Skip to contentSkip to conversation

Fix loading of platform ROM from 32-bit EFI

PATCH RESEND In review Mikel Rychliski 2 participants 6 messages Mar 13, 2020 – Mar 18, 2020

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

v2 cover Fix loading of platform ROM from 32-bit EFI

This patch series fixes an oops when loading the radeon driver on old Macs with a 32-bit EFI implementation.

Tested on a MacPro 1,1 with a Radeon X1900 XT card and 32-bit kernel.

Changes in v2:
- Add iounmap() call in nouveau
- Update function comment for pci_platform_rom()
- Minor changes to commit messages

Mikel Rychliski (2):
drm/radeon: Stop directly referencing iomem
PCI: Use ioremap(), not phys_to_virt() for platform ROM

 drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c             |  1 +
 drivers/gpu/drm/nouveau/nvkm/subdev/bios/shadowpci.c | 11 ++++++++++-
 drivers/gpu/drm/radeon/radeon_bios.c                 | 12 ++++++++----
 drivers/pci/rom.c                                    |  9 ++++++---
 4 files changed, 25 insertions(+), 8 deletions(-)
Signature · 1 line
2.13.7
#1 lore raw

PATCH RESEND v2 1/2 drm/radeon: Stop directly referencing iomem

pci_platform_rom() returns an __iomem pointer which should not be accessed directly. Change radeon_read_platform_bios() to use memcpy_fromio() instead of calling kmemdup() on the __iomem pointer.

---
 drivers/gpu/drm/radeon/radeon_bios.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/gpu/drm/radeon/radeon_bios.c +7 -4
index c42f73fad3e3..c3ae4c92a115 100644
@@ -118,11 +118,14 @@ static bool radeon_read_platform_bios(struct radeon_device *rdev)
 		return false;
 	}
 
-	if (size == 0 || bios[0] != 0x55 || bios[1] != 0xaa) {
+	rdev->bios = kzalloc(size, GFP_KERNEL);
+	if (!rdev->bios)
 		return false;
-	}
-	rdev->bios = kmemdup(bios, size, GFP_KERNEL);
-	if (rdev->bios == NULL) {
+
+	memcpy_fromio(rdev->bios, bios, size);
+
+	if (size == 0 || rdev->bios[0] != 0x55 || rdev->bios[1] != 0xaa) {
+		kfree(rdev->bios);
 		return false;
 	}
 
Signature · 1 line
2.13.7
#2 lore raw

PATCH RESEND v2 2/2 PCI: Use ioremap(), not phys_to_virt() for platform ROM

In reply to Mikel Rychliski

On some EFI systems, the video BIOS is provided by the EFI firmware. The boot stub code stores the physical address of the ROM image in pdev->rom. Currently we attempt to access this pointer using phys_to_virt(), which doesn't work with CONFIG_HIGHMEM.

On these systems, attempting to load the radeon module on a x86_32 kernel can result in the following:

    BUG: unable to handle page fault for address: 3e8ed03c
    #PF: supervisor read access in kernel mode
    #PF: error_code(0x0000) - not-present page
    *pde = 00000000
    Oops: 0000 [#1] PREEMPT SMP
    CPU: 0 PID: 317 Comm: systemd-udevd Not tainted 5.6.0-rc3-next-20200228 #2
    Hardware name: Apple Computer, Inc. MacPro1,1/Mac-F4208DC8, BIOS     MP11.88Z.005C.B08.0707021221 07/02/07
    EIP: radeon_get_bios+0x5ed/0xe50 [radeon]
    Code: 00 00 84 c0 0f 85 12 fd ff ff c7 87 64 01 00 00 00 00 00 00 8b 47 08 8b 55 b0 e8 1e 83 e1 d6 85 c0 74 1a 8b 55 c0 85 d2 74 13 <80> 38 55 75 0e 80 78 01 aa 0f 84 a4 03 00 00 8d 74 26 00 68 dc 06
    EAX: 3e8ed03c EBX: 00000000 ECX: 3e8ed03c EDX: 00010000
    ESI: 00040000 EDI: eec04000 EBP: eef3fc60 ESP: eef3fbe0
    DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068 EFLAGS: 00010206
    CR0: 80050033 CR2: 3e8ed03c CR3: 2ec77000 CR4: 000006d0
    Call Trace:
     ? register_client+0x34/0xe0
     ? register_client+0xab/0xe0
     r520_init+0x26/0x240 [radeon]
     radeon_device_init+0x533/0xa50 [radeon]
     radeon_driver_load_kms+0x80/0x220 [radeon]
     drm_dev_register+0xa7/0x180 [drm]
     radeon_pci_probe+0x10f/0x1a0 [radeon]
     pci_device_probe+0xd4/0x140
     really_probe+0x13d/0x3b0
     driver_probe_device+0x56/0xd0
     device_driver_attach+0x49/0x50
     __driver_attach+0x79/0x130
     ? device_driver_attach+0x50/0x50
     bus_for_each_dev+0x5b/0xa0
     driver_attach+0x19/0x20
     ? device_driver_attach+0x50/0x50
     bus_add_driver+0x117/0x1d0
     ? pci_bus_num_vf+0x20/0x20
     driver_register+0x66/0xb0
     ? 0xf80f4000
     __pci_register_driver+0x3d/0x40
     radeon_init+0x82/0x1000 [radeon]
     do_one_initcall+0x42/0x200
     ? kvfree+0x25/0x30
     ? __vunmap+0x206/0x230
     ? kmem_cache_alloc_trace+0x16f/0x220
     ? do_init_module+0x21/0x220
     do_init_module+0x50/0x220
     load_module+0x1f26/0x2200
     sys_init_module+0x12d/0x160
     do_fast_syscall_32+0x82/0x250
     entry_SYSENTER_32+0xa5/0xf8

Fix the issue by using ioremap() instead of phys_to_virt() in pci_platform_rom().

Now that pci_platform_rom() creates a new mapping to access the ROM image, update all callers to remove this mapping after extracting the BIOS.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c             |  1 +
 drivers/gpu/drm/nouveau/nvkm/subdev/bios/shadowpci.c | 11 ++++++++++-
 drivers/gpu/drm/radeon/radeon_bios.c                 |  1 +
 drivers/pci/rom.c                                    |  9 ++++++---
 4 files changed, 18 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c +1 -0
index 50dff69a0f6e..ea6a1fa98ad9 100644
@@ -207,6 +207,7 @@ static bool amdgpu_read_platform_bios(struct amdgpu_device *adev)
 		return false;
 
 	memcpy_fromio(adev->bios, bios, size);
+	iounmap(bios);
 
 	if (!check_atom_bios(adev->bios, size)) {
 		kfree(adev->bios);
drivers/gpu/drm/nouveau/nvkm/subdev/bios/shadowpci.c +10 -1
index 9b91da09dc5f..8a60a0db7b14 100644
@@ -111,11 +111,20 @@ platform_init(struct nvkm_bios *bios, const char *name)
 	return ERR_PTR(ret);
 }
 
+static void
+platform_fini(void *data)
+{
+	struct priv *priv = data;
+
+	iounmap(priv->rom);
+	kfree(priv);
+}
+
 const struct nvbios_source
 nvbios_platform = {
 	.name = "PLATFORM",
 	.init = platform_init,
-	.fini = (void(*)(void *))kfree,
+	.fini = platform_fini,
 	.read = pcirom_read,
 	.rw = true,
 };
drivers/gpu/drm/radeon/radeon_bios.c +1 -0
index c3ae4c92a115..712b88d88957 100644
@@ -123,6 +123,7 @@ static bool radeon_read_platform_bios(struct radeon_device *rdev)
 		return false;
 
 	memcpy_fromio(rdev->bios, bios, size);
+	iounmap(bios);
 
 	if (size == 0 || rdev->bios[0] != 0x55 || rdev->bios[1] != 0xaa) {
 		kfree(rdev->bios);
drivers/pci/rom.c +6 -3
index 137bf0cee897..b38257d23e6e 100644
@@ -197,16 +197,19 @@ void pci_unmap_rom(struct pci_dev *pdev, void __iomem *rom)
 EXPORT_SYMBOL(pci_unmap_rom);
 
