Hi Arnd,
Please see inline.
Thanks,
Mayuresh.
On Wed, Jan 7, 2015 at 2:05 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Wednesday 07 January 2015 10:58:57 Mayuresh Chitale wrote:
>> This patch implements DMA engine API for DMA controller on APM
>> X-Gene PCIe controller. DMA engine can support up to 4 channels per port
>> and up to 2048 outstanding requests per channel. This is intended
>> to be used on ports that are configured in EP mode or to transfer
>> data from a RC port that is connected to a X-Gene EP port.
>
> Linux currently does not support PCIe endpoints. Do you plan to submit
> a driver for that as well?
>
>> +
>> +static void xpdma_setup_dma_dev(struct dma_device *dev)
>> +{
>> + dma_cap_zero(dev->cap_mask);
>> + dma_cap_set(DMA_SLAVE, dev->cap_mask);
>> + dma_cap_set(DMA_PRIVATE, dev->cap_mask);
>> +
>> + dev->device_alloc_chan_resources = xpdma_alloc_chan_resources;
>> + dev->device_free_chan_resources = xpdma_free_chan_resources;
>> + dev->device_tx_status = xpdma_tx_status;
>> + dev->device_issue_pending = xpdma_issue_pending;
>> + dev->device_prep_slave_sg = xpdma_prep_slave_sg;
>> + dev->device_control = xpdma_device_control;
>> +}
>
> You are setting up DMA_SLAVE here but do not use the DT binding
> for slave DMA. Can your driver handle slave DMA? If it can,
> you should use the appropriate binding, if not then don't
> set DMA_SLAVE here.
Are you referring to DT DMA helpers? If so, I will add it.
>
>> +static int xpdma_setup_dma_channel(struct platform_device *pdev,
>> + struct xpdma_port *port)
>> +{
>> + int i, ret = 0;
>> + struct xpdma_chan *chan;
>> + resource_size_t dma_base;
>> +
>> + dma_base = MAKE_U64(readl(port->cfg + PORT_CFG_HI), readl(port->cfg)) +
>> + CHAN_REG_BASE;
>> + port->dma = devm_ioremap(&pdev->dev, dma_base,
>> + CHAN_REG_SIZE * MAX_DMA_CHAN);
>> + if (port->dma == NULL) {
>> + dev_err(&pdev->dev, "Could not get base addressc$\n");
>> + return -ENOMEM;
>> + }
>
> The registers are not listed in the 'reg' property. Why is that?
Actually these registers have to accessed via the PCIe config space
region for that port. Since the config space base address is
programmed by host controller driver we can know the base address only
by reading the corresponding config space bar at runtime.
>
> You should probably use devm_request_mem_region or devm_ioremap_resource
> here to ensure manage the mmio range.
>
>> +static int xpdma_probe(struct platform_device *pdev)
>> +{
>> + int err;
>> + u32 mask;
>> + struct resource *res;
>> + struct xpdma_port *port;
>> +
>> + port = devm_kzalloc(&pdev->dev, sizeof(*port), GFP_KERNEL);
>> + if (!port)
>> + return -ENOMEM;
>> +
>> + port->dma_dev.dev = &pdev->dev;
>> + res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
>> + if (!res) {
>> + dev_err(&pdev->dev, "No cfg resource\n");
>> + return -EINVAL;
>> + }
>> +
>> + port->cfg = devm_ioremap(&pdev->dev, res->start, resource_size(res));
>> + if (IS_ERR(port->cfg))
>> + return PTR_ERR(port->cfg);
>> +
>> + res = platform_get_resource(pdev, IORESOURCE_MEM, 1);
>> + if (!res) {
>> + dev_err(&pdev->dev, "No intr resource\n");
>> + return -EINVAL;
>> + }
>
> from binding, these two register ranges look like they are part of
> another device. Which device is that:
>
> + reg = < 0x0 0x1f2b0154 0x0 0xc
> + 0x0 0x1f2b0058 0x0 0x8>;
These are registers inside PCIe host controller that contains
information required such as the dma base address, interrupt
configuration. So this will be used for the channels. I am not aware
about syscon but will see if it better represents these registers.
>
> Is this some kind of syscon node that contains all sorts of
> registers munged together, or is this the same register set that
> is used for the channels?
>
> In the former case, better use a syscon reference instead of 'reg'
> and use syscon_regmap_lookup_by_phandle to get the regmap, in the
> second case, just list the entire registers.
>
>> +
>> + xpdma_setup_dma_dev(&port->dma_dev);
>> + /* Setup DMA mask - 32 for 32-bit system and 64 for 64-bit system */
>> + err = dma_set_mask_and_coherent(&pdev->dev,
>> + DMA_BIT_MASK(8*sizeof(void *)));
>> + if (err) {
>> + dev_err(&pdev->dev, "Unable to set dma mask\n");
>> + return err;
>> + }
>
> Just use a 64-bit dma-mask all the time, and fall back to a 32-bit
> mask if that doesn't work, the dma mask is completely independent
> of the word size of the CPU.
okay.
>
> You also have to make sure that the parent of the device contains
> a proper dma-ranges property that allows a 64-bit mask. We currently
> have a bug in the kernel that makes dma_set_mask_and_coherent
> always succeed, and we need to fix that so it fails if you have
> a 64-bit dma capable device on a 32-bit wide bus.
>
>> +static const struct of_device_id xpdma_match_table[] = {
>> + {.compatible = "apm,xgene-pciedma",},
>> + {},
>> +};
>
> This is a rather generic compatible string. Is this device present
> in identical form on all xgene variants?
Yes.
>
> As xgene is a marketing name and you have now announce the speciifc
> model numbers, better use the first model that has this hardware.
okay.
>
> Arnd
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
