xref: /openbmc/linux/mm/mprotect.c (revision 0a94608f)
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 <asm/cacheflush.h>
34 #include <asm/mmu_context.h>
35 #include <asm/tlbflush.h>
36 
37 #include "internal.h"
38 
39 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
40 		unsigned long addr, unsigned long end, pgprot_t newprot,
41 		unsigned long cp_flags)
42 {
43 	pte_t *pte, oldpte;
44 	spinlock_t *ptl;
45 	unsigned long pages = 0;
46 	int target_node = NUMA_NO_NODE;
47 	bool dirty_accountable = cp_flags & MM_CP_DIRTY_ACCT;
48 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
49 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
50 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
51 
52 	/*
53 	 * Can be called with only the mmap_lock for reading by
54 	 * prot_numa so we must check the pmd isn't constantly
55 	 * changing from under us from pmd_none to pmd_trans_huge
56 	 * and/or the other way around.
57 	 */
58 	if (pmd_trans_unstable(pmd))
59 		return 0;
60 
61 	/*
62 	 * The pmd points to a regular pte so the pmd can't change
63 	 * from under us even if the mmap_lock is only hold for
64 	 * reading.
65 	 */
66 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
67 
68 	/* Get target node for single threaded private VMAs */
69 	if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
70 	    atomic_read(&vma->vm_mm->mm_users) == 1)
71 		target_node = numa_node_id();
72 
73 	flush_tlb_batched_pending(vma->vm_mm);
74 	arch_enter_lazy_mmu_mode();
75 	do {
76 		oldpte = *pte;
77 		if (pte_present(oldpte)) {
78 			pte_t ptent;
79 			bool preserve_write = prot_numa && pte_write(oldpte);
80 
81 			/*
82 			 * Avoid trapping faults against the zero or KSM
83 			 * pages. See similar comment in change_huge_pmd.
84 			 */
85 			if (prot_numa) {
86 				struct page *page;
87 				int nid;
88 
89 				/* Avoid TLB flush if possible */
90 				if (pte_protnone(oldpte))
91 					continue;
92 
93 				page = vm_normal_page(vma, addr, oldpte);
94 				if (!page || PageKsm(page))
95 					continue;
96 
97 				/* Also skip shared copy-on-write pages */
98 				if (is_cow_mapping(vma->vm_flags) &&
99 				    page_count(page) != 1)
100 					continue;
101 
102 				/*
103 				 * While migration can move some dirty pages,
104 				 * it cannot move them all from MIGRATE_ASYNC
105 				 * context.
106 				 */
107 				if (page_is_file_lru(page) && PageDirty(page))
108 					continue;
109 
110 				/*
111 				 * Don't mess with PTEs if page is already on the node
112 				 * a single-threaded process is running on.
113 				 */
114 				nid = page_to_nid(page);
115 				if (target_node == nid)
116 					continue;
117 
118 				/*
119 				 * Skip scanning top tier node if normal numa
120 				 * balancing is disabled
121 				 */
122 				if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
123 				    node_is_toptier(nid))
124 					continue;
125 			}
126 
127 			oldpte = ptep_modify_prot_start(vma, addr, pte);
128 			ptent = pte_modify(oldpte, newprot);
129 			if (preserve_write)
130 				ptent = pte_mk_savedwrite(ptent);
131 
132 			if (uffd_wp) {
133 				ptent = pte_wrprotect(ptent);
134 				ptent = pte_mkuffd_wp(ptent);
135 			} else if (uffd_wp_resolve) {
136 				/*
137 				 * Leave the write bit to be handled
138 				 * by PF interrupt handler, then
139 				 * things like COW could be properly
140 				 * handled.
