]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/mremap.c
x86/msr: Add definitions for new speculation control MSRs
[mirror_ubuntu-artful-kernel.git] / mm / mremap.c
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/uaccess.h>
24 #include <linux/mm-arch-hooks.h>
25 #include <linux/userfaultfd_k.h>
26
27 #include <asm/cacheflush.h>
28 #include <asm/tlbflush.h>
29
30 #include "internal.h"
31
32 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
33 {
34 pgd_t *pgd;
35 p4d_t *p4d;
36 pud_t *pud;
37 pmd_t *pmd;
38
39 pgd = pgd_offset(mm, addr);
40 if (pgd_none_or_clear_bad(pgd))
41 return NULL;
42
43 p4d = p4d_offset(pgd, addr);
44 if (p4d_none_or_clear_bad(p4d))
45 return NULL;
46
47 pud = pud_offset(p4d, addr);
48 if (pud_none_or_clear_bad(pud))
49 return NULL;
50
51 pmd = pmd_offset(pud, addr);
52 if (pmd_none(*pmd))
53 return NULL;
54
55 return pmd;
56 }
57
58 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
59 unsigned long addr)
60 {
61 pgd_t *pgd;
62 p4d_t *p4d;
63 pud_t *pud;
64 pmd_t *pmd;
65
66 pgd = pgd_offset(mm, addr);
67 p4d = p4d_alloc(mm, pgd, addr);
68 if (!p4d)
69 return NULL;
70 pud = pud_alloc(mm, p4d, addr);
71 if (!pud)
72 return NULL;
73
74 pmd = pmd_alloc(mm, pud, addr);
75 if (!pmd)
76 return NULL;
77
78 VM_BUG_ON(pmd_trans_huge(*pmd));
79
80 return pmd;
81 }
82
83 static void take_rmap_locks(struct vm_area_struct *vma)
84 {
85 if (vma->vm_file)
86 i_mmap_lock_write(vma->vm_file->f_mapping);
87 if (vma->anon_vma)
88 anon_vma_lock_write(vma->anon_vma);
89 }
90
91 static void drop_rmap_locks(struct vm_area_struct *vma)
92 {
93 if (vma->anon_vma)
94 anon_vma_unlock_write(vma->anon_vma);
95 if (vma->vm_file)
96 i_mmap_unlock_write(vma->vm_file->f_mapping);
97 }
98
99 static pte_t move_soft_dirty_pte(pte_t pte)
100 {
101 /*
102 * Set soft dirty bit so we can notice
103 * in userspace the ptes were moved.
104 */
105 #ifdef CONFIG_MEM_SOFT_DIRTY
106 if (pte_present(pte))
107 pte = pte_mksoft_dirty(pte);
108 else if (is_swap_pte(pte))
109 pte = pte_swp_mksoft_dirty(pte);
110 #endif
111 return pte;
112 }
113
114 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
115 unsigned long old_addr, unsigned long old_end,
116 struct vm_area_struct *new_vma, pmd_t *new_pmd,
117 unsigned long new_addr, bool need_rmap_locks, bool *need_flush)
118 {
119 struct mm_struct *mm = vma->vm_mm;
120 pte_t *old_pte, *new_pte, pte;
121 spinlock_t *old_ptl, *new_ptl;
122 bool force_flush = false;
123 unsigned long len = old_end - old_addr;
124
125 /*
126 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
127 * locks to ensure that rmap will always observe either the old or the
128 * new ptes. This is the easiest way to avoid races with
129 * truncate_pagecache(), page migration, etc...
130 *
131 * When need_rmap_locks is false, we use other ways to avoid
132 * such races:
133 *
134 * - During exec() shift_arg_pages(), we use a specially tagged vma
135 * which rmap call sites look for using is_vma_temporary_stack().
136 *
137 * - During mremap(), new_vma is often known to be placed after vma
138 * in rmap traversal order. This ensures rmap will always observe
139 * either the old pte, or the new pte, or both (the page table locks
140 * serialize access to individual ptes, but only rmap traversal
141 * order guarantees that we won't miss both the old and new ptes).
142 */
143 if (need_rmap_locks)
144 take_rmap_locks(vma);
145
146 /*
147 * We don't have to worry about the ordering of src and dst
148 * pte locks because exclusive mmap_sem prevents deadlock.
149 */
150 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
151 new_pte = pte_offset_map(new_pmd, new_addr);
152 new_ptl = pte_lockptr(mm, new_pmd);
153 if (new_ptl != old_ptl)
154 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
155 flush_tlb_batched_pending(vma->vm_mm);
156 arch_enter_lazy_mmu_mode();
157
158 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
159 new_pte++, new_addr += PAGE_SIZE) {
160 if (pte_none(*old_pte))
161 continue;
162
163 pte = ptep_get_and_clear(mm, old_addr, old_pte);
164 /*
165 * If we are remapping a dirty PTE, make sure
166 * to flush TLB before we drop the PTL for the
167 * old PTE or we may race with page_mkclean().
168 *
169 * This check has to be done after we removed the
170 * old PTE from page tables or another thread may
171 * dirty it after the check and before the removal.
172 */
173 if (pte_present(pte) && pte_dirty(pte))
174 force_flush = true;
175 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
176 pte = move_soft_dirty_pte(pte);
177 set_pte_at(mm, new_addr, new_pte, pte);
178 }
179
180 arch_leave_lazy_mmu_mode();
181 if (new_ptl != old_ptl)
182 spin_unlock(new_ptl);
183 pte_unmap(new_pte - 1);
184 if (force_flush)
185 flush_tlb_range(vma, old_end - len, old_end);
186 else
187 *need_flush = true;
188 pte_unmap_unlock(old_pte - 1, old_ptl);
189 if (need_rmap_locks)
190 drop_rmap_locks(vma);
191 }
192
193 #define LATENCY_LIMIT (64 * PAGE_SIZE)
194
195 unsigned long move_page_tables(struct vm_area_struct *vma,
196 unsigned long old_addr, struct vm_area_struct *new_vma,
197 unsigned long new_addr, unsigned long len,
198 bool need_rmap_locks)
199 {
200 unsigned long extent, next, old_end;
201 pmd_t *old_pmd, *new_pmd;
202 bool need_flush = false;
203 unsigned long mmun_start; /* For mmu_notifiers */
204 unsigned long mmun_end; /* For mmu_notifiers */
205
206 old_end = old_addr + len;
207 flush_cache_range(vma, old_addr, old_end);
208
209 mmun_start = old_addr;
210 mmun_end = old_end;
211 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
212
213 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
214 cond_resched();
215 next = (old_addr + PMD_SIZE) & PMD_MASK;
216 /* even if next overflowed, extent below will be ok */
217 extent = next - old_addr;
218 if (extent > old_end - old_addr)
219 extent = old_end - old_addr;
220 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
221 if (!old_pmd)
222 continue;
223 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
224 if (!new_pmd)
225 break;
226 if (pmd_trans_huge(*old_pmd)) {
227 if (extent == HPAGE_PMD_SIZE) {
228 bool moved;
229 /* See comment in move_ptes() */
230 if (need_rmap_locks)
231 take_rmap_locks(vma);
232 moved = move_huge_pmd(vma, old_addr, new_addr,
233 old_end, old_pmd, new_pmd,
234 &need_flush);
235 if (need_rmap_locks)
236 drop_rmap_locks(vma);
237 if (moved)
238 continue;
239 }
240 split_huge_pmd(vma, old_pmd, old_addr);
241 if (pmd_trans_unstable(old_pmd))
242 continue;
243 }
244 if (pte_alloc(new_vma->vm_mm, new_pmd, new_addr))
245 break;
246 next = (new_addr + PMD_SIZE) & PMD_MASK;
247 if (extent > next - new_addr)
248 extent = next - new_addr;
249 if (extent > LATENCY_LIMIT)
250 extent = LATENCY_LIMIT;
251 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
252 new_pmd, new_addr, need_rmap_locks, &need_flush);
253 }
254 if (need_flush)
255 flush_tlb_range(vma, old_end-len, old_addr);
256
257 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
258
259 return len + old_addr - old_end; /* how much done */
260 }
261
262 static unsigned long move_vma(struct vm_area_struct *vma,
263 unsigned long old_addr, unsigned long old_len,
264 unsigned long new_len, unsigned long new_addr,
265 bool *locked, struct vm_userfaultfd_ctx *uf,
266 struct list_head *uf_unmap)
267 {
268 struct mm_struct *mm = vma->vm_mm;
269 struct vm_area_struct *new_vma;
270 unsigned long vm_flags = vma->vm_flags;
271 unsigned long new_pgoff;
272 unsigned long moved_len;
273 unsigned long excess = 0;
274 unsigned long hiwater_vm;
275 int split = 0;
276 int err;
277 bool need_rmap_locks;
278
279 /*
280 * We'd prefer to avoid failure later on in do_munmap:
281 * which may split one vma into three before unmapping.
