2014-03-03 10:21 GMT-08:00 Vince Bridgers <vbridgers2013@xxxxxxxxx>:
> Hello Florian, thank you for taking the time to comments. My responses inline.
>
> On Sun, Mar 2, 2014 at 6:59 PM, Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:
>> Hello Vince,
>>
>> It might help reviewing the patches by breaking the patches into:
>>
>> - the SGDMA bits
>> - the MSGDMA bits
>> - the Ethernet MAC driver per-se
>
> I'll break down the next submission.
>
>>
>> BTW, it does look like the SGDMA code could/should be a dmaengine driver?
>
> I did consider this, but after studying the dmaengine api I found the
> API definitions and semantics were not a match to the way the SGDMA
> and MSGDMA behave collectively. Moreover, I could not find an example
> of an Ethernet driver that made use of the dmaengine API - only the
> Micrel driver seems to use it. When studying what components actually
> used the dmaengine API I concluded the dmaengine API was defined for
> use cases different than Ethernet.
To tell you the truth, I have myself been toying with the idea of
using a dmaengine driver for some Ethernet drivers out there (mainly
bcm63xx_enet), but so far have not had any chance to take a stab at
doing a real implementation. My concern was more about the fact that
maybe the SGDMA/MSGDMA code could be reusable for other peripherals in
your system, such as USB device, PCM etc... This is fine anyway.
[snip]
>>
>> Is this the software number of descriptors or hardware number of
>> descriptors?
>
> This number applies to the number of descriptors as limited by the
> MSGDMA capabilities. The SGDMA has different limitations and issues,
> but the maximum number of descriptors for either DMA engine that can
> be used is represented as shown above. This is important since an
> unusual hardware configuration could support both SGDMA and MSGDMA
> simultaneously for more than one TSE instance. I used a design that
> supported a single SGDMA with TSE and a single MSGDMA with TSE for
> testing purposes (among other designs). So this a hardware defined
> number of descriptors and is fixed.
Ok, so this describes a hardware configuration, and as such should be
part of some Device Tree properties, something that could easily be
changed by someone wanting to slightly modify the RTL parameters.
>
>>
>> [snip]
>>
>>
>>> +
>>> +static int altera_tse_mdio_create(struct net_device *dev, unsigned int
>>> id)
>>> +{
>>> + struct altera_tse_private *priv = netdev_priv(dev);
>>> + int ret;
>>> + int i;
>>> + struct device_node *mdio_node;
>>> + struct mii_bus *mdio;
>>> +
>>> + mdio_node = of_find_compatible_node(priv->device->of_node, NULL,
>>> + "altr,tse-mdio");
>>> +
>>> + if (mdio_node) {
>>> + dev_warn(priv->device, "FOUND MDIO subnode\n");
>>> + } else {
>>> + dev_warn(priv->device, "NO MDIO subnode\n");
>>> + return 0;
>>> + }
>>> +
>>> + mdio = mdiobus_alloc();
>>> + if (mdio == NULL) {
>>> + dev_err(priv->device, "Error allocating MDIO bus\n");
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + mdio->name = ALTERA_TSE_RESOURCE_NAME;
>>> + mdio->read = &altera_tse_mdio_read;
>>> + mdio->write = &altera_tse_mdio_write;
>>> + snprintf(mdio->id, MII_BUS_ID_SIZE, "%s-%u", mdio->name, id);
>>
>>
>> You could use something more user-friendly such as mdio_node->full_name.
>
> The full name will exceed the characters available in that particular
> structure data member. MII_BUS_ID_SIZE is defined as (20-3) in
> include/linux/phy.h. The full name would exceed the allocated space in
> that structure. That's why this method was chosen.
Ok, that works too.
>
>>
>>
>>> +
>>> + mdio->irq = kcalloc(PHY_MAX_ADDR, sizeof(int), GFP_KERNEL);
>>> + if (mdio->irq == NULL) {
>>> + dev_err(priv->device, "%s: Cannot allocate memory\n",
>>> __func__);
>>> + ret = -ENOMEM;
>>> + goto out_free_mdio;
>>> + }
>>> + for (i = 0; i < PHY_MAX_ADDR; i++)
>>> + mdio->irq[i] = PHY_POLL;
>>> +
>>> + mdio->priv = (void *)priv->mac_dev;
>>
>>
>> No need for the cast here, this is already a void *.
>
> Noted.
