Re: [PATCH v2] PCI: iproc: Fix BCMA PCIe bus scanning regression

From: Bjorn Helgaas
Date: Wed Jan 27 2016 - 17:53:10 EST


On Tue, Jan 26, 2016 at 03:31:40PM -0800, Ray Jui wrote:
> Commit 943ebae781f5 ("PCI: iproc: Add PAXC interface support") causes
> regression on EP device detection on BCMA based platforms. This patch
> fixes the issue by allowing multiple devices to be configured on the
> same bus, for all PAXB based child buses. In addition, this patch also
> adds check to prevent non-zero function from being used on bus 0 (root
> bus).
>
> Function 'iproc_pcie_device_is_valid' is now removed with checks
> folding into 'iproc_pcie_map_cfg_bus' to make them more clear and less
> error-prone
>
> Reported-by: Rafal Milecki <zajec5@xxxxxxxxx>
> Fixes: 943ebae781f5 ("PCI: iproc: Add PAXC interface support")
> Signed-off-by: Ray Jui <rjui@xxxxxxxxxxxx>

Since this fixes a regression, I applied this to for-linus for v4.5,
thanks, Ray.

I still have one clarification question below.

> ---
> drivers/pci/host/pcie-iproc.c | 28 +++++++++++-----------------
> 1 file changed, 11 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/pci/host/pcie-iproc.c b/drivers/pci/host/pcie-iproc.c
> index 5816bce..67396ab 100644
> --- a/drivers/pci/host/pcie-iproc.c
> +++ b/drivers/pci/host/pcie-iproc.c
> @@ -170,20 +170,6 @@ static inline void iproc_pcie_ob_write(struct iproc_pcie *pcie,
> writel(val, pcie->base + offset + (window * 8));
> }
>
> -static inline bool iproc_pcie_device_is_valid(struct iproc_pcie *pcie,
> - unsigned int slot,
> - unsigned int fn)
> -{
> - if (slot > 0)
> - return false;
> -
> - /* PAXC can only support limited number of functions */
> - if (pcie->type == IPROC_PCIE_PAXC && fn >= MAX_NUM_PAXC_PF)
> - return false;
> -
> - return true;
> -}
> -
> /**
> * Note access to the configuration registers are protected at the higher layer
> * by 'pci_lock' in drivers/pci/access.c
> @@ -199,11 +185,11 @@ static void __iomem *iproc_pcie_map_cfg_bus(struct pci_bus *bus,
> u32 val;
> u16 offset;
>
> - if (!iproc_pcie_device_is_valid(pcie, slot, fn))
> - return NULL;
> -
> /* root complex access */
> if (busno == 0) {
> + if (slot > 0 || fn > 0)
> + return NULL;
> +

This looks good and makes sense since config access to root bus is
fundamentally different from other access.

> iproc_pcie_write_reg(pcie, IPROC_PCIE_CFG_IND_ADDR,
> where & CFG_IND_ADDR_MASK);
> offset = iproc_pcie_reg_offset(pcie, IPROC_PCIE_CFG_IND_DATA);
> @@ -213,6 +199,14 @@ static void __iomem *iproc_pcie_map_cfg_bus(struct pci_bus *bus,
> return (pcie->base + offset);
> }
>
> + /*
> + * PAXC is connected to internally emulated EP within the SoC. It
> + * allows only one device and supports limited number of functions
> + */
> + if (pcie->type == IPROC_PCIE_PAXC)
> + if (slot > 0 || fn >= MAX_NUM_PAXC_PF)
> + return NULL;

Is this really necessary? I assume 00:00.0 is a Root Port leading to
bus 01, and 01:00.0, 01:00.1, 01:00.2, and 01:00.3 are the functions
of the internal EP. So this test prevents us from issuing a config
request to devices like 01:00.4.

I would assume the Root Port is standard and would handle a config
request for 01:00.4 correctly, i.e., convert the type 1 request to
type 0 (since it targets the Root Port's secondary bus), and forward
it to the link.

The endpoint should be responsible for handling it as an Unsupported
Request, since it addresses an unimplemented function. But maybe this
embedded EP doesn't do that correctly?

Also, assuming we *do* need this PAXC testing, do you want to test for
"busno == 1" as well? The PCI core shouldn't try to access bus 2
unless there's a bridge from bus 1 to bus 2, but a user could use
things like setpci to issue random config requests.

> +
> /* EP device access */
> val = (busno << CFG_ADDR_BUS_NUM_SHIFT) |
> (slot << CFG_ADDR_DEV_NUM_SHIFT) |
> --
> 1.9.1
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html