On Thu, 2018-06-28 at 08:07 -0500, Bjorn Helgaas wrote:

On Wed, Jun 27, 2018 at 05:21:35PM +0800, honghui.zhang at mediatek.com wrote:

quoted

From: Honghui Zhang <redacted>

Mediatek's host controller have two slots, each have it's own control
registers. The host driver need to identify which slot was connected
in order to access the device's configuration space. There's problem
for current host driver to find out which slot was connected to for
a given EP device.

  The Mediatek host controller has two slots, each with its own control
  registers.

Thanks, I will update it in the next version.

quoted

Assuming each slot have connect with one EP device as below:

                host bridge
  bus 0 --> __________|_______
           |                  |
           |                  |
         slot 0             slot 1
  bus 1 -->|        bus 2 --> |
           |                  |
         EP 0               EP 1

While PCI emulation, system software will scan all the PCI device

s/While PCI emulation/During PCI enumeration/

Thanks, will update this.

quoted

starting from devfn 0. So it will get the proper port for slot0 and
slot1 device when using PCI_SLOT(devfn) for match. But it will get
the wrong slot for EP1: The devfn will be start from 0 when scanning
EP1 behind slot1, it will get port0 since the PCI_SLOT(EP1) is match
for port0's slot value. So the host driver should not using EP's devfn
but the slot's devfn(the slot which EP was connected to) for match.

This patch fix the mtk_pcie_find_port's logical by using the slot's
devfn for match.

Signed-off-by: Honghui Zhang <redacted>
Reviewed-by: Ryder Lee <ryder.lee@mediatek.com>
---
 drivers/pci/controller/pcie-mediatek.c | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)

diff --git a/drivers/pci/controller/pcie-mediatek.c b/drivers/pci/controller/pcie-mediatek.c
index 0baabe3..9cf7ecf 100644
--- a/drivers/pci/controller/pcie-mediatek.c
+++ b/drivers/pci/controller/pcie-mediatek.c

@@ -337,10 +337,23 @@ static struct mtk_pcie_port *mtk_pcie_find_port(struct pci_bus *bus,
 {
 	struct mtk_pcie *pcie = bus->sysdata;
 	struct mtk_pcie_port *port;
+	struct pci_dev *dev;
+	struct pci_bus *pbus;
 
-	list_for_each_entry(port, &pcie->ports, list)
-		if (port->slot == PCI_SLOT(devfn))
+	list_for_each_entry(port, &pcie->ports, list) {

mvebu has the identical hardware structure but uses an array instead
of a list:

  num = of_get_available_child_count(np);
  pcie->ports = devm_kcalloc(dev, num, sizeof(*pcie->ports), GFP_KERNEL);
  for_each_available_child_of_node(np, child) {
    struct mvebu_pcie_port *port = &pcie->ports[i];
    mvebu_pcie_parse_port(pcie, port, child);
  }

It would be nice if mvebu and mtk used the same strategy so the code
looks the same.

Seems there's different approach existing in current host controller
driver for this, tegra PCIe is using list too.
I guess I can change it to using array after this problem is fixed if
you have strong opinion here.

quoted

+		if (bus->number == 0 && port->slot == PCI_SLOT(devfn)) {

Is the root bus number fixed at 0 or is it programmable?  Many drivers
do something like this:

The primary bus number is fixed at 0.(technically speaking, hardware
will capture the very first value of primary bus number, once it's been
captured, change the value in configuration space will not take effort. 

But the hardware default value is 0, the hardware will not response to
other values except 0 at the first time. And the only way to touch the
value in it's configuration space is using primary bus number 0. So it
could be taken as fixed at 0.)

  if (bus->number == pcie->root_bus_nr)

to handle the case of the root bus number being programmable.

quoted

 			return port;
+		} else if (bus->number != 0) {
+			pbus = bus;
+			do {
+				dev = pbus->self;
+				if (port->slot == PCI_SLOT(dev->devfn))
+					return port;
+
+				pbus = dev->bus;
+			} while (dev->bus->number != 0);

You should not need to search up the tree of dev->bus->self.

mvebu_pcie_find_port() checks the root port secondary and subordinate
bus numbers, which should work here, too.

I checked mvebu's host driver, it using software bridge to abstract a
pseudo bridge. And access to this pseudo bridge's configuration space is
all software behavior.

If I following mvebu's way, I need to touch the hardware to read back
the secondary and subordinary bus number, I think current software
solution is more efficient.

Or I could add per port parameter to store the bus_number, and update
the value as mvebu's driver is doing. Do you think it's better than
current solution?

quoted

+		}
+	}
 
 	return NULL;
 }
-- 
2.6.4


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel