GNU Linux-libre 4.19.264-gnu1
[releases.git] / mm / mprotect.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  mm/mprotect.c
4  *
5  *  (C) Copyright 1994 Linus Torvalds
6  *  (C) Copyright 2002 Christoph Hellwig
7  *
8  *  Address space accounting code       <alan@lxorguk.ukuu.org.uk>
9  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
10  */
11
12 #include <linux/mm.h>
13 #include <linux/hugetlb.h>
14 #include <linux/shm.h>
15 #include <linux/mman.h>
16 #include <linux/fs.h>
17 #include <linux/highmem.h>
18 #include <linux/security.h>
19 #include <linux/mempolicy.h>
20 #include <linux/personality.h>
21 #include <linux/syscalls.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/mmu_notifier.h>
25 #include <linux/migrate.h>
26 #include <linux/perf_event.h>
27 #include <linux/pkeys.h>
28 #include <linux/ksm.h>
29 #include <linux/uaccess.h>
30 #include <linux/mm_inline.h>
31 #include <asm/pgtable.h>
32 #include <asm/cacheflush.h>
33 #include <asm/mmu_context.h>
34 #include <asm/tlbflush.h>
35
36 #include "internal.h"
37
38 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
39                 unsigned long addr, unsigned long end, pgprot_t newprot,
40                 int dirty_accountable, int prot_numa)
41 {
42         struct mm_struct *mm = vma->vm_mm;
43         pte_t *pte, oldpte;
44         spinlock_t *ptl;
45         unsigned long pages = 0;
46         int target_node = NUMA_NO_NODE;
47
48         /*
49          * Can be called with only the mmap_sem for reading by
50          * prot_numa so we must check the pmd isn't constantly
51          * changing from under us from pmd_none to pmd_trans_huge
52          * and/or the other way around.
53          */
54         if (pmd_trans_unstable(pmd))
55                 return 0;
56
57         /*
58          * The pmd points to a regular pte so the pmd can't change
59          * from under us even if the mmap_sem is only hold for
60          * reading.
61          */
62         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
63
64         /* Get target node for single threaded private VMAs */
65         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
66             atomic_read(&vma->vm_mm->mm_users) == 1)
67                 target_node = numa_node_id();
68
69         flush_tlb_batched_pending(vma->vm_mm);
70         arch_enter_lazy_mmu_mode();
71         do {
72                 oldpte = *pte;
73                 if (pte_present(oldpte)) {
74                         pte_t ptent;
75                         bool preserve_write = prot_numa && pte_write(oldpte);
76
77                         /*
78                          * Avoid trapping faults against the zero or KSM
79                          * pages. See similar comment in change_huge_pmd.
80                          */
81                         if (prot_numa) {
82                                 struct page *page;
83
84                                 page = vm_normal_page(vma, addr, oldpte);
85                                 if (!page || PageKsm(page))
86                                         continue;
87
88                                 /* Also skip shared copy-on-write pages */
89                                 if (is_cow_mapping(vma->vm_flags) &&
90                                     page_mapcount(page) != 1)
91                                         continue;
92
93                                 /*
94                                  * While migration can move some dirty pages,
95                                  * it cannot move them all from MIGRATE_ASYNC
96                                  * context.
97                                  */
98                                 if (page_is_file_cache(page) && PageDirty(page))
99                                         continue;
100
101                                 /* Avoid TLB flush if possible */
102                                 if (pte_protnone(oldpte))
103                                         continue;
104
105                                 /*
106                                  * Don't mess with PTEs if page is already on the node
107                                  * a single-threaded process is running on.
108                                  */
109                                 if (target_node == page_to_nid(page))
110                                         continue;
111                         }
112
113                         ptent = ptep_modify_prot_start(mm, addr, pte);
114                         ptent = pte_modify(ptent, newprot);
115                         if (preserve_write)
116                                 ptent = pte_mk_savedwrite(ptent);
117
118                         /* Avoid taking write faults for known dirty pages */
119                         if (dirty_accountable && pte_dirty(ptent) &&
120                                         (pte_soft_dirty(ptent) ||
121                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
122                                 ptent = pte_mkwrite(ptent);
123                         }
124                         ptep_modify_prot_commit(mm, addr, pte, ptent);
125                         pages++;
126                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
127                         swp_entry_t entry = pte_to_swp_entry(oldpte);
128
129                         if (is_write_migration_entry(entry)) {
130                                 pte_t newpte;
131                                 /*
132                                  * A protection check is difficult so
133                                  * just be safe and disable write
134                                  */
135                                 make_migration_entry_read(&entry);
136                                 newpte = swp_entry_to_pte(entry);
137                                 if (pte_swp_soft_dirty(oldpte))
138                                         newpte = pte_swp_mksoft_dirty(newpte);
139                                 set_pte_at(mm, addr, pte, newpte);
140
141                                 pages++;
142                         }
143
144                         if (is_write_device_private_entry(entry)) {
145                                 pte_t newpte;
146
147                                 /*
148                                  * We do not preserve soft-dirtiness. See
149                                  * copy_one_pte() for explanation.
150                                  */
151                                 make_device_private_entry_read(&entry);
152                                 newpte = swp_entry_to_pte(entry);
153                                 set_pte_at(mm, addr, pte, newpte);
154
155                                 pages++;
156                         }
157                 }
158         } while (pte++, addr += PAGE_SIZE, addr != end);
159         arch_leave_lazy_mmu_mode();
160         pte_unmap_unlock(pte - 1, ptl);
161
162         return pages;
163 }
164
165 /*
166  * Used when setting automatic NUMA hinting protection where it is
167  * critical that a numa hinting PMD is not confused with a bad PMD.
168  */
169 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
170 {
171         pmd_t pmdval = pmd_read_atomic(pmd);
172
173         /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
174 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
175         barrier();
176 #endif
177
178         if (pmd_none(pmdval))
179                 return 1;
180         if (pmd_trans_huge(pmdval))
181                 return 0;
182         if (unlikely(pmd_bad(pmdval))) {
183                 pmd_clear_bad(pmd);
184                 return 1;
185         }
186
187         return 0;
188 }
189
190 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
191                 pud_t *pud, unsigned long addr, unsigned long end,
192                 pgprot_t newprot, int dirty_accountable, int prot_numa)
193 {
194         pmd_t *pmd;
195         struct mm_struct *mm = vma->vm_mm;
196         unsigned long next;
197         unsigned long pages = 0;
198         unsigned long nr_huge_updates = 0;
199         unsigned long mni_start = 0;
200
201         pmd = pmd_offset(pud, addr);
202         do {
203                 unsigned long this_pages;
204
205                 next = pmd_addr_end(addr, end);
206
207                 /*
208                  * Automatic NUMA balancing walks the tables with mmap_sem
209                  * held for read. It's possible a parallel update to occur
210                  * between pmd_trans_huge() and a pmd_none_or_clear_bad()
211                  * check leading to a false positive and clearing.
212                  * Hence, it's necessary to atomically read the PMD value
213                  * for all the checks.
