Re: [PATCH pci-next v6 1/2] PCI/VGA: Make the vga_is_firmware_default() less arch-dependent

From: Bjorn Helgaas
Date: Wed Oct 04 2023 - 10:01:12 EST


On Wed, Oct 04, 2023 at 08:55:04PM +0800, Sui Jingfeng wrote:
> On 2023/10/3 23:54, Bjorn Helgaas wrote:
> > On Mon, Oct 02, 2023 at 08:05:10PM +0800, Sui Jingfeng wrote:
> > > Currently, the vga_is_firmware_default() function only works on x86 and
> > > ia64, it is a no-op on the rest of the architectures. This patch completes
> > > the implementation for it, the added code tries to capture the PCI (e) VGA
> > > device that owns the firmware framebuffer, since only one GPU could own
> > > the firmware fb, things are almost done once we have determined the boot
> > > VGA device. As the PCI resource relocation do have a influence on the
> > > results of identification, we make it available on architectures where PCI
> > > resource relocation does happen at first. Because this patch is more
> > > important for those architectures(such as arm, arm64, loongarch, mips and
> > > risc-v etc).
> >
> > There's a little too much going on this this patch. The problem is
> > very simple: currently we compare firmware BAR assignments with BARs
> > that may have been reassigned by Linux.
> >
> > What if we did something like the patch below? I think it will be
> > less confusing if we only have one copy of the code related to
> > screen_info.
> >
> > > Signed-off-by: Sui Jingfeng <suijingfeng@xxxxxxxxxxx>
> > > ---
> > > drivers/pci/vgaarb.c | 76 ++++++++++++++++++++++++++++++++++++++++++++
> > > 1 file changed, 76 insertions(+)
> > >
> > > diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
> > > index 5e6b1eb54c64..02821c0f4cd0 100644
> > > --- a/drivers/pci/vgaarb.c
> > > +++ b/drivers/pci/vgaarb.c
> > > @@ -60,6 +60,9 @@ static bool vga_arbiter_used;
> > > static DEFINE_SPINLOCK(vga_lock);
> > > static DECLARE_WAIT_QUEUE_HEAD(vga_wait_queue);
> > > +/* The PCI(e) VGA device which owns the firmware framebuffer */
> > > +static struct pci_dev *pdev_boot_vga;
> > > +
> > > static const char *vga_iostate_to_str(unsigned int iostate)
> > > {
> > > /* Ignore VGA_RSRC_IO and VGA_RSRC_MEM */
> > > @@ -582,6 +585,9 @@ static bool vga_is_firmware_default(struct pci_dev *pdev)
> > > return true;
> > > }
> > > +#else
> > > + if (pdev_boot_vga && pdev_boot_vga == pdev)
> > > + return true;
> > > #endif
> > > return false;
> > > }
> > > @@ -1557,3 +1563,73 @@ static int __init vga_arb_device_init(void)
> > > return rc;
> > > }
> > > subsys_initcall_sync(vga_arb_device_init);
> > > +
> > > +/*
> > > + * Get the physical address range that the firmware framebuffer occupies.
> > > + *
> > > + * Note that the global screen_info is arch-specific, thus CONFIG_SYSFB is
> > > + * chosen as compile-time conditional to suppress linkage problems on non-x86
> > > + * architectures.
> > > + *
> > > + * Returns true on success, otherwise return false.
> > > + */
> > > +static bool vga_arb_get_firmware_fb_range(u64 *start, u64 *end)
> > > +{
> > > + u64 fb_start = 0;
> > > + u64 fb_size = 0;
> > > + u64 fb_end;
> > > +
> > > +#if defined(CONFIG_X86) || defined(CONFIG_IA64) || defined(CONFIG_SYSFB)
> > > + fb_start = screen_info.lfb_base;
> > > + if (screen_info.capabilities & VIDEO_CAPABILITY_64BIT_BASE)
> > > + fb_start |= (u64)screen_info.ext_lfb_base << 32;
> > > +
> > > + fb_size = screen_info.lfb_size;
> > > +#endif
> > > +
> > > + /* No firmware framebuffer support */
> > > + if (!fb_start || !fb_size)
> > > + return false;
> > > +
> > > + fb_end = fb_start + fb_size - 1;
> > > +
> > > + *start = fb_start;
> > > + *end = fb_end;
> > > +
> > > + return true;
> > > +}
> > > +
> > > +/*
> > > + * Identify the PCI VGA device that contains the firmware framebuffer
> > > + */
> > > +static void pci_boot_vga_capturer(struct pci_dev *pdev)
> > > +{
> > > + u64 fb_start, fb_end;
> > > + struct resource *res;
> > > + unsigned int i;
> > > +
> > > + if (pdev_boot_vga)
> > > + return;
> > > +
> > > + if (!vga_arb_get_firmware_fb_range(&fb_start, &fb_end))
