xref: /linux/mm/mprotect.c (revision 8ccd54fe45713cd458015b5b08d6098545e70543)
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 <asm/cacheflush.h>
36 #include <asm/mmu_context.h>
37 #include <asm/tlbflush.h>
38 #include <asm/tlb.h>
39 
40 #include "internal.h"
41 
42 bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
43 			     pte_t pte)
44 {
45 	struct page *page;
46 
47 	if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
48 		return false;
49 
50 	/* Don't touch entries that are not even readable. */
51 	if (pte_protnone(pte))
52 		return false;
53 
54 	/* Do we need write faults for softdirty tracking? */
55 	if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
56 		return false;
57 
58 	/* Do we need write faults for uffd-wp tracking? */
59 	if (userfaultfd_pte_wp(vma, pte))
60 		return false;
61 
62 	if (!(vma->vm_flags & VM_SHARED)) {
63 		/*
64 		 * Writable MAP_PRIVATE mapping: We can only special-case on
65 		 * exclusive anonymous pages, because we know that our
66 		 * write-fault handler similarly would map them writable without
67 		 * any additional checks while holding the PT lock.
68 		 */
69 		page = vm_normal_page(vma, addr, pte);
70 		return page && PageAnon(page) && PageAnonExclusive(page);
71 	}
72 
73 	/*
74 	 * Writable MAP_SHARED mapping: "clean" might indicate that the FS still
75 	 * needs a real write-fault for writenotify
76 	 * (see vma_wants_writenotify()). If "dirty", the assumption is that the
77 	 * FS was already notified and we can simply mark the PTE writable
78 	 * just like the write-fault handler would do.
79 	 */
80 	return pte_dirty(pte);
81 }
82 
83 static long change_pte_range(struct mmu_gather *tlb,
84 		struct vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
85 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
86 {
87 	pte_t *pte, oldpte;
88 	spinlock_t *ptl;
89 	long pages = 0;
90 	int target_node = NUMA_NO_NODE;
91 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
92 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
93 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
94 
95 	tlb_change_page_size(tlb, PAGE_SIZE);
96 
97 	/*
98 	 * Can be called with only the mmap_lock for reading by
99 	 * prot_numa so we must check the pmd isn't constantly
100 	 * changing from under us from pmd_none to pmd_trans_huge
101 	 * and/or the other way around.
102 	 */
103 	if (pmd_trans_unstable(pmd))
104 		return 0;
105 
106 	/*
107 	 * The pmd points to a regular pte so the pmd can't change
108 	 * from under us even if the mmap_lock is only hold for
109 	 * reading.
110 	 */
111 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
112 
113 	/* Get target node for single threaded private VMAs */
114 	if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
115 	    atomic_read(&vma->vm_mm->mm_users) == 1)
116 		target_node = numa_node_id();
117 
118 	flush_tlb_batched_pending(vma->vm_mm);
119 	arch_enter_lazy_mmu_mode();
120 	do {
121 		oldpte = *pte;
122 		if (pte_present(oldpte)) {
123 			pte_t ptent;
124 
125 			/*
126 			 * Avoid trapping faults against the zero or KSM
127 			 * pages. See similar comment in change_huge_pmd.
128 			 */
129 			if (prot_numa) {
130 				struct page *page;
131 				int nid;
132 				bool toptier;
133 
134 				/* Avoid TLB flush if possible */
135 				if (pte_protnone(oldpte))
136 					continue;
137 
138 				page = vm_normal_page(vma, addr, oldpte);
139 				if (!page || is_zone_device_page(page) || PageKsm(page))
140 					continue;
141 
142 				/* Also skip shared copy-on-write pages */
143 				if (is_cow_mapping(vma->vm_flags) &&
144 				    page_count(page) != 1)
145 					continue;
146 
147 				/*
148 				 * While migration can move some dirty pages,
149 				 * it cannot move them all from MIGRATE_ASYNC
150 				 * context.
