xref: /linux/mm/mprotect.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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/pagewalk.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 <linux/pgtable.h>
32 #include <linux/sched/sysctl.h>
33 #include <linux/userfaultfd_k.h>
34 #include <linux/memory-tiers.h>
35 #include <uapi/linux/mman.h>
36 #include <asm/cacheflush.h>
37 #include <asm/mmu_context.h>
38 #include <asm/tlbflush.h>
39 #include <asm/tlb.h>
40 
41 #include "internal.h"
42 
43 bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
44 			     pte_t pte)
45 {
46 	struct page *page;
47 
48 	if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
49 		return false;
50 
51 	/* Don't touch entries that are not even readable. */
52 	if (pte_protnone(pte))
53 		return false;
54 
55 	/* Do we need write faults for softdirty tracking? */
56 	if (pte_needs_soft_dirty_wp(vma, pte))
57 		return false;
58 
59 	/* Do we need write faults for uffd-wp tracking? */
60 	if (userfaultfd_pte_wp(vma, pte))
61 		return false;
62 
63 	if (!(vma->vm_flags & VM_SHARED)) {
64 		/*
65 		 * Writable MAP_PRIVATE mapping: We can only special-case on
66 		 * exclusive anonymous pages, because we know that our
67 		 * write-fault handler similarly would map them writable without
68 		 * any additional checks while holding the PT lock.
69 		 */
70 		page = vm_normal_page(vma, addr, pte);
71 		return page && PageAnon(page) && PageAnonExclusive(page);
72 	}
73 
74 	VM_WARN_ON_ONCE(is_zero_pfn(pte_pfn(pte)) && pte_dirty(pte));
75 
76 	/*
77 	 * Writable MAP_SHARED mapping: "clean" might indicate that the FS still
78 	 * needs a real write-fault for writenotify
79 	 * (see vma_wants_writenotify()). If "dirty", the assumption is that the
80 	 * FS was already notified and we can simply mark the PTE writable
81 	 * just like the write-fault handler would do.
82 	 */
83 	return pte_dirty(pte);
84 }
85 
86 static long change_pte_range(struct mmu_gather *tlb,
87 		struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
88 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
89 {
90 	pte_t *pte, oldpte;
91 	spinlock_t *ptl;
92 	long pages = 0;
93 	int target_node = NUMA_NO_NODE;
94 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
95 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
96 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
97 
98 	tlb_change_page_size(tlb, PAGE_SIZE);
99 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
100 	if (!pte)
101 		return -EAGAIN;
102 
103 	/* Get target node for single threaded private VMAs */
104 	if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
105 	    atomic_read(&vma->vm_mm->mm_users) == 1)
106 		target_node = numa_node_id();
107 
108 	flush_tlb_batched_pending(vma->vm_mm);
109 	arch_enter_lazy_mmu_mode();
110 	do {
111 		oldpte = ptep_get(pte);
112 		if (pte_present(oldpte)) {
113 			pte_t ptent;
114 
115 			/*
116 			 * Avoid trapping faults against the zero or KSM
117 			 * pages. See similar comment in change_huge_pmd.
118 			 */
119 			if (prot_numa) {
120 				struct folio *folio;
121 				int nid;
122 				bool toptier;
123 
124 				/* Avoid TLB flush if possible */
125 				if (pte_protnone(oldpte))
126 					continue;
127 
128 				folio = vm_normal_folio(vma, addr, oldpte);
129 				if (!folio || folio_is_zone_device(folio) ||
130 				    folio_test_ksm(folio))
131 					continue;
132 
133 				/* Also skip shared copy-on-write pages */
134 				if (is_cow_mapping(vma->vm_flags) &&
135 				    (folio_maybe_dma_pinned(folio) ||
136 				     folio_likely_mapped_shared(folio)))
137 					continue;
138 
139 				/*
140 				 * While migration can move some dirty pages,
141 				 * it cannot move them all from MIGRATE_ASYNC
142 				 * context.
143 				 */
144 				if (folio_is_file_lru(folio) &&
145 				    folio_test_dirty(folio))
146 					continue;
147 
148 				/*
149 				 * Don't mess with PTEs if page is already on the node
150 				 * a single-threaded process is running on.