>
>>
>>
>>> + mdio->parent = priv->device;
>>
>>
>> [snip]
>>
>>
>>> + /* make cache consistent with receive packet buffer */
>>> + dma_sync_single_for_cpu(priv->device,
>>> + priv->rx_ring[entry].dma_addr,
>>> + priv->rx_ring[entry].len,
>>> + DMA_FROM_DEVICE);
>>> +
>>> + dma_unmap_single(priv->device,
>>> priv->rx_ring[entry].dma_addr,
>>> + priv->rx_ring[entry].len,
>>> DMA_FROM_DEVICE);
>>> +
>>> + /* make sure all pending memory updates are complete */
>>> + rmb();
>>
>>
>> Are you sure this does something in your case? I am fairly sure that the
>> dma_unmap_single() call would have done that implicitely for you here.
>
> I wrote the code this way intentionally to be explicit. I'll check the
> API for behavior as you describe for both ARM and NIOS and if not
> handled this barrier should probably remain.
>
[snip]
>>
>> I am not sure this will even be triggered if you want do not advertise
>> NETIF_F_SG, so you might want to drop that entirely.
>
> The intent was to add Scatter Gather capabilities at some point in the
> future, so this was a form of documenting. I'll just drop the code and
> add a comment instead if you agree.
Since you are not advertising the NETIF_F_SG capability bit, you
should probably just drop this code and add it back again under a
different form when SG support is added.
[snip]
>>
>> You might rather do this during your driver probe function rather than in
>> the ndo_open() callback.
>
> This seems to be the normal place to probe and initialize the phy from
> examination of existing code. Perhaps I missed something, could you
> provide an example of where this is done differently?
Most drivers I can think about: tg3, bcmgenet, bcm63xx_enet, r6040 do
probe for the PHY in their driver's probe routine.
>
>>
>> [snip]
>>
>>
>>> + /* Stop and disconnect the PHY */
>>> + if (priv->phydev) {
>>> + phy_stop(priv->phydev);
>>> + phy_disconnect(priv->phydev);
>>> + priv->phydev = NULL;
>>> + }
>>> +
>>> + netif_stop_queue(dev);
>>> + napi_disable(&priv->napi);
>>> +
>>> + /* Disable DMA interrupts */
>>> + spin_lock_irqsave(&priv->rxdma_irq_lock, flags);
>>> + priv->disable_rxirq(priv);
>>> + priv->disable_txirq(priv);
>>> + spin_unlock_irqrestore(&priv->rxdma_irq_lock, flags);
>>> +
>>> + /* Free the IRQ lines */
>>> + free_irq(priv->rx_irq, dev);
>>> + free_irq(priv->tx_irq, dev);
>>> +
>>> + /* disable and reset the MAC, empties fifo */
>>> + spin_lock(&priv->mac_cfg_lock);
>>> + spin_lock(&priv->tx_lock);
>>> +
>>> + ret = reset_mac(priv);
>>> + if (ret)
>>> + netdev_err(dev, "Cannot reset MAC core (error: %d)\n",
>>> ret);
>>> + priv->reset_dma(priv);
>>> + free_skbufs(dev);
>>> +
>>> + spin_unlock(&priv->tx_lock);
>>> + spin_unlock(&priv->mac_cfg_lock);
>>> +
>>> + priv->uninit_dma(priv);
>>> +
>>> + netif_carrier_off(dev);
>>
>>
>> phy_stop() does that already.
>
> If you mean phy_stop calls netif_carrier_off, I don't see it in that
> function (phy/phy.c). Common usage in the intree drivers seems to be
> calling netif_carrier_off in this context, but perhaps the drivers I
> examined have not been updated. Is there some more specific feedback
> or comments that you could provide here? Do you mean that
> netif_carrier_off is unnecessary here?
by calling phy_stop() the PHY state machine will move to the state
PHY_HALTED, if the link state was valid at this point, it will call
netif_carrier_off(). If the link was not valid before, the PHY state
machine would have already called netif_carrier_off().
