forked from ~ljy/RK356X_SDK_RELEASE

blame | history | patch | commit | commitdiff
hc
2024-02-20 102a0743326a03cd1a1202ceda21e175b7d3575c 
 kernel/fs/userfaultfd.c
....@@ -1,12 +1,10 @@
1
+// SPDX-License-Identifier: GPL-2.0-only
12 /*
23  *  fs/userfaultfd.c
34  *
45  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
56  *  Copyright (C) 2008-2009 Red Hat, Inc.
67  *  Copyright (C) 2015  Red Hat, Inc.
7
- *
8
- *  This work is licensed under the terms of the GNU GPL, version 2. See
9
- *  the COPYING file in the top-level directory.
108  *
119  *  Some part derived from fs/eventfd.c (anon inode setup) and
1210  *  mm/ksm.c (mm hashing).
....@@ -17,6 +15,7 @@
1715 #include <linux/sched/signal.h>
1816 #include <linux/sched/mm.h>
1917 #include <linux/mm.h>
18
+#include <linux/mmu_notifier.h>
2019 #include <linux/poll.h>
2120 #include <linux/slab.h>
2221 #include <linux/seq_file.h>
....@@ -29,6 +28,8 @@
2928 #include <linux/ioctl.h>
3029 #include <linux/security.h>
3130 #include <linux/hugetlb.h>
31
+
32
+int sysctl_unprivileged_userfaultfd __read_mostly;
3233 
3334 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
3435 
....@@ -56,9 +57,9 @@
5657 	/* waitqueue head for events */
5758 	wait_queue_head_t event_wqh;
5859 	/* a refile sequence protected by fault_pending_wqh lock */
59
-	struct seqcount refile_seq;
60
+	seqcount_spinlock_t refile_seq;
6061 	/* pseudo fd refcounting */
61
-	atomic_t refcount;
62
+	refcount_t refcount;
6263 	/* userfaultfd syscall flags */
6364 	unsigned int flags;
6465 	/* features requested from the userspace */
....@@ -151,8 +152,7 @@
151152  */
152153 static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx)
153154 {
154
-	if (!atomic_inc_not_zero(&ctx->refcount))
155
-		BUG();
155
+	refcount_inc(&ctx->refcount);
156156 }
157157 
158158 /**
....@@ -165,7 +165,7 @@
165165  */
166166 static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx)
167167 {
168
-	if (atomic_dec_and_test(&ctx->refcount)) {
168
+	if (refcount_dec_and_test(&ctx->refcount)) {
169169 		VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock));
170170 		VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh));
171171 		VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock));
....@@ -198,24 +198,21 @@
198198 	msg_init(&msg);
199199 	msg.event = UFFD_EVENT_PAGEFAULT;
200200 	msg.arg.pagefault.address = address;
201
+	/*
202
+	 * These flags indicate why the userfault occurred:
203
+	 * - UFFD_PAGEFAULT_FLAG_WP indicates a write protect fault.
204
+	 * - UFFD_PAGEFAULT_FLAG_MINOR indicates a minor fault.
205
+	 * - Neither of these flags being set indicates a MISSING fault.
206
+	 *
207
+	 * Separately, UFFD_PAGEFAULT_FLAG_WRITE indicates it was a write
208
+	 * fault. Otherwise, it was a read fault.
209
+	 */
201210 	if (flags & FAULT_FLAG_WRITE)
202
-		/*
203
-		 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
204
-		 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE
205
-		 * was not set in a UFFD_EVENT_PAGEFAULT, it means it
206
-		 * was a read fault, otherwise if set it means it's
207
-		 * a write fault.
208
-		 */
209211 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE;
210212 	if (reason & VM_UFFD_WP)
211
-		/*
212
-		 * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the
213
-		 * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was
214
-		 * not set in a UFFD_EVENT_PAGEFAULT, it means it was
215
-		 * a missing fault, otherwise if set it means it's a
216
-		 * write protect fault.
217
-		 */
218213 		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP;
214
+	if (reason & VM_UFFD_MINOR)
215
+		msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_MINOR;
219216 	if (features & UFFD_FEATURE_THREAD_ID)
220217 		msg.arg.pagefault.feat.ptid = task_pid_vnr(current);
221218 	return msg;
....@@ -236,7 +233,7 @@
236233 	pte_t *ptep, pte;
237234 	bool ret = true;
238235 
239
-	VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
236
+	mmap_assert_locked(mm);
240237 
241238 	ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
242239 
....@@ -288,7 +285,7 @@
288285 	pte_t *pte;
289286 	bool ret = true;
290287 
291
-	VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
288
+	mmap_assert_locked(mm);
292289 
293290 	pgd = pgd_offset(mm, address);
294291 	if (!pgd_present(*pgd))
....@@ -316,8 +313,11 @@
316313 	if (!pmd_present(_pmd))
317314 		goto out;
318315 
319
-	if (pmd_trans_huge(_pmd))
316
+	if (pmd_trans_huge(_pmd)) {
317
+		if (!pmd_write(_pmd) && (reason & VM_UFFD_WP))
318
+			ret = true;
320319 		goto out;
320
+	}
321321 
322322 	/*
323323 	 * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it
....@@ -330,10 +330,23 @@
330330 	 */
331331 	if (pte_none(*pte))
332332 		ret = true;
333
+	if (!pte_write(*pte) && (reason & VM_UFFD_WP))
334
+		ret = true;
333335 	pte_unmap(pte);
334336 
335337 out:
336338 	return ret;
339
+}
340
+
341
+static inline long userfaultfd_get_blocking_state(unsigned int flags)
342
+{
343
+	if (flags & FAULT_FLAG_INTERRUPTIBLE)
344
+		return TASK_INTERRUPTIBLE;
345
+
346
+	if (flags & FAULT_FLAG_KILLABLE)
347
+		return TASK_KILLABLE;
348
+
349
+	return TASK_UNINTERRUPTIBLE;
337350 }
338351 
339352 /*
....@@ -342,13 +355,13 @@
342355  * FAULT_FLAG_KILLABLE are not straightforward. The "Caution"
343356  * recommendation in __lock_page_or_retry is not an understatement.