151 				 */
152 				nid = folio_nid(folio);
153 				if (target_node == nid)
154 					continue;
155 				toptier = node_is_toptier(nid);
156 
157 				/*
158 				 * Skip scanning top tier node if normal numa
159 				 * balancing is disabled
160 				 */
161 				if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
162 				    toptier)
163 					continue;
164 				if (folio_use_access_time(folio))
165 					folio_xchg_access_time(folio,
166 						jiffies_to_msecs(jiffies));
167 			}
168 
169 			oldpte = ptep_modify_prot_start(vma, addr, pte);
170 			ptent = pte_modify(oldpte, newprot);
171 
172 			if (uffd_wp)
173 				ptent = pte_mkuffd_wp(ptent);
174 			else if (uffd_wp_resolve)
175 				ptent = pte_clear_uffd_wp(ptent);
176 
177 			/*
178 			 * In some writable, shared mappings, we might want
179 			 * to catch actual write access -- see
180 			 * vma_wants_writenotify().
181 			 *
182 			 * In all writable, private mappings, we have to
183 			 * properly handle COW.
184 			 *
185 			 * In both cases, we can sometimes still change PTEs
186 			 * writable and avoid the write-fault handler, for
187 			 * example, if a PTE is already dirty and no other
188 			 * COW or special handling is required.
189 			 */
190 			if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
191 			    !pte_write(ptent) &&
192 			    can_change_pte_writable(vma, addr, ptent))
193 				ptent = pte_mkwrite(ptent, vma);
194 
195 			ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
196 			if (pte_needs_flush(oldpte, ptent))
197 				tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
198 			pages++;
199 		} else if (is_swap_pte(oldpte)) {
200 			swp_entry_t entry = pte_to_swp_entry(oldpte);
201 			pte_t newpte;
202 
203 			if (is_writable_migration_entry(entry)) {
204 				struct folio *folio = pfn_swap_entry_folio(entry);
205 
206 				/*
207 				 * A protection check is difficult so
208 				 * just be safe and disable write
209 				 */
210 				if (folio_test_anon(folio))
211 					entry = make_readable_exclusive_migration_entry(
212 							     swp_offset(entry));
213 				else
214 					entry = make_readable_migration_entry(swp_offset(entry));
215 				newpte = swp_entry_to_pte(entry);
216 				if (pte_swp_soft_dirty(oldpte))
217 					newpte = pte_swp_mksoft_dirty(newpte);
218 			} else if (is_writable_device_private_entry(entry)) {
219 				/*
220 				 * We do not preserve soft-dirtiness. See
221 				 * copy_nonpresent_pte() for explanation.
222 				 */
223 				entry = make_readable_device_private_entry(
224 							swp_offset(entry));
225 				newpte = swp_entry_to_pte(entry);
226 				if (pte_swp_uffd_wp(oldpte))
227 					newpte = pte_swp_mkuffd_wp(newpte);
228 			} else if (is_writable_device_exclusive_entry(entry)) {
229 				entry = make_readable_device_exclusive_entry(
230 							swp_offset(entry));
231 				newpte = swp_entry_to_pte(entry);
232 				if (pte_swp_soft_dirty(oldpte))
233 					newpte = pte_swp_mksoft_dirty(newpte);
234 				if (pte_swp_uffd_wp(oldpte))
235 					newpte = pte_swp_mkuffd_wp(newpte);
236 			} else if (is_pte_marker_entry(entry)) {
237 				/*
238 				 * Ignore error swap entries unconditionally,
239 				 * because any access should sigbus/sigsegv
240 				 * anyway.
241 				 */
242 				if (is_poisoned_swp_entry(entry) ||
243 				    is_guard_swp_entry(entry))
244 					continue;
245 				/*
246 				 * If this is uffd-wp pte marker and we'd like
247 				 * to unprotect it, drop it; the next page
248 				 * fault will trigger without uffd trapping.
