Using a vfp_hard_struct per vcpu for the host is a bit overkill
(odds are we have more vcpus than physical CPUs). It then makes
sense to have a global set of vfp_hard_structs for the host.
Define a per-cpu variable holding the host CPU states, and let
each vcpu point to the right state as they're about to run.
Cc: Antonios Motakis <a.motakis@xxxxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
---
arch/arm/include/asm/kvm_host.h | 2 +-
arch/arm/kvm/arm.c | 19 ++++++++++++++++++
arch/arm/kvm/interrupts.S | 42 ++++++++++++++++++++-------------------
3 files changed, 42 insertions(+), 21 deletions(-)
diff --git a/arch/arm/include/asm/kvm_host.h b/arch/arm/include/asm/kvm_host.h
index 1e4366e..678aa33 100644
--- a/arch/arm/include/asm/kvm_host.h
+++ b/arch/arm/include/asm/kvm_host.h
@@ -116,7 +116,7 @@ struct kvm_vcpu_arch {
/* Floating point registers (VFP and Advanced SIMD/NEON) */
struct vfp_hard_struct vfp_guest;
- struct vfp_hard_struct vfp_host;
+ struct vfp_hard_struct *vfp_host;
/*
* Anything that is not used directly from assembly code goes
diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c
index 803b97c..49d916c 100644
--- a/arch/arm/kvm/arm.c
+++ b/arch/arm/kvm/arm.c
@@ -55,6 +55,9 @@ __asm__(".arch_extension virt");
static DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
+/* Per-CPU variable containing the host VFP state */
+static DEFINE_PER_CPU(struct vfp_hard_struct, kvm_host_vfp_state);
+
/* The VMID used in the VTTBR */
static atomic64_t kvm_vmid_gen = ATOMIC64_INIT(1);
static u8 kvm_next_vmid;
@@ -283,6 +286,7 @@ void kvm_arch_vcpu_uninit(struct kvm_vcpu *vcpu)
void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
{
vcpu->cpu = cpu;
+ vcpu->arch.vfp_host = this_cpu_ptr(&kvm_host_vfp_state);
/*
* Check whether this vcpu requires the cache to be flushed on
@@ -771,6 +775,21 @@ static int init_hyp_mode(void)
}
}
+ /*
+ * Map the host VFP structures
+ */
+ for_each_possible_cpu(cpu) {
+ struct vfp_hard_struct *vfp;
+
+ vfp = &per_cpu(kvm_host_vfp_state, cpu);
+ err = create_hyp_mappings(vfp, vfp + 1);
+
+ if (err) {
+ kvm_err("Cannot map host VFP state\n");
+ goto out_free_mappings;
+ }
+ }
+
return 0;
out_free_mappings:
free_hyp_pmds();
diff --git a/arch/arm/kvm/interrupts.S b/arch/arm/kvm/interrupts.S
index 450144a..4d6cc71 100644
--- a/arch/arm/kvm/interrupts.S
+++ b/arch/arm/kvm/interrupts.S
@@ -90,15 +90,13 @@ ENTRY(__kvm_flush_vm_context)
bx lr
ENDPROC(__kvm_flush_vm_context)
-/* Clobbers {r2-r7} */
-.macro store_vfp_state vfp_off, vcpup
- add r6, \vcpup, \vfp_off
-
+/* Clobbers {r2-r6} */
+.macro store_vfp_state vfp_base
@ The VFPFMRX and VFPFMXR macros are the VMRS and VMSR instructions
VFPFMRX r2, FPEXC
@ Make sure VFP is enabled so we can touch the registers.
- orr r7, r2, #FPEXC_EN
- VFPFMXR FPEXC, r7
+ orr r6, r2, #FPEXC_EN
+ VFPFMXR FPEXC, r6
VFPFMRX r3, FPSCR
tst r2, #FPEXC_EX @ Check for VFP Subarchitecture
@@ -108,19 +106,17 @@ ENDPROC(__kvm_flush_vm_context)
VFPFMRX r4, FPINST
tst r2, #FPEXC_FP2V
VFPFMRX r5, FPINST2, ne @ vmrsne
- bic r7, r2, #FPEXC_EX @ FPEXC_EX disable
- VFPFMXR FPEXC, r7
+ bic r6, r2, #FPEXC_EX @ FPEXC_EX disable
+ VFPFMXR FPEXC, r6
1:
- VFPFSTMIA r6, r7 @ Save VFP registers
- stm r6, {r2-r5} @ Save FPEXC, FPSCR, FPINST, FPINST2
+ VFPFSTMIA \vfp_base, r6 @ Save VFP registers
+ stm \vfp_base, {r2-r5} @ Save FPEXC, FPSCR, FPINST, FPINST2
.endm
-/* Assume FPEXC_EN is on and FPEXC_EX is off, clobbers {r2-r7} */
-.macro restore_vfp_state vfp_off, vcpup
- add r6, \vcpup, \vfp_off
-
- VFPFLDMIA r6, r7 @ Load VFP registers
- ldm r6, {r2-r5} @ Load FPEXC, FPSCR, FPINST, FPINST2
+/* Assume FPEXC_EN is on and FPEXC_EX is off, clobbers {r2-r6} */
+.macro restore_vfp_state vfp_base
+ VFPFLDMIA \vfp_base, r6 @ Load VFP registers
+ ldm \vfp_base, {r2-r5} @ Load FPEXC, FPSCR, FPINST, FPINST2
VFPFMXR FPSCR, r3
tst r2, #FPEXC_EX @ Check for VFP Subarchitecture
@@ -456,8 +452,11 @@ __kvm_vcpu_return:
bne after_vfp_restore
@ Switch VFP/NEON hardware state to the host's
- store_vfp_state #VCPU_VFP_GUEST, r1
- restore_vfp_state #VCPU_VFP_HOST, r1
+ add r7, r1, #VCPU_VFP_GUEST
+ store_vfp_state r7
+ add r7, r1, #VCPU_VFP_HOST
+ ldr r7, [r7]
+ restore_vfp_state r7
after_vfp_restore:
@ Restore FPEXC_EN which we clobbered on entry
@@ -728,8 +727,11 @@ switch_to_guest_vfp:
set_hcptr 0, (HCPTR_TCP(10) | HCPTR_TCP(11))
@ Switch VFP/NEON hardware state to the guest's
- store_vfp_state #VCPU_VFP_HOST, r0
- restore_vfp_state #VCPU_VFP_GUEST, r0
+ add r7, r0, #VCPU_VFP_HOST
+ ldr r7, [r7]
+ store_vfp_state r7
+ add r7, r0, #VCPU_VFP_GUEST
+ restore_vfp_state r7
pop {r3-r7}
pop {r0-r2}
--
1.7.10.3
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/cucslists/listinfo/kvmarm