344357  *
345
- * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released
358
+ * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released
346359  * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is
347360  * not set.
348361  *
349362  * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not
350363  * set, VM_FAULT_RETRY can still be returned if and only if there are
351
- * fatal_signal_pending()s, and the mmap_sem must be released before
364
+ * fatal_signal_pending()s, and the mmap_lock must be released before
352365  * returning it.
353366  */
354367 vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
....@@ -357,7 +370,7 @@
357370 	struct userfaultfd_ctx *ctx;
358371 	struct userfaultfd_wait_queue uwq;
359372 	vm_fault_t ret = VM_FAULT_SIGBUS;
360
-	bool must_wait, return_to_userland;
373
+	bool must_wait;
361374 	long blocking_state;
362375 
363376 	/*
....@@ -369,16 +382,16 @@
369382 	 * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with
370383 	 * the no_page_table() helper in follow_page_mask(), but the
371384 	 * shmem_vm_ops->fault method is invoked even during
372
-	 * coredumping without mmap_sem and it ends up here.
385
+	 * coredumping without mmap_lock and it ends up here.
373386 	 */
374387 	if (current->flags & (PF_EXITING|PF_DUMPCORE))
375388 		goto out;
376389 
377390 	/*
378
-	 * Coredumping runs without mmap_sem so we can only check that
379
-	 * the mmap_sem is held, if PF_DUMPCORE was not set.
391
+	 * Coredumping runs without mmap_lock so we can only check that
392
+	 * the mmap_lock is held, if PF_DUMPCORE was not set.
380393 	 */
381
-	WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
394
+	mmap_assert_locked(mm);
382395 
383396 	ctx = vmf->vma->vm_userfaultfd_ctx.ctx;
384397 	if (!ctx)
....@@ -386,16 +399,25 @@
386399 
387400 	BUG_ON(ctx->mm != mm);
388401 
389
-	VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP));
390
-	VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP));
402
+	/* Any unrecognized flag is a bug. */
403
+	VM_BUG_ON(reason & ~__VM_UFFD_FLAGS);
404
+	/* 0 or > 1 flags set is a bug; we expect exactly 1. */
405
+	VM_BUG_ON(!reason || (reason & (reason - 1)));
391406 
392407 	if (ctx->features & UFFD_FEATURE_SIGBUS)
393408 		goto out;
409
+	if ((vmf->flags & FAULT_FLAG_USER) == 0 &&
410
+	    ctx->flags & UFFD_USER_MODE_ONLY) {
411
+		printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
412
+			"sysctl knob to 1 if kernel faults must be handled "
413
+			"without obtaining CAP_SYS_PTRACE capability\n");
414
+		goto out;
415
+	}
394416 
395417 	/*
396418 	 * If it's already released don't get it. This avoids to loop
397419 	 * in __get_user_pages if userfaultfd_release waits on the
398
-	 * caller of handle_userfault to release the mmap_sem.
420
+	 * caller of handle_userfault to release the mmap_lock.
399421 	 */
400422 	if (unlikely(READ_ONCE(ctx->released))) {
401423 		/*
....@@ -454,7 +476,7 @@
454476 	if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
455477 		goto out;
456478 
457
-	/* take the reference before dropping the mmap_sem */
479
+	/* take the reference before dropping the mmap_lock */
458480 	userfaultfd_ctx_get(ctx);
459481 
460482 	init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function);
....@@ -464,11 +486,7 @@
464486 	uwq.ctx = ctx;
465487 	uwq.waken = false;
466488 
467
-	return_to_userland =
468
-		(vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) ==
469
-		(FAULT_FLAG_USER|FAULT_FLAG_KILLABLE);
470
-	blocking_state = return_to_userland ? TASK_INTERRUPTIBLE :
471
-			 TASK_KILLABLE;
489
+	blocking_state = userfaultfd_get_blocking_state(vmf->flags);
472490 
473491 	spin_lock_irq(&ctx->fault_pending_wqh.lock);
474492 	/*
....@@ -491,64 +509,14 @@
491509 		must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma,
492510 						       vmf->address,
493511 						       vmf->flags, reason);
494
-	up_read(&mm->mmap_sem);
512
+	mmap_read_unlock(mm);
495513 
496
-	if (likely(must_wait && !READ_ONCE(ctx->released) &&
497
-		   (return_to_userland ? !signal_pending(current) :
498
-		    !fatal_signal_pending(current)))) {
514
+	if (likely(must_wait && !READ_ONCE(ctx->released))) {
499515 		wake_up_poll(&ctx->fd_wqh, EPOLLIN);
500516 		schedule();
501
-		ret |= VM_FAULT_MAJOR;
502
-
503
-		/*
504
-		 * False wakeups can orginate even from rwsem before
505
-		 * up_read() however userfaults will wait either for a
506
-		 * targeted wakeup on the specific uwq waitqueue from
507
-		 * wake_userfault() or for signals or for uffd
508
-		 * release.
