Re: [PATCH 1/3] [RFC] pci: add new method for register PCI hosts

[Thierry Reding]

On Sat, Apr 30, 2016 at 01:01:37AM +0200, Arnd Bergmann wrote:
> This patch makes the existing 'pci_host_bridge' structure a proper device
> that is usable by PCI host drivers in a more standard way. In addition
> to the existing pci_scan_bus, pci_scan_root_bus, pci_scan_root_bus_msi,
> and pci_create_root_bus interfaces, this unfortunately means having to
> add yet another interface doing basically the same thing, and add some
> extra code in the initial step.
> 
> However, this time it's more likely to be extensible enough that we
> won't have to do another one again in the future, and we should be
> able to reduce code much more as a result.
> 
> The main idea is to pull the allocation of 'struct pci_host_bridge' out
> of the registration, and let individual host drivers and architecture
> code fill the members before calling the registration function.
> 
> There are a number of things we can do based on this:
> 
> * Use a single memory allocation for the driver-specific structure
>   and the generic PCI host bridge
> * consolidate the contents of driver specific structures by moving
>   them into pci_host_bridge
> * Add a consistent interface for removing a PCI host bridge again
>   when unloading a host driver module
> * Replace the architecture specific __weak pcibios_* functions with
>   callbacks in a pci_host_bridge device
> * Move common boilerplate code from host drivers into the generic
>   function, based on contents of the structure
> * Extend pci_host_bridge with additional members when needed without
>   having to add arguments to pci_scan_*.
> * Move members of struct pci_bus into pci_host_bridge to avoid
>   having lots of identical copies.
> 
> As mentioned in a previous email, one open question is whether we want
> to export a function for allocating a pci_host_bridge device in
> combination with the per-device structure or let the driver itself
> call kzalloc.

I think the most common pattern in other parts of the kernel is the
latter. That gives drivers the most flexibility to do whatever they
want or need.

> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> ---
>  drivers/pci/probe.c | 100 ++++++++++++++++++++++++++++++----------------------
>  include/linux/pci.h |   7 +++-
>  2 files changed, 63 insertions(+), 44 deletions(-)
> 
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index ae7daeb83e21..fe9d9221b11e 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -520,19 +520,6 @@ static void pci_release_host_bridge_dev(struct device *dev)
>  	kfree(bridge);
>  }
>  
> -static struct pci_host_bridge *pci_alloc_host_bridge(struct pci_bus *b)
> -{
> -	struct pci_host_bridge *bridge;
> -
> -	bridge = kzalloc(sizeof(*bridge), GFP_KERNEL);
> -	if (!bridge)
> -		return NULL;
> -
> -	INIT_LIST_HEAD(&bridge->windows);
> -	bridge->bus = b;
> -	return bridge;
> -}
> -
>  static const unsigned char pcix_bus_speed[] = {
>  	PCI_SPEED_UNKNOWN,		/* 0 */
>  	PCI_SPEED_66MHz_PCIX,		/* 1 */
> @@ -2108,51 +2095,47 @@ void __weak pcibios_remove_bus(struct pci_bus *bus)
>  {
>  }
>  
> -struct pci_bus *pci_create_root_bus(struct device *parent, int bus,
> -		struct pci_ops *ops, void *sysdata, struct list_head *resources)
> +int pci_register_host(struct pci_host_bridge *bridge)

Perhaps pci_register_host_bridge() to mirror the structure name in the
registration function?

>  {
>  	int error;
> -	struct pci_host_bridge *bridge;
>  	struct pci_bus *b, *b2;
>  	struct resource_entry *window, *n;
> +	LIST_HEAD(resources);
>  	struct resource *res;
>  	resource_size_t offset;
>  	char bus_addr[64];
>  	char *fmt;
> +	struct device *parent = bridge->dev.parent;
>  
>  	b = pci_alloc_bus(NULL);
>  	if (!b)
> -		return NULL;
> +		return -ENOMEM;
> +	bridge->bus = b;
>  
> -	b->sysdata = sysdata;
> -	b->ops = ops;
> -	b->number = b->busn_res.start = bus;
> +	/* temporarily move resources off the list */

Might be worth mentioning why we move the resources off the list.

