap_lock in write mode before calling
3588   * mm_take_all_locks(). The caller isn't allowed to release the
3589   * mmap_lock until mm_drop_all_locks() returns.
3590   *
3591   * mmap_lock in write mode is required in order to block all operations
3592   * that could modify pagetables and free pages without need of
3593   * altering the vma layout. It's also needed in write mode to avoid new
3594   * anon_vmas to be associated with existing vmas.
3595   *
3596   * A single task can't take more than one mm_take_all_locks() in a row
3597   * or it would deadlock.
3598   *
3599   * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3600   * mapping->flags avoid to take the same lock twice, if more than one
3601   * vma in this mm is backed by the same anon_vma or address_space.
3602   *
3603   * We take locks in following order, accordingly to comment at beginning
3604   * of mm/rmap.c:
3605   *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3606   *     hugetlb mapping);
3607   *   - all i_mmap_rwsem locks;
3608   *   - all anon_vma->rwseml
3609   *
3610   * We can take all locks within these types randomly because the VM code
3611   * doesn't nest them and we protected from parallel mm_take_all_locks() by
3612   * mm_all_locks_mutex.
3613   *
3614   * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3615   * that may have to take thousand of locks.
3616   *
3617   * mm_take_all_locks() can fail if it's interrupted by signals.
3618   */
3619  int mm_take_all_locks(struct mm_struct *mm)
3620  {
3621  	struct vm_area_struct *vma;
3622  	struct anon_vma_chain *avc;
3623  
3624  	BUG_ON(mmap_read_trylock(mm));
3625  
3626  	mutex_lock(&mm_all_locks_mutex);
3627  
3628  	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3629  		if (signal_pending(current))
3630  			goto out_unlock;
3631  		if (vma->vm_file && vma->vm_file->f_mapping &&
3632  				is_vm_hugetlb_page(vma))
3633  			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3634  	}
3635  
3636  	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3637  		if (signal_pending(current))
3638  			goto out_unlock;
3639  		if (vma->vm_file && vma->vm_file->f_mapping &&
3640  				!is_vm_hugetlb_page(vma))
3641  			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3642  	}
3643  
3644  	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3645  		if (signal_pending(current))
3646  			goto out_unlock;
3647  		if (vma->anon_vma)
3648  			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3649  				vm_lock_anon_vma(mm, avc->anon_vma);
3650  	}
3651  
3652  	return 0;
3653  
3654  out_unlock:
3655  	mm_drop_all_locks(mm);
3656  	return -EINTR;
3657  }
3658  
3659  static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3660  {
3661  	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3662  		/*
3663  		 * The LSB of head.next can't change to 0 from under
3664  		 * us because we hold the mm_all_locks_mutex.
3665  		 *
3666  		 * We must however clear the bitflag before unlocking
3667  		 * the vma so the users using the anon_vma->rb_root will
3668  		 * never see our bitflag.
3669  		 *
3670  		 * No need of atomic instructions here, head.next
3671  		 * can't change from under us until we release the
3672  		 * anon_vma->root->rwsem.
3673  		 */
3674  		if (!__test_and_clear_bit(0, (unsigned long *)
3675  					  &anon_vma->root->rb_root.rb_root.rb_node))
3676  			BUG();
3677  		anon_vma_unlock_write(anon_vma);
3678  	}
3679  }
3680  
3681  static void vm_unlock_mapping(struct address_space *mapping)
3682  {
3683  	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3684  		/*
3685  		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3686  		 * because we hold the mm_all_locks_mutex.
3687  		 */
3688  		i_mmap_unlock_write(mapping);
3689  		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3690  					&mapping->flags))
3691  			BUG();
3692  	}
3693  }
3694  
3695  /*
3696   * The mmap_lock cannot be released by the caller until
3697   * mm_drop_all_locks() returns.
