
kpatch livepatch build and load test for CVE-2026-99999 on Rocky Linux 8.10 (4.18.0-553.144.1.el8_10)
Fixes CVE-2026-99999 (PwnEscape) on a running Rocky Linux host without a reboot, for the kernels listed below whose modules are released.
curl -fsSL https://raw.githubusercontent.com/nsrg-fmlorg/kpatch-rocky-cve-2026-99999/main/scripts/livepatch.sh | sudo bash -s -- apply
The script picks the module for the running kernel (uname -r), and first
checks the distribution, Secure Boot, and the module's vermagic and
SHA-256. If the kernel has no released module or any check fails, it
changes nothing. check instead of apply only checks; status and
remove show and undo it.
.ko yourself: docs/BUILD.mdA livepatch does not survive a reboot. The real fix is a kernel that
contains it, which Rocky Linux now ships for the latest minor of each
major: from 4.18.0-553.178.1.el8_10, 5.14.0-687.59.1.el9_8 and
6.12.0-211.67.1.el10_2 on (checked 2026-10-27). On those minors,
dnf update and a reboot fix a host.
The table lists one kernel per minor release, the oldest whose SRPM and debuginfo are both still on the mirrors (for 9.0, 9.1 and 9.2 only the last kernel of the minor is left), plus 4.18.0-553.144.1.el8_10.
Every other kernel on the mirrors is released too (2026-10-21, and
4.18.0-553.176.1 and 553.177.1.el8_10, which came out after, on
2026-10-30): 295 in
all, listed in tools/kernels.tsv, every kernel whose source and
debuginfo are on the mirrors, all of whose ports match AlmaLinux's
patch. 5.14.0-611.41.1.el9_7 is left out: its SRPM is in 9.7's vault,
its debuginfo is not. Each module passed the load test after its build,
and the one-line command covers all of them (scripts/livepatch.sh
holds each module's SHA-256), and was tested with that command on each
kernel in a VM: all 295 passed (2026-10-21; 553.176.1 and 553.177.1 on
2026-10-30, run 40911319606).
| Minor | Kernel | Tool | Status (2026-10-20) |
|---|---|---|---|
| 8.4 | 4.18.0-305.3.1.el8_4 | kpatch-build | released, load-tested |
| 8.5 | 4.18.0-348.el8.0.2 | kpatch-build | released, load-tested |
| 8.6 | 4.18.0-372.9.1.el8 | kpatch-build | released, load-tested |
| 8.7 | 4.18.0-425.3.1.el8 | kpatch-build | released, load-tested |
| 8.8 | 4.18.0-477.10.1.el8_8 | kpatch-build | released, load-tested |
| 8.9 | 4.18.0-513.5.1.el8_9 | kpatch-build | released, load-tested |
| 8.10 | 4.18.0-553.el8_10 | kpatch-build | released, load-tested |
| 8.10 | 4.18.0-553.144.1.el8_10 | kpatch-build | released, load-tested |
| 9.0 | 5.14.0-70.30.1.el9_0 | kpatch-build | released, load-tested |
| 9.1 | 5.14.0-162.23.1.el9_1 | kpatch-build | released, load-tested |
| 9.2 | 5.14.0-284.30.1.el9_2 | kpatch-build | released, load-tested |
| 9.3 | 5.14.0-362.8.1.el9_3 | kpatch-build | released, load-tested |
| 9.4 | 5.14.0-427.16.1.el9_4 | kpatch-build | released, load-tested |
| 9.5 | 5.14.0-503.14.1.el9_5 | kpatch-build | released, load-tested |
| 9.6 | 5.14.0-570.18.1.el9_6 | kpatch-build | released, load-tested |
| 9.7 | 5.14.0-611.5.1.el9_7 | kpatch-build | released, load-tested |
| 9.8 | 5.14.0-687.10.1.el9_8.0.1 | kpatch-build | released, load-tested |
| 10.0 | 6.12.0-55.12.1.el10_0 | klp-build (from 10.2, own objtool) | released, load-tested |
| 10.1 | 6.12.0-124.8.1.el10_1 | klp-build (from 10.2) | released, load-tested |
| 10.2 | 6.12.0-211.16.1.el10_2.0.1 | klp-build | released, load-tested |
Only released modules can be applied with the one-line command. Each
one was applied with that command, from the release, to its kernel in a
VM (the table's kernels 2026-10-20, runs 39363810989 and 39391770780; every kernel 2026-10-21, one run per minor; all 295 RESULT: PASS).
8.4 to 8.9 were released later than the rest: the test VM first stopped
in the GRUB menu on a boot entry the image's own kernel had left behind
(tools/loadtest-bootprep.sh explains it and now removes such entries).
Every other kernel on the mirrors has its patch in patches/; their
builds come next.
CVE-2026-99999 lets a guest VM, or a local user, make the host kernel's
checksum code overwrite an IPv4 header it has already parsed. The
upstream fix is torvalds/linux 6785011f8b16, "ipv4: validate
checksum_start before completing checksum". None of the kernels above
contains it (ip_check_csum_start is absent from their sources).
The change is what AlmaLinux ships in almalinux-testing (kept in
patches/almalinux-testing/):
nf_conntrack_bridge.c, so that hunk is left out there.patches/<kernel>.diff is that change against Rocky's source of each
kernel (tools/port.py); its sorted changed lines must equal AlmaLinux's
patches, or upstream's for 10 (tools/ref/). On 8.4, which lacks
skb_csum_is_sctp(), the port reads skb->csum_not_inet, which is all
that helper does on EL8.
patches/kpatch/<kernel>.diff is the same change in the form
kpatch-build accepts (tools/kpatchify.py), used for 8 and 9:
ip_check_csum_start() is defined in each file that uses it instead of
include/net/ip.h. A changed header recompiles everything that
includes it, and kpatch-build then stops on unrelated objects.skb_crc32c_csum_help that 26a77d02891a adds is left out;
kpatch-build cannot add exports. nfnetlink_queue reaches the function
through kpatch_skb_crc32c_csum_help(), added at the end of
net/core/skbuff.c, which only calls it.The code that runs is the same as AlmaLinux's.
8 and 9 use kpatch-build, pinned to dynup/kpatch 74ac9ffe96 (2026-09-15): Rocky support came after the v0.9.11 release, whose kpatch-build stops with "Unsupported distribution".
10 uses the kernel's own klp-build, because 10 enables
CONFIG_X86_KERNEL_IBT, which kpatch-build refuses. Only 10.2 ships
klp-build; 10.0 and 10.1 borrow it from 10.2's source
(tools/klp-overlay.sh), without changing the kernel's own code. On
10.0, 10.2's objtool misreads the kernel's objects, so those keep 10.0's
own objtool, and 10.2's only adds the checksums klp-build compares and
processes the livepatch module itself (tools/klp-overlay/objtool-split).
Every klp-build module is checked against the shipped kernel's symbols
and modversions CRCs (tools/klpsymcheck.py).
In every case gcc is pinned to the version the kernel was built with.
The build-and-load workflow builds each kernel and then loads the
module into that kernel, installed on the latest Rocky GenericCloud
image of its major in a VM (so a 10.0 kernel runs on 10.2's userland): the transition must finish, vmlinux and the patched modules must be
patched, a fragmented ping goes through ip_do_fragment(), no BUG,
WARNING or Oops may appear, and the patch must unload cleanly. It has not
been tested against attack packets.