Public bug reported:

[Impact]
On Intel Nova Lake systems with RAID ON and 'VMD Global Mapping' enabled in 
BIOS, the Ubuntu 26.04 OEM installer sees no disks. The install fails, and only 
AHCI mode installs. The installer log shows:

Failed to find matching device for [{'path': '/dev/dm*'}, {'path': '/dev/md*'}, 
{'path': '/dev/nvme0n1'}, {'path': '/dev/pmem*'}, {'path': '/dev/mmcblk0'}, 
{'path': '/dev/sda', 'size': 'largest'}, {'path': '/dev/sdb'}, {'path': 
'/dev/vda'}]
AutoinstallError

Starting with Intel Arrow Lake client VMD, the VMD controller reports
Bus Restriction type 3 and exposes a second root bus (BUS1). The vmd
driver treats type 3 like type 2 and only creates BUS0. The NVMe drives
behind BUS1 are never enumerated, so the installer finds no block
devices.

The Nova Lake VMD device ID is already in this tree (commit
55aa45154fe4). The ID alone does not fix the problem.

[Fix]
Add support for the second VMD root bus (BUS1). When the firmware reports 
restriction type 3 on a client VMD, the driver creates and scans BUS1 in its 
own emulated PCI domain ('domain1'). BUS1 gets its own bus numbers and 
resources, separate from BUS0. If BUS1 cannot be created, the driver warns and 
keeps BUS0 only.

The fix is 7 upstream commits, applied in this order with cherry-pick -x:
9ec146e311d1 PCI: vmd: Add vmd_bus_enumeration() helper function
2f3f85305a90 PCI: vmd: Add vmd_configure_cfgbar() helper function
ed34664daffa PCI: vmd: Add vmd_configure_membar() and 
vmd_configure_membar1_membar2()
b294e381f1de PCI: vmd: Add vmd_create_bus()
7aa850769255 PCI: vmd: Replace hardcoded values with enum and defines
99b617df5c36 PCI: vmd: Convert bus, busn_start and sysdata to an array
e71e4d9a2a36 PCI: vmd: Add support for second rootbus under VMD

The first six commits are helper refactors with no functional change.
The last commit adds the feature.

Upstream status: the series is in the pci.git pci/controller/vmd branch
(merged via 6633c535abc3) and in linux-next. It is not in a released
kernel tag yet.

Downstream changes: three conflicts in drivers/pci/controller/vmd.c. The
local pci_lock_rescan_remove()/pci_unlock_rescan_remove() calls (from
the SAUCE patch 'PCI: vmd: guard device addition and removal') are kept
around the bus creation and enumeration. The VMD_FEAT_INTERRUPT_QUIRK
code is not in this tree and is not added. VMD_FEAT_HAS_BUS1_ROOTBUS
keeps the upstream value (1 << 8).

[Test Plan]
Reproducer, from the ticket:
1. Set the BIOS to RAID ON with 'VMD Global Mapping' enabled.
2. Plug in the USB drive with the Ubuntu 26.04 OEM kernel image.
3. Choose 'Restore Ubuntu to factory state' and start the install.
Expected result: the install completes in RAID ON mode.
Result without the fix: the install cannot find the disks and fails. Only AHCI 
mode installs (ticket 'Actual Result').

Reporter results from the ticket (not re-run by the automation):
- 2026-09-16: a VMD test kernel (v5 series) booted with RAID ON and VMD Global 
Mapping. The reporter then installed with VMD disabled and enabled VMD 
afterwards.
- 2026-10-22: the reporter verified the RAID ON install with the v8 patch set, 
and the OS installed in RAID ON mode.
- 2026-10-30: the reporter saw an install failure on another board. On 
2026-11-02 they confirmed it was a display issue, not RAID ON.

Note: the v8 series was verified. The final v10 series (the 7 commits in
this report) has not been confirmed by a test result in the ticket.

The automation did not reproduce or re-test this fix on any machine.
These results come from the reporter's comments only.

To check a build with this fix, repeat the reproducer above. With RAID
ON, the install should find the NVMe disks. After boot, 'lspci -k'
should show the VMD controller bound to the vmd driver, and 'lsblk'
should list the NVMe disks.

[Where problems could occur]
The change is in the VMD driver (drivers/pci/controller/vmd.c) and affects 
every VMD platform, not only Nova Lake. The first six commits refactor the VMD 
bus setup and resource (CFGBAR, MEMBAR1/MEMBAR2) code. If the refactor is 
wrong, NVMe drives behind VMD on RAID ON systems could stop showing up at boot, 
and the installer or system would see no root disk.

On platforms that report restriction type 3, a second PCI domain
('domain1') is now created. A bug here could change PCI domain or bus
numbering for devices on those systems.

The downstream lock/unlock calls are moved into the new enable_domain
flow. Wrong placement could cause a rescan race, showing up as missing
or duplicate PCI devices at boot.

** Affects: hwe-next
     Importance: Undecided
         Status: New

** Affects: linux-oem-7.3 (Ubuntu)
     Importance: Undecided
         Status: New


** Tags: jira-wtn-460 oem-priority resolute wtn

** Tags added: jira-wtn-460 oem-priority wtn

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2174044

Title:
  RAID ON install fails: no disks found behind Intel VMD

To manage notifications about this bug go to:
https://bugs.launchpad.net/hwe-next/+bug/2174044/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to