249 				 */
250 				if (uffd_wp_resolve) {
251 					pte_clear(vma->vm_mm, addr, pte);
252 					pages++;
253 				}
254 				continue;
255 			} else {
256 				newpte = oldpte;
257 			}
258 
259 			if (uffd_wp)
260 				newpte = pte_swp_mkuffd_wp(newpte);
261 			else if (uffd_wp_resolve)
262 				newpte = pte_swp_clear_uffd_wp(newpte);
263 
264 			if (!pte_same(oldpte, newpte)) {
265 				set_pte_at(vma->vm_mm, addr, pte, newpte);
266 				pages++;
267 			}
268 		} else {
269 			/* It must be an none page, or what else?.. */
270 			WARN_ON_ONCE(!pte_none(oldpte));
271 
272 			/*
273 			 * Nobody plays with any none ptes besides
274 			 * userfaultfd when applying the protections.
275 			 */
276 			if (likely(!uffd_wp))
277 				continue;
278 
279 			if (userfaultfd_wp_use_markers(vma)) {
280 				/*
281 				 * For file-backed mem, we need to be able to
282 				 * wr-protect a none pte, because even if the
283 				 * pte is none, the page/swap cache could
284 				 * exist.  Doing that by install a marker.
285 				 */
286 				set_pte_at(vma->vm_mm, addr, pte,
287 					   make_pte_marker(PTE_MARKER_UFFD_WP));
288 				pages++;
289 			}
290 		}
291 	} while (pte++, addr += PAGE_SIZE, addr != end);
292 	arch_leave_lazy_mmu_mode();
293 	pte_unmap_unlock(pte - 1, ptl);
294 
295 	return pages;
296 }
297 
298 /*
299  * Return true if we want to split THPs into PTE mappings in change
300  * protection procedure, false otherwise.
301  */
302 static inline bool
303 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags)
304 {
305 	/*
306 	 * pte markers only resides in pte level, if we need pte markers,
307 	 * we need to split.  For example, we cannot wr-protect a file thp
308 	 * (e.g. 2M shmem) because file thp is handled differently when
309 	 * split by erasing the pmd so far.
310 	 */
311 	return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
312 }
313 
314 /*
315  * Return true if we want to populate pgtables in change protection
316  * procedure, false otherwise
317  */
318 static inline bool
319 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags)
320 {
321 	/* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */
322 	if (!(cp_flags & MM_CP_UFFD_WP))
323 		return false;
324 
325 	/* Populate if the userfaultfd mode requires pte markers */
326 	return userfaultfd_wp_use_markers(vma);
327 }
328 
329 /*
330  * Populate the pgtable underneath for whatever reason if requested.
331  * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable
332  * allocation failures during page faults by kicking OOM and returning
333  * error.
334  */
335 #define  change_pmd_prepare(vma, pmd, cp_flags)				\
336 	({								\
337 		long err = 0;						\
338 		if (unlikely(pgtable_populate_needed(vma, cp_flags))) {	\
339 			if (pte_alloc(vma->vm_mm, pmd))			\
340 				err = -ENOMEM;				\
341 		}							\
342 		err;							\
343 	})
344 
345 /*
346  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
347  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
348  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
349  */
350 #define  change_prepare(vma, high, low, addr, cp_flags)			\
351 	  ({								\
352 		long err = 0;						\
353 		if (unlikely(pgtable_populate_needed(vma, cp_flags))) {	\
354 			low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
355 			if (p == NULL)					\
356 				err = -ENOMEM;				\
357 		}							\
358 		err;							\
359 	})
360 
361 static inline long change_pmd_range(struct mmu_gather *tlb,
362 		struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
363 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
364 {
365 	pmd_t *pmd;
366 	unsigned long next;
367 	long pages = 0;
368 	unsigned long nr_huge_updates = 0;
369 
370 	pmd = pmd_offset(pud, addr);
371 	do {
372 		long ret;
373 		pmd_t _pmd;
374 again:
375 		next = pmd_addr_end(addr, end);
376 
377 		ret = change_pmd_prepare(vma, pmd, cp_flags);
378 		if (ret) {
379 			pages = ret;
380 			break;
381 		}
382 
383 		if (pmd_none(*pmd))
384 			goto next;
385 
386 		_pmd = pmdp_get_lockless(pmd);
387 		if (is_swap_pmd(_pmd) || pmd_trans_huge(_pmd) || pmd_devmap(_pmd)) {
388 			if ((next - addr != HPAGE_PMD_SIZE) ||
389 			    pgtable_split_needed(vma, cp_flags)) {
390 				__split_huge_pmd(vma, pmd, addr, false, NULL);
391 				/*
392 				 * For file-backed, the pmd could have been
393 				 * cleared; make sure pmd populated if
394 				 * necessary, then fall-through to pte level.
