xref: /openbmc/linux/mm/mremap.c (revision 92a2c6b2)
1 /*
2  *	mm/mremap.c
3  *
4  *	(C) Copyright 1996 Linus Torvalds
5  *
6  *	Address space accounting code	<alan@lxorguk.ukuu.org.uk>
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/shm.h>
13 #include <linux/ksm.h>
14 #include <linux/mman.h>
15 #include <linux/swap.h>
16 #include <linux/capability.h>
17 #include <linux/fs.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 
26 #include <asm/cacheflush.h>
27 #include <asm/tlbflush.h>
28 
29 #include "internal.h"
30 
31 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
32 {
33 	pgd_t *pgd;
34 	pud_t *pud;
35 	pmd_t *pmd;
36 
37 	pgd = pgd_offset(mm, addr);
38 	if (pgd_none_or_clear_bad(pgd))
39 		return NULL;
40 
41 	pud = pud_offset(pgd, addr);
42 	if (pud_none_or_clear_bad(pud))
43 		return NULL;
44 
45 	pmd = pmd_offset(pud, addr);
46 	if (pmd_none(*pmd))
47 		return NULL;
48 
49 	return pmd;
50 }
51 
52 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
53 			    unsigned long addr)
54 {
55 	pgd_t *pgd;
56 	pud_t *pud;
57 	pmd_t *pmd;
58 
59 	pgd = pgd_offset(mm, addr);
60 	pud = pud_alloc(mm, pgd, addr);
61 	if (!pud)
62 		return NULL;
63 
64 	pmd = pmd_alloc(mm, pud, addr);
65 	if (!pmd)
66 		return NULL;
67 
68 	VM_BUG_ON(pmd_trans_huge(*pmd));
69 
70 	return pmd;
71 }
72 
73 static pte_t move_soft_dirty_pte(pte_t pte)
74 {
75 	/*
76 	 * Set soft dirty bit so we can notice
77 	 * in userspace the ptes were moved.
78 	 */
79 #ifdef CONFIG_MEM_SOFT_DIRTY
80 	if (pte_present(pte))
81 		pte = pte_mksoft_dirty(pte);
82 	else if (is_swap_pte(pte))
83 		pte = pte_swp_mksoft_dirty(pte);
84 #endif
85 	return pte;
86 }
87 
88 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
89 		unsigned long old_addr, unsigned long old_end,
90 		struct vm_area_struct *new_vma, pmd_t *new_pmd,
91 		unsigned long new_addr, bool need_rmap_locks)
92 {
93 	struct address_space *mapping = NULL;
94 	struct anon_vma *anon_vma = NULL;
95 	struct mm_struct *mm = vma->vm_mm;
96 	pte_t *old_pte, *new_pte, pte;
97 	spinlock_t *old_ptl, *new_ptl;
98 
99 	/*
100 	 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
101 	 * locks to ensure that rmap will always observe either the old or the
102 	 * new ptes. This is the easiest way to avoid races with
103 	 * truncate_pagecache(), page migration, etc...
104 	 *
105 	 * When need_rmap_locks is false, we use other ways to avoid
106 	 * such races:
107 	 *
108 	 * - During exec() shift_arg_pages(), we use a specially tagged vma
109 	 *   which rmap call sites look for using is_vma_temporary_stack().
110 	 *
111 	 * - During mremap(), new_vma is often known to be placed after vma
112 	 *   in rmap traversal order. This ensures rmap will always observe
113 	 *   either the old pte, or the new pte, or both (the page table locks
114 	 *   serialize access to individual ptes, but only rmap traversal
115 	 *   order guarantees that we won't miss both the old and new ptes).
116 	 */
117 	if (need_rmap_locks) {
118 		if (vma->vm_file) {
119 			mapping = vma->vm_file->f_mapping;
120 			i_mmap_lock_write(mapping);
121 		}
122 		if (vma->anon_vma) {
123 			anon_vma = vma->anon_vma;
124 			anon_vma_lock_write(anon_vma);
125 		}
126 	}
127 
128 	/*
129 	 * We don't have to worry about the ordering of src and dst
130 	 * pte locks because exclusive mmap_sem prevents deadlock.
