4 * (C) Copyright 1996 Linus Torvalds
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
11 #include <linux/hugetlb.h>
12 #include <linux/shm.h>
13 #include <linux/ksm.h>
14 #include <linux/mman.h>
15 #include <linux/swap.h>
16 #include <linux/capability.h>
18 #include <linux/swapops.h>
19 #include <linux/highmem.h>
20 #include <linux/security.h>
21 #include <linux/syscalls.h>
22 #include <linux/mmu_notifier.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/uaccess.h>
25 #include <linux/mm-arch-hooks.h>
27 #include <asm/cacheflush.h>
28 #include <asm/tlbflush.h>
32 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
38 pgd = pgd_offset(mm, addr);
39 if (pgd_none_or_clear_bad(pgd))
42 pud = pud_offset(pgd, addr);
43 if (pud_none_or_clear_bad(pud))
46 pmd = pmd_offset(pud, addr);
53 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
60 pgd = pgd_offset(mm, addr);
61 pud = pud_alloc(mm, pgd, addr);
65 pmd = pmd_alloc(mm, pud, addr);
69 VM_BUG_ON(pmd_trans_huge(*pmd));
74 static pte_t move_soft_dirty_pte(pte_t pte)
77 * Set soft dirty bit so we can notice
78 * in userspace the ptes were moved.
80 #ifdef CONFIG_MEM_SOFT_DIRTY
82 pte = pte_mksoft_dirty(pte);
83 else if (is_swap_pte(pte))
84 pte = pte_swp_mksoft_dirty(pte);
89 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
90 unsigned long old_addr, unsigned long old_end,
91 struct vm_area_struct *new_vma, pmd_t *new_pmd,
92 unsigned long new_addr, bool need_rmap_locks)
94 struct address_space *mapping = NULL;
95 struct anon_vma *anon_vma = NULL;
96 struct mm_struct *mm = vma->vm_mm;
97 pte_t *old_pte, *new_pte, pte;
98 spinlock_t *old_ptl, *new_ptl;
101 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
102 * locks to ensure that rmap will always observe either the old or the
103 * new ptes. This is the easiest way to avoid races with
104 * truncate_pagecache(), page migration, etc...
106 * When need_rmap_locks is false, we use other ways to avoid
109 * - During exec() shift_arg_pages(), we use a specially tagged vma
110 * which rmap call sites look for using is_vma_temporary_stack().
112 * - During mremap(), new_vma is often known to be placed after vma
113 * in rmap traversal order. This ensures rmap will always observe
114 * either the old pte, or the new pte, or both (the page table locks
115 * serialize access to individual ptes, but only rmap traversal
116 * order guarantees that we won't miss both the old and new ptes).
118 if (need_rmap_locks) {
120 mapping = vma->vm_file->f_mapping;
121 i_mmap_lock_write(mapping);
124 anon_vma = vma->anon_vma;
125 anon_vma_lock_write(anon_vma);
130 * We don't have to worry about the ordering of src and dst
131 * pte locks because exclusive mmap_sem prevents deadlock.
133 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
134 new_pte = pte_offset_map(new_pmd, new_addr);
135 new_ptl = pte_lockptr(mm, new_pmd);
136 if (new_ptl != old_ptl)
137 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
138 arch_enter_lazy_mmu_mode();
140 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
141 new_pte++, new_addr += PAGE_SIZE) {
142 if (pte_none(*old_pte))
144 pte = ptep_get_and_clear(mm, old_addr, old_pte);
145 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
146 pte = move_soft_dirty_pte(pte);
147 set_pte_at(mm, new_addr, new_pte, pte);
150 arch_leave_lazy_mmu_mode();
151 if (new_ptl != old_ptl)
152 spin_unlock(new_ptl);
153 pte_unmap(new_pte - 1);
154 pte_unmap_unlock(old_pte - 1, old_ptl);
156 anon_vma_unlock_write(anon_vma);
158 i_mmap_unlock_write(mapping);
161 #define LATENCY_LIMIT (64 * PAGE_SIZE)
163 unsigned long move_page_tables(struct vm_area_struct *vma,
164 unsigned long old_addr, struct vm_area_struct *new_vma,
165 unsigned long new_addr, unsigned long len,
166 bool need_rmap_locks)
168 unsigned long extent, next, old_end;
169 pmd_t *old_pmd, *new_pmd;
170 bool need_flush = false;
