On May 17, Jakub Kicinski wrote:
> On Mon, 16 May 2022 18:06:39 +0200 Lorenzo Bianconi wrote:
> > +	struct mtk_mac *mac = netdev_priv(dev);
> > +	struct mtk_tx_dma_v2 *desc = txd;
> > +	struct mtk_eth *eth = mac->hw;
> > +	u32 data;
> > +
> > +	WRITE_ONCE(desc->txd1, info->addr);
> > +
> > +	data = TX_DMA_PLEN0(info->size);
> > +	if (info->last)
> > +		data |= TX_DMA_LS0;
> > +	WRITE_ONCE(desc->txd3, data);
> > +
> > +	if (!info->qid && mac->id)
> > +		info->qid = MTK_QDMA_GMAC2_QID;
> > +
> > +	data = (mac->id + 1) << TX_DMA_FPORT_SHIFT_V2; /* forward port */
> > +	data |= TX_DMA_SWC_V2 | QID_BITS_V2(info->qid);
> > +	WRITE_ONCE(desc->txd4, data);
> > +
> > +	data = 0;
> > +	if (info->first) {
> > +		if (info->gso)
> > +			data |= TX_DMA_TSO_V2;
> > +		/* tx checksum offload */
> > +		if (info->csum)
> > +			data |= TX_DMA_CHKSUM_V2;
> > +	}
> > +	WRITE_ONCE(desc->txd5, data);
> > +
> > +	data = 0;
> > +	if (info->first && info->vlan)
> > +		data |= TX_DMA_INS_VLAN_V2 | info->vlan_tci;
> > +	WRITE_ONCE(desc->txd6, data);
> > +
> > +	WRITE_ONCE(desc->txd7, 0);
> > +	WRITE_ONCE(desc->txd8, 0);
> 
> Why all the WRITE_ONCE()? Don't you just need a barrier between writing
> the descriptor and kicking the HW? 

I used this approach just to be aligned with current codebase:
https://github.com/torvalds/linux/blob/master/drivers/net/ethernet/mediatek/mtk_eth_soc.c#L1006
https://github.com/torvalds/linux/blob/master/drivers/net/ethernet/mediatek/mtk_eth_soc.c#L1031

but I guess we can even convert the code to use barrier instead. Agree?

Regards,
Lorenzo