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>
25 #include <asm/uaccess.h>
26 #include <asm/cacheflush.h>
27 #include <asm/tlbflush.h>
31 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
37 pgd = pgd_offset(mm, addr);
38 if (pgd_none_or_clear_bad(pgd))
41 pud = pud_offset(pgd, addr);
42 if (pud_none_or_clear_bad(pud))
45 pmd = pmd_offset(pud, addr);
52 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
59 pgd = pgd_offset(mm, addr);
60 pud = pud_alloc(mm, pgd, addr);
64 pmd = pmd_alloc(mm, pud, addr);
68 VM_BUG_ON(pmd_trans_huge(*pmd));
73 static pte_t move_soft_dirty_pte(pte_t pte)
76 * Set soft dirty bit so we can notice
77 * in userspace the ptes were moved.
79 #ifdef CONFIG_MEM_SOFT_DIRTY
81 pte = pte_mksoft_dirty(pte);
82 else if (is_swap_pte(pte))
83 pte = pte_swp_mksoft_dirty(pte);
84 else if (pte_file(pte))
85 pte = pte_file_mksoft_dirty(pte);
90 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
91 unsigned long old_addr, unsigned long old_end,
92 struct vm_area_struct *new_vma, pmd_t *new_pmd,
93 unsigned long new_addr, bool need_rmap_locks)
95 struct address_space *mapping = NULL;
96 struct anon_vma *anon_vma = NULL;
97 struct mm_struct *mm = vma->vm_mm;
98 pte_t *old_pte, *new_pte, pte;
99 spinlock_t *old_ptl, *new_ptl;
102 * When need_rmap_locks is true, we take the i_mmap_mutex and anon_vma
103 * locks to ensure that rmap will always observe either the old or the
104 * new ptes. This is the easiest way to avoid races with
105 * truncate_pagecache(), page migration, etc...
107 * When need_rmap_locks is false, we use other ways to avoid
110 * - During exec() shift_arg_pages(), we use a specially tagged vma
111 * which rmap call sites look for using is_vma_temporary_stack().
113 * - During mremap(), new_vma is often known to be placed after vma
114 * in rmap traversal order. This ensures rmap will always observe
115 * either the old pte, or the new pte, or both (the page table locks
116 * serialize access to individual ptes, but only rmap traversal
117 * order guarantees that we won't miss both the old and new ptes).
119 if (need_rmap_locks) {
121 mapping = vma->vm_file->f_mapping;
122 mutex_lock(&mapping->i_mmap_mutex);
125 anon_vma = vma->anon_vma;
126 anon_vma_lock_write(anon_vma);
131 * We don't have to worry about the ordering of src and dst
132 * pte locks because exclusive mmap_sem prevents deadlock.
134 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
135 new_pte = pte_offset_map(new_pmd, new_addr);
136 new_ptl = pte_lockptr(mm, new_pmd);
137 if (new_ptl != old_ptl)
138 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
139 arch_enter_lazy_mmu_mode();
141 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
142 new_pte++, new_addr += PAGE_SIZE) {
143 if (pte_none(*old_pte))
145 pte = ptep_get_and_clear(mm, old_addr, old_pte);
146 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
147 pte = move_soft_dirty_pte(pte);
148 set_pte_at(mm, new_addr, new_pte, pte);
151 arch_leave_lazy_mmu_mode();
152 if (new_ptl != old_ptl)
153 spin_unlock(new_ptl);
154 pte_unmap(new_pte - 1);
155 pte_unmap_unlock(old_pte - 1, old_ptl);
157 anon_vma_unlock_write(anon_vma);
159 mutex_unlock(&mapping->i_mmap_mutex);
162 #define LATENCY_LIMIT (64 * PAGE_SIZE)
164 unsigned long move_page_tables(struct vm_area_struct *vma,
165 unsigned long old_addr, struct vm_area_struct *new_vma,
166 unsigned long new_addr, unsigned long len,
167 bool need_rmap_locks)
169 unsigned long extent, next, old_end;
170 pmd_t *old_pmd, *new_pmd;
171 bool need_flush = false;
172 unsigned long mmun_start; /* For mmu_notifiers */
173 unsigned long mmun_end; /* For mmu_notifiers */
175 old_end = old_addr + len;
176 flush_cache_range(vma, old_addr, old_end);
178 mmun_start = old_addr;
180 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
182 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
184 next = (old_addr + PMD_SIZE) & PMD_MASK;
185 /* even if next overflowed, extent below will be ok */
186 extent = next - old_addr;
187 if (extent > old_end - old_addr)
188 extent = old_end - old_addr;
189 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
192 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
195 if (pmd_trans_huge(*old_pmd)) {
197 if (extent == HPAGE_PMD_SIZE)
198 err = move_huge_pmd(vma, new_vma, old_addr,
205 split_huge_page_pmd(vma, old_addr, old_pmd);
207 VM_BUG_ON(pmd_trans_huge(*old_pmd));
209 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
212 next = (new_addr + PMD_SIZE) & PMD_MASK;
213 if (extent > next - new_addr)
214 extent = next - new_addr;
215 if (extent > LATENCY_LIMIT)
216 extent = LATENCY_LIMIT;
217 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
