Re: [PATCH net-next v2] net: airoha: Implement BQL support

[Lorenzo Bianconi <lorenzo@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 1539 bytes --]

On Oct 15, Jakub Kicinski wrote:
> On Tue, 15 Oct 2024 16:39:08 +0200 Lorenzo Bianconi wrote:
> > On Oct 15, Jakub Kicinski wrote:
> > > On Sat, 12 Oct 2024 11:01:11 +0200 Lorenzo Bianconi wrote:  
> > > > Introduce BQL support in the airoha_eth driver reporting to the kernel
> > > > info about tx hw DMA queues in order to avoid bufferbloat and keep the
> > > > latency small.  
> > > 
> > > TBH I haven't looked at the code again, but when I looked at v1 I was
> > > surprised you don't have a reset in airoha_qdma_cleanup_tx_queue().
> > > Are you sure it's okay? It's a common bug not to reset the BQL state
> > > when queue is purged while stopping the interface.  
> > 
> > So far airoha_qdma_cleanup_tx_queue() is called just in airoha_hw_cleanup()
> > that in turn runs just when the module is removed (airoha_remove()).
> > Do we need it?
> 
> Oh, thought its called on stop. In that case we're probably good
> from BQL perspective.
> 
> But does it mean potentially very stale packets can sit on the Tx
> ring when the device is stopped, until it's started again?

Do you mean the packets that the stack is transmitting when the .ndo_stop() is
run? 
In airoha_dev_stop() we call netif_tx_disable() to disable the transmission on
new packets and inflight packets will be consumed by the completion napi,
is it not enough? I guess we can even add netdev_tx_reset_subqueue() for all netdev
queues in airoha_dev_stop(), I do not have a strong opinion about it. What
do you prefer?

Regards,
Lorenzo

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]