509
-		 */
510
-		while (!READ_ONCE(uwq.waken)) {
511
-			/*
512
-			 * This needs the full smp_store_mb()
513
-			 * guarantee as the state write must be
514
-			 * visible to other CPUs before reading
515
-			 * uwq.waken from other CPUs.
516
-			 */
517
-			set_current_state(blocking_state);
518
-			if (READ_ONCE(uwq.waken) ||
519
-			    READ_ONCE(ctx->released) ||
520
-			    (return_to_userland ? signal_pending(current) :
521
-			     fatal_signal_pending(current)))
522
-				break;
523
-			schedule();
524
-		}
525517 	}
526518 
527519 	__set_current_state(TASK_RUNNING);
528
-
529
-	if (return_to_userland) {
530
-		if (signal_pending(current) &&
531
-		    !fatal_signal_pending(current)) {
532
-			/*
533
-			 * If we got a SIGSTOP or SIGCONT and this is
534
-			 * a normal userland page fault, just let
535
-			 * userland return so the signal will be
536
-			 * handled and gdb debugging works.  The page
537
-			 * fault code immediately after we return from
538
-			 * this function is going to release the
539
-			 * mmap_sem and it's not depending on it
540
-			 * (unlike gup would if we were not to return
541
-			 * VM_FAULT_RETRY).
542
-			 *
543
-			 * If a fatal signal is pending we still take
544
-			 * the streamlined VM_FAULT_RETRY failure path
545
-			 * and there's no need to retake the mmap_sem
546
-			 * in such case.
547
-			 */
548
-			down_read(&mm->mmap_sem);
549
-			ret = VM_FAULT_NOPAGE;
550
-		}
551
-	}
552520 
553521 	/*
554522 	 * Here we race with the list_del; list_add in
....@@ -640,15 +608,13 @@
640608 		struct mm_struct *mm = release_new_ctx->mm;
641609 
642610 		/* the various vma->vm_userfaultfd_ctx still points to it */
643
-		down_write(&mm->mmap_sem);
644
-		/* no task can run (and in turn coredump) yet */
645
-		VM_WARN_ON(!mmget_still_valid(mm));
611
+		mmap_write_lock(mm);
646612 		for (vma = mm->mmap; vma; vma = vma->vm_next)
647613 			if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) {
648614 				vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
649
-				vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
615
+				vma->vm_flags &= ~__VM_UFFD_FLAGS;
650616 			}
651
-		up_write(&mm->mmap_sem);
617
+		mmap_write_unlock(mm);
652618 
653619 		userfaultfd_ctx_put(release_new_ctx);
654620 	}
....@@ -677,8 +643,11 @@
677643 
678644 	octx = vma->vm_userfaultfd_ctx.ctx;
679645 	if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
646
+		vm_write_begin(vma);
680647 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
681
-		vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
648
+		WRITE_ONCE(vma->vm_flags,
649
+			   vma->vm_flags & ~__VM_UFFD_FLAGS);
650
+		vm_write_end(vma);
682651 		return 0;
683652 	}
684653 
....@@ -699,7 +668,7 @@
699668 			return -ENOMEM;
700669 		}
701670 
702
-		atomic_set(&ctx->refcount, 1);
671
+		refcount_set(&ctx->refcount, 1);
703672 		ctx->flags = octx->flags;
704673 		ctx->features = octx->features;
705674 		ctx->released = false;
....@@ -759,7 +728,7 @@
759728 	} else {
760729 		/* Drop uffd context if remap feature not enabled */
761730 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
762
-		vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
731
+		vma->vm_flags &= ~__VM_UFFD_FLAGS;
763732 	}
764733 }
765734 
....@@ -801,7 +770,7 @@
801770 
802771 	userfaultfd_ctx_get(ctx);
803772 	WRITE_ONCE(ctx->mmap_changing, true);
804
-	up_read(&mm->mmap_sem);
773
+	mmap_read_unlock(mm);
805774 
806775 	msg_init(&ewq.msg);
807776 
....@@ -881,7 +850,6 @@
881850 	/* len == 0 means wake all */
882851 	struct userfaultfd_wake_range range = { .len = 0, };
883852 	unsigned long new_flags;
884
-	bool still_valid;
885853 
886854 	WRITE_ONCE(ctx->released, true);
887855 
....@@ -892,38 +860,37 @@
892860 	 * Flush page faults out of all CPUs. NOTE: all page faults
893861 	 * must be retried without returning VM_FAULT_SIGBUS if
894862 	 * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx
895
-	 * changes while handle_userfault released the mmap_sem. So
863
+	 * changes while handle_userfault released the mmap_lock. So
896864 	 * it's critical that released is set to true (above), before
897
-	 * taking the mmap_sem for writing.
