net: prevent torn reads in netdev_tc_txq

netdev_set_tc_queue() (and related helpers/drivers such as
netdev_bind_sb_channel_queue(), netdev_reset_tc(), and
netdev_unbind_sb_channel()) perform separate 16-bit writes to
dev->tc_to_txq[tc].count and dev->tc_to_txq[tc].offset.

Furthermore, memset() in netdev_reset_tc() and
netdev_unbind_sb_channel() provides no guarantee of performing
full 32-bit word stores.

Concurrent lockless readers (e.g. skb_tx_hash(), netdev_txq_to_tc(),
ixgbe_select_queue(), taprio, mqprio, FPE drivers) can observe torn
values where offset and count belong to inconsistent configurations.

Redefine struct netdev_tc_txq to embed count and offset inside a union
with a u32 combined field, allowing atomic manipulation via
READ_ONCE() and WRITE_ONCE().

Update all lockless readers and writers across the kernel to use
READ_ONCE() and WRITE_ONCE() on the combined field.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260812085440.3917924-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

4 files changed, 56 insertions, 27 deletions

diff --git a/net/core/dev.c b/net/core/dev.c
index 517ac6a575c4..694a5ab773f7 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -2652,11 +2652,13 @@ EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
  */
 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
 {
+	struct netdev_tc_txq res;
 	int i;
-	struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
+
+	res.combined = READ_ONCE(dev->tc_to_txq[0].combined);
 
 	/* If TC0 is invalidated disable TC mapping */
-	if (tc->offset + tc->count > txq) {
+	if (res.offset + res.count > txq) {
 		netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
 		dev->num_tc = 0;
 		return;
@@ -2666,8 +2668,8 @@ static void netif_setup_tc(struct net_device *dev, unsigned int txq)
 	for (i = 1; i < TC_BITMASK + 1; i++) {
 		int q = netdev_get_prio_tc_map(dev, i);
 
-		tc = &dev->tc_to_txq[q];
-		if (tc->offset + tc->count > txq) {
+		res.combined = READ_ONCE(dev->tc_to_txq[q].combined);
+		if (res.offset + res.count > txq) {
 			netdev_warn(dev, "Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
 				    i, q);
 			netdev_set_prio_tc_map(dev, i, 0);
@@ -2683,7 +2685,10 @@ int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
 
 		/* walk through the TCs and see if it falls into any of them */
 		for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
-			if ((txq - tc->offset) < tc->count)
+			struct netdev_tc_txq res;
+
+			res.combined = READ_ONCE(tc->combined);
+			if ((txq - res.offset) < res.count)
 				return i;
 		}
 
@@ -3103,6 +3108,8 @@ static void netdev_unbind_all_sb_channels(struct net_device *dev)
 
 void netdev_reset_tc(struct net_device *dev)
 {
+	int i;
+
 #ifdef CONFIG_XPS
 	netif_reset_xps_queues_gt(dev, 0);
 #endif
@@ -3110,21 +3117,26 @@ void netdev_reset_tc(struct net_device *dev)
 
 	/* Reset TC configuration of device */
 	dev->num_tc = 0;
-	memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
+	for (i = 0; i < TC_MAX_QUEUE; i++)
+		WRITE_ONCE(dev->tc_to_txq[i].combined, 0);
 	memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
 }
 EXPORT_SYMBOL(netdev_reset_tc);
 
 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
 {
+	struct netdev_tc_txq res = {
+		.count = count,
+		.offset = offset,
+	};
+
 	if (tc >= dev->num_tc)
 		return -EINVAL;
 
 #ifdef CONFIG_XPS
 	netif_reset_xps_queues(dev, offset, count);
 #endif
-	dev->tc_to_txq[tc].count = count;
-	dev->tc_to_txq[tc].offset = offset;
+	WRITE_ONCE(dev->tc_to_txq[tc].combined, res.combined);
 	return 0;
 }
 EXPORT_SYMBOL(netdev_set_tc_queue);
@@ -3148,11 +3160,13 @@ void netdev_unbind_sb_channel(struct net_device *dev,
 			      struct net_device *sb_dev)
 {
 	struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
+	int i;
 
 #ifdef CONFIG_XPS
 	netif_reset_xps_queues_gt(sb_dev, 0);
 #endif
-	memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
+	for (i = 0; i < TC_MAX_QUEUE; i++)
+		WRITE_ONCE(sb_dev->tc_to_txq[i].combined, 0);
 	memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
 
 	while (txq-- != &dev->_tx[0]) {
@@ -3175,8 +3189,12 @@ int netdev_bind_sb_channel_queue(struct net_device *dev,
 		return -EINVAL;
 
 	/* Record the mapping */
-	sb_dev->tc_to_txq[tc].count = count;
-	sb_dev->tc_to_txq[tc].offset = offset;
+	struct netdev_tc_txq res = {
+		.count = count,
+		.offset = offset,
+	};
+
+	WRITE_ONCE(sb_dev->tc_to_txq[tc].combined, res.combined);
 
 	/* Provide a way for Tx queue to find the tc_to_txq map or
 	 * XPS map for itself.
@@ -3542,9 +3560,11 @@ static u16 skb_tx_hash(const struct net_device *dev,
 
 	if (dev->num_tc) {
 		u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
+		struct netdev_tc_txq res;
 
