Re: [PATCH V10 2/5] PCI: Create device tree node for bridge

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Jun 29, 2023 at 05:56:31PM -0500, Bjorn Helgaas wrote:
> On Thu, Jun 29, 2023 at 10:19:47AM -0700, Lizhi Hou wrote:
> > The PCI endpoint device such as Xilinx Alveo PCI card maps the register
> > spaces from multiple hardware peripherals to its PCI BAR. Normally,
> > the PCI core discovers devices and BARs using the PCI enumeration process.
> > There is no infrastructure to discover the hardware peripherals that are
> > present in a PCI device, and which can be accessed through the PCI BARs.
> 
> IIUC this is basically a multi-function device except that instead of
> each device being a separate PCI Function, they all appear in a single
> Function.  That would mean all the devices share the same config space
> so a single PCI Command register controls all of them, they all share
> the same IRQs (either INTx or MSI/MSI-X), any MMIO registers are likely
> in a shared BAR, etc., right?

Could be multiple BARs, but yes.

> Obviously PCI enumeration only sees the single Function and binds a
> single driver to it.  But IIUC, you want to use existing drivers for
> each of these sub-devices, so this series adds a DT node for the
> single Function (using the quirks that call of_pci_make_dev_node()).
> And I assume that when the PCI driver claims the single Function, it
> will use that DT node to add platform devices, and those existing
> drivers can claim those?

Yes. It will call some variant of of_platform_populate().

> I don't see the PCI driver for the single Function in this series.  Is
> that coming?  Is this series useful without it?

https://lore.kernel.org/all/20220305052304.726050-4-lizhi.hou@xxxxxxxxxx/

I asked for things to be split up as the original series did a lot 
of new things at once. This series only works with the QEMU PCI test 
device which the DT unittest will use.

> > Apparently, the device tree framework requires a device tree node for the
> > PCI device. Thus, it can generate the device tree nodes for hardware
> > peripherals underneath. Because PCI is self discoverable bus, there might
> > not be a device tree node created for PCI devices. Furthermore, if the PCI
> > device is hot pluggable, when it is plugged in, the device tree nodes for
> > its parent bridges are required. Add support to generate device tree node
> > for PCI bridges.
> 
> Can you remind me why hot-adding a PCI device requires DT nodes for
> parent bridges?

Because the PCI device needs a DT node and we can't just put PCI devices 
in the DT root. We have to create the bus hierarchy.

> I don't think we have those today, so maybe the DT
> node for the PCI device requires a DT parent?  How far up does that
> go?

All the way.

>  From this patch, I guess a Root Port would be the top DT node on
> a PCIe system, since that's the top PCI-to-PCI bridge?

Yes. Plus above the host bridge could have a hierarchy of nodes.

> This patch adds a DT node for *every* PCI bridge in the system.  We
> only actually need that node for these unusual devices.  Is there some
> way the driver for the single PCI Function could add that node when it
> is needed?  Sorry if you've answered this in the past; maybe the
> answer could be in the commit log or a code comment in case somebody
> else wonders.

This was discussed early on. I don't think it would work to create the 
nodes at the time we discover we have a device that wants a DT node. The 
issue is decisions are made in the code based on whether there's a DT 
node for a PCI device or not. It might work, but I think it's fragile to 
have nodes attached to devices at different points in time.

> 
> > @@ -340,6 +340,8 @@ void pci_bus_add_device(struct pci_dev *dev)
> >  	 */
> >  	pcibios_bus_add_device(dev);
> >  	pci_fixup_device(pci_fixup_final, dev);
> > +	if (pci_is_bridge(dev))
> > +		of_pci_make_dev_node(dev);
> 
> It'd be nice to have a clue here about why we need this, since this is
> executed for *every* system, even ACPI platforms that typically don't
> use OF things.
> 
> >  	pci_create_sysfs_dev_files(dev);
> >  	pci_proc_attach_device(dev);
> >  	pci_bridge_d3_update(dev);
> > diff --git a/drivers/pci/of.c b/drivers/pci/of.c
> > index 2c25f4fa0225..9786ae407948 100644
> > --- a/drivers/pci/of.c
> > +++ b/drivers/pci/of.c
> > @@ -487,6 +487,15 @@ static int of_irq_parse_pci(const struct pci_dev *pdev, struct of_phandle_args *
> >  		} else {
> >  			/* We found a P2P bridge, check if it has a node */
> >  			ppnode = pci_device_to_OF_node(ppdev);
> > +#if IS_ENABLED(CONFIG_PCI_DYNAMIC_OF_NODES)
> 
> I would use plain #ifdef here instead of IS_ENABLED(), as you did in
> pci.h below.  IS_ENABLED() is true if the Kconfig symbol is set to
> either "y" or "m".
> 
> Using IS_ENABLED() suggests that the config option *could* be a
> module, which is not the case here because CONFIG_PCI_DYNAMIC_OF_NODES
> is a bool.
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/kconfig.h?id=v6.4#n69
> 
> > @@ -617,6 +626,85 @@ int devm_of_pci_bridge_init(struct device *dev, struct pci_host_bridge *bridge)
> >  	return pci_parse_request_of_pci_ranges(dev, bridge);
> >  }
> >  
> > +#if IS_ENABLED(CONFIG_PCI_DYNAMIC_OF_NODES)
> 
> Same here, of course.
> 
> > +void of_pci_remove_node(struct pci_dev *pdev)
> > +{
> > +	struct device_node *np;
> > +
> > +	np = pci_device_to_OF_node(pdev);
> > +	if (!np || !of_node_check_flag(np, OF_DYNAMIC))
> 
> > + * Each entry in the ranges table is a tuple containing the child address,
> > + * the parent address, and the size of the region in the child address space.
> > + * Thus, for PCI, in each entry parent address is an address on the primary
> > + * side and the child address is the corresponding address on the secondary
> > + * side.
> > + */
> > +struct of_pci_range {
> > +	u32		child_addr[OF_PCI_ADDRESS_CELLS];
> > +	u32		parent_addr[OF_PCI_ADDRESS_CELLS];
> > +	u32		size[OF_PCI_SIZE_CELLS];
> 
> > +		if (pci_is_bridge(pdev)) {
> > +			memcpy(rp[i].child_addr, rp[i].parent_addr,
> > +			       sizeof(rp[i].child_addr));
> > +		} else {
> > +			/*
> > +			 * For endpoint device, the lower 64-bits of child
> > +			 * address is always zero.
> 
> I think this connects with the secondary side comment above, right?  I
> think I would comment this as:
> 
>   /*
>    * PCI-PCI bridges don't translate addresses, so the child
>    * (secondary side) address is identical to the parent (primary
>    * side) address.
>    */
> 
> and
> 
>   /*
>    * Non-bridges have no child (secondary side) address, so clear it
>    * out.
>    */
> 
> > +			 */
> > +			rp[i].child_addr[0] = j;
> 
> > +	ret = of_changeset_add_empty_prop(ocs, np, "dynamic");
> 
> It seems slightly confusing to use a "dynamic" property here when we
> also have the OF_DYNAMIC dynamic flag above.  I think they have
> different meanings, don't they?

Hum, what's the property for? It's new in this version. Any DT property 
needs to be documented, but I don't see why we need it.

Rob



[Index of Archives]     [Device Tree Compilter]     [Device Tree Spec]     [Linux Driver Backports]     [Video for Linux]     [Linux USB Devel]     [Linux PCI Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Yosemite Backpacking]


  Powered by Linux