Merge branch 'for-4.15/upstream' into for-linus
[platform/kernel/linux-starfive.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 (is_swap_pmd(*old_pmd) || 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         /*
388          * !old_len is a special case where an attempt is made to 'duplicate'
389          * a mapping.  This makes no sense for private mappings as it will
390          * instead create a fresh/new mapping unrelated to the original.  This
391          * is contrary to the basic idea of mremap which creates new mappings
392          * based on the original.  There are no known use cases for this
393          * behavior.  As a result, fail such attempts.
394          */
395         if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
396                 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
397                 return ERR_PTR(-EINVAL);
398         }
399
400         if (is_vm_hugetlb_page(vma))
401                 return ERR_PTR(-EINVAL);
402
403         /* We can't remap across vm area boundaries */
404         if (old_len > vma->vm_end - addr)
405                 return ERR_PTR(-EFAULT);
406
407         if (new_len == old_len)
408                 return vma;
409
410         /* Need to be careful about a growing mapping */
411         pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
412         pgoff += vma->vm_pgoff;
413         if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
414                 return ERR_PTR(-EINVAL);
415
416         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
417                 return ERR_PTR(-EFAULT);
418
419         if (vma->vm_flags & VM_LOCKED) {
420                 unsigned long locked, lock_limit;
421                 locked = mm->locked_vm << PAGE_SHIFT;
422                 lock_limit = rlimit(RLIMIT_MEMLOCK);
423                 locked += new_len - old_len;
424                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
425                         return ERR_PTR(-EAGAIN);
426         }
427
428         if (!may_expand_vm(mm, vma->vm_flags,
429                                 (new_len - old_len) >> PAGE_SHIFT))
430                 return ERR_PTR(-ENOMEM);
431
432         if (vma->vm_flags & VM_ACCOUNT) {
433                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
434                 if (security_vm_enough_memory_mm(mm, charged))
435                         return ERR_PTR(-ENOMEM);
436                 *p = charged;
437         }
438
439         return vma;
440 }
441
442 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
443                 unsigned long new_addr, unsigned long new_len, bool *locked,
444                 struct vm_userfaultfd_ctx *uf,
445                 struct list_head *uf_unmap_early,
446                 struct list_head *uf_unmap)
447 {
448         struct mm_struct *mm = current->mm;
449         struct vm_area_struct *vma;
450         unsigned long ret = -EINVAL;
451         unsigned long charged = 0;
452         unsigned long map_flags;
453
454         if (offset_in_page(new_addr))
455                 goto out;
456
457         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
458                 goto out;
459
460         /* Ensure the old/new locations do not overlap */
461         if (addr + old_len > new_addr && new_addr + new_len > addr)
462                 goto out;
463
464         ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
465         if (ret)
466                 goto out;
467
468         if (old_len >= new_len) {
469                 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
470                 if (ret && old_len != new_len)
471                         goto out;
472                 old_len = new_len;
473         }
474
475         vma = vma_to_resize(addr, old_len, new_len, &charged);
476         if (IS_ERR(vma)) {
477                 ret = PTR_ERR(vma);
478                 goto out;
479         }
480
481         map_flags = MAP_FIXED;
482         if (vma->vm_flags & VM_MAYSHARE)
483                 map_flags |= MAP_SHARED;
484
485         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
486                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
487                                 map_flags);
488         if (offset_in_page(ret))
489                 goto out1;
490
491         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
492                        uf_unmap);
493         if (!(offset_in_page(ret)))
494                 goto out;
495 out1:
496         vm_unacct_memory(charged);
497
498 out:
499         return ret;
500 }
501
502 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
503 {
504         unsigned long end = vma->vm_end + delta;
505         if (end < vma->vm_end) /* overflow */
506                 return 0;
507         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
508                 return 0;
509         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
510                               0, MAP_FIXED) & ~PAGE_MASK)
511                 return 0;
512         return 1;
513 }
514
515 /*
516  * Expand (or shrink) an existing mapping, potentially moving it at the
517  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
518  *
519  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
520  * This option implies MREMAP_MAYMOVE.
