Re: [PATCH net 1/3] net: macb: never give hardware a NULL RX buffer

[Théo Lebrun]

On Thu Sep 24, 2026 at 5:41 PM CEST, Jakub Kicinski wrote:
> On Thu, 24 Sep 2026 17:13:58 +0200 Théo Lebrun wrote:
>> > If the AI interprets the rx poll side correctly these seem like a major
>> > flaw in setting the USED bit :(  
>> 
>> Yes the AI is correct (unsurprisingly). But two things:
>
> If it's correct then patch needs to be changed.
> Please don't write puzzles for me to solve.
>
>>  - before this patch what happened was that HW read the NULL descriptor,
>>    so not much better
>>  - we can detect in gem_rx() using our ring buffer cursors that tell us
>>    what we've managed to allocate, that we gem_rx() doesn't iterate
>>    over descriptors that haven't been allocated
>
> Grammatically dubious puzzles at that.

Sorry for the fuzzy message.
Yes, Sashiko is correct and, yes, the patch will change.

To explain more in detail: gem_rx() iterates over descriptors using
RX_USED as signal that a descriptor was filled by the hardware. The
risk is to reach a descriptor where allocation failed (and we used
RX_USED as signal for the hardware to not fill it up). We can avoid
that using our cursors (queue->rx_tail and queue->rx_prepared_head)
that tell us the range of allocated descriptors.

In concrete terms, we stop iterating in gem_rx() if

	queue->rx_tail == queue->rx_prepared_head

or if

	desc->addr & MACB_BIT(RX_USED)

Thanks,

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