On Thu Jul 2, 2026 at 12:37 PM CEST, Nicolai Buchwitz wrote:

On 1.7.2026 17:59, Théo Lebrun wrote:

quoted

ethtool_ops.set_ringparam() is implemented using the primitive close /
update ring size / reopen sequence. Under memory pressure this does not
fly: we free our buffers at close and cannot reallocate new ones at
open. Also, it triggers a slow PHY reinit.

Instead, exploit the new context mechanism and improve our sequence to:
 - allocate a new context (including buffers) first
 - if it fails, early return without any impact to the interface
 - stop interface
 - update global state (bp, netdev, etc)
 - pass buffer pointers to the hardware
 - start interface
 - free old context.

The HW disable sequence is inspired by macb_reset_hw() but avoids
(1) setting NCR bit CLRSTAT and (2) clearing register PBUFRXCUT.

The HW re-enable sequence is inspired by macb_mac_link_up(), skipping
over register writes which would be redundant (because values have not
changed).

The generic context swapping parts are isolated into helper functions
macb_context_swap_start|end(), reusable by other operations 
(change_mtu,
set_channels, etc).

Introduce a new locking primitive (mac_cfg_lock mutex) to serialise 
swap
with phylink MAC callbacks. Avoid stopping phylink to avoid a slow PHY
retrain. Those callbacks grab phydev->lock if it exists so we could
imagine grabbing that from the swap op, but phydev->lock doesn't exist
in the SFP case.

AT91 EMAC is handled differently as their buffer management is separate
and they don't do NAPI. We refuse them (-EBUSY) to avoid implementing
context swapping for them.

Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
---
 drivers/net/ethernet/cadence/macb.h      |   2 +
 drivers/net/ethernet/cadence/macb_main.c | 142 
+++++++++++++++++++++++++++++--

quoted

[...]

quoted

+static void macb_context_swap_start(struct macb *bp)
+{
+	struct macb_queue *queue;
+	unsigned long flags;
+	unsigned int q;
+	u32 ctrl;
+
+	mutex_lock(&bp->mac_cfg_lock);
+
+	/* Mask interrupts before disabling BH features. */
+	spin_lock_irqsave(&bp->lock, flags);
+	for (q = 0, queue = bp->queues; q < bp->num_queues; ++q, ++queue) {
+		queue_writel(queue, IDR, -1);
+		queue_readl(queue, ISR);
+		macb_queue_isr_clear(bp, queue, -1);
+	}
+	spin_unlock_irqrestore(&bp->lock, flags);
+
+	/* Drain BH features. HW is still active and usable at this point. */
+
+	cancel_work_sync(&bp->hresp_err_bh_work);
+	cancel_delayed_work_sync(&bp->tx_lpi_work);
+
+	for (q = 0, queue = bp->queues; q < bp->num_queues; ++q, ++queue) {
+		napi_disable(&queue->napi_rx);
+		napi_disable(&queue->napi_tx);
+		cancel_work_sync(&queue->tx_error_task);
+		netdev_tx_reset_queue(netdev_get_tx_queue(bp->netdev, q));
+	}

Can this deadlock against a pending tx_error_task?

AFAIU macb_tx_error_task() does napi_disable(&queue->napi_tx) and later
napi_enable() on the same napi, and it can already be queued 
(macb_interrupt()
schedules it on a TX error) by the time the swap runs:

   swap_start: napi_disable(napi_tx)          /* sets SCHED, returns */
   worker:     tx_error_task: napi_disable(napi_tx)  /* spins on SCHED */
   swap_start: cancel_work_sync(tx_error_task)       /* waits on worker 
*/

napi_disable() spins until napi_enable() clears SCHED, but here the swap 
won't
re-enable until macb_context_swap_end(), and cancel_work_sync() is 
what's
holding it up. Nothing clears it.

Maybe cancel_work_sync() before the napi_disable() calls would work 
instead? IRQs
are masked just above, so AFAICT nothing can reschedule tx_error_task by 
then.

Ah yes, good catch. macb_tx_error_task() should never enter if NAPI is
disabled, else its napi_disable() will hang.

The fix sounds easy enough, as you indicated. We'll move the
cancel_work_sync(queue->tx_error_task) call to be above our swap start
napi_disable().

Thanks,

--
Théo Lebrun, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com