[PATCH] PCI/VGA: Fix boot device detection after framebuffer relocation

[PATCH] PCI/VGA: Fix boot device detection after framebuffer relocation

[wuqianhai]

From: wuqianhai <wuqianhai@loongson.cn>

PCI resource assignment can relocate the firmware framebuffer BAR, which
leaves the base address in screen_info stale.  The VGA arbiter then calls
screen_info_pci_dev() from vga_is_firmware_default(), which fails to match
the framebuffer against any PCI device's BARs.  As a result, the boot VGA
device is not recognized and the arbiter may select the wrong default
device.

Fix this by calling screen_info_apply_fixups() in vga_arb_device_init()
before adding the VGA devices.  This relocates lfb_base to the assigned
BAR address so that vga_is_firmware_default() matches the boot device
again.

Tested on a LoongArch platform, with vga_is_firmware_default()
additionally enabled for CONFIG_LOONGARCH via local changes.

Fixes: 337bf13aa9ddab8ae47f2677ab75cf69ae33437b ("PCI/VGA: Replace vga_is_firmware_default() with a screen info check")
Signed-off-by: wuqianhai <wuqianhai@loongson.cn>
---
 drivers/pci/vgaarb.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
index 3e58b2c05c85..aca594651f58 100644
--- a/drivers/pci/vgaarb.c
+++ b/drivers/pci/vgaarb.c
@@ -1527,6 +1527,14 @@ static int __init vga_arb_device_init(void)
 	int rc;
 	struct pci_dev *pdev;
 
+	/*
+	 * PCI resource assignment may have relocated the firmware
+	 * framebuffer BAR. Update screen_info's lfb_base so that
+	 * vga_is_firmware_default() can match it against the boot
+	 * device's relocated resource.
+	 */
+	screen_info_apply_fixups();
+
 	rc = misc_register(&vga_arb_device);
 	if (rc < 0)
 		pr_err("error %d registering device\n", rc);
-- 
2.34.1

Re: [PATCH] PCI/VGA: Fix boot device detection after framebuffer relocation

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] PCI/VGA: premature and duplicate screen_info_apply_fixups() in vga_arb_device_init()
- [Medium] PCI/VGA: undefined reference to screen_info_apply_fixups() when !SCREEN_INFO
--

commit da40d2d05299ef123ed4789d2ecb58829165fbe3
Author: wuqianhai <wuqianhai@loongson.cn>

PCI/VGA: Fix boot device detection after framebuffer relocation

This commit calls screen_info_apply_fixups() in vga_arb_device_init() so that
screen_info's lfb_base is updated before VGA devices are registered and
checked by vga_is_firmware_default().

> diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
> index 3e58b2c05c854..aca594651f580 100644
> --- a/drivers/pci/vgaarb.c
> +++ b/drivers/pci/vgaarb.c
> @@ -1527,6 +1527,14 @@ static int __init vga_arb_device_init(void)
>  	int rc;
>  	struct pci_dev *pdev;
>  
> +	/*
> +	 * PCI resource assignment may have relocated the firmware
> +	 * framebuffer BAR. Update screen_info's lfb_base so that
> +	 * vga_is_firmware_default() can match it against the boot
> +	 * device's relocated resource.
> +	 */
> +	screen_info_apply_fixups();

[Severity: Medium]
Will calling screen_info_apply_fixups() unconditionally in
vga_arb_device_init() fail to link when CONFIG_PCI=y, CONFIG_VGA_ARB=y, and
CONFIG_SCREEN_INFO=n?

In drivers/pci/Kconfig, VGA_ARB only selects SCREEN_INFO on x86:

drivers/pci/Kconfig:
	config VGA_ARB
		bool "VGA Arbitration" if EXPERT
		default y
		depends on !S390
		select SCREEN_INFO if X86

which matches the CONFIG_X86 guard in vga_is_firmware_default():

drivers/pci/vgaarb.c:vga_is_firmware_default() {
#if defined CONFIG_X86
	return pdev == screen_info_pci_dev(&sysfb_primary_display.screen);
#else
	return false;
#endif
}

However, include/linux/screen_info.h declares screen_info_apply_fixups() as an
external function whenever CONFIG_PCI is enabled, while
drivers/video/Makefile only builds screen_info_pci.o when CONFIG_SCREEN_INFO
is enabled:

include/linux/screen_info.h:
#if defined(CONFIG_PCI)
void screen_info_apply_fixups(void);
struct pci_dev *screen_info_pci_dev(const struct screen_info *si);
#else
static inline void screen_info_apply_fixups(void)
{ }

drivers/video/Makefile:
obj-$(CONFIG_SCREEN_INFO)         += screen_info.o
...
screen_info-$(CONFIG_PCI)         += screen_info_pci.o

On non-x86 architectures with CONFIG_PCI=y, CONFIG_VGA_ARB=y, and
CONFIG_SCREEN_INFO=n, doesn't this result in an undefined reference to
screen_info_apply_fixups()?

