Loading...
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/mm_inline.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/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48#include <linux/sched/mm.h>
49#include <linux/ksm.h>
50#include <linux/memfd.h>
51
52#include <linux/uaccess.h>
53#include <asm/cacheflush.h>
54#include <asm/tlb.h>
55#include <asm/mmu_context.h>
56
57#define CREATE_TRACE_POINTS
58#include <trace/events/mmap.h>
59
60#include "internal.h"
61
62#ifndef arch_mmap_check
63#define arch_mmap_check(addr, len, flags) (0)
64#endif
65
66#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
67const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
68int mmap_rnd_bits_max __ro_after_init = CONFIG_ARCH_MMAP_RND_BITS_MAX;
69int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
70#endif
71#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
72const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
73const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
74int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
75#endif
76
77static bool ignore_rlimit_data;
78core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
79
80/* Update vma->vm_page_prot to reflect vma->vm_flags. */
81void vma_set_page_prot(struct vm_area_struct *vma)
82{
83 unsigned long vm_flags = vma->vm_flags;
84 pgprot_t vm_page_prot;
85
86 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
87 if (vma_wants_writenotify(vma, vm_page_prot)) {
88 vm_flags &= ~VM_SHARED;
89 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
90 }
91 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
92 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
93}
94
95/*
96 * check_brk_limits() - Use platform specific check of range & verify mlock
97 * limits.
98 * @addr: The address to check
99 * @len: The size of increase.
100 *
101 * Return: 0 on success.
102 */
103static int check_brk_limits(unsigned long addr, unsigned long len)
104{
105 unsigned long mapped_addr;
106
107 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
108 if (IS_ERR_VALUE(mapped_addr))
109 return mapped_addr;
110
111 return mlock_future_ok(current->mm, current->mm->def_flags, len)
112 ? 0 : -EAGAIN;
113}
114static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
115 unsigned long addr, unsigned long request, unsigned long flags);
116SYSCALL_DEFINE1(brk, unsigned long, brk)
117{
118 unsigned long newbrk, oldbrk, origbrk;
119 struct mm_struct *mm = current->mm;
120 struct vm_area_struct *brkvma, *next = NULL;
121 unsigned long min_brk;
122 bool populate = false;
123 LIST_HEAD(uf);
124 struct vma_iterator vmi;
125
126 if (mmap_write_lock_killable(mm))
127 return -EINTR;
128
129 origbrk = mm->brk;
130
131#ifdef CONFIG_COMPAT_BRK
132 /*
133 * CONFIG_COMPAT_BRK can still be overridden by setting
134 * randomize_va_space to 2, which will still cause mm->start_brk
135 * to be arbitrarily shifted
136 */
137 if (current->brk_randomized)
138 min_brk = mm->start_brk;
139 else
140 min_brk = mm->end_data;
141#else
142 min_brk = mm->start_brk;
143#endif
144 if (brk < min_brk)
145 goto out;
146
147 /*
148 * Check against rlimit here. If this check is done later after the test
149 * of oldbrk with newbrk then it can escape the test and let the data
150 * segment grow beyond its set limit the in case where the limit is
151 * not page aligned -Ram Gupta
152 */
153 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
154 mm->end_data, mm->start_data))
155 goto out;
156
157 newbrk = PAGE_ALIGN(brk);
158 oldbrk = PAGE_ALIGN(mm->brk);
159 if (oldbrk == newbrk) {
160 mm->brk = brk;
161 goto success;
162 }
163
164 /* Always allow shrinking brk. */
165 if (brk <= mm->brk) {
166 /* Search one past newbrk */
167 vma_iter_init(&vmi, mm, newbrk);
168 brkvma = vma_find(&vmi, oldbrk);
169 if (!brkvma || brkvma->vm_start >= oldbrk)
170 goto out; /* mapping intersects with an existing non-brk vma. */
171 /*
172 * mm->brk must be protected by write mmap_lock.
173 * do_vmi_align_munmap() will drop the lock on success, so
174 * update it before calling do_vma_munmap().
175 */
176 mm->brk = brk;
177 if (do_vmi_align_munmap(&vmi, brkvma, mm, newbrk, oldbrk, &uf,
178 /* unlock = */ true))
179 goto out;
180
181 goto success_unlocked;
182 }
183
184 if (check_brk_limits(oldbrk, newbrk - oldbrk))
185 goto out;
186
187 /*
188 * Only check if the next VMA is within the stack_guard_gap of the
189 * expansion area
190 */
191 vma_iter_init(&vmi, mm, oldbrk);
192 next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
193 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
194 goto out;
195
196 brkvma = vma_prev_limit(&vmi, mm->start_brk);
197 /* Ok, looks good - let it rip. */
198 if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
199 goto out;
200
201 mm->brk = brk;
202 if (mm->def_flags & VM_LOCKED)
203 populate = true;
204
205success:
206 mmap_write_unlock(mm);
207success_unlocked:
208 userfaultfd_unmap_complete(mm, &uf);
209 if (populate)
210 mm_populate(oldbrk, newbrk - oldbrk);
211 return brk;
212
213out:
214 mm->brk = origbrk;
215 mmap_write_unlock(mm);
216 return origbrk;
217}
218
219/*
220 * If a hint addr is less than mmap_min_addr change hint to be as
221 * low as possible but still greater than mmap_min_addr
222 */
223static inline unsigned long round_hint_to_min(unsigned long hint)
224{
225 hint &= PAGE_MASK;
226 if (((void *)hint != NULL) &&
227 (hint < mmap_min_addr))
228 return PAGE_ALIGN(mmap_min_addr);
229 return hint;
230}
231
232bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
233 unsigned long bytes)
234{
235 unsigned long locked_pages, limit_pages;
236
237 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
238 return true;
239
240 locked_pages = bytes >> PAGE_SHIFT;
241 locked_pages += mm->locked_vm;
242
243 limit_pages = rlimit(RLIMIT_MEMLOCK);
244 limit_pages >>= PAGE_SHIFT;
245
246 return locked_pages <= limit_pages;
247}
248
249static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
250{
251 if (S_ISREG(inode->i_mode))
252 return MAX_LFS_FILESIZE;
253
254 if (S_ISBLK(inode->i_mode))
255 return MAX_LFS_FILESIZE;
256
257 if (S_ISSOCK(inode->i_mode))
258 return MAX_LFS_FILESIZE;
259
260 /* Special "we do even unsigned file positions" case */
261 if (file->f_op->fop_flags & FOP_UNSIGNED_OFFSET)
262 return 0;
263
264 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
265 return ULONG_MAX;
266}
267
268static inline bool file_mmap_ok(struct file *file, struct inode *inode,
269 unsigned long pgoff, unsigned long len)
270{
271 u64 maxsize = file_mmap_size_max(file, inode);
272
273 if (maxsize && len > maxsize)
274 return false;
275 maxsize -= len;
276 if (pgoff > maxsize >> PAGE_SHIFT)
277 return false;
278 return true;
279}
280
281/*
282 * The caller must write-lock current->mm->mmap_lock.
283 */
284unsigned long do_mmap(struct file *file, unsigned long addr,
285 unsigned long len, unsigned long prot,
286 unsigned long flags, vm_flags_t vm_flags,
287 unsigned long pgoff, unsigned long *populate,
288 struct list_head *uf)
289{
290 struct mm_struct *mm = current->mm;
291 int pkey = 0;
292
293 *populate = 0;
294
295 if (!len)
296 return -EINVAL;
297
298 /*
299 * Does the application expect PROT_READ to imply PROT_EXEC?
300 *
301 * (the exception is when the underlying filesystem is noexec
302 * mounted, in which case we don't add PROT_EXEC.)
303 */
304 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
305 if (!(file && path_noexec(&file->f_path)))
306 prot |= PROT_EXEC;
307
308 /* force arch specific MAP_FIXED handling in get_unmapped_area */
309 if (flags & MAP_FIXED_NOREPLACE)
310 flags |= MAP_FIXED;
311
312 if (!(flags & MAP_FIXED))
313 addr = round_hint_to_min(addr);
314
315 /* Careful about overflows.. */
316 len = PAGE_ALIGN(len);
317 if (!len)
318 return -ENOMEM;
319
320 /* offset overflow? */
321 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
322 return -EOVERFLOW;
323
324 /* Too many mappings? */
325 if (mm->map_count > sysctl_max_map_count)
326 return -ENOMEM;
327
328 /*
329 * addr is returned from get_unmapped_area,
330 * There are two cases:
331 * 1> MAP_FIXED == false
332 * unallocated memory, no need to check sealing.
333 * 1> MAP_FIXED == true
334 * sealing is checked inside mmap_region when
335 * do_vmi_munmap is called.
336 */
337
338 if (prot == PROT_EXEC) {
339 pkey = execute_only_pkey(mm);
340 if (pkey < 0)
341 pkey = 0;
342 }
343
344 /* Do simple checking here so the lower-level routines won't have
345 * to. we assume access permissions have been handled by the open
346 * of the memory object, so we don't do any here.
347 */
348 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(file, flags) |
349 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
350
351 /* Obtain the address to map to. we verify (or select) it and ensure
352 * that it represents a valid section of the address space.
353 */
354 addr = __get_unmapped_area(file, addr, len, pgoff, flags, vm_flags);
355 if (IS_ERR_VALUE(addr))
356 return addr;
357
358 if (flags & MAP_FIXED_NOREPLACE) {
359 if (find_vma_intersection(mm, addr, addr + len))
360 return -EEXIST;
361 }
362
363 if (flags & MAP_LOCKED)
364 if (!can_do_mlock())
365 return -EPERM;
366
367 if (!mlock_future_ok(mm, vm_flags, len))
368 return -EAGAIN;
369
370 if (file) {
371 struct inode *inode = file_inode(file);
372 unsigned int seals = memfd_file_seals(file);
373 unsigned long flags_mask;
374
375 if (!file_mmap_ok(file, inode, pgoff, len))
376 return -EOVERFLOW;
377
378 flags_mask = LEGACY_MAP_MASK;
379 if (file->f_op->fop_flags & FOP_MMAP_SYNC)
380 flags_mask |= MAP_SYNC;
381
382 switch (flags & MAP_TYPE) {
383 case MAP_SHARED:
384 /*
385 * Force use of MAP_SHARED_VALIDATE with non-legacy
386 * flags. E.g. MAP_SYNC is dangerous to use with
387 * MAP_SHARED as you don't know which consistency model
388 * you will get. We silently ignore unsupported flags
389 * with MAP_SHARED to preserve backward compatibility.
390 */
391 flags &= LEGACY_MAP_MASK;
392 fallthrough;
393 case MAP_SHARED_VALIDATE:
394 if (flags & ~flags_mask)
395 return -EOPNOTSUPP;
396 if (prot & PROT_WRITE) {
397 if (!(file->f_mode & FMODE_WRITE))
398 return -EACCES;
399 if (IS_SWAPFILE(file->f_mapping->host))
400 return -ETXTBSY;
401 }
402
403 /*
404 * Make sure we don't allow writing to an append-only
405 * file..
406 */
407 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
408 return -EACCES;
409
410 vm_flags |= VM_SHARED | VM_MAYSHARE;
411 if (!(file->f_mode & FMODE_WRITE))
412 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
413 else if (is_readonly_sealed(seals, vm_flags))
414 vm_flags &= ~VM_MAYWRITE;
415 fallthrough;
416 case MAP_PRIVATE:
417 if (!(file->f_mode & FMODE_READ))
418 return -EACCES;
419 if (path_noexec(&file->f_path)) {
420 if (vm_flags & VM_EXEC)
421 return -EPERM;
422 vm_flags &= ~VM_MAYEXEC;
423 }
424
425 if (!file->f_op->mmap)
426 return -ENODEV;
427 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
428 return -EINVAL;
429 break;
430
431 default:
432 return -EINVAL;
433 }
434 } else {
435 switch (flags & MAP_TYPE) {
436 case MAP_SHARED:
437 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
438 return -EINVAL;
439 /*
440 * Ignore pgoff.
441 */
442 pgoff = 0;
443 vm_flags |= VM_SHARED | VM_MAYSHARE;
444 break;
445 case MAP_DROPPABLE:
446 if (VM_DROPPABLE == VM_NONE)
447 return -ENOTSUPP;
448 /*
449 * A locked or stack area makes no sense to be droppable.
450 *
451 * Also, since droppable pages can just go away at any time
452 * it makes no sense to copy them on fork or dump them.
453 *
454 * And don't attempt to combine with hugetlb for now.
455 */
456 if (flags & (MAP_LOCKED | MAP_HUGETLB))
457 return -EINVAL;
458 if (vm_flags & (VM_GROWSDOWN | VM_GROWSUP))
459 return -EINVAL;
460
461 vm_flags |= VM_DROPPABLE;
462
463 /*
464 * If the pages can be dropped, then it doesn't make
465 * sense to reserve them.
466 */
467 vm_flags |= VM_NORESERVE;
468
469 /*
470 * Likewise, they're volatile enough that they
471 * shouldn't survive forks or coredumps.
472 */
473 vm_flags |= VM_WIPEONFORK | VM_DONTDUMP;
474 fallthrough;
475 case MAP_PRIVATE:
476 /*
477 * Set pgoff according to addr for anon_vma.
478 */
479 pgoff = addr >> PAGE_SHIFT;
480 break;
481 default:
482 return -EINVAL;
483 }
484 }
485
486 /*
487 * Set 'VM_NORESERVE' if we should not account for the
488 * memory use of this mapping.
489 */
490 if (flags & MAP_NORESERVE) {
491 /* We honor MAP_NORESERVE if allowed to overcommit */
492 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
493 vm_flags |= VM_NORESERVE;
494
495 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
496 if (file && is_file_hugepages(file))
497 vm_flags |= VM_NORESERVE;
498 }
499
500 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
501 if (!IS_ERR_VALUE(addr) &&
502 ((vm_flags & VM_LOCKED) ||
503 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
504 *populate = len;
505 return addr;
506}
507
508unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
509 unsigned long prot, unsigned long flags,
510 unsigned long fd, unsigned long pgoff)
511{
512 struct file *file = NULL;
513 unsigned long retval;
514
515 if (!(flags & MAP_ANONYMOUS)) {
516 audit_mmap_fd(fd, flags);
517 file = fget(fd);
518 if (!file)
519 return -EBADF;
520 if (is_file_hugepages(file)) {
521 len = ALIGN(len, huge_page_size(hstate_file(file)));
522 } else if (unlikely(flags & MAP_HUGETLB)) {
523 retval = -EINVAL;
524 goto out_fput;
525 }
526 } else if (flags & MAP_HUGETLB) {
527 struct hstate *hs;
528
529 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
530 if (!hs)
531 return -EINVAL;
532
533 len = ALIGN(len, huge_page_size(hs));
534 /*
535 * VM_NORESERVE is used because the reservations will be
536 * taken when vm_ops->mmap() is called
537 */
538 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
539 VM_NORESERVE,
540 HUGETLB_ANONHUGE_INODE,
541 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
542 if (IS_ERR(file))
543 return PTR_ERR(file);
544 }
545
546 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
547out_fput:
548 if (file)
549 fput(file);
550 return retval;
551}
552
553SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
554 unsigned long, prot, unsigned long, flags,
555 unsigned long, fd, unsigned long, pgoff)
556{
557 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
558}
559
560#ifdef __ARCH_WANT_SYS_OLD_MMAP
561struct mmap_arg_struct {
562 unsigned long addr;
563 unsigned long len;
564 unsigned long prot;
565 unsigned long flags;
566 unsigned long fd;
567 unsigned long offset;
568};
569
570SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
571{
572 struct mmap_arg_struct a;
573
574 if (copy_from_user(&a, arg, sizeof(a)))
575 return -EFAULT;
576 if (offset_in_page(a.offset))
577 return -EINVAL;
578
579 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
580 a.offset >> PAGE_SHIFT);
581}
582#endif /* __ARCH_WANT_SYS_OLD_MMAP */
583
584/**
585 * unmapped_area() - Find an area between the low_limit and the high_limit with
586 * the correct alignment and offset, all from @info. Note: current->mm is used
587 * for the search.
588 *
589 * @info: The unmapped area information including the range [low_limit -
590 * high_limit), the alignment offset and mask.
591 *
592 * Return: A memory address or -ENOMEM.
593 */
594static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
595{
596 unsigned long length, gap;
597 unsigned long low_limit, high_limit;
598 struct vm_area_struct *tmp;
599 VMA_ITERATOR(vmi, current->mm, 0);
600
601 /* Adjust search length to account for worst case alignment overhead */
602 length = info->length + info->align_mask + info->start_gap;
603 if (length < info->length)
604 return -ENOMEM;
605
606 low_limit = info->low_limit;
607 if (low_limit < mmap_min_addr)
608 low_limit = mmap_min_addr;
609 high_limit = info->high_limit;
610retry:
611 if (vma_iter_area_lowest(&vmi, low_limit, high_limit, length))
612 return -ENOMEM;
613
614 /*
615 * Adjust for the gap first so it doesn't interfere with the
616 * later alignment. The first step is the minimum needed to
617 * fulill the start gap, the next steps is the minimum to align
618 * that. It is the minimum needed to fulill both.
619 */
620 gap = vma_iter_addr(&vmi) + info->start_gap;
621 gap += (info->align_offset - gap) & info->align_mask;
622 tmp = vma_next(&vmi);
623 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
624 if (vm_start_gap(tmp) < gap + length - 1) {
625 low_limit = tmp->vm_end;
626 vma_iter_reset(&vmi);
627 goto retry;
628 }
629 } else {
630 tmp = vma_prev(&vmi);
631 if (tmp && vm_end_gap(tmp) > gap) {
632 low_limit = vm_end_gap(tmp);
633 vma_iter_reset(&vmi);
634 goto retry;
635 }
636 }
637
638 return gap;
639}
640
641/**
642 * unmapped_area_topdown() - Find an area between the low_limit and the
643 * high_limit with the correct alignment and offset at the highest available
644 * address, all from @info. Note: current->mm is used for the search.
645 *
646 * @info: The unmapped area information including the range [low_limit -
647 * high_limit), the alignment offset and mask.
648 *
649 * Return: A memory address or -ENOMEM.
650 */
651static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
652{
653 unsigned long length, gap, gap_end;
654 unsigned long low_limit, high_limit;
655 struct vm_area_struct *tmp;
656 VMA_ITERATOR(vmi, current->mm, 0);
657
658 /* Adjust search length to account for worst case alignment overhead */
659 length = info->length + info->align_mask + info->start_gap;
660 if (length < info->length)
661 return -ENOMEM;
662
663 low_limit = info->low_limit;
664 if (low_limit < mmap_min_addr)
665 low_limit = mmap_min_addr;
666 high_limit = info->high_limit;
667retry:
668 if (vma_iter_area_highest(&vmi, low_limit, high_limit, length))
669 return -ENOMEM;
670
671 gap = vma_iter_end(&vmi) - info->length;
672 gap -= (gap - info->align_offset) & info->align_mask;
673 gap_end = vma_iter_end(&vmi);
674 tmp = vma_next(&vmi);
675 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
676 if (vm_start_gap(tmp) < gap_end) {
677 high_limit = vm_start_gap(tmp);
678 vma_iter_reset(&vmi);
679 goto retry;
680 }
681 } else {
682 tmp = vma_prev(&vmi);
683 if (tmp && vm_end_gap(tmp) > gap) {
684 high_limit = tmp->vm_start;
685 vma_iter_reset(&vmi);
686 goto retry;
687 }
688 }
689
690 return gap;
691}
692
693/*
694 * Determine if the allocation needs to ensure that there is no
695 * existing mapping within it's guard gaps, for use as start_gap.
696 */
697static inline unsigned long stack_guard_placement(vm_flags_t vm_flags)
698{
699 if (vm_flags & VM_SHADOW_STACK)
700 return PAGE_SIZE;
701
702 return 0;
703}
704
705/*
706 * Search for an unmapped address range.
707 *
708 * We are looking for a range that:
709 * - does not intersect with any VMA;
710 * - is contained within the [low_limit, high_limit) interval;
711 * - is at least the desired size.
712 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
713 */
714unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
715{
716 unsigned long addr;
717
718 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
719 addr = unmapped_area_topdown(info);
720 else
721 addr = unmapped_area(info);
722
723 trace_vm_unmapped_area(addr, info);
724 return addr;
725}
726
727/* Get an address range which is currently unmapped.
728 * For shmat() with addr=0.
729 *
730 * Ugly calling convention alert:
731 * Return value with the low bits set means error value,
732 * ie
733 * if (ret & ~PAGE_MASK)
734 * error = ret;
735 *
736 * This function "knows" that -ENOMEM has the bits set.
737 */
738unsigned long
739generic_get_unmapped_area(struct file *filp, unsigned long addr,
740 unsigned long len, unsigned long pgoff,
741 unsigned long flags, vm_flags_t vm_flags)
742{
743 struct mm_struct *mm = current->mm;
744 struct vm_area_struct *vma, *prev;
745 struct vm_unmapped_area_info info = {};
746 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
747
748 if (len > mmap_end - mmap_min_addr)
749 return -ENOMEM;
750
751 if (flags & MAP_FIXED)
752 return addr;
753
754 if (addr) {
755 addr = PAGE_ALIGN(addr);
756 vma = find_vma_prev(mm, addr, &prev);
757 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
758 (!vma || addr + len <= vm_start_gap(vma)) &&
759 (!prev || addr >= vm_end_gap(prev)))
760 return addr;
761 }
762
763 info.length = len;
764 info.low_limit = mm->mmap_base;
765 info.high_limit = mmap_end;
766 info.start_gap = stack_guard_placement(vm_flags);
767 if (filp && is_file_hugepages(filp))
768 info.align_mask = huge_page_mask_align(filp);
769 return vm_unmapped_area(&info);
770}
771
772#ifndef HAVE_ARCH_UNMAPPED_AREA
773unsigned long
774arch_get_unmapped_area(struct file *filp, unsigned long addr,
775 unsigned long len, unsigned long pgoff,
776 unsigned long flags, vm_flags_t vm_flags)
777{
778 return generic_get_unmapped_area(filp, addr, len, pgoff, flags,
779 vm_flags);
780}
781#endif
782
783/*
784 * This mmap-allocator allocates new areas top-down from below the
785 * stack's low limit (the base):
786 */
787unsigned long
788generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
789 unsigned long len, unsigned long pgoff,
790 unsigned long flags, vm_flags_t vm_flags)
791{
792 struct vm_area_struct *vma, *prev;
793 struct mm_struct *mm = current->mm;
794 struct vm_unmapped_area_info info = {};
795 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
796
797 /* requested length too big for entire address space */
798 if (len > mmap_end - mmap_min_addr)
799 return -ENOMEM;
800
801 if (flags & MAP_FIXED)
802 return addr;
803
804 /* requesting a specific address */
805 if (addr) {
806 addr = PAGE_ALIGN(addr);
807 vma = find_vma_prev(mm, addr, &prev);
808 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
809 (!vma || addr + len <= vm_start_gap(vma)) &&
810 (!prev || addr >= vm_end_gap(prev)))
811 return addr;
812 }
813
814 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
815 info.length = len;
816 info.low_limit = PAGE_SIZE;
817 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
818 info.start_gap = stack_guard_placement(vm_flags);
819 if (filp && is_file_hugepages(filp))
820 info.align_mask = huge_page_mask_align(filp);
821 addr = vm_unmapped_area(&info);
822
823 /*
824 * A failed mmap() very likely causes application failure,
825 * so fall back to the bottom-up function here. This scenario
826 * can happen with large stack limits and large mmap()
827 * allocations.
