On Thu, Dec 01, 2022, Alexey Kardashevskiy wrote:
Reading DR[0-3]_ADDR_MASK MSRs takes about 250 cycles which is going to
be noticeable when the AMD KVM SEV-ES's DebugSwap feature is enabled and
KVM needs to store these before switching to a guest; the DebugSwitch
hardware support restores them as type B swap.
This stores MSR values from set_dr_addr_mask() in percpu values and
returns them via new get_dr_addr_mask(). The gain here is about 10x.
Signed-off-by: Alexey Kardashevskiy <aik@xxxxxxx>
---
arch/x86/include/asm/debugreg.h | 1 +
arch/x86/kernel/cpu/amd.c | 32 ++++++++++++++++++++
2 files changed, 33 insertions(+)
diff --git a/arch/x86/include/asm/debugreg.h b/arch/x86/include/asm/debugreg.h
index cfdf307ddc01..c4324d0205b5 100644
--- a/arch/x86/include/asm/debugreg.h
+++ b/arch/x86/include/asm/debugreg.h
@@ -127,6 +127,7 @@ static __always_inline void local_db_restore(unsigned long dr7)
#ifdef CONFIG_CPU_SUP_AMD
extern void set_dr_addr_mask(unsigned long mask, int dr);
+extern unsigned long get_dr_addr_mask(int dr);
#else
static inline void set_dr_addr_mask(unsigned long mask, int dr) { }
KVM_AMD doesn't depend on CPU_SUP_AMD, i.e. this needs a stub. Or we need to add
a dependency.
diff --git a/arch/x86/kernel/cpu/amd.c b/arch/x86/kernel/cpu/amd.c
index c75d75b9f11a..ec7efcef4e14 100644
--- a/arch/x86/kernel/cpu/amd.c
+++ b/arch/x86/kernel/cpu/amd.c
@@ -1158,6 +1158,11 @@ static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
return false;
}
+DEFINE_PER_CPU_READ_MOSTLY(unsigned long, dr0_addr_mask);
+DEFINE_PER_CPU_READ_MOSTLY(unsigned long, dr1_addr_mask);
+DEFINE_PER_CPU_READ_MOSTLY(unsigned long, dr2_addr_mask);
+DEFINE_PER_CPU_READ_MOSTLY(unsigned long, dr3_addr_mask);
+
void set_dr_addr_mask(unsigned long mask, int dr)
{
if (!boot_cpu_has(X86_FEATURE_BPEXT))
@@ -1166,17 +1171,44 @@ void set_dr_addr_mask(unsigned long mask, int dr)
switch (dr) {
case 0:
wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
LOL, I'd love to hear how MSR_F16H_DR0_ADDR_MASK ended up with a completely
different MSR index.
+ per_cpu(dr0_addr_mask, smp_processor_id()) = mask;
Use an array to avoid the copy+paste? And if you're going to add a cache, might
as well use it to avoid unnecessary writes.
break;
case 1: