[PATCH v2 0/9] kho: add read-only Kexec Handover mode

tarunsahu at google.com tarunsahu at google.com
Sun Nov 15 13:26:54 PST 2026


IIUC, this series will be pciked after the kho feature flags [1]
Should it be rebased on top of [1]?

[1]: https://lore.kernel.org/all/20261110050202.529473-1-tarunsahu@google.com/

~Tarun

Pasha Tatashin <pasha.tatashin at soleen.com> writes:

> This series introduces a read-only Kexec Handover mode, enabled at boot
> via 'kho_ro=<pa>' pointing to the physical address of the KHO FDT.
>
> In standard KHO, the outgoing kernel shuts down or serializes all
> devices into LUO state, releases machine control, and passes the KHO
> FDT physical address through the kexec_file_load() reboot path to the
> next kernel, which becomes the sole "Machine Owner" (MO), initializes
> struct pages for preserved folios during early boot, and reclaims KHO
> metadata and preserved pages back into the page allocator as they are
> restored.
>
> In the "Linux Caretaker" (LC) model (as well as inspection kernels or
> firmware), a minimal LC kernel runs concurrently on dedicated CPUs
> offlined from the MO scheduler while MO is still active or undergoing a
> standard KHO-preserved kexec reboot (Live Update). LC is given
> read-only access to KHO so a minimal VMM inside LC can keep VMs
> responsive (<30ms VM blackout time) by servicing software-only, fully
> virtualized devices (such as virtio-console, pvpanic, virtio-rng, etc.)
> that do not require host hardware interrupts or global machine state,
> while leaving the preserved KHO tree intact for the incoming MO kernel
> to retrieve normally once it boots.
>
> LC is never MO: it has no PCI support, does not handle device
> interrupts, and is booted inside a small LUO-preserved memfd region
> (e.g., ~128MB of RAM on a host with terabytes of memory). Because LC's
> direct map only covers its own small memory footprint (both for
> isolation and to avoid struct page overhead for host memory), the KHO
> FDT, radix tree, and preserved memory all reside outside of LC's direct
> map. Furthermore, because LC is configured and launched from userspace
> while MO is still running rather than through a kexec reboot handoff
> from a shutting-down kernel, userspace reads the KHO FDT physical
> address from /sys/kernel/debug/kho/out/fdt_phys and explicitly passes
> 'kho_ro=<pa>' on LC's kernel command line.
>
> Both CONFIG_KEXEC_HANDOVER and CONFIG_KEXEC_HANDOVER_READONLY can be
> compiled into the same kernel image, with the active mode selected at
> boot via the kernel command line ('kho=on' vs 'kho_ro=<pa>'; passing
> 'kho_ro=<pa>' disables regular KHO).
>
> Overview of the series:
> - Patch 1 exposes the physical address of the outgoing KHO FDT via
>   /sys/kernel/debug/kho/out/fdt_phys so userspace tooling can pass it
>   to a 'kho_ro=<pa>' kernel.
> - Patch 2 moves the shared radix-tree key encoding, indexing, leaf
>   lookup (kho_radix_find_leaf()), FDT validation, and
>   kho_retrieve_subtree() helpers into a shared
>   kernel/liveupdate/kexec_handover_common.c compilation unit with
>   declarations in kernel/liveupdate/kexec_handover_internal.h, while
>   also checking that fdt_totalsize(fdt) does not exceed PAGE_SIZE in
>   kho_check_fdt() and using get_unaligned() for 64-bit FDT property
>   reads in kho_retrieve_subtree().
> - Patch 3 hardens kho_restore_vmalloc() and factors out shared vmalloc
>   helpers by introducing kho_vmalloc_validate() and kho_vmalloc_walk()
>   in kernel/liveupdate/kexec_handover_common.c to validate the
>   preservation metadata header and walk the chunk list.
> - Patch 4 avoids modifying ser->next and bs->head_pa in kho_block_add()
>   and kho_block_set_shrink_one() when operating on an incoming block
>   set, as the chain is already linked and incoming blocks must not be
>   modified in KHO read-only mode.
> - Patch 5 adds CONFIG_KEXEC_HANDOVER_READONLY and
>   kernel/liveupdate/kexec_handover_ro.c. When booted with
>   'kho_ro=<pa>', the KHO FDT is mapped and validated once in an
>   early_initcall, and radix-tree nodes, leaves, and folio/vmalloc
>   mappings are cached on first visit in a unified RCU-protected hash
>   table for the lifetime of the read-only KHO kernel, providing
>   kho_is_ro(), kho_ro_lookup_folio_order(), kho_ro_map_folio(), and
>   kho_ro_map_vmalloc().
> - Patch 6 wires read-only KHO mode into KHO core (kho_populate(),
>   kho_memory_init(), and kho_retrieve_subtree()).
> - Patch 7 updates kho_block to map incoming serialized blocks read-only
>   via kho_ro_map_folio() in read-only KHO mode.
> - Patches 8 and 9 extend test_kho and tools/testing/selftests/kho/vmtest.sh
>   to build a single kernel with both CONFIG_KEXEC_HANDOVER=y and