171 unsigned long mmun_start; /* For mmu_notifiers */
172 unsigned long mmun_end; /* For mmu_notifiers */
174 old_end = old_addr + len;
175 flush_cache_range(vma, old_addr, old_end);
177 mmun_start = old_addr;
179 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
181 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
183 next = (old_addr + PMD_SIZE) & PMD_MASK;
184 /* even if next overflowed, extent below will be ok */
185 extent = next - old_addr;
186 if (extent > old_end - old_addr)
187 extent = old_end - old_addr;
188 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
191 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
194 if (pmd_trans_huge(*old_pmd)) {
195 if (extent == HPAGE_PMD_SIZE) {
197 VM_BUG_ON_VMA(vma->vm_file || !vma->anon_vma,
199 /* See comment in move_ptes() */
201 anon_vma_lock_write(vma->anon_vma);
202 moved = move_huge_pmd(vma, new_vma, old_addr,
206 anon_vma_unlock_write(vma->anon_vma);
212 split_huge_pmd(vma, old_pmd, old_addr);
213 VM_BUG_ON(pmd_trans_huge(*old_pmd));
215 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
218 next = (new_addr + PMD_SIZE) & PMD_MASK;
219 if (extent > next - new_addr)
220 extent = next - new_addr;
221 if (extent > LATENCY_LIMIT)
222 extent = LATENCY_LIMIT;
223 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
224 new_vma, new_pmd, new_addr, need_rmap_locks);
227 if (likely(need_flush))
228 flush_tlb_range(vma, old_end-len, old_addr);
230 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
232 return len + old_addr - old_end; /* how much done */
235 static unsigned long move_vma(struct vm_area_struct *vma,
236 unsigned long old_addr, unsigned long old_len,
237 unsigned long new_len, unsigned long new_addr, bool *locked)
239 struct mm_struct *mm = vma->vm_mm;
240 struct vm_area_struct *new_vma;
241 unsigned long vm_flags = vma->vm_flags;
242 unsigned long new_pgoff;
243 unsigned long moved_len;
244 unsigned long excess = 0;
245 unsigned long hiwater_vm;
248 bool need_rmap_locks;
251 * We'd prefer to avoid failure later on in do_munmap:
252 * which may split one vma into three before unmapping.
254 if (mm->map_count >= sysctl_max_map_count - 3)
258 * Advise KSM to break any KSM pages in the area to be moved:
259 * it would be confusing if they were to turn up at the new
260 * location, where they happen to coincide with different KSM
261 * pages recently unmapped. But leave vma->vm_flags as it was,
262 * so KSM can come around to merge on vma and new_vma afterwards.
264 err = ksm_madvise(vma, old_addr, old_addr + old_len,
265 MADV_UNMERGEABLE, &vm_flags);
269 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
270 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
275 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
277 if (moved_len < old_len) {
279 } else if (vma->vm_ops && vma->vm_ops->mremap) {
280 err = vma->vm_ops->mremap(new_vma);
285 * On error, move entries back from new area to old,
286 * which will succeed since page tables still there,
287 * and then proceed to unmap new area instead of old.
289 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
296 arch_remap(mm, old_addr, old_addr + old_len,
297 new_addr, new_addr + new_len);
300 /* Conceal VM_ACCOUNT so old reservation is not undone */
301 if (vm_flags & VM_ACCOUNT) {
302 vma->vm_flags &= ~VM_ACCOUNT;
303 excess = vma->vm_end - vma->vm_start - old_len;
304 if (old_addr > vma->vm_start &&
305 old_addr + old_len < vma->vm_end)
310 * If we failed to move page tables we still do total_vm increment
311 * since do_munmap() will decrement it by old_len == new_len.
313 * Since total_vm is about to be raised artificially high for a
314 * moment, we need to restore high watermark afterwards: if stats
315 * are taken meanwhile, total_vm and hiwater_vm appear too high.
316 * If this were a serious issue, we'd add a flag to do_munmap().
