[PATCH 2/3] userfaultfd: prevent concurrent API initialization

[Nadav Amit <nadav.amit@gmail.com>]

From: Nadav Amit <namit@vmware.com>

userfaultfd assumes that the enabled features are set once and never
changed after UFFDIO_API ioctl succeeded.

However, currently, UFFDIO_API can be called concurrently from two
different threads, succeed on both threads and leave userfaultfd's
features in non-deterministic state. Theoretically, other uffd
operations (ioctl's and page-faults) can be dispatched while adversely
affected by such changes of features.

Moreover, the writes to ctx->state and ctx->features are not ordered,
which can - theoretically, again - let userfaultfd_ioctl() think
that userfaultfd API completed, while the features are still not
initialized.

To avoid races, it is arguably best to get rid of ctx->state. Since
there are only 2 states, record the API initialization in ctx->features
as the uppermost bit and remove ctx->state.

Cc: Pavel Emelyanov <xemul@parallels.com>
Fixes: 9cd75c3cd4c3d ("userfaultfd: non-cooperative: add ability to report non-PF events from uffd descriptor")
Signed-off-by: Nadav Amit <namit@vmware.com>
---
 fs/userfaultfd.c | 91 +++++++++++++++++++++++-------------------------
 1 file changed, 44 insertions(+), 47 deletions(-)

diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
index 29a3016f16c9..003f0d31743e 100644
--- a/fs/userfaultfd.c
+++ b/fs/userfaultfd.c
@@ -33,11 +33,6 @@ int sysctl_unprivileged_userfaultfd __read_mostly;
 
 static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly;
 
-enum userfaultfd_state {
-	UFFD_STATE_WAIT_API,
-	UFFD_STATE_RUNNING,
-};
-
 /*
  * Start with fault_pending_wqh and fault_wqh so they're more likely
  * to be in the same cacheline.
@@ -69,8 +64,6 @@ struct userfaultfd_ctx {
 	unsigned int flags;
 	/* features requested from the userspace */
 	unsigned int features;
-	/* state machine */
-	enum userfaultfd_state state;
 	/* released */
 	bool released;
 	/* memory mappings are changing because of non-cooperative event */
@@ -104,6 +97,14 @@ struct userfaultfd_wake_range {
 	unsigned long len;
 };
 
+/* internal indication that UFFD_API ioctl was successfully executed */
+#define UFFD_FEATURE_INITIALIZED		(1u << 31)
+
+static bool userfaultfd_is_initialized(struct userfaultfd_ctx *ctx)
+{
+	return ctx->features & UFFD_FEATURE_INITIALIZED;
+}
+
 static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode,
 				     int wake_flags, void *key)
 {
@@ -667,7 +668,6 @@ int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
 
 		refcount_set(&ctx->refcount, 1);
 		ctx->flags = octx->flags;
-		ctx->state = UFFD_STATE_RUNNING;
 		ctx->features = octx->features;
 		ctx->released = false;
 		atomic_set(&ctx->mmap_changing, 0);
@@ -944,38 +944,33 @@ static __poll_t userfaultfd_poll(struct file *file, poll_table *wait)
 
 	poll_wait(file, &ctx->fd_wqh, wait);
 
-	switch (ctx->state) {
-	case UFFD_STATE_WAIT_API:
+	if (!userfaultfd_is_initialized(ctx))
 		return EPOLLERR;
-	case UFFD_STATE_RUNNING:
-		/*
-		 * poll() never guarantees that read won't block.
-		 * userfaults can be waken before they're read().
-		 */
-		if (unlikely(!(file->f_flags & O_NONBLOCK)))
-			return EPOLLERR;
-		/*
-		 * lockless access to see if there are pending faults
-		 * __pollwait last action is the add_wait_queue but
-		 * the spin_unlock would allow the waitqueue_active to
-		 * pass above the actual list_add inside
-		 * add_wait_queue critical section. So use a full
-		 * memory barrier to serialize the list_add write of
-		 * add_wait_queue() with the waitqueue_active read
-		 * below.
-		 */
-		ret = 0;
-		smp_mb();
-		if (waitqueue_active(&ctx->fault_pending_wqh))
-			ret = EPOLLIN;
-		else if (waitqueue_active(&ctx->event_wqh))
-			ret = EPOLLIN;
 