141 				 */
142 				ptent = pte_clear_uffd_wp(ptent);
143 			}
144 
145 			/* Avoid taking write faults for known dirty pages */
146 			if (dirty_accountable && pte_dirty(ptent) &&
147 					(pte_soft_dirty(ptent) ||
148 					 !(vma->vm_flags & VM_SOFTDIRTY))) {
149 				ptent = pte_mkwrite(ptent);
150 			}
151 			ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
152 			pages++;
153 		} else if (is_swap_pte(oldpte)) {
154 			swp_entry_t entry = pte_to_swp_entry(oldpte);
155 			pte_t newpte;
156 
157 			if (is_writable_migration_entry(entry)) {
158 				/*
159 				 * A protection check is difficult so
160 				 * just be safe and disable write
161 				 */
162 				entry = make_readable_migration_entry(
163 							swp_offset(entry));
164 				newpte = swp_entry_to_pte(entry);
165 				if (pte_swp_soft_dirty(oldpte))
166 					newpte = pte_swp_mksoft_dirty(newpte);
167 				if (pte_swp_uffd_wp(oldpte))
168 					newpte = pte_swp_mkuffd_wp(newpte);
169 			} else if (is_writable_device_private_entry(entry)) {
170 				/*
171 				 * We do not preserve soft-dirtiness. See
172 				 * copy_one_pte() for explanation.
173 				 */
174 				entry = make_readable_device_private_entry(
175 							swp_offset(entry));
176 				newpte = swp_entry_to_pte(entry);
177 				if (pte_swp_uffd_wp(oldpte))
178 					newpte = pte_swp_mkuffd_wp(newpte);
179 			} else if (is_writable_device_exclusive_entry(entry)) {
180 				entry = make_readable_device_exclusive_entry(
181 							swp_offset(entry));
182 				newpte = swp_entry_to_pte(entry);
183 				if (pte_swp_soft_dirty(oldpte))
184 					newpte = pte_swp_mksoft_dirty(newpte);
185 				if (pte_swp_uffd_wp(oldpte))
186 					newpte = pte_swp_mkuffd_wp(newpte);
187 			} else {
188 				newpte = oldpte;
189 			}
190 
191 			if (uffd_wp)
192 				newpte = pte_swp_mkuffd_wp(newpte);
193 			else if (uffd_wp_resolve)
194 				newpte = pte_swp_clear_uffd_wp(newpte);
195 
196 			if (!pte_same(oldpte, newpte)) {
197 				set_pte_at(vma->vm_mm, addr, pte, newpte);
198 				pages++;
199 			}
200 		}
201 	} while (pte++, addr += PAGE_SIZE, addr != end);
202 	arch_leave_lazy_mmu_mode();
203 	pte_unmap_unlock(pte - 1, ptl);
204 
205 	return pages;
206 }
207 
208 /*
209  * Used when setting automatic NUMA hinting protection where it is
210  * critical that a numa hinting PMD is not confused with a bad PMD.
211  */
212 static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
213 {
214 	pmd_t pmdval = pmd_read_atomic(pmd);
215 
216 	/* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
217 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
218 	barrier();
219 #endif
220 
221 	if (pmd_none(pmdval))
222 		return 1;
223 	if (pmd_trans_huge(pmdval))
224 		return 0;
225 	if (unlikely(pmd_bad(pmdval))) {
226 		pmd_clear_bad(pmd);
227 		return 1;
228 	}
229 
230 	return 0;
231 }
232 
233 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
234 		pud_t *pud, unsigned long addr, unsigned long end,
235 		pgprot_t newprot, unsigned long cp_flags)
236 {
237 	pmd_t *pmd;
238 	unsigned long next;
239 	unsigned long pages = 0;
240 	unsigned long nr_huge_updates = 0;
241 	struct mmu_notifier_range range;
242 
243 	range.start = 0;
244 
245 	pmd = pmd_offset(pud, addr);
246 	do {
247 		unsigned long this_pages;
248 
249 		next = pmd_addr_end(addr, end);
250 
251 		/*
252 		 * Automatic NUMA balancing walks the tables with mmap_lock
253 		 * held for read. It's possible a parallel update to occur
254 		 * between pmd_trans_huge() and a pmd_none_or_clear_bad()
255 		 * check leading to a false positive and clearing.
