amdgpu fix for HP ProBook 445 G7 — AMD Renoir black screen

Status: ✅ RESOLVED AND PERMANENT. On normal boot amdgpu loads from initramfs, the internal panel (eDP-1) lights up at 1920×1080 native.

🇧🇷 Resumo em português disponível abaixo.

---

Summary

Item	Value
Machine	HP ProBook 445 G7
GPU	AMD Renoir iGPU (1002:1636, subsystem HP 103C:8730), PCI 0000:05:00.0
OS	Ubuntu 26.04, kernel 7.0.0-27-generic
Symptom	100% black screen with amdgpu; amdgpu_get_bios() / display stage failed
Root cause	The VBIOS image in the HP flash is a generic template with no eDP object; the real VBIOS (with the panel) only exists in the ROM BAR, populated by the SBIOS during POST
Fix	(a) amdgpu recompiled with file-load of VBIOS + HPD underflow fix; (b) vbios.bin = dump of the real ROM BAR; (c) early load via dracut
Result	card1 → amdgpu, eDP-1 = 1920×1080/1680×1050/…; no Failed to create link encoder, no wait timed out

---

The problem

On all tested kernels (6.8, 6.17, 7.0) amdgpu failed to probe with -22 (EINVAL), leaving the display on simple-framebuffer (simpledrm) — or, with the patched module probing deeper, in a black screen with no fallback.

amdgpu_get_bios() tries to obtain the VBIOS in this order: ATRM → VFCT → VRAM BAR → ROM BAR → platform. On this machine, in normal boot state:

• ATRM: no ACPI method → absent.
• VFCT: no VFCT ACPI table in this ACPI.
• VRAM shadow: absent.
• ROM BAR: returned I/O error when the GPU had not been POSTed.
• platform: picked up the copy the SBIOS shadows at 0xC0000 — the generic VBIOS.

Even when a VBIOS was obtained, the display (DC) stage failed:

create_links: BIOS object table - number of connectors: 5
... link_id: 21, is_internal_display: 0, hpd_int_gpio_uid id: 0
*ERROR* Failed to create link encoder!
construct_phy failed.

Root cause of "Failed to create link encoder"

• link_factory.c → construct_phy: without HPD in the VBIOS, executes link->hpd_src = hpd_info.hpd_int_gpio_uid - 1. With uid = 0 (generic VBIOS) → hpd_src = -1 (unsigned underflow → huge value).
• dcn21_link_encoder_create (dcn21_resource.c): if (... || enc_init_data->hpd_source >= ARRAY_SIZE(link_enc_hpd_regs)) return NULL; link_enc_hpd_regs[] has 5 entries; huge hpd_source → NULL → "Failed to create link encoder" for all links → no CRTC → black screen.

The decisive discovery: two different VBIOSes

Comparing the VBIOS extracted from the HP firmware with the dump of the ROM BAR (read with the GPU already POSTed for simpledrm), the images differ in ~351 bytes, all in the 0xCC00–0xD300 region (LCD_Info @0xCCD8 + Display Object Table @0xD248):

Region	HP flash (renoir_vbios_113-RENOIR-037.rom)	Real ROM BAR (vbios_rombar.bin)
LCD_Info @0xCCD8	zeroed (no panel timing)	real timing 1920×1080 (0x780×0x438)
Connector table @0xD250	5× 0x15 (MXM) — generic placeholder	0x14 (eDP, internal panel) + 0x0C (HDMI) + 2× 0x13 (DP), with valid HPD/GPIO
Strings	"Renoir Generic VBIOS"	real OEM mapping

Conclusion: the eDP/panel mapping is not in the flash; it is assembled by the SBIOS during iGPU POST. The correct VBIOS always existed — in the ROM BAR.

---

The fix (3 components)

1. Patched amdgpu module (2 patches)

Source downloaded from apt pool linux_7.0.0-27.27 and extracted with dpkg-source.

Patch 1 — VBIOS file-load (drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c):

• New amdgpu_read_bios_from_file() that calls request_firmware("amdgpu/vbios.bin") and validates with check_atom_bios.
• Called first in amdgpu_get_bios_apu().
• MODULE_FIRMWARE("amdgpu/vbios.bin") → ensures vbios.bin is pulled into initramfs.

Patch 2 — HPD underflow fix (drivers/gpu/drm/amd/display/dc/link/link_factory.c):

/* before */ link->hpd_src = hpd_info.hpd_int_gpio_uid - 1;
/* after  */ link->hpd_src = hpd_info.hpd_int_gpio_uid ? hpd_info.hpd_int_gpio_uid - 1 : 0;

With the real VBIOS this fix is a no-op (VBIOS already provides valid HPD), but it protects against regression.

2. vbios.bin = real ROM BAR dump

• Source: /sys/bus/pci/devices/0000:05:00.0/rom (readable with GPU already POSTed).
• 54784 bytes, mod256 checksum = 0 (valid), 113-RENOIR-037, ATOMBIOSBK-AMD VER017.010.000.031.
• sha256: e40fd9a8f8a244d3dbea3c7e9834032740a2c0d71000aa1d81d61989cf95425a
• Install to /lib/firmware/amdgpu/vbios.bin.

3. Early load (dracut)

Ubuntu 26.04 uses dracut, not mkinitramfs.

• /etc/dracut.conf.d/amdgpu.conf → add_drivers+=" amdgpu "
• update-initramfs -u -k 7.0.0-27-generic embeds amdgpu.ko.zst + vbios.bin in the initramfs.

---

Build notes (lessons learned)

Error 1 — modversions CRC mismatch

• Symptom: amdgpu: disagrees about version of symbol module_layout.
• Cause: single-module in-tree build (make drivers/.../amdgpu.ko) without vmlinux.o generates CRCs from source headers, not the binary kernel.
• Fix: EXTMOD build (make -C $SRC M=drivers/gpu/drm/amd/amdgpu modules) with Module.symvers from the binary kernel (/lib/modules/.../build/Module.symvers, module_layout=0xb0c84d61) copied into the source tree.

Error 2 — rejected relocation

• Symptom: x86/modules: Invalid relocation target, existing value is nonzero ... (fatal, -ENOEXEC) → module aborts before probe.
• Cause: kernel has CONFIG_DEBUG_INFO_BTF_MODULES=y, which adds 24 bytes to struct module (fields btf_*) before the exit field. Our .config had the option disabled → cleanup_module landed at the wrong offset (0x490 vs 0x4a8).
• Fix: rebuild with KCFLAGS="-DCONFIG_DEBUG_INFO_BTF_MODULES=1" (no need for pahole/vmlinux). cleanup_module moved to 0x4a8. ✅
• Lesson: when copying Module.symvers from the kernel to match CRCs, you must also match every .config option that changes the layout of an exported struct — otherwise CRC passes but the binary/relocation diverges.

Validated build

• vermagic 7.0.0-27-generic SMP preempt mod_unload modversions ✅
• module_layout = 0xb0c84d61 (== kernel), 0/1130 divergent symbols ✅
• firmware: amdgpu/vbios.bin present ✅
• Build env: gcc 15.2.0 (identical to kernel build), Secure Boot OFF, MODULE_SIG_FORCE off, lockdown [none].
• Required packages: flex bison libelf-dev.

---

Installation — quick start

Precondition: boot with the GPU already POSTed and the amdgpu module not driving the display (the broken state qualifies: simpledrm active, or recovery / nomodeset boot). The ROM BAR is only readable after POST.

sudo apt install dkms patch flex bison libelf-dev linux-headers-$(uname -r)
git clone https://github.com/Jeanmarcus93/hp-probook-445g7-amdgpu-fix.git
cd hp-probook-445g7-amdgpu-fix
sudo bash install.sh
# reboot via the normal GRUB entry

install.sh runs the four steps in order:

1. scripts/dump_vbios_rom.sh — auto-detects the AMD GPU on the PCI bus, enables the ROM BAR, dumps the real VBIOS and validates it (0x55AA signature, mod-256 checksum, ATOMBIOSBK marker, object table v1.4 content).
2. scripts/install_vbios.sh — installs the dump as /lib/firmware/amdgpu/vbios.bin (with backup of any previous file).
3. scripts/build_dkms.sh — downloads the matching linux-source package, extracts only drivers/gpu/drm/amd/ (the subtree is self-contained: FULL_AMD_PATH=$(src)/..), applies the patches from patches/, installs the tree in /usr/src/amdgpu-rombar-<version>/ and runs dkms add/build/install.
4. scripts/make_amdgpu_permanent.sh — dracut early-load conf (add_drivers+=" amdgpu ") + initramfs regeneration + sanity checks.

Each script can also be run individually; all take optional arguments (PCI BDF, dump path, kernel version) and default to auto-detection.

Optional no-reboot test with auto-recovery: sudo bash scripts/test_amdgpu_real.sh loads the module manually and, if the panel doesn't light up in 90 s, reboots back to simpledrm on its own.

---

DKMS

The module is packaged as amdgpu-rombar with AUTOINSTALL="yes": every new kernel installed via apt triggers an automatic rebuild against that kernel's headers — no apt-mark hold needed, no manual rebuild.

The build is EXTMOD against the headers package of the target kernel, so modversions CRCs, autoconf (CONFIG_DEBUG_INFO_BTF_MODULES) and the struct module layout come out right automatically (see Build notes for why that matters).

Secure Boot: Ubuntu's dkms signs modules automatically with the MOK key in /var/lib/shim-signed/mok/ if one is enrolled. With Secure Boot enabled and no enrolled MOK, the module will not load — enroll one first (sudo update-secureboot-policy --new-key + mokutil --import).

---

Patches

The patches/ directory contains unified diffs against the vanilla linux_7.0.0-27.27 source:

• 0001-amdgpu-bios-add-file-load-vbios.patch — VBIOS file-load in amdgpu_bios.c
• 0002-link-factory-fix-hpd-underflow.patch — HPD underflow fix in link_factory.c
• 0003-amdgpu-trace-fix-out-of-tree-include-path.patch — TRACE_INCLUDE_PATH . in amdgpu_trace.h. Only needed for the out-of-tree/DKMS build (the in-tree relative path ../../drivers/gpu/drm/amd/amdgpu doesn't exist when building as an external module). Skip it if you're patching a full kernel tree.

Apply to a full kernel tree with patch -p1, or to the extracted amd/ subtree with patch -p4 (which is what build_dkms.sh does).

The reconstructed tree (vanilla source + these patches) was verified byte-identical to the DKMS tree running on the reference machine.

---

Verification (current confirmed state)

uname -r
# 7.0.0-27-generic

readlink -f /sys/class/drm/card1/device/driver
# .../bus/pci/drivers/amdgpu

dmesg | grep -iE 'Fetched VBIOS|Display Core|eDP-1'
# amdgpu 0000:05:00.0: Fetched VBIOS from file (amdgpu/vbios.bin)
# [drm] Display Core v3.2.369 initialized on DCN 2.1
# [drm] Using ACPI provided EDID for eDP-1

cat /sys/class/drm/card1-eDP-1/modes | head
# 1920x1080
# 1680x1050

sha256sum /lib/firmware/amdgpu/vbios.bin
# e40fd9a8f8a244d3dbea3c7e9834032740a2c0d71000aa1d81d61989cf95425a

The Fetched VBIOS from file timestamp at ~8.5 s confirms the load happens on boot (via initramfs/dracut), not via manual modprobe.

---

Rollback / recovery

If the screen goes black:

1. In GRUB (hold Shift/Esc), choose Advanced options → recovery (nomodeset): amdgpu won't probe and simpledrm restores video.

2. Remove the DKMS module (the stock module in /lib/modules/.../kernel/... takes over again):

   sudo dkms remove amdgpu-rombar/$(uname -r | cut -d- -f1) --all
   sudo rm -f /etc/dracut.conf.d/amdgpu.conf
   sudo depmod -a
   sudo update-initramfs -u -k "$(uname -r)"

3. Reboot.

---

Hardware / IDs

• GPU: AMD Renoir iGPU, PCI 0000:05:00.0 (bus 5, slot 0, func 0)
• 1002:1636, subsystem HP 103C:8730
• HP BIOS S79 (password-locked: F10 setup and flashrom inaccessible → flash immutable)
• Generic VBIOS from HP flash: 113-RENOIR-037, sha256 fb51268390b6ee18a20a2f5d6840f72844ddc88715dd91f6cce9c686182445a7
• Real VBIOS (ROM BAR): same string, sha256 e40fd9a8f8a244d3dbea3c7e9834032740a2c0d71000aa1d81d61989cf95425a

---

Resumo em Português

Correção de tela preta no HP ProBook 445 G7 (AMD Renoir Ryzen 4500U)

Este repositório contém a solução completa para o bug de tela preta que afeta o HP ProBook 445 G7 com processador AMD Ryzen 5 4500U (GPU Renoir integrada) no Ubuntu 26.04 / kernel 7.0.0-27.

Problema: A HP armazenou um VBIOS genérico (template sem painel eDP) na flash do firmware. O VBIOS real — com os timings do painel 1920×1080 e os conectores eDP/HDMI/DisplayPort corretos — só existe na ROM BAR da GPU e é preenchido pelo SBIOS durante o POST. O driver amdgpu do kernel carrega o VBIOS do lugar errado (platform/0xC0000) e recebe o template genérico, causando um underflow no HPD que derruba o display completamente.

Solução (3 partes):

1. Módulo amdgpu patcheado — dois patches no kernel 7.0.0-27:
   • amdgpu_bios.c: nova função que carrega o VBIOS de /lib/firmware/amdgpu/vbios.bin
   • link_factory.c: corrige underflow do hpd_src (proteção contra regressão)
2. vbios.bin real — dump da ROM BAR com a GPU já POSTada (contém eDP 0x14, HDMI 0x0C, 2× DP 0x13, LCD_Info 1920×1080)
3. Carga antecipada via dracut — add_drivers+=" amdgpu " no initramfs

Instalação (comando único):

git clone https://github.com/Jeanmarcus93/hp-probook-445g7-amdgpu-fix.git
cd hp-probook-445g7-amdgpu-fix
sudo bash install.sh
# reinicie pela entrada normal do GRUB

Pré-requisito: boot com a GPU já POSTada (recovery, nomodeset ou simpledrm ativo). A ROM BAR só é legível após o POST da iGPU.

Modelo exato testado: HP ProBook 445 G7, BIOS HP S79 (protegida por senha), AMD Ryzen 5 4500U com Radeon Graphics, PCI 0000:05:00.0, 1002:1636 subsystem 103C:8730.

Rollback: boot via recovery do GRUB (nomodeset), remova o módulo DKMS (dkms remove amdgpu-rombar/7.0.0 --all), apague o conf do dracut e regenere o initramfs.

⚠️ Nota para outros modelos: Este fix é específico para o HP ProBook 445 G7. O script validate_vbios.py --require-edp verifica se o dump contém painel eDP/LVDS. Em outros laptops com o mesmo problema (VBIOS genérico na flash), a abordagem é a mesma mas os patches podem precisar de ajustes para kernels diferentes.