xref: /openbmc/linux/mm/mprotect.c (revision 2bad466cc9d9b4c3b4b16eb9c03c919b59561316)
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 
280 			/*
281 			 * Nobody plays with any none ptes besides
282 			 * userfaultfd when applying the protections.
283 			 */
284 			if (likely(!uffd_wp))
285 				continue;
286 
287 			if (userfaultfd_wp_use_markers(vma)) {
288 				/*
289 				 * For file-backed mem, we need to be able to
290 				 * wr-protect a none pte, because even if the
291 				 * pte is none, the page/swap cache could
292 				 * exist.  Doing that by install a marker.
293 				 */
294 				set_pte_at(vma->vm_mm, addr, pte,
295 					   make_pte_marker(PTE_MARKER_UFFD_WP));
296 				pages++;
297 			}
298 		}
299 	} while (pte++, addr += PAGE_SIZE, addr != end);
300 	arch_leave_lazy_mmu_mode();
301 	pte_unmap_unlock(pte - 1, ptl);
302 
303 	return pages;
304 }
305 
306 /*
307  * Used when setting automatic NUMA hinting protection where it is
308  * critical that a numa hinting PMD is not confused with a bad PMD.
309  */
310 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
311 {
312 	pmd_t pmdval = pmdp_get_lockless(pmd);
313 
314 	/* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
315 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
316 	barrier();
317 #endif
318 
319 	if (pmd_none(pmdval))
320 		return 1;
321 	if (pmd_trans_huge(pmdval))
322 		return 0;
323 	if (unlikely(pmd_bad(pmdval))) {
324 		pmd_clear_bad(pmd);
325 		return 1;
326 	}
327 
328 	return 0;
329 }
330 
331 /*
332  * Return true if we want to split THPs into PTE mappings in change
333  * protection procedure, false otherwise.
334  */
335 static inline bool
336 pgtable_split_needed(struct vm_area_struct *vma, unsigned long cp_flags)
337 {
338 	/*
339 	 * pte markers only resides in pte level, if we need pte markers,
340 	 * we need to split.  We cannot wr-protect shmem thp because file
341 	 * thp is handled differently when split by erasing the pmd so far.
342 	 */
343 	return (cp_flags & MM_CP_UFFD_WP) && !vma_is_anonymous(vma);
344 }
345 
346 /*
347  * Return true if we want to populate pgtables in change protection
348  * procedure, false otherwise
349  */
350 static inline bool
351 pgtable_populate_needed(struct vm_area_struct *vma, unsigned long cp_flags)
352 {
353 	/* If not within ioctl(UFFDIO_WRITEPROTECT), then don't bother */
354 	if (!(cp_flags & MM_CP_UFFD_WP))
355 		return false;
356 
357 	/* Populate if the userfaultfd mode requires pte markers */
358 	return userfaultfd_wp_use_markers(vma);
359 }
360 
361 /*
362  * Populate the pgtable underneath for whatever reason if requested.
363  * When {pte|pmd|...}_alloc() failed we treat it the same way as pgtable
364  * allocation failures during page faults by kicking OOM and returning
365  * error.
366  */
367 #define  change_pmd_prepare(vma, pmd, cp_flags)				\
368 	({								\
369 		long err = 0;						\
370 		if (unlikely(pgtable_populate_needed(vma, cp_flags))) {	\
371 			if (pte_alloc(vma->vm_mm, pmd))			\
372 				err = -ENOMEM;				\
373 		}							\
374 		err;							\
375 	})
376 
377 /*
378  * This is the general pud/p4d/pgd version of change_pmd_prepare(). We need to
379  * have separate change_pmd_prepare() because pte_alloc() returns 0 on success,
380  * while {pmd|pud|p4d}_alloc() returns the valid pointer on success.