218 new_vma, new_pmd, new_addr, need_rmap_locks);
221 if (likely(need_flush))
222 flush_tlb_range(vma, old_end-len, old_addr);
224 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
226 return len + old_addr - old_end; /* how much done */
229 static unsigned long move_vma(struct vm_area_struct *vma,
230 unsigned long old_addr, unsigned long old_len,
231 unsigned long new_len, unsigned long new_addr, bool *locked)
233 struct mm_struct *mm = vma->vm_mm;
234 struct vm_area_struct *new_vma;
235 unsigned long vm_flags = vma->vm_flags;
236 unsigned long new_pgoff;
237 unsigned long moved_len;
238 unsigned long excess = 0;
239 unsigned long hiwater_vm;
242 bool need_rmap_locks;
245 * We'd prefer to avoid failure later on in do_munmap:
246 * which may split one vma into three before unmapping.
248 if (mm->map_count >= sysctl_max_map_count - 3)
252 * Advise KSM to break any KSM pages in the area to be moved:
253 * it would be confusing if they were to turn up at the new
254 * location, where they happen to coincide with different KSM
255 * pages recently unmapped. But leave vma->vm_flags as it was,
256 * so KSM can come around to merge on vma and new_vma afterwards.
258 err = ksm_madvise(vma, old_addr, old_addr + old_len,
259 MADV_UNMERGEABLE, &vm_flags);
263 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
264 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
269 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
271 if (moved_len < old_len) {
273 * On error, move entries back from new area to old,
274 * which will succeed since page tables still there,
275 * and then proceed to unmap new area instead of old.
277 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
285 /* Conceal VM_ACCOUNT so old reservation is not undone */
286 if (vm_flags & VM_ACCOUNT) {
287 vma->vm_flags &= ~VM_ACCOUNT;
288 excess = vma->vm_end - vma->vm_start - old_len;
289 if (old_addr > vma->vm_start &&
290 old_addr + old_len < vma->vm_end)
295 * If we failed to move page tables we still do total_vm increment
296 * since do_munmap() will decrement it by old_len == new_len.
298 * Since total_vm is about to be raised artificially high for a
299 * moment, we need to restore high watermark afterwards: if stats
300 * are taken meanwhile, total_vm and hiwater_vm appear too high.
301 * If this were a serious issue, we'd add a flag to do_munmap().
303 hiwater_vm = mm->hiwater_vm;
304 vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
306 if (do_munmap(mm, old_addr, old_len) < 0) {
307 /* OOM: unable to split vma, just get accounts right */
308 vm_unacct_memory(excess >> PAGE_SHIFT);
311 mm->hiwater_vm = hiwater_vm;
313 /* Restore VM_ACCOUNT if one or two pieces of vma left */
315 vma->vm_flags |= VM_ACCOUNT;
317 vma->vm_next->vm_flags |= VM_ACCOUNT;
320 if (vm_flags & VM_LOCKED) {
321 mm->locked_vm += new_len >> PAGE_SHIFT;
328 static struct vm_area_struct *vma_to_resize(unsigned long addr,
329 unsigned long old_len, unsigned long new_len, unsigned long *p)
331 struct mm_struct *mm = current->mm;
332 struct vm_area_struct *vma = find_vma(mm, addr);
334 if (!vma || vma->vm_start > addr)
337 if (is_vm_hugetlb_page(vma))
340 /* We can't remap across vm area boundaries */
341 if (old_len > vma->vm_end - addr)
344 /* Need to be careful about a growing mapping */
345 if (new_len > old_len) {
348 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
350 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
351 pgoff += vma->vm_pgoff;
352 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
356 if (vma->vm_flags & VM_LOCKED) {
357 unsigned long locked, lock_limit;
358 locked = mm->locked_vm << PAGE_SHIFT;
359 lock_limit = rlimit(RLIMIT_MEMLOCK);
360 locked += new_len - old_len;
361 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
365 if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
368 if (vma->vm_flags & VM_ACCOUNT) {
369 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
370 if (security_vm_enough_memory_mm(mm, charged))
377 Efault: /* very odd choice for most of the cases, but... */
378 return ERR_PTR(-EFAULT);
380 return ERR_PTR(-EINVAL);
382 return ERR_PTR(-ENOMEM);
384 return ERR_PTR(-EAGAIN);
387 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
388 unsigned long new_addr, unsigned long new_len, bool *locked)
390 struct mm_struct *mm = current->mm;
391 struct vm_area_struct *vma;
392 unsigned long ret = -EINVAL;
393 unsigned long charged = 0;
394 unsigned long map_flags;
396 if (new_addr & ~PAGE_MASK)
399 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
402 /* Check if the location we're moving into overlaps the
403 * old location at all, and fail if it does.
