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>
16 #include <linux/capability.h>
18 #include <linux/highmem.h>
19 #include <linux/security.h>
20 #include <linux/syscalls.h>
21 #include <linux/vs_base.h>
22 #include <linux/vs_memory.h>
24 #include <asm/uaccess.h>
25 #include <asm/cacheflush.h>
26 #include <asm/tlbflush.h>
29 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
35 pgd = pgd_offset(mm, addr);
36 if (pgd_none_or_clear_bad(pgd))
39 pud = pud_offset(pgd, addr);
40 if (pud_none_or_clear_bad(pud))
43 pmd = pmd_offset(pud, addr);
44 if (pmd_none_or_clear_bad(pmd))
50 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
56 pgd = pgd_offset(mm, addr);
57 pud = pud_alloc(mm, pgd, addr);
61 pmd = pmd_alloc(mm, pud, addr);
65 if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
71 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
72 unsigned long old_addr, unsigned long old_end,
73 struct vm_area_struct *new_vma, pmd_t *new_pmd,
74 unsigned long new_addr)
76 struct address_space *mapping = NULL;
77 struct mm_struct *mm = vma->vm_mm;
78 pte_t *old_pte, *new_pte, pte;
79 spinlock_t *old_ptl, *new_ptl;
83 * Subtle point from Rajesh Venkatasubramanian: before
84 * moving file-based ptes, we must lock vmtruncate out,
85 * since it might clean the dst vma before the src vma,
86 * and we propagate stale pages into the dst afterward.
88 mapping = vma->vm_file->f_mapping;
89 spin_lock(&mapping->i_mmap_lock);
90 if (new_vma->vm_truncate_count &&
91 new_vma->vm_truncate_count != vma->vm_truncate_count)
92 new_vma->vm_truncate_count = 0;
96 * We don't have to worry about the ordering of src and dst
97 * pte locks because exclusive mmap_sem prevents deadlock.
99 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
100 new_pte = pte_offset_map_nested(new_pmd, new_addr);
101 new_ptl = pte_lockptr(mm, new_pmd);
102 if (new_ptl != old_ptl)
105 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
106 new_pte++, new_addr += PAGE_SIZE) {
107 if (pte_none(*old_pte))
109 pte = ptep_clear_flush(vma, old_addr, old_pte);
110 /* ZERO_PAGE can be dependant on virtual addr */
111 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
112 set_pte_at(mm, new_addr, new_pte, pte);
115 if (new_ptl != old_ptl)
116 spin_unlock(new_ptl);
117 pte_unmap_nested(new_pte - 1);
118 pte_unmap_unlock(old_pte - 1, old_ptl);
120 spin_unlock(&mapping->i_mmap_lock);
123 #define LATENCY_LIMIT (64 * PAGE_SIZE)
125 static unsigned long move_page_tables(struct vm_area_struct *vma,
126 unsigned long old_addr, struct vm_area_struct *new_vma,
127 unsigned long new_addr, unsigned long len)
129 unsigned long extent, next, old_end;
130 pmd_t *old_pmd, *new_pmd;
132 old_end = old_addr + len;
133 flush_cache_range(vma, old_addr, old_end);
135 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
137 next = (old_addr + PMD_SIZE) & PMD_MASK;
138 if (next - 1 > old_end)
140 extent = next - old_addr;
141 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
144 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
147 next = (new_addr + PMD_SIZE) & PMD_MASK;
148 if (extent > next - new_addr)
149 extent = next - new_addr;
150 if (extent > LATENCY_LIMIT)
151 extent = LATENCY_LIMIT;
152 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
153 new_vma, new_pmd, new_addr);
156 return len + old_addr - old_end; /* how much done */
159 static unsigned long move_vma(struct vm_area_struct *vma,
160 unsigned long old_addr, unsigned long old_len,
161 unsigned long new_len, unsigned long new_addr)
163 struct mm_struct *mm = vma->vm_mm;
164 struct vm_area_struct *new_vma;
165 unsigned long vm_flags = vma->vm_flags;
166 unsigned long new_pgoff;
167 unsigned long moved_len;
168 unsigned long excess = 0;
169 unsigned long hiwater_vm;
173 * We'd prefer to avoid failure later on in do_munmap:
174 * which may split one vma into three before unmapping.
176 if (mm->map_count >= sysctl_max_map_count - 3)
179 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
180 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
184 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
185 if (moved_len < old_len) {
187 * On error, move entries back from new area to old,
188 * which will succeed since page tables still there,
189 * and then proceed to unmap new area instead of old.
191 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
198 /* Conceal VM_ACCOUNT so old reservation is not undone */
199 if (vm_flags & VM_ACCOUNT) {
200 vma->vm_flags &= ~VM_ACCOUNT;
201 excess = vma->vm_end - vma->vm_start - old_len;
202 if (old_addr > vma->vm_start &&
203 old_addr + old_len < vma->vm_end)
208 * If we failed to move page tables we still do total_vm increment
209 * since do_munmap() will decrement it by old_len == new_len.
211 * Since total_vm is about to be raised artificially high for a
212 * moment, we need to restore high watermark afterwards: if stats
213 * are taken meanwhile, total_vm and hiwater_vm appear too high.
