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)) {
196 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 err = move_huge_pmd(vma, new_vma, old_addr,
206 anon_vma_unlock_write(vma->anon_vma);
212 split_huge_page_pmd(vma, old_addr, old_pmd);
214 VM_BUG_ON(pmd_trans_huge(*old_pmd));
216 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
219 next = (new_addr + PMD_SIZE) & PMD_MASK;
220 if (extent > next - new_addr)
221 extent = next - new_addr;
222 if (extent > LATENCY_LIMIT)
223 extent = LATENCY_LIMIT;
224 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
225 new_vma, new_pmd, new_addr, need_rmap_locks);
228 if (likely(need_flush))
229 flush_tlb_range(vma, old_end-len, old_addr);
231 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
233 return len + old_addr - old_end; /* how much done */
236 static unsigned long move_vma(struct vm_area_struct *vma,
237 unsigned long old_addr, unsigned long old_len,
238 unsigned long new_len, unsigned long new_addr, bool *locked)
240 struct mm_struct *mm = vma->vm_mm;
241 struct vm_area_struct *new_vma;
242 unsigned long vm_flags = vma->vm_flags;
243 unsigned long new_pgoff;
244 unsigned long moved_len;
245 unsigned long excess = 0;
246 unsigned long hiwater_vm;
249 bool need_rmap_locks;
252 * We'd prefer to avoid failure later on in do_munmap:
253 * which may split one vma into three before unmapping.
255 if (mm->map_count >= sysctl_max_map_count - 3)
259 * Advise KSM to break any KSM pages in the area to be moved:
260 * it would be confusing if they were to turn up at the new
261 * location, where they happen to coincide with different KSM
262 * pages recently unmapped. But leave vma->vm_flags as it was,
263 * so KSM can come around to merge on vma and new_vma afterwards.
265 err = ksm_madvise(vma, old_addr, old_addr + old_len,
266 MADV_UNMERGEABLE, &vm_flags);
270 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
271 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
276 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
278 if (moved_len < old_len) {
280 } else if (vma->vm_ops && vma->vm_ops->mremap) {
281 err = vma->vm_ops->mremap(new_vma);
286 * On error, move entries back from new area to old,
287 * which will succeed since page tables still there,
288 * and then proceed to unmap new area instead of old.
290 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
297 arch_remap(mm, old_addr, old_addr + old_len,
298 new_addr, new_addr + new_len);
301 /* Conceal VM_ACCOUNT so old reservation is not undone */
302 if (vm_flags & VM_ACCOUNT) {
303 vma->vm_flags &= ~VM_ACCOUNT;
304 excess = vma->vm_end - vma->vm_start - old_len;
305 if (old_addr > vma->vm_start &&
306 old_addr + old_len < vma->vm_end)
311 * If we failed to move page tables we still do total_vm increment
312 * since do_munmap() will decrement it by old_len == new_len.
314 * Since total_vm is about to be raised artificially high for a
315 * moment, we need to restore high watermark afterwards: if stats
316 * are taken meanwhile, total_vm and hiwater_vm appear too high.
317 * If this were a serious issue, we'd add a flag to do_munmap().
319 hiwater_vm = mm->hiwater_vm;
320 vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
322 if (do_munmap(mm, old_addr, old_len) < 0) {
323 /* OOM: unable to split vma, just get accounts right */
324 vm_unacct_memory(excess >> PAGE_SHIFT);
327 mm->hiwater_vm = hiwater_vm;
329 /* Restore VM_ACCOUNT if one or two pieces of vma left */
331 vma->vm_flags |= VM_ACCOUNT;
333 vma->vm_next->vm_flags |= VM_ACCOUNT;
336 if (vm_flags & VM_LOCKED) {
337 mm->locked_vm += new_len >> PAGE_SHIFT;
344 static struct vm_area_struct *vma_to_resize(unsigned long addr,
345 unsigned long old_len, unsigned long new_len, unsigned long *p)
347 struct mm_struct *mm = current->mm;
348 struct vm_area_struct *vma = find_vma(mm, addr);
351 if (!vma || vma->vm_start > addr)
352 return ERR_PTR(-EFAULT);
354 if (is_vm_hugetlb_page(vma))
355 return ERR_PTR(-EINVAL);
357 /* We can't remap across vm area boundaries */
358 if (old_len > vma->vm_end - addr)
359 return ERR_PTR(-EFAULT);
361 if (new_len == old_len)
364 /* Need to be careful about a growing mapping */
365 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
366 pgoff += vma->vm_pgoff;
367 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
368 return ERR_PTR(-EINVAL);