828 */
829 if (offset_in_page(addr)) {
830 VM_BUG_ON(addr != -ENOMEM);
831 info.flags = 0;
832 info.low_limit = TASK_UNMAPPED_BASE;
833 info.high_limit = mmap_end;
834 addr = vm_unmapped_area(&info);
835 }
836
837 return addr;
838}
839
840#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
841unsigned long
842arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
843 unsigned long len, unsigned long pgoff,
844 unsigned long flags, vm_flags_t vm_flags)
845{
846 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags,
847 vm_flags);
848}
849#endif
850
851unsigned long mm_get_unmapped_area_vmflags(struct mm_struct *mm, struct file *filp,
852 unsigned long addr, unsigned long len,
853 unsigned long pgoff, unsigned long flags,
854 vm_flags_t vm_flags)
855{
856 if (test_bit(MMF_TOPDOWN, &mm->flags))
857 return arch_get_unmapped_area_topdown(filp, addr, len, pgoff,
858 flags, vm_flags);
859 return arch_get_unmapped_area(filp, addr, len, pgoff, flags, vm_flags);
860}
861
862unsigned long
863__get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
864 unsigned long pgoff, unsigned long flags, vm_flags_t vm_flags)
865{
866 unsigned long (*get_area)(struct file *, unsigned long,
867 unsigned long, unsigned long, unsigned long)
868 = NULL;
869
870 unsigned long error = arch_mmap_check(addr, len, flags);
871 if (error)
872 return error;
873
874 /* Careful about overflows.. */
875 if (len > TASK_SIZE)
876 return -ENOMEM;
877
878 if (file) {
879 if (file->f_op->get_unmapped_area)
880 get_area = file->f_op->get_unmapped_area;
881 } else if (flags & MAP_SHARED) {
882 /*
883 * mmap_region() will call shmem_zero_setup() to create a file,
884 * so use shmem's get_unmapped_area in case it can be huge.
885 */
886 get_area = shmem_get_unmapped_area;
887 }
888
889 /* Always treat pgoff as zero for anonymous memory. */
890 if (!file)
891 pgoff = 0;
892
893 if (get_area) {
894 addr = get_area(file, addr, len, pgoff, flags);
895 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && !file
896 && !addr /* no hint */
897 && IS_ALIGNED(len, PMD_SIZE)) {
898 /* Ensures that larger anonymous mappings are THP aligned. */
899 addr = thp_get_unmapped_area_vmflags(file, addr, len,
900 pgoff, flags, vm_flags);
901 } else {
902 addr = mm_get_unmapped_area_vmflags(current->mm, file, addr, len,
903 pgoff, flags, vm_flags);
904 }
905 if (IS_ERR_VALUE(addr))
906 return addr;
907
908 if (addr > TASK_SIZE - len)
909 return -ENOMEM;
910 if (offset_in_page(addr))
911 return -EINVAL;
912
913 error = security_mmap_addr(addr);
914 return error ? error : addr;
915}
916
917unsigned long
918mm_get_unmapped_area(struct mm_struct *mm, struct file *file,
919 unsigned long addr, unsigned long len,
920 unsigned long pgoff, unsigned long flags)
921{
922 if (test_bit(MMF_TOPDOWN, &mm->flags))
923 return arch_get_unmapped_area_topdown(file, addr, len, pgoff, flags, 0);
924 return arch_get_unmapped_area(file, addr, len, pgoff, flags, 0);
925}
926EXPORT_SYMBOL(mm_get_unmapped_area);
927
928/**
929 * find_vma_intersection() - Look up the first VMA which intersects the interval
930 * @mm: The process address space.
931 * @start_addr: The inclusive start user address.
932 * @end_addr: The exclusive end user address.
933 *
934 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
935 * start_addr < end_addr.
936 */
937struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
938 unsigned long start_addr,
939 unsigned long end_addr)
940{
941 unsigned long index = start_addr;
942
943 mmap_assert_locked(mm);
944 return mt_find(&mm->mm_mt, &index, end_addr - 1);
945}
946EXPORT_SYMBOL(find_vma_intersection);
947
948/**
949 * find_vma() - Find the VMA for a given address, or the next VMA.
950 * @mm: The mm_struct to check
951 * @addr: The address
952 *
953 * Returns: The VMA associated with addr, or the next VMA.
954 * May return %NULL in the case of no VMA at addr or above.
955 */
956struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
957{
958 unsigned long index = addr;
959
960 mmap_assert_locked(mm);
961 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
962}
963EXPORT_SYMBOL(find_vma);
964
965/**
966 * find_vma_prev() - Find the VMA for a given address, or the next vma and
967 * set %pprev to the previous VMA, if any.
968 * @mm: The mm_struct to check
969 * @addr: The address
970 * @pprev: The pointer to set to the previous VMA
971 *
972 * Note that RCU lock is missing here since the external mmap_lock() is used
973 * instead.
974 *
975 * Returns: The VMA associated with @addr, or the next vma.
976 * May return %NULL in the case of no vma at addr or above.
977 */
978struct vm_area_struct *
979find_vma_prev(struct mm_struct *mm, unsigned long addr,
980 struct vm_area_struct **pprev)
981{
982 struct vm_area_struct *vma;
983 VMA_ITERATOR(vmi, mm, addr);
984
985 vma = vma_iter_load(&vmi);
986 *pprev = vma_prev(&vmi);
987 if (!vma)
988 vma = vma_next(&vmi);
989 return vma;
990}
991
992/*
993 * Verify that the stack growth is acceptable and
994 * update accounting. This is shared with both the
995 * grow-up and grow-down cases.
996 */
997static int acct_stack_growth(struct vm_area_struct *vma,
998 unsigned long size, unsigned long grow)
999{
1000 struct mm_struct *mm = vma->vm_mm;
1001 unsigned long new_start;
1002
1003 /* address space limit tests */
1004 if (!may_expand_vm(mm, vma->vm_flags, grow))
1005 return -ENOMEM;
1006
1007 /* Stack limit test */
1008 if (size > rlimit(RLIMIT_STACK))
1009 return -ENOMEM;
1010
1011 /* mlock limit tests */
1012 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1013 return -ENOMEM;
1014
1015 /* Check to ensure the stack will not grow into a hugetlb-only region */
1016 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1017 vma->vm_end - size;
1018 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1019 return -EFAULT;
1020
1021 /*
1022 * Overcommit.. This must be the final test, as it will
1023 * update security statistics.
1024 */
1025 if (security_vm_enough_memory_mm(mm, grow))
1026 return -ENOMEM;
1027
1028 return 0;
1029}
1030
1031#if defined(CONFIG_STACK_GROWSUP)
1032/*
1033 * PA-RISC uses this for its stack.
1034 * vma is the last one with address > vma->vm_end. Have to extend vma.
1035 */
1036static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1037{
1038 struct mm_struct *mm = vma->vm_mm;
1039 struct vm_area_struct *next;
1040 unsigned long gap_addr;
1041 int error = 0;
1042 VMA_ITERATOR(vmi, mm, vma->vm_start);
1043
1044 if (!(vma->vm_flags & VM_GROWSUP))
1045 return -EFAULT;
1046
1047 mmap_assert_write_locked(mm);
1048
1049 /* Guard against exceeding limits of the address space. */
1050 address &= PAGE_MASK;
1051 if (address >= (TASK_SIZE & PAGE_MASK))
1052 return -ENOMEM;
1053 address += PAGE_SIZE;
1054
1055 /* Enforce stack_guard_gap */
1056 gap_addr = address + stack_guard_gap;
1057
1058 /* Guard against overflow */
1059 if (gap_addr < address || gap_addr > TASK_SIZE)
1060 gap_addr = TASK_SIZE;
1061
1062 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1063 if (next && vma_is_accessible(next)) {
1064 if (!(next->vm_flags & VM_GROWSUP))
1065 return -ENOMEM;
1066 /* Check that both stack segments have the same anon_vma? */
1067 }
1068
1069 if (next)
1070 vma_iter_prev_range_limit(&vmi, address);
1071
1072 vma_iter_config(&vmi, vma->vm_start, address);
1073 if (vma_iter_prealloc(&vmi, vma))
1074 return -ENOMEM;
1075
1076 /* We must make sure the anon_vma is allocated. */
1077 if (unlikely(anon_vma_prepare(vma))) {
1078 vma_iter_free(&vmi);
1079 return -ENOMEM;
1080 }
1081
1082 /* Lock the VMA before expanding to prevent concurrent page faults */
1083 vma_start_write(vma);
1084 /* We update the anon VMA tree. */
1085 anon_vma_lock_write(vma->anon_vma);
1086
1087 /* Somebody else might have raced and expanded it already */
1088 if (address > vma->vm_end) {
1089 unsigned long size, grow;
1090
1091 size = address - vma->vm_start;
1092 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1093
1094 error = -ENOMEM;
1095 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1096 error = acct_stack_growth(vma, size, grow);
1097 if (!error) {
1098 if (vma->vm_flags & VM_LOCKED)
1099 mm->locked_vm += grow;
1100 vm_stat_account(mm, vma->vm_flags, grow);
1101 anon_vma_interval_tree_pre_update_vma(vma);
1102 vma->vm_end = address;
1103 /* Overwrite old entry in mtree. */
1104 vma_iter_store(&vmi, vma);
1105 anon_vma_interval_tree_post_update_vma(vma);
1106
1107 perf_event_mmap(vma);
1108 }
1109 }
1110 }
1111 anon_vma_unlock_write(vma->anon_vma);
1112 vma_iter_free(&vmi);
1113 validate_mm(mm);
1114 return error;
1115}
1116#endif /* CONFIG_STACK_GROWSUP */
1117
1118/*
1119 * vma is the first one with address < vma->vm_start. Have to extend vma.
1120 * mmap_lock held for writing.
1121 */
1122int expand_downwards(struct vm_area_struct *vma, unsigned long address)
1123{
1124 struct mm_struct *mm = vma->vm_mm;
1125 struct vm_area_struct *prev;
1126 int error = 0;
1127 VMA_ITERATOR(vmi, mm, vma->vm_start);
1128
1129 if (!(vma->vm_flags & VM_GROWSDOWN))
1130 return -EFAULT;
1131
1132 mmap_assert_write_locked(mm);
1133
1134 address &= PAGE_MASK;
1135 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
1136 return -EPERM;
1137
1138 /* Enforce stack_guard_gap */
1139 prev = vma_prev(&vmi);
1140 /* Check that both stack segments have the same anon_vma? */
1141 if (prev) {
1142 if (!(prev->vm_flags & VM_GROWSDOWN) &&
1143 vma_is_accessible(prev) &&
1144 (address - prev->vm_end < stack_guard_gap))
1145 return -ENOMEM;
1146 }
1147
1148 if (prev)
1149 vma_iter_next_range_limit(&vmi, vma->vm_start);
1150
1151 vma_iter_config(&vmi, address, vma->vm_end);
1152 if (vma_iter_prealloc(&vmi, vma))
1153 return -ENOMEM;
1154
1155 /* We must make sure the anon_vma is allocated. */
1156 if (unlikely(anon_vma_prepare(vma))) {
1157 vma_iter_free(&vmi);
1158 return -ENOMEM;
1159 }
1160
1161 /* Lock the VMA before expanding to prevent concurrent page faults */
1162 vma_start_write(vma);
1163 /* We update the anon VMA tree. */
1164 anon_vma_lock_write(vma->anon_vma);
1165
1166 /* Somebody else might have raced and expanded it already */
1167 if (address < vma->vm_start) {
1168 unsigned long size, grow;
1169
1170 size = vma->vm_end - address;
1171 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1172
1173 error = -ENOMEM;
1174 if (grow <= vma->vm_pgoff) {
1175 error = acct_stack_growth(vma, size, grow);
1176 if (!error) {
1177 if (vma->vm_flags & VM_LOCKED)
1178 mm->locked_vm += grow;
1179 vm_stat_account(mm, vma->vm_flags, grow);
1180 anon_vma_interval_tree_pre_update_vma(vma);
1181 vma->vm_start = address;
1182 vma->vm_pgoff -= grow;
1183 /* Overwrite old entry in mtree. */
1184 vma_iter_store(&vmi, vma);
1185 anon_vma_interval_tree_post_update_vma(vma);
1186
1187 perf_event_mmap(vma);
1188 }
1189 }
1190 }
1191 anon_vma_unlock_write(vma->anon_vma);
1192 vma_iter_free(&vmi);
1193 validate_mm(mm);
1194 return error;
1195}
1196
1197/* enforced gap between the expanding stack and other mappings. */
1198unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
1199
1200static int __init cmdline_parse_stack_guard_gap(char *p)
1201{
1202 unsigned long val;
1203 char *endptr;
1204
1205 val = simple_strtoul(p, &endptr, 10);
1206 if (!*endptr)
1207 stack_guard_gap = val << PAGE_SHIFT;
1208
1209 return 1;
1210}
1211__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
1212
1213#ifdef CONFIG_STACK_GROWSUP
1214int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
1215{
1216 return expand_upwards(vma, address);
1217}
1218
1219struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
1220{
1221 struct vm_area_struct *vma, *prev;
1222
1223 addr &= PAGE_MASK;
1224 vma = find_vma_prev(mm, addr, &prev);
1225 if (vma && (vma->vm_start <= addr))
1226 return vma;
1227 if (!prev)
1228 return NULL;
1229 if (expand_stack_locked(prev, addr))
1230 return NULL;
1231 if (prev->vm_flags & VM_LOCKED)
1232 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
1233 return prev;
1234}
1235#else
1236int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
1237{
1238 return expand_downwards(vma, address);
1239}
1240
1241struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
1242{
1243 struct vm_area_struct *vma;
1244 unsigned long start;
1245
1246 addr &= PAGE_MASK;
1247 vma = find_vma(mm, addr);
1248 if (!vma)
1249 return NULL;
1250 if (vma->vm_start <= addr)
1251 return vma;
1252 start = vma->vm_start;
1253 if (expand_stack_locked(vma, addr))
1254 return NULL;
1255 if (vma->vm_flags & VM_LOCKED)
1256 populate_vma_page_range(vma, addr, start, NULL);
1257 return vma;
1258}
1259#endif
1260
1261#if defined(CONFIG_STACK_GROWSUP)
1262
1263#define vma_expand_up(vma,addr) expand_upwards(vma, addr)
1264#define vma_expand_down(vma, addr) (-EFAULT)
1265
1266#else
1267
1268#define vma_expand_up(vma,addr) (-EFAULT)
1269#define vma_expand_down(vma, addr) expand_downwards(vma, addr)
1270
1271#endif
1272
1273/*
1274 * expand_stack(): legacy interface for page faulting. Don't use unless
1275 * you have to.
1276 *
1277 * This is called with the mm locked for reading, drops the lock, takes
1278 * the lock for writing, tries to look up a vma again, expands it if
1279 * necessary, and downgrades the lock to reading again.
1280 *
1281 * If no vma is found or it can't be expanded, it returns NULL and has
1282 * dropped the lock.
1283 */
1284struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
1285{
1286 struct vm_area_struct *vma, *prev;
1287
1288 mmap_read_unlock(mm);
1289 if (mmap_write_lock_killable(mm))
1290 return NULL;
1291
1292 vma = find_vma_prev(mm, addr, &prev);
1293 if (vma && vma->vm_start <= addr)
1294 goto success;
1295
1296 if (prev && !vma_expand_up(prev, addr)) {
1297 vma = prev;
1298 goto success;
1299 }
1300
1301 if (vma && !vma_expand_down(vma, addr))
1302 goto success;
1303
1304 mmap_write_unlock(mm);
1305 return NULL;
1306
1307success:
1308 mmap_write_downgrade(mm);
1309 return vma;
1310}
1311
1312/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
1313 * @mm: The mm_struct
1314 * @start: The start address to munmap
1315 * @len: The length to be munmapped.
1316 * @uf: The userfaultfd list_head
1317 *
1318 * Return: 0 on success, error otherwise.
1319 */
1320int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
1321 struct list_head *uf)
1322{
1323 VMA_ITERATOR(vmi, mm, start);
1324
1325 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
1326}
1327
1328unsigned long mmap_region(struct file *file, unsigned long addr,
1329 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1330 struct list_head *uf)
1331{
1332 unsigned long ret;
1333 bool writable_file_mapping = false;
1334
1335 /* Check to see if MDWE is applicable. */
1336 if (map_deny_write_exec(vm_flags, vm_flags))
1337 return -EACCES;
1338
1339 /* Allow architectures to sanity-check the vm_flags. */
1340 if (!arch_validate_flags(vm_flags))
1341 return -EINVAL;
1342
1343 /* Map writable and ensure this isn't a sealed memfd. */
1344 if (file && is_shared_maywrite(vm_flags)) {
1345 int error = mapping_map_writable(file->f_mapping);
1346
1347 if (error)
1348 return error;
1349 writable_file_mapping = true;
1350 }
1351
1352 ret = __mmap_region(file, addr, len, vm_flags, pgoff, uf);
1353
1354 /* Clear our write mapping regardless of error. */
1355 if (writable_file_mapping)
1356 mapping_unmap_writable(file->f_mapping);
1357
1358 validate_mm(current->mm);
1359 return ret;
1360}
1361
1362static int __vm_munmap(unsigned long start, size_t len, bool unlock)
1363{
1364 int ret;
1365 struct mm_struct *mm = current->mm;
1366 LIST_HEAD(uf);
1367 VMA_ITERATOR(vmi, mm, start);
1368
1369 if (mmap_write_lock_killable(mm))
1370 return -EINTR;
1371
1372 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
1373 if (ret || !unlock)
1374 mmap_write_unlock(mm);
1375
1376 userfaultfd_unmap_complete(mm, &uf);
1377 return ret;
1378}
1379
1380int vm_munmap(unsigned long start, size_t len)
1381{
1382 return __vm_munmap(start, len, false);
1383}
1384EXPORT_SYMBOL(vm_munmap);
1385
1386SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1387{
1388 addr = untagged_addr(addr);
1389 return __vm_munmap(addr, len, true);
1390}
1391
1392
1393/*
1394 * Emulation of deprecated remap_file_pages() syscall.
1395 */
1396SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
1397 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
1398{
1399
1400 struct mm_struct *mm = current->mm;
1401 struct vm_area_struct *vma;
1402 unsigned long populate = 0;
1403 unsigned long ret = -EINVAL;
1404 struct file *file;
1405 vm_flags_t vm_flags;
1406
1407 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
1408 current->comm, current->pid);
1409
1410 if (prot)
1411 return ret;
1412 start = start & PAGE_MASK;
1413 size = size & PAGE_MASK;
1414
1415 if (start + size <= start)
1416 return ret;
1417
1418 /* Does pgoff wrap? */
1419 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
1420 return ret;
1421
1422 if (mmap_read_lock_killable(mm))
1423 return -EINTR;
1424
1425 /*
1426 * Look up VMA under read lock first so we can perform the security
1427 * without holding locks (which can be problematic). We reacquire a
1428 * write lock later and check nothing changed underneath us.
1429 */
1430 vma = vma_lookup(mm, start);
1431
1432 if (!vma || !(vma->vm_flags & VM_SHARED)) {
1433 mmap_read_unlock(mm);
1434 return -EINVAL;
1435 }
1436
1437 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
1438 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
1439 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
1440
1441 flags &= MAP_NONBLOCK;
1442 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
1443 if (vma->vm_flags & VM_LOCKED)
1444 flags |= MAP_LOCKED;
1445
1446 /* Save vm_flags used to calculate prot and flags, and recheck later. */
1447 vm_flags = vma->vm_flags;
1448 file = get_file(vma->vm_file);
1449
1450 mmap_read_unlock(mm);
1451
1452 /* Call outside mmap_lock to be consistent with other callers. */
1453 ret = security_mmap_file(file, prot, flags);
1454 if (ret) {
1455 fput(file);
1456 return ret;
1457 }
1458
1459 ret = -EINVAL;
1460
1461 /* OK security check passed, take write lock + let it rip. */
1462 if (mmap_write_lock_killable(mm)) {
1463 fput(file);
1464 return -EINTR;
1465 }
1466
1467 vma = vma_lookup(mm, start);
1468
1469 if (!vma)
1470 goto out;
1471
1472 /* Make sure things didn't change under us. */
1473 if (vma->vm_flags != vm_flags)
1474 goto out;
1475 if (vma->vm_file != file)
1476 goto out;
1477
1478 if (start + size > vma->vm_end) {
1479 VMA_ITERATOR(vmi, mm, vma->vm_end);
1480 struct vm_area_struct *next, *prev = vma;
1481
1482 for_each_vma_range(vmi, next, start + size) {
1483 /* hole between vmas ? */
1484 if (next->vm_start != prev->vm_end)
1485 goto out;
1486
1487 if (next->vm_file != vma->vm_file)
1488 goto out;
1489
1490 if (next->vm_flags != vma->vm_flags)
1491 goto out;
1492
1493 if (start + size <= next->vm_end)
1494 break;
1495
1496 prev = next;
1497 }
1498
1499 if (!next)
1500 goto out;
1501 }
1502
1503 ret = do_mmap(vma->vm_file, start, size,
1504 prot, flags, 0, pgoff, &populate, NULL);
1505out:
1506 mmap_write_unlock(mm);
1507 fput(file);
1508 if (populate)
1509 mm_populate(ret, populate);
1510 if (!IS_ERR_VALUE(ret))
1511 ret = 0;
1512 return ret;
1513}
1514
1515/*
1516 * do_brk_flags() - Increase the brk vma if the flags match.
1517 * @vmi: The vma iterator
1518 * @addr: The start address
1519 * @len: The length of the increase
1520 * @vma: The vma,
1521 * @flags: The VMA Flags
1522 *
1523 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
1524 * do not match then create a new anonymous VMA. Eventually we may be able to
1525 * do some brk-specific accounting here.