214 * If this were a serious issue, we'd add a flag to do_munmap().
216 hiwater_vm = mm->hiwater_vm;
217 vx_vmpages_add(mm, new_len >> PAGE_SHIFT);
218 vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
220 if (do_munmap(mm, old_addr, old_len) < 0) {
221 /* OOM: unable to split vma, just get accounts right */
222 vm_unacct_memory(excess >> PAGE_SHIFT);
225 mm->hiwater_vm = hiwater_vm;
227 /* Restore VM_ACCOUNT if one or two pieces of vma left */
229 vma->vm_flags |= VM_ACCOUNT;
231 vma->vm_next->vm_flags |= VM_ACCOUNT;
234 if (vm_flags & VM_LOCKED) {
235 vx_vmlocked_add(mm, new_len >> PAGE_SHIFT);
236 if (new_len > old_len)
237 make_pages_present(new_addr + old_len,
245 * Expand (or shrink) an existing mapping, potentially moving it at the
246 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
248 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
249 * This option implies MREMAP_MAYMOVE.
251 unsigned long do_mremap(unsigned long addr,
252 unsigned long old_len, unsigned long new_len,
253 unsigned long flags, unsigned long new_addr)
255 struct mm_struct *mm = current->mm;
256 struct vm_area_struct *vma;
257 unsigned long ret = -EINVAL;
258 unsigned long charged = 0;
260 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
263 if (addr & ~PAGE_MASK)
266 old_len = PAGE_ALIGN(old_len);
267 new_len = PAGE_ALIGN(new_len);
270 * We allow a zero old-len as a special case
271 * for DOS-emu "duplicate shm area" thing. But
272 * a zero new-len is nonsensical.
277 /* new_addr is only valid if MREMAP_FIXED is specified */
278 if (flags & MREMAP_FIXED) {
279 if (new_addr & ~PAGE_MASK)
281 if (!(flags & MREMAP_MAYMOVE))
284 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
287 /* Check if the location we're moving into overlaps the
288 * old location at all, and fail if it does.
290 if ((new_addr <= addr) && (new_addr+new_len) > addr)
293 if ((addr <= new_addr) && (addr+old_len) > new_addr)
296 ret = do_munmap(mm, new_addr, new_len);
302 * Always allow a shrinking remap: that just unmaps
303 * the unnecessary pages..
304 * do_munmap does all the needed commit accounting
306 if (old_len >= new_len) {
307 ret = do_munmap(mm, addr+new_len, old_len - new_len);
308 if (ret && old_len != new_len)
311 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
317 * Ok, we need to grow.. or relocate.
320 vma = find_vma(mm, addr);
321 if (!vma || vma->vm_start > addr)
323 if (is_vm_hugetlb_page(vma)) {
327 /* We can't remap across vm area boundaries */
328 if (old_len > vma->vm_end - addr)
330 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
331 if (new_len > old_len)
334 if (vma->vm_flags & VM_LOCKED) {
335 unsigned long locked, lock_limit;
336 locked = mm->locked_vm << PAGE_SHIFT;
337 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
338 locked += new_len - old_len;
340 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
342 if (!vx_vmlocked_avail(current->mm,
343 (new_len - old_len) >> PAGE_SHIFT))
346 if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
351 if (vma->vm_flags & VM_ACCOUNT) {
352 charged = (new_len - old_len) >> PAGE_SHIFT;
353 if (security_vm_enough_memory(charged))
357 /* old_len exactly to the end of the area..
358 * And we're not relocating the area.
360 if (old_len == vma->vm_end - addr &&
361 !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
362 (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
363 unsigned long max_addr = TASK_SIZE;
365 max_addr = vma->vm_next->vm_start;
366 /* can we just expand the current mapping? */
367 if (max_addr - addr >= new_len) {
368 int pages = (new_len - old_len) >> PAGE_SHIFT;
370 vma_adjust(vma, vma->vm_start,
371 addr + new_len, vma->vm_pgoff, NULL);
373 vx_vmpages_add(mm, pages);
374 vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
375 if (vma->vm_flags & VM_LOCKED) {
376 vx_vmlocked_add(mm, pages);
377 make_pages_present(addr + old_len,
386 * We weren't able to just expand or shrink the area,
387 * we need to create a new one and move it..
390 if (flags & MREMAP_MAYMOVE) {
391 if (!(flags & MREMAP_FIXED)) {
392 unsigned long map_flags = 0;
393 if (vma->vm_flags & VM_MAYSHARE)
394 map_flags |= MAP_SHARED;
396 new_addr = get_unmapped_area_prot(vma->vm_file, 0, new_len,
397 vma->vm_pgoff, map_flags, vma->vm_flags & VM_EXEC);
399 if (new_addr & ~PAGE_MASK)
402 ret = move_vma(vma, addr, old_len, new_len, new_addr);
405 if (ret & ~PAGE_MASK)
406 vm_unacct_memory(charged);
411 asmlinkage unsigned long sys_mremap(unsigned long addr,
412 unsigned long old_len, unsigned long new_len,
413 unsigned long flags, unsigned long new_addr)
417 down_write(¤t->mm->mmap_sem);
418 ret = do_mremap(addr, old_len, new_len, flags, new_addr);
419 up_write(¤t->mm->mmap_sem);