151 				 */
152 				if (page_is_file_lru(page) && PageDirty(page))
153 					continue;
154 
155 				/*
156 				 * Don't mess with PTEs if page is already on the node
157 				 * a single-threaded process is running on.
158 				 */
159 				nid = page_to_nid(page);
160 				if (target_node == nid)
161 					continue;
162 				toptier = node_is_toptier(nid);
163 
164 				/*
165 				 * Skip scanning top tier node if normal numa
166 				 * balancing is disabled
167 				 */
168 				if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
169 				    toptier)
170 					continue;
171 				if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
172 				    !toptier)
173 					xchg_page_access_time(page,
174 						jiffies_to_msecs(jiffies));
175 			}
176 
177 			oldpte = ptep_modify_prot_start(vma, addr, pte);
178 			ptent = pte_modify(oldpte, newprot);
179 
180 			if (uffd_wp)
181 				ptent = pte_mkuffd_wp(ptent);
182 			else if (uffd_wp_resolve)
183 				ptent = pte_clear_uffd_wp(ptent);
184 
185 			/*
186 			 * In some writable, shared mappings, we might want
187 			 * to catch actual write access -- see
188 			 * vma_wants_writenotify().
189 			 *
190 			 * In all writable, private mappings, we have to
191 			 * properly handle COW.
192 			 *
193 			 * In both cases, we can sometimes still change PTEs
194 			 * writable and avoid the write-fault handler, for
195 			 * example, if a PTE is already dirty and no other
196 			 * COW or special handling is required.
197 			 */
198 			if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) &&
199 			    !pte_write(ptent) &&
200 			    can_change_pte_writable(vma, addr, ptent))
201 				ptent = pte_mkwrite(ptent);
202 
203 			ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
204 			if (pte_needs_flush(oldpte, ptent))
205 				tlb_flush_pte_range(tlb, addr, PAGE_SIZE);
206 			pages++;
207 		} else if (is_swap_pte(oldpte)) {
208 			swp_entry_t entry = pte_to_swp_entry(oldpte);
209 			pte_t newpte;
210 
211 			if (is_writable_migration_entry(entry)) {
212 				struct page *page = pfn_swap_entry_to_page(entry);
213 
214 				/*
215 				 * A protection check is difficult so
216 				 * just be safe and disable write
217 				 */
218 				if (PageAnon(page))
219 					entry = make_readable_exclusive_migration_entry(
220 							     swp_offset(entry));
221 				else
222 					entry = make_readable_migration_entry(swp_offset(entry));
223 				newpte = swp_entry_to_pte(entry);
224 				if (pte_swp_soft_dirty(oldpte))
225 					newpte = pte_swp_mksoft_dirty(newpte);
226 				if (pte_swp_uffd_wp(oldpte))
227 					newpte = pte_swp_mkuffd_wp(newpte);
228 			} else if (is_writable_device_private_entry(entry)) {
229 				/*
230 				 * We do not preserve soft-dirtiness. See
231 				 * copy_one_pte() for explanation.
232 				 */
233 				entry = make_readable_device_private_entry(
234 							swp_offset(entry));
235 				newpte = swp_entry_to_pte(entry);
236 				if (pte_swp_uffd_wp(oldpte))
237 					newpte = pte_swp_mkuffd_wp(newpte);
238 			} else if (is_writable_device_exclusive_entry(entry)) {
239 				entry = make_readable_device_exclusive_entry(
240 							swp_offset(entry));
241 				newpte = swp_entry_to_pte(entry);
242 				if (pte_swp_soft_dirty(oldpte))
243 					newpte = pte_swp_mksoft_dirty(newpte);
244 				if (pte_swp_uffd_wp(oldpte))
245 					newpte = pte_swp_mkuffd_wp(newpte);
246 			} else if (is_pte_marker_entry(entry)) {
247 				/*
248 				 * Ignore swapin errors unconditionally,
249 				 * because any access should sigbus anyway.