395 				 */
396 				ret = change_pmd_prepare(vma, pmd, cp_flags);
397 				if (ret) {
398 					pages = ret;
399 					break;
400 				}
401 			} else {
402 				ret = change_huge_pmd(tlb, vma, pmd,
403 						addr, newprot, cp_flags);
404 				if (ret) {
405 					if (ret == HPAGE_PMD_NR) {
406 						pages += HPAGE_PMD_NR;
407 						nr_huge_updates++;
408 					}
409 
410 					/* huge pmd was handled */
411 					goto next;
412 				}
413 			}
414 			/* fall through, the trans huge pmd just split */
415 		}
416 
417 		ret = change_pte_range(tlb, vma, pmd, addr, next, newprot,
418 				       cp_flags);
419 		if (ret < 0)
420 			goto again;
421 		pages += ret;
422 next:
423 		cond_resched();
424 	} while (pmd++, addr = next, addr != end);
425 
426 	if (nr_huge_updates)
427 		count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
428 	return pages;
429 }
430 
431 static inline long change_pud_range(struct mmu_gather *tlb,
432 		struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
433 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
434 {
435 	struct mmu_notifier_range range;
436 	pud_t *pudp, pud;
437 	unsigned long next;
438 	long pages = 0, ret;
439 
440 	range.start = 0;
441 
442 	pudp = pud_offset(p4d, addr);
443 	do {
444 again:
445 		next = pud_addr_end(addr, end);
446 		ret = change_prepare(vma, pudp, pmd, addr, cp_flags);
447 		if (ret) {
448 			pages = ret;
449 			break;
450 		}
451 
452 		pud = READ_ONCE(*pudp);
453 		if (pud_none(pud))
454 			continue;
455 
456 		if (!range.start) {
457 			mmu_notifier_range_init(&range,
458 						MMU_NOTIFY_PROTECTION_VMA, 0,
459 						vma->vm_mm, addr, end);
460 			mmu_notifier_invalidate_range_start(&range);
461 		}
462 
463 		if (pud_leaf(pud)) {
464 			if ((next - addr != PUD_SIZE) ||
465 			    pgtable_split_needed(vma, cp_flags)) {
466 				__split_huge_pud(vma, pudp, addr);
467 				goto again;
468 			} else {
469 				ret = change_huge_pud(tlb, vma, pudp,
470 						      addr, newprot, cp_flags);
471 				if (ret == 0)
472 					goto again;
473 				/* huge pud was handled */
474 				if (ret == HPAGE_PUD_NR)
475 					pages += HPAGE_PUD_NR;
476 				continue;
477 			}
478 		}
479 
480 		pages += change_pmd_range(tlb, vma, pudp, addr, next, newprot,
481 					  cp_flags);
482 	} while (pudp++, addr = next, addr != end);
483 
484 	if (range.start)
485 		mmu_notifier_invalidate_range_end(&range);
486 
487 	return pages;
488 }
489 
490 static inline long change_p4d_range(struct mmu_gather *tlb,
491 		struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
492 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
493 {
494 	p4d_t *p4d;
495 	unsigned long next;
496 	long pages = 0, ret;
497 
498 	p4d = p4d_offset(pgd, addr);
499 	do {
500 		next = p4d_addr_end(addr, end);
501 		ret = change_prepare(vma, p4d, pud, addr, cp_flags);
502 		if (ret)
503 			return ret;
504 		if (p4d_none_or_clear_bad(p4d))
505 			continue;
506 		pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
507 					  cp_flags);
508 	} while (p4d++, addr = next, addr != end);
509 
510 	return pages;
511 }
512 
513 static long change_protection_range(struct mmu_gather *tlb,
514 		struct vm_area_struct *vma, unsigned long addr,
515 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
516 {
517 	struct mm_struct *mm = vma->vm_mm;
518 	pgd_t *pgd;
519 	unsigned long next;
520 	long pages = 0, ret;
521 
522 	BUG_ON(addr >= end);
523 	pgd = pgd_offset(mm, addr);
524 	tlb_start_vma(tlb, vma);
525 	do {
526 		next = pgd_addr_end(addr, end);
527 		ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
528 		if (ret) {
529 			pages = ret;
530 			break;
531 		}
