Hi Brian,
> On Nov 17, 2015, at 11:16 AM, Brian Norris <computersforpeace@xxxxxxxxx> wrote:
>
> Hi,
>
> On Fri, Oct 30, 2015 at 11:01:16PM +0900, Jaedon Shin wrote:
>> Add quirk for broken ncq. Some chipsets (eg. BCM7349A0, BCM7445A0,
>> BCM7445B0, and all 40nm chipsets including BCM7425) need a workaround
>> disabling NCQ.
>>
>> Signed-off-by: Jaedon Shin <jaedon.shin@xxxxxxxxx>
>> ---
>> drivers/ata/ahci_brcmstb.c | 46 ++++++++++++++++++++++++++++++++++++++++++++++
>> 1 file changed, 46 insertions(+)
>>
>> diff --git a/drivers/ata/ahci_brcmstb.c b/drivers/ata/ahci_brcmstb.c
>> index 73e3b0b2a3c2..194aeda8f14d 100644
>> --- a/drivers/ata/ahci_brcmstb.c
>> +++ b/drivers/ata/ahci_brcmstb.c
>> @@ -69,10 +69,15 @@
>> (DATA_ENDIAN << DMADESC_ENDIAN_SHIFT) | \
>> (MMIO_ENDIAN << MMIO_ENDIAN_SHIFT))
>>
>> +enum brcm_ahci_quirks {
>> + BRCM_AHCI_QUIRK_NONCQ = BIT(0),
>> +};
>> +
>> struct brcm_ahci_priv {
>> struct device *dev;
>> void __iomem *top_ctrl;
>> u32 port_mask;
>> + u32 quirks;
>> };
>>
>> static const struct ata_port_info ahci_brcm_port_info = {
>> @@ -202,6 +207,42 @@ static u32 brcm_ahci_get_portmask(struct platform_device *pdev,
>> return impl;
>> }
>>
>> +static void brcm_sata_quirks(struct platform_device *pdev,
>> + struct brcm_ahci_priv *priv)
>> +{
>> + if (priv->quirks & BRCM_AHCI_QUIRK_NONCQ) {
>> + void __iomem *ctrl = priv->top_ctrl + SATA_TOP_CTRL_BUS_CTRL;
>> + void __iomem *ahci;
>> + struct resource *res;
>> + u32 reg;
>> +
>> + res = platform_get_resource_byname(pdev, IORESOURCE_MEM,
>> + "ahci");
>> + ahci = devm_ioremap_resource(&pdev->dev, res);
>> + if (IS_ERR(ahci))
>> + return;
>> +
>> + reg = brcm_sata_readreg(ctrl);
>> + reg |= OVERRIDE_HWINIT;
>> + brcm_sata_writereg(reg, ctrl);
>> +
>> + /* Clear out the NCQ bit so the AHCI driver will not issue
>> + * FPDMA/NCQ commands.
>> + */
>> + reg = readl(ahci + HOST_CAP);
>> + reg &= ~HOST_CAP_NCQ;
>> + writel(reg, ahci + HOST_CAP);
>
> You're using readl()/writel() to access the AHCI block, but...
>
>> +
>> + reg = brcm_sata_readreg(ctrl);
>> + reg &= ~OVERRIDE_HWINIT;
>> + brcm_sata_writereg(reg, ctrl);
>> +
>> + devm_iounmap(&pdev->dev, ahci);
>> + devm_release_mem_region(&pdev->dev, res->start,
>> + resource_size(res));
>> + }
>> +}
>> +
>> static void brcm_sata_init(struct brcm_ahci_priv *priv)
>> {
>> /* Configure endianness */
>> @@ -256,6 +297,11 @@ static int brcm_ahci_probe(struct platform_device *pdev)
>> if (IS_ERR(priv->top_ctrl))
>> return PTR_ERR(priv->top_ctrl);
>>
>> + if (of_device_is_compatible(dev->of_node, "brcm,bcm7425-ahci"))
>> + priv->quirks |= BRCM_AHCI_QUIRK_NONCQ;
>> +
>> + brcm_sata_quirks(pdev, priv);
>> +
>> brcm_sata_init(priv);
>
> ...the MMIO endianness is only configured in brcm_sata_init(). You won't
> see this problem on ARM LE, but you should on MIPS BE. Maybe
> brcm_sata_quirks() should be after brcm_sata_init()?
>
Florian already pointed out, the NCQ disabling occurs prior to initializing
the SATA controller endianness in the original BSP. Therefore I think it's better
to change to brcm_sata_{read,write}reg() instead of {read,write}l() for
HOST_CAP overwriting.
>>
>> priv->port_mask = brcm_ahci_get_portmask(pdev, priv);
>
> Brian