381  */
382 #define  change_prepare(vma, high, low, addr, cp_flags)			\
383 	  ({								\
384 		long err = 0;						\
385 		if (unlikely(pgtable_populate_needed(vma, cp_flags))) {	\
386 			low##_t *p = low##_alloc(vma->vm_mm, high, addr); \
387 			if (p == NULL)					\
388 				err = -ENOMEM;				\
389 		}							\
390 		err;							\
391 	})
392 
393 static inline long change_pmd_range(struct mmu_gather *tlb,
394 		struct vm_area_struct *vma, pud_t *pud, unsigned long addr,
395 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
396 {
397 	pmd_t *pmd;
398 	unsigned long next;
399 	long pages = 0;
400 	unsigned long nr_huge_updates = 0;
401 	struct mmu_notifier_range range;
402 
403 	range.start = 0;
404 
405 	pmd = pmd_offset(pud, addr);
406 	do {
407 		long ret;
408 
409 		next = pmd_addr_end(addr, end);
410 
411 		ret = change_pmd_prepare(vma, pmd, cp_flags);
412 		if (ret) {
413 			pages = ret;
414 			break;
415 		}
416 		/*
417 		 * Automatic NUMA balancing walks the tables with mmap_lock
418 		 * held for read. It's possible a parallel update to occur
419 		 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
420 		 * check leading to a false positive and clearing.
421 		 * Hence, it's necessary to atomically read the PMD value
422 		 * for all the checks.
423 		 */
424 		if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
425 		     pmd_none_or_clear_bad_unless_trans_huge(pmd))
426 			goto next;
427 
428 		/* invoke the mmu notifier if the pmd is populated */
429 		if (!range.start) {
430 			mmu_notifier_range_init(&range,
431 				MMU_NOTIFY_PROTECTION_VMA, 0,
432 				vma->vm_mm, addr, end);
433 			mmu_notifier_invalidate_range_start(&range);
434 		}
435 
436 		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
437 			if ((next - addr != HPAGE_PMD_SIZE) ||
438 			    pgtable_split_needed(vma, cp_flags)) {
439 				__split_huge_pmd(vma, pmd, addr, false, NULL);
440 				/*
441 				 * For file-backed, the pmd could have been
442 				 * cleared; make sure pmd populated if
443 				 * necessary, then fall-through to pte level.
444 				 */
445 				ret = change_pmd_prepare(vma, pmd, cp_flags);
446 				if (ret) {
447 					pages = ret;
448 					break;
449 				}
450 			} else {
451 				/*
452 				 * change_huge_pmd() does not defer TLB flushes,
453 				 * so no need to propagate the tlb argument.
454 				 */
455 				int nr_ptes = change_huge_pmd(tlb, vma, pmd,
456 						addr, newprot, cp_flags);
457 
458 				if (nr_ptes) {
459 					if (nr_ptes == HPAGE_PMD_NR) {
460 						pages += HPAGE_PMD_NR;
461 						nr_huge_updates++;
462 					}
463 
464 					/* huge pmd was handled */
465 					goto next;
466 				}
467 			}
468 			/* fall through, the trans huge pmd just split */
469 		}
470 		pages += change_pte_range(tlb, vma, pmd, addr, next,
471 					  newprot, cp_flags);
472 next:
473 		cond_resched();
474 	} while (pmd++, addr = next, addr != end);
475 
476 	if (range.start)
477 		mmu_notifier_invalidate_range_end(&range);
478 
479 	if (nr_huge_updates)
480 		count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
481 	return pages;
482 }
483 
484 static inline long change_pud_range(struct mmu_gather *tlb,
485 		struct vm_area_struct *vma, p4d_t *p4d, unsigned long addr,
486 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
487 {
488 	pud_t *pud;
489 	unsigned long next;
490 	long pages = 0, ret;
491 
492 	pud = pud_offset(p4d, addr);
493 	do {
494 		next = pud_addr_end(addr, end);
495 		ret = change_prepare(vma, pud, pmd, addr, cp_flags);
496 		if (ret)
497 			return ret;
498 		if (pud_none_or_clear_bad(pud))
499 			continue;
500 		pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
501 					  cp_flags);
502 	} while (pud++, addr = next, addr != end);
503 
504 	return pages;
505 }
506 
507 static inline long change_p4d_range(struct mmu_gather *tlb,
508 		struct vm_area_struct *vma, pgd_t *pgd, unsigned long addr,
