Hi Mario,
pci=realloc,assign-busses was the fix for my problem.
Without it the dGPU is unusable.
+ Oz Tiram
On 10/14/26 17:19, Deucher, Alexander wrote:
Public
+ Mario
-----Original Message-----
From: Reza Nikzad Chaleshori <rezanikzadch@gmail.com>
Sent: Wednesday, October 14, 2026 4:05 AM
To: amd-gfx@lists.freedesktop.org
Cc: Reza Nikzad Chaleshori <rezanikzadch@gmail.com>; Deucher, Alexander
<Alexander.Deucher@amd.com>; Koenig, Christian
<Christian.Koenig@amd.com>; David Airlie <airlied@gmail.com>; Simona
Vetter <simona@ffwll.ch>; Zhang, Hawking <Hawking.Zhang@amd.com>;
Min, Frank <Frank.Min@amd.com>; Luo, Zhigang <Zhigang.Luo@amd.com>;
Wang, YuBiao (Ethan) <YuBiao.Wang@amd.com>
Subject: [PATCH] drm/amdgpu: prefer exact bus match when selecting VFCT
VBIOS image
amdgpu_acpi_vfct_match() accepts an entry whose bus number differs from
the runtime bus number as long as vendor, device, slot and function match.
The caller takes the first entry that matches and stops.
On systems with several GPUs of the same model (e.g. mining boards) every
card matches every entry on vendor/device/function, so an earlier entry is
selected for a later card even when the table contains an entry with the exact
bus number. The later card then runs with another board's VBIOS (different
power/clock/memory tables) and can fail to initialize, for example with "UVD
not responding" and a fatal error during GPU init.
Search the table first for an exact bus match, and only if none exists fall back to
matching by device identity.
Reproduced with two RX 580 (Polaris10, 1002:67df) from different vendors
on an ASUS B250 mining board: VFCT holds entries for bus 1 and bus 12, the
card at bus 12 was given the bus 1 image.
Signed-off-by: Reza Nikzad Chaleshori <rezanikzadch@gmail.com>
I think I'd prefer to just revert the change the loosened the mapping. The loosening was done to fix when someone sets pci=realloc on the kernel command line. I'd consider that to be more of a corner case and presumably even that would potentially break on a system like yours anyway.
Alex
Oz,
Just to double check, it was pci=realloc that caused the issue right? I seem to think so.
Another way we could do this is to only do the soft match when pci=realloc was set on kernel command line. I think that would help both cases.
---
drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c | 71 +++++++++++++++-----
----
1 file changed, 46 insertions(+), 25 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c
index f9315e047..d42485e6b 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c
@@ -441,10 +441,15 @@ static bool amdgpu_read_disabled_bios(struct
amdgpu_device *adev)
* the bus numbers disagree we accept the entry if the device identity
* matches.
*
+ * @exact: if true, only an exact bus number match is accepted. The
+ caller
+ * first searches the whole table with @exact set, and only retries
+ with it
+ * cleared when no entry matched exactly, so that an earlier entry for
+ an
+ * identical card on a different bus cannot shadow the correct one.
+ *
* Returns: 0 on match, -ENODEV on no match
*/
static int amdgpu_acpi_vfct_match(struct amdgpu_device *adev,
- VFCT_IMAGE_HEADER *vhdr)
+ VFCT_IMAGE_HEADER *vhdr, bool exact)
{
/* Vendor and device IDs must always match */
if (vhdr->VendorID != adev->pdev->vendor || @@ -459,6 +464,9 @@
static int amdgpu_acpi_vfct_match(struct amdgpu_device *adev,
if (vhdr->PCIBus == adev->pdev->bus->number)
return 0;
+ if (exact)
+ return -ENODEV;
+
/* Bus mismatch but device identity matches (PCI renumbering case)
*/
dev_notice(adev->dev,
"VFCT bus number mismatch: table %u != runtime %u,
matching by device identity (vendor 0x%04x device 0x%04x)\n", @@ -474,6
+482,7 @@ static bool amdgpu_acpi_vfct_bios(struct amdgpu_device *adev)
UEFI_ACPI_VFCT *vfct;
unsigned int offset;
bool r = false;
+ bool exact;
if (!ACPI_SUCCESS(acpi_get_table("VFCT", 1, &hdr)))
return false;
@@ -484,38 +493,50 @@ static bool amdgpu_acpi_vfct_bios(struct
amdgpu_device *adev)
}
vfct = (UEFI_ACPI_VFCT *)hdr;
- offset = vfct->VBIOSImageOffset;
- while (offset < tbl_size) {
- GOP_VBIOS_CONTENT *vbios = (GOP_VBIOS_CONTENT
*)((char *)hdr + offset);
- VFCT_IMAGE_HEADER *vhdr = &vbios->VbiosHeader;
+ /*
+ * Pass 1 looks for an entry whose bus number matches exactly. Only
if
+ * there is none does pass 2 accept an entry that matches by device
+ * identity alone. Without this, on a system with several identical
+ * GPUs an earlier entry would be picked for a later card.
+ */
+ for (exact = true; ; exact = false) {
+ offset = vfct->VBIOSImageOffset;
- offset += sizeof(VFCT_IMAGE_HEADER);
- if (offset > tbl_size) {
- dev_info(adev->dev, "ACPI VFCT image header
truncated,skipping\n");
- goto out;
- }
+ while (offset < tbl_size) {
+ GOP_VBIOS_CONTENT *vbios =
(GOP_VBIOS_CONTENT *)((char *)hdr + offset);
+ VFCT_IMAGE_HEADER *vhdr = &vbios->VbiosHeader;
- offset += vhdr->ImageLength;
- if (offset > tbl_size) {
- dev_info(adev->dev, "ACPI VFCT image
truncated,skipping\n");
- goto out;
- }
+ offset += sizeof(VFCT_IMAGE_HEADER);
+ if (offset > tbl_size) {
+ dev_info(adev->dev, "ACPI VFCT image
header truncated,skipping\n");
+ goto out;
+ }
- if (vhdr->ImageLength &&
- !amdgpu_acpi_vfct_match(adev, vhdr)) {
- adev->bios = kmemdup(&vbios->VbiosContent,
- vhdr->ImageLength,
- GFP_KERNEL);
+ offset += vhdr->ImageLength;
+ if (offset > tbl_size) {
+ dev_info(adev->dev, "ACPI VFCT image
truncated,skipping\n");
+ goto out;
+ }
- if (!check_atom_bios(adev, vhdr->ImageLength)) {
- amdgpu_bios_release(adev);
+ if (vhdr->ImageLength &&
+ !amdgpu_acpi_vfct_match(adev, vhdr, exact)) {
+ adev->bios = kmemdup(&vbios-
VbiosContent,+ vhdr->ImageLength,
+ GFP_KERNEL);
+
+ if (!check_atom_bios(adev, vhdr-
ImageLength)) {+ amdgpu_bios_release(adev);
+ goto out;
+ }
+ adev->bios_size = vhdr->ImageLength;
+ r = true;
goto out;
}
- adev->bios_size = vhdr->ImageLength;
- r = true;
- goto out;
}
+
+ if (!exact)
+ break;
}
dev_info(adev->dev, "ACPI VFCT table present but broken (too short
#2),skipping\n");
--
2.53.0