Xen Security Advisory 392 v4 (CVE-2021-28714,CVE-2021-28715) - Guest can force Linux netback driver to hog large amounts of kernel memory

From 8b01ec26b69dfb8fc033600456c7522777c8e183 Mon Sep 17 00:00:00 2001
From: Juergen Gross <[email protected]>
Date: Fri, 26 Nov 2021 10:36:57 +0100
Subject: [PATCH 1/2] xen/netback: fix rx queue stall detection

Commit 1d5d48523900a4b ("xen-netback: require fewer guest Rx slots when
not using GSO") introduced a security problem in netback, as an
interface would only be regarded to be stalled if no slot is available
in the rx queue ring page. In case the SKB at the head of the queued
requests will need more than one rx slot and only one slot is free the
stall detection logic will never trigger, as the test for that is only
looking for at least one slot to be free.

Fix that by testing for the needed number of slots instead of only one
slot being available.

In order to not have to take the rx queue lock that often, store the
number of needed slots in the queue data. As all SKB dequeue operations
happen in the rx queue kernel thread this is safe, as long as the
number of needed slots is accessed via READ/WRITE_ONCE() only and
updates are always done with the rx queue lock held.

Add a small helper for obtaining the number of free slots.

This is CVE-2021-28714, part of XSA-392

Fixes: 1d5d48523900a4b ("xen-netback: require fewer guest Rx slots when not using GSO")
Signed-off-by: Juergen Gross <[email protected]>
Reviewed-by: Jan Beulich <[email protected]>
---
 drivers/net/xen-netback/common.h |  1 +
 drivers/net/xen-netback/rx.c     | 65 ++++++++++++++++++++------------
 2 files changed, 42 insertions(+), 24 deletions(-)

diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
index 4a16d6e33c09..d9dea4829c86 100644
--- a/drivers/net/xen-netback/common.h
+++ b/drivers/net/xen-netback/common.h
@@ -203,6 +203,7 @@ struct xenvif_queue { /* Per-queue data for xenvif */
 	unsigned int rx_queue_max;
 	unsigned int rx_queue_len;
 	unsigned long last_rx_time;
+	unsigned int rx_slots_needed;
 	bool stalled;
 
 	struct xenvif_copy_state rx_copy;
diff --git a/drivers/net/xen-netback/rx.c b/drivers/net/xen-netback/rx.c
index accc991d153f..a8511e27d6c1 100644
--- a/drivers/net/xen-netback/rx.c
+++ b/drivers/net/xen-netback/rx.c
@@ -33,28 +33,36 @@
 #include <xen/xen.h>
 #include <xen/events.h>
 
-static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
+/*
+ * Update the needed ring page slots for the first SKB queued.
+ * Note that any call sequence outside the RX thread calling this function
+ * needs to wake up the RX thread via a call of xenvif_kick_thread()
+ * afterwards in order to avoid a race with putting the thread to sleep.
+ */
+static void xenvif_update_needed_slots(struct xenvif_queue *queue,
+				       const struct sk_buff *skb)
 {
-	RING_IDX prod, cons;
-	struct sk_buff *skb;
-	int needed;
-	unsigned long flags;
-
-	spin_lock_irqsave(&queue->rx_queue.lock, flags);
+	unsigned int needed = 0;
 
-	skb = skb_peek(&queue->rx_queue);
-	if (!skb) {
-		spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
-		return false;
+	if (skb) {
+		needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
+		if (skb_is_gso(skb))
+			needed++;
+		if (skb->sw_hash)
+			needed++;
 	}
 
-	needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
-	if (skb_is_gso(skb))
-		needed++;
-	if (skb->sw_hash)
-		needed++;
+	WRITE_ONCE(queue->rx_slots_needed, needed);
+}
 
-	spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
+static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
+{
+	RING_IDX prod, cons;
+	unsigned int needed;
+
+	needed = READ_ONCE(queue->rx_slots_needed);
+	if (!needed)
+		return false;
 
 	do {
 		prod = queue->rx.sring->req_prod;
@@ -80,6 +88,9 @@ void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
 
 	spin_lock_irqsave(&queue->rx_queue.lock, flags);
 