509 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
510 {
511 	p4d_t *p4d;
512 	unsigned long next;
513 	long pages = 0, ret;
514 
515 	p4d = p4d_offset(pgd, addr);
516 	do {
517 		next = p4d_addr_end(addr, end);
518 		ret = change_prepare(vma, p4d, pud, addr, cp_flags);
519 		if (ret)
520 			return ret;
521 		if (p4d_none_or_clear_bad(p4d))
522 			continue;
523 		pages += change_pud_range(tlb, vma, p4d, addr, next, newprot,
524 					  cp_flags);
525 	} while (p4d++, addr = next, addr != end);
526 
527 	return pages;
528 }
529 
530 static long change_protection_range(struct mmu_gather *tlb,
531 		struct vm_area_struct *vma, unsigned long addr,
532 		unsigned long end, pgprot_t newprot, unsigned long cp_flags)
533 {
534 	struct mm_struct *mm = vma->vm_mm;
535 	pgd_t *pgd;
536 	unsigned long next;
537 	long pages = 0, ret;
538 
539 	BUG_ON(addr >= end);
540 	pgd = pgd_offset(mm, addr);
541 	tlb_start_vma(tlb, vma);
542 	do {
543 		next = pgd_addr_end(addr, end);
544 		ret = change_prepare(vma, pgd, p4d, addr, cp_flags);
545 		if (ret) {
546 			pages = ret;
547 			break;
548 		}
549 		if (pgd_none_or_clear_bad(pgd))
550 			continue;
551 		pages += change_p4d_range(tlb, vma, pgd, addr, next, newprot,
552 					  cp_flags);
553 	} while (pgd++, addr = next, addr != end);
554 
555 	tlb_end_vma(tlb, vma);
556 
557 	return pages;
558 }
559 
560 long change_protection(struct mmu_gather *tlb,
561 		       struct vm_area_struct *vma, unsigned long start,
562 		       unsigned long end, unsigned long cp_flags)
563 {
564 	pgprot_t newprot = vma->vm_page_prot;
565 	long pages;
566 
567 	BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
568 
569 #ifdef CONFIG_NUMA_BALANCING
570 	/*
571 	 * Ordinary protection updates (mprotect, uffd-wp, softdirty tracking)
572 	 * are expected to reflect their requirements via VMA flags such that
573 	 * vma_set_page_prot() will adjust vma->vm_page_prot accordingly.
574 	 */
575 	if (cp_flags & MM_CP_PROT_NUMA)
576 		newprot = PAGE_NONE;
577 #else
578 	WARN_ON_ONCE(cp_flags & MM_CP_PROT_NUMA);
579 #endif
580 
581 	if (is_vm_hugetlb_page(vma))
582 		pages = hugetlb_change_protection(vma, start, end, newprot,
583 						  cp_flags);
584 	else
585 		pages = change_protection_range(tlb, vma, start, end, newprot,
586 						cp_flags);
587 
588 	return pages;
589 }
590 
591 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
592 			       unsigned long next, struct mm_walk *walk)
593 {
594 	return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
595 		0 : -EACCES;
596 }
597 
598 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
599 				   unsigned long addr, unsigned long next,
600 				   struct mm_walk *walk)
601 {
602 	return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
603 		0 : -EACCES;
604 }
605 
606 static int prot_none_test(unsigned long addr, unsigned long next,
607 			  struct mm_walk *walk)
608 {
609 	return 0;
610 }
611 
612 static const struct mm_walk_ops prot_none_walk_ops = {
613 	.pte_entry		= prot_none_pte_entry,
614 	.hugetlb_entry		= prot_none_hugetlb_entry,
615 	.test_walk		= prot_none_test,
616 };
617 
618 int
619 mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
620 	       struct vm_area_struct *vma, struct vm_area_struct **pprev,
621 	       unsigned long start, unsigned long end, unsigned long newflags)
622 {
623 	struct mm_struct *mm = vma->vm_mm;
624 	unsigned long oldflags = vma->vm_flags;
625 	long nrpages = (end - start) >> PAGE_SHIFT;
626 	unsigned int mm_cp_flags = 0;
627 	unsigned long charged = 0;
628 	pgoff_t pgoff;
629 	int error;
630 
631 	if (newflags == oldflags) {
632 		*pprev = vma;
633 		return 0;
634 	}
635 
636 	/*
637 	 * Do PROT_NONE PFN permission checks here when we can still
638 	 * bail out without undoing a lot of state. This is a rather
639 	 * uncommon case, so doesn't need to be very optimized.
