Re: [PATCH] efi: arm: defer probe of PCIe backed efifb on DT systems

[Saravana Kannan]

On Tue, Nov 26, 2019 at 8:30 AM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
>
> The new of_devlink support breaks PCIe probing on ARM platforms booting
> via UEFI if the firmware exposes a EFI framebuffer that is backed by a
> PCI device.

Thanks for testing with of_devlink enabled!

> The reason is that the probing order gets reversed,
> resulting in a resource conflict on the framebuffer memory window when
> the PCIe probes last, causing it to give up entirely.

Just so I understand it clearly, the probe order reversal is only
between this efi-framebuffer device and the PCIe device right? Not all
PCI devices or something like that, right? Do you have any info on
what dependency causes this reversal? Just curious.

> Given that we rely on PCI quirks to deal with EFI framebuffers that get
> moved around in memory, we cannot simply drop the memory reservation, so
> instead, let's use the device link infrastructure to register this
> dependency, and force the probing to occur in the expected order.
>
> Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> Cc: Saravana Kannan <saravanak@xxxxxxxxxx>
> Signed-off-by: Ard Biesheuvel <ardb@xxxxxxxxxx>
> ---
>  drivers/firmware/efi/arm-init.c | 66 ++++++++++++++++++--
>  1 file changed, 61 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/firmware/efi/arm-init.c b/drivers/firmware/efi/arm-init.c
> index 311cd349a862..617226d50774 100644
> --- a/drivers/firmware/efi/arm-init.c
> +++ b/drivers/firmware/efi/arm-init.c
> @@ -14,6 +14,7 @@
>  #include <linux/memblock.h>
>  #include <linux/mm_types.h>
>  #include <linux/of.h>
> +#include <linux/of_address.h>
>  #include <linux/of_fdt.h>
>  #include <linux/platform_device.h>
>  #include <linux/screen_info.h>
> @@ -267,15 +268,70 @@ void __init efi_init(void)
>                 efi_memmap_unmap();
>  }
>
> +static bool __init efifb_overlaps_pci_range(const struct of_pci_range *range)
> +{
> +       u64 fb_base = screen_info.lfb_base;
> +
> +       if (screen_info.capabilities & VIDEO_CAPABILITY_64BIT_BASE)
> +               fb_base |= (u64)(unsigned long)screen_info.ext_lfb_base << 32;
> +
> +       return fb_base >= range->cpu_addr &&
> +              fb_base < (range->cpu_addr + range->size);
> +}
> +
>  static int __init register_gop_device(void)
>  {
> -       void *pd;
> +       struct platform_device *pd;
> +       struct device_node *np;
> +       bool found = false;
> +       int err;
>
>         if (screen_info.orig_video_isVGA != VIDEO_TYPE_EFI)
>                 return 0;
>
> -       pd = platform_device_register_data(NULL, "efi-framebuffer", 0,
> -                                          &screen_info, sizeof(screen_info));
> -       return PTR_ERR_OR_ZERO(pd);
> +       pd = platform_device_alloc("efi-framebuffer", 0);
> +       if (!pd)
> +               return -ENOMEM;
> +
> +       err = platform_device_add_data(pd, &screen_info, sizeof(screen_info));
> +       if (err)
> +               return err;
> +
> +       /*
> +        * If the efifb framebuffer is backed by a PCI graphics controller, we
> +        * have to ensure that this relation is expressed using a device link
> +        * when running in DT mode, or the probe order may be reversed,
> +        * resulting in a resource reservation conflict on the memory window
> +        * that the efifb framebuffer steals from the PCIe host bridge.
> +        */
> +       for_each_node_by_type(np, "pci") {
> +               struct of_pci_range_parser parser;
> +               struct of_pci_range range;
> +               struct device *sup_dev;
> +
> +               if (found) {
> +                       of_node_put(np);
> +                       break;
> +               }

It looks like you are doing this here because you can't break out of
two loops when you set found = true. Is that right? If so, I think
doing this at the end of the loop would make it more obvious on what's
going on.

> +
> +               err = of_pci_range_parser_init(&parser, np);
> +               if (err) {
> +                       pr_warn("of_pci_range_parser_init() failed: %d\n", err);
> +                       continue;
> +               }
> +
> +               sup_dev = get_dev_from_fwnode(&np->fwnode);
> +
> +               for_each_of_pci_range(&parser, &range) {
> +                       if (efifb_overlaps_pci_range(&range)) {
> +                               found = true;
> +                               if (!device_link_add(&pd->dev, sup_dev, 0))
> +                                       pr_warn("device_link_add() failed\n");

I think dev_warn(&pd->dev,...) might make the message more useful.
Otherwise, it's so confusing.

> +                               break;
> +                       }
> +               }
> +               put_device(sup_dev);

Can't you do the if (found) here? Another option is to simply do a
"goto out;" at the end of the if block where you set found = true.

> +       }
> +       return platform_device_add(pd);
>  }
> -subsys_initcall(register_gop_device);
> +device_initcall(register_gop_device);

Looks like you are doing this so that this efi-framebuffer device gets
added after the PCIe device? So that device_add_link() succeeds?

I'm wondering if it would be better to implement this as a
fwnode_operations.add_links(). Since this efi-framebuffer device won't have any
fwnode, you can create your own fwnode and implement the add_links()
property. Not a strong opinion on this, but some food for thought.

Thanks,
Saravana