250 				 */
251 				if (is_swapin_error_entry(entry))
252 					continue;
253 				/*
254 				 * If this is uffd-wp pte marker and we'd like
255 				 * to unprotect it, drop it; the next page
256 				 * fault will trigger without uffd trapping.
257 				 */
258 				if (uffd_wp_resolve) {
259 					pte_clear(vma->vm_mm, addr, pte);
260 					pages++;
261 				}
262 				continue;
263 			} else {
264 				newpte = oldpte;
265 			}
266 
267 			if (uffd_wp)
268 				newpte = pte_swp_mkuffd_wp(newpte);
269 			else if (uffd_wp_resolve)
270 				newpte = pte_swp_clear_uffd_wp(newpte);
271 
272 			if (!pte_same(oldpte, newpte)) {
273 				set_pte_at(vma->vm_mm, addr, pte, newpte);
274 				pages++;
275 			}
276 		} else {
277 			/* It must be an none page, or what else?.. */
278 			WARN_ON_ONCE(!pte_none(oldpte));
279 			if (unlikely(uffd_wp && !vma_is_anonymous(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  * Used when setting automatic NUMA hinting protection where it is
300  * critical that a numa hinting PMD is not confused with a bad PMD.
301  */
302 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
303 {
304 	pmd_t pmdval = pmdp_get_lockless(pmd);
305 
306 	/* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
307 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
308 	barrier();
309 #endif
310 
311 	if (pmd_none(pmdval))
312 		return 1;
313 	if (pmd_trans_huge(pmdval))
314 		return 0;
315 	if (unlikely(pmd_bad(pmdval))) {
316 		pmd_clear_bad(pmd);
317 		return 1;
318 	}
319 
320 	return 0;
321 }
322 
323 /* Return true if we're uffd wr-protecting file-backed memory, or false */
324 static inline bool
325 uffd_wp_protect_file(struct vm_area_struct *vma, unsigned long cp_flags)
326 {
327 	return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
328 }
329 
330 /*
331  * If wr-protecting the range for file-backed, populate pgtable for the case
332  * when pgtable is empty but page cache exists.  When {pte|pmd|...}_alloc()
333  * failed we treat it the same way as pgtable allocation failures during
334  * page faults by kicking OOM and returning error.
335  */
336 #define  change_pmd_prepare(vma, pmd, cp_flags)				\
337 	({								\
338 		long err = 0;						\
339 		if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {	\
340 			if (pte_alloc(vma->vm_mm, pmd))			\
341 				err = -ENOMEM;				\
342 		}							\
343 		err;							\
344 	})
345 
346 /*
347  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
348  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
349  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
350  */
351 #define  change_prepare(vma, high, low, addr, cp_flags)			\
352 	  ({								\
353 		long err = 0;						\
354 		if (unlikely(uffd_wp_protect_file(vma, cp_flags))) {	\
355 			low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
356 			if (p == NULL)					\
357 				err = -ENOMEM;				\
358 		}							\
359 		err;							\
360 	})
361 
362 static inline long change_pmd_range(struct mmu_gather *tlb,
363 		struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
364 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
365 {
366 	pmd_t *pmd;
367 	unsigned long next;
368 	long pages = 0;
369 	unsigned long nr_huge_updates = 0;
370 	struct mmu_notifier_range range;
371 
372 	range.start = 0;
373 
374 	pmd = pmd_offset(pud, addr);
375 	do {
376 		long ret;
377 
378 		next = pmd_addr_end(addr, end);
379 
380 		ret = change_pmd_prepare(vma, pmd, cp_flags);
381 		if (ret) {
382 			pages = ret;
383 			break;
384 		}
385 		/*
386 		 * Automatic NUMA balancing walks the tables with mmap_lock
387 		 * held for read. It's possible a parallel update to occur
388 		 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
389 		 * check leading to a false positive and clearing.
390 		 * Hence, it's necessary to atomically read the PMD value
391 		 * for all the checks.