214                  */
215                 if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
216                      pmd_none_or_clear_bad_unless_trans_huge(pmd))
217                         goto next;
218
219                 /* invoke the mmu notifier if the pmd is populated */
220                 if (!mni_start) {
221                         mni_start = addr;
222                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
223                 }
224
225                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
226                         if (next - addr != HPAGE_PMD_SIZE) {
227                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
228                         } else {
229                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
230                                                 newprot, prot_numa);
231
232                                 if (nr_ptes) {
233                                         if (nr_ptes == HPAGE_PMD_NR) {
234                                                 pages += HPAGE_PMD_NR;
235                                                 nr_huge_updates++;
236                                         }
237
238                                         /* huge pmd was handled */
239                                         goto next;
240                                 }
241                         }
242                         /* fall through, the trans huge pmd just split */
243                 }
244                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
245                                  dirty_accountable, prot_numa);
246                 pages += this_pages;
247 next:
248                 cond_resched();
249         } while (pmd++, addr = next, addr != end);
250
251         if (mni_start)
252                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
253
254         if (nr_huge_updates)
255                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
256         return pages;
257 }
258
259 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
260                 p4d_t *p4d, unsigned long addr, unsigned long end,
261                 pgprot_t newprot, int dirty_accountable, int prot_numa)
262 {
263         pud_t *pud;
264         unsigned long next;
265         unsigned long pages = 0;
266
267         pud = pud_offset(p4d, addr);
268         do {
269                 next = pud_addr_end(addr, end);
270                 if (pud_none_or_clear_bad(pud))
271                         continue;
272                 pages += change_pmd_range(vma, pud, addr, next, newprot,
273                                  dirty_accountable, prot_numa);
274         } while (pud++, addr = next, addr != end);
275
276         return pages;
277 }
278
279 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
280                 pgd_t *pgd, unsigned long addr, unsigned long end,
281                 pgprot_t newprot, int dirty_accountable, int prot_numa)
282 {
283         p4d_t *p4d;
284         unsigned long next;
285         unsigned long pages = 0;
286
287         p4d = p4d_offset(pgd, addr);
288         do {
289                 next = p4d_addr_end(addr, end);
290                 if (p4d_none_or_clear_bad(p4d))
291                         continue;
292                 pages += change_pud_range(vma, p4d, addr, next, newprot,
293                                  dirty_accountable, prot_numa);
294         } while (p4d++, addr = next, addr != end);
295
296         return pages;
297 }
298
299 static unsigned long change_protection_range(struct vm_area_struct *vma,
300                 unsigned long addr, unsigned long end, pgprot_t newprot,
301                 int dirty_accountable, int prot_numa)
302 {
303         struct mm_struct *mm = vma->vm_mm;
304         pgd_t *pgd;
305         unsigned long next;
306         unsigned long start = addr;
307         unsigned long pages = 0;
308
309         BUG_ON(addr >= end);
310         pgd = pgd_offset(mm, addr);
311         flush_cache_range(vma, addr, end);
312         inc_tlb_flush_pending(mm);
313         do {
314                 next = pgd_addr_end(addr, end);
315                 if (pgd_none_or_clear_bad(pgd))
316                         continue;
317                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
318                                  dirty_accountable, prot_numa);
319         } while (pgd++, addr = next, addr != end);
320
321         /* Only flush the TLB if we actually modified any entries: */
322         if (pages)
323                 flush_tlb_range(vma, start, end);
324         dec_tlb_flush_pending(mm);
325
326         return pages;
327 }
328
329 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
330                        unsigned long end, pgprot_t newprot,
331                        int dirty_accountable, int prot_numa)
332 {
333         unsigned long pages;
334
335         if (is_vm_hugetlb_page(vma))
336                 pages = hugetlb_change_protection(vma, start, end, newprot);
337         else
338                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
339
340         return pages;
341 }
342
343 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
344                                unsigned long next, struct mm_walk *walk)
345 {
346         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
347                 0 : -EACCES;
348 }
349
350 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
351                                    unsigned long addr, unsigned long next,
352                                    struct mm_walk *walk)
353 {
354         return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
355                 0 : -EACCES;
356 }
357
358 static int prot_none_test(unsigned long addr, unsigned long next,
359                           struct mm_walk *walk)
360 {
361         return 0;
362 }
363
364 static int prot_none_walk(struct vm_area_struct *vma, unsigned long start,
365                            unsigned long end, unsigned long newflags)
366 {
367         pgprot_t new_pgprot = vm_get_page_prot(newflags);
368         struct mm_walk prot_none_walk = {
369                 .pte_entry = prot_none_pte_entry,
370                 .hugetlb_entry = prot_none_hugetlb_entry,
371                 .test_walk = prot_none_test,
372                 .mm = current->mm,
373                 .private = &new_pgprot,
374         };
375
376         return walk_page_range(start, end, &prot_none_walk);
377 }
378
379 int
380 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
381         unsigned long start, unsigned long end, unsigned long newflags)
382 {
383         struct mm_struct *mm = vma->vm_mm;
384         unsigned long oldflags = vma->vm_flags;
385         long nrpages = (end - start) >> PAGE_SHIFT;
386         unsigned long charged = 0;
387         pgoff_t pgoff;
388         int error;
389         int dirty_accountable = 0;
390
391         if (newflags == oldflags) {
392                 *pprev = vma;
393                 return 0;
394         }
395
396         /*
397          * Do PROT_NONE PFN permission checks here when we can still
398          * bail out without undoing a lot of state. This is a rather
399          * uncommon case, so doesn't need to be very optimized.