370 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
371 return ERR_PTR(-EFAULT);
373 if (vma->vm_flags & VM_LOCKED) {
374 unsigned long locked, lock_limit;
375 locked = mm->locked_vm << PAGE_SHIFT;
376 lock_limit = rlimit(RLIMIT_MEMLOCK);
377 locked += new_len - old_len;
378 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
379 return ERR_PTR(-EAGAIN);
382 if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
383 return ERR_PTR(-ENOMEM);
385 if (vma->vm_flags & VM_ACCOUNT) {
386 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
387 if (security_vm_enough_memory_mm(mm, charged))
388 return ERR_PTR(-ENOMEM);
395 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
396 unsigned long new_addr, unsigned long new_len, bool *locked)
398 struct mm_struct *mm = current->mm;
399 struct vm_area_struct *vma;
400 unsigned long ret = -EINVAL;
401 unsigned long charged = 0;
402 unsigned long map_flags;
404 if (offset_in_page(new_addr))
407 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
410 /* Ensure the old/new locations do not overlap */
411 if (addr + old_len > new_addr && new_addr + new_len > addr)
414 ret = do_munmap(mm, new_addr, new_len);
418 if (old_len >= new_len) {
419 ret = do_munmap(mm, addr+new_len, old_len - new_len);
420 if (ret && old_len != new_len)
425 vma = vma_to_resize(addr, old_len, new_len, &charged);
431 map_flags = MAP_FIXED;
432 if (vma->vm_flags & VM_MAYSHARE)
433 map_flags |= MAP_SHARED;
435 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
436 ((addr - vma->vm_start) >> PAGE_SHIFT),
438 if (offset_in_page(ret))
441 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
442 if (!(offset_in_page(ret)))
445 vm_unacct_memory(charged);
451 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
453 unsigned long end = vma->vm_end + delta;
454 if (end < vma->vm_end) /* overflow */
456 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
458 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
459 0, MAP_FIXED) & ~PAGE_MASK)
465 * Expand (or shrink) an existing mapping, potentially moving it at the
466 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
468 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
469 * This option implies MREMAP_MAYMOVE.
471 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
472 unsigned long, new_len, unsigned long, flags,
473 unsigned long, new_addr)
475 struct mm_struct *mm = current->mm;
476 struct vm_area_struct *vma;
477 unsigned long ret = -EINVAL;
478 unsigned long charged = 0;
481 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
484 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
487 if (offset_in_page(addr))
490 old_len = PAGE_ALIGN(old_len);
491 new_len = PAGE_ALIGN(new_len);
494 * We allow a zero old-len as a special case
495 * for DOS-emu "duplicate shm area" thing. But
496 * a zero new-len is nonsensical.
501 down_write(¤t->mm->mmap_sem);
503 if (flags & MREMAP_FIXED) {
504 ret = mremap_to(addr, old_len, new_addr, new_len,
510 * Always allow a shrinking remap: that just unmaps
511 * the unnecessary pages..
512 * do_munmap does all the needed commit accounting
514 if (old_len >= new_len) {
515 ret = do_munmap(mm, addr+new_len, old_len - new_len);
516 if (ret && old_len != new_len)
523 * Ok, we need to grow..
525 vma = vma_to_resize(addr, old_len, new_len, &charged);
531 /* old_len exactly to the end of the area..
533 if (old_len == vma->vm_end - addr) {
534 /* can we just expand the current mapping? */
535 if (vma_expandable(vma, new_len - old_len)) {
536 int pages = (new_len - old_len) >> PAGE_SHIFT;
538 if (vma_adjust(vma, vma->vm_start, addr + new_len,
539 vma->vm_pgoff, NULL)) {
544 vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
545 if (vma->vm_flags & VM_LOCKED) {
546 mm->locked_vm += pages;
556 * We weren't able to just expand or shrink the area,
557 * we need to create a new one and move it..
560 if (flags & MREMAP_MAYMOVE) {
561 unsigned long map_flags = 0;
562 if (vma->vm_flags & VM_MAYSHARE)
563 map_flags |= MAP_SHARED;
565 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
567 ((addr - vma->vm_start) >> PAGE_SHIFT),
569 if (offset_in_page(new_addr)) {
574 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
577 if (offset_in_page(ret)) {
578 vm_unacct_memory(charged);
581 up_write(¤t->mm->mmap_sem);
582 if (locked && new_len > old_len)
583 mm_populate(new_addr + old_len, new_len - old_len);