4 * (C) Copyright 1996 Linus Torvalds
6 * Address space accounting code <alan@redhat.com>
7 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
11 #include <linux/hugetlb.h>
12 #include <linux/slab.h>
13 #include <linux/shm.h>
14 #include <linux/mman.h>
15 #include <linux/swap.h>
17 #include <linux/highmem.h>
18 #include <linux/security.h>
20 #include <asm/uaccess.h>
21 #include <asm/pgalloc.h>
22 #include <asm/cacheflush.h>
23 #include <asm/tlbflush.h>
25 static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
31 pgd = pgd_offset(mm, addr);
40 pmd = pmd_offset(pgd, addr);
49 pte = pte_offset_map_nested(pmd, addr);
51 pte_unmap_nested(pte);
58 static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
63 pgd = pgd_offset(mm, addr);
66 pmd = pmd_offset(pgd, addr);
67 if (!pmd_present(*pmd))
69 return pte_offset_map(pmd, addr);
72 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
77 pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
79 pte = pte_alloc_map(mm, pmd, addr);
84 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
85 unsigned long new_addr)
87 struct address_space *mapping = NULL;
88 struct mm_struct *mm = vma->vm_mm;
94 * Subtle point from Rajesh Venkatasubramanian: before
95 * moving file-based ptes, we must lock vmtruncate out,
96 * since it might clean the dst vma before the src vma,
97 * and we propagate stale pages into the dst afterward.
99 mapping = vma->vm_file->f_mapping;
100 spin_lock(&mapping->i_mmap_lock);
102 spin_lock(&mm->page_table_lock);
104 src = get_one_pte_map_nested(mm, old_addr);
107 * Look to see whether alloc_one_pte_map needs to perform a
108 * memory allocation. If it does then we need to drop the
111 dst = get_one_pte_map(mm, new_addr);
112 if (unlikely(!dst)) {
113 pte_unmap_nested(src);
115 spin_unlock(&mapping->i_mmap_lock);
116 dst = alloc_one_pte_map(mm, new_addr);
117 if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
118 spin_unlock(&mm->page_table_lock);
119 spin_lock(&mapping->i_mmap_lock);
120 spin_lock(&mm->page_table_lock);
122 src = get_one_pte_map_nested(mm, old_addr);
125 * Since alloc_one_pte_map can drop and re-acquire
126 * page_table_lock, we should re-check the src entry...
131 pte = ptep_clear_flush(vma, old_addr, src);
135 pte_unmap_nested(src);
140 spin_unlock(&mm->page_table_lock);
142 spin_unlock(&mapping->i_mmap_lock);
146 static unsigned long move_page_tables(struct vm_area_struct *vma,
147 unsigned long new_addr, unsigned long old_addr,
150 unsigned long offset;
152 flush_cache_range(vma, old_addr, old_addr + len);
155 * This is not the clever way to do this, but we're taking the
156 * easy way out on the assumption that most remappings will be
157 * only a few pages.. This also makes error recovery easier.
159 for (offset = 0; offset < len; offset += PAGE_SIZE) {
160 if (move_one_page(vma, old_addr+offset, new_addr+offset) < 0)
167 static unsigned long move_vma(struct vm_area_struct *vma,
168 unsigned long old_addr, unsigned long old_len,
169 unsigned long new_len, unsigned long new_addr)
171 struct mm_struct *mm = vma->vm_mm;
172 struct vm_area_struct *new_vma;
173 unsigned long vm_flags = vma->vm_flags;
174 unsigned long new_pgoff;
175 unsigned long moved_len;
176 unsigned long excess = 0;
180 * We'd prefer to avoid failure later on in do_munmap:
181 * which may split one vma into three before unmapping.
183 if (mm->map_count >= sysctl_max_map_count - 3)
186 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
187 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
191 moved_len = move_page_tables(vma, new_addr, old_addr, old_len);
192 if (moved_len < old_len) {
194 * On error, move entries back from new area to old,
195 * which will succeed since page tables still there,
196 * and then proceed to unmap new area instead of old.