>   CONFIG_KEXEC_HANDOVER_READONLY=y and test both standard KHO and
>   read-only KHO handover when preserved memory lies outside the boot
>   kernel's initial memory range on x86_64 and arm64.
>
> This series is based on liveupdate/next commit 27be83a12b24 ("Merge
> branch 'lu-pci-preservation' into next").
>
> Changelog:
> v2:
> - Expand the cover letter with the LC architecture, explaining why
>   'kho_ro=<pa>' takes the FDT physical address on the command line and
>   describing the expected users of the read-only KHO APIs (Suggested by
>   Mike Rapoport).
> - Patch 2 (Suggested by Mike Rapoport):
>   - Move the shared radix-tree, FDT, and subtree retrieval helpers into
>     a dedicated kernel/liveupdate/kexec_handover_common.c file instead
>     of static inline functions in
>     kernel/liveupdate/kexec_handover_internal.h, and factor out
>     kho_radix_find_leaf() for radix leaf lookup.
>   - Check that fdt_totalsize(fdt) does not exceed PAGE_SIZE in
>     kho_check_fdt(), drop the "setup: " log prefix from FDT validation
>     warnings, and use get_unaligned() when reading 64-bit FDT properties
>     in kho_retrieve_subtree().
> - Patch 3:
>   - Move kho_flags_to_vmalloc() to kexec_handover_common.c alongside
>     kho_vmalloc_validate() and add a shared kho_vmalloc_walk() chunk
>     walker used by both kho_restore_vmalloc() and kho_ro_map_vmalloc().
> - Patches 5 and 6:
>   - Allow CONFIG_KEXEC_HANDOVER and CONFIG_KEXEC_HANDOVER_READONLY to be
>     compiled into the same kernel image and select the active mode via
>     the kernel command line ('kho_ro=<pa>' forces regular KHO disabled).
>   - Split the KHO core integration (kho_populate(), kho_memory_init(),
>     and kho_retrieve_subtree()) into a separate patch (Patch 6) so
>     Patch 5 only introduces the read-only KHO backend.
>   - Map and validate the KHO FDT and preserved-memory-map root once in
>     an early_initcall so kho_ro_get_fdt() and kho_retrieve_subtree() do
>     not sleep or map lazily.
>   - Replace kho_ro_restore_vmalloc() (which allocated RAM and copied
>     pages) with kho_ro_map_vmalloc() to map preserved kho_vmalloc
>     regions read-only in place using kho_vmalloc_walk().
>   - Cache radix-tree node, leaf, folio, and vmalloc mappings in a
>     unified RCU-protected hash table (kho_ro_map_ht) keyed by address,
>     size, and protection for the lifetime of the read-only KHO kernel,
>     and drop kho_ro_unmap().
> - Patch 9:
>   - Build the kernel and initrd only once in vmtest.sh with both
>     CONFIG_KEXEC_HANDOVER=y and CONFIG_KEXEC_HANDOVER_READONLY=y
>     enabled, running both the standard KHO test and the read-only KHO
>     test on the same kernel image.
> Link: https://lore.kernel.org/all/20261019021111.990394-1-pasha.tatashin@soleen.com/ # v1
>
> Pasha Tatashin (9):
>   kho: expose physical address of outgoing FDT
>   kho: move radix tree and FDT helpers to common file
>   kho: validate preservation metadata in kho_restore_vmalloc()
>   kho: do not modify incoming serialized blocks
>   kho: add read-only Kexec Handover mode
>   kho: support read-only KHO mode in KHO core
>   kho: support read-only KHO mode in kho_block
>   kho: test: support read-only KHO mode in test_kho
>   selftests: kho: add read-only KHO test to vmtest.sh
>
>  .../admin-guide/kernel-parameters.txt         |   8 +
>  Documentation/admin-guide/mm/kho.rst          |   3 +
>  Documentation/core-api/kho/index.rst          |   6 +
>  include/linux/kexec_handover.h                |  44 +-
>  kernel/liveupdate/Kconfig                     |  10 +
>  kernel/liveupdate/Makefile                    |   3 +-
>  kernel/liveupdate/kexec_handover.c            | 284 +++---------
>  kernel/liveupdate/kexec_handover_common.c     | 298 ++++++++++++
>  kernel/liveupdate/kexec_handover_debugfs.c    |  22 +-
>  kernel/liveupdate/kexec_handover_internal.h   |  47 +-
>  kernel/liveupdate/kexec_handover_ro.c         | 429 ++++++++++++++++++
>  kernel/liveupdate/kho_block.c                 |  78 +++-
>  lib/Kconfig.debug                             |   2 +-
>  lib/test_kho.c                                |  71 ++-
>  tools/testing/selftests/kho/arm64.conf        |   1 +
>  tools/testing/selftests/kho/init.c            |  38 +-
>  tools/testing/selftests/kho/vmtest.sh         |  65 ++-
>  tools/testing/selftests/kho/x86.conf          |   1 +
>  18 files changed, 1121 insertions(+), 289 deletions(-)
>  create mode 100644 kernel/liveupdate/kexec_handover_common.c
>  create mode 100644 kernel/liveupdate/kexec_handover_ro.c
>
>
> base-commit: 27be83a12b246b6b3783e7d3ac5c378ff7520f25
> -- 
> 2.56.0.307.g5fe7bd680c-goog



More information about the kexec mailing list