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

From: Saravana Kannan
Date: Tue Dec 17 2019 - 21:15:01 EST


On Thu, Nov 28, 2019 at 12:19 PM Ard Biesheuvel
<ard.biesheuvel@xxxxxxxxxx> wrote:
>
> 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.

I haven't forgotten this thread -- it's in my TODO list. I'm hoping to
get to this during the holiday weeks. I plan on sending an example
patch with some of your code in it and you can take it from there.
Does that sound good?

Thanks,
Saravana