Also clear all accessed bit in pagetables when the VM do wbinvd. Then
the next wbinvd will benefit from it.
Tracked-On: #4703
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
This is a preparation patch for adding support for VT-d PI
related vCPU scheduling.
ACRN does not support vCPU migration, one vCPU always runs on
the same pCPU, so PI's ndst is never changed after startup.
VCPUs of a VM won’t share same pCPU. So the maximum possible number
of VCPUs that can run on a pCPU is CONFIG_MAX_VM_NUM.
Allocate unique Activation Notification Vectors (ANV) for each vCPU
that belongs to the same pCPU, the ANVs need only be unique within each
pCPU, not across all vCPUs. This reduces # of pre-allocated ANVs for
posted interrupts to CONFIG_MAX_VM_NUM, and enables ACRN to avoid
switching between active and wake-up vector values in the posted
interrupt descriptor on vCPU scheduling state changes.
A total of CONFIG_MAX_VM_NUM consecutive IRQs/vectors are reserved
for posted interrupts use.
The code first initializes vcpu->arch.pid.control.bits.nv dynamically
(will be added in subsequent patch), the other code shall use
vcpu->arch.pid.control.bits.nv instead of the hard-coded notification vectors.
Rename some functions:
apicv_post_intr --> apicv_trigger_pi_anv
posted_intr_notification --> handle_pi_notification
setup_posted_intr_notification --> setup_pi_notification
Tracked-On: #4506
Signed-off-by: dongshen <dongsheng.x.zhang@intel.com>
Reviewed-by: Eddie Dong <eddie.dong@Intel.com>
The posted interrupt descriptor is more of a vmx/vmcs concept than a vlapic
concept. struct acrn_vcpu_arch stores the vmx/vmcs info, so put struct pi_desc
in struct acrn_vcpu_arch.
Remove the function apicv_get_pir_desc_paddr()
A few coding style/typo fixes
Tracked-On: #4506
Signed-off-by: dongshen <dongsheng.x.zhang@intel.com>
Reviewed-by: Eddie Dong <eddie.dong@Intel.com>
1. Align the coding style for these MACROs
2. Align the values of fixed VECTORs
Tracked-On: #4348
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
In lapic passthrough mode, it should passthrough HLT/PAUSE execution
too. This patch disable their emulation when switch to lapic passthrough mode.
Tracked-On: #4329
Tested-by: Dongsheng Zhang <dongsheng.x.zhang@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
HLT emulation is import to CPU resource maximum utilization. vcpu
doing HLT means it is idle and can give up CPU proactively. Thus, we
pause the vcpu thread in HLT emulation and resume it while event happens.
When vcpu enter HLT, its vcpu thread will sleep, but the vcpu state is
still 'Running'.
VM ID PCPU ID VCPU ID VCPU ROLE VCPU STATE
===== ======= ======= ========= ==========
0 0 0 PRIMARY Running
0 1 1 SECONDARY Running
Tracked-On: #4329
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
As we enabled cpu sharing, PAUSE-loop exiting can help vcpu
to release its pcpu proactively. It's good for performance.
VMX_PLE_GAP: upper bound on the amount of time between two successive
executions of PAUSE in a loop.
VMX_PLE_WINDOW: upper bound on the amount of time a guest is allowed to
execute in a PAUSE loop
Tracked-On: #4329
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
According to SDM Vol.3 Section 25.3, behavior of the INVPCID
instruction is determined first by the setting of the “enable
INVPCID” VM-execution control:
- If the “enable INVPCID” VM-execution control is 0, INVPCID
causes an invalid-opcode exception (#UD).
- If the “enable INVPCID” VM-execution control is 1, treatment
is based on the setting of the “INVLPG exiting” VM-execution
control:
* If the “INVLPG exiting” VM-execution control is 0, INVPCID
operates normally.
* If the “INVLPG exiting” VM-execution control is 1, INVPCID
causes a VM exit.
In current implementation, hypervisor doesn't set “INVLPG exiting”
VM-execution control, this patch sets “enable INVPCID” VM-execution
control to 1 when the instruction is supported by physical cpu.
If INVPCID is supported by physical cpu, INVPCID will not cause VM
exit in VM.
