@@ -62,8 +62,26 @@
#define CHOOSE_VHE_SYM(sym) sym
#define CHOOSE_NVHE_SYM(sym) kvm_nvhe_sym(sym)
-#define CHOOSE_HYP_SYM(sym) (has_vhe() ? CHOOSE_VHE_SYM(sym) \
+
+#ifndef __KVM_NVHE_HYPERVISOR__
+/*
+ * BIG FAT WARNINGS:
+ *
+ * - Don't be tempted to change the following is_kernel_in_hyp_mode()
+ * to has_vhe(). has_vhe() is implemented as a *final* capability,
+ * while this is used early at boot time, when the capabilities are
+ * not final yet....
+ *
+ * - Don't let the nVHE hypervisor have access to this, as it will
+ * pick the *wrong* symbol (yes, it runs at EL2...).
+ */
+#define CHOOSE_HYP_SYM(sym) (is_kernel_in_hyp_mode() ? CHOOSE_VHE_SYM(sym) \
: CHOOSE_NVHE_SYM(sym))
+#else
+/* The nVHE hypervisor shouldn't even try to access anything */
+extern void *__nvhe_undefined_symbol;
+#define CHOOSE_HYP_SYM(sym) __nvhe_undefined_symbol
+#endif
/* Translate a kernel address @ptr into its equivalent linear mapping */
#define kvm_ksym_ref(ptr) \
The recently introduced CHOOSE_HYP_SYM() macro picks one symbol or another, depending on whether the kernel run as a VHE hypervisor or not. For that, it uses the has_vhe() helper, which is itself implemented as a final capability. Unfortunately, __copy_hyp_vect_bpi now indirectly uses CHOOSE_HYP_SYM to get the __bp_harden_hyp_vecs symbol, using has_vhe() in the process. At this stage, the capability isn't final and things explode: [ 0.000000] ACPI: SRAT not present [ 0.000000] percpu: Embedded 34 pages/cpu s101264 r8192 d29808 u139264 [ 0.000000] Detected PIPT I-cache on CPU0 [ 0.000000] ------------[ cut here ]------------ [ 0.000000] kernel BUG at arch/arm64/include/asm/cpufeature.h:459! [ 0.000000] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP [ 0.000000] Modules linked in: [ 0.000000] CPU: 0 PID: 0 Comm: swapper Not tainted 5.8.0-rc4-00080-gd630681366e5 #1388 [ 0.000000] pstate: 80000085 (Nzcv daIf -PAN -UAO BTYPE=--) [ 0.000000] pc : check_branch_predictor+0x3a4/0x408 [ 0.000000] lr : check_branch_predictor+0x2a4/0x408 [ 0.000000] sp : ffff800011693e90 [ 0.000000] x29: ffff800011693e90 x28: ffff8000116a1530 [ 0.000000] x27: ffff8000112c1008 x26: ffff800010ca6ff8 [ 0.000000] x25: ffff8000112c1000 x24: ffff8000116a1320 [ 0.000000] x23: 0000000000000000 x22: ffff8000112c1000 [ 0.000000] x21: ffff800010177120 x20: ffff8000116ae108 [ 0.000000] x19: 0000000000000000 x18: ffff800011965c90 [ 0.000000] x17: 0000000000022000 x16: 0000000000000003 [ 0.000000] x15: 00000000ffffffff x14: ffff8000118c3a38 [ 0.000000] x13: 0000000000000021 x12: 0000000000000022 [ 0.000000] x11: d37a6f4de9bd37a7 x10: 000000000000001d [ 0.000000] x9 : 0000000000000000 x8 : ffff800011f8dad8 [ 0.000000] x7 : ffff800011965ad0 x6 : 0000000000000003 [ 0.000000] x5 : 0000000000000000 x4 : 0000000000000000 [ 0.000000] x3 : 0000000000000100 x2 : 0000000000000004 [ 0.000000] x1 : ffff8000116ae148 x0 : 0000000000000000 [ 0.000000] Call trace: [ 0.000000] check_branch_predictor+0x3a4/0x408 [ 0.000000] update_cpu_capabilities+0x84/0x138 [ 0.000000] init_cpu_features+0x2c0/0x2d8 [ 0.000000] cpuinfo_store_boot_cpu+0x54/0x64 [ 0.000000] smp_prepare_boot_cpu+0x2c/0x60 [ 0.000000] start_kernel+0x16c/0x574 [ 0.000000] Code: 17ffffc7 91010281 14000198 17ffffca (d4210000) This is addressed using a two-fold process: - Replace has_vhe() with is_kernel_in_hyp_mode(), which tests whether we are running at EL2. - Make CHOOSE_HYP_SYM() return an *undefined* symbol when compiled in the nVHE hypervisor, as we really should never use this helper in the nVHE-specific code. With this in place, we're back to a bootable kernel again. Fixes: b877e9849d41 ("KVM: arm64: Build hyp-entry.S separately for VHE/nVHE") Signed-off-by: Marc Zyngier <maz@kernel.org> --- arch/arm64/include/asm/kvm_asm.h | 20 +++++++++++++++++++- 1 file changed, 19 insertions(+), 1 deletion(-)