mremap.c revision 119f657c72fc07d6fd28c61de59cfba1566970a9 
1/*
2 *	mm/mremap.c
3 *
4 *	(C) Copyright 1996 Linus Torvalds
5 *
6 *	Address space accounting code	<alan@redhat.com>
7 *	(C) Copyright 2002 Red Hat Inc, All Rights Reserved
8 */
9
10#include <linux/mm.h>
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/fs.h>
17#include <linux/highmem.h>
18#include <linux/security.h>
19#include <linux/syscalls.h>
20
21#include <asm/uaccess.h>
22#include <asm/cacheflush.h>
23#include <asm/tlbflush.h>
24
25static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
26{
27	pgd_t *pgd;
28	pud_t *pud;
29	pmd_t *pmd;
30	pte_t *pte = NULL;
31
32	pgd = pgd_offset(mm, addr);
33	if (pgd_none_or_clear_bad(pgd))
34		goto end;
35
36	pud = pud_offset(pgd, addr);
37	if (pud_none_or_clear_bad(pud))
38		goto end;
39
40	pmd = pmd_offset(pud, addr);
41	if (pmd_none_or_clear_bad(pmd))
42		goto end;
43
44	pte = pte_offset_map_nested(pmd, addr);
45	if (pte_none(*pte)) {
46		pte_unmap_nested(pte);
47		pte = NULL;
48	}
49end:
50	return pte;
51}
52
53static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
54{
55	pgd_t *pgd;
56	pud_t *pud;
57	pmd_t *pmd;
58
59	pgd = pgd_offset(mm, addr);
60	if (pgd_none_or_clear_bad(pgd))
61		return NULL;
62
63	pud = pud_offset(pgd, addr);
64	if (pud_none_or_clear_bad(pud))
65		return NULL;
66
67	pmd = pmd_offset(pud, addr);
68	if (pmd_none_or_clear_bad(pmd))
69		return NULL;
70
71	return pte_offset_map(pmd, addr);
72}
73
74static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
75{
76	pgd_t *pgd;
77	pud_t *pud;
78	pmd_t *pmd;
79	pte_t *pte = NULL;
80
81	pgd = pgd_offset(mm, addr);
82
83	pud = pud_alloc(mm, pgd, addr);
84	if (!pud)
85		return NULL;
86	pmd = pmd_alloc(mm, pud, addr);
87	if (pmd)
88		pte = pte_alloc_map(mm, pmd, addr);
89	return pte;
90}
91
92static int
93move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
94		struct vm_area_struct *new_vma, unsigned long new_addr)
95{
96	struct address_space *mapping = NULL;
97	struct mm_struct *mm = vma->vm_mm;
98	int error = 0;
99	pte_t *src, *dst;
100
101	if (vma->vm_file) {
102		/*
103		 * Subtle point from Rajesh Venkatasubramanian: before
104		 * moving file-based ptes, we must lock vmtruncate out,
105		 * since it might clean the dst vma before the src vma,
106		 * and we propagate stale pages into the dst afterward.
107		 */
108		mapping = vma->vm_file->f_mapping;
109		spin_lock(&mapping->i_mmap_lock);
110		if (new_vma->vm_truncate_count &&
111		    new_vma->vm_truncate_count != vma->vm_truncate_count)
112			new_vma->vm_truncate_count = 0;
113	}
114	spin_lock(&mm->page_table_lock);
115
116	src = get_one_pte_map_nested(mm, old_addr);
117	if (src) {
118		/*
119		 * Look to see whether alloc_one_pte_map needs to perform a
120		 * memory allocation.  If it does then we need to drop the
121		 * atomic kmap
122		 */
123		dst = get_one_pte_map(mm, new_addr);
124		if (unlikely(!dst)) {
125			pte_unmap_nested(src);
126			if (mapping)
127				spin_unlock(&mapping->i_mmap_lock);
128			dst = alloc_one_pte_map(mm, new_addr);
129			if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
130				spin_unlock(&mm->page_table_lock);
131				spin_lock(&mapping->i_mmap_lock);
132				spin_lock(&mm->page_table_lock);
133			}
134			src = get_one_pte_map_nested(mm, old_addr);
135		}
136		/*
137		 * Since alloc_one_pte_map can drop and re-acquire
138		 * page_table_lock, we should re-check the src entry...