-		qoffset = sb_dev->tc_to_txq[tc].offset;
-		qcount = sb_dev->tc_to_txq[tc].count;
+		res.combined = READ_ONCE(sb_dev->tc_to_txq[tc].combined);
+		qoffset = res.offset;
+		qcount = res.count;
 		if (unlikely(!qcount)) {
 			net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n",
 					     sb_dev->name, qoffset, tc);
diff --git a/net/sched/sch_mqprio.c b/net/sched/sch_mqprio.c
index ae991fc25b43..6ced7008ef5c 100644
--- a/net/sched/sch_mqprio.c
+++ b/net/sched/sch_mqprio.c
@@ -679,12 +679,14 @@ static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 	rcu_read_lock();
 	if (cl >= TC_H_MIN_PRIORITY) {
 		struct net_device *dev = qdisc_dev(sch);
-		struct netdev_tc_txq tc = dev->tc_to_txq[cl & TC_BITMASK];
+		struct netdev_tc_txq tc;
 		struct gnet_stats_queue qstats = {0};
 		struct gnet_stats_basic_sync bstats;
 		u32 qlen = 0;
 		int i;
 
+		tc.combined = READ_ONCE(dev->tc_to_txq[cl & TC_BITMASK].combined);
+
 		gnet_stats_basic_sync_init(&bstats);
 
 		for (i = tc.offset; i < tc.offset + tc.count; i++) {
diff --git a/net/sched/sch_mqprio_lib.c b/net/sched/sch_mqprio_lib.c
index b3a5572c167b..b60e130c7078 100644
--- a/net/sched/sch_mqprio_lib.c
+++ b/net/sched/sch_mqprio_lib.c
@@ -108,8 +108,11 @@ void mqprio_qopt_reconstruct(struct net_device *dev, struct tc_mqprio_qopt *qopt
 	memcpy(qopt->prio_tc_map, dev->prio_tc_map, sizeof(qopt->prio_tc_map));
 
 	for (tc = 0; tc < num_tc; tc++) {
-		qopt->count[tc] = dev->tc_to_txq[tc].count;
-		qopt->offset[tc] = dev->tc_to_txq[tc].offset;
+		struct netdev_tc_txq res;
+
+		res.combined = READ_ONCE(dev->tc_to_txq[tc].combined);
+		qopt->count[tc] = res.count;
+		qopt->offset[tc] = res.offset;
 	}
 }
 EXPORT_SYMBOL_GPL(mqprio_qopt_reconstruct);
diff --git a/net/sched/sch_taprio.c b/net/sched/sch_taprio.c
index 299234a5f0fe..7d5fe93a4c12 100644
--- a/net/sched/sch_taprio.c
+++ b/net/sched/sch_taprio.c
@@ -762,12 +762,13 @@ skip_peek_checks:
 
 static void taprio_next_tc_txq(struct net_device *dev, int tc, int *txq)
 {
-	int offset = dev->tc_to_txq[tc].offset;
-	int count = dev->tc_to_txq[tc].count;
+	struct netdev_tc_txq res;
+
+	res.combined = READ_ONCE(dev->tc_to_txq[tc].combined);
 
 	(*txq)++;
-	if (*txq == offset + count)
-		*txq = offset;
+	if (*txq == res.offset + res.count)
+		*txq = res.offset;
 }
 
 /* Prioritize higher traffic classes, and select among TXQs belonging to the
@@ -1441,15 +1442,14 @@ static u32 tc_map_to_queue_mask(struct net_device *dev, u32 tc_mask)
 	u32 i, queue_mask = 0;
 
 	for (i = 0; i < dev->num_tc; i++) {
-		u32 offset, count;
+		struct netdev_tc_txq res;
 
 		if (!(tc_mask & BIT(i)))
 			continue;
 
-		offset = dev->tc_to_txq[i].offset;
-		count = dev->tc_to_txq[i].count;
+		res.combined = READ_ONCE(dev->tc_to_txq[i].combined);
 
-		queue_mask |= GENMASK(offset + count - 1, offset);
+		queue_mask |= GENMASK(res.offset + res.count - 1, res.offset);
 	}
 
 	return queue_mask;
@@ -1802,10 +1802,14 @@ static int taprio_mqprio_cmp(const struct net_device *dev,
 	if (!mqprio || mqprio->num_tc != dev->num_tc)
 		return -1;
 
-	for (i = 0; i < mqprio->num_tc; i++)
-		if (dev->tc_to_txq[i].count != mqprio->count[i] ||
-		    dev->tc_to_txq[i].offset != mqprio->offset[i])
+	for (i = 0; i < mqprio->num_tc; i++) {
+		struct netdev_tc_txq res;
+
+		res.combined = READ_ONCE(dev->tc_to_txq[i].combined);
+		if (res.count != mqprio->count[i] ||
+		    res.offset != mqprio->offset[i])
 			return -1;
+	}
 
 	for (i = 0; i <= TC_BITMASK; i++)
 		if (dev->prio_tc_map[i] != mqprio->prio_tc_map[i])