[PATCH V7 04/17] PCI: Take additional IOV BAR alignment in sizing and assigning

[Bjorn Helgaas]

On Thu, Jul 24, 2014 at 02:22:14PM +0800, Wei Yang wrote:
> At resource sizing/assigning stage, resources are divided into two lists,
> requested list and additional list, while the alignement of the additional
> IOV BAR is not taken into the sizeing and assigning procedure.
> 
> This is reasonable in the original implementation, since IOV BAR's alignment is
> mostly the size of a PF BAR alignemt. This means the alignment is already taken
> into consideration. While this rule may be violated on some platform.
> 
> This patch take the additional IOV BAR alignment in sizing and assigning stage
> explicitly.
> 
> Signed-off-by: Wei Yang <weiyang at linux.vnet.ibm.com>
> ---
>  drivers/pci/setup-bus.c |   68 +++++++++++++++++++++++++++++++++++++++++------
>  1 file changed, 60 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/pci/setup-bus.c b/drivers/pci/setup-bus.c
> index a5a63ec..d83681f 100644
> --- a/drivers/pci/setup-bus.c
> +++ b/drivers/pci/setup-bus.c
> @@ -120,6 +120,28 @@ static resource_size_t get_res_add_size(struct list_head *head,
>  	return 0;
>  }
>  
> +static resource_size_t get_res_add_align(struct list_head *head,
> +		struct resource *res)
> +{
> +	struct pci_dev_resource *dev_res;
> +
> +	list_for_each_entry(dev_res, head, list) {
> +		if (dev_res->res == res) {
> +			int idx = res - &dev_res->dev->resource[0];
> +
> +			dev_printk(KERN_DEBUG, &dev_res->dev->dev,
> +				   "res[%d]=%pR get_res_add_align min_align %llx\n",
> +				   idx, dev_res->res,
> +				   (unsigned long long)dev_res->min_align);
> +
> +			return dev_res->min_align;
> +		}
> +	}
> +
> +	return 0;
> +}

I see that you copied the structure of the existing get_res_add_size()
here.  But I don't understand *that* function.  It looks basically like
this:

  resource_size_t get_res_add_size(list, res)
  {
    list_for_each_entry(dev_res, head, list) {
      if (dev_res->res == res)
        return dev_res->add_size;
    }
    return 0;
  }

and we call it like this:

  dev_res->res->end += get_res_add_size(realloc_head, dev_res->res);

So we start out with dev_res", pass in dev_res->res, search the
realloc_head list to find dev_res again, and return dev_res->add_size.
That looks equivalent to just:

  dev_res->res->end += dev_res->add_size;

It looks like get_res_add_size() merely adds a printk and some complexity.
Am I missing something?

I do see that there are other callers where we don't actually start with
dev_res, which makes it a little more complicated.  But I think you should
either add something like this:

  struct pci_dev_resource *res_to_dev_res(list, res)
  {
    list_for_each_entry(dev_res, head, list) {
      if (dev_res->res == res)
        return dev_res;
    }
    return NULL;
  }

which can be used to replace get_res_add_size() and get_res_add_align(), OR
figure out whether the dev_res of interest is always one we've just added.
If it is, maybe you can just make add_to_list() return the dev_res pointer
instead of an errno, and hang onto the pointer.  I'd like that much better
if that's possible.