256 		 * Hence, it's necessary to atomically read the PMD value
257 		 * for all the checks.
258 		 */
259 		if (!is_swap_pmd(*pmd) && !pmd_devmap(*pmd) &&
260 		     pmd_none_or_clear_bad_unless_trans_huge(pmd))
261 			goto next;
262 
263 		/* invoke the mmu notifier if the pmd is populated */
264 		if (!range.start) {
265 			mmu_notifier_range_init(&range,
266 				MMU_NOTIFY_PROTECTION_VMA, 0,
267 				vma, vma->vm_mm, addr, end);
268 			mmu_notifier_invalidate_range_start(&range);
269 		}
270 
271 		if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
272 			if (next - addr != HPAGE_PMD_SIZE) {
273 				__split_huge_pmd(vma, pmd, addr, false, NULL);
274 			} else {
275 				int nr_ptes = change_huge_pmd(vma, pmd, addr,
276 							      newprot, cp_flags);
277 
278 				if (nr_ptes) {
279 					if (nr_ptes == HPAGE_PMD_NR) {
280 						pages += HPAGE_PMD_NR;
281 						nr_huge_updates++;
282 					}
283 
284 					/* huge pmd was handled */
285 					goto next;
286 				}
287 			}
288 			/* fall through, the trans huge pmd just split */
289 		}
290 		this_pages = change_pte_range(vma, pmd, addr, next, newprot,
291 					      cp_flags);
292 		pages += this_pages;
293 next:
294 		cond_resched();
295 	} while (pmd++, addr = next, addr != end);
296 
297 	if (range.start)
298 		mmu_notifier_invalidate_range_end(&range);
299 
300 	if (nr_huge_updates)
301 		count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
302 	return pages;
303 }
304 
305 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
306 		p4d_t *p4d, unsigned long addr, unsigned long end,
307 		pgprot_t newprot, unsigned long cp_flags)
308 {
309 	pud_t *pud;
310 	unsigned long next;
311 	unsigned long pages = 0;
312 
313 	pud = pud_offset(p4d, addr);
314 	do {
315 		next = pud_addr_end(addr, end);
316 		if (pud_none_or_clear_bad(pud))
317 			continue;
318 		pages += change_pmd_range(vma, pud, addr, next, newprot,
319 					  cp_flags);
320 	} while (pud++, addr = next, addr != end);
321 
322 	return pages;
323 }
324 
325 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
326 		pgd_t *pgd, unsigned long addr, unsigned long end,
327 		pgprot_t newprot, unsigned long cp_flags)
328 {
329 	p4d_t *p4d;
330 	unsigned long next;
331 	unsigned long pages = 0;
332 
333 	p4d = p4d_offset(pgd, addr);
334 	do {
335 		next = p4d_addr_end(addr, end);
336 		if (p4d_none_or_clear_bad(p4d))
337 			continue;
338 		pages += change_pud_range(vma, p4d, addr, next, newprot,
339 					  cp_flags);
340 	} while (p4d++, addr = next, addr != end);
341 
342 	return pages;
343 }
344 
345 static unsigned long change_protection_range(struct vm_area_struct *vma,
346 		unsigned long addr, unsigned long end, pgprot_t newprot,
347 		unsigned long cp_flags)
348 {
349 	struct mm_struct *mm = vma->vm_mm;
350 	pgd_t *pgd;
351 	unsigned long next;
352 	unsigned long start = addr;
353 	unsigned long pages = 0;
354 
355 	BUG_ON(addr >= end);
356 	pgd = pgd_offset(mm, addr);
357 	flush_cache_range(vma, addr, end);
358 	inc_tlb_flush_pending(mm);
359 	do {
360 		next = pgd_addr_end(addr, end);
361 		if (pgd_none_or_clear_bad(pgd))
362 			continue;
363 		pages += change_p4d_range(vma, pgd, addr, next, newprot,
364 					  cp_flags);
365 	} while (pgd++, addr = next, addr != end);
366 
367 	/* Only flush the TLB if we actually modified any entries: */