640 	 */
641 	if (arch_has_pfn_modify_check() &&
642 	    (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
643 	    (newflags & VM_ACCESS_FLAGS) == 0) {
644 		pgprot_t new_pgprot = vm_get_page_prot(newflags);
645 
646 		error = walk_page_range(current->mm, start, end,
647 				&prot_none_walk_ops, &new_pgprot);
648 		if (error)
649 			return error;
650 	}
651 
652 	/*
653 	 * If we make a private mapping writable we increase our commit;
654 	 * but (without finer accounting) cannot reduce our commit if we
655 	 * make it unwritable again. hugetlb mapping were accounted for
656 	 * even if read-only so there is no need to account for them here
657 	 */
658 	if (newflags & VM_WRITE) {
659 		/* Check space limits when area turns into data. */
660 		if (!may_expand_vm(mm, newflags, nrpages) &&
661 				may_expand_vm(mm, oldflags, nrpages))
662 			return -ENOMEM;
663 		if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
664 						VM_SHARED|VM_NORESERVE))) {
665 			charged = nrpages;
666 			if (security_vm_enough_memory_mm(mm, charged))
667 				return -ENOMEM;
668 			newflags |= VM_ACCOUNT;
669 		}
670 	}
671 
672 	/*
673 	 * First try to merge with previous and/or next vma.
674 	 */
675 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
676 	*pprev = vma_merge(vmi, mm, *pprev, start, end, newflags,
677 			   vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
678 			   vma->vm_userfaultfd_ctx, anon_vma_name(vma));
679 	if (*pprev) {
680 		vma = *pprev;
681 		VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
682 		goto success;
683 	}
684 
685 	*pprev = vma;
686 
687 	if (start != vma->vm_start) {
688 		error = split_vma(vmi, vma, start, 1);
689 		if (error)
690 			goto fail;
691 	}
692 
693 	if (end != vma->vm_end) {
694 		error = split_vma(vmi, vma, end, 0);
695 		if (error)
696 			goto fail;
697 	}
698 
699 success:
700 	/*
701 	 * vm_flags and vm_page_prot are protected by the mmap_lock
702 	 * held in write mode.
703 	 */
704 	vm_flags_reset(vma, newflags);
705 	if (vma_wants_manual_pte_write_upgrade(vma))
706 		mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
707 	vma_set_page_prot(vma);
708 
709 	change_protection(tlb, vma, start, end, mm_cp_flags);
710 
711 	/*
712 	 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
713 	 * fault on access.
714 	 */
715 	if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
716 			(newflags & VM_WRITE)) {
717 		populate_vma_page_range(vma, start, end, NULL);
718 	}
719 
720 	vm_stat_account(mm, oldflags, -nrpages);
721 	vm_stat_account(mm, newflags, nrpages);
722 	perf_event_mmap(vma);
723 	return 0;
724 
725 fail:
726 	vm_unacct_memory(charged);
727 	return error;
728 }
729 
730 /*
731  * pkey==-1 when doing a legacy mprotect()
732  */
733 static int do_mprotect_pkey(unsigned long start, size_t len,
734 		unsigned long prot, int pkey)
735 {
736 	unsigned long nstart, end, tmp, reqprot;
737 	struct vm_area_struct *vma, *prev;
738 	int error;
739 	const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
740 	const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
741 				(prot & PROT_READ);
742 	struct mmu_gather tlb;
743 	struct vma_iterator vmi;
744 
745 	start = untagged_addr(start);
746 
747 	prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
748 	if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
749 		return -EINVAL;
750 
751 	if (start & ~PAGE_MASK)
752 		return -EINVAL;
753 	if (!len)
754 		return 0;
755 	len = PAGE_ALIGN(len);
756 	end = start + len;
757 	if (end <= start)
758 		return -ENOMEM;
759 	if (!arch_validate_prot(prot, start))
760 		return -EINVAL;
761 
762 	reqprot = prot;
763 
764 	if (mmap_write_lock_killable(current->mm))
765 		return -EINTR;
766 
767 	/*
768 	 * If userspace did not allocate the pkey, do not let
769 	 * them use it here.