>
>>
>>
>>> +
>>> + return 0;
>>> +}
>>> +
>>> +static struct net_device_ops altera_tse_netdev_ops = {
>>> + .ndo_open = tse_open,
>>> + .ndo_stop = tse_shutdown,
>>> + .ndo_start_xmit = tse_start_xmit,
>>> + .ndo_set_mac_address = eth_mac_addr,
>>> + .ndo_set_rx_mode = tse_set_rx_mode,
>>> + .ndo_change_mtu = tse_change_mtu,
>>> + .ndo_validate_addr = eth_validate_addr,
>>> +};
>>> +
>>> +static int altera_tse_get_of_prop(struct platform_device *pdev,
>>> + const char *name, unsigned int *val)
>>> +{
>>> + const __be32 *tmp;
>>> + int len;
>>> + char buf[strlen(name)+1];
>>> +
>>> + tmp = of_get_property(pdev->dev.of_node, name, &len);
>>> + if (!tmp && !strncmp(name, "altr,", 5)) {
>>> + strcpy(buf, name);
>>> + strncpy(buf, "ALTR,", 5);
>>> + tmp = of_get_property(pdev->dev.of_node, buf, &len);
>>> + }
>>> + if (!tmp || (len < sizeof(__be32)))
>>> + return -ENODEV;
>>> +
>>> + *val = be32_to_cpup(tmp);
>>> + return 0;
>>> +}
>>
>>
>> Do we really need that abstration?
>
> The intent here is to support legacy device trees that were created
> with upper case "ALTR".
Oh Device Tree fun, welcome to the club.
>
>>
>>
>>> +
>>> +static int altera_tse_get_phy_iface_prop(struct platform_device *pdev,
>>> + phy_interface_t *iface)
>>> +{
>>> + const void *prop;
>>> + int len;
>>> +
>>> + prop = of_get_property(pdev->dev.of_node, "phy-mode", &len);
>>> + if (!prop)
>>> + return -ENOENT;
>>> + if (len < 4)
>>> + return -EINVAL;
>>> +
>>> + if (!strncmp((char *)prop, "mii", 3)) {
>>> + *iface = PHY_INTERFACE_MODE_MII;
>>> + return 0;
>>> + } else if (!strncmp((char *)prop, "gmii", 4)) {
>>> + *iface = PHY_INTERFACE_MODE_GMII;
>>> + return 0;
>>> + } else if (!strncmp((char *)prop, "rgmii-id", 8)) {
>>> + *iface = PHY_INTERFACE_MODE_RGMII_ID;
>>> + return 0;
>>> + } else if (!strncmp((char *)prop, "rgmii", 5)) {
>>> + *iface = PHY_INTERFACE_MODE_RGMII;
>>> + return 0;
>>> + } else if (!strncmp((char *)prop, "sgmii", 5)) {
>>> + *iface = PHY_INTERFACE_MODE_SGMII;
>>> + return 0;
>>> + }
>>
>>
>> of_get_phy_mode() does that for you.
>
> Will address this. Thank you.
>
>>
>>
>>> +
>>> + return -EINVAL;
>>> +}
>>> +
>>> +static int request_and_map(struct platform_device *pdev, const char
>>> *name,
>>> + struct resource **res, void __iomem **ptr)
>>> +{
>>> + struct resource *region;
>>> + struct device *device = &pdev->dev;
>>> +
>>> + *res = platform_get_resource_byname(pdev, IORESOURCE_MEM, name);
>>> + if (*res == NULL) {
>>> + dev_err(device, "resource %s not defined\n", name);
>>> + return -ENODEV;
>>> + }
>>> +
>>> + region = devm_request_mem_region(device, (*res)->start,
>>> + resource_size(*res),
>>> dev_name(device));
>>> + if (region == NULL) {
>>> + dev_err(device, "unable to request %s\n", name);
>>> + return -EBUSY;
>>> + }
>>> +
>>> + *ptr = devm_ioremap_nocache(device, region->start,
>>> + resource_size(region));
>>> + if (*ptr == NULL) {
>>> + dev_err(device, "ioremap_nocache of %s failed!", name);
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + return 0;
>>> +}
>>> +
>>> +/* Probe Altera TSE MAC device
>>> + */
>>> +static int altera_tse_probe(struct platform_device *pdev)
>>> +{
>>> + struct net_device *ndev;
>>> + int ret = -ENODEV;
>>> + struct resource *control_port;
>>> + struct resource *dma_res;
>>> + struct altera_tse_private *priv;
>>> + int len;
>>> + const unsigned char *macaddr;
>>> + struct device_node *np = pdev->dev.of_node;
>>> + unsigned int descmap;
>>> +
>>> + ndev = alloc_etherdev(sizeof(struct altera_tse_private));
>>> + if (!ndev) {
>>> + dev_err(&pdev->dev, "Could not allocate network
>>> device\n");
>>> + return -ENODEV;
>>> + }
>>> +
>>> + SET_NETDEV_DEV(ndev, &pdev->dev);
>>> +
>>> + priv = netdev_priv(ndev);
>>> + priv->device = &pdev->dev;
>>> + priv->dev = ndev;
>>> + priv->msg_enable = netif_msg_init(debug, default_msg_level);
>>> +
>>> + if (of_device_is_compatible(np, "altr,tse-1.0") ||
>>> + of_device_is_compatible(np, "ALTR,tse-1.0")) {
>>
>>
>> Use the .data pointer associated with the compatible string to help you do
>> that, see below.