532 		if (pgd_none_or_clear_bad(pgd))
533 			continue;
534 		pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
535 					  cp_flags);
536 	} while (pgd++, addr = next, addr != end);
537 
538 	tlb_end_vma(tlb, vma);
539 
540 	return pages;
541 }
542 
543 long change_protection(struct mmu_gather *tlb,
544 		       struct vm_area_struct *vma, unsigned long start,
545 		       unsigned long end, unsigned long cp_flags)
546 {
547 	pgprot_t newprot = vma->vm_page_prot;
548 	long pages;
549 
550 	BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
551 
552 #ifdef CONFIG_NUMA_BALANCING
553 	/*
554 	 * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
555 	 * are expected to reflect their requirements via VMA flags such that
556 	 * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
557 	 */
558 	if (cp_flags & MM_CP_PROT_NUMA)
559 		newprot = PAGE_NONE;
560 #else
561 	WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
562 #endif
563 
564 	if (is_vm_hugetlb_page(vma))
565 		pages = hugetlb_change_protection(vma, start, end, newprot,
566 						  cp_flags);
567 	else
568 		pages = change_protection_range(tlb, vma, start, end, newprot,
569 						cp_flags);
570 
571 	return pages;
572 }
573 
574 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
575 			       unsigned long next, struct mm_walk *walk)
576 {
577 	return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
578 				  *(pgprot_t *)(walk->private)) ?
579 		0 : -EACCES;
580 }
581 
582 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
583 				   unsigned long addr, unsigned long next,
584 				   struct mm_walk *walk)
585 {
586 	return pfn_modify_allowed(pte_pfn(ptep_get(pte)),
587 				  *(pgprot_t *)(walk->private)) ?
588 		0 : -EACCES;
589 }
590 
591 static int prot_none_test(unsigned long addr, unsigned long next,
592 			  struct mm_walk *walk)
593 {
594 	return 0;
595 }
596 
597 static const struct mm_walk_ops prot_none_walk_ops = {
598 	.pte_entry		= prot_none_pte_entry,
599 	.hugetlb_entry		= prot_none_hugetlb_entry,
600 	.test_walk		= prot_none_test,
601 	.walk_lock		= PGWALK_WRLOCK,
602 };
603 
604 int
605 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
606 	       struct vm_area_struct *vma, struct vm_area_struct **pprev,
607 	       unsigned long start, unsigned long end, unsigned long newflags)
608 {
609 	struct mm_struct *mm = vma->vm_mm;
610 	unsigned long oldflags = vma->vm_flags;
611 	long nrpages = (end - start) >> PAGE_SHIFT;
612 	unsigned int mm_cp_flags = 0;
613 	unsigned long charged = 0;
614 	int error;
615 
616 	if (!can_modify_vma(vma))
617 		return -EPERM;
618 
619 	if (newflags == oldflags) {
620 		*pprev = vma;
621 		return 0;
622 	}
623 
624 	/*
625 	 * Do PROT_NONE PFN permission checks here when we can still
626 	 * bail out without undoing a lot of state. This is a rather
627 	 * uncommon case, so doesn't need to be very optimized.
628 	 */
629 	if (arch_has_pfn_modify_check() &&
630 	    (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
631 	    (newflags & VM_ACCESS_FLAGS) == 0) {
632 		pgprot_t new_pgprot = vm_get_page_prot(newflags);
633 
634 		error = walk_page_range(current->mm, start, end,
635 				&prot_none_walk_ops, &new_pgprot);
636 		if (error)
637 			return error;
638 	}
639 
640 	/*
641 	 * If we make a private mapping writable we increase our commit;
642 	 * but (without finer accounting) cannot reduce our commit if we
643 	 * make it unwritable again except in the anonymous case where no
644 	 * anon_vma has yet to be assigned.
645 	 *
646 	 * hugetlb mapping were accounted for even if read-only so there is
647 	 * no need to account for them here.
