[PATCH 1/3] ethernet: Add new driver for Marvell Armada 375 network unit

[Ezequiel Garcia]

Hello Francois,

First of all, thanks a lot for such a detailed review!

On 05 Jul 11:03 PM, Francois Romieu wrote:
> > diff --git a/drivers/net/ethernet/marvell/mvpp2.c b/drivers/net/ethernet/marvell/mvpp2.c
> > new file mode 100644
> > index 0000000..154b866
> > --- /dev/null
> > +++ b/drivers/net/ethernet/marvell/mvpp2.c
> [...]
> > +static int mvpp2_prs_tcam_first_free(struct mvpp2 *pp2, int start, int end)
> > +{
> > +	int tid;
> > +	bool found = false;
> > +
> > +	if (start < end)
> > +		for (tid = start; tid <= end; tid++) {
> > +			if (!pp2->prs_shadow[tid].valid) {
> > +				found = true;
> > +				break;
> > +			}
> > +		}
> > +	else
> > +		for (tid = start; tid >= end; tid--) {
> > +			if (!pp2->prs_shadow[tid].valid) {
> > +				found = true;
> > +				break;
> > +			}
> > +		}
> 
> Missing parenthsesis in "if ... else ..." block.
> 
> [...]

Right, I'll do as Joe suggests and use swap() to have just one loop.

> > +static void mvpp2_defaults_set(struct mvpp2_port *pp)
> > +{
> [...]
> > +	/* At default, mask all interrupts to all present cpus */
> > +	for_each_present_cpu(cpu)
> > +		mvpp2_cpu_interrupts_disable(pp, cpu);
> 
> Would it hurt to issue a single write and disable all irqs ?
> 

No, that sounds just fine. I'll fix it.

> [...]
> > +static int mvpp2_tx_frag_process(struct mvpp2_port *pp, struct sk_buff *skb,
> > +				 struct mvpp2_tx_queue *aggr_txq,
> > +				 struct mvpp2_tx_queue *txq)
> > +{
> [...]
> > +error:
> > +	/* Release all descriptors that were used to map fragments of
> > +	 * this packet, as well as the corresponding DMA mappings
> > +	 */
> > +	for (i = i - 1; i >= 0; i--) {
> 
> You may consider "while (--i >= 0) {" as an idiom to balance a
> "for (i = 0; .." loop. Your call.
> 

Hm... we've used the same idiom in the mvneta driver so I'd rather keep it
just to have some consistency between the mvneta and the mvpp2 drivers.

> > +		tx_desc = txq->descs + i;
> > +		dma_unmap_single(pp->dev->dev.parent,
> > +				 tx_desc->buf_phys_addr,
> > +				 tx_desc->data_size,
> > +				 DMA_TO_DEVICE);
> 
> 		dma_unmap_single(pp->dev->dev.parent, tx_desc->buf_phys_addr,
> 				 tx_desc->data_size, DMA_TO_DEVICE);
> 
> You may also factor out mvpp2_dma_unmap_single(pp, tx_desc) or the whole
> dma_unmap_single + mvpp2_txq_desc_put sequence.
> 

Refactoring those two sounds a good cleanup.

> [...]
> > +/* Main tx processing */
> > +static int mvpp2_tx(struct sk_buff *skb, struct net_device *dev)
> > +{
> > +	struct mvpp2_port *pp = netdev_priv(dev);
> > +	struct mvpp2_tx_queue *txq, *aggr_txq;
> > +	struct mvpp2_txq_pcpu *txq_pcpu;
> > +	struct mvpp2_tx_desc *tx_desc;
> > +	dma_addr_t buf_phys_addr;
> > +	int frags = 0;
> > +	u16 txq_id;
> > +	u32 tx_cmd;
> > +
> > +	if (!netif_running(dev))
> > +		goto out;
> 
> The driver is expected to netif_stop_queue when the device goes down, not
> the opposite.
> 

Right, these netif_running() checks are redundant: the driver already stops
the queues when it goes down, so it seems safe to drop them.

> [...]
> > +static int mvpp2_poll(struct napi_struct *napi, int budget)
> > +{
> > +	u32 cause_rx_tx, cause_rx;
> > +	int rx_done = 0;
> > +	struct mvpp2_port *pp = netdev_priv(napi->dev);
> > +
> > +	if (!netif_running(pp->dev)) {
> > +		napi_complete(napi);
> > +		return rx_done;
> > +	}
> 
> Same thing as above.
> 
> [...]
> > +static int mvpp2_ethtool_set_ringparam(struct net_device *dev,
> > +				       struct ethtool_ringparam *ring)
> > +{
> > +	struct mvpp2_port *pp = netdev_priv(dev);
> > +
> > +	if ((ring->rx_pending == 0) || (ring->tx_pending == 0))
> > +		return -EINVAL;
> > +	pp->rx_ring_size = ring->rx_pending < MVPP2_MAX_RXD ?
> > +		ring->rx_pending : MVPP2_MAX_RXD;
> > +	pp->tx_ring_size = ring->tx_pending < MVPP2_MAX_TXD ?
> > +		ring->tx_pending : MVPP2_MAX_TXD;
> > +
> > +	if (netif_running(dev)) {
> > +		mvpp2_stop(dev);
> > +		if (mvpp2_open(dev)) {
> 
> You aren't supposed to (ab)use net_device_ops.{ndo_open / stop} like that.
> mvpp2_tx() is still racing with mvpp2_cleanup_txqs(). Even if you go through
> the hassle of (1) avoiding this race, then (2) enforcing a safe double call
> of mvpp2_stop without any mvpp2_open inbetween, nobody wants its device to
> become unresponsive because of a failed ring parameter change: the driver
> must stop and configure the physical device more gently.
> 
> Depending on the rework, you may consider postponing ethtool ring change to
> a separate patch series.
> 

Right, we will rework this.