282 */
283 if (mm->map_count >= sysctl_max_map_count - 3)
284 return -ENOMEM;
285
286 /*
287 * Advise KSM to break any KSM pages in the area to be moved:
288 * it would be confusing if they were to turn up at the new
289 * location, where they happen to coincide with different KSM
290 * pages recently unmapped. But leave vma->vm_flags as it was,
291 * so KSM can come around to merge on vma and new_vma afterwards.
292 */
293 err = ksm_madvise(vma, old_addr, old_addr + old_len,
294 MADV_UNMERGEABLE, &vm_flags);
295 if (err)
296 return err;
297
298 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
299 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
300 &need_rmap_locks);
301 if (!new_vma)
302 return -ENOMEM;
303
304 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
305 need_rmap_locks);
306 if (moved_len < old_len) {
307 err = -ENOMEM;
308 } else if (vma->vm_ops && vma->vm_ops->mremap) {
309 err = vma->vm_ops->mremap(new_vma);
310 }
311
312 if (unlikely(err)) {
313 /*
314 * On error, move entries back from new area to old,
315 * which will succeed since page tables still there,
316 * and then proceed to unmap new area instead of old.
317 */
318 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
319 true);
320 vma = new_vma;
321 old_len = new_len;
322 old_addr = new_addr;
323 new_addr = err;
324 } else {
325 mremap_userfaultfd_prep(new_vma, uf);
326 arch_remap(mm, old_addr, old_addr + old_len,
327 new_addr, new_addr + new_len);
328 }
329
330 /* Conceal VM_ACCOUNT so old reservation is not undone */
331 if (vm_flags & VM_ACCOUNT) {
332 vma->vm_flags &= ~VM_ACCOUNT;
333 excess = vma->vm_end - vma->vm_start - old_len;
334 if (old_addr > vma->vm_start &&
335 old_addr + old_len < vma->vm_end)
336 split = 1;
337 }
338
339 /*
340 * If we failed to move page tables we still do total_vm increment
341 * since do_munmap() will decrement it by old_len == new_len.
342 *
343 * Since total_vm is about to be raised artificially high for a
344 * moment, we need to restore high watermark afterwards: if stats
345 * are taken meanwhile, total_vm and hiwater_vm appear too high.
346 * If this were a serious issue, we'd add a flag to do_munmap().