368 	if (pages)
369 		flush_tlb_range(vma, start, end);
370 	dec_tlb_flush_pending(mm);
371 
372 	return pages;
373 }
374 
375 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
376 		       unsigned long end, pgprot_t newprot,
377 		       unsigned long cp_flags)
378 {
379 	unsigned long pages;
380 
381 	BUG_ON((cp_flags & MM_CP_UFFD_WP_ALL) == MM_CP_UFFD_WP_ALL);
382 
383 	if (is_vm_hugetlb_page(vma))
384 		pages = hugetlb_change_protection(vma, start, end, newprot);
385 	else
386 		pages = change_protection_range(vma, start, end, newprot,
387 						cp_flags);
388 
389 	return pages;
390 }
391 
392 static int prot_none_pte_entry(pte_t *pte, unsigned long addr,
393 			       unsigned long next, struct mm_walk *walk)
394 {
395 	return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
396 		0 : -EACCES;
397 }
398 
399 static int prot_none_hugetlb_entry(pte_t *pte, unsigned long hmask,
400 				   unsigned long addr, unsigned long next,
401 				   struct mm_walk *walk)
402 {
403 	return pfn_modify_allowed(pte_pfn(*pte), *(pgprot_t *)(walk->private)) ?
404 		0 : -EACCES;
405 }
406 
407 static int prot_none_test(unsigned long addr, unsigned long next,
408 			  struct mm_walk *walk)
409 {
410 	return 0;
411 }
412 
413 static const struct mm_walk_ops prot_none_walk_ops = {
414 	.pte_entry		= prot_none_pte_entry,
415 	.hugetlb_entry		= prot_none_hugetlb_entry,
416 	.test_walk		= prot_none_test,
417 };
418 
419 int
420 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
421 	unsigned long start, unsigned long end, unsigned long newflags)
422 {
423 	struct mm_struct *mm = vma->vm_mm;
424 	unsigned long oldflags = vma->vm_flags;
425 	long nrpages = (end - start) >> PAGE_SHIFT;
426 	unsigned long charged = 0;
427 	pgoff_t pgoff;
428 	int error;
429 	int dirty_accountable = 0;
430 
431 	if (newflags == oldflags) {
432 		*pprev = vma;
433 		return 0;
434 	}
435 
436 	/*
437 	 * Do PROT_NONE PFN permission checks here when we can still
438 	 * bail out without undoing a lot of state. This is a rather
439 	 * uncommon case, so doesn't need to be very optimized.
440 	 */
441 	if (arch_has_pfn_modify_check() &&
442 	    (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
443 	    (newflags & VM_ACCESS_FLAGS) == 0) {
444 		pgprot_t new_pgprot = vm_get_page_prot(newflags);
445 
446 		error = walk_page_range(current->mm, start, end,
447 				&prot_none_walk_ops, &new_pgprot);
448 		if (error)
449 			return error;
450 	}
451 
452 	/*
453 	 * If we make a private mapping writable we increase our commit;
454 	 * but (without finer accounting) cannot reduce our commit if we
455 	 * make it unwritable again. hugetlb mapping were accounted for
456 	 * even if read-only so there is no need to account for them here
457 	 */
458 	if (newflags & VM_WRITE) {
459 		/* Check space limits when area turns into data. */
460 		if (!may_expand_vm(mm, newflags, nrpages) &&
461 				may_expand_vm(mm, oldflags, nrpages))
462 			return -ENOMEM;
463 		if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
464 						VM_SHARED|VM_NORESERVE))) {
465 			charged = nrpages;
466 			if (security_vm_enough_memory_mm(mm, charged))
467 				return -ENOMEM;
468 			newflags |= VM_ACCOUNT;
469 		}
470 	}
471 
472 	/*
473 	 * First try to merge with previous and/or next vma.
