We currently spend around ~400 cycles on each entry/exit to the guest
dealing with arch timer registers, even when the timer is not pending
and not doing anything (on certain architectures).
We can do much better by moving the arch timer save/restore to the
vcpu_load and vcpu_put functions, but this means that if we don't read
back the timer state on every exit from the guest, then we have to be
able to start taking timer interrupts for the virtual timer in KVM and
handle that properly.
That has a number of entertaining consequences, such as having to make
sure we don't deadlock between any of the vgic code and interrupt
injection happening from an ISR. On the plus side, being able to inject
virtual interrupts corresponding to a physical interrupt directly from
an ISR is probably a good system design change overall.
We also have to change the use of the physical vs. virtual counter in
the arm64 kernel to avoid having to save/restore the CNTVOFF_EL2
register on every return to the hypervisor. The only reason I could
find for using the virtual counter for the kernel on systems with access
to the physical counter is to detect if firmware did not properly clear
CNTVOFF_EL2, and this change has to weighed against the existing check
(assuming I got this right).
On a non-VHE system (AMD Seattle) I have measured this to improve the
world-switch time by about ~100 cycles, but on an EL2 kernel (emulating
VHE behavior on the same hardware) this gives us around ~250 cycles
worth of improvement, because we can avoid the extra configuration of
trapping accesses to the physical timer from EL1 on every switch.
These patches require that the GICv2 hardware (on such systems) is
properly reported by firmware to have the extra CPU interface page for
the deactivate register.
Based on v4.13-rc1
Code is also available here:
git://git.kernel.org/pub/scm/linux/kernel/git/cdall/linux.git timer-optimize-rfc-v2
Thanks,
Christoffer
Christoffer Dall (19):
arm64: Use physical counter for in-kernel reads
arm64: Use the physical counter when available for read_cycles
KVM: arm/arm64: Guard kvm_vgic_map_is_active against !vgic_initialized
KVM: arm/arm64: Support calling vgic_update_irq_pending from irq
context
KVM: arm/arm64: Check that system supports split eoi/deactivate
KVM: arm/arm64: Make timer_arm and timer_disarm helpers more generic
KVM: arm/arm64: Rename soft timer to bg_timer
KVM: arm/arm64: Use separate timer for phys timer emulation
KVM: arm/arm64: Move timer/vgic flush/sync under disabled irq
KVM: arm/arm64: Move timer save/restore out of the hyp code
genirq: Document vcpu_info usage for per-CPU interrupts
KVM: arm/arm64: Set VCPU affinity for virt timer irq
KVM: arm/arm64: Avoid timer save/restore in vcpu entry/exit
KVM: arm/arm64: Support EL1 phys timer register access in set/get reg
KVM: arm/arm64: Use kvm_arm_timer_set/get_reg for guest register traps
KVM: arm/arm64: Move phys_timer_emulate function
KVM: arm/arm64: Avoid phys timer emulation in vcpu entry/exit
KVM: arm/arm64: Get rid of kvm_timer_flush_hwstate
KVM: arm/arm64: Rework kvm_timer_should_fire
arch/arm/include/asm/kvm_asm.h | 2 +
arch/arm/include/asm/kvm_hyp.h | 4 +-
arch/arm/include/uapi/asm/kvm.h | 6 +
arch/arm/kvm/hyp/switch.c | 7 +-
arch/arm64/include/asm/arch_timer.h | 18 +-
arch/arm64/include/asm/kvm_asm.h | 2 +
arch/arm64/include/asm/kvm_hyp.h | 4 +-
arch/arm64/include/asm/timex.h | 2 +-
arch/arm64/include/uapi/asm/kvm.h | 6 +
arch/arm64/kvm/hyp/switch.c | 6 +-
arch/arm64/kvm/sys_regs.c | 41 ++--
drivers/clocksource/arm_arch_timer.c | 33 ++-
drivers/irqchip/irq-gic.c | 12 +-
include/kvm/arm_arch_timer.h | 19 +-
kernel/irq/manage.c | 3 +-
virt/kvm/arm/arch_timer.c | 446 ++++++++++++++++++++++++-----------
virt/kvm/arm/arm.c | 45 ++--
virt/kvm/arm/hyp/timer-sr.c | 74 +++---
virt/kvm/arm/vgic/vgic-its.c | 17 +-
virt/kvm/arm/vgic/vgic-mmio-v2.c | 22 +-
virt/kvm/arm/vgic/vgic-mmio-v3.c | 17 +-
virt/kvm/arm/vgic/vgic-mmio.c | 44 ++--
virt/kvm/arm/vgic/vgic-v2.c | 5 +-
virt/kvm/arm/vgic/vgic-v3.c | 12 +-
virt/kvm/arm/vgic/vgic.c | 63 +++--
virt/kvm/arm/vgic/vgic.h | 3 +-
26 files changed, 586 insertions(+), 327 deletions(-)
--
2.9.0