400          */
401         if (arch_has_pfn_modify_check() &&
402             (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
403             (newflags & (VM_READ|VM_WRITE|VM_EXEC)) == 0) {
404                 error = prot_none_walk(vma, start, end, newflags);
405                 if (error)
406                         return error;
407         }
408
409         /*
410          * If we make a private mapping writable we increase our commit;
411          * but (without finer accounting) cannot reduce our commit if we
412          * make it unwritable again. hugetlb mapping were accounted for
413          * even if read-only so there is no need to account for them here
414          */
415         if (newflags & VM_WRITE) {
416                 /* Check space limits when area turns into data. */
417                 if (!may_expand_vm(mm, newflags, nrpages) &&
418                                 may_expand_vm(mm, oldflags, nrpages))
419                         return -ENOMEM;
420                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
421                                                 VM_SHARED|VM_NORESERVE))) {
422                         charged = nrpages;
423                         if (security_vm_enough_memory_mm(mm, charged))
424                                 return -ENOMEM;
425                         newflags |= VM_ACCOUNT;
426                 }
427         }
428
429         /*
430          * First try to merge with previous and/or next vma.
431          */
432         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
433         *pprev = vma_merge(mm, *pprev, start, end, newflags,
434                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
435                            vma->vm_userfaultfd_ctx);
436         if (*pprev) {
437                 vma = *pprev;
438                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
439                 goto success;
440         }
441
442         *pprev = vma;
443
444         if (start != vma->vm_start) {
445                 error = split_vma(mm, vma, start, 1);
446                 if (error)
447                         goto fail;
448         }
449
450         if (end != vma->vm_end) {
451                 error = split_vma(mm, vma, end, 0);
452                 if (error)
453                         goto fail;
454         }
455
456 success:
457         /*
458          * vm_flags and vm_page_prot are protected by the mmap_sem
459          * held in write mode.
460          */
461         vma->vm_flags = newflags;
462         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
463         vma_set_page_prot(vma);
464
465         change_protection(vma, start, end, vma->vm_page_prot,
466                           dirty_accountable, 0);
467
468         /*
469          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
470          * fault on access.
471          */
472         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
473                         (newflags & VM_WRITE)) {
474                 populate_vma_page_range(vma, start, end, NULL);
475         }
476
477         vm_stat_account(mm, oldflags, -nrpages);
478         vm_stat_account(mm, newflags, nrpages);
479         perf_event_mmap(vma);
480         return 0;
481
482 fail:
483         vm_unacct_memory(charged);
484         return error;
485 }
486
487 /*
488  * pkey==-1 when doing a legacy mprotect()
489  */
490 static int do_mprotect_pkey(unsigned long start, size_t len,
491                 unsigned long prot, int pkey)
492 {
493         unsigned long nstart, end, tmp, reqprot;
494         struct vm_area_struct *vma, *prev;
495         int error = -EINVAL;
496         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
497         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
498                                 (prot & PROT_READ);
499
500         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
501         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
502                 return -EINVAL;
503
504         if (start & ~PAGE_MASK)
505                 return -EINVAL;
506         if (!len)
507                 return 0;
508         len = PAGE_ALIGN(len);
509         end = start + len;
510         if (end <= start)
511                 return -ENOMEM;
512         if (!arch_validate_prot(prot, start))
513                 return -EINVAL;
514
515         reqprot = prot;
516
517         if (down_write_killable(&current->mm->mmap_sem))
518                 return -EINTR;
519
520         /*
521          * If userspace did not allocate the pkey, do not let
522          * them use it here.