> +
> +
>  /* Sort resources by alignment */
>  static void pdev_sort_resources(struct pci_dev *dev, struct list_head *head)
>  {
> @@ -368,8 +390,9 @@ static void __assign_resources_sorted(struct list_head *head,
>  	LIST_HEAD(save_head);
>  	LIST_HEAD(local_fail_head);
>  	struct pci_dev_resource *save_res;
> -	struct pci_dev_resource *dev_res, *tmp_res;
> +	struct pci_dev_resource *dev_res, *tmp_res, *dev_res2;
>  	unsigned long fail_type;
> +	resource_size_t add_align, align;
>  
>  	/* Check if optional add_size is there */
>  	if (!realloc_head || list_empty(realloc_head))
> @@ -384,10 +407,31 @@ static void __assign_resources_sorted(struct list_head *head,
>  	}
>  
>  	/* Update res in head list with add_size in realloc_head list */
> -	list_for_each_entry(dev_res, head, list)
> +	list_for_each_entry_safe(dev_res, tmp_res, head, list) {
>  		dev_res->res->end += get_res_add_size(realloc_head,
>  							dev_res->res);
>  
> +		if (!(dev_res->res->flags & IORESOURCE_STARTALIGN))
> +			continue;
> +
> +		add_align = get_res_add_align(realloc_head, dev_res->res);
> +
> +		if (add_align > dev_res->res->start) {
> +			dev_res->res->start = add_align;
> +			dev_res->res->end = add_align +
> +				            resource_size(dev_res->res);
> +
> +			list_for_each_entry(dev_res2, head, list) {
> +				align = pci_resource_alignment(dev_res2->dev,
> +							       dev_res2->res);
> +				if (add_align > align)
> +					list_move_tail(&dev_res->list,
> +						       &dev_res2->list);
> +			}
> +               }
> +
> +	}
> +
>  	/* Try updated head list with add_size added */
>  	assign_requested_resources_sorted(head, &local_fail_head);
>  
> @@ -930,6 +974,8 @@ static int pbus_size_mem(struct pci_bus *bus, unsigned long mask,
>  	struct resource *b_res = find_free_bus_resource(bus,
>  					mask | IORESOURCE_PREFETCH, type);
>  	resource_size_t children_add_size = 0;
> +	resource_size_t children_add_align = 0;
> +	resource_size_t add_align = 0;
>  
>  	if (!b_res)
>  		return -ENOSPC;
> @@ -954,6 +1000,7 @@ static int pbus_size_mem(struct pci_bus *bus, unsigned long mask,
>  			/* put SRIOV requested res to the optional list */
>  			if (realloc_head && i >= PCI_IOV_RESOURCES &&
>  					i <= PCI_IOV_RESOURCE_END) {
> +				add_align = max(pci_resource_alignment(dev, r), add_align);
>  				r->end = r->start - 1;
>  				add_to_list(realloc_head, dev, r, r_size, 0/* don't care */);
>  				children_add_size += r_size;
> @@ -984,8 +1031,11 @@ static int pbus_size_mem(struct pci_bus *bus, unsigned long mask,
>  			if (order > max_order)
>  				max_order = order;
>  
> -			if (realloc_head)
> +			if (realloc_head) {
>  				children_add_size += get_res_add_size(realloc_head, r);
> +				children_add_align = get_res_add_align(realloc_head, r);
> +				add_align = max(add_align, children_add_align);
> +			}
>  		}
>  	}
>  
> @@ -996,7 +1046,7 @@ static int pbus_size_mem(struct pci_bus *bus, unsigned long mask,
>  		add_size = children_add_size;
>  	size1 = (!realloc_head || (realloc_head && !add_size)) ? size0 :
>  		calculate_memsize(size, min_size, add_size,
> -				resource_size(b_res), min_align);
> +				resource_size(b_res), max(min_align, add_align));
>  	if (!size0 && !size1) {
>  		if (b_res->start || b_res->end)
>  			dev_info(&bus->self->dev, "disabling bridge window %pR to %pR (unused)\n",
> @@ -1008,10 +1058,12 @@ static int pbus_size_mem(struct pci_bus *bus, unsigned long mask,
>  	b_res->end = size0 + min_align - 1;
>  	b_res->flags |= IORESOURCE_STARTALIGN;
>  	if (size1 > size0 && realloc_head) {
> -		add_to_list(realloc_head, bus->self, b_res, size1-size0, min_align);
> -		dev_printk(KERN_DEBUG, &bus->self->dev, "bridge window %pR to %pR add_size %llx\n",
> -			   b_res, &bus->busn_res,
> -			   (unsigned long long)size1-size0);
> +		add_to_list(realloc_head, bus->self, b_res, size1-size0,
> +				max(min_align, add_align));
> +		dev_printk(KERN_DEBUG, &bus->self->dev, "bridge window "
> +				 "%pR to %pR add_size %llx add_align %llx\n", b_res,
> +				 &bus->busn_res, (unsigned long long)size1-size0,
> +				 max(min_align, add_align));

Factor out this "max(min_align, add_align)" thing so we don't have to
change these lines.  Bonus points if you can also factor it out of the
calculate_memsize() call above.  That one is a pretty complicated ternary
expression that should probably be turned into an "if" instead anyway.

>  	}
>  	return 0;
>  }
> -- 
> 1.7.9.5
>