[Severity: High]
Also, is subsys_initcall_sync(vga_arb_device_init) the right place in the
initcall sequence to call screen_info_apply_fixups()?

Upstream vga_is_firmware_default() in drivers/pci/vgaarb.c only calls
screen_info_pci_dev() under CONFIG_X86. On x86, subsys_initcall(pci_subsys_init)
runs pcibios_resource_survey() -> pcibios_allocate_dev_resources(). When the
firmware framebuffer BAR has a resource conflict and awaits reassignment,
pcibios_allocate_dev_resources() saves the firmware address and resets
r->start to 0 while keeping resource_size(r) intact:

arch/x86/pci/i386.c:pcibios_allocate_dev_resources() {
	...
	/* We'll assign a new address later */
	pcibios_save_fw_addr(dev, idx, r->start);
	r->end -= r->start;
	r->start = 0;
	...
}

The new BAR address on x86 is not assigned until
fs_initcall(pcibios_assign_resources) -> pci_assign_unassigned_resources(),
which runs after subsys_initcall_sync(vga_arb_device_init).

When screen_info_apply_fixups() runs at subsys_initcall_sync while a
conflicting BAR has pr->start == 0, pr->start != screen_info_lfb_res_start is
true and __screen_info_relocation_is_valid(si, pr) succeeds because
resource_size(pr) is unchanged:

drivers/video/screen_info_pci.c:screen_info_apply_fixups() {
	if (screen_info_lfb_pdev) {
		struct resource *pr = &screen_info_lfb_pdev->resource[screen_info_lfb_bar];

		if (pr->start != screen_info_lfb_res_start) {
			if (__screen_info_relocation_is_valid(si, pr)) {
				__screen_info_set_lfb_base(si, pr->start + screen_info_lfb_offset);
				pr_info("Relocating firmware framebuffer to offset %pa[d] within %pr\n",
					&screen_info_lfb_offset, pr);
			}
			...
		}
	}
}

Wouldn't this clobber si->lfb_base to 0 + screen_info_lfb_offset with a
[mem 0x00000000-...] message before fs_initcall(pcibios_assign_resources)
assigns the actual BAR address, while still failing to match in
vga_is_firmware_default()?

Similarly, for PCI host bridges probed at device_initcall via
pci_host_probe(), vga_arb_device_init() runs before the PCI bus is scanned
and before VGA devices are added via pci_notify().

Finally, since screen_info_apply_fixups() does not update
screen_info_lfb_res_start or clear screen_info_lfb_pdev after applying the
fixup, won't calling it in both vga_arb_device_init() and
device_initcall(sysfb_init) in drivers/firmware/sysfb.c cause the relocation
and log messages to run twice?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261028094634.766723-1-wuqianhai@loongson.cn?part=1

Re: [PATCH] PCI/VGA: Fix boot device detection after framebuffer relocation

[Bjorn Helgaas]

On Wed, Oct 28, 2026 at 05:46:34PM +0800, wuqianhai@loongson.cn wrote:
> From: wuqianhai <wuqianhai@loongson.cn>
> 
> PCI resource assignment can relocate the firmware framebuffer BAR, which
> leaves the base address in screen_info stale.  The VGA arbiter then calls
> screen_info_pci_dev() from vga_is_firmware_default(), which fails to match
> the framebuffer against any PCI device's BARs.  As a result, the boot VGA
> device is not recognized and the arbiter may select the wrong default
> device.
> 
> Fix this by calling screen_info_apply_fixups() in vga_arb_device_init()
> before adding the VGA devices.  This relocates lfb_base to the assigned
> BAR address so that vga_is_firmware_default() matches the boot device
> again.
> 
> Tested on a LoongArch platform, with vga_is_firmware_default()
> additionally enabled for CONFIG_LOONGARCH via local changes.
> 
> Fixes: 337bf13aa9ddab8ae47f2677ab75cf69ae33437b ("PCI/VGA: Replace vga_is_firmware_default() with a screen info check")
> Signed-off-by: wuqianhai <wuqianhai@loongson.cn>
> ---
>  drivers/pci/vgaarb.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
> index 3e58b2c05c85..aca594651f58 100644
> --- a/drivers/pci/vgaarb.c
> +++ b/drivers/pci/vgaarb.c
> @@ -1527,6 +1527,14 @@ static int __init vga_arb_device_init(void)
>  	int rc;
>  	struct pci_dev *pdev;
>  
> +	/*
> +	 * PCI resource assignment may have relocated the firmware
> +	 * framebuffer BAR. Update screen_info's lfb_base so that
> +	 * vga_is_firmware_default() can match it against the boot
> +	 * device's relocated resource.
> +	 */
> +	screen_info_apply_fixups();

I'm really confused about this.  vga_is_firmware_default() relies on
screen_info_pci_dev(&sysfb_primary_display.screen):

  vga_is_firmware_default
    screen_info_pci_dev(&sysfb_primary_display.screen)
      numres = screen_info_resources(&sysfb_primary_display.screen, res, ...)
      for (i = 0; i < numres; ++i)
        pdev = __screen_info_pci_dev(&res[i])

screen_info_resources() returns resource structs with PCI bus
addresses in them.  These are not valid CPU addresses, so that seems
broken on any system where bus addresses are not the same as CPU
addresses.

screen_info_fixup_lfb() and screen_info_apply_fixups() may fix the
issue you see, but I think the way this happens is questionable.

screen_info_fixup_lfb() captures the framebuffer PCI bus address in
screen_info_lfb_res_start.  Then screen_info_apply_fixups() compares
that PCI bus address with the CPU address in pdev->resource[].  That's
broken on any system where bus addresses are not the same as CPU
addresses.  I think it's also broken because screen_info_lfb_res_start
may not be at the beginning of the BAR.

I think screen_info_fixup_lfb() should only capture the pci_dev, the
BAR number, and the framebuffer offset into the BAR.  When we need the
CPU or bus address, we can always compute it from those values.

>  	rc = misc_register(&vga_arb_device);
>  	if (rc < 0)
>  		pr_err("error %d registering device\n", rc);
> -- 
> 2.34.1
> 

[PATCH v2] PCI/VGA: Enable firmware default device detection on LoongArch

[wuqianhai]

From: wuqianhai <wuqianhai@loongson.cn>

On LoongArch, PCI resource assignment completes during subsys_initcall
(acpi_init), before vga_arb_device_init() runs.  By that time the
firmware framebuffer BAR may already have been relocated, leaving
screen_info's lfb_base stale.  Enable vga_is_firmware_default() for
LoongArch and refresh screen_info in vga_arb_device_init() before the
boot device is matched.

Unlike x86, sysfb_primary_display is only defined when CONFIG_SYSFB is
enabled on LoongArch, so have VGA_ARB select SYSFB there.

Signed-off-by: wuqianhai <wuqianhai@loongson.cn>
---
Changes in v2:
- Add VGA_ARB's select of SYSFB for LoongArch to avoid an undefined
  reference when CONFIG_SCREEN_INFO is not set.
- Restrict the screen_info_apply_fixups() call to CONFIG_LOONGARCH so
  x86 and other architectures are unaffected.
- Enable vga_is_firmware_default() for LoongArch.
---
 drivers/pci/Kconfig  |  1 +
 drivers/pci/vgaarb.c | 12 +++++++++++-
 2 files changed, 12 insertions(+), 1 deletion(-)

diff --git a/drivers/pci/Kconfig b/drivers/pci/Kconfig
index 0c7408509ba2..55cc60f5d312 100644
--- a/drivers/pci/Kconfig
+++ b/drivers/pci/Kconfig
@@ -256,6 +256,7 @@ config VGA_ARB
 	default y
 	depends on !S390
 	select SCREEN_INFO if X86
+	select SYSFB if LOONGARCH
 	help
 	  Some "legacy" VGA devices implemented on PCI typically have the same
 	  hard-decoded addresses as they did on ISA. When multiple PCI devices
diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
index 3e58b2c05c85..1b7e915629f9 100644
--- a/drivers/pci/vgaarb.c
+++ b/drivers/pci/vgaarb.c
@@ -566,7 +566,7 @@ EXPORT_SYMBOL(vga_put);
 
 static bool vga_is_firmware_default(struct pci_dev *pdev)
 {
-#if defined CONFIG_X86
+#if defined(CONFIG_X86) || defined(CONFIG_LOONGARCH)
 	return pdev == screen_info_pci_dev(&sysfb_primary_display.screen);
 #else
 	return false;
@@ -1527,6 +1527,16 @@ static int __init vga_arb_device_init(void)
 	int rc;
 	struct pci_dev *pdev;
 
+#if defined(CONFIG_LOONGARCH)
+	/*
+	 * On LoongArch, PCI resource assignment completes during
+	 * subsys_initcall (acpi_init), before the VGA arbiter runs.  Refresh
+	 * screen_info so vga_is_firmware_default() can match the relocated
+	 * BAR.
+	 */
+	screen_info_apply_fixups();
+#endif
+
 	rc = misc_register(&vga_arb_device);
 	if (rc < 0)
 		pr_err("error %d registering device\n", rc);
-- 
2.34.1

Re: [PATCH] PCI/VGA: Fix boot device detection after framebuffer relocation

[wuqianhai]

Hi Bjorn,

Thanks for the detailed analysis.