198 move_page_tables(new_vma, old_addr, new_addr, moved_len);
205 /* Conceal VM_ACCOUNT so old reservation is not undone */
206 if (vm_flags & VM_ACCOUNT) {
207 vma->vm_flags &= ~VM_ACCOUNT;
208 excess = vma->vm_end - vma->vm_start - old_len;
209 if (old_addr > vma->vm_start &&
210 old_addr + old_len < vma->vm_end)
214 if (do_munmap(mm, old_addr, old_len) < 0) {
215 /* OOM: unable to split vma, just get accounts right */
216 vm_unacct_memory(excess >> PAGE_SHIFT);
220 /* Restore VM_ACCOUNT if one or two pieces of vma left */
222 vma->vm_flags |= VM_ACCOUNT;
224 vma->vm_next->vm_flags |= VM_ACCOUNT;
227 mm->total_vm += new_len >> PAGE_SHIFT;
228 if (vm_flags & VM_LOCKED) {
229 mm->locked_vm += new_len >> PAGE_SHIFT;
230 if (new_len > old_len)
231 make_pages_present(new_addr + old_len,
239 * Expand (or shrink) an existing mapping, potentially moving it at the
240 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
242 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
243 * This option implies MREMAP_MAYMOVE.
245 unsigned long do_mremap(unsigned long addr,
246 unsigned long old_len, unsigned long new_len,
247 unsigned long flags, unsigned long new_addr)
249 struct vm_area_struct *vma;
250 unsigned long ret = -EINVAL;
251 unsigned long charged = 0;
253 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
256 if (addr & ~PAGE_MASK)
259 old_len = PAGE_ALIGN(old_len);
260 new_len = PAGE_ALIGN(new_len);
263 * We allow a zero old-len as a special case
264 * for DOS-emu "duplicate shm area" thing. But
265 * a zero new-len is nonsensical.
270 /* new_addr is only valid if MREMAP_FIXED is specified */
271 if (flags & MREMAP_FIXED) {
272 if (new_addr & ~PAGE_MASK)
274 if (!(flags & MREMAP_MAYMOVE))
277 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
280 /* Check if the location we're moving into overlaps the
281 * old location at all, and fail if it does.
283 if ((new_addr <= addr) && (new_addr+new_len) > addr)
286 if ((addr <= new_addr) && (addr+old_len) > new_addr)
289 ret = do_munmap(current->mm, new_addr, new_len);
295 * Always allow a shrinking remap: that just unmaps
296 * the unnecessary pages..
297 * do_munmap does all the needed commit accounting
299 if (old_len >= new_len) {
300 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
301 if (ret && old_len != new_len)
304 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
310 * Ok, we need to grow.. or relocate.
313 vma = find_vma(current->mm, addr);
314 if (!vma || vma->vm_start > addr)
316 if (is_vm_hugetlb_page(vma)) {
320 /* We can't remap across vm area boundaries */
321 if (old_len > vma->vm_end - addr)
323 if (vma->vm_flags & VM_DONTEXPAND) {
324 if (new_len > old_len)
327 if (vma->vm_flags & VM_LOCKED) {
328 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
329 locked += new_len - old_len;
331 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
335 if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
336 > current->rlim[RLIMIT_AS].rlim_cur)
339 if (vma->vm_flags & VM_ACCOUNT) {
340 charged = (new_len - old_len) >> PAGE_SHIFT;
341 if (security_vm_enough_memory(charged))
345 /* old_len exactly to the end of the area..
346 * And we're not relocating the area.
348 if (old_len == vma->vm_end - addr &&
349 !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
350 (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
351 unsigned long max_addr = TASK_SIZE;
353 max_addr = vma->vm_next->vm_start;
354 /* can we just expand the current mapping? */
355 if (max_addr - addr >= new_len) {
356 int pages = (new_len - old_len) >> PAGE_SHIFT;
358 vma_adjust(vma, vma->vm_start,
359 addr + new_len, vma->vm_pgoff, NULL);
361 current->mm->total_vm += pages;
362 if (vma->vm_flags & VM_LOCKED) {
363 current->mm->locked_vm += pages;
364 make_pages_present(addr + old_len,
373 * We weren't able to just expand or shrink the area,
374 * we need to create a new one and move it..
377 if (flags & MREMAP_MAYMOVE) {
378 if (!(flags & MREMAP_FIXED)) {
379 unsigned long map_flags = 0;
380 if (vma->vm_flags & VM_MAYSHARE)
381 map_flags |= MAP_SHARED;
383 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
384 vma->vm_pgoff, map_flags,
385 vma->vm_flags & VM_EXEC);
387 if (new_addr & ~PAGE_MASK)
390 ret = move_vma(vma, addr, old_len, new_len, new_addr);
393 if (ret & ~PAGE_MASK)
394 vm_unacct_memory(charged);
399 asmlinkage unsigned long sys_mremap(unsigned long addr,
400 unsigned long old_len, unsigned long new_len,
401 unsigned long flags, unsigned long new_addr)
405 down_write(¤t->mm->mmap_sem);
406 ret = do_mremap(addr, old_len, new_len, flags, new_addr);
407 up_write(¤t->mm->mmap_sem);