523          */
524         error = -EINVAL;
525         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
526                 goto out;
527
528         vma = find_vma(current->mm, start);
529         error = -ENOMEM;
530         if (!vma)
531                 goto out;
532         prev = vma->vm_prev;
533         if (unlikely(grows & PROT_GROWSDOWN)) {
534                 if (vma->vm_start >= end)
535                         goto out;
536                 start = vma->vm_start;
537                 error = -EINVAL;
538                 if (!(vma->vm_flags & VM_GROWSDOWN))
539                         goto out;
540         } else {
541                 if (vma->vm_start > start)
542                         goto out;
543                 if (unlikely(grows & PROT_GROWSUP)) {
544                         end = vma->vm_end;
545                         error = -EINVAL;
546                         if (!(vma->vm_flags & VM_GROWSUP))
547                                 goto out;
548                 }
549         }
550         if (start > vma->vm_start)
551                 prev = vma;
552
553         for (nstart = start ; ; ) {
554                 unsigned long mask_off_old_flags;
555                 unsigned long newflags;
556                 int new_vma_pkey;
557
558                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
559
560                 /* Does the application expect PROT_READ to imply PROT_EXEC */
561                 if (rier && (vma->vm_flags & VM_MAYEXEC))
562                         prot |= PROT_EXEC;
563
564                 /*
565                  * Each mprotect() call explicitly passes r/w/x permissions.
566                  * If a permission is not passed to mprotect(), it must be
567                  * cleared from the VMA.
568                  */
569                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
570                                         VM_FLAGS_CLEAR;
571
572                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
573                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
574                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
575
576                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
577                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
578                         error = -EACCES;
579                         goto out;
580                 }
581
582                 error = security_file_mprotect(vma, reqprot, prot);
583                 if (error)
584                         goto out;
585
586                 tmp = vma->vm_end;
587                 if (tmp > end)
588                         tmp = end;
589                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
590                 if (error)
591                         goto out;
592                 nstart = tmp;
593
594                 if (nstart < prev->vm_end)
595                         nstart = prev->vm_end;
596                 if (nstart >= end)
597                         goto out;
598
599                 vma = prev->vm_next;
600                 if (!vma || vma->vm_start != nstart) {
601                         error = -ENOMEM;
602                         goto out;
603                 }
604                 prot = reqprot;
605         }
606 out:
607         up_write(&current->mm->mmap_sem);
608         return error;
609 }
610
611 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
612                 unsigned long, prot)
613 {
614         return do_mprotect_pkey(start, len, prot, -1);
615 }
616
617 #ifdef CONFIG_ARCH_HAS_PKEYS
618
619 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
620                 unsigned long, prot, int, pkey)
621 {
622         return do_mprotect_pkey(start, len, prot, pkey);
623 }
624
625 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
626 {
627         int pkey;
628         int ret;
629
630         /* No flags supported yet. */
631         if (flags)
632                 return -EINVAL;
633         /* check for unsupported init values */
634         if (init_val & ~PKEY_ACCESS_MASK)
635                 return -EINVAL;
636
637         down_write(&current->mm->mmap_sem);
638         pkey = mm_pkey_alloc(current->mm);
639
640         ret = -ENOSPC;
641         if (pkey == -1)
642                 goto out;
643
644         ret = arch_set_user_pkey_access(current, pkey, init_val);
645         if (ret) {
646                 mm_pkey_free(current->mm, pkey);
647                 goto out;
648         }
649         ret = pkey;
650 out:
651         up_write(&current->mm->mmap_sem);
652         return ret;
653 }
654
655 SYSCALL_DEFINE1(pkey_free, int, pkey)
656 {
657         int ret;
658
659         down_write(&current->mm->mmap_sem);
660         ret = mm_pkey_free(current->mm, pkey);
661         up_write(&current->mm->mmap_sem);
662
663         /*
664          * We could provie warnings or errors if any VMA still
665          * has the pkey set here.
666          */
667         return ret;
668 }
669
670 #endif /* CONFIG_ARCH_HAS_PKEYS */