[PATCH net 1/3] igb: check __IGB_DOWN in igb_clean_rx_irq_zc()

[PATCH net 1/3] igb: check __IGB_DOWN in igb_clean_rx_irq_zc()

[Alex Dvoretsky]

When an AF_XDP zero-copy application terminates abruptly (e.g.,
kill -9), the XSK buffer pool is destroyed but NAPI polling continues.
igb_clean_rx_irq_zc() keeps returning budget (no descriptors, no
buffers to allocate, xsk_buff_alloc() returns NULL) which makes
napi_complete_done() re-arm the poll indefinitely.

Meanwhile, igb_down() → napi_synchronize() waits for a NAPI poll cycle
that signals completion with done < budget — which never happens. This
blocks igb_down() forever, and the 5-second TX watchdog fires because
no TX completions are processed while NAPI is stuck. Since igb_down()
never finishes, igb_up() is never called, and the TX queue remains
permanently stalled.

Fix this by adding an __IGB_DOWN check at the top of
igb_clean_rx_irq_zc(), returning 0 immediately when the adapter is
going down. This allows napi_synchronize() in igb_down() to complete,
matching the pattern already used in igb_clean_tx_irq().

Fixes: 2c6196013f84 ("igb: Add AF_XDP zero-copy Rx support")
Cc: stable@vger.kernel.org
Signed-off-by: Alex Dvoretsky <advoretsky@gmail.com>
---
 drivers/net/ethernet/intel/igb/igb_xsk.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/net/ethernet/intel/igb/igb_xsk.c b/drivers/net/ethernet/intel/igb/igb_xsk.c
index 30ce5fbb5b77..ca4aa4d935d5 100644
--- a/drivers/net/ethernet/intel/igb/igb_xsk.c
+++ b/drivers/net/ethernet/intel/igb/igb_xsk.c
@@ -351,6 +351,9 @@ int igb_clean_rx_irq_zc(struct igb_q_vector *q_vector,
 	u16 entries_to_alloc;
 	struct sk_buff *skb;
 
+	if (test_bit(__IGB_DOWN, &adapter->state))
+		return 0;
+
 	/* xdp_prog cannot be NULL in the ZC path */
 	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
 
-- 
2.51.0

[PATCH net 2/3] igb: skip reset in igb_tx_timeout() during XDP transition

[Alex Dvoretsky]

When igb_xdp_setup() transitions between XDP and non-XDP mode on a
running device, it calls igb_close() followed by igb_open(). During
this window the adapter is down and trans_start is stale, so the TX
watchdog can fire a spurious timeout.

The resulting schedule_work(&adapter->reset_task) races with the
igb_open() path: the reset task may run while the device is being
brought back up, or immediately after, causing unexpected ring
reinitialisation and register writes.

Fix this by checking __IGB_DOWN at the top of igb_tx_timeout(). If the
adapter is down (either during normal close or during the XDP close/open
transition), there is nothing useful a reset can do — the subsequent
igb_open() will reinitialise everything.

Fixes: 9cbc948b5a20 ("igb: add XDP support")
Cc: stable@vger.kernel.org
Signed-off-by: Alex Dvoretsky <advoretsky@gmail.com>
---
 drivers/net/ethernet/intel/igb/igb_main.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c
index 223a10cae4a9..ddb7ce9e97bf 100644
--- a/drivers/net/ethernet/intel/igb/igb_main.c
+++ b/drivers/net/ethernet/intel/igb/igb_main.c
@@ -6651,6 +6651,15 @@ static void igb_tx_timeout(struct net_device *netdev, unsigned int __always_unus
 	struct igb_adapter *adapter = netdev_priv(netdev);
 	struct e1000_hw *hw = &adapter->hw;
 
+	/* Do not schedule a reset if the adapter is already going down or
+	 * being reconfigured (e.g., XDP program transition via igb_close/
+	 * igb_open). The stale trans_start from before the close will
+	 * trigger a spurious timeout that resolves once igb_open()
+	 * completes.
+	 */
+	if (test_bit(__IGB_DOWN, &adapter->state))
+		return;
+
 	/* Do the reset outside of interrupt context */
 	adapter->tx_timeout_count++;
 
-- 
2.51.0

[PATCH net 3/3] igb: add XDP transition guards in igb_xdp_setup()

[Alex Dvoretsky]

igb_xdp_setup() calls igb_close() + igb_open() when transitioning
between XDP and non-XDP mode on a running device. This has two issues:

1. When removing an XDP program that has AF_XDP zero-copy sockets,
   ndo_xsk_wakeup() may be executing concurrently under rcu_read_lock().
   If igb_close() tears down the rings while ndo_xsk_wakeup() is still
   accessing them, it races with the teardown. Add synchronize_rcu()
   before igb_close() when removing an XDP program to ensure all
   in-flight ndo_xsk_wakeup() calls complete first.

