Re: [PATCH v3 net-next 14/15] net: aquantia: fixups on 64bit dma counters

[Andrew Lunn <andrew@lunn.ch>]

On Sat, Apr 27, 2019 at 06:28:12AM +0000, Igor Russkikh wrote:
> 
> 
> On 27.04.2019 0:36, Andrew Lunn wrote:
> >> +	self->curr_stats.dma_pkt_rc = hw_atl_stats_rx_dma_good_pkt_counterlsw_get(self) +
> >> +				      ((u64)hw_atl_stats_rx_dma_good_pkt_countermsw_get(self) << 32);
> > 
> > Don't you need to do something to avoid issue with overflow from lsw
> > into msw? I've often seen code get the msw, the lsw and then the msm
> > again. If the two msw reads are different, it repeats it all again.
> 
> Hardware latches msw when host reads lsw register. So thats safe.
> 
> However looking into spec it says to always read lsw first.
> The syntax above does not guarantee that in general.
> I'll change that with something like
> 
> dma_pkt_rc = hw_atl_stats_rx_dma_good_pkt_counterlsw_get(self);
> dma_pkt_rc |= ((u64)hw_atl_stats_rx_dma_good_pkt_countermsw_get(self) << 32);

Hi Igor

I think i would probably have a hw_atl_stats_rx_dma_good_pkt_counter()
which returns a u64.

Under that, have a aq_hw_read_reg64() which does things in the correct
ordering, and a comment about reading the lsw latches the msw.

	Andrew