1526 */
1527static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
1528 unsigned long addr, unsigned long len, unsigned long flags)
1529{
1530 struct mm_struct *mm = current->mm;
1531
1532 /*
1533 * Check against address space limits by the changed size
1534 * Note: This happens *after* clearing old mappings in some code paths.
1535 */
1536 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1537 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
1538 return -ENOMEM;
1539
1540 if (mm->map_count > sysctl_max_map_count)
1541 return -ENOMEM;
1542
1543 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
1544 return -ENOMEM;
1545
1546 /*
1547 * Expand the existing vma if possible; Note that singular lists do not
1548 * occur after forking, so the expand will only happen on new VMAs.
1549 */
1550 if (vma && vma->vm_end == addr) {
1551 VMG_STATE(vmg, mm, vmi, addr, addr + len, flags, PHYS_PFN(addr));
1552
1553 vmg.prev = vma;
1554 /* vmi is positioned at prev, which this mode expects. */
1555 vmg.merge_flags = VMG_FLAG_JUST_EXPAND;
1556
1557 if (vma_merge_new_range(&vmg))
1558 goto out;
1559 else if (vmg_nomem(&vmg))
1560 goto unacct_fail;
1561 }
1562
1563 if (vma)
1564 vma_iter_next_range(vmi);
1565 /* create a vma struct for an anonymous mapping */
1566 vma = vm_area_alloc(mm);
1567 if (!vma)
1568 goto unacct_fail;
1569
1570 vma_set_anonymous(vma);
1571 vma_set_range(vma, addr, addr + len, addr >> PAGE_SHIFT);
1572 vm_flags_init(vma, flags);
1573 vma->vm_page_prot = vm_get_page_prot(flags);
1574 vma_start_write(vma);
1575 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
1576 goto mas_store_fail;
1577
1578 mm->map_count++;
1579 validate_mm(mm);
1580 ksm_add_vma(vma);
1581out:
1582 perf_event_mmap(vma);
1583 mm->total_vm += len >> PAGE_SHIFT;
1584 mm->data_vm += len >> PAGE_SHIFT;
1585 if (flags & VM_LOCKED)
1586 mm->locked_vm += (len >> PAGE_SHIFT);
1587 vm_flags_set(vma, VM_SOFTDIRTY);
1588 return 0;
1589
1590mas_store_fail:
1591 vm_area_free(vma);
1592unacct_fail:
1593 vm_unacct_memory(len >> PAGE_SHIFT);
1594 return -ENOMEM;
1595}
1596
1597int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
1598{
1599 struct mm_struct *mm = current->mm;
1600 struct vm_area_struct *vma = NULL;
1601 unsigned long len;
1602 int ret;
1603 bool populate;
1604 LIST_HEAD(uf);
1605 VMA_ITERATOR(vmi, mm, addr);
1606
1607 len = PAGE_ALIGN(request);
1608 if (len < request)
1609 return -ENOMEM;
1610 if (!len)
1611 return 0;
1612
1613 /* Until we need other flags, refuse anything except VM_EXEC. */
1614 if ((flags & (~VM_EXEC)) != 0)
1615 return -EINVAL;
1616
1617 if (mmap_write_lock_killable(mm))
1618 return -EINTR;
1619
1620 ret = check_brk_limits(addr, len);
1621 if (ret)
1622 goto limits_failed;
1623
1624 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
1625 if (ret)
1626 goto munmap_failed;
1627
1628 vma = vma_prev(&vmi);
1629 ret = do_brk_flags(&vmi, vma, addr, len, flags);
1630 populate = ((mm->def_flags & VM_LOCKED) != 0);
1631 mmap_write_unlock(mm);
1632 userfaultfd_unmap_complete(mm, &uf);
1633 if (populate && !ret)
1634 mm_populate(addr, len);
1635 return ret;
1636
1637munmap_failed:
1638limits_failed:
1639 mmap_write_unlock(mm);
1640 return ret;
1641}
1642EXPORT_SYMBOL(vm_brk_flags);
1643
1644/* Release all mmaps. */
1645void exit_mmap(struct mm_struct *mm)
1646{
1647 struct mmu_gather tlb;
1648 struct vm_area_struct *vma;
1649 unsigned long nr_accounted = 0;
1650 VMA_ITERATOR(vmi, mm, 0);
1651 int count = 0;
1652
1653 /* mm's last user has gone, and its about to be pulled down */
1654 mmu_notifier_release(mm);
1655
1656 mmap_read_lock(mm);
1657 arch_exit_mmap(mm);
1658
1659 vma = vma_next(&vmi);
1660 if (!vma || unlikely(xa_is_zero(vma))) {
1661 /* Can happen if dup_mmap() received an OOM */
1662 mmap_read_unlock(mm);
1663 mmap_write_lock(mm);
1664 goto destroy;
1665 }
1666
1667 lru_add_drain();
1668 flush_cache_mm(mm);
1669 tlb_gather_mmu_fullmm(&tlb, mm);
1670 /* update_hiwater_rss(mm) here? but nobody should be looking */
1671 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
1672 unmap_vmas(&tlb, &vmi.mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
1673 mmap_read_unlock(mm);
1674
1675 /*
1676 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
1677 * because the memory has been already freed.
1678 */
1679 set_bit(MMF_OOM_SKIP, &mm->flags);
1680 mmap_write_lock(mm);
1681 mt_clear_in_rcu(&mm->mm_mt);
1682 vma_iter_set(&vmi, vma->vm_end);
1683 free_pgtables(&tlb, &vmi.mas, vma, FIRST_USER_ADDRESS,
1684 USER_PGTABLES_CEILING, true);
1685 tlb_finish_mmu(&tlb);
1686
1687 /*
1688 * Walk the list again, actually closing and freeing it, with preemption
1689 * enabled, without holding any MM locks besides the unreachable
1690 * mmap_write_lock.
1691 */
1692 vma_iter_set(&vmi, vma->vm_end);
1693 do {
1694 if (vma->vm_flags & VM_ACCOUNT)
1695 nr_accounted += vma_pages(vma);
1696 remove_vma(vma, /* unreachable = */ true);
1697 count++;
1698 cond_resched();
1699 vma = vma_next(&vmi);
1700 } while (vma && likely(!xa_is_zero(vma)));
1701
1702 BUG_ON(count != mm->map_count);
1703
1704 trace_exit_mmap(mm);
1705destroy:
1706 __mt_destroy(&mm->mm_mt);
1707 mmap_write_unlock(mm);
1708 vm_unacct_memory(nr_accounted);
1709}
1710
1711/* Insert vm structure into process list sorted by address
1712 * and into the inode's i_mmap tree. If vm_file is non-NULL
1713 * then i_mmap_rwsem is taken here.
1714 */
1715int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
1716{
1717 unsigned long charged = vma_pages(vma);
1718
1719
1720 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
1721 return -ENOMEM;
1722
1723 if ((vma->vm_flags & VM_ACCOUNT) &&
1724 security_vm_enough_memory_mm(mm, charged))
1725 return -ENOMEM;
1726
1727 /*
1728 * The vm_pgoff of a purely anonymous vma should be irrelevant
1729 * until its first write fault, when page's anon_vma and index
1730 * are set. But now set the vm_pgoff it will almost certainly
1731 * end up with (unless mremap moves it elsewhere before that
1732 * first wfault), so /proc/pid/maps tells a consistent story.
1733 *
1734 * By setting it to reflect the virtual start address of the
1735 * vma, merges and splits can happen in a seamless way, just
1736 * using the existing file pgoff checks and manipulations.
1737 * Similarly in do_mmap and in do_brk_flags.
1738 */
1739 if (vma_is_anonymous(vma)) {
1740 BUG_ON(vma->anon_vma);
1741 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
1742 }
1743
1744 if (vma_link(mm, vma)) {
1745 if (vma->vm_flags & VM_ACCOUNT)
1746 vm_unacct_memory(charged);
1747 return -ENOMEM;
1748 }
1749
1750 return 0;
1751}
1752
1753/*
1754 * Return true if the calling process may expand its vm space by the passed
1755 * number of pages
1756 */
1757bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
1758{
1759 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
1760 return false;
1761
1762 if (is_data_mapping(flags) &&
1763 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
1764 /* Workaround for Valgrind */
1765 if (rlimit(RLIMIT_DATA) == 0 &&
1766 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
1767 return true;
1768
1769 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
1770 current->comm, current->pid,
1771 (mm->data_vm + npages) << PAGE_SHIFT,
1772 rlimit(RLIMIT_DATA),
1773 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
1774
1775 if (!ignore_rlimit_data)
1776 return false;
1777 }
1778
1779 return true;
1780}
1781
1782void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
1783{
1784 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
1785
1786 if (is_exec_mapping(flags))
1787 mm->exec_vm += npages;
1788 else if (is_stack_mapping(flags))
1789 mm->stack_vm += npages;
1790 else if (is_data_mapping(flags))
1791 mm->data_vm += npages;
1792}
1793
1794static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
1795
1796/*
1797 * Close hook, called for unmap() and on the old vma for mremap().
1798 *
1799 * Having a close hook prevents vma merging regardless of flags.
1800 */
1801static void special_mapping_close(struct vm_area_struct *vma)
1802{
1803 const struct vm_special_mapping *sm = vma->vm_private_data;
1804
1805 if (sm->close)
1806 sm->close(sm, vma);
1807}
1808
1809static const char *special_mapping_name(struct vm_area_struct *vma)
1810{
1811 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
1812}
1813
1814static int special_mapping_mremap(struct vm_area_struct *new_vma)
1815{
1816 struct vm_special_mapping *sm = new_vma->vm_private_data;
1817
1818 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
1819 return -EFAULT;
1820
1821 if (sm->mremap)
1822 return sm->mremap(sm, new_vma);
1823
1824 return 0;
1825}
1826
1827static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
1828{
1829 /*
1830 * Forbid splitting special mappings - kernel has expectations over
1831 * the number of pages in mapping. Together with VM_DONTEXPAND
1832 * the size of vma should stay the same over the special mapping's
1833 * lifetime.
1834 */
1835 return -EINVAL;
1836}
1837
1838static const struct vm_operations_struct special_mapping_vmops = {
1839 .close = special_mapping_close,
1840 .fault = special_mapping_fault,
1841 .mremap = special_mapping_mremap,
1842 .name = special_mapping_name,
1843 /* vDSO code relies that VVAR can't be accessed remotely */
1844 .access = NULL,
1845 .may_split = special_mapping_split,
1846};
1847
1848static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
1849{
1850 struct vm_area_struct *vma = vmf->vma;
1851 pgoff_t pgoff;
1852 struct page **pages;
1853 struct vm_special_mapping *sm = vma->vm_private_data;
1854
1855 if (sm->fault)
1856 return sm->fault(sm, vmf->vma, vmf);
1857
1858 pages = sm->pages;
1859
1860 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
1861 pgoff--;
1862
1863 if (*pages) {
1864 struct page *page = *pages;
1865 get_page(page);
1866 vmf->page = page;
1867 return 0;
1868 }
1869
1870 return VM_FAULT_SIGBUS;
1871}
1872
1873static struct vm_area_struct *__install_special_mapping(
1874 struct mm_struct *mm,
1875 unsigned long addr, unsigned long len,
1876 unsigned long vm_flags, void *priv,
1877 const struct vm_operations_struct *ops)
1878{
1879 int ret;
1880 struct vm_area_struct *vma;
1881
1882 vma = vm_area_alloc(mm);
1883 if (unlikely(vma == NULL))
1884 return ERR_PTR(-ENOMEM);
1885
1886 vma_set_range(vma, addr, addr + len, 0);
1887 vm_flags_init(vma, (vm_flags | mm->def_flags |
1888 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
1889 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
1890
1891 vma->vm_ops = ops;
1892 vma->vm_private_data = priv;
1893
1894 ret = insert_vm_struct(mm, vma);
1895 if (ret)
1896 goto out;
1897
1898 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
1899
1900 perf_event_mmap(vma);
1901
1902 return vma;
1903
1904out:
1905 vm_area_free(vma);
1906 return ERR_PTR(ret);
1907}
1908
1909bool vma_is_special_mapping(const struct vm_area_struct *vma,
1910 const struct vm_special_mapping *sm)
1911{
1912 return vma->vm_private_data == sm &&
1913 vma->vm_ops == &special_mapping_vmops;
1914}
1915
1916/*
1917 * Called with mm->mmap_lock held for writing.
1918 * Insert a new vma covering the given region, with the given flags.
1919 * Its pages are supplied by the given array of struct page *.
1920 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
1921 * The region past the last page supplied will always produce SIGBUS.
1922 * The array pointer and the pages it points to are assumed to stay alive
1923 * for as long as this mapping might exist.
1924 */
1925struct vm_area_struct *_install_special_mapping(
1926 struct mm_struct *mm,
1927 unsigned long addr, unsigned long len,
1928 unsigned long vm_flags, const struct vm_special_mapping *spec)
1929{
1930 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
1931 &special_mapping_vmops);
1932}
1933
1934/*
1935 * initialise the percpu counter for VM
1936 */
1937void __init mmap_init(void)
1938{
1939 int ret;
1940
1941 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
1942 VM_BUG_ON(ret);
1943}
1944
1945/*
1946 * Initialise sysctl_user_reserve_kbytes.
1947 *
1948 * This is intended to prevent a user from starting a single memory hogging
1949 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
1950 * mode.
1951 *
1952 * The default value is min(3% of free memory, 128MB)
1953 * 128MB is enough to recover with sshd/login, bash, and top/kill.
1954 */
1955static int init_user_reserve(void)
1956{
1957 unsigned long free_kbytes;
1958
1959 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
1960
1961 sysctl_user_reserve_kbytes = min(free_kbytes / 32, SZ_128K);
1962 return 0;
1963}
1964subsys_initcall(init_user_reserve);
1965
1966/*
1967 * Initialise sysctl_admin_reserve_kbytes.
1968 *
1969 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
1970 * to log in and kill a memory hogging process.
1971 *
1972 * Systems with more than 256MB will reserve 8MB, enough to recover
1973 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
1974 * only reserve 3% of free pages by default.
1975 */
1976static int init_admin_reserve(void)
1977{
1978 unsigned long free_kbytes;
1979
1980 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
1981
1982 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, SZ_8K);
1983 return 0;
1984}
1985subsys_initcall(init_admin_reserve);
1986
1987/*
1988 * Reinititalise user and admin reserves if memory is added or removed.
1989 *
1990 * The default user reserve max is 128MB, and the default max for the
1991 * admin reserve is 8MB. These are usually, but not always, enough to
1992 * enable recovery from a memory hogging process using login/sshd, a shell,
1993 * and tools like top. It may make sense to increase or even disable the
1994 * reserve depending on the existence of swap or variations in the recovery
1995 * tools. So, the admin may have changed them.
1996 *
1997 * If memory is added and the reserves have been eliminated or increased above
1998 * the default max, then we'll trust the admin.
1999 *
2000 * If memory is removed and there isn't enough free memory, then we
2001 * need to reset the reserves.
2002 *
2003 * Otherwise keep the reserve set by the admin.
2004 */
2005static int reserve_mem_notifier(struct notifier_block *nb,
2006 unsigned long action, void *data)
2007{
2008 unsigned long tmp, free_kbytes;
2009
2010 switch (action) {
2011 case MEM_ONLINE:
2012 /* Default max is 128MB. Leave alone if modified by operator. */
2013 tmp = sysctl_user_reserve_kbytes;
2014 if (tmp > 0 && tmp < SZ_128K)
2015 init_user_reserve();
2016
2017 /* Default max is 8MB. Leave alone if modified by operator. */
2018 tmp = sysctl_admin_reserve_kbytes;
2019 if (tmp > 0 && tmp < SZ_8K)
2020 init_admin_reserve();
2021
2022 break;
2023 case MEM_OFFLINE:
2024 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
2025
2026 if (sysctl_user_reserve_kbytes > free_kbytes) {
2027 init_user_reserve();
2028 pr_info("vm.user_reserve_kbytes reset to %lu\n",
2029 sysctl_user_reserve_kbytes);
2030 }
2031
2032 if (sysctl_admin_reserve_kbytes > free_kbytes) {
2033 init_admin_reserve();
2034 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
2035 sysctl_admin_reserve_kbytes);
2036 }
2037 break;
2038 default:
2039 break;
2040 }
2041 return NOTIFY_OK;
2042}
2043
2044static int __meminit init_reserve_notifier(void)
2045{
2046 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
2047 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
2048
2049 return 0;
2050}
2051subsys_initcall(init_reserve_notifier);
2052
2053/*
2054 * Relocate a VMA downwards by shift bytes. There cannot be any VMAs between
2055 * this VMA and its relocated range, which will now reside at [vma->vm_start -
2056 * shift, vma->vm_end - shift).
2057 *
2058 * This function is almost certainly NOT what you want for anything other than
2059 * early executable temporary stack relocation.
2060 */
2061int relocate_vma_down(struct vm_area_struct *vma, unsigned long shift)
2062{
2063 /*
2064 * The process proceeds as follows:
2065 *
2066 * 1) Use shift to calculate the new vma endpoints.
2067 * 2) Extend vma to cover both the old and new ranges. This ensures the
2068 * arguments passed to subsequent functions are consistent.
2069 * 3) Move vma's page tables to the new range.
2070 * 4) Free up any cleared pgd range.
2071 * 5) Shrink the vma to cover only the new range.
2072 */
2073
2074 struct mm_struct *mm = vma->vm_mm;
2075 unsigned long old_start = vma->vm_start;
2076 unsigned long old_end = vma->vm_end;
2077 unsigned long length = old_end - old_start;
2078 unsigned long new_start = old_start - shift;
2079 unsigned long new_end = old_end - shift;
2080 VMA_ITERATOR(vmi, mm, new_start);
2081 VMG_STATE(vmg, mm, &vmi, new_start, old_end, 0, vma->vm_pgoff);
2082 struct vm_area_struct *next;
2083 struct mmu_gather tlb;
2084
2085 BUG_ON(new_start > new_end);
2086
2087 /*
2088 * ensure there are no vmas between where we want to go
2089 * and where we are
2090 */
2091 if (vma != vma_next(&vmi))
2092 return -EFAULT;
2093
2094 vma_iter_prev_range(&vmi);
2095 /*
2096 * cover the whole range: [new_start, old_end)
2097 */
2098 vmg.vma = vma;
2099 if (vma_expand(&vmg))
2100 return -ENOMEM;
2101
2102 /*
2103 * move the page tables downwards, on failure we rely on
2104 * process cleanup to remove whatever mess we made.
2105 */
2106 if (length != move_page_tables(vma, old_start,
2107 vma, new_start, length, false, true))
2108 return -ENOMEM;
2109
2110 lru_add_drain();
2111 tlb_gather_mmu(&tlb, mm);
2112 next = vma_next(&vmi);
2113 if (new_end > old_start) {
2114 /*
2115 * when the old and new regions overlap clear from new_end.
2116 */
2117 free_pgd_range(&tlb, new_end, old_end, new_end,
2118 next ? next->vm_start : USER_PGTABLES_CEILING);
2119 } else {
2120 /*
2121 * otherwise, clean from old_start; this is done to not touch
2122 * the address space in [new_end, old_start) some architectures
2123 * have constraints on va-space that make this illegal (IA64) -
2124 * for the others its just a little faster.
2125 */
2126 free_pgd_range(&tlb, old_start, old_end, new_end,
2127 next ? next->vm_start : USER_PGTABLES_CEILING);
2128 }
2129 tlb_finish_mmu(&tlb);
2130
2131 vma_prev(&vmi);
2132 /* Shrink the vma to just the new range */
2133 return vma_shrink(&vmi, vma, new_start, new_end, vma->vm_pgoff);
2134}
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/mm_inline.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/notifier.h>
42#include <linux/memory.h>
43#include <linux/printk.h>
44#include <linux/userfaultfd_k.h>
45#include <linux/moduleparam.h>
46#include <linux/pkeys.h>
47#include <linux/oom.h>
48#include <linux/sched/mm.h>
49
50#include <linux/uaccess.h>
51#include <asm/cacheflush.h>
52#include <asm/tlb.h>
53#include <asm/mmu_context.h>
54
55#define CREATE_TRACE_POINTS
56#include <trace/events/mmap.h>
57
58#include "internal.h"
59
60#ifndef arch_mmap_check
61#define arch_mmap_check(addr, len, flags) (0)
62#endif
63
64#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
65const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
66const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
67int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
68#endif
69#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
70const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
71const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
72int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
73#endif
74
75static bool ignore_rlimit_data;
76core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
77
78static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
79 struct vm_area_struct *vma, struct vm_area_struct *prev,
80 struct vm_area_struct *next, unsigned long start,
81 unsigned long end);
82
83static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
84{
85 return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
86}
87
88/* Update vma->vm_page_prot to reflect vma->vm_flags. */
89void vma_set_page_prot(struct vm_area_struct *vma)
90{
91 unsigned long vm_flags = vma->vm_flags;
92 pgprot_t vm_page_prot;
93
94 vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
95 if (vma_wants_writenotify(vma, vm_page_prot)) {
96 vm_flags &= ~VM_SHARED;
97 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
98 }
99 /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
100 WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
101}
102
103/*
104 * Requires inode->i_mapping->i_mmap_rwsem
105 */
106static void __remove_shared_vm_struct(struct vm_area_struct *vma,
107 struct file *file, struct address_space *mapping)
108{
109 if (vma->vm_flags & VM_SHARED)
110 mapping_unmap_writable(mapping);
111
112 flush_dcache_mmap_lock(mapping);
113 vma_interval_tree_remove(vma, &mapping->i_mmap);
114 flush_dcache_mmap_unlock(mapping);
115}
116
117/*
118 * Unlink a file-based vm structure from its interval tree, to hide
119 * vma from rmap and vmtruncate before freeing its page tables.
120 */
121void unlink_file_vma(struct vm_area_struct *vma)
122{
123 struct file *file = vma->vm_file;
124
125 if (file) {
126 struct address_space *mapping = file->f_mapping;
127 i_mmap_lock_write(mapping);
128 __remove_shared_vm_struct(vma, file, mapping);
129 i_mmap_unlock_write(mapping);
130 }
131}
132
133/*
134 * Close a vm structure and free it.
135 */
136static void remove_vma(struct vm_area_struct *vma)
137{
138 might_sleep();
139 if (vma->vm_ops && vma->vm_ops->close)
140 vma->vm_ops->close(vma);
141 if (vma->vm_file)
142 fput(vma->vm_file);
143 mpol_put(vma_policy(vma));
144 vm_area_free(vma);
145}
146
147/*
148 * check_brk_limits() - Use platform specific check of range & verify mlock
149 * limits.
150 * @addr: The address to check
151 * @len: The size of increase.
152 *
153 * Return: 0 on success.