> > > + return;
> > > +
> > > + pci_dev_for_each_resource(pdev, res, i) {
> > > + if (resource_type(res) != IORESOURCE_MEM)
> > > + continue;
> > > +
> > > + if (!res->start || !res->end)
> > > + continue;
> > > +
> > > + if (res->start <= fb_start && fb_end <= res->end) {
> > > + pdev_boot_vga = pdev;
> > > +
> > > + vgaarb_info(&pdev->dev,
> > > + "BAR %u: %pR contains firmware FB [0x%llx-0x%llx]\n",
> > > + i, res, fb_start, fb_end);
> > > + break;
> > > + }
> > > + }
> > > +}
> > > +DECLARE_PCI_FIXUP_CLASS_HEADER(PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_DISPLAY_VGA,
> > > + 8, pci_boot_vga_capturer);
> >
> > PCI/VGA: Match firmware framebuffer before BAR reassignment
> >
> > vga_is_firmware_default() decides a device is the firmware default VGA
> > device if it has a BAR that contains the framebuffer described by
> > screen_info.
> >
> > Previously this was unreliable because the screen_info framebuffer address
> > comes from firmware, and the Linux PCI core may reassign device BARs before
> > vga_is_firmware_default() runs. This reassignment means the BAR may not
> > match the screen_info values, but we still want to select the device as the
> > firmware default.
> >
> > Make vga_is_firmware_default() more reliable by running it as a quirk so it
> > happens before any BAR reassignment.
> >
> > diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
> > index 5e6b1eb54c64..4a53e76caddd 100644
> > --- a/drivers/pci/vgaarb.c
> > +++ b/drivers/pci/vgaarb.c
> > @@ -60,6 +60,8 @@ static bool vga_arbiter_used;
> > static DEFINE_SPINLOCK(vga_lock);
> > static DECLARE_WAIT_QUEUE_HEAD(vga_wait_queue);
> > +static struct pci_dev *vga_firmware_device;
> > +
> > static const char *vga_iostate_to_str(unsigned int iostate)
> > {
> > /* Ignore VGA_RSRC_IO and VGA_RSRC_MEM */
> > @@ -560,6 +562,7 @@ static bool vga_is_firmware_default(struct pci_dev *pdev)
> > u64 base = screen_info.lfb_base;
> > u64 size = screen_info.lfb_size;
> > struct resource *r;
> > + unsigned int i;
> > u64 limit;
> > /* Select the device owning the boot framebuffer if there is one */
> > @@ -570,7 +573,7 @@ static bool vga_is_firmware_default(struct pci_dev *pdev)
> > limit = base + size;
> > /* Does firmware framebuffer belong to us? */
> > - pci_dev_for_each_resource(pdev, r) {
> > + pci_dev_for_each_resource(pdev, r, i) {
> > if (resource_type(r) != IORESOURCE_MEM)
> > continue;
> > @@ -580,6 +583,8 @@ static bool vga_is_firmware_default(struct pci_dev *pdev)
> > if (base < r->start || limit >= r->end)
> > continue;
> > + vgaarb_info(&pdev->dev, "BAR %u: %pR contains firmware framebuffer [%#010llx-%#010llx]\n",
> > + i, r, base, limit - 1);
> > return true;
> > }
> > #endif
> > @@ -623,7 +628,7 @@ static bool vga_is_boot_device(struct vga_device *vgadev)
> > if (boot_vga && boot_vga->is_firmware_default)
> > return false;
> > - if (vga_is_firmware_default(pdev)) {
> > + if (pdev == vga_firmware_device) {
> > vgadev->is_firmware_default = true;
> > return true;
> > }
> > @@ -1557,3 +1562,14 @@ static int __init vga_arb_device_init(void)
> > return rc;
> > }
> > subsys_initcall_sync(vga_arb_device_init);
> > +
> > +static void vga_match_firmware_framebuffer(struct pci_dev *pdev)
> > +{
> > + if (vga_firmware_device)
> > + return;
> > +
> > + if (vga_is_firmware_default(pdev))
> > + vga_firmware_device = pdev;
> > +}
> > +DECLARE_PCI_FIXUP_CLASS_HEADER(PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_DISPLAY_VGA,
> > + 8, vga_match_firmware_framebuffer);
>
>
> Q: What if we did something like the patch below?
>
> A:
>
> But the vga_is_firmware_default() function only works on X86 and IA64,
> you patch doesn't solve the problems on ARM64 and LoongArch.

Yes, that's true. Ideally a patch solves a single problem. This one
solves an issue for x86 and ia64. A subsequent patch can solve the
problems on ARM64 and LoongArch. Doing them separately means that
each patch is easier to understand and if we accidentally break
something, a bisection can give more specific information about what
broke.

Bjorn