Re: [PATCH v8 08/22] PCI, acpiphp: Separate out hot-add support of pci host bridge

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Friday, January 11, 2013 02:40:35 PM Yinghai Lu wrote:
> It causes confusion.
> 
> We may only need acpi hp for pci host bridge.

What does this mean?

> Split host bridge hot-add support to pci_root_hp, and keep acpiphp simple.

s/Split/Move/ I suppose?

In any case that's not telling the whole story, because the patch doesn't just
move code from one file to another.

> -v2: put back pci_root_hp change in one patch
> -v3: add pcibios_resource_survey_bus() calling
> -v4: remove not needed code with remove_bridge
> -v5: put back support for acpiphp support for slots just on root bus.
> -v6: change some functions to *_p2p_* to make it more clean.
> -v7: split hot_added change out.
> -v8: Move to pci_root.c instead of adding another file requested by Bjorn.
> 
> Signed-off-by: Yinghai Lu <yinghai@kernel.org>
> ---
>  drivers/acpi/pci_root.c            |  220 ++++++++++++++++++++++++++++++++++++
>  drivers/pci/hotplug/acpiphp_glue.c |   59 +++-------
>  2 files changed, 235 insertions(+), 44 deletions(-)
> 
> diff --git a/drivers/acpi/pci_root.c b/drivers/acpi/pci_root.c
> index 471b2dc..5c1f462c 100644
> --- a/drivers/acpi/pci_root.c
> +++ b/drivers/acpi/pci_root.c
> @@ -673,3 +673,223 @@ int __init acpi_pci_root_init(void)
>  
>  	return 0;
>  }
> +
> +/*
> + * Separated from drivers/pci/hotplug/acpiphp_glue.c
> + *	only support root bridge
> + */

This comment will be useless after applying the patch.

> +
> +static LIST_HEAD(acpi_root_bridge_list);
> +struct acpi_root_bridge {
> +	struct list_head list;
> +	acpi_handle handle;
> +	u32 flags;
> +};

We have struct acpi_pci_root already.  Why do we need this in addition?

Also, we have acpi_pci_roots, so why do we need another list of root bridges?

> +
> +/* bridge flags */
> +#define ROOT_BRIDGE_HAS_EJ0	(0x00000002)
> +#define ROOT_BRIDGE_HAS_PS3	(0x00000080)

What is that needed for?

> +
> +#define ACPI_STA_FUNCTIONING	(0x00000008)
> +
> +static struct acpi_root_bridge *acpi_root_handle_to_bridge(acpi_handle handle)
> +{
> +	struct acpi_root_bridge *bridge;
> +
> +	list_for_each_entry(bridge, &acpi_root_bridge_list, list)
> +		if (bridge->handle == handle)
> +			return bridge;
> +
> +	return NULL;
> +}
> +
> +/* allocate and initialize host bridge data structure */
> +static void add_acpi_root_bridge(acpi_handle handle)
> +{
> +	struct acpi_root_bridge *bridge;
> +	acpi_handle dummy_handle;
> +	acpi_status status;
> +

Why do we need to evaluate all of the methods directly here?

Don't we have a struct acpi_device for handle already?

> +	/* if the bridge doesn't have _STA, we assume it is always there */
> +	status = acpi_get_handle(handle, "_STA", &dummy_handle);
> +	if (ACPI_SUCCESS(status)) {
> +		unsigned long long tmp;
> +
> +		status = acpi_evaluate_integer(handle, "_STA", NULL, &tmp);
> +		if (ACPI_FAILURE(status)) {
> +			printk(KERN_DEBUG "%s: _STA evaluation failure\n",
> +						 __func__);
> +			return;
> +		}
> +		if ((tmp & ACPI_STA_FUNCTIONING) == 0)
> +			/* don't register this object */
> +			return;
> +	}
> +
> +	bridge = kzalloc(sizeof(struct acpi_root_bridge), GFP_KERNEL);
> +	if (!bridge)
> +		return;
> +
> +	bridge->handle = handle;
> +
> +	if (ACPI_SUCCESS(acpi_get_handle(handle, "_EJ0", &dummy_handle)))
> +		bridge->flags |= ROOT_BRIDGE_HAS_EJ0;
> +	if (ACPI_SUCCESS(acpi_get_handle(handle, "_PS3", &dummy_handle)))
> +		bridge->flags |= ROOT_BRIDGE_HAS_PS3;

All of this attempts to duplicate the scanning code from scan.c in a very
incomplete and questionable way.

For example, what if the root bridge has _PR0?

> +
> +	list_add(&bridge->list, &acpi_root_bridge_list);
> +}
> +
> +struct acpi_root_hp_work {
> +	struct work_struct work;
> +	acpi_handle handle;
> +	u32 type;
> +	void *context;
> +};
> +
> +static void alloc_acpi_root_hp_work(acpi_handle handle, u32 type,
> +					void *context,
> +					void (*func)(struct work_struct *work))
> +{
> +	struct acpi_root_hp_work *hp_work;
> +	int ret;
> +
> +	hp_work = kmalloc(sizeof(*hp_work), GFP_KERNEL);
> +	if (!hp_work)
> +		return;
> +
> +	hp_work->handle = handle;
> +	hp_work->type = type;
> +	hp_work->context = context;
> +
> +	INIT_WORK(&hp_work->work, func);
> +	ret = queue_work(kacpi_hotplug_wq, &hp_work->work);
> +	if (!ret)
> +		kfree(hp_work);
> +}

The function above is called only once and used by __init stuff only.
Why don't we move it to the caller and mark that caller as __init too?