770 	 */
771 	error = -EINVAL;
772 	if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
773 		goto out;
774 
775 	vma_iter_init(&vmi, current->mm, start);
776 	vma = vma_find(&vmi, end);
777 	error = -ENOMEM;
778 	if (!vma)
779 		goto out;
780 
781 	if (unlikely(grows & PROT_GROWSDOWN)) {
782 		if (vma->vm_start >= end)
783 			goto out;
784 		start = vma->vm_start;
785 		error = -EINVAL;
786 		if (!(vma->vm_flags & VM_GROWSDOWN))
787 			goto out;
788 	} else {
789 		if (vma->vm_start > start)
790 			goto out;
791 		if (unlikely(grows & PROT_GROWSUP)) {
792 			end = vma->vm_end;
793 			error = -EINVAL;
794 			if (!(vma->vm_flags & VM_GROWSUP))
795 				goto out;
796 		}
797 	}
798 
799 	prev = vma_prev(&vmi);
800 	if (start > vma->vm_start)
801 		prev = vma;
802 
803 	tlb_gather_mmu(&tlb, current->mm);
804 	nstart = start;
805 	tmp = vma->vm_start;
806 	for_each_vma_range(vmi, vma, end) {
807 		unsigned long mask_off_old_flags;
808 		unsigned long newflags;
809 		int new_vma_pkey;
810 
811 		if (vma->vm_start != tmp) {
812 			error = -ENOMEM;
813 			break;
814 		}
815 
816 		/* Does the application expect PROT_READ to imply PROT_EXEC */
817 		if (rier && (vma->vm_flags & VM_MAYEXEC))
818 			prot |= PROT_EXEC;
819 
820 		/*
821 		 * Each mprotect() call explicitly passes r/w/x permissions.
822 		 * If a permission is not passed to mprotect(), it must be
823 		 * cleared from the VMA.
824 		 */
825 		mask_off_old_flags = VM_ACCESS_FLAGS | VM_FLAGS_CLEAR;
826 
827 		new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
828 		newflags = calc_vm_prot_bits(prot, new_vma_pkey);
829 		newflags |= (vma->vm_flags & ~mask_off_old_flags);
830 
831 		/* newflags >> 4 shift VM_MAY% in place of VM_% */
832 		if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
833 			error = -EACCES;
834 			break;
835 		}
836 
837 		if (map_deny_write_exec(vma, newflags)) {
838 			error = -EACCES;
839 			break;
840 		}
841 
842 		/* Allow architectures to sanity-check the new flags */
843 		if (!arch_validate_flags(newflags)) {
844 			error = -EINVAL;
845 			break;
846 		}
847 
848 		error = security_file_mprotect(vma, reqprot, prot);
849 		if (error)
850 			break;
851 
852 		tmp = vma->vm_end;
853 		if (tmp > end)
854 			tmp = end;
855 
856 		if (vma->vm_ops && vma->vm_ops->mprotect) {
857 			error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
858 			if (error)
859 				break;
860 		}
861 
862 		error = mprotect_fixup(&vmi, &tlb, vma, &prev, nstart, tmp, newflags);
863 		if (error)
864 			break;
865 
866 		tmp = vma_iter_end(&vmi);
867 		nstart = tmp;
868 		prot = reqprot;
869 	}
870 	tlb_finish_mmu(&tlb);
871 
872 	if (vma_iter_end(&vmi) < end)
873 		error = -ENOMEM;
874 
875 out:
876 	mmap_write_unlock(current->mm);
877 	return error;
878 }
879 
880 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
881 		unsigned long, prot)
882 {
883 	return do_mprotect_pkey(start, len, prot, -1);
884 }
885 
886 #ifdef CONFIG_ARCH_HAS_PKEYS
887 
888 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
889 		unsigned long, prot, int, pkey)
890 {
891 	return do_mprotect_pkey(start, len, prot, pkey);
892 }
893 
894 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
895 {
896 	int pkey;
897 	int ret;
898 
899 	/* No flags supported yet. */
900 	if (flags)
901 		return -EINVAL;
902 	/* check for unsupported init values */
903 	if (init_val & ~PKEY_ACCESS_MASK)
904 		return -EINVAL;
905 
906 	mmap_write_lock(current->mm);
907 	pkey = mm_pkey_alloc(current->mm);
908 
909 	ret = -ENOSPC;
910 	if (pkey == -1)
911 		goto out;
912 
913 	ret = arch_set_user_pkey_access(current, pkey, init_val);
914 	if (ret) {
915 		mm_pkey_free(current->mm, pkey);
916 		goto out;
917 	}
918 	ret = pkey;
919 out:
920 	mmap_write_unlock(current->mm);
921 	return ret;
922 }
923 
924 SYSCALL_DEFINE1(pkey_free, int, pkey)
925 {
926 	int ret;
927 
928 	mmap_write_lock(current->mm);
929 	ret = mm_pkey_free(current->mm, pkey);
930 	mmap_write_unlock(current->mm);
931 
932 	/*
933 	 * We could provide warnings or errors if any VMA still
934 	 * has the pkey set here.
935 	 */
936 	return ret;
937 }
938 
939 #endif /* CONFIG_ARCH_HAS_PKEYS */
940