If INVPCID is not supported by physical cpu, INVPCID causes an #UD
in VM.
When INVPCID is passed-through to VM, According to SDM Vol.3 28.3.3.1,
INVPCID instruction invalidates linear mappings and combined mappings.
They are required to do so only for the current VPID.
HV assigned a unique vpid for each vCPU, if guest uses wrong PCID,
it would not affect other vCPUs.
Tracked-On: #4296
Signed-off-by: Binbin Wu <binbin.wu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
ACRN relies on the capability of VPID to avoid EPT flushes during VMX transitions.
This capability is checked as a must have hardware capability, otherwise, ACRN will
refuse to boot.
Also, the current code has already made sure each vpid for a virtual cpu is valid.
So, no need to check the validity of vpid for vcpu and enable VPID for vCPU by default.
Tracked-On: #4296
Signed-off-by: Binbin Wu <binbin.wu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
We have implemented a new notification method using NMI.
So replace the INIT notification method with the NMI one.
Then we can remove INIT notification related code later.
Tracked-On: #3886
Signed-off-by: Kaige Fu <kaige.fu@intel.com>
There is a window where we may miss the current request in the
notification period when the work flow is as the following:
CPUx + + CPUr
| |
| +--+
| | | Handle pending req
| <--+
+--+ |
| | Set req flag |
<--+ |
+------------------>---+
| Send NMI | | Handle NMI
| <--+
| |
| |
| +--> vCPU enter
| |
+ +
So, this patch enables the NMI-window exiting to trigger the next vmexit
once there is no "virtual-NMI blocking" after vCPU enter into VMX non-root
mode. Then we can process the pending request on time.
Tracked-On: #3886
Acked-by: Eddie Dong <eddie.dong@intel.com>
Signed-off-by: Kaige Fu <kaige.fu@intel.com>
ACRN hypervisor needs to kick vCPU off VMX non-root mode to do some
operations in hypervisor, such as interrupt/exception injection, EPT
flush etc. For non lapic-pt vCPUs, we can use IPI to do so. But, it
doesn't work for lapic-pt vCPUs as the IPI will be injected to VMs
directly without vmexit.
Without the way to kick the vCPU off VMX non-root mode to handle pending
request on time, there may be fatal errors triggered.
1). Certain operation may not be carried out on time which may further
lead to fatal errors. Taking the EPT flush request as an example, once we
don't flush the EPT on time and the guest access the out-of-date EPT,
fatal error happens.
2). ACRN now will send an IPI with vector 0xF0 to target vCPU to kick the vCPU
off VMX non-root mode if it wants to do some operations on target vCPU.
However, this way doesn't work for lapic-pt vCPUs. The IPI will be delivered
to the guest directly without vmexit and the guest will receive a unexpected
interrupt. Consequently, if the guest can't handle this interrupt properly,
fatal error may happen.
The NMI can be used as the notification signal to kick the vCPU off VMX
non-root mode for lapic-pt vCPUs. So, this patch uses NMI as notification signal
to address the above issues for lapic-pt vCPUs.
Tracked-On: #3886
Acked-by: Eddie Dong <eddie.dong@intel.com>
Signed-off-by: Kaige Fu <kaige.fu@intel.com>
With cpu-sharing enabled, there are more than 1 vcpu on 1 pcpu, so the
smp_call handler should switch the vmcs to the target vcpu's vmcs. Then
get the info.
dump_vcpu_reg and dump_guest_mem should run on certain vmcs, otherwise,
there will be #GP error.
Renaming:
vcpu_dumpreg -> dump_vcpu_reg
switch_vmcs -> load_vmcs
Tracked-On: #4178
Signed-off-by: Conghui Chen <conghui.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
PMU is hidden from any guest, UD is expected when guest
try to execute 'rdpmc' instruction.
this patch sets 'RDPMC exiting' in Processorbased
VM-execution control.
Tracked-On: #3453
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
After changing init_vmcs to smp call approach and do it before
launch_vcpu, it could work with noop scheduler. On real sharing
scheudler, it has problem.
pcpu0 pcpu1 pcpu1
vmBvcpu0 vmAvcpu1 vmBvcpu1
vmentry
init_vmcs(vmBvcpu1) vmexit->do_init_vmcs
corrupt current vmcs
vmentry fail
launch_vcpu(vmBvcpu1)
This patch mark a event flag when request vmcs init for specific vcpu. When
it is running and checking pending events, will do init_vmcs firstly.
