Message ID | 1461188810-24083-1-git-send-email-swarren@wwwdotorg.org |
---|---|
State | Accepted |
Commit | f5c6db84e71c60bfc7ae746bfb2cd1090d0b8765 |
Delegated to: | Tom Warren |
Headers | show |
On Wed, Apr 20, 2016 at 03:46:50PM -0600, Stephen Warren wrote: > From: Stephen Warren <swarren@nvidia.com> > > Tegra20's PCIe controller has a couple of quirks. There are workarounds in > the driver for these, but they don't work after the DM conversion: > > 1) The PCI_CLASS value is wrong in HW. > > This is worked around in pci_tegra_read_config() by patching up the value > read from that register. Pre-DM, the PCIe core always read this via a > 16-bit access to the 16-bit offset 0xa. With DM, 32-bit accesses are used, > so we need to check for offset 0x8 instead. Mask the offset value back to > 32-bit alignment to make this work in all cases. > > 2) Accessing devices other than dev 1 causes a data abort. > > Pre-DM, this was worked around in pci_skip_dev(), which the PCIe core code > called during enumeration while iterating over a bus. The DM PCIe core > doesn't use this function. Instead, enhance tegra_pcie_conf_address() to > validate the bdf being accessed, and refuse to access invalid devices. > Since pci_skip_dev() isn't used, delete it. > > I've also validated that both these WARs are only needed for Tegra20, by > testing on Tegra30/Cardhu and Tegra124/Jetson TKx. So, compile them in > conditionally. > > Fixes: e81ca88451cf ("dm: tegra: pci: Convert tegra boards to driver model for PCI") > Signed-off-by: Stephen Warren <swarren@nvidia.com> > --- > drivers/pci/pci_tegra.c | 21 ++++++++++----------- > 1 file changed, 10 insertions(+), 11 deletions(-) Neat: Reviewed-by: Thierry Reding <treding@nvidia.com>
Hi Stephen, On 20 April 2016 at 15:46, Stephen Warren <swarren@wwwdotorg.org> wrote: > From: Stephen Warren <swarren@nvidia.com> > > Tegra20's PCIe controller has a couple of quirks. There are workarounds in > the driver for these, but they don't work after the DM conversion: > > 1) The PCI_CLASS value is wrong in HW. > > This is worked around in pci_tegra_read_config() by patching up the value > read from that register. Pre-DM, the PCIe core always read this via a > 16-bit access to the 16-bit offset 0xa. With DM, 32-bit accesses are used, > so we need to check for offset 0x8 instead. Mask the offset value back to > 32-bit alignment to make this work in all cases. > > 2) Accessing devices other than dev 1 causes a data abort. > > Pre-DM, this was worked around in pci_skip_dev(), which the PCIe core code > called during enumeration while iterating over a bus. The DM PCIe core > doesn't use this function. Instead, enhance tegra_pcie_conf_address() to > validate the bdf being accessed, and refuse to access invalid devices. > Since pci_skip_dev() isn't used, delete it. > > I've also validated that both these WARs are only needed for Tegra20, by > testing on Tegra30/Cardhu and Tegra124/Jetson TKx. So, compile them in > conditionally. > > Fixes: e81ca88451cf ("dm: tegra: pci: Convert tegra boards to driver model for PCI") > Signed-off-by: Stephen Warren <swarren@nvidia.com> > --- > drivers/pci/pci_tegra.c | 21 ++++++++++----------- > 1 file changed, 10 insertions(+), 11 deletions(-) Nice patch. Reviewed-by: Simon Glass <sjg@chromium.org> - Simon
On 04/20/2016 03:46 PM, Stephen Warren wrote: > From: Stephen Warren <swarren@nvidia.com> > > Tegra20's PCIe controller has a couple of quirks. There are workarounds in > the driver for these, but they don't work after the DM conversion: > > 1) The PCI_CLASS value is wrong in HW. > > This is worked around in pci_tegra_read_config() by patching up the value > read from that register. Pre-DM, the PCIe core always read this via a > 16-bit access to the 16-bit offset 0xa. With DM, 32-bit accesses are used, > so we need to check for offset 0x8 instead. Mask the offset value back to > 32-bit alignment to make this work in all cases. > > 2) Accessing devices other than dev 1 causes a data abort. > > Pre-DM, this was worked around in pci_skip_dev(), which the PCIe core code > called during enumeration while iterating over a bus. The DM PCIe core > doesn't use this function. Instead, enhance tegra_pcie_conf_address() to > validate the bdf being accessed, and refuse to access invalid devices. > Since pci_skip_dev() isn't used, delete it. > > I've also validated that both these WARs are only needed for Tegra20, by > testing on Tegra30/Cardhu and Tegra124/Jetson TKx. So, compile them in > conditionally. > > Fixes: e81ca88451cf ("dm: tegra: pci: Convert tegra boards to driver model for PCI") > Signed-off-by: Stephen Warren <swarren@nvidia.com> TomW, I'd like to make sure this one gets into the upcoming release, since it's a functional bug-fix.
diff --git a/drivers/pci/pci_tegra.c b/drivers/pci/pci_tegra.c index 5dadf6fa62a7..c5b04daffa65 100644 --- a/drivers/pci/pci_tegra.c +++ b/drivers/pci/pci_tegra.c @@ -275,12 +275,17 @@ static int tegra_pcie_conf_address(struct tegra_pcie *pcie, pci_dev_t bdf, return 0; } } + return -EFAULT; } else { +#ifdef CONFIG_TEGRA20 + unsigned int dev = PCI_DEV(bdf); + if (dev != 0) + return -EFAULT; +#endif + *address = pcie->cs.start + tegra_pcie_conf_offset(bdf, where); return 0; } - - return -EFAULT; } static int pci_tegra_read_config(struct udevice *bus, pci_dev_t bdf, @@ -299,13 +304,15 @@ static int pci_tegra_read_config(struct udevice *bus, pci_dev_t bdf, value = readl(address); +#ifdef CONFIG_TEGRA20 /* fixup root port class */ if (PCI_BUS(bdf) == 0) { - if (offset == PCI_CLASS_REVISION) { + if ((offset & ~3) == PCI_CLASS_REVISION) { value &= ~0x00ff0000; value |= PCI_CLASS_BRIDGE_PCI << 16; } } +#endif done: *valuep = pci_conv_32_to_size(value, offset, size); @@ -1041,11 +1048,3 @@ U_BOOT_DRIVER(pci_tegra) = { .probe = pci_tegra_probe, .priv_auto_alloc_size = sizeof(struct tegra_pcie), }; - -int pci_skip_dev(struct pci_controller *hose, pci_dev_t dev) -{ - if (PCI_BUS(dev) != 0 && PCI_DEV(dev) > 0) - return 1; - - return 0; -}