131 	 */
132 	old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
133 	new_pte = pte_offset_map(new_pmd, new_addr);
134 	new_ptl = pte_lockptr(mm, new_pmd);
135 	if (new_ptl != old_ptl)
136 		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
137 	arch_enter_lazy_mmu_mode();
138 
139 	for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
140 				   new_pte++, new_addr += PAGE_SIZE) {
141 		if (pte_none(*old_pte))
142 			continue;
143 		pte = ptep_get_and_clear(mm, old_addr, old_pte);
144 		pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
145 		pte = move_soft_dirty_pte(pte);
146 		set_pte_at(mm, new_addr, new_pte, pte);
147 	}
148 
149 	arch_leave_lazy_mmu_mode();
150 	if (new_ptl != old_ptl)
151 		spin_unlock(new_ptl);
152 	pte_unmap(new_pte - 1);
153 	pte_unmap_unlock(old_pte - 1, old_ptl);
154 	if (anon_vma)
155 		anon_vma_unlock_write(anon_vma);
156 	if (mapping)
157 		i_mmap_unlock_write(mapping);
158 }
159 
160 #define LATENCY_LIMIT	(64 * PAGE_SIZE)
161 
162 unsigned long move_page_tables(struct vm_area_struct *vma,
163 		unsigned long old_addr, struct vm_area_struct *new_vma,
164 		unsigned long new_addr, unsigned long len,
165 		bool need_rmap_locks)
166 {
167 	unsigned long extent, next, old_end;
168 	pmd_t *old_pmd, *new_pmd;
169 	bool need_flush = false;
170 	unsigned long mmun_start;	/* For mmu_notifiers */
171 	unsigned long mmun_end;		/* For mmu_notifiers */
172 
173 	old_end = old_addr + len;
174 	flush_cache_range(vma, old_addr, old_end);
175 
176 	mmun_start = old_addr;
177 	mmun_end   = old_end;
178 	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
179 
180 	for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
181 		cond_resched();
182 		next = (old_addr + PMD_SIZE) & PMD_MASK;
183 		/* even if next overflowed, extent below will be ok */
184 		extent = next - old_addr;
185 		if (extent > old_end - old_addr)
186 			extent = old_end - old_addr;
187 		old_pmd = get_old_pmd(vma->vm_mm, old_addr);
188 		if (!old_pmd)
189 			continue;
190 		new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
191 		if (!new_pmd)
192 			break;
193 		if (pmd_trans_huge(*old_pmd)) {
194 			int err = 0;
195 			if (extent == HPAGE_PMD_SIZE) {
196 				VM_BUG_ON_VMA(vma->vm_file || !vma->anon_vma,
197 					      vma);
198 				/* See comment in move_ptes() */
199 				if (need_rmap_locks)
200 					anon_vma_lock_write(vma->anon_vma);
201 				err = move_huge_pmd(vma, new_vma, old_addr,
202 						    new_addr, old_end,
203 						    old_pmd, new_pmd);
204 				if (need_rmap_locks)
205 					anon_vma_unlock_write(vma->anon_vma);
206 			}
207 			if (err > 0) {
208 				need_flush = true;
209 				continue;
210 			} else if (!err) {
211 				split_huge_page_pmd(vma, old_addr, old_pmd);
212 			}
213 			VM_BUG_ON(pmd_trans_huge(*old_pmd));
214 		}
215 		if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
216 						      new_pmd, new_addr))
217 			break;
218 		next = (new_addr + PMD_SIZE) & PMD_MASK;
219 		if (extent > next - new_addr)
220 			extent = next - new_addr;
221 		if (extent > LATENCY_LIMIT)
222 			extent = LATENCY_LIMIT;
223 		move_ptes(vma, old_pmd, old_addr, old_addr + extent,
224 			  new_vma, new_pmd, new_addr, need_rmap_locks);
225 		need_flush = true;
226 	}
227 	if (likely(need_flush))
228 		flush_tlb_range(vma, old_end-len, old_addr);
229 
230 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
231 
232 	return len + old_addr - old_end;	/* how much done */
233 }
234 
235 static unsigned long move_vma(struct vm_area_struct *vma,
236 		unsigned long old_addr, unsigned long old_len,
237 		unsigned long new_len, unsigned long new_addr, bool *locked)
238 {
239 	struct mm_struct *mm = vma->vm_mm;
240 	struct vm_area_struct *new_vma;
241 	unsigned long vm_flags = vma->vm_flags;
242 	unsigned long new_pgoff;
243 	unsigned long moved_len;
244 	unsigned long excess = 0;
245 	unsigned long hiwater_vm;
246 	int split = 0;
247 	int err;
248 	bool need_rmap_locks;
249 
250 	/*
251 	 * We'd prefer to avoid failure later on in do_munmap:
252 	 * which may split one vma into three before unmapping.
