Re: [PATCH net-next v6 1/2] net: macb: implement ethtool_ops.get|set_channels()

From: Théo Lebrun

Date: Tue Nov 17 2026 - 05:20:41 EST


Hello Nicolai,

On Sun Nov 15, 2026 at 8:39 PM CET, Nicolai Buchwitz wrote:
> On 13.11.2026 18:32, Théo Lebrun wrote:
>> @@ -3351,16 +3374,21 @@ static void macb_context_swap_end(struct macb
>> *bp,
>>
>> lockdep_assert_held(&bp->mac_cfg_lock);
>>
>> - /* Swap contexts & give buffer pointers to HW. */
>> + /* Swap contexts & give buffer pointers to HW.
>> + * !new_ctx means that the swap failed and we keep the old context.
>> + */
>> +
>> + if (new_ctx) {
>> + old_ctx = bp->ctx;
>> + WRITE_ONCE(bp->ctx, new_ctx);
>> + wmb(); /* ensure IRQ enabled below read the new context */
>> + }
>>
>> - old_ctx = bp->ctx;
>> - WRITE_ONCE(bp->ctx, new_ctx);
>> - wmb(); /* ensure IRQ enabled below read the new context */
>> macb_init_buffers(bp);
>
> The NULL case reuses the old rings without mog_init_rings(). The HW
> restarts at descriptor 0 but head/tail don't, so Tx stalls after a
> failed ethtool -L? macb_hresp_error_task() reinits the rings first.

Hum, so you are totally right on that. Both Tx (txq->head + txq->tail)
and Rx (rxq->tail + rxq->prepared_head) cursors are wrong by being out
of sync with the hardware cursors.

But the approach taken by macb_hresp_error_task() is wrong: it resets
the Tx cursors and Tx descriptors in the existing context but by doing
so leak all Tx buffers. Rx buffers aren't leaking because
gem_init_rx_ring() sets RX_USED on all slots then refills.

Sashiko [0] points out the same issue; I'll wait for the end of the
Clashiko [1] embargo period before re-sending, to see what it has to
say.

---

There can be two solutions to this:

- We don't touch the HW cursors RBQPH/TBQPH/RBQP/TBQP by not calling
macb_init_buffers() at swap_end. So here the swap operation will
have been as close as possible to a no-op where we just stopped the
HW for a small window.

- We do reset the HW cursors. But that means we need to make sure both
rings are in a good shape (which they are not with v6).

It isn't a massive amount of work, but it still is more code (that
might be buggy). We need to free pending Tx slots or bring them to
the front of the ring. For Rx we should consume RX_USED slots.

I prefer the first approach. It is simpler and alloc-free (ENOMEM is
probably the reason why the set_channels context swap failed when
calling netif_set_real_num_queues).

[0]: https://sashiko.dev/#/patchset/20261113-macb-set-channels-v6-0-5de59d9e7f02%40bootlin.com?part=1
[1]: https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261113-macb-set-channels-v6-0-5de59d9e7f02%40bootlin.com

Thanks,

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