154 */
155static int check_brk_limits(unsigned long addr, unsigned long len)
156{
157 unsigned long mapped_addr;
158
159 mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
160 if (IS_ERR_VALUE(mapped_addr))
161 return mapped_addr;
162
163 return mlock_future_check(current->mm, current->mm->def_flags, len);
164}
165static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
166 unsigned long newbrk, unsigned long oldbrk,
167 struct list_head *uf);
168static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *brkvma,
169 unsigned long addr, unsigned long request, unsigned long flags);
170SYSCALL_DEFINE1(brk, unsigned long, brk)
171{
172 unsigned long newbrk, oldbrk, origbrk;
173 struct mm_struct *mm = current->mm;
174 struct vm_area_struct *brkvma, *next = NULL;
175 unsigned long min_brk;
176 bool populate;
177 bool downgraded = false;
178 LIST_HEAD(uf);
179 MA_STATE(mas, &mm->mm_mt, 0, 0);
180
181 if (mmap_write_lock_killable(mm))
182 return -EINTR;
183
184 origbrk = mm->brk;
185
186#ifdef CONFIG_COMPAT_BRK
187 /*
188 * CONFIG_COMPAT_BRK can still be overridden by setting
189 * randomize_va_space to 2, which will still cause mm->start_brk
190 * to be arbitrarily shifted
191 */
192 if (current->brk_randomized)
193 min_brk = mm->start_brk;
194 else
195 min_brk = mm->end_data;
196#else
197 min_brk = mm->start_brk;
198#endif
199 if (brk < min_brk)
200 goto out;
201
202 /*
203 * Check against rlimit here. If this check is done later after the test
204 * of oldbrk with newbrk then it can escape the test and let the data
205 * segment grow beyond its set limit the in case where the limit is
206 * not page aligned -Ram Gupta
207 */
208 if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
209 mm->end_data, mm->start_data))
210 goto out;
211
212 newbrk = PAGE_ALIGN(brk);
213 oldbrk = PAGE_ALIGN(mm->brk);
214 if (oldbrk == newbrk) {
215 mm->brk = brk;
216 goto success;
217 }
218
219 /*
220 * Always allow shrinking brk.
221 * do_brk_munmap() may downgrade mmap_lock to read.
222 */
223 if (brk <= mm->brk) {
224 int ret;
225
226 /* Search one past newbrk */
227 mas_set(&mas, newbrk);
228 brkvma = mas_find(&mas, oldbrk);
229 if (!brkvma || brkvma->vm_start >= oldbrk)
230 goto out; /* mapping intersects with an existing non-brk vma. */
231 /*
232 * mm->brk must be protected by write mmap_lock.
233 * do_brk_munmap() may downgrade the lock, so update it
234 * before calling do_brk_munmap().
235 */
236 mm->brk = brk;
237 ret = do_brk_munmap(&mas, brkvma, newbrk, oldbrk, &uf);
238 if (ret == 1) {
239 downgraded = true;
240 goto success;
241 } else if (!ret)
242 goto success;
243
244 mm->brk = origbrk;
245 goto out;
246 }
247
248 if (check_brk_limits(oldbrk, newbrk - oldbrk))
249 goto out;
250
251 /*
252 * Only check if the next VMA is within the stack_guard_gap of the
253 * expansion area
254 */
255 mas_set(&mas, oldbrk);
256 next = mas_find(&mas, newbrk - 1 + PAGE_SIZE + stack_guard_gap);
257 if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
258 goto out;
259
260 brkvma = mas_prev(&mas, mm->start_brk);
261 /* Ok, looks good - let it rip. */
262 if (do_brk_flags(&mas, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
263 goto out;
264
265 mm->brk = brk;
266
267success:
268 populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
269 if (downgraded)
270 mmap_read_unlock(mm);
271 else
272 mmap_write_unlock(mm);
273 userfaultfd_unmap_complete(mm, &uf);
274 if (populate)
275 mm_populate(oldbrk, newbrk - oldbrk);
276 return brk;
277
278out:
279 mmap_write_unlock(mm);
280 return origbrk;
281}
282
283#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
284extern void mt_validate(struct maple_tree *mt);
285extern void mt_dump(const struct maple_tree *mt);
286
287/* Validate the maple tree */
288static void validate_mm_mt(struct mm_struct *mm)
289{
290 struct maple_tree *mt = &mm->mm_mt;
291 struct vm_area_struct *vma_mt;
292
293 MA_STATE(mas, mt, 0, 0);
294
295 mt_validate(&mm->mm_mt);
296 mas_for_each(&mas, vma_mt, ULONG_MAX) {
297 if ((vma_mt->vm_start != mas.index) ||
298 (vma_mt->vm_end - 1 != mas.last)) {
299 pr_emerg("issue in %s\n", current->comm);
300 dump_stack();
301 dump_vma(vma_mt);
302 pr_emerg("mt piv: %p %lu - %lu\n", vma_mt,
303 mas.index, mas.last);
304 pr_emerg("mt vma: %p %lu - %lu\n", vma_mt,
305 vma_mt->vm_start, vma_mt->vm_end);
306
307 mt_dump(mas.tree);
308 if (vma_mt->vm_end != mas.last + 1) {
309 pr_err("vma: %p vma_mt %lu-%lu\tmt %lu-%lu\n",
310 mm, vma_mt->vm_start, vma_mt->vm_end,
311 mas.index, mas.last);
312 mt_dump(mas.tree);
313 }
314 VM_BUG_ON_MM(vma_mt->vm_end != mas.last + 1, mm);
315 if (vma_mt->vm_start != mas.index) {
316 pr_err("vma: %p vma_mt %p %lu - %lu doesn't match\n",
317 mm, vma_mt, vma_mt->vm_start, vma_mt->vm_end);
318 mt_dump(mas.tree);
319 }
320 VM_BUG_ON_MM(vma_mt->vm_start != mas.index, mm);
321 }
322 }
323}
324
325static void validate_mm(struct mm_struct *mm)
326{
327 int bug = 0;
328 int i = 0;
329 struct vm_area_struct *vma;
330 MA_STATE(mas, &mm->mm_mt, 0, 0);
331
332 validate_mm_mt(mm);
333
334 mas_for_each(&mas, vma, ULONG_MAX) {
335#ifdef CONFIG_DEBUG_VM_RB
336 struct anon_vma *anon_vma = vma->anon_vma;
337 struct anon_vma_chain *avc;
338
339 if (anon_vma) {
340 anon_vma_lock_read(anon_vma);
341 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
342 anon_vma_interval_tree_verify(avc);
343 anon_vma_unlock_read(anon_vma);
344 }
345#endif
346 i++;
347 }
348 if (i != mm->map_count) {
349 pr_emerg("map_count %d mas_for_each %d\n", mm->map_count, i);
350 bug = 1;
351 }
352 VM_BUG_ON_MM(bug, mm);
353}
354
355#else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
356#define validate_mm_mt(root) do { } while (0)
357#define validate_mm(mm) do { } while (0)
358#endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
359
360/*
361 * vma has some anon_vma assigned, and is already inserted on that
362 * anon_vma's interval trees.
363 *
364 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
365 * vma must be removed from the anon_vma's interval trees using
366 * anon_vma_interval_tree_pre_update_vma().
367 *
368 * After the update, the vma will be reinserted using
369 * anon_vma_interval_tree_post_update_vma().
370 *
371 * The entire update must be protected by exclusive mmap_lock and by
372 * the root anon_vma's mutex.
373 */
374static inline void
375anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
376{
377 struct anon_vma_chain *avc;
378
379 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
380 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
381}
382
383static inline void
384anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
385{
386 struct anon_vma_chain *avc;
387
388 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
389 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
390}
391
392static unsigned long count_vma_pages_range(struct mm_struct *mm,
393 unsigned long addr, unsigned long end)
394{
395 VMA_ITERATOR(vmi, mm, addr);
396 struct vm_area_struct *vma;
397 unsigned long nr_pages = 0;
398
399 for_each_vma_range(vmi, vma, end) {
400 unsigned long vm_start = max(addr, vma->vm_start);
401 unsigned long vm_end = min(end, vma->vm_end);
402
403 nr_pages += PHYS_PFN(vm_end - vm_start);
404 }
405
406 return nr_pages;
407}
408
409static void __vma_link_file(struct vm_area_struct *vma,
410 struct address_space *mapping)
411{
412 if (vma->vm_flags & VM_SHARED)
413 mapping_allow_writable(mapping);
414
415 flush_dcache_mmap_lock(mapping);
416 vma_interval_tree_insert(vma, &mapping->i_mmap);
417 flush_dcache_mmap_unlock(mapping);
418}
419
420/*
421 * vma_mas_store() - Store a VMA in the maple tree.
422 * @vma: The vm_area_struct
423 * @mas: The maple state
424 *
425 * Efficient way to store a VMA in the maple tree when the @mas has already
426 * walked to the correct location.
427 *
428 * Note: the end address is inclusive in the maple tree.
429 */
430void vma_mas_store(struct vm_area_struct *vma, struct ma_state *mas)
431{
432 trace_vma_store(mas->tree, vma);
433 mas_set_range(mas, vma->vm_start, vma->vm_end - 1);
434 mas_store_prealloc(mas, vma);
435}
436
437/*
438 * vma_mas_remove() - Remove a VMA from the maple tree.
439 * @vma: The vm_area_struct
440 * @mas: The maple state
441 *
442 * Efficient way to remove a VMA from the maple tree when the @mas has already
443 * been established and points to the correct location.
444 * Note: the end address is inclusive in the maple tree.
445 */
446void vma_mas_remove(struct vm_area_struct *vma, struct ma_state *mas)
447{
448 trace_vma_mas_szero(mas->tree, vma->vm_start, vma->vm_end - 1);
449 mas->index = vma->vm_start;
450 mas->last = vma->vm_end - 1;
451 mas_store_prealloc(mas, NULL);
452}
453
454/*
455 * vma_mas_szero() - Set a given range to zero. Used when modifying a
456 * vm_area_struct start or end.
457 *
458 * @mas: The maple tree ma_state
459 * @start: The start address to zero
460 * @end: The end address to zero.
461 */
462static inline void vma_mas_szero(struct ma_state *mas, unsigned long start,
463 unsigned long end)
464{
465 trace_vma_mas_szero(mas->tree, start, end - 1);
466 mas_set_range(mas, start, end - 1);
467 mas_store_prealloc(mas, NULL);
468}
469
470static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
471{
472 MA_STATE(mas, &mm->mm_mt, 0, 0);
473 struct address_space *mapping = NULL;
474
475 if (mas_preallocate(&mas, vma, GFP_KERNEL))
476 return -ENOMEM;
477
478 if (vma->vm_file) {
479 mapping = vma->vm_file->f_mapping;
480 i_mmap_lock_write(mapping);
481 }
482
483 vma_mas_store(vma, &mas);
484
485 if (mapping) {
486 __vma_link_file(vma, mapping);
487 i_mmap_unlock_write(mapping);
488 }
489
490 mm->map_count++;
491 validate_mm(mm);
492 return 0;
493}
494
495/*
496 * vma_expand - Expand an existing VMA
497 *
498 * @mas: The maple state
499 * @vma: The vma to expand
500 * @start: The start of the vma
501 * @end: The exclusive end of the vma
502 * @pgoff: The page offset of vma
503 * @next: The current of next vma.
504 *
505 * Expand @vma to @start and @end. Can expand off the start and end. Will
506 * expand over @next if it's different from @vma and @end == @next->vm_end.
507 * Checking if the @vma can expand and merge with @next needs to be handled by
508 * the caller.
509 *
510 * Returns: 0 on success
511 */
512inline int vma_expand(struct ma_state *mas, struct vm_area_struct *vma,
513 unsigned long start, unsigned long end, pgoff_t pgoff,
514 struct vm_area_struct *next)
515{
516 struct mm_struct *mm = vma->vm_mm;
517 struct address_space *mapping = NULL;
518 struct rb_root_cached *root = NULL;
519 struct anon_vma *anon_vma = vma->anon_vma;
520 struct file *file = vma->vm_file;
521 bool remove_next = false;
522
523 if (next && (vma != next) && (end == next->vm_end)) {
524 remove_next = true;
525 if (next->anon_vma && !vma->anon_vma) {
526 int error;
527
528 anon_vma = next->anon_vma;
529 vma->anon_vma = anon_vma;
530 error = anon_vma_clone(vma, next);
531 if (error)
532 return error;
533 }
534 }
535
536 /* Not merging but overwriting any part of next is not handled. */
537 VM_BUG_ON(next && !remove_next && next != vma && end > next->vm_start);
538 /* Only handles expanding */
539 VM_BUG_ON(vma->vm_start < start || vma->vm_end > end);
540
541 if (mas_preallocate(mas, vma, GFP_KERNEL))
542 goto nomem;
543
544 vma_adjust_trans_huge(vma, start, end, 0);
545
546 if (file) {
547 mapping = file->f_mapping;
548 root = &mapping->i_mmap;
549 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
550 i_mmap_lock_write(mapping);
551 }
552
553 if (anon_vma) {
554 anon_vma_lock_write(anon_vma);
555 anon_vma_interval_tree_pre_update_vma(vma);
556 }
557
558 if (file) {
559 flush_dcache_mmap_lock(mapping);
560 vma_interval_tree_remove(vma, root);
561 }
562
563 vma->vm_start = start;
564 vma->vm_end = end;
565 vma->vm_pgoff = pgoff;
566 /* Note: mas must be pointing to the expanding VMA */
567 vma_mas_store(vma, mas);
568
569 if (file) {
570 vma_interval_tree_insert(vma, root);
571 flush_dcache_mmap_unlock(mapping);
572 }
573
574 /* Expanding over the next vma */
575 if (remove_next && file) {
576 __remove_shared_vm_struct(next, file, mapping);
577 }
578
579 if (anon_vma) {
580 anon_vma_interval_tree_post_update_vma(vma);
581 anon_vma_unlock_write(anon_vma);
582 }
583
584 if (file) {
585 i_mmap_unlock_write(mapping);
586 uprobe_mmap(vma);
587 }
588
589 if (remove_next) {
590 if (file) {
591 uprobe_munmap(next, next->vm_start, next->vm_end);
592 fput(file);
593 }
594 if (next->anon_vma)
595 anon_vma_merge(vma, next);
596 mm->map_count--;
597 mpol_put(vma_policy(next));
598 vm_area_free(next);
599 }
600
601 validate_mm(mm);
602 return 0;
603
604nomem:
605 return -ENOMEM;
606}
607
608/*
609 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
610 * is already present in an i_mmap tree without adjusting the tree.
611 * The following helper function should be used when such adjustments
612 * are necessary. The "insert" vma (if any) is to be inserted
613 * before we drop the necessary locks.
614 */
615int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
616 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
617 struct vm_area_struct *expand)
618{
619 struct mm_struct *mm = vma->vm_mm;
620 struct vm_area_struct *next_next = NULL; /* uninit var warning */
621 struct vm_area_struct *next = find_vma(mm, vma->vm_end);
622 struct vm_area_struct *orig_vma = vma;
623 struct address_space *mapping = NULL;
624 struct rb_root_cached *root = NULL;
625 struct anon_vma *anon_vma = NULL;
626 struct file *file = vma->vm_file;
627 bool vma_changed = false;
628 long adjust_next = 0;
629 int remove_next = 0;
630 MA_STATE(mas, &mm->mm_mt, 0, 0);
631 struct vm_area_struct *exporter = NULL, *importer = NULL;
632
633 if (next && !insert) {
634 if (end >= next->vm_end) {
635 /*
636 * vma expands, overlapping all the next, and
637 * perhaps the one after too (mprotect case 6).
638 * The only other cases that gets here are
639 * case 1, case 7 and case 8.
640 */
641 if (next == expand) {
642 /*
643 * The only case where we don't expand "vma"
644 * and we expand "next" instead is case 8.
645 */
646 VM_WARN_ON(end != next->vm_end);
647 /*
648 * remove_next == 3 means we're
649 * removing "vma" and that to do so we
650 * swapped "vma" and "next".
651 */
652 remove_next = 3;
653 VM_WARN_ON(file != next->vm_file);
654 swap(vma, next);
655 } else {
656 VM_WARN_ON(expand != vma);
657 /*
658 * case 1, 6, 7, remove_next == 2 is case 6,
659 * remove_next == 1 is case 1 or 7.
660 */
661 remove_next = 1 + (end > next->vm_end);
662 if (remove_next == 2)
663 next_next = find_vma(mm, next->vm_end);
664
665 VM_WARN_ON(remove_next == 2 &&
666 end != next_next->vm_end);
667 }
668
669 exporter = next;
670 importer = vma;
671
672 /*
673 * If next doesn't have anon_vma, import from vma after
674 * next, if the vma overlaps with it.
675 */
676 if (remove_next == 2 && !next->anon_vma)
677 exporter = next_next;
678
679 } else if (end > next->vm_start) {
680 /*
681 * vma expands, overlapping part of the next:
682 * mprotect case 5 shifting the boundary up.
683 */
684 adjust_next = (end - next->vm_start);
685 exporter = next;
686 importer = vma;
687 VM_WARN_ON(expand != importer);
688 } else if (end < vma->vm_end) {
689 /*
690 * vma shrinks, and !insert tells it's not
691 * split_vma inserting another: so it must be
692 * mprotect case 4 shifting the boundary down.
693 */
694 adjust_next = -(vma->vm_end - end);
695 exporter = vma;
696 importer = next;
697 VM_WARN_ON(expand != importer);
698 }
699
700 /*
701 * Easily overlooked: when mprotect shifts the boundary,
702 * make sure the expanding vma has anon_vma set if the
703 * shrinking vma had, to cover any anon pages imported.
704 */
705 if (exporter && exporter->anon_vma && !importer->anon_vma) {
706 int error;
707
708 importer->anon_vma = exporter->anon_vma;
709 error = anon_vma_clone(importer, exporter);
710 if (error)
711 return error;
712 }
713 }
714
715 if (mas_preallocate(&mas, vma, GFP_KERNEL))
716 return -ENOMEM;
717
718 vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
719 if (file) {
720 mapping = file->f_mapping;
721 root = &mapping->i_mmap;
722 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
723
724 if (adjust_next)
725 uprobe_munmap(next, next->vm_start, next->vm_end);
726
727 i_mmap_lock_write(mapping);
728 if (insert && insert->vm_file) {
729 /*
730 * Put into interval tree now, so instantiated pages
731 * are visible to arm/parisc __flush_dcache_page
732 * throughout; but we cannot insert into address
733 * space until vma start or end is updated.
734 */
735 __vma_link_file(insert, insert->vm_file->f_mapping);
736 }
737 }
738
739 anon_vma = vma->anon_vma;
740 if (!anon_vma && adjust_next)
741 anon_vma = next->anon_vma;
742 if (anon_vma) {
743 VM_WARN_ON(adjust_next && next->anon_vma &&
744 anon_vma != next->anon_vma);
745 anon_vma_lock_write(anon_vma);
746 anon_vma_interval_tree_pre_update_vma(vma);
747 if (adjust_next)
748 anon_vma_interval_tree_pre_update_vma(next);
749 }
750
751 if (file) {
752 flush_dcache_mmap_lock(mapping);
753 vma_interval_tree_remove(vma, root);
754 if (adjust_next)
755 vma_interval_tree_remove(next, root);
756 }
757
758 if (start != vma->vm_start) {
759 if ((vma->vm_start < start) &&
760 (!insert || (insert->vm_end != start))) {
761 vma_mas_szero(&mas, vma->vm_start, start);
762 VM_WARN_ON(insert && insert->vm_start > vma->vm_start);
763 } else {
764 vma_changed = true;
765 }
766 vma->vm_start = start;
767 }
768 if (end != vma->vm_end) {
769 if (vma->vm_end > end) {
770 if (!insert || (insert->vm_start != end)) {
771 vma_mas_szero(&mas, end, vma->vm_end);
772 mas_reset(&mas);
773 VM_WARN_ON(insert &&
774 insert->vm_end < vma->vm_end);
775 }
776 } else {
777 vma_changed = true;
778 }
779 vma->vm_end = end;
780 }
781
782 if (vma_changed)
783 vma_mas_store(vma, &mas);
784
785 vma->vm_pgoff = pgoff;
786 if (adjust_next) {
787 next->vm_start += adjust_next;
788 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
789 vma_mas_store(next, &mas);
790 }
791
792 if (file) {
793 if (adjust_next)
794 vma_interval_tree_insert(next, root);
795 vma_interval_tree_insert(vma, root);
796 flush_dcache_mmap_unlock(mapping);
797 }
798
799 if (remove_next && file) {
800 __remove_shared_vm_struct(next, file, mapping);
801 if (remove_next == 2)
802 __remove_shared_vm_struct(next_next, file, mapping);
803 } else if (insert) {
804 /*
805 * split_vma has split insert from vma, and needs
806 * us to insert it before dropping the locks
807 * (it may either follow vma or precede it).
808 */
809 mas_reset(&mas);
810 vma_mas_store(insert, &mas);
811 mm->map_count++;
812 }
813
814 if (anon_vma) {
815 anon_vma_interval_tree_post_update_vma(vma);
816 if (adjust_next)
817 anon_vma_interval_tree_post_update_vma(next);
818 anon_vma_unlock_write(anon_vma);
819 }
820
821 if (file) {
822 i_mmap_unlock_write(mapping);
823 uprobe_mmap(vma);
824
825 if (adjust_next)
826 uprobe_mmap(next);
827 }
828
829 if (remove_next) {
830again:
831 if (file) {
832 uprobe_munmap(next, next->vm_start, next->vm_end);
833 fput(file);
834 }
835 if (next->anon_vma)
836 anon_vma_merge(vma, next);
837 mm->map_count--;
838 mpol_put(vma_policy(next));
839 if (remove_next != 2)
840 BUG_ON(vma->vm_end < next->vm_end);
841 vm_area_free(next);
842
843 /*
844 * In mprotect's case 6 (see comments on vma_merge),
845 * we must remove next_next too.
846 */
847 if (remove_next == 2) {
848 remove_next = 1;
849 next = next_next;
850 goto again;
851 }
852 }
853 if (insert && file)
854 uprobe_mmap(insert);
855
856 mas_destroy(&mas);
857 validate_mm(mm);
858
859 return 0;
860}
861
862/*
863 * If the vma has a ->close operation then the driver probably needs to release
864 * per-vma resources, so we don't attempt to merge those.
865 */
866static inline int is_mergeable_vma(struct vm_area_struct *vma,
867 struct file *file, unsigned long vm_flags,
868 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
869 struct anon_vma_name *anon_name)
870{
871 /*
872 * VM_SOFTDIRTY should not prevent from VMA merging, if we
873 * match the flags but dirty bit -- the caller should mark
874 * merged VMA as dirty. If dirty bit won't be excluded from
875 * comparison, we increase pressure on the memory system forcing
876 * the kernel to generate new VMAs when old one could be
877 * extended instead.
