Re: [PATCH] Add support the Korina (IDT RC32434) Ethernet MAC

[Francois Romieu <romieu@fr.zoreil.com>]

Florian Fainelli <florian.fainelli@telecomint.eu> :
> Le mercredi 19 mars 2008, Francois Romieu a ?crit?:
[...]
> > If the allocation fails, the driver will have to handle a shorter
> > than expected ring (it is probably ok, I have not checked it). You
> > may consider initializing the link field for each descriptor if
> > possible. See epic_init_ring() for some inspiration.
> 
> That is why I use the i variable such that if a break in the loop happens,
> the ring and last/first decscriptors are properly set.

I was more worried that rx_next_done (in korina_rx) does not seem to
know that the descriptor ring can actually be shorter than expected.

static void korina_alloc_ring(struct net_device *dev)
{
[...]
	for (i = 0; i < KORINA_NUM_RDS; i++) {
                struct sk_buff *skb = lp->rx_skb[i];

                skb = dev_alloc_skb(KORINA_RBSIZE + 2);
                if (!skb)
                        break;
[...]
                lp->rd_ring[i].link = CPHYSADDR(&lp->rd_ring[i+1]);
[...]
static int korina_rx(struct net_device *dev, int limit)
{
[...]
                        rd->control = DMA_COUNT(KORINA_RBSIZE) |
                                DMA_DESC_COD | DMA_DESC_IOD;
                        lp->rd_ring[(lp->rx_next_done - 1) &
                                KORINA_RDS_MASK].control &= ~DMA_DESC_COD;

                        lp->rx_next_done = (lp->rx_next_done + 1) & KORINA_RDS_MASK;

-> Sooner than later, rx_next_done will force the code through a
   descriptor ring entry where rd->link is not set while the
   hardware has already returned to the start of the ring.

-- 
Ueimor