648 	 */
649 	if (newflags & VM_WRITE) {
650 		/* Check space limits when area turns into data. */
651 		if (!may_expand_vm(mm, newflags, nrpages) &&
652 				may_expand_vm(mm, oldflags, nrpages))
653 			return -ENOMEM;
654 		if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
655 						VM_SHARED|VM_NORESERVE))) {
656 			charged = nrpages;
657 			if (security_vm_enough_memory_mm(mm, charged))
658 				return -ENOMEM;
659 			newflags |= VM_ACCOUNT;
660 		}
661 	} else if ((oldflags & VM_ACCOUNT) && vma_is_anonymous(vma) &&
662 		   !vma->anon_vma) {
663 		newflags &= ~VM_ACCOUNT;
664 	}
665 
666 	vma = vma_modify_flags(vmi, *pprev, vma, start, end, newflags);
667 	if (IS_ERR(vma)) {
668 		error = PTR_ERR(vma);
669 		goto fail;
670 	}
671 
672 	*pprev = vma;
673 
674 	/*
675 	 * vm_flags and vm_page_prot are protected by the mmap_lock
676 	 * held in write mode.
677 	 */
678 	vma_start_write(vma);
679 	vm_flags_reset(vma, newflags);
680 	if (vma_wants_manual_pte_write_upgrade(vma))
681 		mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
682 	vma_set_page_prot(vma);
683 
684 	change_protection(tlb, vma, start, end, mm_cp_flags);
685 
686 	if ((oldflags & VM_ACCOUNT) && !(newflags & VM_ACCOUNT))
687 		vm_unacct_memory(nrpages);
688 
689 	/*
690 	 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
691 	 * fault on access.
692 	 */
693 	if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
694 			(newflags & VM_WRITE)) {
695 		populate_vma_page_range(vma, start, end, NULL);
696 	}
697 
698 	vm_stat_account(mm, oldflags, -nrpages);
699 	vm_stat_account(mm, newflags, nrpages);
700 	perf_event_mmap(vma);
701 	return 0;
702 
703 fail:
704 	vm_unacct_memory(charged);
705 	return error;
706 }
707 
708 /*
709  * pkey==-1 when doing a legacy mprotect()
710  */
711 static int do_mprotect_pkey(unsigned long start, size_t len,
712 		unsigned long prot, int pkey)
713 {
714 	unsigned long nstart, end, tmp, reqprot;
715 	struct vm_area_struct *vma, *prev;
716 	int error;
717 	const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
718 	const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
719 				(prot & PROT_READ);
720 	struct mmu_gather tlb;
721 	struct vma_iterator vmi;
722 
723 	start = untagged_addr(start);
724 
725 	prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
726 	if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
727 		return -EINVAL;
728 
729 	if (start & ~PAGE_MASK)
730 		return -EINVAL;
731 	if (!len)
732 		return 0;
733 	len = PAGE_ALIGN(len);
734 	end = start + len;
735 	if (end <= start)
736 		return -ENOMEM;
737 	if (!arch_validate_prot(prot, start))
738 		return -EINVAL;
739 
740 	reqprot = prot;
741 
742 	if (mmap_write_lock_killable(current->mm))
743 		return -EINTR;
744 
745 	/*
746 	 * If userspace did not allocate the pkey, do not let
747 	 * them use it here.
