Re: [PATCH v2 4/4] bgmac: fix DMA rx corruption

[Felix Fietkau <nbd@openwrt.org>]

On 2015-04-12 21:04, Eric Dumazet wrote:
> On Sun, 2015-04-12 at 20:46 +0200, Felix Fietkau wrote:
>> On 2015-04-12 20:09, Eric Dumazet wrote:
> 
>> > 
>> > Given the fact that previous driver kind of worked, have you tried not
>> > doing the modulo at all here ?
>> The previous driver was causing reproducible DMA corruption issues under
>> load. My patch fixes this, and I've verified this with hours of stress
>> testing (previously I could reproduce the issue in minutes).
>> 
>> > ring->end = desc_idx + 1;
>> > 
>> > It looks like the hardware accepted setting BGMAC_DMA_RX_INDEX to
>> > ring->index_base + ring->num_slots * sizeof(struct bgmac_dma_desc)
>> The hardware accepts it, but with that value, it considers all rx slot
>> entries filled, there's no CPU ownership bit in the descriptor.
>> According to documentation, that register indicates the first *invalid*
>> rx descriptor (i.e. one where the CPU hasn't posted a valid buffer yet).
>> 
>> > So apparently the presence of BGMAC_DESC_CTL0_EOT in last rx desc was
>> > enough to avoid any problem with this.
>> EOT serves a completely different purpose. It simply indicates the end
>> of the ring, telling the hardware to go back to the beginning for the
>> next descriptor.
> 
> Right, but after your patch, we program the hardware with end == start
> == 0
> 
> So I do not really understand how the hardware can still deliver first
> frame.
> 
> If it was really taking care of 'end' value, then it should not even
> deliver one frame.
> 
> The dma corruption happened because hardware was simply reusing frames,
> that driver already took to upper stack, since as you said, there is no
> ownership bit.
> 
> You might try your patch with BGMAC_RX_RING_SLOTS = 3 and check if it is
> still working properly.
> 
> I find BGMAC_RX_RING_SLOTS==511 very suspect, and a comment of why this
> magic value is not 512 would be nice.
I completely reworked the code and simplified handling
BGMAC_DMA_RX_INDEX by setting it to the last filled rx slot (keeping the
entire ring filled). I now get better performance and it stays reliable
even with smaller rings. Sending v3 now.

- Felix