347 */
348 hiwater_vm = mm->hiwater_vm;
349 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
350
351 /* Tell pfnmap has moved from this vma */
352 if (unlikely(vma->vm_flags & VM_PFNMAP))
353 untrack_pfn_moved(vma);
354
355 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
356 /* OOM: unable to split vma, just get accounts right */
357 vm_unacct_memory(excess >> PAGE_SHIFT);
358 excess = 0;
359 }
360 mm->hiwater_vm = hiwater_vm;
361
362 /* Restore VM_ACCOUNT if one or two pieces of vma left */
363 if (excess) {
364 vma->vm_flags |= VM_ACCOUNT;
365 if (split)
366 vma->vm_next->vm_flags |= VM_ACCOUNT;
367 }
368
369 if (vm_flags & VM_LOCKED) {
370 mm->locked_vm += new_len >> PAGE_SHIFT;
371 *locked = true;
372 }
373
374 return new_addr;
375 }
376
377 static struct vm_area_struct *vma_to_resize(unsigned long addr,
378 unsigned long old_len, unsigned long new_len, unsigned long *p)
379 {
380 struct mm_struct *mm = current->mm;
381 struct vm_area_struct *vma = find_vma(mm, addr);
382 unsigned long pgoff;
383
384 if (!vma || vma->vm_start > addr)
385 return ERR_PTR(-EFAULT);
386
387 if (is_vm_hugetlb_page(vma))
388 return ERR_PTR(-EINVAL);
389
390 /* We can't remap across vm area boundaries */
391 if (old_len > vma->vm_end - addr)
392 return ERR_PTR(-EFAULT);
393
394 if (new_len == old_len)
395 return vma;
396
397 /* Need to be careful about a growing mapping */
398 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
399 pgoff += vma->vm_pgoff;
400 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
401 return ERR_PTR(-EINVAL);
402
403 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
404 return ERR_PTR(-EFAULT);
405
406 if (vma->vm_flags & VM_LOCKED) {
407 unsigned long locked, lock_limit;
408 locked = mm->locked_vm << PAGE_SHIFT;
409 lock_limit = rlimit(RLIMIT_MEMLOCK);
410 locked += new_len - old_len;
411 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
412 return ERR_PTR(-EAGAIN);
413 }
414
415 if (!may_expand_vm(mm, vma->vm_flags,
416 (new_len - old_len) >> PAGE_SHIFT))
417 return ERR_PTR(-ENOMEM);
418
419 if (vma->vm_flags & VM_ACCOUNT) {
420 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
421 if (security_vm_enough_memory_mm(mm, charged))
422 return ERR_PTR(-ENOMEM);
423 *p = charged;
424 }
425
426 return vma;
427 }
428
429 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
430 unsigned long new_addr, unsigned long new_len, bool *locked,
431 struct vm_userfaultfd_ctx *uf,
432 struct list_head *uf_unmap_early,
433 struct list_head *uf_unmap)
434 {
435 struct mm_struct *mm = current->mm;
436 struct vm_area_struct *vma;
437 unsigned long ret = -EINVAL;
438 unsigned long charged = 0;
439 unsigned long map_flags;
440
441 if (offset_in_page(new_addr))
442 goto out;
443
444 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
445 goto out;
446
447 /* Ensure the old/new locations do not overlap */
448 if (addr + old_len > new_addr && new_addr + new_len > addr)
449 goto out;
450
451 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
452 if (ret)
453 goto out;
454
455 if (old_len >= new_len) {
456 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
457 if (ret && old_len != new_len)
458 goto out;
459 old_len = new_len;
460 }
461
462 vma = vma_to_resize(addr, old_len, new_len, &charged);
463 if (IS_ERR(vma)) {
464 ret = PTR_ERR(vma);
465 goto out;
466 }
467
468 map_flags = MAP_FIXED;
469 if (vma->vm_flags & VM_MAYSHARE)
470 map_flags |= MAP_SHARED;
471
472 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
473 ((addr - vma->vm_start) >> PAGE_SHIFT),
474 map_flags);
475 if (offset_in_page(ret))
476 goto out1;
477
478 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
479 uf_unmap);
480 if (!(offset_in_page(ret)))
481 goto out;
482 out1:
483 vm_unacct_memory(charged);
484
485 out:
486 return ret;
487 }
488
489 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
490 {
491 unsigned long end = vma->vm_end + delta;
492 if (end < vma->vm_end) /* overflow */
493 return 0;
494 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
495 return 0;
496 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
497 0, MAP_FIXED) & ~PAGE_MASK)
498 return 0;
499 return 1;
500 }
501
502 /*
503 * Expand (or shrink) an existing mapping, potentially moving it at the
504 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
505 *
506 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
507 * This option implies MREMAP_MAYMOVE.