748 	 */
749 	error = -EINVAL;
750 	if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
751 		goto out;
752 
753 	vma_iter_init(&vmi, current->mm, start);
754 	vma = vma_find(&vmi, end);
755 	error = -ENOMEM;
756 	if (!vma)
757 		goto out;
758 
759 	if (unlikely(grows & PROT_GROWSDOWN)) {
760 		if (vma->vm_start >= end)
761 			goto out;
762 		start = vma->vm_start;
763 		error = -EINVAL;
764 		if (!(vma->vm_flags & VM_GROWSDOWN))
765 			goto out;
766 	} else {
767 		if (vma->vm_start > start)
768 			goto out;
769 		if (unlikely(grows & PROT_GROWSUP)) {
770 			end = vma->vm_end;
771 			error = -EINVAL;
772 			if (!(vma->vm_flags & VM_GROWSUP))
773 				goto out;
774 		}
775 	}
776 
777 	prev = vma_prev(&vmi);
778 	if (start > vma->vm_start)
779 		prev = vma;
780 
781 	tlb_gather_mmu(&tlb, current->mm);
782 	nstart = start;
783 	tmp = vma->vm_start;
784 	for_each_vma_range(vmi, vma, end) {
785 		unsigned long mask_off_old_flags;
786 		unsigned long newflags;
787 		int new_vma_pkey;
788 
789 		if (vma->vm_start != tmp) {
790 			error = -ENOMEM;
791 			break;
792 		}
793 
794 		/* Does the application expect PROT_READ to imply PROT_EXEC */
795 		if (rier && (vma->vm_flags & VM_MAYEXEC))
796 			prot |= PROT_EXEC;
797 
798 		/*
799 		 * Each mprotect() call explicitly passes r/w/x permissions.
800 		 * If a permission is not passed to mprotect(), it must be
801 		 * cleared from the VMA.
802 		 */
803 		mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
804 
805 		new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
806 		newflags = calc_vm_prot_bits(prot, new_vma_pkey);
807 		newflags |= (vma->vm_flags & ~mask_off_old_flags);
808 
809 		/* newflags >> 4 shift VM_MAY% in place of VM_% */
810 		if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
811 			error = -EACCES;
812 			break;
813 		}
814 
815 		if (map_deny_write_exec(vma->vm_flags, newflags)) {
816 			error = -EACCES;
817 			break;
818 		}
819 
820 		/* Allow architectures to sanity-check the new flags */
821 		if (!arch_validate_flags(newflags)) {
822 			error = -EINVAL;
823 			break;
824 		}
825 
826 		error = security_file_mprotect(vma, reqprot, prot);
827 		if (error)
828 			break;
829 
830 		tmp = vma->vm_end;
831 		if (tmp > end)
832 			tmp = end;
833 
834 		if (vma->vm_ops && vma->vm_ops->mprotect) {
835 			error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
836 			if (error)
837 				break;
838 		}
839 
840 		error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
841 		if (error)
842 			break;
843 
844 		tmp = vma_iter_end(&vmi);
845 		nstart = tmp;
846 		prot = reqprot;
847 	}
848 	tlb_finish_mmu(&tlb);
849 
850 	if (!error && tmp < end)
851 		error = -ENOMEM;
852 
853 out:
854 	mmap_write_unlock(current->mm);
855 	return error;
856 }
857 
858 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
859 		unsigned long, prot)
860 {
861 	return do_mprotect_pkey(start, len, prot, -1);
862 }
863 
864 #ifdef CONFIG_ARCH_HAS_PKEYS
865 
866 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
867 		unsigned long, prot, int, pkey)
868 {
869 	return do_mprotect_pkey(start, len, prot, pkey);
870 }
871 
872 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
873 {
874 	int pkey;
875 	int ret;
876 
877 	/* No flags supported yet. */
878 	if (flags)
879 		return -EINVAL;
880 	/* check for unsupported init values */
881 	if (init_val & ~PKEY_ACCESS_MASK)
882 		return -EINVAL;
883 
884 	mmap_write_lock(current->mm);
885 	pkey = mm_pkey_alloc(current->mm);
886 
887 	ret = -ENOSPC;
888 	if (pkey == -1)
889 		goto out;
890 
891 	ret = arch_set_user_pkey_access(current, pkey, init_val);
892 	if (ret) {
893 		mm_pkey_free(current->mm, pkey);
894 		goto out;
895 	}
896 	ret = pkey;
897 out:
898 	mmap_write_unlock(current->mm);
899 	return ret;
900 }
901 
902 SYSCALL_DEFINE1(pkey_free, int, pkey)
903 {
904 	int ret;
905 
906 	mmap_write_lock(current->mm);
907 	ret = mm_pkey_free(current->mm, pkey);
908 	mmap_write_unlock(current->mm);
909 
910 	/*
911 	 * We could provide warnings or errors if any VMA still
912 	 * has the pkey set here.
913 	 */
914 	return ret;
915 }
916 
917 #endif /* CONFIG_ARCH_HAS_PKEYS */
918