On 09/04/26 at 06:08pm, George Guo wrote: > This series enables Kexec HandOver (KHO) on 64-bit LoongArch and adds > LoongArch coverage to the KHO QEMU selftest. KHO is the remaining > architecture gate needed by Live Update (LUO) on LoongArch. > > The transport is the central problem. Arm64 and riscv can pass the KHO > state FDT and scratch area through properties under /chosen in a boot > FDT. LoongArch boots through EFI without a boot FDT, so that path is not > available.
I didn't look into your patchset very carefully, but I doubt you are reinventing the wheel. X86 didn't have FDT, seems it introduces FDT for KHO support. LoongArch doesn't have FDT, so you reworks the EFI transport. Is there a reason you don't want to follow the standard KHO way using FDT too? Please correct me if I am wrong. > > EFI KHO channel: > > v5 reworks the EFI transport following Pratyush's and Huacai's reviews > [1, 2]. Instead of rebuilding the EFI configuration table in > LoongArch code and switching st->tables before kexec, the common EFI stub > allocates and installs a persistent table under > LINUX_EFI_KEXEC_HANDOVER_GUID. This follows the model used by > LINUX_EFI_MEMRESERVE_TABLE_GUID. > > The common EFI table parser reserves the stub-allocated table, reads it > during early boot, and calls kho_populate(). An early initcall keeps the > table persistently mapped so it can also be updated on the crash kexec > path. The implementation is architecture-independent; LoongArch only > selects EFI_KHO when KEXEC_HANDOVER is enabled. > > The EFI table is global, but kexec can keep separate normal and crash > images and can replace a loaded candidate. Updating the table while a > candidate image is being loaded could therefore leave it pointing at a > failed candidate while the previously installed image remains the one > that will be executed. v5 records the KHO data in struct kimage and > synchronizes the channel from the image selected for execution. Images > without KHO data, including legacy and crash images, clear the channel so > the next kernel boots cold rather than reviving stale state. > > The GUID name is changed to LINUX_EFI_KEXEC_HANDOVER_GUID as suggested by > Huacai and Pratyush [3]. Existing x86 KHO support continues to use its > setup_data transport. Besides supporting non-UEFI boot, that entry is > consumed by the x86 decompressor before normal EFI table parsing, so KASLR > only uses KHO scratch areas and does not overwrite preserved memory. The > new channel is installed by the generic EFI-stub entry path and relies on > EFI_GENERIC_STUB, which x86 does not select. Moving x86 would therefore > require installing the table from the x86-specific EFI stub and scanning it > in the decompressor, not just selecting EFI_KHO. > > LoongArch enablement and build fix: > > The architecture patch now only enables ARCH_SUPPORTS_KEXEC_HANDOVER and > selects the generic EFI channel. No LoongArch setup reader, kexec writer, > or EFI system-table replacement remains. > > The linux/mm.h build fix is moved from luo_session.c to kho_block.c, where > the virt_to_phys() and phys_to_virt() calls reside after the liveupdate > refactoring. This addresses Pratyush's review [4]. Adding linux/mm.h to > asm/io.h was previously attempted, but creates an include cycle under > CONFIG_KFENCE [5]. The fix precedes LoongArch enablement so every > intermediate commit builds with CONFIG_KFENCE=y. > > Selftest: > > The LoongArch vmtest boots vmlinux.efi through EDK2 because the initial > EFI stub invocation installs the channel being tested. QEMU_EFI can > override the firmware filename for distributions that install it under a > different name or path. > > CONFIG_ACPI_SPCR_TABLE lets the bare earlycon parameter obtain the UART > description from EDK2's SPCR table. CONFIG_PCI_HOST_GENERIC keeps i8042 > enabled while ensuring that the QEMU virt machine's PCI I/O space is > mapped before the legacy fallback probe, addressing Huacai's concern that > the same kernel configuration should remain usable for hosts and guests. > > With EFI runtime reset services available, panic=-1 and QEMU's existing > -no-reboot option terminate the test normally. The generic QEMU runner > therefore remains unchanged and does not mask failures by forcibly > terminating QEMU after a timeout. > > The LoongArch vmtest uses the default KHO scratch-sizing path. This > series is based on the generic fix that calculates per-node scratch sizes > before scratch allocation [6]. > > The KHO vmtest passes on: > > * LoongArch with KVM using QEMU_EFI.fd from edk2-loongarch64. > * x86_64 with KVM, as regression coverage for the generic kexec/KHO > changes. > > Patch layout: > > 1. Add the generic EFI KHO channel and its ABI. > 2. Synchronize the channel from the selected image at execution. > 3. Fix the kho_block.c header dependency exposed by LoongArch/KFENCE. > 4. Enable KHO on 64-bit LoongArch through the generic EFI channel. > 5. Add LoongArch to the KHO QEMU selftest. > > Changes since v4: > > * Allocate and install the KHO table from the common EFI stub, following > the EFI memreserve-table model (Pratyush). > * Parse, reserve, map, and update the channel in generic EFI code; remove > the LoongArch-specific reader, writer, and st->tables replacement > (Pratyush, Huacai). > * Synchronize the global channel at execution rather than image load; > clear it for cold and crash images and use the scratch payload size. > * Rename the GUID to LINUX_EFI_KEXEC_HANDOVER_GUID (Huacai, Pratyush). > * Move the linux/mm.h include from luo_session.c to the actual caller, > kho_block.c, and place the fix before architecture enablement > (Pratyush, Huacai). > * Boot the vmtest through EDK2, allow QEMU_EFI to override the firmware, > and use vmlinux.efi to exercise the EFI stub and channel. > * Keep i8042 enabled and select PCI_HOST_GENERIC instead of disabling > the host configuration in the selftest (Huacai). > * Enable ACPI_SPCR_TABLE and use the SPCR-provided early console. > * Exercise the default KHO scratch-sizing path in the LoongArch vmtest; > depend on the separately posted generic sizing fix [6]. > * Drop the generic QEMU timeout now that EFI runtime reset services make > the LoongArch guest terminate normally. > * Rebase onto Linux 7.3-rc1. > > [1] Pratyush Yadav: > https://lore.kernel.org/r/[email protected]/ > [2] Huacai Chen: > > https://lore.kernel.org/r/CAAhV-H4PKRNCa4g=6tgepp+trtm82jimowgvr3o7pqbzyp9...@mail.gmail.com/ > [3] https://lore.kernel.org/r/[email protected]/ > [4] https://lore.kernel.org/r/[email protected]/ > [5] https://lore.kernel.org/r/[email protected]/ > [6] > https://lore.kernel.org/loongarch/[email protected]/ > > Previous version: > > v4: > https://lore.kernel.org/loongarch/[email protected]/ > > George Guo (5): > efi: add a KHO configuration table channel > liveupdate: synchronize EFI KHO channel at execution > liveupdate: kho_block: include linux/mm.h for virt/phys translation > LoongArch: enable kexec handover (KHO) > selftests/kho: add LoongArch vmtest support > > arch/loongarch/Kconfig | 4 ++ > drivers/firmware/efi/Kconfig | 12 ++++ > drivers/firmware/efi/efi.c | 78 ++++++++++++++++++++++ > drivers/firmware/efi/libstub/efi-stub.c | 25 +++++++ > include/linux/efi.h | 36 ++++++++++ > kernel/crash_core.c | 7 ++ > kernel/kexec_core.c | 5 ++ > kernel/kexec_internal.h | 3 + > kernel/liveupdate/kexec_handover.c | 33 +++++++++ > kernel/liveupdate/kho_block.c | 1 + > tools/testing/selftests/kho/loongarch.conf | 11 +++ > tools/testing/selftests/kho/vmtest.sh | 3 +- > 12 files changed, 217 insertions(+), 1 deletion(-) > create mode 100644 tools/testing/selftests/kho/loongarch.conf > > > base-commit: cee9395acd8043be0644b25c34bfa86623f2b935 > prerequisite-patch-id: a2f06addb17cc379b9f98674f087a1642588dc6e > -- > 2.53.0