318 hiwater_vm = mm->hiwater_vm;
319 vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
321 if (do_munmap(mm, old_addr, old_len) < 0) {
322 /* OOM: unable to split vma, just get accounts right */
323 vm_unacct_memory(excess >> PAGE_SHIFT);
326 mm->hiwater_vm = hiwater_vm;
328 /* Restore VM_ACCOUNT if one or two pieces of vma left */
330 vma->vm_flags |= VM_ACCOUNT;
332 vma->vm_next->vm_flags |= VM_ACCOUNT;
335 if (vm_flags & VM_LOCKED) {
336 mm->locked_vm += new_len >> PAGE_SHIFT;
343 static struct vm_area_struct *vma_to_resize(unsigned long addr,
344 unsigned long old_len, unsigned long new_len, unsigned long *p)
346 struct mm_struct *mm = current->mm;
347 struct vm_area_struct *vma = find_vma(mm, addr);
350 if (!vma || vma->vm_start > addr)
351 return ERR_PTR(-EFAULT);
353 if (is_vm_hugetlb_page(vma))
354 return ERR_PTR(-EINVAL);
356 /* We can't remap across vm area boundaries */
357 if (old_len > vma->vm_end - addr)
358 return ERR_PTR(-EFAULT);
360 if (new_len == old_len)
363 /* Need to be careful about a growing mapping */
364 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
365 pgoff += vma->vm_pgoff;
366 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
367 return ERR_PTR(-EINVAL);
369 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
370 return ERR_PTR(-EFAULT);
372 if (vma->vm_flags & VM_LOCKED) {
373 unsigned long locked, lock_limit;
374 locked = mm->locked_vm << PAGE_SHIFT;
375 lock_limit = rlimit(RLIMIT_MEMLOCK);
376 locked += new_len - old_len;
377 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
378 return ERR_PTR(-EAGAIN);
381 if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
382 return ERR_PTR(-ENOMEM);
384 if (vma->vm_flags & VM_ACCOUNT) {
385 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
386 if (security_vm_enough_memory_mm(mm, charged))
387 return ERR_PTR(-ENOMEM);
394 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
395 unsigned long new_addr, unsigned long new_len, bool *locked)
397 struct mm_struct *mm = current->mm;
398 struct vm_area_struct *vma;
399 unsigned long ret = -EINVAL;
400 unsigned long charged = 0;
401 unsigned long map_flags;
403 if (offset_in_page(new_addr))
406 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
409 /* Ensure the old/new locations do not overlap */
410 if (addr + old_len > new_addr && new_addr + new_len > addr)
413 ret = do_munmap(mm, new_addr, new_len);
417 if (old_len >= new_len) {
418 ret = do_munmap(mm, addr+new_len, old_len - new_len);
419 if (ret && old_len != new_len)
424 vma = vma_to_resize(addr, old_len, new_len, &charged);
430 map_flags = MAP_FIXED;
431 if (vma->vm_flags & VM_MAYSHARE)
432 map_flags |= MAP_SHARED;
434 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
435 ((addr - vma->vm_start) >> PAGE_SHIFT),
437 if (offset_in_page(ret))
440 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
441 if (!(offset_in_page(ret)))
444 vm_unacct_memory(charged);
450 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
452 unsigned long end = vma->vm_end + delta;
453 if (end < vma->vm_end) /* overflow */
455 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
457 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
458 0, MAP_FIXED) & ~PAGE_MASK)
464 * Expand (or shrink) an existing mapping, potentially moving it at the
465 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
467 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
468 * This option implies MREMAP_MAYMOVE.
470 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
471 unsigned long, new_len, unsigned long, flags,
472 unsigned long, new_addr)
474 struct mm_struct *mm = current->mm;
475 struct vm_area_struct *vma;
476 unsigned long ret = -EINVAL;
477 unsigned long charged = 0;
480 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
483 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
486 if (offset_in_page(addr))
489 old_len = PAGE_ALIGN(old_len);
490 new_len = PAGE_ALIGN(new_len);
493 * We allow a zero old-len as a special case
494 * for DOS-emu "duplicate shm area" thing. But
495 * a zero new-len is nonsensical.
500 down_write(¤t->mm->mmap_sem);
502 if (flags & MREMAP_FIXED) {
503 ret = mremap_to(addr, old_len, new_addr, new_len,
509 * Always allow a shrinking remap: that just unmaps
510 * the unnecessary pages..
511 * do_munmap does all the needed commit accounting
513 if (old_len >= new_len) {
514 ret = do_munmap(mm, addr+new_len, old_len - new_len);
515 if (ret && old_len != new_len)
522 * Ok, we need to grow..
524 vma = vma_to_resize(addr, old_len, new_len, &charged);
530 /* old_len exactly to the end of the area..
532 if (old_len == vma->vm_end - addr) {
533 /* can we just expand the current mapping? */
534 if (vma_expandable(vma, new_len - old_len)) {
535 int pages = (new_len - old_len) >> PAGE_SHIFT;
537 if (vma_adjust(vma, vma->vm_start, addr + new_len,
538 vma->vm_pgoff, NULL)) {
543 vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
544 if (vma->vm_flags & VM_LOCKED) {
545 mm->locked_vm += pages;
555 * We weren't able to just expand or shrink the area,
556 * we need to create a new one and move it..
559 if (flags & MREMAP_MAYMOVE) {
560 unsigned long map_flags = 0;
561 if (vma->vm_flags & VM_MAYSHARE)
562 map_flags |= MAP_SHARED;
564 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
566 ((addr - vma->vm_start) >> PAGE_SHIFT),
568 if (offset_in_page(new_addr)) {
573 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
576 if (offset_in_page(ret)) {
577 vm_unacct_memory(charged);
580 up_write(¤t->mm->mmap_sem);
581 if (locked && new_len > old_len)
582 mm_populate(new_addr + old_len, new_len - old_len);