392 		 */
393 		if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
394 		     pmd_none_or_clear_bad_unless_trans_huge(pmd))
395 			goto next;
396 
397 		/* invoke the mmu notifier if the pmd is populated */
398 		if (!range.start) {
399 			mmu_notifier_range_init(&range,
400 				MMU_NOTIFY_PROTECTION_VMA, 0,
401 				vma->vm_mm, addr, end);
402 			mmu_notifier_invalidate_range_start(&range);
403 		}
404 
405 		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
406 			if ((next - addr != HPAGE_PMD_SIZE) ||
407 			    uffd_wp_protect_file(vma, cp_flags)) {
408 				__split_huge_pmd(vma, pmd, addr, false, NULL);
409 				/*
410 				 * For file-backed, the pmd could have been
411 				 * cleared; make sure pmd populated if
412 				 * necessary, then fall-through to pte level.
413 				 */
414 				ret = change_pmd_prepare(vma, pmd, cp_flags);
415 				if (ret) {
416 					pages = ret;
417 					break;
418 				}
419 			} else {
420 				/*
421 				 * change_huge_pmd() does not defer TLB flushes,
422 				 * so no need to propagate the tlb argument.
423 				 */
424 				int nr_ptes = change_huge_pmd(tlb, vma, pmd,
425 						addr, newprot, cp_flags);
426 
427 				if (nr_ptes) {
428 					if (nr_ptes == HPAGE_PMD_NR) {
429 						pages += HPAGE_PMD_NR;
430 						nr_huge_updates++;
431 					}
432 
433 					/* huge pmd was handled */
434 					goto next;
435 				}
436 			}
437 			/* fall through, the trans huge pmd just split */
438 		}
439 		pages += change_pte_range(tlb, vma, pmd, addr, next,
440 					  newprot, cp_flags);
441 next:
442 		cond_resched();
443 	} while (pmd++, addr = next, addr != end);
444 
445 	if (range.start)
446 		mmu_notifier_invalidate_range_end(&range);
447 
448 	if (nr_huge_updates)
449 		count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
450 	return pages;
451 }
452 
453 static inline long change_pud_range(struct mmu_gather *tlb,
454 		struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
455 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
456 {
457 	pud_t *pud;
458 	unsigned long next;
459 	long pages = 0, ret;
460 
461 	pud = pud_offset(p4d, addr);
462 	do {
463 		next = pud_addr_end(addr, end);
464 		ret = change_prepare(vma, pud, pmd, addr, cp_flags);
465 		if (ret)
466 			return ret;
467 		if (pud_none_or_clear_bad(pud))
468 			continue;
469 		pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
470 					  cp_flags);
471 	} while (pud++, addr = next, addr != end);
472 
473 	return pages;
474 }
475 
476 static inline long change_p4d_range(struct mmu_gather *tlb,
477 		struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
478 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
479 {
480 	p4d_t *p4d;
481 	unsigned long next;
482 	long pages = 0, ret;
483 
484 	p4d = p4d_offset(pgd, addr);
485 	do {
486 		next = p4d_addr_end(addr, end);
487 		ret = change_prepare(vma, p4d, pud, addr, cp_flags);
488 		if (ret)
489 			return ret;
490 		if (p4d_none_or_clear_bad(p4d))
491 			continue;
492 		pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
493 					  cp_flags);
494 	} while (p4d++, addr = next, addr != end);
495 
496 	return pages;
497 }
498 
499 static long change_protection_range(struct mmu_gather *tlb,
500 		struct vm_area_struct *vma, unsigned long addr,
501 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
502 {
503 	struct mm_struct *mm = vma->vm_mm;
504 	pgd_t *pgd;
505 	unsigned long next;
506 	long pages = 0, ret;
507 
508 	BUG_ON(addr >= end);
509 	pgd = pgd_offset(mm, addr);
510 	tlb_start_vma(tlb, vma);
511 	do {
512 		next = pgd_addr_end(addr, end);
513 		ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
514 		if (ret) {
515 			pages = ret;
516 			break;
517 		}
518 		if (pgd_none_or_clear_bad(pgd))
519 			continue;
520 		pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