865
+	 * taking the mmap_lock for writing.
898866 	 */
899
-	down_write(&mm->mmap_sem);
900
-	still_valid = mmget_still_valid(mm);
867
+	mmap_write_lock(mm);
901868 	prev = NULL;
902869 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
903870 		cond_resched();
904871 		BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^
905
-		       !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
872
+		       !!(vma->vm_flags & __VM_UFFD_FLAGS));
906873 		if (vma->vm_userfaultfd_ctx.ctx != ctx) {
907874 			prev = vma;
908875 			continue;
909876 		}
910
-		new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
911
-		if (still_valid) {
912
-			prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
913
-					 new_flags, vma->anon_vma,
914
-					 vma->vm_file, vma->vm_pgoff,
915
-					 vma_policy(vma),
916
-					 NULL_VM_UFFD_CTX,
917
-					 vma_get_anon_name(vma));
918
-			if (prev)
919
-				vma = prev;
920
-			else
921
-				prev = vma;
922
-		}
923
-		vma->vm_flags = new_flags;
877
+		new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
878
+		prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end,
879
+				 new_flags, vma->anon_vma,
880
+				 vma->vm_file, vma->vm_pgoff,
881
+				 vma_policy(vma),
882
+				 NULL_VM_UFFD_CTX,
883
+				 vma_get_anon_name(vma));
884
+		if (prev)
885
+			vma = prev;
886
+		else
887
+			prev = vma;
888
+		vm_write_begin(vma);
889
+		WRITE_ONCE(vma->vm_flags, new_flags);
924890 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
891
+		vm_write_end(vma);
925892 	}
926
-	up_write(&mm->mmap_sem);
893
+	mmap_write_unlock(mm);
927894 	mmput(mm);
928895 wakeup:
929896 	/*
....@@ -951,7 +918,7 @@
951918 	wait_queue_entry_t *wq;
952919 	struct userfaultfd_wait_queue *uwq;
953920 
954
-	VM_BUG_ON(!spin_is_locked(&wqh->lock));
921
+	lockdep_assert_held(&wqh->lock);
955922 
956923 	uwq = NULL;
957924 	if (!waitqueue_active(wqh))
....@@ -1013,14 +980,14 @@
1013980 
1014981 static const struct file_operations userfaultfd_fops;
1015982 
1016
-static int resolve_userfault_fork(struct userfaultfd_ctx *ctx,
1017
-				  struct userfaultfd_ctx *new,
983
+static int resolve_userfault_fork(struct userfaultfd_ctx *new,
984
+				  struct inode *inode,
1018985 				  struct uffd_msg *msg)
1019986 {
1020987 	int fd;
1021988 
1022
-	fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new,
1023
-			      O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
989
+	fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new,
990
+			O_RDONLY | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode);
1024991 	if (fd < 0)
1025992 		return fd;
1026993 
....@@ -1030,7 +997,7 @@
1030997 }
1031998 
1032999 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
1033
-				    struct uffd_msg *msg)
1000
+				    struct uffd_msg *msg, struct inode *inode)
10341001 {
10351002 	ssize_t ret;
10361003 	DECLARE_WAITQUEUE(wait, current);
....@@ -1141,7 +1108,7 @@
11411108 	spin_unlock_irq(&ctx->fd_wqh.lock);
11421109 
11431110 	if (!ret && msg->event == UFFD_EVENT_FORK) {
1144
-		ret = resolve_userfault_fork(ctx, fork_nctx, msg);
1111
+		ret = resolve_userfault_fork(fork_nctx, inode, msg);
11451112 		spin_lock_irq(&ctx->event_wqh.lock);
11461113 		if (!list_empty(&fork_event)) {
11471114 			/*
....@@ -1201,6 +1168,7 @@
12011168 	ssize_t _ret, ret = 0;
12021169 	struct uffd_msg msg;
12031170 	int no_wait = file->f_flags & O_NONBLOCK;
1171
+	struct inode *inode = file_inode(file);
12041172 
12051173 	if (!userfaultfd_is_initialized(ctx))
12061174 		return -EINVAL;
....@@ -1208,7 +1176,7 @@
12081176 	for (;;) {
12091177 		if (count < sizeof(msg))
12101178 			return ret ? ret : -EINVAL;
1211
-		_ret = userfaultfd_ctx_read(ctx, no_wait, &msg);
1179
+		_ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
12121180 		if (_ret < 0)
12131181 			return ret ? ret : _ret;
12141182 		if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
....@@ -1246,7 +1214,7 @@
12461214 	/*
12471215 	 * To be sure waitqueue_active() is not reordered by the CPU
12481216 	 * before the pagetable update, use an explicit SMP memory
1249
-	 * barrier here. PT lock release or up_read(mmap_sem) still
1217
+	 * barrier here. PT lock release or mmap_read_unlock(mm) still
12501218 	 * have release semantics that can allow the
12511219 	 * waitqueue_active() to be reordered before the pte update.
12521220 	 */
....@@ -1269,31 +1237,41 @@
12691237 }
12701238 
12711239 static __always_inline int validate_range(struct mm_struct *mm,
1272
-					  __u64 *start, __u64 len)
1240
+					  __u64 start, __u64 len)
12731241 {
12741242 	__u64 task_size = mm->task_size;
12751243 
1276
-	*start = untagged_addr(*start);
1277
-
1278
-	if (*start & ~PAGE_MASK)
1244
+	if (start & ~PAGE_MASK)
12791245 		return -EINVAL;
12801246 	if (len & ~PAGE_MASK)
12811247 		return -EINVAL;
12821248 	if (!len)
12831249 		return -EINVAL;
1284
-	if (*start < mmap_min_addr)
1250
+	if (start < mmap_min_addr)
12851251 		return -EINVAL;
1286
-	if (*start >= task_size)
1252
+	if (start >= task_size)
12871253 		return -EINVAL;
1288
-	if (len > task_size - *start)
1254
+	if (len > task_size - start)
12891255 		return -EINVAL;
12901256 	return 0;
12911257 }
12921258 
1293
-static inline bool vma_can_userfault(struct vm_area_struct *vma)
1259
+static inline bool vma_can_userfault(struct vm_area_struct *vma,
1260
+				     unsigned long vm_flags)
12941261 {
1262
+	/* FIXME: add WP support to hugetlbfs and shmem */
1263
+	if (vm_flags & VM_UFFD_WP) {
1264
+		if (is_vm_hugetlb_page(vma) || vma_is_shmem(vma))
1265
+			return false;
1266
+	}
1267
+
1268
+	if (vm_flags & VM_UFFD_MINOR) {
1269
+		if (!(is_vm_hugetlb_page(vma) || vma_is_shmem(vma)))
1270
+			return false;
1271
+	}
1272
+
12951273 	return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
1296
-		vma_is_shmem(vma);
1274
+	       vma_is_shmem(vma);
12971275 }
12981276 
12991277 static int userfaultfd_register(struct userfaultfd_ctx *ctx,
....@@ -1319,23 +1297,21 @@
13191297 	ret = -EINVAL;
13201298 	if (!uffdio_register.mode)
13211299 		goto out;
1322
-	if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING|
1323
-				     UFFDIO_REGISTER_MODE_WP))
1300
+	if (uffdio_register.mode & ~UFFD_API_REGISTER_MODES)
13241301 		goto out;
13251302 	vm_flags = 0;
13261303 	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING)
13271304 		vm_flags |= VM_UFFD_MISSING;
1328
-	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) {
1305
+	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)
13291306 		vm_flags |= VM_UFFD_WP;
1330
-		/*
1331
-		 * FIXME: remove the below error constraint by
1332
-		 * implementing the wprotect tracking mode.
1333
-		 */
1334
-		ret = -EINVAL;
1307
+	if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR) {
1308
+#ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
13351309 		goto out;
1310
+#endif
1311
+		vm_flags |= VM_UFFD_MINOR;
13361312 	}
13371313 
1338
-	ret = validate_range(mm, &uffdio_register.range.start,
1314
+	ret = validate_range(mm, uffdio_register.range.start,
13391315 			     uffdio_register.range.len);
13401316 	if (ret)
13411317 		goto out;
....@@ -1347,9 +1323,7 @@
13471323 	if (!mmget_not_zero(mm))
13481324 		goto out;
13491325 
1350
-	down_write(&mm->mmap_sem);
1351
-	if (!mmget_still_valid(mm))
1352
-		goto out_unlock;
1326
+	mmap_write_lock(mm);
13531327 	vma = find_vma_prev(mm, start, &prev);
13541328 	if (!vma)
13551329 		goto out_unlock;
....@@ -1379,11 +1353,11 @@
13791353 		cond_resched();
13801354 
13811355 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1382
-		       !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1356
+		       !!(cur->vm_flags & __VM_UFFD_FLAGS));
13831357 
13841358 		/* check not compatible vmas */
13851359 		ret = -EINVAL;
1386
-		if (!vma_can_userfault(cur))
1360
+		if (!vma_can_userfault(cur, vm_flags))
13871361 			goto out_unlock;
13881362 
13891363 		/*
....@@ -1411,6 +1385,8 @@
14111385 			if (end & (vma_hpagesize - 1))
14121386 				goto out_unlock;
14131387 		}
1388
+		if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE))
1389
+			goto out_unlock;
14141390 
14151391 		/*
14161392 		 * Check that this vma isn't already owned by a
....@@ -1440,7 +1416,7 @@
14401416 	do {
14411417 		cond_resched();
14421418 
1443
-		BUG_ON(!vma_can_userfault(vma));
1419
+		BUG_ON(!vma_can_userfault(vma, vm_flags));
14441420 		BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
14451421 		       vma->vm_userfaultfd_ctx.ctx != ctx);
14461422 		WARN_ON(!(vma->vm_flags & VM_MAYWRITE));
....@@ -1457,7 +1433,7 @@
14571433 			start = vma->vm_start;
14581434 		vma_end = min(end, vma->vm_end);
14591435 
1460
-		new_flags = (vma->vm_flags & ~vm_flags) | vm_flags;
1436
+		new_flags = (vma->vm_flags & ~__VM_UFFD_FLAGS) | vm_flags;
14611437 		prev = vma_merge(mm, prev, start, vma_end, new_flags,
14621438 				 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
14631439 				 vma_policy(vma),
....@@ -1483,8 +1459,13 @@
14831459 		 * the next vma was merged into the current one and
14841460 		 * the current one has not been updated yet.