> +	list_splice_init(&bridge->windows, &resources);
> +	b->sysdata = bridge->sysdata;

Does the sysdata not become effectively obsolete after this series? My
understanding is that it's primarily used to store driver-specific data
along with a PCI bus, but if drivers can embed struct pci_host_bridge
they can simply upcast bus->bridge. I do notice that bus->bridge is
currently a struct device *, perhaps we can replace it by a back pointer
to the parent struct pci_host_bridge, which would have to gain a struct
device *parent to point at the device that instantiated the bridge. This
is becoming somewhat complicated, but maybe that can be simplified at
some point.

> +	b->msi = bridge->msi;
> +	b->ops = bridge->ops;
> +	b->number = b->busn_res.start = bridge->busnr;
>  	pci_bus_assign_domain_nr(b, parent);
> -	b2 = pci_find_bus(pci_domain_nr(b), bus);
> +	b2 = pci_find_bus(pci_domain_nr(b), bridge->busnr);
>  	if (b2) {
>  		/* If we already got to this bus through a different bridge, ignore it */
>  		dev_dbg(&b2->dev, "bus already known\n");
> +		error = -EEXIST;
>  		goto err_out;
>  	}
>  
> -	bridge = pci_alloc_host_bridge(b);
> -	if (!bridge)
> -		goto err_out;
> -
> -	bridge->dev.parent = parent;
> -	bridge->dev.release = pci_release_host_bridge_dev;
> -	dev_set_name(&bridge->dev, "pci%04x:%02x", pci_domain_nr(b), bus);
> +	dev_set_name(&bridge->dev, "pci%04x:%02x", pci_domain_nr(b), bridge->busnr);
>  	error = pcibios_root_bridge_prepare(bridge);
> -	if (error) {
> -		kfree(bridge);
> +	if (error)
>  		goto err_out;
> -	}
>  
>  	error = device_register(&bridge->dev);
> -	if (error) {
> +	if (error)
>  		put_device(&bridge->dev);
> -		goto err_out;
> -	}
> +
>  	b->bridge = get_device(&bridge->dev);

I'm not sure I understand why we continue after failing to register the
device. Is the usage model here that drivers set up bridge->dev with an
initial set of values here, such as what the bridge->dev.parent is? One
complication I can imagine with that is that drivers would need to have
an implementation for the bridge device's ->release() callback. Perhaps
this could be simplified by having a default release callback (maybe
set up by pci_register_host() if none was specified by the driver) that
calls a callback in struct pci_host_bridge which gets passed a struct
pci_host_bridge. I think that would make usage more uniform from the
driver perspective.

On a side-note, perhaps it would be worth adding a structure that
carries host bridge operations (struct pci_host_bridge_ops)?

> diff --git a/include/linux/pci.h b/include/linux/pci.h
> index 81f070a47ee7..8bb5dff617a1 100644
> --- a/include/linux/pci.h
> +++ b/include/linux/pci.h
> @@ -400,10 +400,14 @@ static inline int pci_channel_offline(struct pci_dev *pdev)
>  
>  struct pci_host_bridge {
>  	struct device dev;
> -	struct pci_bus *bus;		/* root bus */
> +	struct pci_ops *ops;
> +	void *sysdata;
> +	int busnr;

While at it, is there any reason why this can't be made unsigned? I know
this must sound pedantic, but whenever I see a signed integer variable I
immediately ask myself what the meaning of negative values would be, and
I can't think of any scenario where this one could possible be negative.
But perhaps I'm missing something?

Thierry

Attachment: signature.asc
Description: PGP signature