521 					  cp_flags);
522 	} while (pgd++, addr = next, addr != end);
523 
524 	tlb_end_vma(tlb, vma);
525 
526 	return pages;
527 }
528 
529 long change_protection(struct mmu_gather *tlb,
530 		       struct vm_area_struct *vma, unsigned long start,
531 		       unsigned long end, unsigned long cp_flags)
532 {
533 	pgprot_t newprot = vma->vm_page_prot;
534 	long pages;
535 
536 	BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
537 
538 #ifdef CONFIG_NUMA_BALANCING
539 	/*
540 	 * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
541 	 * are expected to reflect their requirements via VMA flags such that
542 	 * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
543 	 */
544 	if (cp_flags & MM_CP_PROT_NUMA)
545 		newprot = PAGE_NONE;
546 #else
547 	WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
548 #endif
549 
550 	if (is_vm_hugetlb_page(vma))
551 		pages = hugetlb_change_protection(vma, start, end, newprot,
552 						  cp_flags);
553 	else
554 		pages = change_protection_range(tlb, vma, start, end, newprot,
555 						cp_flags);
556 
557 	return pages;
558 }
559 
560 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
561 			       unsigned long next, struct mm_walk *walk)
562 {
563 	return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
564 		0 : -EACCES;
565 }
566 
567 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
568 				   unsigned long addr, unsigned long next,
569 				   struct mm_walk *walk)
570 {
571 	return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
572 		0 : -EACCES;
573 }
574 
575 static int prot_none_test(unsigned long addr, unsigned long next,
576 			  struct mm_walk *walk)
577 {
578 	return 0;
579 }
580 
581 static const struct mm_walk_ops prot_none_walk_ops = {
582 	.pte_entry		= prot_none_pte_entry,
583 	.hugetlb_entry		= prot_none_hugetlb_entry,
584 	.test_walk		= prot_none_test,
585 };
586 
587 int
588 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
589 	       struct vm_area_struct *vma, struct vm_area_struct **pprev,
590 	       unsigned long start, unsigned long end, unsigned long newflags)
591 {
592 	struct mm_struct *mm = vma->vm_mm;
593 	unsigned long oldflags = vma->vm_flags;
594 	long nrpages = (end - start) >> PAGE_SHIFT;
595 	unsigned int mm_cp_flags = 0;
596 	unsigned long charged = 0;
597 	pgoff_t pgoff;
598 	int error;
599 
600 	if (newflags == oldflags) {
601 		*pprev = vma;
602 		return 0;
603 	}
604 
605 	/*
606 	 * Do PROT_NONE PFN permission checks here when we can still
607 	 * bail out without undoing a lot of state. This is a rather
608 	 * uncommon case, so doesn't need to be very optimized.
609 	 */
610 	if (arch_has_pfn_modify_check() &&
611 	    (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
612 	    (newflags & VM_ACCESS_FLAGS) == 0) {
613 		pgprot_t new_pgprot = vm_get_page_prot(newflags);
614 
615 		error = walk_page_range(current->mm, start, end,
616 				&prot_none_walk_ops, &new_pgprot);
617 		if (error)
618 			return error;
619 	}
620 
621 	/*
622 	 * If we make a private mapping writable we increase our commit;
623 	 * but (without finer accounting) cannot reduce our commit if we
624 	 * make it unwritable again. hugetlb mapping were accounted for
625 	 * even if read-only so there is no need to account for them here
626 	 */
627 	if (newflags & VM_WRITE) {
628 		/* Check space limits when area turns into data. */
629 		if (!may_expand_vm(mm, newflags, nrpages) &&
630 				may_expand_vm(mm, oldflags, nrpages))
631 			return -ENOMEM;
632 		if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
633 						VM_SHARED|VM_NORESERVE))) {
634 			charged = nrpages;
635 			if (security_vm_enough_memory_mm(mm, charged))
636 				return -ENOMEM;
637 			newflags |= VM_ACCOUNT;
638 		}
639 	}
640 
641 	/*
642 	 * First try to merge with previous and/or next vma.
