Linux原始碼版本: 5.3.0
相關資料結構
#define KVM_DEFAULT_PLE_GAP 128 // ple_gap
#define KVM_VMX_DEFAULT_PLE_WINDOW 4096 //ple_window
// ple_window的增大系數,每次呼叫grow_ple_window時,ple_window增大2倍
#define KVM_DEFAULT_PLE_WINDOW_GROW 2
// ple_window的縮小系數
#define KVM_DEFAULT_PLE_WINDOW_SHRINK 0
// ple_window最大不能超過這麼大,該值在32bit和64bit機器上取值不同
#define KVM_VMX_DEFAULT_PLE_WINDOW_MAX UINT_MAX
// ple_window和ple_gap的初始化,前提是該vcpu沒有禁用ple
if (!kvm_pause_in_guest(vmx->vcpu.kvm)) {
vmcs_write32(PLE_GAP, ple_gap);
vmx->ple_window = ple_window;
vmx->ple_window_dirty = true;
}
PAUSE Exit的處理
Intel的cpu上,使用的VMM為kvm時,當guest的vcpu變為busy-waiting狀態,也就是loop-wait狀態,就會在一定情況下觸發vmexit.
觸發條件: 由於kvm中不會使能"PAUSE exiting"feature,因此單一的PAUSE指令不會導致vmexit,kvm中只使用"PAUSE-loop exiting" feature,即迴圈(loop-wait)中的PAUSE指令會導致vmexit,具體情境為:當一個迴圈中的兩次PAUSE之間的時間差不超過PLE_gap常量,且該迴圈中某次PAUSE指令與第一次PAUSE指令的時間差超過了PLE_window,那麼就會產生一個vmexit,觸發原因field會填為PAUSE指令.
kvm程式碼中,如果進入了handle_pause()
函式,說明已經觸發了pause_vmexit.
handle_pause()的大致結構:
其中,grow_ple_window()是為了讓"沒有禁用PLE的guest"調整PLE_window
/*
* Indicate a busy-waiting vcpu in spinlock. We do not enable the PAUSE
* exiting, so only get here on cpu with PAUSE-Loop-Exiting.
*/
static int handle_pause(struct kvm_vcpu *vcpu)
{
if (!kvm_pause_in_guest(vcpu->kvm)) // 1. 如果該vm沒有禁用PLE,則增大PLE_window的值
grow_ple_window(vcpu);
/*
* Intel sdm vol3 ch-25.1.3 says: The "PAUSE-loop exiting"
* VM-execution control is ignored if CPL > 0. OTOH, KVM
* never set PAUSE_EXITING and just set PLE if supported,
* so the vcpu must be CPL=0 if it gets a PAUSE exit.
*/
kvm_vcpu_on_spin(vcpu, true); // 2. 找一個之前被搶佔,目前又可以執行的vcpu,繼續執行spin-loop
// 結束當前vcpu的spin狀態
/*
* 3. 如果guest在debug狀態,則產生了單步中斷,(vcpu_enter_guest)返回0,exit到userspace繼續處理
* 如果guest不在debug狀態,則(vcpu_enter_guest)返回1,無需exit到userspace處理
*/
return kvm_skip_emulated_instruction(vcpu);
}
/*
* 返回值為:當前guest是否禁用PLE feature
* 禁用:返回true
* 沒有禁用: 返回false
*/
static inline bool kvm_pause_in_guest(struct kvm *kvm)
{
return kvm->arch.pause_in_guest;
}
/* 增加PLE_window的值,new_ple_window *= 2 */
static void grow_ple_window(struct kvm_vcpu *vcpu)
{
struct vcpu_vmx *vmx = to_vmx(vcpu);
int old = vmx->ple_window;
vmx->ple_window = __grow_ple_window(old, ple_window,
ple_window_grow,
ple_window_max);
if (vmx->ple_window != old)
vmx->ple_window_dirty = true;
trace_kvm_ple_window_grow(vcpu->vcpu_id, vmx->ple_window, old);
}
kvm_pause_in_guest()
分析過程
不知道這個kvm_pause_in_guest()是什麼意思,但在handle_pause()中可以看到的是,每次發生PAUSE vmexit,都會檢查kvm_pause_in_guest()的返回值,如果返回值為false,則要增大PLE_window的值。
分析1:
思考一下什麼條件下需要增大PLE_window的值呢?
只有在kvm覺得這個guest提前exit了的時候,才需要增大PLE_window,因為再多等一下就可以等到那個鎖了。
結合kvm_pause_in_guest()函式的名字,猜測該函式返回的是PAUSE vmexit前等待的那個鎖是否還沒有開啟,如果沒開啟,返回true,如果開啟了,就返回false.
分析2:
kvm development mail list中對kvm_pause_in_guest()返回的arch.pause_in_guest的說明是:"Allow to disable pause loop exit/pause filtering on a per VM basis. If some VMs have dedicated host CPUs, they won't be negatively affected due to needlessly intercepted PAUSE instructions.", 大意為,允許在特定guest上禁用PLE(intel)/PF(amd). 如果有些guest擁有繫結的host cpu,則不會由於不必要地攔截PAUSE指令而對它們產生負面影響。
什麼意思呢?假設有guestA和guestB,guestA有2個固定vcpu,繫結在host的cpu0,cpu1上,guestB有2個vcpu,不固定host cpu。
當在guestB上的vcpu0上發生spin-loop時,需要vcpu1上的lock,但是vcpu1由於排程原因去做其他事情了,該lock無法處理,guestB只能攔截PAUSE指令,exit到host.
當在guestA上的vcpu0上發生spin-loop時,需要vcpu1上的lock,因為vcpu1固定屬於guestA,不會被排程去做其他事情,相比與guestB,lock的平均解鎖時間肯定小於guestB,所以就沒必要exit到host,spin-wait就行.