3698   */
3699  void mm_drop_all_locks(struct mm_struct *mm)
3700  {
3701  	struct vm_area_struct *vma;
3702  	struct anon_vma_chain *avc;
3703  
3704  	BUG_ON(mmap_read_trylock(mm));
3705  	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3706  
3707  	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3708  		if (vma->anon_vma)
3709  			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3710  				vm_unlock_anon_vma(avc->anon_vma);
3711  		if (vma->vm_file && vma->vm_file->f_mapping)
3712  			vm_unlock_mapping(vma->vm_file->f_mapping);
3713  	}
3714  
3715  	mutex_unlock(&mm_all_locks_mutex);
3716  }
3717  
3718  /*
3719   * initialise the percpu counter for VM
3720   */
3721  void __init mmap_init(void)
3722  {
3723  	int ret;
3724  
3725  	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3726  	VM_BUG_ON(ret);
3727  }
3728  
3729  /*
3730   * Initialise sysctl_user_reserve_kbytes.
3731   *
3732   * This is intended to prevent a user from starting a single memory hogging
3733   * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3734   * mode.
3735   *
3736   * The default value is min(3% of free memory, 128MB)
3737   * 128MB is enough to recover with sshd/login, bash, and top/kill.
3738   */
3739  static int init_user_reserve(void)
3740  {
3741  	unsigned long free_kbytes;
3742  
3743  	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3744  
3745  	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3746  	return 0;
3747  }
3748  subsys_initcall(init_user_reserve);
3749  
3750  /*
3751   * Initialise sysctl_admin_reserve_kbytes.
3752   *
3753   * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3754   * to log in and kill a memory hogging process.
3755   *
3756   * Systems with more than 256MB will reserve 8MB, enough to recover
3757   * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3758   * only reserve 3% of free pages by default.
3759   */
3760  static int init_admin_reserve(void)
3761  {
3762  	unsigned long free_kbytes;
3763  
3764  	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3765  
3766  	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3767  	return 0;
3768  }
3769  subsys_initcall(init_admin_reserve);
3770  
3771  /*
3772   * Reinititalise user and admin reserves if memory is added or removed.
3773   *
3774   * The default user reserve max is 128MB, and the default max for the
3775   * admin reserve is 8MB. These are usually, but not always, enough to
3776   * enable recovery from a memory hogging process using login/sshd, a shell,
3777   * and tools like top. It may make sense to increase or even disable the
3778   * reserve depending on the existence of swap or variations in the recovery
3779   * tools. So, the admin may have changed them.
3780   *
3781   * If memory is added and the reserves have been eliminated or increased above
3782   * the default max, then we'll trust the admin.
3783   *
3784   * If memory is removed and there isn't enough free memory, then we
3785   * need to reset the reserves.
3786   *
3787   * Otherwise keep the reserve set by the admin.
3788   */
3789  static int reserve_mem_notifier(struct notifier_block *nb,
3790  			     unsigned long action, void *data)
3791  {
3792  	unsigned long tmp, free_kbytes;
3793  
3794  	switch (action) {
3795  	case MEM_ONLINE:
3796  		/* Default max is 128MB. Leave alone if modified by operator. */
3797  		tmp = sysctl_user_reserve_kbytes;
3798  		if (0 < tmp && tmp < (1UL << 17))
3799  			init_user_reserve();
3800  
3801  		/* Default max is 8MB.  Leave alone if modified by operator. */
3802  		tmp = sysctl_admin_reserve_kbytes;
3803  		if (0 < tmp && tmp < (1UL << 13))
3804  			init_admin_reserve();
3805  
3806  		break;
3807  	case MEM_OFFLINE:
3808  		free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3809  
3810  		if (sysctl_user_reserve_kbytes > free_kbytes) {
3811  			init_user_reserve();
3812  			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3813  				sysctl_user_reserve_kbytes);
3814  		}
3815  
3816  		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3817  			init_admin_reserve();
3818  			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3819  				sysctl_admin_reserve_kbytes);
3820  		}
3821  		break;
3822  	default:
3823  		break;
3824  	}
3825  	return NOTIFY_OK;
3826  }
3827  
3828  static struct notifier_block reserve_mem_nb = {
3829  	.notifier_call = reserve_mem_notifier,
3830  };
3831  
3832  static int __meminit init_reserve_notifier(void)
3833  {
3834  	if (register_hotmemory_notifier(&reserve_mem_nb))
3835  		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3836  
3837  	return 0;
3838  }
3839  subsys_initcall(init_reserve_notifier);
3840