Hi guys, When looking at Masami Hiramatsu's kprobe cleanup series, it occurred to me the 'no KVM' isn't just about the EL1/EL2 split on non-VHE systems, but whether KVM is prepared to handle stepping on a breakpoint. It's not.
This labels all the VHE-only KVM functions that run during world switch with NOKPROBE_SYMBOL(), and always blacklists __hyp_text for kprobes to cover the common guest entry/exit code. I anticipate patch 1 going via the KVM tree to avoid conflicts. Patch 2 will conflict with [0]. I'm not sure what the best thing to do with the hyp-stub is. Patch 3 moves it to __hyp_text, and patch 4 covers the hibernate fallout from doing that. We don't have any other mmu-off but not idmap'd text. Probing the hyp-stub has to be done by address as the symbol names alias those in the __entry_text, which is blacklisted. (although this might depend on link order). I think anyone doing this is trying to shoot themselves in the foot. Know issues: * Other regions we should blacklist are the kexec and hibernate 'copy everything' code, as the vectors may have been overwritten by the time we step on the probe. cpu-suspend needs investigating... Thanks, James [0] lore.kernel.org/r/154753341900.31541.8135985235882849464.stgit@devbox James Morse (4): KVM: arm64: Forbid kprobing of the VHE world-switch code arm64: kprobe: Always blacklist the KVM world-switch code arm64: hyp-stub: Forbid kprobing of the hyp-stub arm64: hibernate: Clean the __hyp_text to PoC after resume arch/arm64/kernel/hibernate.c | 4 +++- arch/arm64/kernel/hyp-stub.S | 2 ++ arch/arm64/kernel/probes/kprobes.c | 6 +++--- arch/arm64/kvm/hyp/switch.c | 5 +++++ arch/arm64/kvm/hyp/sysreg-sr.c | 5 +++++ 5 files changed, 18 insertions(+), 4 deletions(-) -- 2.20.1 _______________________________________________ kvmarm mailing list kvmarm@lists.cs.columbia.edu https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