You're right that screen_info_fixup_lfb()/screen_info_apply_fixups()
stores a PCI bus address in screen_info_lfb_res_start and compares it
with the CPU address in pdev->resource[].start, which is broken on
systems where bus and CPU addresses differ.

On LoongArch, though, bus and CPU addresses are identical: with modern
firmware, acpi_prepare_root_resources() takes the PCIH path and leaves
the window offset at 0 (see commit 49f33840dcc9).  So the existing
screen_info_apply_fixups() works correctly for LoongArch, and the
address-space confusion you describe does not trigger there.

I've sent a v2 that:
- enables vga_is_firmware_default() for LoongArch,
- restricts the screen_info_apply_fixups() call to LoongArch so x86 and
  other architectures are unaffected,
- makes VGA_ARB select SYSFB on LoongArch to avoid the undefined
  reference when CONFIG_SCREEN_INFO is not set.

The general bus != CPU problem in screen_info_pci.c is still worth
fixing for ARM and friends, but I'd rather keep that separate from this
LoongArch enablement.  Happy to help with that follow-up if useful.

Regards,
wuqianhai

Re: [PATCH v2] PCI/VGA: Enable firmware default device detection on LoongArch

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:
- [High] PCI/VGA: pci_fixup_vgadev() overrides vga_is_firmware_default() on LoongArch
- [Medium] PCI/VGA: duplicate non-idempotent screen_info_apply_fixups() call on LoongArch
- [Medium] PCI/VGA: vga_is_firmware_default() fails to match Loongson integrated DC
--

commit 6de860408bce3c5a20f5a4c23b073eafdc942117
Author: wuqianhai <wuqianhai@loongson.cn>

PCI/VGA: Enable firmware default device detection on LoongArch

This commit enables vga_is_firmware_default() on LoongArch, selects
CONFIG_SYSFB from CONFIG_VGA_ARB for LoongArch, and calls
screen_info_apply_fixups() in vga_arb_device_init() so relocated firmware
framebuffer BARs can be matched when selecting the default boot VGA device.

> diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
> index 3e58b2c05c854..1b7e915629f97 100644
> --- a/drivers/pci/vgaarb.c
> +++ b/drivers/pci/vgaarb.c
> @@ -566,7 +566,7 @@ EXPORT_SYMBOL(vga_put);
>  
>  static bool vga_is_firmware_default(struct pci_dev *pdev)
>  {
> -#if defined CONFIG_X86
> +#if defined(CONFIG_X86) || defined(CONFIG_LOONGARCH)
>  	return pdev == screen_info_pci_dev(&sysfb_primary_display.screen);

[Severity: High]
On LoongArch systems with a Loongson display controller (PCI device IDs
0x7a06, 0x7a36, or 0x7a46) and one or more discrete VGA devices, will
pci_fixup_vgadev() in arch/loongarch/pci/pci.c later overwrite the firmware
default VGA device selected here?

During ACPI boot on LoongArch, PCI enumeration runs in
subsys_initcall(acpi_init), and vga_arb_device_init() runs next at
subsys_initcall_sync:

vga_arb_device_init()
  vga_arbiter_add_pci_device()
    vga_is_boot_device()
      vga_is_firmware_default()
    vga_set_default_device(pdev)

After vga_arb_device_init() finishes, pci_apply_final_quirks() in
drivers/pci/quirks.c runs at fs_initcall_sync and invokes the
DECLARE_PCI_FIXUP_FINAL hooks in arch/loongarch/pci/pci.c:

static void pci_fixup_vgadev(struct pci_dev *pdev)
{
	struct pci_dev *devp = NULL;

	while ((devp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, devp))) {
		if (devp->vendor != PCI_VENDOR_ID_LOONGSON) {
			vga_set_default_device(devp);
			dev_info(&pdev->dev,
				"Overriding boot device as %X:%X\n",
				devp->vendor, devp->device);
		}
	}
}
DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_LOONGSON, PCI_DEVICE_ID_LOONGSON_DC1, pci_fixup_vgadev);
DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_LOONGSON, PCI_DEVICE_ID_LOONGSON_DC2, pci_fixup_vgadev);
DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_LOONGSON, PCI_DEVICE_ID_LOONGSON_DC3, pci_fixup_vgadev);

Because pci_fixup_vgadev() unconditionally calls vga_set_default_device(devp)
for every non-Loongson VGA device without checking whether a firmware default
device was already chosen, won't it clobber vga_default whenever firmware
selected a Loongson display device or the first of multiple non-Loongson
discrete GPUs as the primary display?