405 if ((new_addr <= addr) && (new_addr+new_len) > addr)
408 if ((addr <= new_addr) && (addr+old_len) > new_addr)
411 ret = do_munmap(mm, new_addr, new_len);
415 if (old_len >= new_len) {
416 ret = do_munmap(mm, addr+new_len, old_len - new_len);
417 if (ret && old_len != new_len)
422 vma = vma_to_resize(addr, old_len, new_len, &charged);
428 map_flags = MAP_FIXED;
429 if (vma->vm_flags & VM_MAYSHARE)
430 map_flags |= MAP_SHARED;
432 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
433 ((addr - vma->vm_start) >> PAGE_SHIFT),
435 if (ret & ~PAGE_MASK)
438 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
439 if (!(ret & ~PAGE_MASK))
442 vm_unacct_memory(charged);
448 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
450 unsigned long end = vma->vm_end + delta;
451 if (end < vma->vm_end) /* overflow */
453 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
455 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
456 0, MAP_FIXED) & ~PAGE_MASK)
462 * Expand (or shrink) an existing mapping, potentially moving it at the
463 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
465 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
466 * This option implies MREMAP_MAYMOVE.
468 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
469 unsigned long, new_len, unsigned long, flags,
470 unsigned long, new_addr)
472 struct mm_struct *mm = current->mm;
473 struct vm_area_struct *vma;
474 unsigned long ret = -EINVAL;
475 unsigned long charged = 0;
478 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
481 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
484 if (addr & ~PAGE_MASK)
487 old_len = PAGE_ALIGN(old_len);
488 new_len = PAGE_ALIGN(new_len);
491 * We allow a zero old-len as a special case
492 * for DOS-emu "duplicate shm area" thing. But
493 * a zero new-len is nonsensical.
498 down_write(¤t->mm->mmap_sem);
500 if (flags & MREMAP_FIXED) {
501 ret = mremap_to(addr, old_len, new_addr, new_len,
507 * Always allow a shrinking remap: that just unmaps
508 * the unnecessary pages..
509 * do_munmap does all the needed commit accounting
511 if (old_len >= new_len) {
512 ret = do_munmap(mm, addr+new_len, old_len - new_len);
513 if (ret && old_len != new_len)
520 * Ok, we need to grow..
522 vma = vma_to_resize(addr, old_len, new_len, &charged);
528 /* old_len exactly to the end of the area..
530 if (old_len == vma->vm_end - addr) {
531 /* can we just expand the current mapping? */
532 if (vma_expandable(vma, new_len - old_len)) {
533 int pages = (new_len - old_len) >> PAGE_SHIFT;
535 if (vma_adjust(vma, vma->vm_start, addr + new_len,
536 vma->vm_pgoff, NULL)) {
541 vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
542 if (vma->vm_flags & VM_LOCKED) {
543 mm->locked_vm += pages;
553 * We weren't able to just expand or shrink the area,
554 * we need to create a new one and move it..
557 if (flags & MREMAP_MAYMOVE) {
558 unsigned long map_flags = 0;
559 if (vma->vm_flags & VM_MAYSHARE)
560 map_flags |= MAP_SHARED;
562 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
564 ((addr - vma->vm_start) >> PAGE_SHIFT),
566 if (new_addr & ~PAGE_MASK) {
571 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
574 if (ret & ~PAGE_MASK)
575 vm_unacct_memory(charged);
576 up_write(¤t->mm->mmap_sem);
577 if (locked && new_len > old_len)
578 mm_populate(new_addr + old_len, new_len - old_len);