> [...]
> > +static const struct net_device_ops mvpp2_netdev_ops = {
> > +	.ndo_open            = mvpp2_open,
> > +	.ndo_stop            = mvpp2_stop,
> > +	.ndo_start_xmit      = mvpp2_tx,
> > +	.ndo_set_rx_mode     = mvpp2_set_rx_mode,
> > +	.ndo_set_mac_address = mvpp2_set_mac_address,
> > +	.ndo_change_mtu      = mvpp2_change_mtu,
> > +	.ndo_get_stats64     = mvpp2_get_stats64,
> > +};
> 
> Please replace spaces by tabs before "=".
> 

Done.

> > +
> > +static const struct ethtool_ops mvpp2_eth_tool_ops = {
> > +	.get_link       = ethtool_op_get_link,
> > +	.get_settings   = mvpp2_ethtool_get_settings,
> > +	.set_settings   = mvpp2_ethtool_set_settings,
> > +	.set_coalesce   = mvpp2_ethtool_set_coalesce,
> > +	.get_coalesce   = mvpp2_ethtool_get_coalesce,
> > +	.get_drvinfo    = mvpp2_ethtool_get_drvinfo,
> > +	.get_ringparam  = mvpp2_ethtool_get_ringparam,
> > +	.set_ringparam	= mvpp2_ethtool_set_ringparam,
> > +};
> 
> Same as above.
> 

Done.

> [...]
> > +static int mvpp2_port_init(struct mvpp2_port *pp)
> > +{
> > +	struct device *dev = pp->dev->dev.parent;
> > +	struct mvpp2 *pp2 = pp->pp2;
> > +	struct mvpp2_txq_pcpu *txq_pcpu;
> > +	int queue, txp, cpu;
> > +
> > +	if (pp->first_rxq + rxq_number > MVPP2_RXQ_TOTAL_NUM)
> > +		return -EINVAL;
> > +
> > +	/* Disable port */
> > +	mvpp2_egress_disable(pp);
> > +	mvpp2_port_disable(pp);
> > +
> > +	pp->txqs = devm_kzalloc(dev, pp->txp_num * txq_number *
> > +				sizeof(struct mvpp2_tx_queue *), GFP_KERNEL);
> > +	if (!pp->txqs)
> > +		return -ENOMEM;
> > +
> > +	/* Associate physical Tx queues to this port and initialize.
> > +	 * The mapping is predefined.
> > +	 */
> > +	for (txp = 0; txp < pp->txp_num; txp++) {
> > +		for (queue = 0; queue < txq_number; queue++) {
> > +			int txq_idx = txp * txq_number + queue;
> > +			int queue_phy_id = mvpp2_txq_phys(pp->id, queue);
> > +			struct mvpp2_tx_queue *txq;
> > +
> > +			txq = devm_kzalloc(dev, sizeof(*txq), GFP_KERNEL);
> > +			if (!txq)
> > +				return -ENOMEM;
> > +
> > +			txq->pcpu = alloc_percpu(struct mvpp2_txq_pcpu);
> > +			if (!txq->pcpu)
> > +				return -ENOMEM;
> 
> There is a per_cpu leak if mvpp2_port_init fails.
> 

Yes, good catch.

> [...]
> > +/* Ports initialization */
> > +static int mvpp2_port_probe(struct platform_device *pdev,
> > +			    struct device_node *port_node,
> > +			    struct mvpp2 *pp2,
> > +			    int *next_first_rxq)
> > +{
> > +	struct device_node *phy_node;
> > +	struct mvpp2_port *pp;
> 
> (nit below)
> 
> I am not a huge fan of the "pp2" variable as there's no "pp1" but I've seen
> worse. However naming "pp" some completely unrelated data imho verges on
> confusing
> 

Hm, yes, I see what you mean. These names bother me too, and I think we could
probably change it so "pp2" is "dev", and "pp" is "port". But on the other
side, I'm not sure it's worth doing such an intrusive change just to fix these
names.

What do you think?

> [...]
> > +	dev->irq = irq_of_parse_and_map(port_node, 0);
> > +	if (dev->irq == 0) {
> > +		err = -EINVAL;
> > +		goto err_free_netdev;
> > +	}
> 
> net_device.irq should be considered legacy. It would be nice to avoid
> more uses of it. You may store it near pp->base for instance.
> 

Done.

Thanks a lot for the feedback!
-- 
Ezequiel García, Free Electrons
Embedded Linux, Kernel and Android Engineering
http://free-electrons.com