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

[Ard Biesheuvel]

On Thu, 28 Nov 2019 at 20:29, Saravana Kannan <saravanak@xxxxxxxxxx> wrote:
>
> 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!
>

Sure, no trouble at all.

> > 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.
>

It is the probe reversal between the efi-framebuffer on the one hand
and the entire PCIe hierarchy on the other.

For some reason, PCIe host controllers are usually probed very early,
and I wouldn't be surprised if deferring that may cause other issues
as well. However, of_devlink is presumably specific to DT systems,
where PCIe does not play such a fundamental role like it does on x86,
for instance.

> > 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.
>

Yeah, I realized that after I posted it.

> > +
> > +               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.
>

OK

> > +                               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.
>

Indeed.

> > +       }
> > +       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 should have mentioned this in the commit log, I suppose: I copied
this from the x86 code that registers the efifb platform device, it
also uses device_initcall() to prevent probing too early.

> 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.
>

I have no idea how that would look, Could you elaborate? I'd prefer it
if we could have a solution where this logic is only invoked when
necessary, i.e., when we are using device links in the first place.

Thanks,
Ard.