14851461 		 */
1486
-		vma->vm_flags = new_flags;
1462
+		vm_write_begin(vma);
1463
+		WRITE_ONCE(vma->vm_flags, new_flags);
14871464 		vma->vm_userfaultfd_ctx.ctx = ctx;
1465
+		vm_write_end(vma);
1466
+
1467
+		if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma))
1468
+			hugetlb_unshare_all_pmds(vma);
14881469 
14891470 	skip:
14901471 		prev = vma;
....@@ -1492,17 +1473,31 @@
14921473 		vma = vma->vm_next;
14931474 	} while (vma && vma->vm_start < end);
14941475 out_unlock:
1495
-	up_write(&mm->mmap_sem);
1476
+	mmap_write_unlock(mm);
14961477 	mmput(mm);
14971478 	if (!ret) {
1479
+		__u64 ioctls_out;
1480
+
1481
+		ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1482
+		    UFFD_API_RANGE_IOCTLS;
1483
+
1484
+		/*
1485
+		 * Declare the WP ioctl only if the WP mode is
1486
+		 * specified and all checks passed with the range
1487
+		 */
1488
+		if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP))
1489
+			ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT);
1490
+
1491
+		/* CONTINUE ioctl is only supported for MINOR ranges. */
1492
+		if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_MINOR))
1493
+			ioctls_out &= ~((__u64)1 << _UFFDIO_CONTINUE);
1494
+
14981495 		/*
14991496 		 * Now that we scanned all vmas we can already tell
15001497 		 * userland which ioctls methods are guaranteed to
15011498 		 * succeed on this range.
15021499 		 */
1503
-		if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC :
1504
-			     UFFD_API_RANGE_IOCTLS,
1505
-			     &user_uffdio_register->ioctls))
1500
+		if (put_user(ioctls_out, &user_uffdio_register->ioctls))
15061501 			ret = -EFAULT;
15071502 	}
15081503 out:
....@@ -1525,7 +1520,7 @@
15251520 	if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister)))
15261521 		goto out;
15271522 
1528
-	ret = validate_range(mm, &uffdio_unregister.start,
1523
+	ret = validate_range(mm, uffdio_unregister.start,
15291524 			     uffdio_unregister.len);
15301525 	if (ret)
15311526 		goto out;
....@@ -1537,9 +1532,7 @@
15371532 	if (!mmget_not_zero(mm))
15381533 		goto out;
15391534 
1540
-	down_write(&mm->mmap_sem);
1541
-	if (!mmget_still_valid(mm))
1542
-		goto out_unlock;
1535
+	mmap_write_lock(mm);
15431536 	vma = find_vma_prev(mm, start, &prev);
15441537 	if (!vma)
15451538 		goto out_unlock;
....@@ -1569,7 +1562,7 @@
15691562 		cond_resched();
15701563 
15711564 		BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^
1572
-		       !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP)));
1565
+		       !!(cur->vm_flags & __VM_UFFD_FLAGS));
15731566 
15741567 		/*
15751568 		 * Check not compatible vmas, not strictly required
....@@ -1578,7 +1571,7 @@
15781571 		 * provides for more strict behavior to notice
15791572 		 * unregistration errors.
15801573 		 */
1581
-		if (!vma_can_userfault(cur))
1574
+		if (!vma_can_userfault(cur, cur->vm_flags))
15821575 			goto out_unlock;
15831576 
15841577 		found = true;
....@@ -1592,7 +1585,7 @@
15921585 	do {
15931586 		cond_resched();
15941587 
1595
-		BUG_ON(!vma_can_userfault(vma));
1588
+		BUG_ON(!vma_can_userfault(vma, vma->vm_flags));
15961589 
15971590 		/*
15981591 		 * Nothing to do: this vma is already registered into this
....@@ -1620,7 +1613,7 @@
16201613 			wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range);
16211614 		}
16221615 
1623
-		new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP);
1616
+		new_flags = vma->vm_flags & ~__VM_UFFD_FLAGS;
16241617 		prev = vma_merge(mm, prev, start, vma_end, new_flags,
16251618 				 vma->anon_vma, vma->vm_file, vma->vm_pgoff,
16261619 				 vma_policy(vma),
....@@ -1646,8 +1639,10 @@
16461639 		 * the next vma was merged into the current one and
16471640 		 * the current one has not been updated yet.
16481641 		 */
1649
-		vma->vm_flags = new_flags;
1642
+		vm_write_begin(vma);
1643
+		WRITE_ONCE(vma->vm_flags, new_flags);
16501644 		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
1645
+		vm_write_end(vma);
16511646 
16521647 	skip:
16531648 		prev = vma;
....@@ -1655,7 +1650,7 @@
16551650 		vma = vma->vm_next;
16561651 	} while (vma && vma->vm_start < end);
16571652 out_unlock:
1658
-	up_write(&mm->mmap_sem);
1653
+	mmap_write_unlock(mm);
16591654 	mmput(mm);
16601655 out:
16611656 	return ret;
....@@ -1677,7 +1672,7 @@
16771672 	if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake)))
16781673 		goto out;
16791674 
1680
-	ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len);
1675
+	ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len);
16811676 	if (ret)
16821677 		goto out;
16831678 
....@@ -1717,7 +1712,7 @@
17171712 			   sizeof(uffdio_copy)-sizeof(__s64)))
17181713 		goto out;
17191714 
1720
-	ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len);
1715
+	ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len);
17211716 	if (ret)
17221717 		goto out;
17231718 	/*
....@@ -1728,11 +1723,12 @@
17281723 	ret = -EINVAL;
17291724 	if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src)
17301725 		goto out;
1731
-	if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE)
1726
+	if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP))
17321727 		goto out;
17331728 	if (mmget_not_zero(ctx->mm)) {
17341729 		ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src,
1735
-				   uffdio_copy.len, &ctx->mmap_changing);
1730
+				   uffdio_copy.len, &ctx->mmap_changing,
1731
+				   uffdio_copy.mode);
17361732 		mmput(ctx->mm);
17371733 	} else {
17381734 		return -ESRCH;
....@@ -1773,7 +1769,7 @@
17731769 			   sizeof(uffdio_zeropage)-sizeof(__s64)))
17741770 		goto out;
17751771 
1776
-	ret = validate_range(ctx->mm, &uffdio_zeropage.range.start,
1772
+	ret = validate_range(ctx->mm, uffdio_zeropage.range.start,
17771773 			     uffdio_zeropage.range.len);
17781774 	if (ret)
17791775 		goto out;
....@@ -1801,6 +1797,119 @@
18011797 		wake_userfault(ctx, &range);
18021798 	}
18031799 	ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN;
1800
+out:
1801
+	return ret;
1802
+}
1803
+
1804
+static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx,
1805
+				    unsigned long arg)
1806
+{
1807
+	int ret;
1808
+	struct uffdio_writeprotect uffdio_wp;
1809
+	struct uffdio_writeprotect __user *user_uffdio_wp;
1810
+	struct userfaultfd_wake_range range;
1811
+	bool mode_wp, mode_dontwake;
1812
+
1813
+	if (READ_ONCE(ctx->mmap_changing))
1814
+		return -EAGAIN;
1815
+
1816
+	user_uffdio_wp = (struct uffdio_writeprotect __user *) arg;
1817
+
1818
+	if (copy_from_user(&uffdio_wp, user_uffdio_wp,
1819
+			   sizeof(struct uffdio_writeprotect)))
1820
+		return -EFAULT;
1821
+
1822
+	ret = validate_range(ctx->mm, uffdio_wp.range.start,
1823
+			     uffdio_wp.range.len);
1824
+	if (ret)
1825
+		return ret;
1826
+
1827
+	if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE |
1828
+			       UFFDIO_WRITEPROTECT_MODE_WP))
1829
+		return -EINVAL;
1830
+
1831
+	mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP;
1832
+	mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE;
1833
+
1834
+	if (mode_wp && mode_dontwake)
1835
+		return -EINVAL;
1836
+
1837
+	if (mmget_not_zero(ctx->mm)) {
1838
+		ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start,
1839
+					  uffdio_wp.range.len, mode_wp,
1840
+					  &ctx->mmap_changing);
1841
+		mmput(ctx->mm);
1842
+	} else {
1843
+		return -ESRCH;
1844
+	}
1845
+
1846
+	if (ret)
1847
+		return ret;
1848
+
1849
+	if (!mode_wp && !mode_dontwake) {
1850
+		range.start = uffdio_wp.range.start;
1851
+		range.len = uffdio_wp.range.len;
1852
+		wake_userfault(ctx, &range);
1853
+	}
1854
+	return ret;
1855
+}
1856
+
1857
+static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
1858
+{
1859
+	__s64 ret;
1860
+	struct uffdio_continue uffdio_continue;
1861
+	struct uffdio_continue __user *user_uffdio_continue;
1862
+	struct userfaultfd_wake_range range;
1863
+
1864
+	user_uffdio_continue = (struct uffdio_continue __user *)arg;
1865
+
1866
+	ret = -EAGAIN;
1867
+	if (READ_ONCE(ctx->mmap_changing))
1868
+		goto out;
1869
+
1870
+	ret = -EFAULT;
1871
+	if (copy_from_user(&uffdio_continue, user_uffdio_continue,
1872
+			   /* don't copy the output fields */
1873
+			   sizeof(uffdio_continue) - (sizeof(__s64))))