>
> Noted. If we can agree that use of the dmaengine api is inappropriate
> in this case, I'll update the code to use a .data pointer as
> suggested. Thank you for the comment.
I think not using the dmaengine API is fine.
>
>>
>> [snip]
>>
>>
>>> + /* get supplemental address settings for this instance */
>>> + ret = altera_tse_get_of_prop(pdev, "altr,enable-sup-addr",
>>> + &priv->added_unicast);
>>> + if (ret)
>>> + priv->added_unicast = 0;
>>> +
>>> + /* Max MTU is 1500, ETH_DATA_LEN */
>>> + priv->max_mtu = ETH_DATA_LEN;
>>
>>
>> How about VLANs? If this is always 1500, just let the core ethernet
>> functions configure the MTU for your interface.
>
> The TSE core handles frame size expansion for VLAN tagged frames, so
> it's ok (tested). At some point, frame sizes > 1518 may be supported
> (the core supports Jumbo frames, the driver is intentionally simple
> for initial submission). Your comment is noted, I accept the
> suggestion.
In case this needs to be matched against a newer version of the RTL,
or whatever HW configuration some RTL user is allowed to make, you
could probaly use the 'max-frame-size' ePAPR-defined property for
this?
>
>>
>>
>>> +
>>> + /* The DMA buffer size already accounts for an alignment bias
>>> + * to avoid unaligned access exceptions for the NIOS processor,
>>> + */
>>> + priv->rx_dma_buf_sz = ALTERA_RXDMABUFFER_SIZE;
>>> +
>>> + /* get default MAC address from device tree */
>>> + macaddr = of_get_property(pdev->dev.of_node, "local-mac-address",
>>> &len);
>>> + if (macaddr && len == ETH_ALEN)
>>> + memcpy(ndev->dev_addr, macaddr, ETH_ALEN);
>>> +
>>> + /* If we didn't get a valid address, generate a random one */
>>> + if (!is_valid_ether_addr(ndev->dev_addr))
>>> + eth_hw_addr_random(ndev);
>>> +
>>> + ret = altera_tse_get_phy_iface_prop(pdev, &priv->phy_iface);
>>> + if (ret == -ENOENT) {
>>> + /* backward compatability, assume RGMII */
>>> + dev_warn(&pdev->dev,
>>> + "cannot obtain PHY interface mode, assuming
>>> RGMII\n");
>>> + priv->phy_iface = PHY_INTERFACE_MODE_RGMII;
>>> + } else if (ret) {
>>> + dev_err(&pdev->dev, "unknown PHY interface mode\n");
>>> + goto out_free;
>>> + }
>>> +
>>> + /* try to get PHY address from device tree, use PHY autodetection
>>> if
>>> + * no valid address is given
>>> + */
>>> + ret = altera_tse_get_of_prop(pdev, "altr,phy-addr",
>>> &priv->phy_addr);
>>> + if (ret)
>>> + priv->phy_addr = POLL_PHY;
>>
>>
>> Please do not use such as custom property, either you use an Ethernet PHY
>> device tree node as described in ePAPR; or you do not and use a fixed-link
>> property instead.
>
> Agreed, the code tries phy handles as described in the ePAPR v1.1
> specification, then falls back to the method in question. The intent
> is to support legacy device trees as well. Is there a preferred way to
> handle legacy configurations that we may encounter in the wild?
Well, ideally a new driver should have no legacy at all, but I
understand that situation might not be the case, I believe it should
be duly noted in the Device Tree binding documentation (have not
reviewed that patch yet). Issuing a warning might be beneficial as
well to spot "old" DT bindings and help troubleshooting setups?
--
Florian
--
To unsubscribe from this list: send the line "unsubscribe linux-doc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