253 	 */
254 	if (mm->map_count >= sysctl_max_map_count - 3)
255 		return -ENOMEM;
256 
257 	/*
258 	 * Advise KSM to break any KSM pages in the area to be moved:
259 	 * it would be confusing if they were to turn up at the new
260 	 * location, where they happen to coincide with different KSM
261 	 * pages recently unmapped.  But leave vma->vm_flags as it was,
262 	 * so KSM can come around to merge on vma and new_vma afterwards.
263 	 */
264 	err = ksm_madvise(vma, old_addr, old_addr + old_len,
265 						MADV_UNMERGEABLE, &vm_flags);
266 	if (err)
267 		return err;
268 
269 	new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
270 	new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
271 			   &need_rmap_locks);
272 	if (!new_vma)
273 		return -ENOMEM;
274 
275 	moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
276 				     need_rmap_locks);
277 	if (moved_len < old_len) {
278 		/*
279 		 * On error, move entries back from new area to old,
280 		 * which will succeed since page tables still there,
281 		 * and then proceed to unmap new area instead of old.
282 		 */
283 		move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
284 				 true);
285 		vma = new_vma;
286 		old_len = new_len;
287 		old_addr = new_addr;
288 		new_addr = -ENOMEM;
289 	} else if (vma->vm_file && vma->vm_file->f_op->mremap)
290 		vma->vm_file->f_op->mremap(vma->vm_file, new_vma);
291 
292 	/* Conceal VM_ACCOUNT so old reservation is not undone */
293 	if (vm_flags & VM_ACCOUNT) {
294 		vma->vm_flags &= ~VM_ACCOUNT;
295 		excess = vma->vm_end - vma->vm_start - old_len;
296 		if (old_addr > vma->vm_start &&
297 		    old_addr + old_len < vma->vm_end)
298 			split = 1;
299 	}
300 
301 	/*
302 	 * If we failed to move page tables we still do total_vm increment
303 	 * since do_munmap() will decrement it by old_len == new_len.
304 	 *
305 	 * Since total_vm is about to be raised artificially high for a
306 	 * moment, we need to restore high watermark afterwards: if stats
307 	 * are taken meanwhile, total_vm and hiwater_vm appear too high.
308 	 * If this were a serious issue, we'd add a flag to do_munmap().