2. The igb_close()/igb_open() window leaves trans_start stale from
   before the close: the TX watchdog can fire a spurious timeout and
   queue a reset_task that races with igb_open(). Add
   netif_trans_update() after igb_open() to refresh the timestamp, and
   cancel_work() to drain any reset_task queued during the window.

Fixes: 9cbc948b5a20 ("igb: add XDP support")
Cc: stable@vger.kernel.org
Signed-off-by: Alex Dvoretsky <advoretsky@gmail.com>
---
 drivers/net/ethernet/intel/igb/igb_main.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c
index ddb7ce9e97bf..9ba944bf67b4 100644
--- a/drivers/net/ethernet/intel/igb/igb_main.c
+++ b/drivers/net/ethernet/intel/igb/igb_main.c
@@ -2913,6 +2913,9 @@ static int igb_xdp_setup(struct net_device *dev, struct netdev_bpf *bpf)
 
 	/* device is up and bpf is added/removed, must setup the RX queues */
 	if (need_reset && running) {
+		if (!prog)
+			/* Wait until ndo_xsk_wakeup completes. */
+			synchronize_rcu();
 		igb_close(dev);
 	} else {
 		for (i = 0; i < adapter->num_rx_queues; i++)
@@ -2936,6 +2939,16 @@ static int igb_xdp_setup(struct net_device *dev, struct netdev_bpf *bpf)
 	if (running)
 		igb_open(dev);
 
+	/* Refresh trans_start to prevent the TX watchdog from firing on a
+	 * stale timestamp from before igb_close(). Cancel any reset_task
+	 * that igb_tx_timeout() may have queued between igb_close() setting
+	 * __IGB_DOWN and the actual napi_synchronize() completion.
+	 */
+	if (need_reset && running) {
+		netif_trans_update(dev);
+		cancel_work(&adapter->reset_task);
+	}
+
 	return 0;
 }
 
-- 
2.51.0

[PATCH net 2/3] igb: skip reset in igb_tx_timeout() during XDP transition

[Alex Dvoretsky]

When igb_xdp_setup() transitions between XDP and non-XDP mode on a
running device, it calls igb_close() followed by igb_open(). During
this window the adapter is down while trans_start still contains the
timestamp from before igb_close(), so the TX watchdog can fire a
spurious timeout.

The resulting schedule_work(&adapter->reset_task) races with the
igb_open() path: the reset task may run while the device is being
brought back up, or immediately after, causing unexpected ring
reinitialisation and register writes.

Fix this by checking __IGB_DOWN at the top of igb_tx_timeout(). A
reset is unnecessary because the device will be fully reinitialised
by the subsequent igb_open().

Fixes: 9cbc948b5a20 ("igb: add XDP support")
Cc: stable@vger.kernel.org
Signed-off-by: Alex Dvoretsky <advoretsky@gmail.com>
---
 drivers/net/ethernet/intel/igb/igb_main.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c
index 223a10cae4a9..ddb7ce9e97bf 100644
--- a/drivers/net/ethernet/intel/igb/igb_main.c
+++ b/drivers/net/ethernet/intel/igb/igb_main.c
@@ -6651,6 +6651,10 @@ static void igb_tx_timeout(struct net_device *netdev, unsigned int __always_unus
 	struct igb_adapter *adapter = netdev_priv(netdev);
 	struct e1000_hw *hw = &adapter->hw;
 
+	/* Ignore timeout if the adapter is going down. */
+	if (test_bit(__IGB_DOWN, &adapter->state))
+		return;
+
 	/* Do the reset outside of interrupt context */
 	adapter->tx_timeout_count++;
 
-- 
2.51.0