508 */
509 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
510 unsigned long, new_len, unsigned long, flags,
511 unsigned long, new_addr)
512 {
513 struct mm_struct *mm = current->mm;
514 struct vm_area_struct *vma;
515 unsigned long ret = -EINVAL;
516 unsigned long charged = 0;
517 bool locked = false;
518 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
519 LIST_HEAD(uf_unmap_early);
520 LIST_HEAD(uf_unmap);
521
522 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
523 return ret;
524
525 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
526 return ret;
527
528 if (offset_in_page(addr))
529 return ret;
530
531 old_len = PAGE_ALIGN(old_len);
532 new_len = PAGE_ALIGN(new_len);
533
534 /*
535 * We allow a zero old-len as a special case
536 * for DOS-emu "duplicate shm area" thing. But
537 * a zero new-len is nonsensical.
538 */
539 if (!new_len)
540 return ret;
541
542 if (down_write_killable(&current->mm->mmap_sem))
543 return -EINTR;
544
545 if (flags & MREMAP_FIXED) {
546 ret = mremap_to(addr, old_len, new_addr, new_len,
547 &locked, &uf, &uf_unmap_early, &uf_unmap);
548 goto out;
549 }
550
551 /*
552 * Always allow a shrinking remap: that just unmaps
553 * the unnecessary pages..
554 * do_munmap does all the needed commit accounting
555 */
556 if (old_len >= new_len) {
557 ret = do_munmap(mm, addr+new_len, old_len - new_len, &uf_unmap);
558 if (ret && old_len != new_len)
559 goto out;
560 ret = addr;
561 goto out;
562 }
563
564 /*
565 * Ok, we need to grow..
566 */
567 vma = vma_to_resize(addr, old_len, new_len, &charged);
568 if (IS_ERR(vma)) {
569 ret = PTR_ERR(vma);
570 goto out;
571 }
572
573 /* old_len exactly to the end of the area..
574 */
575 if (old_len == vma->vm_end - addr) {
576 /* can we just expand the current mapping? */
577 if (vma_expandable(vma, new_len - old_len)) {
578 int pages = (new_len - old_len) >> PAGE_SHIFT;
579
580 if (vma_adjust(vma, vma->vm_start, addr + new_len,
581 vma->vm_pgoff, NULL)) {
582 ret = -ENOMEM;
583 goto out;
584 }
585
586 vm_stat_account(mm, vma->vm_flags, pages);
587 if (vma->vm_flags & VM_LOCKED) {
588 mm->locked_vm += pages;
589 locked = true;
590 new_addr = addr;
591 }
592 ret = addr;
593 goto out;
594 }
595 }
596
597 /*
598 * We weren't able to just expand or shrink the area,
599 * we need to create a new one and move it..
600 */
601 ret = -ENOMEM;
602 if (flags & MREMAP_MAYMOVE) {
603 unsigned long map_flags = 0;
604 if (vma->vm_flags & VM_MAYSHARE)
605 map_flags |= MAP_SHARED;
606
607 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
608 vma->vm_pgoff +
609 ((addr - vma->vm_start) >> PAGE_SHIFT),
610 map_flags);
611 if (offset_in_page(new_addr)) {
612 ret = new_addr;
613 goto out;
614 }
615
616 ret = move_vma(vma, addr, old_len, new_len, new_addr,
617 &locked, &uf, &uf_unmap);
618 }
619 out:
620 if (offset_in_page(ret)) {
621 vm_unacct_memory(charged);
622 locked = 0;
623 }
624 up_write(&current->mm->mmap_sem);
625 if (locked && new_len > old_len)
626 mm_populate(new_addr + old_len, new_len - old_len);
627 userfaultfd_unmap_complete(mm, &uf_unmap_early);
628 mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
629 userfaultfd_unmap_complete(mm, &uf_unmap);
630 return ret;
631 }