474 	 */
475 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
476 	*pprev = vma_merge(mm, *pprev, start, end, newflags,
477 			   vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
478 			   vma->vm_userfaultfd_ctx, anon_vma_name(vma));
479 	if (*pprev) {
480 		vma = *pprev;
481 		VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
482 		goto success;
483 	}
484 
485 	*pprev = vma;
486 
487 	if (start != vma->vm_start) {
488 		error = split_vma(mm, vma, start, 1);
489 		if (error)
490 			goto fail;
491 	}
492 
493 	if (end != vma->vm_end) {
494 		error = split_vma(mm, vma, end, 0);
495 		if (error)
496 			goto fail;
497 	}
498 
499 success:
500 	/*
501 	 * vm_flags and vm_page_prot are protected by the mmap_lock
502 	 * held in write mode.
503 	 */
504 	vma->vm_flags = newflags;
505 	dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
506 	vma_set_page_prot(vma);
507 
508 	change_protection(vma, start, end, vma->vm_page_prot,
509 			  dirty_accountable ? MM_CP_DIRTY_ACCT : 0);
510 
511 	/*
512 	 * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
513 	 * fault on access.
514 	 */
515 	if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
516 			(newflags & VM_WRITE)) {
517 		populate_vma_page_range(vma, start, end, NULL);
518 	}
519 
520 	vm_stat_account(mm, oldflags, -nrpages);
521 	vm_stat_account(mm, newflags, nrpages);
522 	perf_event_mmap(vma);
523 	return 0;
524 
525 fail:
526 	vm_unacct_memory(charged);
527 	return error;
528 }
529 
530 /*
531  * pkey==-1 when doing a legacy mprotect()
532  */
533 static int do_mprotect_pkey(unsigned long start, size_t len,
534 		unsigned long prot, int pkey)
535 {
536 	unsigned long nstart, end, tmp, reqprot;
537 	struct vm_area_struct *vma, *prev;
538 	int error = -EINVAL;
539 	const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
540 	const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
541 				(prot & PROT_READ);
542 
543 	start = untagged_addr(start);
544 
545 	prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
546 	if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
547 		return -EINVAL;
548 
549 	if (start & ~PAGE_MASK)
550 		return -EINVAL;
551 	if (!len)
552 		return 0;
553 	len = PAGE_ALIGN(len);
554 	end = start + len;
555 	if (end <= start)
556 		return -ENOMEM;
557 	if (!arch_validate_prot(prot, start))
558 		return -EINVAL;
559 
560 	reqprot = prot;
561 
562 	if (mmap_write_lock_killable(current->mm))
563 		return -EINTR;
564 
565 	/*
566 	 * If userspace did not allocate the pkey, do not let
567 	 * them use it here.
568 	 */
569 	error = -EINVAL;
570 	if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
571 		goto out;
572 
573 	vma = find_vma(current->mm, start);
574 	error = -ENOMEM;
575 	if (!vma)
576 		goto out;
577 
578 	if (unlikely(grows & PROT_GROWSDOWN)) {
579 		if (vma->vm_start >= end)
580 			goto out;
581 		start = vma->vm_start;
582 		error = -EINVAL;
583 		if (!(vma->vm_flags & VM_GROWSDOWN))
584 			goto out;
585 	} else {
586 		if (vma->vm_start > start)
587 			goto out;
588 		if (unlikely(grows & PROT_GROWSUP)) {
589 			end = vma->vm_end;
590 			error = -EINVAL;
591 			if (!(vma->vm_flags & VM_GROWSUP))
592 				goto out;
593 		}
594 	}
595 
596 	if (start > vma->vm_start)
597 		prev = vma;
598 	else
599 		prev = vma->vm_prev;
600 
601 	for (nstart = start ; ; ) {
602 		unsigned long mask_off_old_flags;
603 		unsigned long newflags;
604 		int new_vma_pkey;
605 
606 		/* Here we know that vma->vm_start <= nstart < vma->vm_end. */
607 
608 		/* Does the application expect PROT_READ to imply PROT_EXEC */
609 		if (rier && (vma->vm_flags & VM_MAYEXEC))
610 			prot |= PROT_EXEC;
611 
612 		/*
613 		 * Each mprotect() call explicitly passes r/w/x permissions.
614 		 * If a permission is not passed to mprotect(), it must be
615 		 * cleared from the VMA.
616 		 */
617 		mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
618 					VM_FLAGS_CLEAR;
619 
620 		new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
621 		newflags = calc_vm_prot_bits(prot, new_vma_pkey);
622 		newflags |= (vma->vm_flags & ~mask_off_old_flags);
623 
624 		/* newflags >> 4 shift VM_MAY% in place of VM_% */
625 		if ((newflags & ~(newflags >> 4)) & VM_ACCESS_FLAGS) {
626 			error = -EACCES;
627 			goto out;
628 		}
629 
630 		/* Allow architectures to sanity-check the new flags */
631 		if (!arch_validate_flags(newflags)) {
632 			error = -EINVAL;
633 			goto out;
634 		}
635 
636 		error = security_file_mprotect(vma, reqprot, prot);
637 		if (error)
638 			goto out;
639 
640 		tmp = vma->vm_end;
641 		if (tmp > end)
642 			tmp = end;
643 
644 		if (vma->vm_ops && vma->vm_ops->mprotect) {
645 			error = vma->vm_ops->mprotect(vma, nstart, tmp, newflags);
646 			if (error)
647 				goto out;
648 		}
649 
650 		error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
651 		if (error)
652 			goto out;
653 
654 		nstart = tmp;
655 
656 		if (nstart < prev->vm_end)
657 			nstart = prev->vm_end;
658 		if (nstart >= end)
659 			goto out;
660 
661 		vma = prev->vm_next;
662 		if (!vma || vma->vm_start != nstart) {
663 			error = -ENOMEM;
664 			goto out;
665 		}
666 		prot = reqprot;
667 	}
668 out:
669 	mmap_write_unlock(current->mm);
670 	return error;
671 }
672 
673 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
674 		unsigned long, prot)
675 {
676 	return do_mprotect_pkey(start, len, prot, -1);
677 }
678 
679 #ifdef CONFIG_ARCH_HAS_PKEYS
680 
681 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
682 		unsigned long, prot, int, pkey)
683 {
684 	return do_mprotect_pkey(start, len, prot, pkey);
685 }
686 
687 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
688 {
689 	int pkey;
690 	int ret;
691 
692 	/* No flags supported yet. */
693 	if (flags)
694 		return -EINVAL;
695 	/* check for unsupported init values */
696 	if (init_val & ~PKEY_ACCESS_MASK)
697 		return -EINVAL;
698 
699 	mmap_write_lock(current->mm);
700 	pkey = mm_pkey_alloc(current->mm);
701 
702 	ret = -ENOSPC;
703 	if (pkey == -1)
704 		goto out;
705 
706 	ret = arch_set_user_pkey_access(current, pkey, init_val);
707 	if (ret) {
708 		mm_pkey_free(current->mm, pkey);
709 		goto out;
710 	}
711 	ret = pkey;
712 out:
713 	mmap_write_unlock(current->mm);
714 	return ret;
715 }
716 
717 SYSCALL_DEFINE1(pkey_free, int, pkey)
718 {
719 	int ret;
720 
721 	mmap_write_lock(current->mm);
722 	ret = mm_pkey_free(current->mm, pkey);
723 	mmap_write_unlock(current->mm);
724 
725 	/*
726 	 * We could provide warnings or errors if any VMA still
727 	 * has the pkey set here.
728 	 */
729 	return ret;
730 }
731 
732 #endif /* CONFIG_ARCH_HAS_PKEYS */
733