309 	 */
310 	hiwater_vm = mm->hiwater_vm;
311 	vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
312 
313 	if (do_munmap(mm, old_addr, old_len) < 0) {
314 		/* OOM: unable to split vma, just get accounts right */
315 		vm_unacct_memory(excess >> PAGE_SHIFT);
316 		excess = 0;
317 	}
318 	mm->hiwater_vm = hiwater_vm;
319 
320 	/* Restore VM_ACCOUNT if one or two pieces of vma left */
321 	if (excess) {
322 		vma->vm_flags |= VM_ACCOUNT;
323 		if (split)
324 			vma->vm_next->vm_flags |= VM_ACCOUNT;
325 	}
326 
327 	if (vm_flags & VM_LOCKED) {
328 		mm->locked_vm += new_len >> PAGE_SHIFT;
329 		*locked = true;
330 	}
331 
332 	return new_addr;
333 }
334 
335 static struct vm_area_struct *vma_to_resize(unsigned long addr,
336 	unsigned long old_len, unsigned long new_len, unsigned long *p)
337 {
338 	struct mm_struct *mm = current->mm;
339 	struct vm_area_struct *vma = find_vma(mm, addr);
340 
341 	if (!vma || vma->vm_start > addr)
342 		goto Efault;
343 
344 	if (is_vm_hugetlb_page(vma))
345 		goto Einval;
346 
347 	/* We can't remap across vm area boundaries */
348 	if (old_len > vma->vm_end - addr)
349 		goto Efault;
350 
351 	/* Need to be careful about a growing mapping */
352 	if (new_len > old_len) {
353 		unsigned long pgoff;
354 
355 		if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
356 			goto Efault;
357 		pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
358 		pgoff += vma->vm_pgoff;
359 		if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
360 			goto Einval;
361 	}
362 
363 	if (vma->vm_flags & VM_LOCKED) {
364 		unsigned long locked, lock_limit;
365 		locked = mm->locked_vm << PAGE_SHIFT;
366 		lock_limit = rlimit(RLIMIT_MEMLOCK);
367 		locked += new_len - old_len;
368 		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
369 			goto Eagain;
370 	}
371 
372 	if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
373 		goto Enomem;
374 
375 	if (vma->vm_flags & VM_ACCOUNT) {
376 		unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
377 		if (security_vm_enough_memory_mm(mm, charged))
378 			goto Efault;
379 		*p = charged;
380 	}
381 
382 	return vma;
383 
384 Efault:	/* very odd choice for most of the cases, but... */
385 	return ERR_PTR(-EFAULT);
386 Einval:
387 	return ERR_PTR(-EINVAL);
388 Enomem:
389 	return ERR_PTR(-ENOMEM);
390 Eagain:
391 	return ERR_PTR(-EAGAIN);
392 }
393 
394 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
395 		unsigned long new_addr, unsigned long new_len, bool *locked)
396 {
397 	struct mm_struct *mm = current->mm;
398 	struct vm_area_struct *vma;
399 	unsigned long ret = -EINVAL;
400 	unsigned long charged = 0;
401 	unsigned long map_flags;
402 
403 	if (new_addr & ~PAGE_MASK)
404 		goto out;
405 
406 	if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
407 		goto out;
408 
409 	/* Check if the location we're moving into overlaps the
410 	 * old location at all, and fail if it does.
411 	 */
412 	if ((new_addr <= addr) && (new_addr+new_len) > addr)
413 		goto out;
414 
415 	if ((addr <= new_addr) && (addr+old_len) > new_addr)
416 		goto out;
417 
418 	ret = do_munmap(mm, new_addr, new_len);
419 	if (ret)
420 		goto out;
421 
422 	if (old_len >= new_len) {
423 		ret = do_munmap(mm, addr+new_len, old_len - new_len);
424 		if (ret && old_len != new_len)
425 			goto out;
426 		old_len = new_len;
427 	}
428 
429 	vma = vma_to_resize(addr, old_len, new_len, &charged);
430 	if (IS_ERR(vma)) {
431 		ret = PTR_ERR(vma);
432 		goto out;
433 	}
434 
435 	map_flags = MAP_FIXED;
436 	if (vma->vm_flags & VM_MAYSHARE)
437 		map_flags |= MAP_SHARED;
438 
439 	ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
440 				((addr - vma->vm_start) >> PAGE_SHIFT),
441 				map_flags);
442 	if (ret & ~PAGE_MASK)
443 		goto out1;
444 
445 	ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
446 	if (!(ret & ~PAGE_MASK))
447 		goto out;
448 out1:
449 	vm_unacct_memory(charged);
450 
451 out:
452 	return ret;
453 }
454 
455 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
456 {
457 	unsigned long end = vma->vm_end + delta;
458 	if (end < vma->vm_end) /* overflow */
459 		return 0;
460 	if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
461 		return 0;
462 	if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
463 			      0, MAP_FIXED) & ~PAGE_MASK)
464 		return 0;
465 	return 1;
466 }
467 
468 /*
469  * Expand (or shrink) an existing mapping, potentially moving it at the
470  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
471  *
472  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
473  * This option implies MREMAP_MAYMOVE.
