[PATCH v3 0/2] can: enable NAPI before starting the controller in two drivers

[PATCH v3 0/2] can: enable NAPI before starting the controller in two drivers

[Patrick Zhang]

Hi,

ctucanfd and xilinx_can open the device in the order "start the
controller and unmask its RX interrupt, then napi_enable()". Each of
them has an ISR that masks the RX interrupt before calling
napi_schedule(), and a poll that is the only place the RX interrupt is
unmasked again. A frame that arrives between the controller start and
napi_enable() is handled by the ISR, which masks RX and calls
napi_schedule() on a NAPI instance that is still disabled.
napi_schedule_prep() sees NAPI_STATE_SCHED set and drops the schedule.
napi_enable() clears SCHED without scheduling. Nothing ever polls, so
RX stays masked and the interface receives no data frames until the
next close. Transmit keeps working. ctucanfd counts no errors, so the
failure is silent there. xilinx_can counts RX FIFO overflows once the
FIFO fills.

The fix is the same in both: call napi_enable() before the controller
start, so that an early RX interrupt can schedule the poll, and call
napi_disable() on the error path if the start fails.

Patch 1 (ctucanfd) was tested on QEMU with the ctucan_pci device.
Patch 2 (xilinx_can) was tested on a Versal board.

The bug has been present since each driver was added, so each patch
has a Fixes: tag pointing at the commit that added the driver, and both
are marked for stable backport. The patches are independent and can be
applied separately.

I looked at the other NAPI users under drivers/net/can. at91_can had
the same persistent-mask pattern in 6.6 but no longer masks RX in its
ISR since the conversion to rx-offload, so I left it alone; ti_hecc and
virtio_can start the device before napi_enable() but do not leave RX
masked, so a dropped first schedule only delays the first frame.
ifi_canfd has the same race; see the changes in v3. All the others
enable NAPI before the device can interrupt.

Changes in v3:
- Drop the ifi_canfd patch. The netdev AI review of v2 showed that a
  safe fix has to reorder the reset, error counter and interrupt mask
  writes in ifi_canfd_start(), on the open and the bus-off restart
  paths. That change is no longer obviously correct by inspection, and
  I cannot test it without IFI CAN FD hardware. The race in
  ifi_canfd_open() remains.
- Patch 1: no changes
- Patch 2: add Reviewed-by from Harini
- v2: https://lore.kernel.org/r/20261030233330.3483988-1-patrzhan@amazon.com

Changes in v2:
- Patch 1: add Acked-by and Tested-by from Pavel
- Patch 2: correct the commit message, which said no error counter
  moves during the stall; the RX overflow interrupt stays enabled and
  counts rx_over_errors
- Patch 3: split the reset out of ifi_canfd_start() and enable NAPI
  after it, with synchronize_irq() in between, so the poll cannot race
  with the reset (reported by the netdev AI review). This replaces the
  ifi_canfd_stop() call proposed in my reply to that review; see the
  patch notes
- Patches 2 and 3: name the related fixes by subject
- Rebased on current linux-can main
- v1: https://lore.kernel.org/r/20261021205924.2538511-1-patrzhan@amazon.com

Patrick Zhang (2):
  can: ctucanfd: enable NAPI before starting the controller
  can: xilinx_can: enable NAPI before starting the controller

 drivers/net/can/ctucanfd/ctucanfd_base.c | 6 +++++-
 drivers/net/can/xilinx_can.c             | 6 +++++-
 2 files changed, 10 insertions(+), 2 deletions(-)


base-commit: abeacca61eea4f6803df4a595b656ea45c2d3249
-- 
2.50.1

[PATCH v3 1/2] can: ctucanfd: enable NAPI before starting the controller

[Patrick Zhang]

ctucan_open() starts the controller and unmasks RBNEI in
ctucan_chip_start(), prints "ctu_can_fd device registered", and only
then calls napi_enable(). A frame that arrives in that window raises
RBNEI. ctucan_interrupt() masks RBNEI and calls napi_schedule(). NAPI is
still disabled, so napi_schedule_prep() finds NAPI_STATE_SCHED set and
the schedule is dropped. napi_enable() then clears SCHED without
scheduling anything. Only ctucan_rx_poll() unmasks RBNEI again, and it
never runs. The interface receives nothing until the next close.
Transmit still works. The RX FIFO fills up and the core sets DOR, but
the overrun is only accounted in ctucan_rx_poll(), so rx_errors and
rx_over_errors stay at zero and nothing is logged. Every re-open (ip
link down/up, a driver rebind) runs the same race again.

The netdev_info() in the window makes it much wider on a slow console.
This can be reproduced by sending continuous traffic with cangen from
another node and repeating "ip link set down/up" on the test device.

Enable NAPI before the controller starts, so that an early RX interrupt
can schedule the poll. NAPI can be enabled while the controller is in
reset, because nothing schedules it until an interrupt arrives. If
ctucan_chip_start() fails, disable NAPI again. Otherwise the next open
would call napi_enable() on an already enabled instance and trip the
BUG_ON() in napi_enable_locked().

Tested on QEMU with the ctucan_pci device.