+	if (skb_queue_empty(&queue->rx_queue))
+		xenvif_update_needed_slots(queue, skb);
+
 	__skb_queue_tail(&queue->rx_queue, skb);
 
 	queue->rx_queue_len += skb->len;
@@ -100,6 +111,8 @@ static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
 
 	skb = __skb_dequeue(&queue->rx_queue);
 	if (skb) {
+		xenvif_update_needed_slots(queue, skb_peek(&queue->rx_queue));
+
 		queue->rx_queue_len -= skb->len;
 		if (queue->rx_queue_len < queue->rx_queue_max) {
 			struct netdev_queue *txq;
@@ -487,27 +500,31 @@ void xenvif_rx_action(struct xenvif_queue *queue)
 	xenvif_rx_copy_flush(queue);
 }
 
-static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
+static RING_IDX xenvif_rx_queue_slots(const struct xenvif_queue *queue)
 {
 	RING_IDX prod, cons;
 
 	prod = queue->rx.sring->req_prod;
 	cons = queue->rx.req_cons;
 
+	return prod - cons;
+}
+
+static bool xenvif_rx_queue_stalled(const struct xenvif_queue *queue)
+{
+	unsigned int needed = READ_ONCE(queue->rx_slots_needed);
+
 	return !queue->stalled &&
-		prod - cons < 1 &&
+		xenvif_rx_queue_slots(queue) < needed &&
 		time_after(jiffies,
 			   queue->last_rx_time + queue->vif->stall_timeout);
 }
 
 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
 {
-	RING_IDX prod, cons;
-
-	prod = queue->rx.sring->req_prod;
-	cons = queue->rx.req_cons;
+	unsigned int needed = READ_ONCE(queue->rx_slots_needed);
 
-	return queue->stalled && prod - cons >= 1;
+	return queue->stalled && xenvif_rx_queue_slots(queue) >= needed;
 }
 
 bool xenvif_have_rx_work(struct xenvif_queue *queue, bool test_kthread)
-- 
2.26.2

From 71492f1f12db956a9eb49e727d12b225dad74205 Mon Sep 17 00:00:00 2001
From: Juergen Gross <[email protected]>
Date: Fri, 26 Nov 2021 10:36:57 +0100
Subject: [PATCH 2/2] xen/netback: don't queue unlimited number of packages

In case a guest isn't consuming incoming network traffic as fast as it
is coming in, xen-netback is buffering network packages in unlimited
numbers today. This can result in host OOM situations.

Commit f48da8b14d04ca8 ("xen-netback: fix unlimited guest Rx internal
queue and carrier flapping") meant to introduce a mechanism to limit
the amount of buffered data by stopping the Tx queue when reaching the
data limit, but this doesn't work for cases like UDP.

When hitting the limit don't queue further SKBs, but drop them instead.
In order to be able to tell Rx packages have been dropped increment the
rx_dropped statistics counter in this case.

It should be noted that the old solution to continue queueing SKBs had
the additional problem of an overflow of the 32-bit rx_queue_len value
would result in intermittent Tx queue enabling.

This is CVE-2021-28715, part of XSA-392

Fixes: f48da8b14d04ca8 ("xen-netback: fix unlimited guest Rx internal queue and carrier flapping")
Signed-off-by: Juergen Gross <[email protected]>
Reviewed-by: Jan Beulich <[email protected]>
---
 drivers/net/xen-netback/rx.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)

diff --git a/drivers/net/xen-netback/rx.c b/drivers/net/xen-netback/rx.c
index a8511e27d6c1..dbac4c03d21a 100644
--- a/drivers/net/xen-netback/rx.c
+++ b/drivers/net/xen-netback/rx.c
@@ -88,16 +88,19 @@ void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
 
 	spin_lock_irqsave(&queue->rx_queue.lock, flags);
 
-	if (skb_queue_empty(&queue->rx_queue))
-		xenvif_update_needed_slots(queue, skb);
-
-	__skb_queue_tail(&queue->rx_queue, skb);
-
-	queue->rx_queue_len += skb->len;
-	if (queue->rx_queue_len > queue->rx_queue_max) {
+	if (queue->rx_queue_len >= queue->rx_queue_max) {
 		struct net_device *dev = queue->vif->dev;
 
 		netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
+		kfree_skb(skb);
+		queue->vif->dev->stats.rx_dropped++;
+	} else {
+		if (skb_queue_empty(&queue->rx_queue))
+			xenvif_update_needed_slots(queue, skb);
+
+		__skb_queue_tail(&queue->rx_queue, skb);
+
+		queue->rx_queue_len += skb->len;
 	}
 
 	spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
@@ -147,6 +150,7 @@ static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
 			break;
 		xenvif_rx_dequeue(queue);
 		kfree_skb(skb);
+		queue->vif->dev->stats.rx_dropped++;
 	}
 }
 
-- 
2.26.2