521  */
522 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
523                 unsigned long, new_len, unsigned long, flags,
524                 unsigned long, new_addr)
525 {
526         struct mm_struct *mm = current->mm;
527         struct vm_area_struct *vma;
528         unsigned long ret = -EINVAL;
529         unsigned long charged = 0;
530         bool locked = false;
531         struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
532         LIST_HEAD(uf_unmap_early);
533         LIST_HEAD(uf_unmap);
534
535         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
536                 return ret;
537
538         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
539                 return ret;
540
541         if (offset_in_page(addr))
542                 return ret;
543
544         old_len = PAGE_ALIGN(old_len);
545         new_len = PAGE_ALIGN(new_len);
546
547         /*
548          * We allow a zero old-len as a special case
549          * for DOS-emu "duplicate shm area" thing. But
550          * a zero new-len is nonsensical.
551          */
552         if (!new_len)
553                 return ret;
554
555         if (down_write_killable(&current->mm->mmap_sem))
556                 return -EINTR;
557
558         if (flags & MREMAP_FIXED) {
559                 ret = mremap_to(addr, old_len, new_addr, new_len,
560                                 &locked, &uf, &uf_unmap_early, &uf_unmap);
561                 goto out;
562         }
563
564         /*
565          * Always allow a shrinking remap: that just unmaps
566          * the unnecessary pages..
567          * do_munmap does all the needed commit accounting
568          */
569         if (old_len >= new_len) {
570                 ret = do_munmap(mm, addr+new_len, old_len - new_len, &uf_unmap);
571                 if (ret && old_len != new_len)
572                         goto out;
573                 ret = addr;
574                 goto out;
575         }
576
577         /*
578          * Ok, we need to grow..
579          */
580         vma = vma_to_resize(addr, old_len, new_len, &charged);
581         if (IS_ERR(vma)) {
582                 ret = PTR_ERR(vma);
583                 goto out;
584         }
585
586         /* old_len exactly to the end of the area..
587          */
588         if (old_len == vma->vm_end - addr) {
589                 /* can we just expand the current mapping? */
590                 if (vma_expandable(vma, new_len - old_len)) {
591                         int pages = (new_len - old_len) >> PAGE_SHIFT;
592
593                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
594                                        vma->vm_pgoff, NULL)) {
595                                 ret = -ENOMEM;
596                                 goto out;
597                         }
598
599                         vm_stat_account(mm, vma->vm_flags, pages);
600                         if (vma->vm_flags & VM_LOCKED) {
601                                 mm->locked_vm += pages;
602                                 locked = true;
603                                 new_addr = addr;
604                         }
605                         ret = addr;
606                         goto out;
607                 }
608         }
609
610         /*
611          * We weren't able to just expand or shrink the area,
612          * we need to create a new one and move it..
613          */
614         ret = -ENOMEM;
615         if (flags & MREMAP_MAYMOVE) {
616                 unsigned long map_flags = 0;
617                 if (vma->vm_flags & VM_MAYSHARE)
618                         map_flags |= MAP_SHARED;
619
620                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
621                                         vma->vm_pgoff +
622                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
623                                         map_flags);
624                 if (offset_in_page(new_addr)) {
625                         ret = new_addr;
626                         goto out;
627                 }
628
629                 ret = move_vma(vma, addr, old_len, new_len, new_addr,
630                                &locked, &uf, &uf_unmap);
631         }
632 out:
633         if (offset_in_page(ret)) {
634                 vm_unacct_memory(charged);
635                 locked = 0;
636         }
637         up_write(&current->mm->mmap_sem);
638         if (locked && new_len > old_len)
639                 mm_populate(new_addr + old_len, new_len - old_len);
640         userfaultfd_unmap_complete(mm, &uf_unmap_early);
641         mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
642         userfaultfd_unmap_complete(mm, &uf_unmap);
643         return ret;
644 }