Fixes: 2dcb8e8782d8 ("can: ctucanfd: add support for CTU CAN FD open-source IP core - bus independent part.")
Cc: stable@vger.kernel.org
Assisted-by: LLM
Signed-off-by: Patrick Zhang <patrzhan@amazon.com>
Acked-by: Pavel Pisa <pisa@fel.cvut.cz>
Tested-by: Pavel Pisa <pisa@fel.cvut.cz>
---

Notes:
    v3:
    - No changes
    - v2: https://lore.kernel.org/r/20261030233330.3483988-2-patrzhan@amazon.com
    
    v2:
    - Add Acked-by and Tested-by from Pavel
    - v1: https://lore.kernel.org/r/20261021205924.2538511-2-patrzhan@amazon.com

 drivers/net/can/ctucanfd/ctucanfd_base.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/net/can/ctucanfd/ctucanfd_base.c b/drivers/net/can/ctucanfd/ctucanfd_base.c
index 07d4aa43c..82bf6c645 100644
--- a/drivers/net/can/ctucanfd/ctucanfd_base.c
+++ b/drivers/net/can/ctucanfd/ctucanfd_base.c
@@ -1231,6 +1231,10 @@ static int ctucan_open(struct net_device *ndev)
 		goto err_irq;
 	}
 
+	/* Keep before ctucan_chip_start(): the ISR masks RBNEI and only the
+	 * poll unmasks it.
+	 */
+	napi_enable(&priv->napi);
 	ret = ctucan_chip_start(ndev);
 	if (ret < 0) {
 		netdev_err(ndev, "ctucan_chip_start failed!\n");
@@ -1238,12 +1242,12 @@ static int ctucan_open(struct net_device *ndev)
 	}
 
 	netdev_info(ndev, "ctu_can_fd device registered\n");
-	napi_enable(&priv->napi);
 	netif_start_queue(ndev);
 
 	return 0;
 
 err_chip_start:
+	napi_disable(&priv->napi);
 	free_irq(ndev->irq, ndev);
 err_irq:
 	close_candev(ndev);
-- 
2.50.1

[PATCH v3 2/2] can: xilinx_can: enable NAPI before starting the controller

[Patrick Zhang]

xcan_open() calls xcan_chip_start(), which writes XCAN_IER_OFFSET to
unmask the RX interrupts and then sets XCAN_SRR_CEN_MASK to enable the
controller, and only then calls napi_enable(). A frame that arrives in
that window raises the RX interrupt. xcan_interrupt() clears the RX
bits in XCAN_IER_OFFSET and calls napi_schedule(). NAPI is still
disabled, so napi_schedule_prep() finds NAPI_STATE_SCHED set and the
schedule is dropped. Only xcan_rx_poll() unmasks RX again, and it never
runs. The interface receives no data frames until the next close.
Transmit still works. The RX overflow interrupt stays enabled, so once
the RX FIFO fills, xcan_err_interrupt() counts rx_over_errors and
reports overflow error frames, but nothing reads the FIFO. Every
re-open (ip link down/up, a driver rebind) runs the same race again.

The window is only the tail of xcan_chip_start(), so the failure is
rare. Adding an mdelay(50) at the end of xcan_chip_start() makes it
reproducible with continuous traffic from another node and a single
"ip link set down/up" on the test device.

Enable NAPI before the controller starts, so that an early RX interrupt
can schedule the poll. NAPI can be enabled while the controller is in
reset, because nothing schedules it until an interrupt arrives. If
xcan_chip_start() fails, disable NAPI again. Otherwise the next open
would call napi_enable() on an already enabled instance and trip the
BUG_ON() in napi_enable_locked().

This is the same race as in "can: ctucanfd: enable NAPI before starting
the controller".

Tested on a Versal board.

Fixes: b1201e44f50b ("can: xilinx CAN controller support")
Cc: stable@vger.kernel.org
Assisted-by: LLM
Signed-off-by: Patrick Zhang <patrzhan@amazon.com>
Reviewed-by: Harini T <harini.t@amd.com>
---

Notes:
    v3:
    - Add Reviewed-by from Harini
    - v2: https://lore.kernel.org/r/20261030233330.3483988-3-patrzhan@amazon.com
    
    v2:
    - Correct the commit message: the RX overflow interrupt stays enabled
      during the stall, so rx_over_errors is counted once the FIFO fills
    - Name the ctucanfd fix by subject instead of "the previous patch"
    - v1: https://lore.kernel.org/r/20261021205924.2538511-3-patrzhan@amazon.com

 drivers/net/can/xilinx_can.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/net/can/xilinx_can.c b/drivers/net/can/xilinx_can.c
index 43d7f2282..5c2412bf9 100644
--- a/drivers/net/can/xilinx_can.c
+++ b/drivers/net/can/xilinx_can.c
@@ -1570,18 +1570,22 @@ static int xcan_open(struct net_device *ndev)
 	if (ret)
 		goto err_irq;
 
+	/* Keep before xcan_chip_start(): the ISR masks RX interrupts and only
+	 * the poll unmasks them.
+	 */
+	napi_enable(&priv->napi);
 	ret = xcan_chip_start(ndev);
 	if (ret < 0) {
 		netdev_err(ndev, "xcan_chip_start failed!\n");
 		goto err_candev;
 	}
 
-	napi_enable(&priv->napi);
 	netif_start_queue(ndev);
 
 	return 0;
 
 err_candev:
+	napi_disable(&priv->napi);
 	close_candev(ndev);
 err_irq:
 	free_irq(ndev->irq, ndev);
-- 
2.50.1