haltpoll
-
kvm 参数中存在 haltpolling 操作: https://docs.kernel.org/virt/guest-halt-polling.html https://liujunming.top/2022/10/29/Notes-about-kvm-halt-polling/
-
此外，存在一个 cstate governor 来控制:
➜ cpuidle cpupower idle-info
CPUidle driver: none
CPUidle governor: haltpoll
analyzing CPU 18:
CPU 18: No idle states
这个到底做啥来的
- drivers/cpuidle/governors/haltpoll.c
- 这个显然，有一个 governors 就是 haltpoll，代码量很小，只有几百行
- drivers/cpuidle/cpuidle-haltpoll.c
- 这个就是 idle driver 是 haltpoll ，也是代码量很小的
默认状态
- 99.98% 0.00% swapper [kernel.kallsyms] [k] common_startup_64
- common_startup_64
- 96.86% start_secondary
cpu_startup_entry
- do_idle
- 96.86% default_idle_call
default_idle
- 3.13% x86_64_start_kernel
x86_64_start_reservations
start_kernel
__pfx_kernel_init
cpu_startup_entry
do_idle
default_idle_call
default_idle
修改 governor
echo haltpoll > /sys/devices/system/cpu/cpuidle/current_governor
➜ cpuidle cpupower idle-info
CPUidle driver: none
CPUidle governor: haltpoll
analyzing CPU 18:
CPU 18: No idle states
添加启动参数，重启
不过，为什么默认不打开?
if (!kvm_para_available() || !haltpoll_want()){
➜ ~ cpupower idle-info
CPUidle driver: haltpoll
CPUidle governor: haltpoll
analyzing CPU 26:
Number of idle states: 2
Available idle states: POLL haltpoll idle
POLL:
Flags/Description: CPUIDLE CORE POLL IDLE
Latency: 0
Usage: 951
Duration: 50468
haltpoll idle:
Flags/Description: default architecture idle
Latency: 1
Usage: 10032
Duration: 9535233
perf 的结果为:
- 99.98% 0.00% swapper [kernel.kallsyms] [k] common_startup_64
- common_startup_64
- 96.85% start_secondary
cpu_startup_entry
- do_idle
- 96.85% cpuidle_enter
- cpuidle_enter_state
- 93.31% default_enter_idle
default_idle
3.54% poll_idle
- 3.12% x86_64_start_kernel
x86_64_start_reservations
start_kernel
__pfx_kernel_init
cpu_startup_entry
do_idle
cpuidle_enter
- cpuidle_enter_state
- 2.35% default_enter_idle
default_idle
0.78% poll_idle
HALTPOLL_CPUIDLE
config HALTPOLL_CPUIDLE
tristate "Halt poll cpuidle driver"
depends on X86 && KVM_GUEST
default y
help
This option enables halt poll cpuidle driver, which allows to poll
before halting in the guest (more efficient than polling in the
host via halt_poll_ns for some scenarios).
那么 haltpoll 和 poll 有什么区别，如何控制?
Documentation/virt/kvm/halt-polling.rst
基本理解:
- halt polling interval :
kvm_vcpu->halt_poll_ns，guest halt 后，虚拟机还会 halt 多长时间 - total block time : 从 halt 到虚拟机接受中断需要的时间
- global max polling interval :
halt_poll_ns命令行参数，这个是认为的应该 polling 的时间
测试一下，这个说法是真的吗?
Halt polling will only be conducted by the host when no other tasks are runnable on that cpu, otherwise the polling will cease immediately and schedule will be invoked to allow that other task to run. Thus this doesn’t allow a guest to cause denial of service of the cpu.
如果系统中的负载很高，如何保证，kvm 的 polling 不会给其他的业务带来压力，也就是，现在切换到其他的 业务是需要时间的!
lab
debugfs 中的内容也是需要看下:
执行 iperf3 的时候:
kvm statistics - summary
Event Total %Total CurAvg/s
exits 1256164 44.3 71764
halt_exits 520599 18.4 29651
halt_attempted_poll 509684 18.0 29078
halt_successful_poll 508457 17.9 28995
irq_exits 14712 0.5 847
host_state_reload 12149 0.4 659
halt_wakeup 12102 0.4 656
preemption_reported 48 0.0 3
preemption_other 3 0.0 0
halt_poll_success_hist 2 0.0 0
blocking -2 -0.0 0
Total 2833918 161655
[ ] kvm_emulate_halt
halt 到底要不要退出?
在 arch/x86/include/asm/vmx.h 中是定义了这个的:
#define CPU_BASED_HLT_EXITING VMCS_CONTROL_BIT(HLT_EXITING)
最好还是要退出的，不然把物理 CPU 占用了
一个机器上有很多 vCPU 都是退出的状态的。
似乎 handle_fastpath_hlt 和 kvm_emulate_halt 都有被调用
测试发现
当打开 iperf 的时候，可以发现:
Tracing 1 functions for "b'kvm_vcpu_check_block'"... Hit Ctrl-C to end.
FUNC COUNT
kvm_vcpu_check_block 7794616
FUNC COUNT
kvm_vcpu_check_block 7725424
FUNC COUNT
kvm_vcpu_check_block 8343019
而当 guest 没有负载的时候，这个函数的调用次数几乎为 0 的。
- 100.00% 0.00% qemu-system-x86 libc.so.6 [.] __GI___ioctl
- __GI___ioctl
- 98.37% entry_SYSCALL_64_after_hwframe
do_syscall_64
- __x64_sys_ioctl
- 98.37% kvm_vcpu_ioctl
- kvm_arch_vcpu_ioctl_run
- 90.46% kvm_vcpu_halt
- 61.50% kvm_vcpu_check_block
- 47.89% kvm_arch_vcpu_runnable
- 30.90% kvm_cpu_has_interrupt
- 19.17% apic_has_interrupt_for_ppr
- 18.32% vmx_sync_pir_to_irr
8.97% vmx_set_rvi
6.01% kvm_lapic_find_highest_irr
- 8.42% kvm_apic_has_interrupt
6.94% __apic_update_ppr
1.64% kvm_apic_accept_pic_intr
- 10.58% vmx_interrupt_allowed
- 8.84% vmx_interrupt_blocked
0.63% vmx_get_rflags
5.20% __srcu_read_lock
4.82% __srcu_read_unlock
0.74% apic_has_pending_timer
- 18.51% ktime_get
11.46% read_tsc
- 4.73% vmx_vcpu_run
- 1.63% kvm_load_host_xsave_state
native_write_msr
- 1.60% kvm_load_guest_xsave_state
native_write_msr
0.67% handle_fastpath_set_msr_irqoff
1.49% vmx_vmexit
本站所有文章转发 CSDN 将按侵权追究法律责任，其它情况随意。