1874
+		goto out;
1875
+
1876
+	ret = validate_range(ctx->mm, uffdio_continue.range.start,
1877
+			     uffdio_continue.range.len);
1878
+	if (ret)
1879
+		goto out;
1880
+
1881
+	ret = -EINVAL;
1882
+	/* double check for wraparound just in case. */
1883
+	if (uffdio_continue.range.start + uffdio_continue.range.len <=
1884
+	    uffdio_continue.range.start) {
1885
+		goto out;
1886
+	}
1887
+	if (uffdio_continue.mode & ~UFFDIO_CONTINUE_MODE_DONTWAKE)
1888
+		goto out;
1889
+
1890
+	if (mmget_not_zero(ctx->mm)) {
1891
+		ret = mcopy_continue(ctx->mm, uffdio_continue.range.start,
1892
+				     uffdio_continue.range.len,
1893
+				     &ctx->mmap_changing);
1894
+		mmput(ctx->mm);
1895
+	} else {
1896
+		return -ESRCH;
1897
+	}
1898
+
1899
+	if (unlikely(put_user(ret, &user_uffdio_continue->mapped)))
1900
+		return -EFAULT;
1901
+	if (ret < 0)
1902
+		goto out;
1903
+
1904
+	/* len == 0 would wake all */
1905
+	BUG_ON(!ret);
1906
+	range.len = ret;
1907
+	if (!(uffdio_continue.mode & UFFDIO_CONTINUE_MODE_DONTWAKE)) {
1908
+		range.start = uffdio_continue.range.start;
1909
+		wake_userfault(ctx, &range);
1910
+	}
1911
+	ret = range.len == uffdio_continue.range.len ? 0 : -EAGAIN;
1912
+
18041913 out:
18051914 	return ret;
18061915 }
....@@ -1840,6 +1949,10 @@
18401949 		goto err_out;
18411950 	/* report all available features and ioctls to userland */
18421951 	uffdio_api.features = UFFD_API_FEATURES;
1952
+#ifndef CONFIG_HAVE_ARCH_USERFAULTFD_MINOR
1953
+	uffdio_api.features &=
1954
+		~(UFFD_FEATURE_MINOR_HUGETLBFS | UFFD_FEATURE_MINOR_SHMEM);
1955
+#endif
18431956 	uffdio_api.ioctls = UFFD_API_IOCTLS;
18441957 	ret = -EFAULT;
18451958 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
....@@ -1889,6 +2002,12 @@
18892002 	case UFFDIO_ZEROPAGE:
18902003 		ret = userfaultfd_zeropage(ctx, arg);
18912004 		break;
2005
+	case UFFDIO_WRITEPROTECT:
2006
+		ret = userfaultfd_writeprotect(ctx, arg);
2007
+		break;
2008
+	case UFFDIO_CONTINUE:
2009
+		ret = userfaultfd_continue(ctx, arg);
2010
+		break;
18922011 	}
18932012 	return ret;
18942013 }
....@@ -1929,7 +2048,7 @@
19292048 	.poll		= userfaultfd_poll,
19302049 	.read		= userfaultfd_read,
19312050 	.unlocked_ioctl = userfaultfd_ioctl,
1932
-	.compat_ioctl	= userfaultfd_ioctl,
2051
+	.compat_ioctl	= compat_ptr_ioctl,
19332052 	.llseek		= noop_llseek,
19342053 };
19352054 
....@@ -1941,7 +2060,7 @@
19412060 	init_waitqueue_head(&ctx->fault_wqh);
19422061 	init_waitqueue_head(&ctx->event_wqh);
19432062 	init_waitqueue_head(&ctx->fd_wqh);
1944
-	seqcount_init(&ctx->refile_seq);
2063
+	seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock);
19452064 }
19462065 
19472066 SYSCALL_DEFINE1(userfaultfd, int, flags)
....@@ -1949,20 +2068,30 @@
19492068 	struct userfaultfd_ctx *ctx;
19502069 	int fd;
19512070 
2071
+	if (!sysctl_unprivileged_userfaultfd &&
2072
+	    (flags & UFFD_USER_MODE_ONLY) == 0 &&
2073
+	    !capable(CAP_SYS_PTRACE)) {
2074
+		printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd "
2075
+			"sysctl knob to 1 if kernel faults must be handled "
2076
+			"without obtaining CAP_SYS_PTRACE capability\n");
2077
+		return -EPERM;
2078
+	}
2079
+
19522080 	BUG_ON(!current->mm);
19532081 
19542082 	/* Check the UFFD_* constants for consistency.  */
2083
+	BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS);
19552084 	BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC);
19562085 	BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK);
19572086 
1958
-	if (flags & ~UFFD_SHARED_FCNTL_FLAGS)
2087
+	if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY))
19592088 		return -EINVAL;
19602089 
19612090 	ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL);
19622091 	if (!ctx)
19632092 		return -ENOMEM;
19642093 
1965
-	atomic_set(&ctx->refcount, 1);
2094
+	refcount_set(&ctx->refcount, 1);
19662095 	ctx->flags = flags;
19672096 	ctx->features = 0;
19682097 	ctx->released = false;
....@@ -1971,8 +2100,8 @@
19712100 	/* prevent the mm struct to be freed */
19722101 	mmgrab(ctx->mm);
19732102 
1974
-	fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
1975
-			      O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS));
2103
+	fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx,
2104
+			O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
19762105 	if (fd < 0) {
19772106 		mmdrop(ctx->mm);
19782107 		kmem_cache_free(userfaultfd_ctx_cachep, ctx);