 /**
- * pci_platform_rom - provides a pointer to any ROM image provided by the
- * platform
+ * pci_platform_rom - ioremap() the ROM image provided by the platform
  * @pdev: pointer to pci device struct
  * @size: pointer to receive size of pci window over ROM
+ *
+ * Return: kernel virtual pointer to image of ROM
+ *
+ * The caller is responsible for removing the mapping with iounmap()
  */
 void __iomem *pci_platform_rom(struct pci_dev *pdev, size_t *size)
 {
 	if (pdev->rom && pdev->romlen) {
 		*size = pdev->romlen;
-		return phys_to_virt((phys_addr_t)pdev->rom);
+		return ioremap(pdev->rom, pdev->romlen);
 	}
 
 	return NULL;
Signature · 1 line
2.13.7
#3 lore raw

Re: [PATCH RESEND v2 2/2] PCI: Use ioremap(), not phys_to_virt() for platform ROM

Quoted 11 lines · Mikel Rychliski

/**
+ * pci_platform_rom - ioremap() the ROM image provided by the platform
* @pdev: pointer to pci device struct
* @size: pointer to receive size of pci window over ROM
+ *
+ * Return: kernel virtual pointer to image of ROM
+ *
+ * The caller is responsible for removing the mapping with iounmap() */
void __iomem *pci_platform_rom(struct pci_dev *pdev, size_t *size) {

drivers/pci/rom.c void pci_unmap_rom(struct pci_dev *pdev, void __iomem *rom) · line 213

Quoted 5 lines
drivers/pci/rom.c +1 -1
 	if (pdev->rom && pdev->romlen) {
 		*size = pdev->romlen;
-		return phys_to_virt((phys_addr_t)pdev->rom);
+		return ioremap(pdev->rom, pdev->romlen);
 	}

What is the value of this helper over just open coding an ioremap of pdev->rom in the callers?

#4 lore raw

Re: [PATCH RESEND v2 2/2] PCI: Use ioremap(), not phys_to_virt() for platform ROM

Hi Christoph,

Thanks for your comments. I'm also replying here to your comments on the previous series.

Quoted 2 lines · EDT Christoph Hellwig

Any reason drivers can't just use pci_map_rom insteadἅ which already does the right thing?

Some machines don't expose the video BIOS in the PCI BAR and instead only make it available via EFI boot services calls. So drivers need to be able to use the ROM provided by EFI calls, but only if they can't find a valid one anywhere else.

At one point, the EFI provided ROM in pdev->rom *was* exposed via pci_map_rom(). However it had to be split out into a separate function so that drivers could have more control over which sources were preferred.

Quoted 3 lines · EDT Christoph Hellwig

This and the next patch really need to be folded into the previous one to avoid regressions (assuming my other suggestion doesn't work for some reason).

Addressed in v2

Quoted 2 lines · EDT Christoph Hellwig

What is the value of this helper over just open coding an ioremap of pdev->rom in the callers?

I think the direct access to pdev->rom you suggest makes sense, especially because users of the pci_platform_rom() are exposed to the fact that it just calls ioremap().

I decided against wrapping the iounmap() with a pci_unmap_platform_rom(), but I didn't apply the same consideration to the existing function.

How about something like this (with pci_platform_rom() removed)?

static bool radeon_read_platform_bios(struct radeon_device *rdev)

{
	phys_addr_t rom = rdev->pdev->rom;
	size_t romlen = rdev->pdev->romlen;
	void __iomem *bios;

	rdev->bios = NULL;

	if (!rom || romlen == 0)
		return false;

	rdev->bios = kzalloc(romlen, GFP_KERNEL);
	if (!rdev->bios)
		return false;

	bios = ioremap(rom, romlen);
	if (!bios)
		goto free_bios;

	memcpy_fromio(rdev->bios, bios, romlen);
	iounmap(bios);

	if (rdev->bios[0] != 0x55 || rdev->bios[1] != 0xaa)
		goto free_bios;

	return true;
free_bios:
	kfree(rdev->bios);
	return false;
}

If this is preferred, I'll send an updated series. I'm assuming that the removal of pci_platform_rom() and updating of all the callers should be combined into this patch.

Thanks,
Mikel

#5 lore raw

Re: [PATCH RESEND v2 2/2] PCI: Use ioremap(), not phys_to_virt() for platform ROM

Quoted 8 lines · Mikel Rychliski

I think the direct access to pdev->rom you suggest makes sense, especially because users of the pci_platform_rom() are exposed to the fact that it just calls ioremap().

I decided against wrapping the iounmap() with a pci_unmap_platform_rom(), but I didn't apply the same consideration to the existing function.

How about something like this (with pci_platform_rom() removed)?

That looks sensible to me.