139		 */
140		if (src) {
141			if (dst) {
142				pte_t pte;
143				pte = ptep_clear_flush(vma, old_addr, src);
144				set_pte_at(mm, new_addr, dst, pte);
145			} else
146				error = -ENOMEM;
147			pte_unmap_nested(src);
148		}
149		if (dst)
150			pte_unmap(dst);
151	}
152	spin_unlock(&mm->page_table_lock);
153	if (mapping)
154		spin_unlock(&mapping->i_mmap_lock);
155	return error;
156}
157
158static unsigned long move_page_tables(struct vm_area_struct *vma,
159		unsigned long old_addr, struct vm_area_struct *new_vma,
160		unsigned long new_addr, unsigned long len)
161{
162	unsigned long offset;
163
164	flush_cache_range(vma, old_addr, old_addr + len);
165
166	/*
167	 * This is not the clever way to do this, but we're taking the
168	 * easy way out on the assumption that most remappings will be
169	 * only a few pages.. This also makes error recovery easier.
170	 */
171	for (offset = 0; offset < len; offset += PAGE_SIZE) {
172		if (move_one_page(vma, old_addr + offset,
173				new_vma, new_addr + offset) < 0)
174			break;
175		cond_resched();
176	}
177	return offset;
178}
179
180static unsigned long move_vma(struct vm_area_struct *vma,
181		unsigned long old_addr, unsigned long old_len,
182		unsigned long new_len, unsigned long new_addr)
183{
184	struct mm_struct *mm = vma->vm_mm;
185	struct vm_area_struct *new_vma;
186	unsigned long vm_flags = vma->vm_flags;
187	unsigned long new_pgoff;
188	unsigned long moved_len;
189	unsigned long excess = 0;
190	int split = 0;
191
192	/*
193	 * We'd prefer to avoid failure later on in do_munmap:
194	 * which may split one vma into three before unmapping.
195	 */
196	if (mm->map_count >= sysctl_max_map_count - 3)
197		return -ENOMEM;
198
199	new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
200	new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
201	if (!new_vma)
202		return -ENOMEM;
203
204	moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
205	if (moved_len < old_len) {
206		/*
207		 * On error, move entries back from new area to old,
208		 * which will succeed since page tables still there,
209		 * and then proceed to unmap new area instead of old.
210		 */
211		move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
212		vma = new_vma;
213		old_len = new_len;
214		old_addr = new_addr;
215		new_addr = -ENOMEM;
216	}
217
218	/* Conceal VM_ACCOUNT so old reservation is not undone */
219	if (vm_flags & VM_ACCOUNT) {
220		vma->vm_flags &= ~VM_ACCOUNT;
221		excess = vma->vm_end - vma->vm_start - old_len;
222		if (old_addr > vma->vm_start &&
223		    old_addr + old_len < vma->vm_end)
224			split = 1;
225	}
226
227	if (do_munmap(mm, old_addr, old_len) < 0) {
228		/* OOM: unable to split vma, just get accounts right */
229		vm_unacct_memory(excess >> PAGE_SHIFT);
230		excess = 0;
231	}
232
233	/* Restore VM_ACCOUNT if one or two pieces of vma left */
234	if (excess) {
235		vma->vm_flags |= VM_ACCOUNT;
236		if (split)
237			vma->vm_next->vm_flags |= VM_ACCOUNT;
238	}
239
240	mm->total_vm += new_len >> PAGE_SHIFT;
241	__vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
242	if (vm_flags & VM_LOCKED) {
243		mm->locked_vm += new_len >> PAGE_SHIFT;
244		if (new_len > old_len)
245			make_pages_present(new_addr + old_len,
246					   new_addr + new_len);
247	}
248
249	return new_addr;
250}
251
252/*
253 * Expand (or shrink) an existing mapping, potentially moving it at the
254 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
255 *
256 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
257 * This option implies MREMAP_MAYMOVE.
258 */
259unsigned long do_mremap(unsigned long addr,
260	unsigned long old_len, unsigned long new_len,
261	unsigned long flags, unsigned long new_addr)
262{
263	struct vm_area_struct *vma;
264	unsigned long ret = -EINVAL;
265	unsigned long charged = 0;
266
267	if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
268		goto out;
269
270	if (addr & ~PAGE_MASK)
271		goto out;
272
273	old_len = PAGE_ALIGN(old_len);
274	new_len = PAGE_ALIGN(new_len);
275
276	/*
277	 * We allow a zero old-len as a special case
278	 * for DOS-emu "duplicate shm area" thing. But
279	 * a zero new-len is nonsensical.