878 */
879 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
880 return 0;
881 if (vma->vm_file != file)
882 return 0;
883 if (vma->vm_ops && vma->vm_ops->close)
884 return 0;
885 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
886 return 0;
887 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
888 return 0;
889 return 1;
890}
891
892static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
893 struct anon_vma *anon_vma2,
894 struct vm_area_struct *vma)
895{
896 /*
897 * The list_is_singular() test is to avoid merging VMA cloned from
898 * parents. This can improve scalability caused by anon_vma lock.
899 */
900 if ((!anon_vma1 || !anon_vma2) && (!vma ||
901 list_is_singular(&vma->anon_vma_chain)))
902 return 1;
903 return anon_vma1 == anon_vma2;
904}
905
906/*
907 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
908 * in front of (at a lower virtual address and file offset than) the vma.
909 *
910 * We cannot merge two vmas if they have differently assigned (non-NULL)
911 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
912 *
913 * We don't check here for the merged mmap wrapping around the end of pagecache
914 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
915 * wrap, nor mmaps which cover the final page at index -1UL.
916 */
917static int
918can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
919 struct anon_vma *anon_vma, struct file *file,
920 pgoff_t vm_pgoff,
921 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
922 struct anon_vma_name *anon_name)
923{
924 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
925 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
926 if (vma->vm_pgoff == vm_pgoff)
927 return 1;
928 }
929 return 0;
930}
931
932/*
933 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
934 * beyond (at a higher virtual address and file offset than) the vma.
935 *
936 * We cannot merge two vmas if they have differently assigned (non-NULL)
937 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
938 */
939static int
940can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
941 struct anon_vma *anon_vma, struct file *file,
942 pgoff_t vm_pgoff,
943 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
944 struct anon_vma_name *anon_name)
945{
946 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
947 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
948 pgoff_t vm_pglen;
949 vm_pglen = vma_pages(vma);
950 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
951 return 1;
952 }
953 return 0;
954}
955
956/*
957 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
958 * figure out whether that can be merged with its predecessor or its
959 * successor. Or both (it neatly fills a hole).
960 *
961 * In most cases - when called for mmap, brk or mremap - [addr,end) is
962 * certain not to be mapped by the time vma_merge is called; but when
963 * called for mprotect, it is certain to be already mapped (either at
964 * an offset within prev, or at the start of next), and the flags of
965 * this area are about to be changed to vm_flags - and the no-change
966 * case has already been eliminated.
967 *
968 * The following mprotect cases have to be considered, where AAAA is
969 * the area passed down from mprotect_fixup, never extending beyond one
970 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
971 *
972 * AAAA AAAA AAAA
973 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
974 * cannot merge might become might become
975 * PPNNNNNNNNNN PPPPPPPPPPNN
976 * mmap, brk or case 4 below case 5 below
977 * mremap move:
978 * AAAA AAAA
979 * PPPP NNNN PPPPNNNNXXXX
980 * might become might become
981 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
982 * PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
983 * PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
984 *
985 * It is important for case 8 that the vma NNNN overlapping the
986 * region AAAA is never going to extended over XXXX. Instead XXXX must
987 * be extended in region AAAA and NNNN must be removed. This way in
988 * all cases where vma_merge succeeds, the moment vma_adjust drops the
989 * rmap_locks, the properties of the merged vma will be already
990 * correct for the whole merged range. Some of those properties like
991 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
992 * be correct for the whole merged range immediately after the
993 * rmap_locks are released. Otherwise if XXXX would be removed and
994 * NNNN would be extended over the XXXX range, remove_migration_ptes
995 * or other rmap walkers (if working on addresses beyond the "end"
996 * parameter) may establish ptes with the wrong permissions of NNNN
997 * instead of the right permissions of XXXX.
998 */
999struct vm_area_struct *vma_merge(struct mm_struct *mm,
1000 struct vm_area_struct *prev, unsigned long addr,
1001 unsigned long end, unsigned long vm_flags,
1002 struct anon_vma *anon_vma, struct file *file,
1003 pgoff_t pgoff, struct mempolicy *policy,
1004 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1005 struct anon_vma_name *anon_name)
1006{
1007 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1008 struct vm_area_struct *mid, *next, *res;
1009 int err = -1;
1010 bool merge_prev = false;
1011 bool merge_next = false;
1012
1013 /*
1014 * We later require that vma->vm_flags == vm_flags,
1015 * so this tests vma->vm_flags & VM_SPECIAL, too.
1016 */
1017 if (vm_flags & VM_SPECIAL)
1018 return NULL;
1019
1020 next = find_vma(mm, prev ? prev->vm_end : 0);
1021 mid = next;
1022 if (next && next->vm_end == end) /* cases 6, 7, 8 */
1023 next = find_vma(mm, next->vm_end);
1024
1025 /* verify some invariant that must be enforced by the caller */
1026 VM_WARN_ON(prev && addr <= prev->vm_start);
1027 VM_WARN_ON(mid && end > mid->vm_end);
1028 VM_WARN_ON(addr >= end);
1029
1030 /* Can we merge the predecessor? */
1031 if (prev && prev->vm_end == addr &&
1032 mpol_equal(vma_policy(prev), policy) &&
1033 can_vma_merge_after(prev, vm_flags,
1034 anon_vma, file, pgoff,
1035 vm_userfaultfd_ctx, anon_name)) {
1036 merge_prev = true;
1037 }
1038 /* Can we merge the successor? */
1039 if (next && end == next->vm_start &&
1040 mpol_equal(policy, vma_policy(next)) &&
1041 can_vma_merge_before(next, vm_flags,
1042 anon_vma, file, pgoff+pglen,
1043 vm_userfaultfd_ctx, anon_name)) {
1044 merge_next = true;
1045 }
1046 /* Can we merge both the predecessor and the successor? */
1047 if (merge_prev && merge_next &&
1048 is_mergeable_anon_vma(prev->anon_vma,
1049 next->anon_vma, NULL)) { /* cases 1, 6 */
1050 err = __vma_adjust(prev, prev->vm_start,
1051 next->vm_end, prev->vm_pgoff, NULL,
1052 prev);
1053 res = prev;
1054 } else if (merge_prev) { /* cases 2, 5, 7 */
1055 err = __vma_adjust(prev, prev->vm_start,
1056 end, prev->vm_pgoff, NULL, prev);
1057 res = prev;
1058 } else if (merge_next) {
1059 if (prev && addr < prev->vm_end) /* case 4 */
1060 err = __vma_adjust(prev, prev->vm_start,
1061 addr, prev->vm_pgoff, NULL, next);
1062 else /* cases 3, 8 */
1063 err = __vma_adjust(mid, addr, next->vm_end,
1064 next->vm_pgoff - pglen, NULL, next);
1065 res = next;
1066 }
1067
1068 /*
1069 * Cannot merge with predecessor or successor or error in __vma_adjust?
1070 */
1071 if (err)
1072 return NULL;
1073 khugepaged_enter_vma(res, vm_flags);
1074 return res;
1075}
1076
1077/*
1078 * Rough compatibility check to quickly see if it's even worth looking
1079 * at sharing an anon_vma.
1080 *
1081 * They need to have the same vm_file, and the flags can only differ
1082 * in things that mprotect may change.
1083 *
1084 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1085 * we can merge the two vma's. For example, we refuse to merge a vma if
1086 * there is a vm_ops->close() function, because that indicates that the
1087 * driver is doing some kind of reference counting. But that doesn't
1088 * really matter for the anon_vma sharing case.
1089 */
1090static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1091{
1092 return a->vm_end == b->vm_start &&
1093 mpol_equal(vma_policy(a), vma_policy(b)) &&
1094 a->vm_file == b->vm_file &&
1095 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1096 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1097}
1098
1099/*
1100 * Do some basic sanity checking to see if we can re-use the anon_vma
1101 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1102 * the same as 'old', the other will be the new one that is trying
1103 * to share the anon_vma.
1104 *
1105 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1106 * the anon_vma of 'old' is concurrently in the process of being set up
1107 * by another page fault trying to merge _that_. But that's ok: if it
1108 * is being set up, that automatically means that it will be a singleton
1109 * acceptable for merging, so we can do all of this optimistically. But
1110 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1111 *
1112 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1113 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1114 * is to return an anon_vma that is "complex" due to having gone through
1115 * a fork).
1116 *
1117 * We also make sure that the two vma's are compatible (adjacent,
1118 * and with the same memory policies). That's all stable, even with just
1119 * a read lock on the mmap_lock.
1120 */
1121static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1122{
1123 if (anon_vma_compatible(a, b)) {
1124 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1125
1126 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1127 return anon_vma;
1128 }
1129 return NULL;
1130}
1131
1132/*
1133 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1134 * neighbouring vmas for a suitable anon_vma, before it goes off
1135 * to allocate a new anon_vma. It checks because a repetitive
1136 * sequence of mprotects and faults may otherwise lead to distinct
1137 * anon_vmas being allocated, preventing vma merge in subsequent
1138 * mprotect.
1139 */
1140struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1141{
1142 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1143 struct anon_vma *anon_vma = NULL;
1144 struct vm_area_struct *prev, *next;
1145
1146 /* Try next first. */
1147 next = mas_walk(&mas);
1148 if (next) {
1149 anon_vma = reusable_anon_vma(next, vma, next);
1150 if (anon_vma)
1151 return anon_vma;
1152 }
1153
1154 prev = mas_prev(&mas, 0);
1155 VM_BUG_ON_VMA(prev != vma, vma);
1156 prev = mas_prev(&mas, 0);
1157 /* Try prev next. */
1158 if (prev)
1159 anon_vma = reusable_anon_vma(prev, prev, vma);
1160
1161 /*
1162 * We might reach here with anon_vma == NULL if we can't find
1163 * any reusable anon_vma.
1164 * There's no absolute need to look only at touching neighbours:
1165 * we could search further afield for "compatible" anon_vmas.
1166 * But it would probably just be a waste of time searching,
1167 * or lead to too many vmas hanging off the same anon_vma.
1168 * We're trying to allow mprotect remerging later on,
1169 * not trying to minimize memory used for anon_vmas.
1170 */
1171 return anon_vma;
1172}
1173
1174/*
1175 * If a hint addr is less than mmap_min_addr change hint to be as
1176 * low as possible but still greater than mmap_min_addr
1177 */
1178static inline unsigned long round_hint_to_min(unsigned long hint)
1179{
1180 hint &= PAGE_MASK;
1181 if (((void *)hint != NULL) &&
1182 (hint < mmap_min_addr))
1183 return PAGE_ALIGN(mmap_min_addr);
1184 return hint;
1185}
1186
1187int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1188 unsigned long len)
1189{
1190 unsigned long locked, lock_limit;
1191
1192 /* mlock MCL_FUTURE? */
1193 if (flags & VM_LOCKED) {
1194 locked = len >> PAGE_SHIFT;
1195 locked += mm->locked_vm;
1196 lock_limit = rlimit(RLIMIT_MEMLOCK);
1197 lock_limit >>= PAGE_SHIFT;
1198 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1199 return -EAGAIN;
1200 }
1201 return 0;
1202}
1203
1204static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1205{
1206 if (S_ISREG(inode->i_mode))
1207 return MAX_LFS_FILESIZE;
1208
1209 if (S_ISBLK(inode->i_mode))
1210 return MAX_LFS_FILESIZE;
1211
1212 if (S_ISSOCK(inode->i_mode))
1213 return MAX_LFS_FILESIZE;
1214
1215 /* Special "we do even unsigned file positions" case */
1216 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1217 return 0;
1218
1219 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1220 return ULONG_MAX;
1221}
1222
1223static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1224 unsigned long pgoff, unsigned long len)
1225{
1226 u64 maxsize = file_mmap_size_max(file, inode);
1227
1228 if (maxsize && len > maxsize)
1229 return false;
1230 maxsize -= len;
1231 if (pgoff > maxsize >> PAGE_SHIFT)
1232 return false;
1233 return true;
1234}
1235
1236/*
1237 * The caller must write-lock current->mm->mmap_lock.
1238 */
1239unsigned long do_mmap(struct file *file, unsigned long addr,
1240 unsigned long len, unsigned long prot,
1241 unsigned long flags, unsigned long pgoff,
1242 unsigned long *populate, struct list_head *uf)
1243{
1244 struct mm_struct *mm = current->mm;
1245 vm_flags_t vm_flags;
1246 int pkey = 0;
1247
1248 validate_mm(mm);
1249 *populate = 0;
1250
1251 if (!len)
1252 return -EINVAL;
1253
1254 /*
1255 * Does the application expect PROT_READ to imply PROT_EXEC?
1256 *
1257 * (the exception is when the underlying filesystem is noexec
1258 * mounted, in which case we dont add PROT_EXEC.)
1259 */
1260 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1261 if (!(file && path_noexec(&file->f_path)))
1262 prot |= PROT_EXEC;
1263
1264 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1265 if (flags & MAP_FIXED_NOREPLACE)
1266 flags |= MAP_FIXED;
1267
1268 if (!(flags & MAP_FIXED))
1269 addr = round_hint_to_min(addr);
1270
1271 /* Careful about overflows.. */
1272 len = PAGE_ALIGN(len);
1273 if (!len)
1274 return -ENOMEM;
1275
1276 /* offset overflow? */
1277 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1278 return -EOVERFLOW;
1279
1280 /* Too many mappings? */
1281 if (mm->map_count > sysctl_max_map_count)
1282 return -ENOMEM;
1283
1284 /* Obtain the address to map to. we verify (or select) it and ensure
1285 * that it represents a valid section of the address space.
1286 */
1287 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1288 if (IS_ERR_VALUE(addr))
1289 return addr;
1290
1291 if (flags & MAP_FIXED_NOREPLACE) {
1292 if (find_vma_intersection(mm, addr, addr + len))
1293 return -EEXIST;
1294 }
1295
1296 if (prot == PROT_EXEC) {
1297 pkey = execute_only_pkey(mm);
1298 if (pkey < 0)
1299 pkey = 0;
1300 }
1301
1302 /* Do simple checking here so the lower-level routines won't have
1303 * to. we assume access permissions have been handled by the open
1304 * of the memory object, so we don't do any here.
1305 */
1306 vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1307 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1308
1309 if (flags & MAP_LOCKED)
1310 if (!can_do_mlock())
1311 return -EPERM;
1312
1313 if (mlock_future_check(mm, vm_flags, len))
1314 return -EAGAIN;
1315
1316 if (file) {
1317 struct inode *inode = file_inode(file);
1318 unsigned long flags_mask;
1319
1320 if (!file_mmap_ok(file, inode, pgoff, len))
1321 return -EOVERFLOW;
1322
1323 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1324
1325 switch (flags & MAP_TYPE) {
1326 case MAP_SHARED:
1327 /*
1328 * Force use of MAP_SHARED_VALIDATE with non-legacy
1329 * flags. E.g. MAP_SYNC is dangerous to use with
1330 * MAP_SHARED as you don't know which consistency model
1331 * you will get. We silently ignore unsupported flags
1332 * with MAP_SHARED to preserve backward compatibility.
1333 */
1334 flags &= LEGACY_MAP_MASK;
1335 fallthrough;
1336 case MAP_SHARED_VALIDATE:
1337 if (flags & ~flags_mask)
1338 return -EOPNOTSUPP;
1339 if (prot & PROT_WRITE) {
1340 if (!(file->f_mode & FMODE_WRITE))
1341 return -EACCES;
1342 if (IS_SWAPFILE(file->f_mapping->host))
1343 return -ETXTBSY;
1344 }
1345
1346 /*
1347 * Make sure we don't allow writing to an append-only
1348 * file..
1349 */
1350 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1351 return -EACCES;
1352
1353 vm_flags |= VM_SHARED | VM_MAYSHARE;
1354 if (!(file->f_mode & FMODE_WRITE))
1355 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1356 fallthrough;
1357 case MAP_PRIVATE:
1358 if (!(file->f_mode & FMODE_READ))
1359 return -EACCES;
1360 if (path_noexec(&file->f_path)) {
1361 if (vm_flags & VM_EXEC)
1362 return -EPERM;
1363 vm_flags &= ~VM_MAYEXEC;
1364 }
1365
1366 if (!file->f_op->mmap)
1367 return -ENODEV;
1368 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1369 return -EINVAL;
1370 break;
1371
1372 default:
1373 return -EINVAL;
1374 }
1375 } else {
1376 switch (flags & MAP_TYPE) {
1377 case MAP_SHARED:
1378 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1379 return -EINVAL;
1380 /*
1381 * Ignore pgoff.
1382 */
1383 pgoff = 0;
1384 vm_flags |= VM_SHARED | VM_MAYSHARE;
1385 break;
1386 case MAP_PRIVATE:
1387 /*
1388 * Set pgoff according to addr for anon_vma.
1389 */
1390 pgoff = addr >> PAGE_SHIFT;
1391 break;
1392 default:
1393 return -EINVAL;
1394 }
1395 }
1396
1397 /*
1398 * Set 'VM_NORESERVE' if we should not account for the
1399 * memory use of this mapping.
1400 */
1401 if (flags & MAP_NORESERVE) {
1402 /* We honor MAP_NORESERVE if allowed to overcommit */
1403 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1404 vm_flags |= VM_NORESERVE;
1405
1406 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1407 if (file && is_file_hugepages(file))
1408 vm_flags |= VM_NORESERVE;
1409 }
1410
1411 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1412 if (!IS_ERR_VALUE(addr) &&
1413 ((vm_flags & VM_LOCKED) ||
1414 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1415 *populate = len;
1416 return addr;
1417}
1418
1419unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1420 unsigned long prot, unsigned long flags,
1421 unsigned long fd, unsigned long pgoff)
1422{
1423 struct file *file = NULL;
1424 unsigned long retval;
1425
1426 if (!(flags & MAP_ANONYMOUS)) {
1427 audit_mmap_fd(fd, flags);
1428 file = fget(fd);
1429 if (!file)
1430 return -EBADF;
1431 if (is_file_hugepages(file)) {
1432 len = ALIGN(len, huge_page_size(hstate_file(file)));
1433 } else if (unlikely(flags & MAP_HUGETLB)) {
1434 retval = -EINVAL;
1435 goto out_fput;
1436 }
1437 } else if (flags & MAP_HUGETLB) {
1438 struct hstate *hs;
1439
1440 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1441 if (!hs)
1442 return -EINVAL;
1443
1444 len = ALIGN(len, huge_page_size(hs));
1445 /*
1446 * VM_NORESERVE is used because the reservations will be
1447 * taken when vm_ops->mmap() is called
1448 */
1449 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1450 VM_NORESERVE,
1451 HUGETLB_ANONHUGE_INODE,
1452 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1453 if (IS_ERR(file))
1454 return PTR_ERR(file);
1455 }
1456
1457 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1458out_fput:
1459 if (file)
1460 fput(file);
1461 return retval;
1462}
1463
1464SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1465 unsigned long, prot, unsigned long, flags,
1466 unsigned long, fd, unsigned long, pgoff)
1467{
1468 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1469}
1470
1471#ifdef __ARCH_WANT_SYS_OLD_MMAP
1472struct mmap_arg_struct {
1473 unsigned long addr;
1474 unsigned long len;
1475 unsigned long prot;
1476 unsigned long flags;
1477 unsigned long fd;
1478 unsigned long offset;
1479};
1480
1481SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1482{
1483 struct mmap_arg_struct a;
1484
1485 if (copy_from_user(&a, arg, sizeof(a)))
1486 return -EFAULT;
1487 if (offset_in_page(a.offset))
1488 return -EINVAL;
1489
1490 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1491 a.offset >> PAGE_SHIFT);
1492}
1493#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1494
1495/*
1496 * Some shared mappings will want the pages marked read-only
1497 * to track write events. If so, we'll downgrade vm_page_prot
1498 * to the private version (using protection_map[] without the
1499 * VM_SHARED bit).
1500 */
1501int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1502{
1503 vm_flags_t vm_flags = vma->vm_flags;
1504 const struct vm_operations_struct *vm_ops = vma->vm_ops;
1505
1506 /* If it was private or non-writable, the write bit is already clear */
1507 if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1508 return 0;
1509
1510 /* The backer wishes to know when pages are first written to? */
1511 if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1512 return 1;
1513
1514 /* The open routine did something to the protections that pgprot_modify
1515 * won't preserve? */
1516 if (pgprot_val(vm_page_prot) !=
1517 pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1518 return 0;
1519
1520 /*
1521 * Do we need to track softdirty? hugetlb does not support softdirty
1522 * tracking yet.
1523 */
1524 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1525 return 1;
1526
1527 /* Do we need write faults for uffd-wp tracking? */
1528 if (userfaultfd_wp(vma))
1529 return 1;
1530
1531 /* Specialty mapping? */
1532 if (vm_flags & VM_PFNMAP)
1533 return 0;
1534
1535 /* Can the mapping track the dirty pages? */
1536 return vma->vm_file && vma->vm_file->f_mapping &&
1537 mapping_can_writeback(vma->vm_file->f_mapping);
1538}
1539
1540/*
1541 * We account for memory if it's a private writeable mapping,
1542 * not hugepages and VM_NORESERVE wasn't set.
1543 */
1544static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1545{
1546 /*
1547 * hugetlb has its own accounting separate from the core VM
1548 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1549 */
1550 if (file && is_file_hugepages(file))
1551 return 0;
1552
1553 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1554}
1555
1556/**
1557 * unmapped_area() - Find an area between the low_limit and the high_limit with
1558 * the correct alignment and offset, all from @info. Note: current->mm is used
1559 * for the search.
1560 *
1561 * @info: The unmapped area information including the range (low_limit -
1562 * hight_limit), the alignment offset and mask.
1563 *
1564 * Return: A memory address or -ENOMEM.
1565 */
1566static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1567{
1568 unsigned long length, gap;
1569
1570 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1571
1572 /* Adjust search length to account for worst case alignment overhead */
1573 length = info->length + info->align_mask;
1574 if (length < info->length)
1575 return -ENOMEM;
1576
1577 if (mas_empty_area(&mas, info->low_limit, info->high_limit - 1,
1578 length))
1579 return -ENOMEM;
1580
1581 gap = mas.index;
1582 gap += (info->align_offset - gap) & info->align_mask;
1583 return gap;
1584}
1585
1586/**
1587 * unmapped_area_topdown() - Find an area between the low_limit and the
1588 * high_limit with * the correct alignment and offset at the highest available
1589 * address, all from @info. Note: current->mm is used for the search.
1590 *
1591 * @info: The unmapped area information including the range (low_limit -
1592 * hight_limit), the alignment offset and mask.
1593 *
1594 * Return: A memory address or -ENOMEM.
