Hi,
Thanks for the review.
Response and further questions below.
On 02/02/16 20:48, Arnd Bergmann wrote:
On Tuesday 02 February 2016 18:24:45 Zubair Lutfullah Kakakhel wrote:
diff --git a/Documentation/devicetree/bindings/mips/cavium/sata-uctl.txt
b/Documentation/devicetree/bindings/mips/cavium/sata-uctl.txt
new file mode 100644
index 0000000..3bd3c2f
--- /dev/null
+++ b/Documentation/devicetree/bindings/mips/cavium/sata-uctl.txt
@@ -0,0 +1,42 @@
+* UCTL SATA controller glue
+
+UCTL is the bridge unit between the I/O interconnect (an internal bus)
+and the SATA AHCI host controller (UAHC). It performs the following functions:
+ - provides interfaces for the applications to access the UAHC AHCI
+ registers on the CN71XX I/O space.
+ - provides a bridge for UAHC to fetch AHCI command table entries and
data
+ buffers from Level 2 Cache.
+ - posts interrupts to the CIU.
+ - contains registers that:
+ - control the behavior of the UAHC
+ - control the clock/reset generation to UAHC
+ - control endian swapping for all UAHC registers and DMA
accesses
Typically we treat those special registers as part of the device itself
and have a single device node for the AHCI controller and that one.
What is your reason for doing it differently here?
Two reasons
1- The hardware is like a proper split rather than additional hidden registers
in
the same memory space.
2- Tons of devices in the field have the following DT node built in the
bootloader.
uctl@118006c000000 {
compatible = "cavium,octeon-7130-sata-uctl";
reg = <0x11800 0x6c000000 0x0 0x100>;
...
sata: sata@16c0000000000 {
compatible = "cavium,octeon-7130-ahci";
reg = <0x16c00 0x00000000 0x0 0x200>;
...
};
};
The patch suggests a way to handle this.
index 04975b8..6bc1de6 100644
--- a/drivers/ata/ahci_platform.c
+++ b/drivers/ata/ahci_platform.c
@@ -76,6 +76,8 @@ static const struct of_device_id ahci_of_match[] = {
{ .compatible = "ibm,476gtr-ahci", },
{ .compatible = "snps,dwc-ahci", },
{ .compatible = "hisilicon,hisi-ahci", },
+ { .compatible = "fsl,qoriq-ahci", },
+ { .compatible = "cavium,octeon-7130-ahci", },
{},
The qoriq-ahci addition seems misplaced in this patch. Maybe split
that out into a separate patch?
oops. Sorry. Will fix.
+/**
+ * cvmx_sata_uctl_shim_cfg
+ * from cvmx-sata-defs.h
+ *
+ * Accessible by: only when A_CLKDIV_EN
+ * Reset by: IOI reset (srst_n) or SATA_UCTL_CTL[SATA_UCTL_RST]
+ * This register allows configuration of various shim (UCTL) features.
+ * Fields XS_NCB_OOB_* are captured when there are no outstanding OOB errors
+ * indicated in INTSTAT and a new OOB error arrives.
+ * Fields XS_BAD_DMA_* are captured when there are no outstanding DMA errors
+ * indicated in INTSTAT and a new DMA error arrives.
+ */
+union cvmx_sata_uctl_shim_cfg {
+ uint64_t u64;
+ struct cvmx_sata_uctl_shim_cfg_s {
+ /*
+ * Read/write error log for out-of-bound UAHC register access.
+ * 0 = read, 1 = write.
+ */
+ __BITFIELD_FIELD(uint64_t xs_ncb_oob_wrn : 1,
+ __BITFIELD_FIELD(uint64_t reserved_57_62 : 6,
+ /*
Try to avoid bitfields when defining structures that interact with
the hardware, bitfields often cause problems with mixed-endian
environments and don't make this easier to understand.
Bitfields for both endians are used and handled by mips.
Mainly used by cavium.
As this is a cavium driver, would it be acceptable?
Or should I replace with the following.
v = cvmx_read_csr((uint64_t)base + CVMX_SATA_UCTL_SHIM_CFG);
v &= ~(SATA_UCTL_ENDIAN_MODE_E_MASK << DMA_ENDIAN_MODE);
v &= ~(SATA_UCTL_ENDIAN_MODE_E_MASK << CSR_ENDIAN_MODE);
#ifdef __BIG_ENDIAN
v |= SATA_UCTL_ENDIAN_MODE_E_BIG << DMA_ENDIAN_MODE;
v |= SATA_UCTL_ENDIAN_MODE_E_BIG << CSR_ENDIAN_MODE;
#else
v |= SATA_UCTL_ENDIAN_MODE_E_LITTLE << DMA_ENDIAN_MODE;
v |= SATA_UCTL_ENDIAN_MODE_E_LITTLE << CSR_ENDIAN_MODE;
#endif
v |= 1 << DMA_READ_CMD;
cvmx_write_csr((uint64_t)base + CVMX_SATA_UCTL_SHIM_CFG, v);
+static int ahci_octeon_probe(struct platform_device *pdev)
+{
+ union cvmx_sata_uctl_shim_cfg shim_cfg;
+ struct device *dev = &pdev->dev;
+ struct device_node *node = dev->of_node;
+ struct resource *res;
+ void __iomem *base;
+ int ret;
+
+ res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+ if (!res) {
+ dev_err(&pdev->dev, "Platform resource[0] is missing\n");
+ return -ENODEV;
+ }
+
+ base = devm_ioremap_resource(&pdev->dev, res);
+ if (IS_ERR(base))
+ return PTR_ERR(base);
+
+ /* set-up endian mode */
+ shim_cfg.u64 = cvmx_read_csr(
+ (__force uint64_t)base + CVMX_SATA_UCTL_SHIM_CFG);
+#ifdef __BIG_ENDIAN
+ shim_cfg.s.dma_endian_mode = 1;
+ shim_cfg.s.csr_endian_mode = 1;
+#else
+ shim_cfg.s.dma_endian_mode = 0;
+ shim_cfg.s.csr_endian_mode = 0;
+#endif
+ shim_cfg.s.dma_read_cmd = 1; /* No allocate L2C */
+ cvmx_write_csr(
+ (__force uint64_t)base + CVMX_SATA_UCTL_SHIM_CFG, shim_cfg.u64);
+
Maybe add a helper function around cvmx_write_csr() that
allows you to write to an __iomem address? Drivers should
generally not need to use __force.
Hmm. Will see.
Thanks
ZubairLK
Arnd