-		return ret;
-	default:
-		WARN_ON_ONCE(1);
+	/*
+	 * poll() never guarantees that read won't block.
+	 * userfaults can be waken before they're read().
+	 */
+	if (unlikely(!(file->f_flags & O_NONBLOCK)))
 		return EPOLLERR;
-	}
+	/*
+	 * lockless access to see if there are pending faults
+	 * __pollwait last action is the add_wait_queue but
+	 * the spin_unlock would allow the waitqueue_active to
+	 * pass above the actual list_add inside
+	 * add_wait_queue critical section. So use a full
+	 * memory barrier to serialize the list_add write of
+	 * add_wait_queue() with the waitqueue_active read
+	 * below.
+	 */
+	ret = 0;
+	smp_mb();
+	if (waitqueue_active(&ctx->fault_pending_wqh))
+		ret = EPOLLIN;
+	else if (waitqueue_active(&ctx->event_wqh))
+		ret = EPOLLIN;
+
+	return ret;
 }
 
 static const struct file_operations userfaultfd_fops;
@@ -1170,7 +1165,7 @@ static ssize_t userfaultfd_read(struct file *file, char __user *buf,
 	int no_wait = file->f_flags & O_NONBLOCK;
 	struct inode *inode = file_inode(file);
 
-	if (ctx->state == UFFD_STATE_WAIT_API)
+	if (!userfaultfd_is_initialized(ctx))
 		return -EINVAL;
 
 	for (;;) {
@@ -1909,9 +1904,10 @@ static int userfaultfd_continue(struct userfaultfd_ctx *ctx, unsigned long arg)
 static inline unsigned int uffd_ctx_features(__u64 user_features)
 {
 	/*
-	 * For the current set of features the bits just coincide
+	 * For the current set of features the bits just coincide. Set
+	 * UFFD_FEATURE_INITIALIZED to mark the features as enabled.
 	 */
-	return (unsigned int)user_features;
+	return (unsigned int)user_features | UFFD_FEATURE_INITIALIZED;
 }
 
 /*
@@ -1924,12 +1920,10 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 {
 	struct uffdio_api uffdio_api;
 	void __user *buf = (void __user *)arg;
+	unsigned int ctx_features;
 	int ret;
 	__u64 features;
 
-	ret = -EINVAL;
-	if (ctx->state != UFFD_STATE_WAIT_API)
-		goto out;
 	ret = -EFAULT;
 	if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api)))
 		goto out;
@@ -1953,9 +1947,13 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 	ret = -EFAULT;
 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
 		goto out;
-	ctx->state = UFFD_STATE_RUNNING;
+
 	/* only enable the requested features for this uffd context */
-	ctx->features = uffd_ctx_features(features);
+	ctx_features = uffd_ctx_features(features);
+	ret = -EINVAL;
+	if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
+		goto err_out;
+
 	ret = 0;
 out:
 	return ret;
@@ -1972,7 +1970,7 @@ static long userfaultfd_ioctl(struct file *file, unsigned cmd,
 	int ret = -EINVAL;
 	struct userfaultfd_ctx *ctx = file->private_data;
 
-	if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API)
+	if (cmd != UFFDIO_API && !userfaultfd_is_initialized(ctx))
 		return -EINVAL;
 
 	switch(cmd) {
@@ -2086,7 +2084,6 @@ SYSCALL_DEFINE1(userfaultfd, int, flags)
 	refcount_set(&ctx->refcount, 1);
 	ctx->flags = flags;
 	ctx->features = 0;
-	ctx->state = UFFD_STATE_WAIT_API;
 	ctx->released = false;
 	atomic_set(&ctx->mmap_changing, 0);
 	ctx->mm = current->mm;
-- 
2.25.1