@@ -408,11 +408,11 @@ static inline bool __hyp_handle_ptrauth(struct kvm_vcpu *vcpu)
}
/*
- * Return true when we were able to fixup the guest exit and should return to
- * the guest, false when we should restore the host state and return to the
- * main run loop.
+ * Prologue for the guest fixup, populating ESR_EL2 and fixing up PC
+ * if required.
*/
-static inline bool fixup_guest_exit(struct kvm_vcpu *vcpu, u64 *exit_code)
+static inline void fixup_guest_exit_prologue(struct kvm_vcpu *vcpu,
+ u64 *exit_code)
{
if (ARM_EXCEPTION_CODE(*exit_code) != ARM_EXCEPTION_IRQ)
vcpu->arch.fault.esr_el2 = read_sysreg_el2(SYS_ESR);
@@ -431,7 +431,15 @@ static inline bool fixup_guest_exit(struct kvm_vcpu *vcpu, u64 *exit_code)
if (esr_ec == ESR_ELx_EC_HVC32 || esr_ec == ESR_ELx_EC_HVC64)
write_sysreg_el2(read_sysreg_el2(SYS_ELR) - 4, SYS_ELR);
}
+}
+/*
+ * Return true when we were able to fixup the guest exit and should return to
+ * the guest, false when we should restore the host state and return to the
+ * main run loop.
+ */
+static inline bool fixup_guest_exit(struct kvm_vcpu *vcpu, u64 *exit_code)
+{
/*
* We're using the raw exception code in order to only process
* the trap if no SError is pending. We will come back to the
@@ -51,32 +51,12 @@ static inline void __sysreg_save_el1_state(struct kvm_cpu_context *ctxt)
ctxt_sys_reg(ctxt, SPSR_EL1) = read_sysreg_el1(SYS_SPSR);
}
-static inline u64 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 inline void __sysreg_save_el2_return_state(struct kvm_cpu_context *ctxt)
{
+ /* On VHE, PSTATE is saved in fixup_guest_exit_vhe() */
+ 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_final_cap(ARM64_HAS_RAS_EXTN))
ctxt_sys_reg(ctxt, DISR_EL1) = read_sysreg_s(SYS_VDISR_EL2);
@@ -166,6 +166,12 @@ static void __pmu_switch_to_host(struct kvm_cpu_context *host_ctxt)
write_sysreg(pmu->events_host, pmcntenset_el0);
}
+static bool fixup_guest_exit_nvhe(struct kvm_vcpu *vcpu, u64 *exit_code)
+{
+ fixup_guest_exit_prologue(vcpu, exit_code);
+ return fixup_guest_exit(vcpu, exit_code);
+}
+
/* Switch to the guest for legacy non-VHE systems */
int __kvm_vcpu_run(struct kvm_vcpu *vcpu)
{
@@ -227,7 +233,7 @@ int __kvm_vcpu_run(struct kvm_vcpu *vcpu)
exit_code = __guest_enter(vcpu);
/* And we're baaack! */
- } while (fixup_guest_exit(vcpu, &exit_code));
+ } while (fixup_guest_exit_nvhe(vcpu, &exit_code));
__sysreg_save_state_nvhe(guest_ctxt);
__sysreg32_save_state(vcpu);
@@ -154,12 +154,60 @@ void deactivate_traps_vhe_put(void)
__deactivate_traps_common();
}
+static bool fixup_guest_exit_vhe(struct kvm_vcpu *vcpu, u64 *exit_code,
+ bool hyp_ctxt)
+{
+ u64 pstate = read_sysreg_el2(SYS_SPSR);
+
+ /*
+ * 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.
+ *
+ * Consider this as the prologue before the fixup prologue...
+ */
+
+ 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;
+
+ fixup_guest_exit_prologue(vcpu, exit_code);
+
+ if (*exit_code == ARM_EXCEPTION_TRAP) {
+ /* more to come here */
+ }
+
+ return fixup_guest_exit(vcpu, exit_code);
+}
+
/* Switch to the guest for VHE systems running in EL2 */
static 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 = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
host_ctxt->__hyp_running_vcpu = vcpu;
@@ -186,12 +234,19 @@ static int __kvm_vcpu_run_vhe(struct kvm_vcpu *vcpu)
sysreg_restore_guest_state_vhe(guest_ctxt);
__debug_switch_to_guest(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.
+ */
+ WRITE_ONCE(hyp_ctxt, is_hyp_ctxt(vcpu));
+
do {
/* Jump in the fire! */
exit_code = __guest_enter(vcpu);
/* And we're baaack! */
- } while (fixup_guest_exit(vcpu, &exit_code));
+ } while (fixup_guest_exit_vhe(vcpu, &exit_code, READ_ONCE(hyp_ctxt)));
sysreg_save_guest_state_vhe(guest_ctxt);
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/include/hyp/switch.h | 16 ++++-- arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h | 26 ++-------- arch/arm64/kvm/hyp/nvhe/switch.c | 8 ++- arch/arm64/kvm/hyp/vhe/switch.c | 57 +++++++++++++++++++++- 4 files changed, 78 insertions(+), 29 deletions(-)