[Severity: Medium]
Also in vga_is_firmware_default() in drivers/pci/vgaarb.c, can this pointer
comparison ever match when the firmware framebuffer resides on Loongson
integrated graphics (such as on LS7A1000, LS7A2000, or LS2K2000)?

As shown in lsdc_get_dedicated_vram() in drivers/gpu/drm/loongson/lsdc_drv.c,
the integrated graphics subsystem is split across two PCI functions:

	/*
	 * The GPU has 00:06.0 as its BDF, while the DC has 00:06.1
	 * This is true for the LS7A1000, LS7A2000 and LS2K2000.
	 */
	pdev_gpu = pci_get_domain_bus_and_slot(pci_domain_nr(pdev_dc->bus),
					       pdev_dc->bus->number,
					       PCI_DEVFN(6, 0));
	...
	base = pci_resource_start(pdev_gpu, 2);
	size = pci_resource_len(pdev_gpu, 2);

When the firmware framebuffer is placed in integrated VRAM,
screen_info_pci_dev(&sysfb_primary_display.screen) matches BAR 2 of the
companion GPU device at 00:06.0 (PCI_DEVICE_ID_LOONGSON_GPU1/2/3).

However, vga_arb_device_init() and pci_notify() filter devices using
pci_is_vga(pdev), so only the VGA-class Display Controller function at
00:06.1 (PCI_DEVICE_ID_LOONGSON_DC1/2/3, whose BAR 0 only maps MMIO
registers) is passed to vga_is_firmware_default(pdev). Doesn't comparing
pdev (00:06.1) against screen_info_pci_dev() (00:06.0) always evaluate to
false for the integrated display controller?

>  #else
>  	return false;
> @@ -1527,6 +1527,16 @@ static int __init vga_arb_device_init(void)
>  	int rc;
>  	struct pci_dev *pdev;
>  
> +#if defined(CONFIG_LOONGARCH)
> +	/*
> +	 * On LoongArch, PCI resource assignment completes during
> +	 * subsys_initcall (acpi_init), before the VGA arbiter runs.  Refresh
> +	 * screen_info so vga_is_firmware_default() can match the relocated
> +	 * BAR.
> +	 */
> +	screen_info_apply_fixups();
> +#endif

[Severity: Medium]
With CONFIG_VGA_ARB selecting CONFIG_SYSFB on LoongArch, calling
screen_info_apply_fixups() here in vga_arb_device_init() at
subsys_initcall_sync means screen_info_apply_fixups() will be called a second
time when sysfb_init() in drivers/firmware/sysfb.c runs at device_initcall:

drivers/firmware/sysfb.c:sysfb_init() {
	...
	screen_info_apply_fixups();

	mutex_lock(&disable_lock);
	if (disabled)
		goto unlock_mutex;
	...
}

Looking at screen_info_apply_fixups() in drivers/video/screen_info_pci.c:

void screen_info_apply_fixups(void)
{
	struct screen_info *si = &sysfb_primary_display.screen;

	if (screen_info_lfb_pdev) {
		struct resource *pr = &screen_info_lfb_pdev->resource[screen_info_lfb_bar];

		if (pr->start != screen_info_lfb_res_start) {
			if (__screen_info_relocation_is_valid(si, pr)) {
				__screen_info_set_lfb_base(si, pr->start + screen_info_lfb_offset);
				pr_info("Relocating firmware framebuffer to offset %pa[d] within %pr\n",
					&screen_info_lfb_offset, pr);
			} else {
				pr_warn("Invalid relocating, disabling firmware framebuffer\n");
			}
		}
	}
}

Because screen_info_apply_fixups() neither updates screen_info_lfb_res_start
nor clears screen_info_lfb_pdev after applying the fixup, won't the second
call in sysfb_init() re-run the relocation logic and print a duplicate
message whenever the framebuffer BAR was relocated, or emit a spurious
warning if a built-in GPU driver probed earlier in device_initcall and
resized the BAR after calling sysfb_disable()?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261029033440.973077-1-wuqianhai@loongson.cn?part=1