@@ -537,8 +537,45 @@ static bool __hyp_text handle_tx2_tvm(struct kvm_vcpu *vcpu)
* the guest, false when we should restore the host state and return to the
* main run loop.
*/
-static bool __hyp_text fixup_guest_exit(struct kvm_vcpu *vcpu, u64 *exit_code)
+static bool __hyp_text fixup_guest_exit(struct kvm_vcpu *vcpu, u64 *exit_code,
+ bool hyp_ctxt)
{
+ /*
+ * Sync pstate back as early as possible, so that is_hyp_ctxt()
+ * reflects the exact context. It is otherwise possible to get
+ * confused with a VHE guest and ARMv8.4-NV, such as:
+ *
+ * (1) enter guest in host EL0
+ * (2) guest traps to guest vEL2 (no hypervisor intervention)
+ * (3) guest clears virtual HCR_EL2.TGE (no trap either)
+ * (4) host interrupt fires, exit
+ * (5) is_hyp_ctxt() now says "guest" (pstate.M==EL1 and TGE==0)
+ *
+ * If host preemption occurs, vcpu_load/put() will be very confused.
+ * This of course only matters to VHE.
+ */
+
+ if (has_vhe()) {
+ u64 pstate = read_sysreg_el2(SYS_SPSR);
+
+ if (unlikely(hyp_ctxt)) {
+ u64 mode = pstate & PSR_MODE_MASK;
+
+ switch (mode) {
+ case PSR_MODE_EL1t:
+ mode = PSR_MODE_EL2t;
+ break;
+ case PSR_MODE_EL1h:
+ mode = PSR_MODE_EL2h;
+ break;
+ }
+
+ pstate = (pstate & ~PSR_MODE_MASK) | mode;
+ }
+
+ *vcpu_cpsr(vcpu) = pstate;
+ }
+
if (ARM_EXCEPTION_CODE(*exit_code) != ARM_EXCEPTION_IRQ)
vcpu->arch.fault.esr_el2 = read_sysreg_el2(SYS_ESR);
@@ -683,6 +720,7 @@ int kvm_vcpu_run_vhe(struct kvm_vcpu *vcpu)
struct kvm_cpu_context *host_ctxt;
struct kvm_cpu_context *guest_ctxt;
u64 exit_code;
+ bool hyp_ctxt;
host_ctxt = vcpu->arch.host_cpu_context;
host_ctxt->__hyp_running_vcpu = vcpu;
@@ -709,12 +747,19 @@ int kvm_vcpu_run_vhe(struct kvm_vcpu *vcpu)
__set_guest_arch_workaround_state(vcpu);
+ /*
+ * Being in HYP context or not is an invariant here. If we enter in
+ * a given context, we exit in the same context. We can thus only
+ * sample the context once.
+ */
+ hyp_ctxt = is_hyp_ctxt(vcpu);
+
do {
/* Jump in the fire! */
exit_code = __guest_enter(vcpu, host_ctxt);
/* And we're baaack! */
- } while (fixup_guest_exit(vcpu, &exit_code));
+ } while (fixup_guest_exit(vcpu, &exit_code, hyp_ctxt));
__set_host_arch_workaround_state(vcpu);
@@ -788,7 +833,7 @@ int __hyp_text __kvm_vcpu_run_nvhe(struct kvm_vcpu *vcpu)
exit_code = __guest_enter(vcpu, host_ctxt);
/* And we're baaack! */
- } while (fixup_guest_exit(vcpu, &exit_code));
+ } while (fixup_guest_exit(vcpu, &exit_code, false));
__set_host_arch_workaround_state(vcpu);
@@ -109,32 +109,12 @@ static void __hyp_text __sysreg_save_el1_state(struct kvm_cpu_context *ctxt)
__sysreg_save_vel1_state(ctxt);
}
-static u64 __hyp_text from_hw_pstate(const struct kvm_cpu_context *ctxt)
-{
- u64 reg = read_sysreg_el2(SYS_SPSR);
-
- if (__is_hyp_ctxt(ctxt)) {
- u64 mode = reg & (PSR_MODE_MASK | PSR_MODE32_BIT);
-
- switch (mode) {
- case PSR_MODE_EL1t:
- mode = PSR_MODE_EL2t;
- break;
- case PSR_MODE_EL1h:
- mode = PSR_MODE_EL2h;
- break;
- }
-
- return (reg & ~(PSR_MODE_MASK | PSR_MODE32_BIT)) | mode;
- }
-
- return reg;
-}
-
static void __hyp_text __sysreg_save_el2_return_state(struct kvm_cpu_context *ctxt)
{
+ /* On VHE, PSTATE is saved in fixup_guest_exit() */
+ if (!has_vhe())
+ ctxt->regs.pstate = read_sysreg_el2(SYS_SPSR);
ctxt->regs.pc = read_sysreg_el2(SYS_ELR);
- ctxt->regs.pstate = from_hw_pstate(ctxt);
if (cpus_have_const_cap(ARM64_HAS_RAS_EXTN))
ctxt_sys_reg(ctxt, DISR_EL1) = read_sysreg_s(SYS_VDISR_EL2);
The NV code relies on predicates such as is_hyp_ctxt() being reliable. In turn, is_hyp_ctxt() relies on things like PSTATE and the virtual HCR_EL2 being accurate. But with ARMv8.4-NV removing trapping for a large part of the EL2 system registers (among which HCR_EL2), we can't use such trapping to synchronize the rest of the state. Let's look at the following sequence for a VHE guest: (1) enter guest in host EL0 (2) guest traps to guest vEL2 (no hypervisor intervention) (3) guest clears virtual HCR_EL2.TGE (no trap either) (4) host interrupt fires, exit (5) is_hyp_ctxt() now says "guest" (PSTATE.M==EL1 and TGE==0) It is obvious that such behaviour would be rather unfortunate, and lead to interesting, difficult to catch bugs specially if preemption kicks in (yes, I wasted a whole week chasing this one). In order to preserve the invariant that a guest entered in host context must exit in the same context, we must make sure that is_hyp_ctxt() works correctly. Since we can always observe the guest value of HCR_EL2.{E2H,TGE} in the VNCR_EL2 page, we solely need to synchronize PSTATE as early as possible. This basically amounts to moving from_hw_pstate() as close as possible to the guest exit point, and fixup_guest_exit() seems as good a place as any. Signed-off-by: Marc Zyngier <maz@kernel.org> --- arch/arm64/kvm/hyp/switch.c | 51 ++++++++++++++++++++++++++++++++-- arch/arm64/kvm/hyp/sysreg-sr.c | 26 ++--------------- 2 files changed, 51 insertions(+), 26 deletions(-)