> +
> +static void handle_root_bridge_insertion(acpi_handle handle)
> +{
> +	struct acpi_device *device, *pdevice;
> +	acpi_handle phandle;
> +	int ret_val;
> +
> +	acpi_get_parent(handle, &phandle);
> +	if (acpi_bus_get_device(phandle, &pdevice)) {
> +		printk(KERN_DEBUG "no parent device, assuming NULL\n");
> +		pdevice = NULL;
> +	}
> +	if (!acpi_bus_get_device(handle, &device)) {
> +		/* check if  pci root_bus is removed */
> +		struct acpi_pci_root *root = acpi_driver_data(device);
> +		if (pci_find_bus(root->segment, root->secondary.start))
> +			return;
> +
> +		printk(KERN_DEBUG "bus exists... trim\n");
> +		/* this shouldn't be in here, so remove
> +		 * the bus then re-add it...
> +		 */

Why?  Shouldn't we just bail out here?

> +		/* remove pci devices at first */
> +		ret_val = acpi_bus_trim(device, 0);
> +		printk(KERN_DEBUG "acpi_bus_trim stop return %x\n", ret_val);
> +		/* remove acpi devices */
> +		ret_val = acpi_bus_trim(device, 1);

Oh, I see why you need the second argument of acpi_bus_trim() now.

Do I think correctly that you want acpi_pci_root_remove() to be executed before
all of the struct acpi_device objects are removed?  In which case why don't we
call acpi_pci_root_remove() directly before doing the acpi_bus_trim(device, 1)
instead of making the code next to impossible to understand?

> +		printk(KERN_DEBUG "acpi_bus_trim remove return %x\n", ret_val);
> +	}
> +	if (acpi_bus_add(handle))
> +		printk(KERN_ERR "cannot add bridge to acpi list\n");
> +}
> +
> +static void _handle_hotplug_event_root(struct work_struct *work)
> +{
> +	struct acpi_root_bridge *bridge;
> +	char objname[64];
> +	struct acpi_buffer buffer = { .length = sizeof(objname),
> +				      .pointer = objname };
> +	struct acpi_root_hp_work *hp_work;
> +	acpi_handle handle;
> +	u32 type;
> +
> +	hp_work = container_of(work, struct acpi_root_hp_work, work);
> +	handle = hp_work->handle;
> +	type = hp_work->type;
> +
> +	bridge = acpi_root_handle_to_bridge(handle);
> +
> +	acpi_get_name(handle, ACPI_FULL_PATHNAME, &buffer);
> +
> +	switch (type) {
> +	case ACPI_NOTIFY_BUS_CHECK:
> +		/* bus enumerate */
> +		printk(KERN_DEBUG "%s: Bus check notify on %s\n", __func__,
> +				 objname);
> +		if (!bridge) {
> +			handle_root_bridge_insertion(handle);

I don't think we should call add_acpi_root_bridge() for handle if the above
fails.  So probably handle_root_bridge_insertion() should return error codes?

> +			add_acpi_root_bridge(handle);
> +		}
> +
> +		break;
> +
> +	case ACPI_NOTIFY_DEVICE_CHECK:
> +		/* device check */
> +		printk(KERN_DEBUG "%s: Device check notify on %s\n", __func__,
> +				 objname);
> +		if (!bridge) {
> +			handle_root_bridge_insertion(handle);
> +			add_acpi_root_bridge(handle);
> +		}
> +		break;
> +
> +	default:
> +		printk(KERN_WARNING "notify_handler: unknown event type 0x%x for %s\n",
> +				 type, objname);
> +		break;
> +	}
> +
> +	kfree(hp_work); /* allocated in handle_hotplug_event_bridge */
> +}
> +
> +static void handle_hotplug_event_root(acpi_handle handle, u32 type,
> +					void *context)
> +{
> +	alloc_acpi_root_hp_work(handle, type, context,
> +				_handle_hotplug_event_root);
> +}
> +
> +static acpi_status __init
> +find_root_bridges(acpi_handle handle, u32 lvl, void *context, void **rv)
> +{
> +	char objname[64];
> +	struct acpi_buffer buffer = { .length = sizeof(objname),
> +				      .pointer = objname };
> +	int *count = (int *)context;
> +
> +	if (!acpi_is_root_bridge(handle))
> +		return AE_OK;
> +
> +	(*count)++;
> +
> +	acpi_get_name(handle, ACPI_FULL_PATHNAME, &buffer);
> +
> +	acpi_install_notify_handler(handle, ACPI_SYSTEM_NOTIFY,
> +				handle_hotplug_event_root, NULL);
> +	printk(KERN_DEBUG "acpi root: %s notify handler installed\n", objname);
> +
> +	add_acpi_root_bridge(handle);
> +
> +	return AE_OK;
> +}
> +
> +static int __init acpi_pci_root_hp_init(void)
> +{
> +	int num = 0;
> +
> +	acpi_walk_namespace(ACPI_TYPE_DEVICE, ACPI_ROOT_OBJECT,
> +		ACPI_UINT32_MAX, find_root_bridges, NULL, &num, NULL);
> +
> +	printk(KERN_DEBUG "Found %d acpi root devices\n", num);
> +
> +	return 0;
> +}

Why do we need to do that from an initcall?  Couldn't we simply hook up
that code to acpi_pci_root_add() somewhere?

And even if not, why don't we call acpi_pci_root_hp_init() from
acpi_pci_root_init()?

All of the changes in acpiphp_glue.c look reasonable to me.

Thanks,
Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.