The load_segments() function changes segment registers, invalidating GS base (which KCOV relies on for per-cpu data). When CONFIG_KCOV is enabled, any subsequent instrumented C code call (e.g. native_gdt_invalidate()) begins crashing the kernel in an endless loop. To reproduce the problem, it's sufficient to do kexec on a KCOV-instrumented kernel: $ kexec -l /boot/otherKernel $ kexec -e The real-world context for this problem is enabling crash dump collection in syzkaller. For this, the tool loads a panic kernel before fuzzing and then calls makedumpfile after the panic. This workflow requires both CONFIG_KEXEC and CONFIG_KCOV to be enabled simultaneously. Adding safeguards directly to the KCOV fast-path (__sanitizer_cov_trace_pc()) is also undesirable as it would introduce an extra performance overhead. Disabling instrumentation for the individual functions would be too fragile, so let's fix the bug by disabling KCOV instrumentation for the entire machine_kexec_64.c and physaddr.c. If coverage-guided fuzzing ever needs these components in the future, we should consider other approaches. The problem is not relevant for 32 bit kernels as CONFIG_KCOV is not supported there. Reviewed-by: Dmitry Vyukov Signed-off-by: Aleksandr Nogikh Cc: stable@vger.kernel.org --- v2: Updated the comments to explain the underlying context. v1: https://lore.kernel.org/all/20260216173716.2279847-1-nogikh@google.com/ --- arch/x86/kernel/Makefile | 10 ++++++++++ arch/x86/mm/Makefile | 10 ++++++++++ 2 files changed, 20 insertions(+) diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile index e9aeeeafad173..41b1333907ded 100644 --- a/arch/x86/kernel/Makefile +++ b/arch/x86/kernel/Makefile @@ -43,6 +43,16 @@ KCOV_INSTRUMENT_dumpstack_$(BITS).o := n KCOV_INSTRUMENT_unwind_orc.o := n KCOV_INSTRUMENT_unwind_frame.o := n KCOV_INSTRUMENT_unwind_guess.o := n +# Disable KCOV to prevent crashes during kexec: load_segments() invalidates +# the GS base, which KCOV relies on for per-CPU data. +# As KCOV && KEXEC compatibility should be preserved (e.g. syzkaller is +# using it to collect crash dumps during kernel fuzzing), we could either +# selectively disable KCOV instrumentation, which can be fragile, or add +# more checks to KCOV, which would slow it down. +# As a compromise solution, let's disable KCOV instrumentation for the +# whole file. If its coverage is ever needed, we should consider other +# approaches. +KCOV_INSTRUMENT_machine_kexec_64.o := n CFLAGS_head32.o := -fno-stack-protector CFLAGS_head64.o := -fno-stack-protector diff --git a/arch/x86/mm/Makefile b/arch/x86/mm/Makefile index 5b9908f13dcfd..ea3a31b54e49e 100644 --- a/arch/x86/mm/Makefile +++ b/arch/x86/mm/Makefile @@ -4,6 +4,16 @@ KCOV_INSTRUMENT_tlb.o := n KCOV_INSTRUMENT_mem_encrypt.o := n KCOV_INSTRUMENT_mem_encrypt_amd.o := n KCOV_INSTRUMENT_pgprot.o := n +# Disable KCOV to prevent crashes during kexec: load_segments() invalidates +# the GS base, which KCOV relies on for per-CPU data. +# As KCOV && KEXEC compatibility should be preserved (e.g. syzkaller is +# using it to collect crash dumps during kernel fuzzing), we could either +# selectively disable KCOV instrumentation, which can be fragile, or add +# more checks to KCOV, which would slow it down. +# As a compromise solution, let's disable KCOV instrumentation for the +# whole file. If its coverage is ever needed, we should consider other +# approaches. +KCOV_INSTRUMENT_physaddr.o := n KASAN_SANITIZE_mem_encrypt.o := n KASAN_SANITIZE_mem_encrypt_amd.o := n base-commit: f338e77383789c0cae23ca3d48adcc5e9e137e3c -- 2.53.0.959.g497ff81fa9-goog