1595 */
1596static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1597{
1598 unsigned long length, gap;
1599
1600 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1601 /* Adjust search length to account for worst case alignment overhead */
1602 length = info->length + info->align_mask;
1603 if (length < info->length)
1604 return -ENOMEM;
1605
1606 if (mas_empty_area_rev(&mas, info->low_limit, info->high_limit - 1,
1607 length))
1608 return -ENOMEM;
1609
1610 gap = mas.last + 1 - info->length;
1611 gap -= (gap - info->align_offset) & info->align_mask;
1612 return gap;
1613}
1614
1615/*
1616 * Search for an unmapped address range.
1617 *
1618 * We are looking for a range that:
1619 * - does not intersect with any VMA;
1620 * - is contained within the [low_limit, high_limit) interval;
1621 * - is at least the desired size.
1622 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1623 */
1624unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1625{
1626 unsigned long addr;
1627
1628 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1629 addr = unmapped_area_topdown(info);
1630 else
1631 addr = unmapped_area(info);
1632
1633 trace_vm_unmapped_area(addr, info);
1634 return addr;
1635}
1636
1637/* Get an address range which is currently unmapped.
1638 * For shmat() with addr=0.
1639 *
1640 * Ugly calling convention alert:
1641 * Return value with the low bits set means error value,
1642 * ie
1643 * if (ret & ~PAGE_MASK)
1644 * error = ret;
1645 *
1646 * This function "knows" that -ENOMEM has the bits set.
1647 */
1648unsigned long
1649generic_get_unmapped_area(struct file *filp, unsigned long addr,
1650 unsigned long len, unsigned long pgoff,
1651 unsigned long flags)
1652{
1653 struct mm_struct *mm = current->mm;
1654 struct vm_area_struct *vma, *prev;
1655 struct vm_unmapped_area_info info;
1656 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1657
1658 if (len > mmap_end - mmap_min_addr)
1659 return -ENOMEM;
1660
1661 if (flags & MAP_FIXED)
1662 return addr;
1663
1664 if (addr) {
1665 addr = PAGE_ALIGN(addr);
1666 vma = find_vma_prev(mm, addr, &prev);
1667 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1668 (!vma || addr + len <= vm_start_gap(vma)) &&
1669 (!prev || addr >= vm_end_gap(prev)))
1670 return addr;
1671 }
1672
1673 info.flags = 0;
1674 info.length = len;
1675 info.low_limit = mm->mmap_base;
1676 info.high_limit = mmap_end;
1677 info.align_mask = 0;
1678 info.align_offset = 0;
1679 return vm_unmapped_area(&info);
1680}
1681
1682#ifndef HAVE_ARCH_UNMAPPED_AREA
1683unsigned long
1684arch_get_unmapped_area(struct file *filp, unsigned long addr,
1685 unsigned long len, unsigned long pgoff,
1686 unsigned long flags)
1687{
1688 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1689}
1690#endif
1691
1692/*
1693 * This mmap-allocator allocates new areas top-down from below the
1694 * stack's low limit (the base):
1695 */
1696unsigned long
1697generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1698 unsigned long len, unsigned long pgoff,
1699 unsigned long flags)
1700{
1701 struct vm_area_struct *vma, *prev;
1702 struct mm_struct *mm = current->mm;
1703 struct vm_unmapped_area_info info;
1704 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1705
1706 /* requested length too big for entire address space */
1707 if (len > mmap_end - mmap_min_addr)
1708 return -ENOMEM;
1709
1710 if (flags & MAP_FIXED)
1711 return addr;
1712
1713 /* requesting a specific address */
1714 if (addr) {
1715 addr = PAGE_ALIGN(addr);
1716 vma = find_vma_prev(mm, addr, &prev);
1717 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1718 (!vma || addr + len <= vm_start_gap(vma)) &&
1719 (!prev || addr >= vm_end_gap(prev)))
1720 return addr;
1721 }
1722
1723 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1724 info.length = len;
1725 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1726 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1727 info.align_mask = 0;
1728 info.align_offset = 0;
1729 addr = vm_unmapped_area(&info);
1730
1731 /*
1732 * A failed mmap() very likely causes application failure,
1733 * so fall back to the bottom-up function here. This scenario
1734 * can happen with large stack limits and large mmap()
1735 * allocations.
1736 */
1737 if (offset_in_page(addr)) {
1738 VM_BUG_ON(addr != -ENOMEM);
1739 info.flags = 0;
1740 info.low_limit = TASK_UNMAPPED_BASE;
1741 info.high_limit = mmap_end;
1742 addr = vm_unmapped_area(&info);
1743 }
1744
1745 return addr;
1746}
1747
1748#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1749unsigned long
1750arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751 unsigned long len, unsigned long pgoff,
1752 unsigned long flags)
1753{
1754 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1755}
1756#endif
1757
1758unsigned long
1759get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1760 unsigned long pgoff, unsigned long flags)
1761{
1762 unsigned long (*get_area)(struct file *, unsigned long,
1763 unsigned long, unsigned long, unsigned long);
1764
1765 unsigned long error = arch_mmap_check(addr, len, flags);
1766 if (error)
1767 return error;
1768
1769 /* Careful about overflows.. */
1770 if (len > TASK_SIZE)
1771 return -ENOMEM;
1772
1773 get_area = current->mm->get_unmapped_area;
1774 if (file) {
1775 if (file->f_op->get_unmapped_area)
1776 get_area = file->f_op->get_unmapped_area;
1777 } else if (flags & MAP_SHARED) {
1778 /*
1779 * mmap_region() will call shmem_zero_setup() to create a file,
1780 * so use shmem's get_unmapped_area in case it can be huge.
1781 * do_mmap() will clear pgoff, so match alignment.
1782 */
1783 pgoff = 0;
1784 get_area = shmem_get_unmapped_area;
1785 }
1786
1787 addr = get_area(file, addr, len, pgoff, flags);
1788 if (IS_ERR_VALUE(addr))
1789 return addr;
1790
1791 if (addr > TASK_SIZE - len)
1792 return -ENOMEM;
1793 if (offset_in_page(addr))
1794 return -EINVAL;
1795
1796 error = security_mmap_addr(addr);
1797 return error ? error : addr;
1798}
1799
1800EXPORT_SYMBOL(get_unmapped_area);
1801
1802/**
1803 * find_vma_intersection() - Look up the first VMA which intersects the interval
1804 * @mm: The process address space.
1805 * @start_addr: The inclusive start user address.
1806 * @end_addr: The exclusive end user address.
1807 *
1808 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1809 * start_addr < end_addr.
1810 */
1811struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1812 unsigned long start_addr,
1813 unsigned long end_addr)
1814{
1815 unsigned long index = start_addr;
1816
1817 mmap_assert_locked(mm);
1818 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1819}
1820EXPORT_SYMBOL(find_vma_intersection);
1821
1822/**
1823 * find_vma() - Find the VMA for a given address, or the next VMA.
1824 * @mm: The mm_struct to check
1825 * @addr: The address
1826 *
1827 * Returns: The VMA associated with addr, or the next VMA.
1828 * May return %NULL in the case of no VMA at addr or above.
1829 */
1830struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1831{
1832 unsigned long index = addr;
1833
1834 mmap_assert_locked(mm);
1835 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1836}
1837EXPORT_SYMBOL(find_vma);
1838
1839/**
1840 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1841 * set %pprev to the previous VMA, if any.
1842 * @mm: The mm_struct to check
1843 * @addr: The address
1844 * @pprev: The pointer to set to the previous VMA
1845 *
1846 * Note that RCU lock is missing here since the external mmap_lock() is used
1847 * instead.
1848 *
1849 * Returns: The VMA associated with @addr, or the next vma.
1850 * May return %NULL in the case of no vma at addr or above.
1851 */
1852struct vm_area_struct *
1853find_vma_prev(struct mm_struct *mm, unsigned long addr,
1854 struct vm_area_struct **pprev)
1855{
1856 struct vm_area_struct *vma;
1857 MA_STATE(mas, &mm->mm_mt, addr, addr);
1858
1859 vma = mas_walk(&mas);
1860 *pprev = mas_prev(&mas, 0);
1861 if (!vma)
1862 vma = mas_next(&mas, ULONG_MAX);
1863 return vma;
1864}
1865
1866/*
1867 * Verify that the stack growth is acceptable and
1868 * update accounting. This is shared with both the
1869 * grow-up and grow-down cases.
1870 */
1871static int acct_stack_growth(struct vm_area_struct *vma,
1872 unsigned long size, unsigned long grow)
1873{
1874 struct mm_struct *mm = vma->vm_mm;
1875 unsigned long new_start;
1876
1877 /* address space limit tests */
1878 if (!may_expand_vm(mm, vma->vm_flags, grow))
1879 return -ENOMEM;
1880
1881 /* Stack limit test */
1882 if (size > rlimit(RLIMIT_STACK))
1883 return -ENOMEM;
1884
1885 /* mlock limit tests */
1886 if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT))
1887 return -ENOMEM;
1888
1889 /* Check to ensure the stack will not grow into a hugetlb-only region */
1890 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1891 vma->vm_end - size;
1892 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1893 return -EFAULT;
1894
1895 /*
1896 * Overcommit.. This must be the final test, as it will
1897 * update security statistics.
1898 */
1899 if (security_vm_enough_memory_mm(mm, grow))
1900 return -ENOMEM;
1901
1902 return 0;
1903}
1904
1905#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1906/*
1907 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1908 * vma is the last one with address > vma->vm_end. Have to extend vma.
1909 */
1910int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1911{
1912 struct mm_struct *mm = vma->vm_mm;
1913 struct vm_area_struct *next;
1914 unsigned long gap_addr;
1915 int error = 0;
1916 MA_STATE(mas, &mm->mm_mt, 0, 0);
1917
1918 if (!(vma->vm_flags & VM_GROWSUP))
1919 return -EFAULT;
1920
1921 /* Guard against exceeding limits of the address space. */
1922 address &= PAGE_MASK;
1923 if (address >= (TASK_SIZE & PAGE_MASK))
1924 return -ENOMEM;
1925 address += PAGE_SIZE;
1926
1927 /* Enforce stack_guard_gap */
1928 gap_addr = address + stack_guard_gap;
1929
1930 /* Guard against overflow */
1931 if (gap_addr < address || gap_addr > TASK_SIZE)
1932 gap_addr = TASK_SIZE;
1933
1934 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1935 if (next && vma_is_accessible(next)) {
1936 if (!(next->vm_flags & VM_GROWSUP))
1937 return -ENOMEM;
1938 /* Check that both stack segments have the same anon_vma? */
1939 }
1940
1941 if (mas_preallocate(&mas, vma, GFP_KERNEL))
1942 return -ENOMEM;
1943
1944 /* We must make sure the anon_vma is allocated. */
1945 if (unlikely(anon_vma_prepare(vma))) {
1946 mas_destroy(&mas);
1947 return -ENOMEM;
1948 }
1949
1950 /*
1951 * vma->vm_start/vm_end cannot change under us because the caller
1952 * is required to hold the mmap_lock in read mode. We need the
1953 * anon_vma lock to serialize against concurrent expand_stacks.
1954 */
1955 anon_vma_lock_write(vma->anon_vma);
1956
1957 /* Somebody else might have raced and expanded it already */
1958 if (address > vma->vm_end) {
1959 unsigned long size, grow;
1960
1961 size = address - vma->vm_start;
1962 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1963
1964 error = -ENOMEM;
1965 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1966 error = acct_stack_growth(vma, size, grow);
1967 if (!error) {
1968 /*
1969 * We only hold a shared mmap_lock lock here, so
1970 * we need to protect against concurrent vma
1971 * expansions. anon_vma_lock_write() doesn't
1972 * help here, as we don't guarantee that all
1973 * growable vmas in a mm share the same root
1974 * anon vma. So, we reuse mm->page_table_lock
1975 * to guard against concurrent vma expansions.
1976 */
1977 spin_lock(&mm->page_table_lock);
1978 if (vma->vm_flags & VM_LOCKED)
1979 mm->locked_vm += grow;
1980 vm_stat_account(mm, vma->vm_flags, grow);
1981 anon_vma_interval_tree_pre_update_vma(vma);
1982 vma->vm_end = address;
1983 /* Overwrite old entry in mtree. */
1984 vma_mas_store(vma, &mas);
1985 anon_vma_interval_tree_post_update_vma(vma);
1986 spin_unlock(&mm->page_table_lock);
1987
1988 perf_event_mmap(vma);
1989 }
1990 }
1991 }
1992 anon_vma_unlock_write(vma->anon_vma);
1993 khugepaged_enter_vma(vma, vma->vm_flags);
1994 mas_destroy(&mas);
1995 return error;
1996}
1997#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1998
1999/*
2000 * vma is the first one with address < vma->vm_start. Have to extend vma.
2001 */
2002int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2003{
2004 struct mm_struct *mm = vma->vm_mm;
2005 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2006 struct vm_area_struct *prev;
2007 int error = 0;
2008
2009 address &= PAGE_MASK;
2010 if (address < mmap_min_addr)
2011 return -EPERM;
2012
2013 /* Enforce stack_guard_gap */
2014 prev = mas_prev(&mas, 0);
2015 /* Check that both stack segments have the same anon_vma? */
2016 if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2017 vma_is_accessible(prev)) {
2018 if (address - prev->vm_end < stack_guard_gap)
2019 return -ENOMEM;
2020 }
2021
2022 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2023 return -ENOMEM;
2024
2025 /* We must make sure the anon_vma is allocated. */
2026 if (unlikely(anon_vma_prepare(vma))) {
2027 mas_destroy(&mas);
2028 return -ENOMEM;
2029 }
2030
2031 /*
2032 * vma->vm_start/vm_end cannot change under us because the caller
2033 * is required to hold the mmap_lock in read mode. We need the
2034 * anon_vma lock to serialize against concurrent expand_stacks.
2035 */
2036 anon_vma_lock_write(vma->anon_vma);
2037
2038 /* Somebody else might have raced and expanded it already */
2039 if (address < vma->vm_start) {
2040 unsigned long size, grow;
2041
2042 size = vma->vm_end - address;
2043 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2044
2045 error = -ENOMEM;
2046 if (grow <= vma->vm_pgoff) {
2047 error = acct_stack_growth(vma, size, grow);
2048 if (!error) {
2049 /*
2050 * We only hold a shared mmap_lock lock here, so
2051 * we need to protect against concurrent vma
2052 * expansions. anon_vma_lock_write() doesn't
2053 * help here, as we don't guarantee that all
2054 * growable vmas in a mm share the same root
2055 * anon vma. So, we reuse mm->page_table_lock
2056 * to guard against concurrent vma expansions.
2057 */
2058 spin_lock(&mm->page_table_lock);
2059 if (vma->vm_flags & VM_LOCKED)
2060 mm->locked_vm += grow;
2061 vm_stat_account(mm, vma->vm_flags, grow);
2062 anon_vma_interval_tree_pre_update_vma(vma);
2063 vma->vm_start = address;
2064 vma->vm_pgoff -= grow;
2065 /* Overwrite old entry in mtree. */
2066 vma_mas_store(vma, &mas);
2067 anon_vma_interval_tree_post_update_vma(vma);
2068 spin_unlock(&mm->page_table_lock);
2069
2070 perf_event_mmap(vma);
2071 }
2072 }
2073 }
2074 anon_vma_unlock_write(vma->anon_vma);
2075 khugepaged_enter_vma(vma, vma->vm_flags);
2076 mas_destroy(&mas);
2077 return error;
2078}
2079
2080/* enforced gap between the expanding stack and other mappings. */
2081unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2082
2083static int __init cmdline_parse_stack_guard_gap(char *p)
2084{
2085 unsigned long val;
2086 char *endptr;
2087
2088 val = simple_strtoul(p, &endptr, 10);
2089 if (!*endptr)
2090 stack_guard_gap = val << PAGE_SHIFT;
2091
2092 return 1;
2093}
2094__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2095
2096#ifdef CONFIG_STACK_GROWSUP
2097int expand_stack(struct vm_area_struct *vma, unsigned long address)
2098{
2099 return expand_upwards(vma, address);
2100}
2101
2102struct vm_area_struct *
2103find_extend_vma(struct mm_struct *mm, unsigned long addr)
2104{
2105 struct vm_area_struct *vma, *prev;
2106
2107 addr &= PAGE_MASK;
2108 vma = find_vma_prev(mm, addr, &prev);
2109 if (vma && (vma->vm_start <= addr))
2110 return vma;
2111 if (!prev || expand_stack(prev, addr))
2112 return NULL;
2113 if (prev->vm_flags & VM_LOCKED)
2114 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2115 return prev;
2116}
2117#else
2118int expand_stack(struct vm_area_struct *vma, unsigned long address)
2119{
2120 return expand_downwards(vma, address);
2121}
2122
2123struct vm_area_struct *
2124find_extend_vma(struct mm_struct *mm, unsigned long addr)
2125{
2126 struct vm_area_struct *vma;
2127 unsigned long start;
2128
2129 addr &= PAGE_MASK;
2130 vma = find_vma(mm, addr);
2131 if (!vma)
2132 return NULL;
2133 if (vma->vm_start <= addr)
2134 return vma;
2135 if (!(vma->vm_flags & VM_GROWSDOWN))
2136 return NULL;
2137 start = vma->vm_start;
2138 if (expand_stack(vma, addr))
2139 return NULL;
2140 if (vma->vm_flags & VM_LOCKED)
2141 populate_vma_page_range(vma, addr, start, NULL);
2142 return vma;
2143}
2144#endif
2145
2146EXPORT_SYMBOL_GPL(find_extend_vma);
2147
2148/*
2149 * Ok - we have the memory areas we should free on a maple tree so release them,
2150 * and do the vma updates.
2151 *
2152 * Called with the mm semaphore held.
2153 */
2154static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2155{
2156 unsigned long nr_accounted = 0;
2157 struct vm_area_struct *vma;
2158
2159 /* Update high watermark before we lower total_vm */
2160 update_hiwater_vm(mm);
2161 mas_for_each(mas, vma, ULONG_MAX) {
2162 long nrpages = vma_pages(vma);
2163
2164 if (vma->vm_flags & VM_ACCOUNT)
2165 nr_accounted += nrpages;
2166 vm_stat_account(mm, vma->vm_flags, -nrpages);
2167 remove_vma(vma);
2168 }
2169 vm_unacct_memory(nr_accounted);
2170 validate_mm(mm);
2171}
2172
2173/*
2174 * Get rid of page table information in the indicated region.
2175 *
2176 * Called with the mm semaphore held.
2177 */
2178static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
2179 struct vm_area_struct *vma, struct vm_area_struct *prev,
2180 struct vm_area_struct *next,
2181 unsigned long start, unsigned long end)
2182{
2183 struct mmu_gather tlb;
2184
2185 lru_add_drain();
2186 tlb_gather_mmu(&tlb, mm);
2187 update_hiwater_rss(mm);
2188 unmap_vmas(&tlb, mt, vma, start, end);
2189 free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2190 next ? next->vm_start : USER_PGTABLES_CEILING);
2191 tlb_finish_mmu(&tlb);
2192}
2193
2194/*
2195 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2196 * has already been checked or doesn't make sense to fail.
2197 */
2198int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2199 unsigned long addr, int new_below)
2200{
2201 struct vm_area_struct *new;
2202 int err;
2203 validate_mm_mt(mm);
2204
2205 if (vma->vm_ops && vma->vm_ops->may_split) {
2206 err = vma->vm_ops->may_split(vma, addr);
2207 if (err)
2208 return err;
2209 }
2210
2211 new = vm_area_dup(vma);
2212 if (!new)
2213 return -ENOMEM;
2214
2215 if (new_below)
2216 new->vm_end = addr;
2217 else {
2218 new->vm_start = addr;
2219 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2220 }
2221
2222 err = vma_dup_policy(vma, new);
2223 if (err)
2224 goto out_free_vma;
2225
2226 err = anon_vma_clone(new, vma);
2227 if (err)
2228 goto out_free_mpol;
2229
2230 if (new->vm_file)
2231 get_file(new->vm_file);
2232
2233 if (new->vm_ops && new->vm_ops->open)
2234 new->vm_ops->open(new);
2235
2236 if (new_below)
2237 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2238 ((addr - new->vm_start) >> PAGE_SHIFT), new);
2239 else
2240 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2241
2242 /* Success. */
2243 if (!err)
2244 return 0;
2245
2246 /* Avoid vm accounting in close() operation */
2247 new->vm_start = new->vm_end;
2248 new->vm_pgoff = 0;
2249 /* Clean everything up if vma_adjust failed. */
2250 if (new->vm_ops && new->vm_ops->close)
2251 new->vm_ops->close(new);
2252 if (new->vm_file)
2253 fput(new->vm_file);
2254 unlink_anon_vmas(new);
2255 out_free_mpol:
2256 mpol_put(vma_policy(new));
2257 out_free_vma:
2258 vm_area_free(new);
2259 validate_mm_mt(mm);
2260 return err;
2261}
2262
2263/*
2264 * Split a vma into two pieces at address 'addr', a new vma is allocated
2265 * either for the first part or the tail.
2266 */
2267int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2268 unsigned long addr, int new_below)
2269{
2270 if (mm->map_count >= sysctl_max_map_count)
2271 return -ENOMEM;
2272
2273 return __split_vma(mm, vma, addr, new_below);
2274}
2275
2276static inline int munmap_sidetree(struct vm_area_struct *vma,
2277 struct ma_state *mas_detach)
2278{
2279 mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1);
2280 if (mas_store_gfp(mas_detach, vma, GFP_KERNEL))
2281 return -ENOMEM;
2282
2283 if (vma->vm_flags & VM_LOCKED)
2284 vma->vm_mm->locked_vm -= vma_pages(vma);
2285
2286 return 0;
2287}
2288
2289/*
2290 * do_mas_align_munmap() - munmap the aligned region from @start to @end.
2291 * @mas: The maple_state, ideally set up to alter the correct tree location.
2292 * @vma: The starting vm_area_struct
2293 * @mm: The mm_struct
2294 * @start: The aligned start address to munmap.
2295 * @end: The aligned end address to munmap.
2296 * @uf: The userfaultfd list_head
2297 * @downgrade: Set to true to attempt a write downgrade of the mmap_lock
2298 *
2299 * If @downgrade is true, check return code for potential release of the lock.