474  */
475 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
476 		unsigned long, new_len, unsigned long, flags,
477 		unsigned long, new_addr)
478 {
479 	struct mm_struct *mm = current->mm;
480 	struct vm_area_struct *vma;
481 	unsigned long ret = -EINVAL;
482 	unsigned long charged = 0;
483 	bool locked = false;
484 
485 	if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
486 		return ret;
487 
488 	if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
489 		return ret;
490 
491 	if (addr & ~PAGE_MASK)
492 		return ret;
493 
494 	old_len = PAGE_ALIGN(old_len);
495 	new_len = PAGE_ALIGN(new_len);
496 
497 	/*
498 	 * We allow a zero old-len as a special case
499 	 * for DOS-emu "duplicate shm area" thing. But
500 	 * a zero new-len is nonsensical.
501 	 */
502 	if (!new_len)
503 		return ret;
504 
505 	down_write(&current->mm->mmap_sem);
506 
507 	if (flags & MREMAP_FIXED) {
508 		ret = mremap_to(addr, old_len, new_addr, new_len,
509 				&locked);
510 		goto out;
511 	}
512 
513 	/*
514 	 * Always allow a shrinking remap: that just unmaps
515 	 * the unnecessary pages..
516 	 * do_munmap does all the needed commit accounting
517 	 */
518 	if (old_len >= new_len) {
519 		ret = do_munmap(mm, addr+new_len, old_len - new_len);
520 		if (ret && old_len != new_len)
521 			goto out;
522 		ret = addr;
523 		goto out;
524 	}
525 
526 	/*
527 	 * Ok, we need to grow..
528 	 */
529 	vma = vma_to_resize(addr, old_len, new_len, &charged);
530 	if (IS_ERR(vma)) {
531 		ret = PTR_ERR(vma);
532 		goto out;
533 	}
534 
535 	/* old_len exactly to the end of the area..
536 	 */
537 	if (old_len == vma->vm_end - addr) {
538 		/* can we just expand the current mapping? */
539 		if (vma_expandable(vma, new_len - old_len)) {
540 			int pages = (new_len - old_len) >> PAGE_SHIFT;
541 
542 			if (vma_adjust(vma, vma->vm_start, addr + new_len,
543 				       vma->vm_pgoff, NULL)) {
544 				ret = -ENOMEM;
545 				goto out;
546 			}
547 
548 			vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
549 			if (vma->vm_flags & VM_LOCKED) {
550 				mm->locked_vm += pages;
551 				locked = true;
552 				new_addr = addr;
553 			}
554 			ret = addr;
555 			goto out;
556 		}
557 	}
558 
559 	/*
560 	 * We weren't able to just expand or shrink the area,
561 	 * we need to create a new one and move it..
562 	 */
563 	ret = -ENOMEM;
564 	if (flags & MREMAP_MAYMOVE) {
565 		unsigned long map_flags = 0;
566 		if (vma->vm_flags & VM_MAYSHARE)
567 			map_flags |= MAP_SHARED;
568 
569 		new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
570 					vma->vm_pgoff +
571 					((addr - vma->vm_start) >> PAGE_SHIFT),
572 					map_flags);
573 		if (new_addr & ~PAGE_MASK) {
574 			ret = new_addr;
575 			goto out;
576 		}
577 
578 		ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
579 	}
580 out:
581 	if (ret & ~PAGE_MASK)
582 		vm_unacct_memory(charged);
583 	up_write(&current->mm->mmap_sem);
584 	if (locked && new_len > old_len)
585 		mm_populate(new_addr + old_len, new_len - old_len);
586 	return ret;
587 }
588