結論
綜上所述,kvm_pause_in_guest()返回的是該guest是否禁用了PLE,如果禁用了就返回true,否則false.
該結論的程式碼支援:
// arch/x86/kvm/x86.c
int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
struct kvm_enable_cap *cap)
{
...
case KVM_CAP_X86_DISABLE_EXITS:
...
if (cap->args[0] & KVM_X86_DISABLE_EXITS_PAUSE) // 如果禁用該vm中的pause exit
kvm->arch.pause_in_guest = true; // 該bool值就為true
...
}
kvm_vcpu_on_spin()
首先查詢了mail list中該函式的相關內容,發現了KVM: introduce kvm_vcpu_on_spin,"Introduce kvm_vcpu_on_spin, to be used by VMX/SVM to yield processing once the cpu detects pause-based looping.",直接說明了kvm_vcpu_on_spin()函式的用意,"一旦cpu檢測到pause-loop,就會進行相關操作。"
該函式主要將當前vcpu中的剩餘spin-loop的剩餘任務切換到新的vcpu中執行
// virt/kvm/kvm_main.c
void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode)
{
...
// 將剛剛pause vmexit的vcpu設定為in-spin-loop狀態
kvm_vcpu_set_in_spin_loop(me, true);
// 將當前未執行,但狀態為可執行的vcpu的優先順序提高,因為這樣的vcpu之前被搶佔了,
// 被搶佔之後又在__vcpu_run()中執行了排程函式,所以我們要提高它的優先順序。
// 希望這些vcpu中包含我們需要的鎖,從最後一個被提升優先順序的vcpu開始迴圈試圖切換
for (pass = 0; pass < 2 && !yielded && try; pass++) {
....
}
// 將當前vcpu設定為非spin-loop狀態
kvm_vcpu_set_in_spin_loop(me, false);
// 確保當前vcpu在下一次spin-loop時不被選為exit的vcpu
// 因為只有大家輪流執行spin-loop,效能才能平均且 高
kvm_vcpu_set_dy_eligible(me, false);
}
kvm_skip_emulated_instruction()
該函式主要獲取當前vcpu的RFLAGS暫存器內容,賦值給當前guest的相應資料結構。同時檢查是否需要產生單步中斷.
使Guest的RIP跳過一個指令.
// arch/x86/kvm/x86.c
int kvm_skip_emulated_instruction(struct kvm_vcpu *vcpu)
{
unsigned long rflags = kvm_x86_ops->get_rflags(vcpu);
int r = EMULATE_DONE;
// 更新rip值,確保guest interruptibiliy state的最後2bit為0,即STI和MOV SS為0,即不接收中斷
kvm_x86_ops->skip_emulated_instruction(vcpu);
/*
* rflags is the old, "raw" value of the flags. The new value has
* not been saved yet.
*
* This is correct even for TF set by the guest, because "the
* processor will not generate this exception after the instruction
* that sets the TF flag".
*/
if (unlikely(rflags & X86_EFLAGS_TF)) //如果guest處於debug狀態,就會產生單步中斷,那麼就會將r置為1
kvm_vcpu_do_singlestep(vcpu, &r);
return r == EMULATE_DONE; // 如果產生單步中斷,則需要exit到VMM中處理
}
static void skip_emulated_instruction(struct kvm_vcpu *vcpu)
{
// 獲取RIP暫存器應該跳躍的值,然後將新的RIP值更新到之前vmexit的vcpu暫存器中
unsigned long rip;
rip = kvm_rip_read(vcpu);
rip += vmcs_read32(VM_EXIT_INSTRUCTION_LEN);
kvm_rip_write(vcpu, rip);
/* skipping an emulated instruction also counts */
vmx_set_interrupt_shadow(vcpu, 0);
}
/* 該函式的本意為:
* 如果在vmexit期間,該vm的GUEST_INTERRUPTIBILITY_INFO發生變化,那麼就將變化寫入vmcs.
* 但在以上skip_emulated_instruction()中,呼叫了vmx_set_interrupt_shadow(vcpu, 0); mask為0時,
* vmx_set_interrupt_shadow的只是在確定GUEST_INTERRUPTIBILITY_INFO的最後2bit,
* 即GUEST_INTR_STATE_STI和GUEST_INTR_STATE_MOV_SS是否一直為0,而這2個bit為除錯使用的狀態
*/
void vmx_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
{
u32 interruptibility_old = vmcs_read32(GUEST_INTERRUPTIBILITY_INFO);
u32 interruptibility = interruptibility_old;
interruptibility &= ~(GUEST_INTR_STATE_STI | GUEST_INTR_STATE_MOV_SS);
if (mask & KVM_X86_SHADOW_INT_MOV_SS)
interruptibility |= GUEST_INTR_STATE_MOV_SS;
else if (mask & KVM_X86_SHADOW_INT_STI)
interruptibility |= GUEST_INTR_STATE_STI;
if ((interruptibility != interruptibility_old))
vmcs_write32(GUEST_INTERRUPTIBILITY_INFO, interruptibility);
}
/* 單步中斷的賦值操作 */
static void kvm_vcpu_do_singlestep(struct kvm_vcpu *vcpu, int *r)
{
struct kvm_run *kvm_run = vcpu->run;
if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP) {
kvm_run->debug.arch.dr6 = DR6_BS | DR6_FIXED_1 | DR6_RTM;
kvm_run->debug.arch.pc = vcpu->arch.singlestep_rip;
kvm_run->debug.arch.exception = DB_VECTOR;
kvm_run->exit_reason = KVM_EXIT_DEBUG;
*r = EMULATE_USER_EXIT;
} else {
kvm_queue_exception_p(vcpu, DB_VECTOR, DR6_BS);
}
}