Yu-hsin Wang has submitted this change. (
https://gem5-review.googlesource.com/c/public/gem5/+/42161 )
Change subject: cpu-kvm,arch-arm: correct arm kvm virtual time
......................................................................
cpu-kvm,arch-arm: correct arm kvm virtual time
According to the kernel code*1, the virtual time is related to physical
timer and cntvoff. When the simulator goes from KVM to gem5, the
physical timer is still ticking. After gem5 simulating models and going
back to KVM, the virtual time also goes away. We should update cntvoff
by ourselve to get correct virtual time.
Moreover, according to update_vtimer_cntvoff*2, setting cntvoff affacts
all vcpus. Instead of puting individual vtime on BaseArmKvmCPU, we
maintain a global vtime, restore it before the first vcpu going into
KVM, and save it after the last vcpu back from KVM.
1. https://code.woboq.org/linux/linux/virt/kvm/arm/arch_timer.c.html#826
2.
https://code.woboq.org/linux/linux/virt/kvm/arm/arch_timer.c.html#update_vtimer_cntvoff
Change-Id: Ie054104642f2a6d5a0740f39b947f5f2c29c36f3
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/42161
Reviewed-by: Earl Ou <shunhsin...@google.com>
Reviewed-by: Andreas Sandberg <andreas.sandb...@arm.com>
Maintainer: Andreas Sandberg <andreas.sandb...@arm.com>
Tested-by: kokoro <noreply+kok...@google.com>
---
M src/arch/arm/kvm/base_cpu.cc
M src/arch/arm/kvm/base_cpu.hh
2 files changed, 39 insertions(+), 0 deletions(-)
Approvals:
Andreas Sandberg: Looks good to me, approved; Looks good to me, approved
Earl Ou: Looks good to me, but someone else must approve
kokoro: Regressions pass
diff --git a/src/arch/arm/kvm/base_cpu.cc b/src/arch/arm/kvm/base_cpu.cc
index 2a948c7..ae53095 100644
--- a/src/arch/arm/kvm/base_cpu.cc
+++ b/src/arch/arm/kvm/base_cpu.cc
@@ -38,8 +38,10 @@
#include "arch/arm/kvm/base_cpu.hh"
#include <linux/kvm.h>
+#include <mutex>
#include "arch/arm/interrupts.hh"
+#include "base/uncontended_mutex.hh"
#include "debug/KvmInt.hh"
#include "dev/arm/generic_timer.hh"
#include "params/BaseArmKvmCPU.hh"
@@ -58,6 +60,21 @@
#define INTERRUPT_VCPU_FIQ(vcpu) \
INTERRUPT_ID(KVM_ARM_IRQ_TYPE_CPU, vcpu, KVM_ARM_IRQ_CPU_FIQ)
+namespace {
+
+/**
+ * When the simulator returns from KVM for simulating other models, the
+ * in-kernel timer doesn't stop. We have to save the virtual time and
+ * restore before going into KVM next time. Moreover, setting virtual time
+ * affacts all vcpus according to the kvm implementation. We maintain a
global
+ * virtual time here, restore it before the first vcpu going into KVM, and
save
+ * it after the last vcpu back from KVM.
+ */
+uint64_t vtime = 0;
+uint64_t vtime_counter = 0;
+UncontendedMutex vtime_mutex;
+
+} // namespace
BaseArmKvmCPU::BaseArmKvmCPU(const BaseArmKvmCPUParams ¶ms)
: BaseKvmCPU(params),
@@ -147,6 +164,26 @@
return kvmRunTicks;
}
+void
+BaseArmKvmCPU::ioctlRun()
+{
+ // Check if it's the first vcpu going into KVM. If yes, it should
restore
+ // the virtual time.
+ {
+ std::lock_guard<UncontendedMutex> l(vtime_mutex);
+ if (vtime_counter++ == 0)
+ setOneReg(KVM_REG_ARM_TIMER_CNT, vtime);
+ }
+ BaseKvmCPU::ioctlRun();
+ // Check if it's the last vcpu back from KVM. If yes, it should save
the
+ // virtual time.
+ {
+ std::lock_guard<UncontendedMutex> l(vtime_mutex);
+ if (--vtime_counter == 0)
+ getOneReg(KVM_REG_ARM_TIMER_CNT, &vtime);
+ }
+}
+
const BaseArmKvmCPU::RegIndexVector &
BaseArmKvmCPU::getRegList() const
{
diff --git a/src/arch/arm/kvm/base_cpu.hh b/src/arch/arm/kvm/base_cpu.hh
index 6419547..e4bc49f 100644
--- a/src/arch/arm/kvm/base_cpu.hh
+++ b/src/arch/arm/kvm/base_cpu.hh
@@ -56,6 +56,8 @@
protected:
Tick kvmRun(Tick ticks) override;
+ /** Override for synchronizing state in kvm_run */
+ void ioctlRun() override;
/** Cached state of the IRQ line */
bool irqAsserted;
--
To view, visit https://gem5-review.googlesource.com/c/public/gem5/+/42161
To unsubscribe, or for help writing mail filters, visit
https://gem5-review.googlesource.com/settings
Gerrit-Project: public/gem5
Gerrit-Branch: develop
Gerrit-Change-Id: Ie054104642f2a6d5a0740f39b947f5f2c29c36f3
Gerrit-Change-Number: 42161
Gerrit-PatchSet: 7
Gerrit-Owner: Yu-hsin Wang <yuhsi...@google.com>
Gerrit-Reviewer: Andreas Sandberg <andreas.sandb...@arm.com>
Gerrit-Reviewer: Earl Ou <shunhsin...@google.com>
Gerrit-Reviewer: Gabe Black <gabebl...@google.com>
Gerrit-Reviewer: Giacomo Travaglini <giacomo.travagl...@arm.com>
Gerrit-Reviewer: Yu-hsin Wang <yuhsi...@google.com>
Gerrit-Reviewer: kokoro <noreply+kok...@google.com>
Gerrit-CC: Gabe Black <gabe.bl...@gmail.com>
Gerrit-MessageType: merged
_______________________________________________
gem5-dev mailing list -- gem5-dev@gem5.org
To unsubscribe send an email to gem5-dev-le...@gem5.org
%(web_page_url)slistinfo%(cgiext)s/%(_internal_name)s