2300 */
2301static int
2302do_mas_align_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2303 struct mm_struct *mm, unsigned long start,
2304 unsigned long end, struct list_head *uf, bool downgrade)
2305{
2306 struct vm_area_struct *prev, *next = NULL;
2307 struct maple_tree mt_detach;
2308 int count = 0;
2309 int error = -ENOMEM;
2310 MA_STATE(mas_detach, &mt_detach, 0, 0);
2311 mt_init_flags(&mt_detach, MT_FLAGS_LOCK_EXTERN);
2312 mt_set_external_lock(&mt_detach, &mm->mmap_lock);
2313
2314 if (mas_preallocate(mas, vma, GFP_KERNEL))
2315 return -ENOMEM;
2316
2317 mas->last = end - 1;
2318 /*
2319 * If we need to split any vma, do it now to save pain later.
2320 *
2321 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2322 * unmapped vm_area_struct will remain in use: so lower split_vma
2323 * places tmp vma above, and higher split_vma places tmp vma below.
2324 */
2325
2326 /* Does it split the first one? */
2327 if (start > vma->vm_start) {
2328
2329 /*
2330 * Make sure that map_count on return from munmap() will
2331 * not exceed its limit; but let map_count go just above
2332 * its limit temporarily, to help free resources as expected.
2333 */
2334 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2335 goto map_count_exceeded;
2336
2337 /*
2338 * mas_pause() is not needed since mas->index needs to be set
2339 * differently than vma->vm_end anyways.
2340 */
2341 error = __split_vma(mm, vma, start, 0);
2342 if (error)
2343 goto start_split_failed;
2344
2345 mas_set(mas, start);
2346 vma = mas_walk(mas);
2347 }
2348
2349 prev = mas_prev(mas, 0);
2350 if (unlikely((!prev)))
2351 mas_set(mas, start);
2352
2353 /*
2354 * Detach a range of VMAs from the mm. Using next as a temp variable as
2355 * it is always overwritten.
2356 */
2357 mas_for_each(mas, next, end - 1) {
2358 /* Does it split the end? */
2359 if (next->vm_end > end) {
2360 struct vm_area_struct *split;
2361
2362 error = __split_vma(mm, next, end, 1);
2363 if (error)
2364 goto end_split_failed;
2365
2366 mas_set(mas, end);
2367 split = mas_prev(mas, 0);
2368 error = munmap_sidetree(split, &mas_detach);
2369 if (error)
2370 goto munmap_sidetree_failed;
2371
2372 count++;
2373 if (vma == next)
2374 vma = split;
2375 break;
2376 }
2377 error = munmap_sidetree(next, &mas_detach);
2378 if (error)
2379 goto munmap_sidetree_failed;
2380
2381 count++;
2382#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2383 BUG_ON(next->vm_start < start);
2384 BUG_ON(next->vm_start > end);
2385#endif
2386 }
2387
2388 if (!next)
2389 next = mas_next(mas, ULONG_MAX);
2390
2391 if (unlikely(uf)) {
2392 /*
2393 * If userfaultfd_unmap_prep returns an error the vmas
2394 * will remain split, but userland will get a
2395 * highly unexpected error anyway. This is no
2396 * different than the case where the first of the two
2397 * __split_vma fails, but we don't undo the first
2398 * split, despite we could. This is unlikely enough
2399 * failure that it's not worth optimizing it for.
2400 */
2401 error = userfaultfd_unmap_prep(mm, start, end, uf);
2402
2403 if (error)
2404 goto userfaultfd_error;
2405 }
2406
2407 /* Point of no return */
2408 mas_set_range(mas, start, end - 1);
2409#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2410 /* Make sure no VMAs are about to be lost. */
2411 {
2412 MA_STATE(test, &mt_detach, start, end - 1);
2413 struct vm_area_struct *vma_mas, *vma_test;
2414 int test_count = 0;
2415
2416 rcu_read_lock();
2417 vma_test = mas_find(&test, end - 1);
2418 mas_for_each(mas, vma_mas, end - 1) {
2419 BUG_ON(vma_mas != vma_test);
2420 test_count++;
2421 vma_test = mas_next(&test, end - 1);
2422 }
2423 rcu_read_unlock();
2424 BUG_ON(count != test_count);
2425 mas_set_range(mas, start, end - 1);
2426 }
2427#endif
2428 mas_store_prealloc(mas, NULL);
2429 mm->map_count -= count;
2430 /*
2431 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2432 * VM_GROWSUP VMA. Such VMAs can change their size under
2433 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2434 */
2435 if (downgrade) {
2436 if (next && (next->vm_flags & VM_GROWSDOWN))
2437 downgrade = false;
2438 else if (prev && (prev->vm_flags & VM_GROWSUP))
2439 downgrade = false;
2440 else
2441 mmap_write_downgrade(mm);
2442 }
2443
2444 unmap_region(mm, &mt_detach, vma, prev, next, start, end);
2445 /* Statistics and freeing VMAs */
2446 mas_set(&mas_detach, start);
2447 remove_mt(mm, &mas_detach);
2448 __mt_destroy(&mt_detach);
2449
2450
2451 validate_mm(mm);
2452 return downgrade ? 1 : 0;
2453
2454userfaultfd_error:
2455munmap_sidetree_failed:
2456end_split_failed:
2457 __mt_destroy(&mt_detach);
2458start_split_failed:
2459map_count_exceeded:
2460 mas_destroy(mas);
2461 return error;
2462}
2463
2464/*
2465 * do_mas_munmap() - munmap a given range.
2466 * @mas: The maple state
2467 * @mm: The mm_struct
2468 * @start: The start address to munmap
2469 * @len: The length of the range to munmap
2470 * @uf: The userfaultfd list_head
2471 * @downgrade: set to true if the user wants to attempt to write_downgrade the
2472 * mmap_lock
2473 *
2474 * This function takes a @mas that is either pointing to the previous VMA or set
2475 * to MA_START and sets it up to remove the mapping(s). The @len will be
2476 * aligned and any arch_unmap work will be preformed.
2477 *
2478 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise.
2479 */
2480int do_mas_munmap(struct ma_state *mas, struct mm_struct *mm,
2481 unsigned long start, size_t len, struct list_head *uf,
2482 bool downgrade)
2483{
2484 unsigned long end;
2485 struct vm_area_struct *vma;
2486
2487 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2488 return -EINVAL;
2489
2490 end = start + PAGE_ALIGN(len);
2491 if (end == start)
2492 return -EINVAL;
2493
2494 /* arch_unmap() might do unmaps itself. */
2495 arch_unmap(mm, start, end);
2496
2497 /* Find the first overlapping VMA */
2498 vma = mas_find(mas, end - 1);
2499 if (!vma)
2500 return 0;
2501
2502 return do_mas_align_munmap(mas, vma, mm, start, end, uf, downgrade);
2503}
2504
2505/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2506 * @mm: The mm_struct
2507 * @start: The start address to munmap
2508 * @len: The length to be munmapped.
2509 * @uf: The userfaultfd list_head
2510 */
2511int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2512 struct list_head *uf)
2513{
2514 MA_STATE(mas, &mm->mm_mt, start, start);
2515
2516 return do_mas_munmap(&mas, mm, start, len, uf, false);
2517}
2518
2519unsigned long mmap_region(struct file *file, unsigned long addr,
2520 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2521 struct list_head *uf)
2522{
2523 struct mm_struct *mm = current->mm;
2524 struct vm_area_struct *vma = NULL;
2525 struct vm_area_struct *next, *prev, *merge;
2526 pgoff_t pglen = len >> PAGE_SHIFT;
2527 unsigned long charged = 0;
2528 unsigned long end = addr + len;
2529 unsigned long merge_start = addr, merge_end = end;
2530 pgoff_t vm_pgoff;
2531 int error;
2532 MA_STATE(mas, &mm->mm_mt, addr, end - 1);
2533
2534 /* Check against address space limit. */
2535 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2536 unsigned long nr_pages;
2537
2538 /*
2539 * MAP_FIXED may remove pages of mappings that intersects with
2540 * requested mapping. Account for the pages it would unmap.
2541 */
2542 nr_pages = count_vma_pages_range(mm, addr, end);
2543
2544 if (!may_expand_vm(mm, vm_flags,
2545 (len >> PAGE_SHIFT) - nr_pages))
2546 return -ENOMEM;
2547 }
2548
2549 /* Unmap any existing mapping in the area */
2550 if (do_mas_munmap(&mas, mm, addr, len, uf, false))
2551 return -ENOMEM;
2552
2553 /*
2554 * Private writable mapping: check memory availability
2555 */
2556 if (accountable_mapping(file, vm_flags)) {
2557 charged = len >> PAGE_SHIFT;
2558 if (security_vm_enough_memory_mm(mm, charged))
2559 return -ENOMEM;
2560 vm_flags |= VM_ACCOUNT;
2561 }
2562
2563 next = mas_next(&mas, ULONG_MAX);
2564 prev = mas_prev(&mas, 0);
2565 if (vm_flags & VM_SPECIAL)
2566 goto cannot_expand;
2567
2568 /* Attempt to expand an old mapping */
2569 /* Check next */
2570 if (next && next->vm_start == end && !vma_policy(next) &&
2571 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2572 NULL_VM_UFFD_CTX, NULL)) {
2573 merge_end = next->vm_end;
2574 vma = next;
2575 vm_pgoff = next->vm_pgoff - pglen;
2576 }
2577
2578 /* Check prev */
2579 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2580 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2581 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2582 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2583 NULL_VM_UFFD_CTX, NULL))) {
2584 merge_start = prev->vm_start;
2585 vma = prev;
2586 vm_pgoff = prev->vm_pgoff;
2587 }
2588
2589
2590 /* Actually expand, if possible */
2591 if (vma &&
2592 !vma_expand(&mas, vma, merge_start, merge_end, vm_pgoff, next)) {
2593 khugepaged_enter_vma(vma, vm_flags);
2594 goto expanded;
2595 }
2596
2597 mas.index = addr;
2598 mas.last = end - 1;
2599cannot_expand:
2600 /*
2601 * Determine the object being mapped and call the appropriate
2602 * specific mapper. the address has already been validated, but
2603 * not unmapped, but the maps are removed from the list.
2604 */
2605 vma = vm_area_alloc(mm);
2606 if (!vma) {
2607 error = -ENOMEM;
2608 goto unacct_error;
2609 }
2610
2611 vma->vm_start = addr;
2612 vma->vm_end = end;
2613 vma->vm_flags = vm_flags;
2614 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2615 vma->vm_pgoff = pgoff;
2616
2617 if (file) {
2618 if (vm_flags & VM_SHARED) {
2619 error = mapping_map_writable(file->f_mapping);
2620 if (error)
2621 goto free_vma;
2622 }
2623
2624 vma->vm_file = get_file(file);
2625 error = call_mmap(file, vma);
2626 if (error)
2627 goto unmap_and_free_vma;
2628
2629 /*
2630 * Expansion is handled above, merging is handled below.
2631 * Drivers should not alter the address of the VMA.
2632 */
2633 if (WARN_ON((addr != vma->vm_start))) {
2634 error = -EINVAL;
2635 goto close_and_free_vma;
2636 }
2637 mas_reset(&mas);
2638
2639 /*
2640 * If vm_flags changed after call_mmap(), we should try merge
2641 * vma again as we may succeed this time.
2642 */
2643 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2644 merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
2645 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2646 if (merge) {
2647 /*
2648 * ->mmap() can change vma->vm_file and fput
2649 * the original file. So fput the vma->vm_file
2650 * here or we would add an extra fput for file
2651 * and cause general protection fault
2652 * ultimately.
2653 */
2654 fput(vma->vm_file);
2655 vm_area_free(vma);
2656 vma = merge;
2657 /* Update vm_flags to pick up the change. */
2658 vm_flags = vma->vm_flags;
2659 goto unmap_writable;
2660 }
2661 }
2662
2663 vm_flags = vma->vm_flags;
2664 } else if (vm_flags & VM_SHARED) {
2665 error = shmem_zero_setup(vma);
2666 if (error)
2667 goto free_vma;
2668 } else {
2669 vma_set_anonymous(vma);
2670 }
2671
2672 /* Allow architectures to sanity-check the vm_flags */
2673 if (!arch_validate_flags(vma->vm_flags)) {
2674 error = -EINVAL;
2675 if (file)
2676 goto close_and_free_vma;
2677 else if (vma->vm_file)
2678 goto unmap_and_free_vma;
2679 else
2680 goto free_vma;
2681 }
2682
2683 if (mas_preallocate(&mas, vma, GFP_KERNEL)) {
2684 error = -ENOMEM;
2685 if (file)
2686 goto close_and_free_vma;
2687 else if (vma->vm_file)
2688 goto unmap_and_free_vma;
2689 else
2690 goto free_vma;
2691 }
2692
2693 if (vma->vm_file)
2694 i_mmap_lock_write(vma->vm_file->f_mapping);
2695
2696 vma_mas_store(vma, &mas);
2697 mm->map_count++;
2698 if (vma->vm_file) {
2699 if (vma->vm_flags & VM_SHARED)
2700 mapping_allow_writable(vma->vm_file->f_mapping);
2701
2702 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2703 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2704 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2705 i_mmap_unlock_write(vma->vm_file->f_mapping);
2706 }
2707
2708 /*
2709 * vma_merge() calls khugepaged_enter_vma() either, the below
2710 * call covers the non-merge case.
2711 */
2712 khugepaged_enter_vma(vma, vma->vm_flags);
2713
2714 /* Once vma denies write, undo our temporary denial count */
2715unmap_writable:
2716 if (file && vm_flags & VM_SHARED)
2717 mapping_unmap_writable(file->f_mapping);
2718 file = vma->vm_file;
2719expanded:
2720 perf_event_mmap(vma);
2721
2722 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2723 if (vm_flags & VM_LOCKED) {
2724 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2725 is_vm_hugetlb_page(vma) ||
2726 vma == get_gate_vma(current->mm))
2727 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
2728 else
2729 mm->locked_vm += (len >> PAGE_SHIFT);
2730 }
2731
2732 if (file)
2733 uprobe_mmap(vma);
2734
2735 /*
2736 * New (or expanded) vma always get soft dirty status.
2737 * Otherwise user-space soft-dirty page tracker won't
2738 * be able to distinguish situation when vma area unmapped,
2739 * then new mapped in-place (which must be aimed as
2740 * a completely new data area).
2741 */
2742 vma->vm_flags |= VM_SOFTDIRTY;
2743
2744 vma_set_page_prot(vma);
2745
2746 validate_mm(mm);
2747 return addr;
2748
2749close_and_free_vma:
2750 if (vma->vm_ops && vma->vm_ops->close)
2751 vma->vm_ops->close(vma);
2752unmap_and_free_vma:
2753 fput(vma->vm_file);
2754 vma->vm_file = NULL;
2755
2756 /* Undo any partial mapping done by a device driver. */
2757 unmap_region(mm, mas.tree, vma, prev, next, vma->vm_start, vma->vm_end);
2758 if (file && (vm_flags & VM_SHARED))
2759 mapping_unmap_writable(file->f_mapping);
2760free_vma:
2761 vm_area_free(vma);
2762unacct_error:
2763 if (charged)
2764 vm_unacct_memory(charged);
2765 validate_mm(mm);
2766 return error;
2767}
2768
2769static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2770{
2771 int ret;
2772 struct mm_struct *mm = current->mm;
2773 LIST_HEAD(uf);
2774 MA_STATE(mas, &mm->mm_mt, start, start);
2775
2776 if (mmap_write_lock_killable(mm))
2777 return -EINTR;
2778
2779 ret = do_mas_munmap(&mas, mm, start, len, &uf, downgrade);
2780 /*
2781 * Returning 1 indicates mmap_lock is downgraded.
2782 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2783 * it to 0 before return.
2784 */
2785 if (ret == 1) {
2786 mmap_read_unlock(mm);
2787 ret = 0;
2788 } else
2789 mmap_write_unlock(mm);
2790
2791 userfaultfd_unmap_complete(mm, &uf);
2792 return ret;
2793}
2794
2795int vm_munmap(unsigned long start, size_t len)
2796{
2797 return __vm_munmap(start, len, false);
2798}
2799EXPORT_SYMBOL(vm_munmap);
2800
2801SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2802{
2803 addr = untagged_addr(addr);
2804 return __vm_munmap(addr, len, true);
2805}
2806
2807
2808/*
2809 * Emulation of deprecated remap_file_pages() syscall.
2810 */
2811SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2812 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2813{
2814
2815 struct mm_struct *mm = current->mm;
2816 struct vm_area_struct *vma;
2817 unsigned long populate = 0;
2818 unsigned long ret = -EINVAL;
2819 struct file *file;
2820
2821 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2822 current->comm, current->pid);
2823
2824 if (prot)
2825 return ret;
2826 start = start & PAGE_MASK;
2827 size = size & PAGE_MASK;
2828
2829 if (start + size <= start)
2830 return ret;
2831
2832 /* Does pgoff wrap? */
2833 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2834 return ret;
2835
2836 if (mmap_write_lock_killable(mm))
2837 return -EINTR;
2838
2839 vma = vma_lookup(mm, start);
2840
2841 if (!vma || !(vma->vm_flags & VM_SHARED))
2842 goto out;
2843
2844 if (start + size > vma->vm_end) {
2845 VMA_ITERATOR(vmi, mm, vma->vm_end);
2846 struct vm_area_struct *next, *prev = vma;
2847
2848 for_each_vma_range(vmi, next, start + size) {
2849 /* hole between vmas ? */
2850 if (next->vm_start != prev->vm_end)
2851 goto out;
2852
2853 if (next->vm_file != vma->vm_file)
2854 goto out;
2855
2856 if (next->vm_flags != vma->vm_flags)
2857 goto out;
2858
2859 if (start + size <= next->vm_end)
2860 break;
2861
2862 prev = next;
2863 }
2864
2865 if (!next)
2866 goto out;
2867 }
2868
2869 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2870 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2871 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2872
2873 flags &= MAP_NONBLOCK;
2874 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2875 if (vma->vm_flags & VM_LOCKED)
2876 flags |= MAP_LOCKED;
2877
2878 file = get_file(vma->vm_file);
2879 ret = do_mmap(vma->vm_file, start, size,
2880 prot, flags, pgoff, &populate, NULL);
2881 fput(file);
2882out:
2883 mmap_write_unlock(mm);
2884 if (populate)
2885 mm_populate(ret, populate);
2886 if (!IS_ERR_VALUE(ret))
2887 ret = 0;
2888 return ret;
2889}
2890
2891/*
2892 * brk_munmap() - Unmap a parital vma.
2893 * @mas: The maple tree state.
2894 * @vma: The vma to be modified
2895 * @newbrk: the start of the address to unmap
2896 * @oldbrk: The end of the address to unmap
2897 * @uf: The userfaultfd list_head
2898 *
2899 * Returns: 1 on success.
2900 * unmaps a partial VMA mapping. Does not handle alignment, downgrades lock if
2901 * possible.
2902 */
2903static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2904 unsigned long newbrk, unsigned long oldbrk,
2905 struct list_head *uf)
2906{
2907 struct mm_struct *mm = vma->vm_mm;
2908 int ret;
2909
2910 arch_unmap(mm, newbrk, oldbrk);
2911 ret = do_mas_align_munmap(mas, vma, mm, newbrk, oldbrk, uf, true);
2912 validate_mm_mt(mm);
2913 return ret;
2914}
2915
2916/*
2917 * do_brk_flags() - Increase the brk vma if the flags match.
2918 * @mas: The maple tree state.
2919 * @addr: The start address
2920 * @len: The length of the increase
2921 * @vma: The vma,
2922 * @flags: The VMA Flags
2923 *
2924 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
2925 * do not match then create a new anonymous VMA. Eventually we may be able to
2926 * do some brk-specific accounting here.
2927 */
2928static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *vma,
2929 unsigned long addr, unsigned long len, unsigned long flags)
2930{
2931 struct mm_struct *mm = current->mm;
2932
2933 validate_mm_mt(mm);
2934 /*
2935 * Check against address space limits by the changed size
2936 * Note: This happens *after* clearing old mappings in some code paths.
2937 */
2938 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2939 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2940 return -ENOMEM;
2941
2942 if (mm->map_count > sysctl_max_map_count)
2943 return -ENOMEM;
2944
2945 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2946 return -ENOMEM;
2947
2948 /*
2949 * Expand the existing vma if possible; Note that singular lists do not
2950 * occur after forking, so the expand will only happen on new VMAs.
2951 */
2952 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
2953 can_vma_merge_after(vma, flags, NULL, NULL,
2954 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
2955 mas_set_range(mas, vma->vm_start, addr + len - 1);
2956 if (mas_preallocate(mas, vma, GFP_KERNEL))
2957 goto unacct_fail;
2958
2959 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
2960 if (vma->anon_vma) {
2961 anon_vma_lock_write(vma->anon_vma);
2962 anon_vma_interval_tree_pre_update_vma(vma);
2963 }
2964 vma->vm_end = addr + len;
2965 vma->vm_flags |= VM_SOFTDIRTY;
2966 mas_store_prealloc(mas, vma);
2967
2968 if (vma->anon_vma) {
2969 anon_vma_interval_tree_post_update_vma(vma);
2970 anon_vma_unlock_write(vma->anon_vma);
2971 }
2972 khugepaged_enter_vma(vma, flags);
2973 goto out;
2974 }
2975
2976 /* create a vma struct for an anonymous mapping */
2977 vma = vm_area_alloc(mm);
2978 if (!vma)
2979 goto unacct_fail;
2980
2981 vma_set_anonymous(vma);
2982 vma->vm_start = addr;
2983 vma->vm_end = addr + len;
2984 vma->vm_pgoff = addr >> PAGE_SHIFT;
2985 vma->vm_flags = flags;
2986 vma->vm_page_prot = vm_get_page_prot(flags);
2987 mas_set_range(mas, vma->vm_start, addr + len - 1);
2988 if (mas_store_gfp(mas, vma, GFP_KERNEL))
2989 goto mas_store_fail;
2990
2991 mm->map_count++;
2992out:
2993 perf_event_mmap(vma);
2994 mm->total_vm += len >> PAGE_SHIFT;
2995 mm->data_vm += len >> PAGE_SHIFT;
2996 if (flags & VM_LOCKED)
2997 mm->locked_vm += (len >> PAGE_SHIFT);
2998 vma->vm_flags |= VM_SOFTDIRTY;
2999 validate_mm(mm);
3000 return 0;
3001
3002mas_store_fail:
3003 vm_area_free(vma);
3004unacct_fail:
3005 vm_unacct_memory(len >> PAGE_SHIFT);
3006 return -ENOMEM;
3007}
3008
3009int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3010{
3011 struct mm_struct *mm = current->mm;
3012 struct vm_area_struct *vma = NULL;
3013 unsigned long len;
3014 int ret;
3015 bool populate;
3016 LIST_HEAD(uf);
3017 MA_STATE(mas, &mm->mm_mt, addr, addr);
3018
3019 len = PAGE_ALIGN(request);
3020 if (len < request)
3021 return -ENOMEM;
3022 if (!len)
3023 return 0;
3024
3025 if (mmap_write_lock_killable(mm))
3026 return -EINTR;
3027
3028 /* Until we need other flags, refuse anything except VM_EXEC. */
3029 if ((flags & (~VM_EXEC)) != 0)
3030 return -EINVAL;
3031
3032 ret = check_brk_limits(addr, len);
3033 if (ret)
3034 goto limits_failed;
3035
3036 ret = do_mas_munmap(&mas, mm, addr, len, &uf, 0);
3037 if (ret)
3038 goto munmap_failed;
3039
3040 vma = mas_prev(&mas, 0);
3041 ret = do_brk_flags(&mas, vma, addr, len, flags);
3042 populate = ((mm->def_flags & VM_LOCKED) != 0);
3043 mmap_write_unlock(mm);
3044 userfaultfd_unmap_complete(mm, &uf);
3045 if (populate && !ret)
3046 mm_populate(addr, len);
3047 return ret;
3048
3049munmap_failed:
3050limits_failed:
3051 mmap_write_unlock(mm);
3052 return ret;
3053}
3054EXPORT_SYMBOL(vm_brk_flags);
3055
3056int vm_brk(unsigned long addr, unsigned long len)
3057{
3058 return vm_brk_flags(addr, len, 0);
3059}
3060EXPORT_SYMBOL(vm_brk);
3061
3062/* Release all mmaps. */
3063void exit_mmap(struct mm_struct *mm)
3064{
3065 struct mmu_gather tlb;
3066 struct vm_area_struct *vma;
3067 unsigned long nr_accounted = 0;
3068 MA_STATE(mas, &mm->mm_mt, 0, 0);
3069 int count = 0;
3070
3071 /* mm's last user has gone, and its about to be pulled down */
3072 mmu_notifier_release(mm);
3073
3074 mmap_read_lock(mm);
3075 arch_exit_mmap(mm);
3076
3077 vma = mas_find(&mas, ULONG_MAX);
3078 if (!vma) {
3079 /* Can happen if dup_mmap() received an OOM */
3080 mmap_read_unlock(mm);
3081 return;
3082 }
3083
3084 lru_add_drain();
3085 flush_cache_mm(mm);
3086 tlb_gather_mmu_fullmm(&tlb, mm);
3087 /* update_hiwater_rss(mm) here? but nobody should be looking */
3088 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3089 unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX);
3090 mmap_read_unlock(mm);
3091
3092 /*
3093 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3094 * because the memory has been already freed.
3095 */
3096 set_bit(MMF_OOM_SKIP, &mm->flags);
3097 mmap_write_lock(mm);
3098 free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS,
3099 USER_PGTABLES_CEILING);
3100 tlb_finish_mmu(&tlb);
3101
3102 /*
3103 * Walk the list again, actually closing and freeing it, with preemption
3104 * enabled, without holding any MM locks besides the unreachable
3105 * mmap_write_lock.
3106 */
3107 do {
3108 if (vma->vm_flags & VM_ACCOUNT)
3109 nr_accounted += vma_pages(vma);
3110 remove_vma(vma);
3111 count++;
3112 cond_resched();
3113 } while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3114
3115 BUG_ON(count != mm->map_count);
3116
3117 trace_exit_mmap(mm);
3118 __mt_destroy(&mm->mm_mt);
3119 mmap_write_unlock(mm);
3120 vm_unacct_memory(nr_accounted);
3121}
3122
3123/* Insert vm structure into process list sorted by address
3124 * and into the inode's i_mmap tree. If vm_file is non-NULL
3125 * then i_mmap_rwsem is taken here.
3126 */
3127int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3128{
3129 unsigned long charged = vma_pages(vma);
3130
3131
3132 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3133 return -ENOMEM;
3134
3135 if ((vma->vm_flags & VM_ACCOUNT) &&
3136 security_vm_enough_memory_mm(mm, charged))
3137 return -ENOMEM;
3138
3139 /*
3140 * The vm_pgoff of a purely anonymous vma should be irrelevant
3141 * until its first write fault, when page's anon_vma and index
3142 * are set. But now set the vm_pgoff it will almost certainly
3143 * end up with (unless mremap moves it elsewhere before that
3144 * first wfault), so /proc/pid/maps tells a consistent story.
3145 *
3146 * By setting it to reflect the virtual start address of the
3147 * vma, merges and splits can happen in a seamless way, just
3148 * using the existing file pgoff checks and manipulations.
3149 * Similarly in do_mmap and in do_brk_flags.
3150 */
3151 if (vma_is_anonymous(vma)) {
3152 BUG_ON(vma->anon_vma);
3153 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3154 }
3155
3156 if (vma_link(mm, vma)) {
3157 vm_unacct_memory(charged);
3158 return -ENOMEM;
3159 }
3160
3161 return 0;
3162}
3163
3164/*
3165 * Copy the vma structure to a new location in the same mm,
3166 * prior to moving page table entries, to effect an mremap move.
3167 */
3168struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3169 unsigned long addr, unsigned long len, pgoff_t pgoff,
3170 bool *need_rmap_locks)
3171{
3172 struct vm_area_struct *vma = *vmap;
3173 unsigned long vma_start = vma->vm_start;
3174 struct mm_struct *mm = vma->vm_mm;
3175 struct vm_area_struct *new_vma, *prev;
3176 bool faulted_in_anon_vma = true;
3177
3178 validate_mm_mt(mm);
3179 /*
3180 * If anonymous vma has not yet been faulted, update new pgoff
3181 * to match new location, to increase its chance of merging.
3182 */
3183 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3184 pgoff = addr >> PAGE_SHIFT;
3185 faulted_in_anon_vma = false;
3186 }
3187
3188 new_vma = find_vma_prev(mm, addr, &prev);
3189 if (new_vma && new_vma->vm_start < addr + len)
3190 return NULL; /* should never get here */
3191
3192 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3193 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3194 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3195 if (new_vma) {
3196 /*
3197 * Source vma may have been merged into new_vma
3198 */
3199 if (unlikely(vma_start >= new_vma->vm_start &&
3200 vma_start < new_vma->vm_end)) {
3201 /*
3202 * The only way we can get a vma_merge with
3203 * self during an mremap is if the vma hasn't
3204 * been faulted in yet and we were allowed to
3205 * reset the dst vma->vm_pgoff to the
3206 * destination address of the mremap to allow
3207 * the merge to happen. mremap must change the
3208 * vm_pgoff linearity between src and dst vmas
3209 * (in turn preventing a vma_merge) to be
3210 * safe. It is only safe to keep the vm_pgoff
3211 * linear if there are no pages mapped yet.
3212 */
3213 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3214 *vmap = vma = new_vma;
3215 }
3216 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3217 } else {
3218 new_vma = vm_area_dup(vma);
3219 if (!new_vma)
3220 goto out;
3221 new_vma->vm_start = addr;
3222 new_vma->vm_end = addr + len;
3223 new_vma->vm_pgoff = pgoff;
3224 if (vma_dup_policy(vma, new_vma))
3225 goto out_free_vma;
3226 if (anon_vma_clone(new_vma, vma))
3227 goto out_free_mempol;
3228 if (new_vma->vm_file)
3229 get_file(new_vma->vm_file);
3230 if (new_vma->vm_ops && new_vma->vm_ops->open)
3231 new_vma->vm_ops->open(new_vma);
3232 if (vma_link(mm, new_vma))
3233 goto out_vma_link;
3234 *need_rmap_locks = false;
3235 }
3236 validate_mm_mt(mm);
3237 return new_vma;
3238
3239out_vma_link:
3240 if (new_vma->vm_ops && new_vma->vm_ops->close)
3241 new_vma->vm_ops->close(new_vma);
3242
3243 if (new_vma->vm_file)
3244 fput(new_vma->vm_file);
3245
3246 unlink_anon_vmas(new_vma);
3247out_free_mempol:
3248 mpol_put(vma_policy(new_vma));
3249out_free_vma:
3250 vm_area_free(new_vma);
3251out:
3252 validate_mm_mt(mm);
3253 return NULL;
3254}
3255
3256/*
3257 * Return true if the calling process may expand its vm space by the passed
3258 * number of pages
3259 */
3260bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3261{
3262 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3263 return false;
3264
3265 if (is_data_mapping(flags) &&
3266 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3267 /* Workaround for Valgrind */
3268 if (rlimit(RLIMIT_DATA) == 0 &&
3269 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3270 return true;
3271
3272 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3273 current->comm, current->pid,
3274 (mm->data_vm + npages) << PAGE_SHIFT,
3275 rlimit(RLIMIT_DATA),
3276 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3277
3278 if (!ignore_rlimit_data)
3279 return false;
3280 }
3281
3282 return true;
3283}
3284
3285void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3286{
3287 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3288
3289 if (is_exec_mapping(flags))
3290 mm->exec_vm += npages;
3291 else if (is_stack_mapping(flags))
3292 mm->stack_vm += npages;
3293 else if (is_data_mapping(flags))
3294 mm->data_vm += npages;
3295}
3296
3297static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3298
3299/*
3300 * Having a close hook prevents vma merging regardless of flags.
3301 */
3302static void special_mapping_close(struct vm_area_struct *vma)
3303{
3304}
3305
3306static const char *special_mapping_name(struct vm_area_struct *vma)
3307{
3308 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3309}
3310
3311static int special_mapping_mremap(struct vm_area_struct *new_vma)
3312{
3313 struct vm_special_mapping *sm = new_vma->vm_private_data;
3314
3315 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3316 return -EFAULT;
3317
3318 if (sm->mremap)
3319 return sm->mremap(sm, new_vma);
3320
3321 return 0;
3322}
3323
3324static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3325{
3326 /*
3327 * Forbid splitting special mappings - kernel has expectations over
3328 * the number of pages in mapping. Together with VM_DONTEXPAND
3329 * the size of vma should stay the same over the special mapping's
3330 * lifetime.
3331 */
3332 return -EINVAL;
3333}
3334
3335static const struct vm_operations_struct special_mapping_vmops = {
3336 .close = special_mapping_close,
3337 .fault = special_mapping_fault,
3338 .mremap = special_mapping_mremap,
3339 .name = special_mapping_name,
3340 /* vDSO code relies that VVAR can't be accessed remotely */
3341 .access = NULL,
3342 .may_split = special_mapping_split,
3343};
3344
3345static const struct vm_operations_struct legacy_special_mapping_vmops = {
3346 .close = special_mapping_close,
3347 .fault = special_mapping_fault,
3348};
3349
3350static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3351{
3352 struct vm_area_struct *vma = vmf->vma;
3353 pgoff_t pgoff;
3354 struct page **pages;
3355
3356 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3357 pages = vma->vm_private_data;
3358 } else {
3359 struct vm_special_mapping *sm = vma->vm_private_data;
3360
3361 if (sm->fault)
3362 return sm->fault(sm, vmf->vma, vmf);
3363
3364 pages = sm->pages;
3365 }
3366
3367 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3368 pgoff--;
3369
3370 if (*pages) {
3371 struct page *page = *pages;
3372 get_page(page);
3373 vmf->page = page;
3374 return 0;
3375 }
3376
3377 return VM_FAULT_SIGBUS;
3378}
3379
3380static struct vm_area_struct *__install_special_mapping(
3381 struct mm_struct *mm,
3382 unsigned long addr, unsigned long len,
3383 unsigned long vm_flags, void *priv,
3384 const struct vm_operations_struct *ops)
3385{
3386 int ret;
3387 struct vm_area_struct *vma;
3388
3389 validate_mm_mt(mm);
3390 vma = vm_area_alloc(mm);
3391 if (unlikely(vma == NULL))
3392 return ERR_PTR(-ENOMEM);
3393
3394 vma->vm_start = addr;
3395 vma->vm_end = addr + len;
3396
3397 vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3398 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
3399 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3400
3401 vma->vm_ops = ops;
3402 vma->vm_private_data = priv;
3403
3404 ret = insert_vm_struct(mm, vma);
3405 if (ret)
3406 goto out;
3407
3408 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3409
3410 perf_event_mmap(vma);
3411
3412 validate_mm_mt(mm);
3413 return vma;
3414
3415out:
3416 vm_area_free(vma);
3417 validate_mm_mt(mm);
3418 return ERR_PTR(ret);
3419}
3420
3421bool vma_is_special_mapping(const struct vm_area_struct *vma,
3422 const struct vm_special_mapping *sm)
3423{
3424 return vma->vm_private_data == sm &&
3425 (vma->vm_ops == &special_mapping_vmops ||
3426 vma->vm_ops == &legacy_special_mapping_vmops);
3427}
3428
3429/*
3430 * Called with mm->mmap_lock held for writing.
3431 * Insert a new vma covering the given region, with the given flags.
3432 * Its pages are supplied by the given array of struct page *.
3433 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3434 * The region past the last page supplied will always produce SIGBUS.
3435 * The array pointer and the pages it points to are assumed to stay alive
3436 * for as long as this mapping might exist.
3437 */
3438struct vm_area_struct *_install_special_mapping(
3439 struct mm_struct *mm,
3440 unsigned long addr, unsigned long len,
3441 unsigned long vm_flags, const struct vm_special_mapping *spec)
3442{
3443 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3444 &special_mapping_vmops);
3445}
3446
3447int install_special_mapping(struct mm_struct *mm,
3448 unsigned long addr, unsigned long len,
3449 unsigned long vm_flags, struct page **pages)
3450{
3451 struct vm_area_struct *vma = __install_special_mapping(
3452 mm, addr, len, vm_flags, (void *)pages,
3453 &legacy_special_mapping_vmops);
3454
3455 return PTR_ERR_OR_ZERO(vma);
3456}
3457
3458static DEFINE_MUTEX(mm_all_locks_mutex);
3459
3460static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3461{
3462 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3463 /*
3464 * The LSB of head.next can't change from under us
3465 * because we hold the mm_all_locks_mutex.
3466 */
3467 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3468 /*
3469 * We can safely modify head.next after taking the
3470 * anon_vma->root->rwsem. If some other vma in this mm shares
3471 * the same anon_vma we won't take it again.
3472 *
3473 * No need of atomic instructions here, head.next
3474 * can't change from under us thanks to the
3475 * anon_vma->root->rwsem.
3476 */
3477 if (__test_and_set_bit(0, (unsigned long *)
3478 &anon_vma->root->rb_root.rb_root.rb_node))
3479 BUG();
3480 }
3481}
3482
3483static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3484{
3485 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3486 /*
3487 * AS_MM_ALL_LOCKS can't change from under us because
3488 * we hold the mm_all_locks_mutex.
3489 *
3490 * Operations on ->flags have to be atomic because
3491 * even if AS_MM_ALL_LOCKS is stable thanks to the
3492 * mm_all_locks_mutex, there may be other cpus
3493 * changing other bitflags in parallel to us.
3494 */
3495 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3496 BUG();
3497 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3498 }
3499}
3500
3501/*
3502 * This operation locks against the VM for all pte/vma/mm related
3503 * operations that could ever happen on a certain mm. This includes
3504 * vmtruncate, try_to_unmap, and all page faults.
3505 *
3506 * The caller must take the mmap_lock in write mode before calling
3507 * mm_take_all_locks(). The caller isn't allowed to release the
3508 * mmap_lock until mm_drop_all_locks() returns.
3509 *
3510 * mmap_lock in write mode is required in order to block all operations
3511 * that could modify pagetables and free pages without need of
3512 * altering the vma layout. It's also needed in write mode to avoid new
3513 * anon_vmas to be associated with existing vmas.
3514 *
3515 * A single task can't take more than one mm_take_all_locks() in a row
3516 * or it would deadlock.
3517 *
3518 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3519 * mapping->flags avoid to take the same lock twice, if more than one
3520 * vma in this mm is backed by the same anon_vma or address_space.
3521 *
3522 * We take locks in following order, accordingly to comment at beginning
3523 * of mm/rmap.c:
3524 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3525 * hugetlb mapping);
3526 * - all i_mmap_rwsem locks;
3527 * - all anon_vma->rwseml
3528 *
3529 * We can take all locks within these types randomly because the VM code
3530 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3531 * mm_all_locks_mutex.
3532 *
3533 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3534 * that may have to take thousand of locks.
3535 *
3536 * mm_take_all_locks() can fail if it's interrupted by signals.
3537 */
3538int mm_take_all_locks(struct mm_struct *mm)
3539{
3540 struct vm_area_struct *vma;
3541 struct anon_vma_chain *avc;
3542 MA_STATE(mas, &mm->mm_mt, 0, 0);
3543
3544 mmap_assert_write_locked(mm);
3545
3546 mutex_lock(&mm_all_locks_mutex);
3547
3548 mas_for_each(&mas, vma, ULONG_MAX) {
3549 if (signal_pending(current))
3550 goto out_unlock;
3551 if (vma->vm_file && vma->vm_file->f_mapping &&
3552 is_vm_hugetlb_page(vma))
3553 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3554 }
3555
3556 mas_set(&mas, 0);
3557 mas_for_each(&mas, vma, ULONG_MAX) {
3558 if (signal_pending(current))
3559 goto out_unlock;
3560 if (vma->vm_file && vma->vm_file->f_mapping &&
3561 !is_vm_hugetlb_page(vma))
3562 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3563 }
3564
3565 mas_set(&mas, 0);
3566 mas_for_each(&mas, vma, ULONG_MAX) {
3567 if (signal_pending(current))
3568 goto out_unlock;
3569 if (vma->anon_vma)
3570 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3571 vm_lock_anon_vma(mm, avc->anon_vma);
3572 }
3573
3574 return 0;
3575
3576out_unlock:
3577 mm_drop_all_locks(mm);
3578 return -EINTR;
3579}
3580
3581static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3582{
3583 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3584 /*
3585 * The LSB of head.next can't change to 0 from under
3586 * us because we hold the mm_all_locks_mutex.
3587 *
3588 * We must however clear the bitflag before unlocking
3589 * the vma so the users using the anon_vma->rb_root will
3590 * never see our bitflag.
3591 *
3592 * No need of atomic instructions here, head.next
3593 * can't change from under us until we release the
3594 * anon_vma->root->rwsem.
3595 */
3596 if (!__test_and_clear_bit(0, (unsigned long *)
3597 &anon_vma->root->rb_root.rb_root.rb_node))
3598 BUG();
3599 anon_vma_unlock_write(anon_vma);
3600 }
3601}
3602
3603static void vm_unlock_mapping(struct address_space *mapping)
3604{
3605 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3606 /*
3607 * AS_MM_ALL_LOCKS can't change to 0 from under us
3608 * because we hold the mm_all_locks_mutex.
3609 */
3610 i_mmap_unlock_write(mapping);
3611 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3612 &mapping->flags))
3613 BUG();
3614 }
3615}
3616
3617/*
3618 * The mmap_lock cannot be released by the caller until
3619 * mm_drop_all_locks() returns.
3620 */
3621void mm_drop_all_locks(struct mm_struct *mm)
3622{
3623 struct vm_area_struct *vma;
3624 struct anon_vma_chain *avc;
3625 MA_STATE(mas, &mm->mm_mt, 0, 0);
3626
3627 mmap_assert_write_locked(mm);
3628 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3629
3630 mas_for_each(&mas, vma, ULONG_MAX) {
3631 if (vma->anon_vma)
3632 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3633 vm_unlock_anon_vma(avc->anon_vma);
3634 if (vma->vm_file && vma->vm_file->f_mapping)
3635 vm_unlock_mapping(vma->vm_file->f_mapping);
3636 }
3637
3638 mutex_unlock(&mm_all_locks_mutex);
3639}
3640
3641/*
3642 * initialise the percpu counter for VM
3643 */
3644void __init mmap_init(void)
3645{
3646 int ret;
3647
3648 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3649 VM_BUG_ON(ret);
3650}
3651
3652/*
3653 * Initialise sysctl_user_reserve_kbytes.
3654 *
3655 * This is intended to prevent a user from starting a single memory hogging
3656 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3657 * mode.
3658 *
3659 * The default value is min(3% of free memory, 128MB)
3660 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3661 */
3662static int init_user_reserve(void)
3663{
3664 unsigned long free_kbytes;
3665
3666 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3667
3668 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3669 return 0;
3670}
3671subsys_initcall(init_user_reserve);
3672
3673/*
3674 * Initialise sysctl_admin_reserve_kbytes.
3675 *
3676 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3677 * to log in and kill a memory hogging process.
3678 *
3679 * Systems with more than 256MB will reserve 8MB, enough to recover
3680 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3681 * only reserve 3% of free pages by default.
3682 */
3683static int init_admin_reserve(void)
3684{
3685 unsigned long free_kbytes;
3686
3687 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3688
3689 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3690 return 0;
3691}
3692subsys_initcall(init_admin_reserve);
3693
3694/*
3695 * Reinititalise user and admin reserves if memory is added or removed.
3696 *
3697 * The default user reserve max is 128MB, and the default max for the
3698 * admin reserve is 8MB. These are usually, but not always, enough to
3699 * enable recovery from a memory hogging process using login/sshd, a shell,
3700 * and tools like top. It may make sense to increase or even disable the
3701 * reserve depending on the existence of swap or variations in the recovery
3702 * tools. So, the admin may have changed them.
3703 *
3704 * If memory is added and the reserves have been eliminated or increased above
3705 * the default max, then we'll trust the admin.
3706 *
3707 * If memory is removed and there isn't enough free memory, then we
3708 * need to reset the reserves.
3709 *
3710 * Otherwise keep the reserve set by the admin.
3711 */
3712static int reserve_mem_notifier(struct notifier_block *nb,
3713 unsigned long action, void *data)
3714{
3715 unsigned long tmp, free_kbytes;
3716
3717 switch (action) {
3718 case MEM_ONLINE:
3719 /* Default max is 128MB. Leave alone if modified by operator. */
3720 tmp = sysctl_user_reserve_kbytes;
3721 if (0 < tmp && tmp < (1UL << 17))
3722 init_user_reserve();
3723
3724 /* Default max is 8MB. Leave alone if modified by operator. */
3725 tmp = sysctl_admin_reserve_kbytes;
3726 if (0 < tmp && tmp < (1UL << 13))
3727 init_admin_reserve();
3728
3729 break;
3730 case MEM_OFFLINE:
3731 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3732
3733 if (sysctl_user_reserve_kbytes > free_kbytes) {
3734 init_user_reserve();
3735 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3736 sysctl_user_reserve_kbytes);
3737 }
3738
3739 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3740 init_admin_reserve();
3741 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3742 sysctl_admin_reserve_kbytes);
3743 }
3744 break;
3745 default:
3746 break;
3747 }
3748 return NOTIFY_OK;
3749}
3750
3751static int __meminit init_reserve_notifier(void)
3752{
3753 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3754 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3755
3756 return 0;
3757}
3758subsys_initcall(init_reserve_notifier);