643 	 */
644 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
645 	*pprev = vma_merge(vmi, mm, *pprev, start, end, newflags,
646 			   vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
647 			   vma->vm_userfaultfd_ctx, anon_vma_name(vma));
648 	if (*pprev) {
649 		vma = *pprev;
650 		VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
651 		goto success;
652 	}
653 
654 	*pprev = vma;
655 
656 	if (start != vma->vm_start) {
657 		error = split_vma(vmi, vma, start, 1);
658 		if (error)
659 			goto fail;
660 	}
661 
662 	if (end != vma->vm_end) {
663 		error = split_vma(vmi, vma, end, 0);
664 		if (error)
665 			goto fail;
666 	}
667 
668 success:
669 	/*
670 	 * vm_flags and vm_page_prot are protected by the mmap_lock
671 	 * held in write mode.
672 	 */
673 	vm_flags_reset(vma, newflags);
674 	if (vma_wants_manual_pte_write_upgrade(vma))
675 		mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
676 	vma_set_page_prot(vma);
677 
678 	change_protection(tlb, vma, start, end, mm_cp_flags);
679 
680 	/*
681 	 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
682 	 * fault on access.
683 	 */
684 	if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
685 			(newflags & VM_WRITE)) {
686 		populate_vma_page_range(vma, start, end, NULL);
687 	}
688 
689 	vm_stat_account(mm, oldflags, -nrpages);
690 	vm_stat_account(mm, newflags, nrpages);
691 	perf_event_mmap(vma);
692 	return 0;
693 
694 fail:
695 	vm_unacct_memory(charged);
696 	return error;
697 }
698 
699 /*
700  * pkey==-1 when doing a legacy mprotect()
701  */
702 static int do_mprotect_pkey(unsigned long start, size_t len,
703 		unsigned long prot, int pkey)
704 {
705 	unsigned long nstart, end, tmp, reqprot;
706 	struct vm_area_struct *vma, *prev;
707 	int error;
708 	const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
709 	const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
710 				(prot & PROT_READ);
711 	struct mmu_gather tlb;
712 	struct vma_iterator vmi;
713 
714 	start = untagged_addr(start);
715 
716 	prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
717 	if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
718 		return -EINVAL;
719 
720 	if (start & ~PAGE_MASK)
721 		return -EINVAL;
722 	if (!len)
723 		return 0;
724 	len = PAGE_ALIGN(len);
725 	end = start + len;
726 	if (end <= start)
727 		return -ENOMEM;
728 	if (!arch_validate_prot(prot, start))
729 		return -EINVAL;
730 
731 	reqprot = prot;
732 
733 	if (mmap_write_lock_killable(current->mm))
734 		return -EINTR;
735 
736 	/*
737 	 * If userspace did not allocate the pkey, do not let
738 	 * them use it here.
739 	 */
740 	error = -EINVAL;
741 	if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
742 		goto out;
743 
744 	vma_iter_init(&vmi, current->mm, start);
745 	vma = vma_find(&vmi, end);
746 	error = -ENOMEM;
747 	if (!vma)
748 		goto out;
749 
750 	if (unlikely(grows & PROT_GROWSDOWN)) {
751 		if (vma->vm_start >= end)
752 			goto out;
753 		start = vma->vm_start;
754 		error = -EINVAL;
755 		if (!(vma->vm_flags & VM_GROWSDOWN))
756 			goto out;
757 	} else {
758 		if (vma->vm_start > start)
759 			goto out;
760 		if (unlikely(grows & PROT_GROWSUP)) {
761 			end = vma->vm_end;
762 			error = -EINVAL;
763 			if (!(vma->vm_flags & VM_GROWSUP))
764 				goto out;
765 		}
766 	}
767 
768 	prev = vma_prev(&vmi);
769 	if (start > vma->vm_start)
770 		prev = vma;
771 
772 	tlb_gather_mmu(&tlb, current->mm);
773 	nstart = start;
774 	tmp = vma->vm_start;
775 	for_each_vma_range(vmi, vma, end) {
776 		unsigned long mask_off_old_flags;
777 		unsigned long newflags;
778 		int new_vma_pkey;
779 
780 		if (vma->vm_start != tmp) {
781 			error = -ENOMEM;
782 			break;
783 		}
784 
785 		/* Does the application expect PROT_READ to imply PROT_EXEC */
786 		if (rier && (vma->vm_flags & VM_MAYEXEC))
787 			prot |= PROT_EXEC;
788 
789 		/*
790 		 * Each mprotect() call explicitly passes r/w/x permissions.
