On Tue, Jul 7, 2026 at 9:51 PM Chen Pei <[email protected]> wrote: > > Hi Alistair, > > On Mon, 6 Jul 2026 14:08:45 +1000, Alistair Francis wrote: > > What happens to kernels that don't have this change? > > Good question. The behavior depends on how old the kernel is: > > - Kernels that support _DEP processing but lack the > acpi_dev_clear_dependencies() call in acpi_pci_root_add(): the > ACPI0017 (CXLM) device will remain deferred indefinitely because > its declared dependency on ACPI0016 is never cleared. CXL memory > would be unusable on such kernels with this QEMU. > > - Very old kernels that ignore _DEP entirely: the behavior is the > same as v1 — the probe-ordering race between acpi_pci_root and > cxl_acpi remains, but nothing is worse than before. > > The kernel-side fix has already merged: > > https://lore.kernel.org/linux-pci/[email protected]/ > > So this is effectively a "new QEMU requires a sufficiently new kernel > for CXL to work" situation, similar to other QEMU features that depend > on kernel support.
That's fine, we can say you need a new kernel for this new feature to work, as long as it doesn't break existing PCIe for old kernels Alistair > > Thanks, > Pei