280	 */
281	if (!new_len)
282		goto out;
283
284	/* new_addr is only valid if MREMAP_FIXED is specified */
285	if (flags & MREMAP_FIXED) {
286		if (new_addr & ~PAGE_MASK)
287			goto out;
288		if (!(flags & MREMAP_MAYMOVE))
289			goto out;
290
291		if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
292			goto out;
293
294		/* Check if the location we're moving into overlaps the
295		 * old location at all, and fail if it does.
296		 */
297		if ((new_addr <= addr) && (new_addr+new_len) > addr)
298			goto out;
299
300		if ((addr <= new_addr) && (addr+old_len) > new_addr)
301			goto out;
302
303		ret = do_munmap(current->mm, new_addr, new_len);
304		if (ret)
305			goto out;
306	}
307
308	/*
309	 * Always allow a shrinking remap: that just unmaps
310	 * the unnecessary pages..
311	 * do_munmap does all the needed commit accounting
312	 */
313	if (old_len >= new_len) {
314		ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
315		if (ret && old_len != new_len)
316			goto out;
317		ret = addr;
318		if (!(flags & MREMAP_FIXED) || (new_addr == addr))
319			goto out;
320		old_len = new_len;
321	}
322
323	/*
324	 * Ok, we need to grow..  or relocate.
325	 */
326	ret = -EFAULT;
327	vma = find_vma(current->mm, addr);
328	if (!vma || vma->vm_start > addr)
329		goto out;
330	if (is_vm_hugetlb_page(vma)) {
331		ret = -EINVAL;
332		goto out;
333	}
334	/* We can't remap across vm area boundaries */
335	if (old_len > vma->vm_end - addr)
336		goto out;
337	if (vma->vm_flags & VM_DONTEXPAND) {
338		if (new_len > old_len)
339			goto out;
340	}
341	if (vma->vm_flags & VM_LOCKED) {
342		unsigned long locked, lock_limit;
343		locked = current->mm->locked_vm << PAGE_SHIFT;
344		lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
345		locked += new_len - old_len;
346		ret = -EAGAIN;
347		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
348			goto out;
349	}
350	if (!may_expand_vm(current->mm, (new_len - old_len) >> PAGE_SHIFT)) {
351		ret = -ENOMEM;
352		goto out;
353	}
354
355	if (vma->vm_flags & VM_ACCOUNT) {
356		charged = (new_len - old_len) >> PAGE_SHIFT;
357		if (security_vm_enough_memory(charged))
358			goto out_nc;
359	}
360
361	/* old_len exactly to the end of the area..
362	 * And we're not relocating the area.
363	 */
364	if (old_len == vma->vm_end - addr &&
365	    !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
366	    (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
367		unsigned long max_addr = TASK_SIZE;
368		if (vma->vm_next)
369			max_addr = vma->vm_next->vm_start;
370		/* can we just expand the current mapping? */
371		if (max_addr - addr >= new_len) {
372			int pages = (new_len - old_len) >> PAGE_SHIFT;
373
374			vma_adjust(vma, vma->vm_start,
375				addr + new_len, vma->vm_pgoff, NULL);
376
377			current->mm->total_vm += pages;
378			__vm_stat_account(vma->vm_mm, vma->vm_flags,
379							vma->vm_file, pages);
380			if (vma->vm_flags & VM_LOCKED) {
381				current->mm->locked_vm += pages;
382				make_pages_present(addr + old_len,
383						   addr + new_len);
384			}
385			ret = addr;
386			goto out;
387		}
388	}
389
390	/*
391	 * We weren't able to just expand or shrink the area,
392	 * we need to create a new one and move it..
393	 */
394	ret = -ENOMEM;
395	if (flags & MREMAP_MAYMOVE) {
396		if (!(flags & MREMAP_FIXED)) {
397			unsigned long map_flags = 0;
398			if (vma->vm_flags & VM_MAYSHARE)
399				map_flags |= MAP_SHARED;
400
401			new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
402						vma->vm_pgoff, map_flags);
403			ret = new_addr;
404			if (new_addr & ~PAGE_MASK)
405				goto out;
406		}
407		ret = move_vma(vma, addr, old_len, new_len, new_addr);
408	}
409out:
410	if (ret & ~PAGE_MASK)
411		vm_unacct_memory(charged);
412out_nc:
413	return ret;
414}
415
416asmlinkage unsigned long sys_mremap(unsigned long addr,
417	unsigned long old_len, unsigned long new_len,
418	unsigned long flags, unsigned long new_addr)
419{
420	unsigned long ret;
421
422	down_write(&current->mm->mmap_sem);
423	ret = do_mremap(addr, old_len, new_len, flags, new_addr);
424	up_write(&current->mm->mmap_sem);
425	return ret;
426}
427