791 		 * If a permission is not passed to mprotect(), it must be
792 		 * cleared from the VMA.
793 		 */
794 		mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
795 
796 		new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
797 		newflags = calc_vm_prot_bits(prot, new_vma_pkey);
798 		newflags |= (vma->vm_flags & ~mask_off_old_flags);
799 
800 		/* newflags >> 4 shift VM_MAY% in place of VM_% */
801 		if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
802 			error = -EACCES;
803 			break;
804 		}
805 
806 		if (map_deny_write_exec(vma, newflags)) {
807 			error = -EACCES;
808 			break;
809 		}
810 
811 		/* Allow architectures to sanity-check the new flags */
812 		if (!arch_validate_flags(newflags)) {
813 			error = -EINVAL;
814 			break;
815 		}
816 
817 		error = security_file_mprotect(vma, reqprot, prot);
818 		if (error)
819 			break;
820 
821 		tmp = vma->vm_end;
822 		if (tmp > end)
823 			tmp = end;
824 
825 		if (vma->vm_ops && vma->vm_ops->mprotect) {
826 			error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
827 			if (error)
828 				break;
829 		}
830 
831 		error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
832 		if (error)
833 			break;
834 
835 		tmp = vma_iter_end(&vmi);
836 		nstart = tmp;
837 		prot = reqprot;
838 	}
839 	tlb_finish_mmu(&tlb);
840 
841 	if (!error && vma_iter_end(&vmi) < end)
842 		error = -ENOMEM;
843 
844 out:
845 	mmap_write_unlock(current->mm);
846 	return error;
847 }
848 
849 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
850 		unsigned long, prot)
851 {
852 	return do_mprotect_pkey(start, len, prot, -1);
853 }
854 
855 #ifdef CONFIG_ARCH_HAS_PKEYS
856 
857 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
858 		unsigned long, prot, int, pkey)
859 {
860 	return do_mprotect_pkey(start, len, prot, pkey);
861 }
862 
863 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
864 {
865 	int pkey;
866 	int ret;
867 
868 	/* No flags supported yet. */
869 	if (flags)
870 		return -EINVAL;
871 	/* check for unsupported init values */
872 	if (init_val & ~PKEY_ACCESS_MASK)
873 		return -EINVAL;
874 
875 	mmap_write_lock(current->mm);
876 	pkey = mm_pkey_alloc(current->mm);
877 
878 	ret = -ENOSPC;
879 	if (pkey == -1)
880 		goto out;
881 
882 	ret = arch_set_user_pkey_access(current, pkey, init_val);
883 	if (ret) {
884 		mm_pkey_free(current->mm, pkey);
885 		goto out;
886 	}
887 	ret = pkey;
888 out:
889 	mmap_write_unlock(current->mm);
890 	return ret;
891 }
892 
893 SYSCALL_DEFINE1(pkey_free, int, pkey)
894 {
895 	int ret;
896 
897 	mmap_write_lock(current->mm);
898 	ret = mm_pkey_free(current->mm, pkey);
899 	mmap_write_unlock(current->mm);
900 
901 	/*
902 	 * We could provide warnings or errors if any VMA still
903 	 * has the pkey set here.
904 	 */
905 	return ret;
906 }
907 
908 #endif /* CONFIG_ARCH_HAS_PKEYS */
909