Tracked-On: #4178
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
We should use INIT signal to notify the vcpu threads when
powering off the hard RTVM. To achive this, we should set
the vcpu->thread_obj.notify_mode as SCHED_NOTIFY_INIT.
Patch (27163df9 hv: sched: add sleep/wake for thread object)
tries to set the notify_mode according `is_lapic_pt_enabled(vcpu)`
in function prepare_vcpu. But at this point, the is_lapic_pt_enabled(vcpu)
will always return false. Consequently, it will set notify_mode
as SCHED_NOTIFY_IPI. Then leads to the failure of powering off
hard RTVM.
This patch fixes it by:
- Initialize the notify_mode as SCHED_NOTIFY_IPI in prepare_vcpu.
- Set notify_mode as SCHED_NOTIFY_INIT after guest is trying to
enable x2apic mode of passthru lapic.
Tracked-On: #3975
Reviewed-by: Yin Fengwei <fengwei.yin@intel.com>
Reviewed-by: Yan, Like <like.yan@intel.com>
Signed-off-by: Kaige Fu <kaige.fu@intel.com>
To support cpu sharing, multiple vcpu can run on same pcpu. We need do
necessary vcpu context switch. This patch add below actions in context
switch.
1) fxsave/fxrstor;
2) save/restore MSRs: MSR_IA32_STAR, MSR_IA32_LSTAR,
MSR_IA32_FMASK, MSR_IA32_KERNEL_GS_BASE;
3) switch vmcs.
Tracked-On: #3813
Signed-off-by: Jason Chen CJ <jason.cj.chen@intel.com>
Signed-off-by: Yu Wang <yu1.wang@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
With cpu sharing enabled, we will map acrn_vcpu to thread_object
in scheduling. From modulization perspective, we'd better hide the
pcpu_id in acrn_vcpu and move it to thread_object.
Tracked-On: #3813
Signed-off-by: Jason Chen CJ <jason.cj.chen@intel.com>
Signed-off-by: Yu Wang <yu1.wang@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
vcpu thread's stack shouldn't follow reset_vcpu to reset.
There is also a bug here:
while vcpu B thread set vcpu->running to false, other vcpu A thread
will treat the vcpu B is paused while it has not been switch out
completely, then reset_vcpu will reset the vcpu B thread's stack and
corrupt its running context.
This patch will remove the vcpu thread's stack reset from reset_vcpu.
With the change, we need do init_vmcs between vcpu startup address be
settled and scheduled in. And switch_to_idle() is not needed anymore
as S3 thread's stack will not be reset.
Tracked-On: #3813
Signed-off-by: Fengwei Yin <fengwei.yin@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
--remove unnecessary includes
--remove unnecssary forward-declaration for 'struct vhm_request'
Tracked-On: #861
Signed-off-by: Mingqiang Chi <mingqiang.chi@intel.com>
Now the structures(run_context & ext_context) are defined
in vcpu.h,and they are used in the lower-layer modules(wakeup.S),
this patch move down the structures from vcpu.h to cpu.h
to avoid reversed dependency.
Tracked-On: #1842
Signed-off-by: Mingqiang Chi <mingqiang.chi@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Currently, the clos id of the cpu cores in vmx root mode is the same as non-root mode.
For RTVM, if hypervisor share the same clos id with non-root mode, the cacheline may
be polluted due to the hypervisor code execution when vmexit.
The patch adds hv_clos in vm_configurations.c
Hypervisor initializes clos setting according to hv_clos during physical cpu cores initialization.
For RTVM, MSR auto load/store areas are used to switch different settings for VMX root/non-root
mode for RTVM.
Tracked-On: #2462
Signed-off-by: Binbin Wu <binbin.wu@intel.com>
Reviewed-by: Eddie Dong <eddie.dong@intel.com>
This patch moves vmx_rdmsr_pat/vmx_wrmsr_pat from vmcs.c to vmsr.c,
so that these two functions would become internal functions inside
vmsr.c.
This approach improves the modularity.
v1 -> v2:
* remove 'vmx_rdmsr_pat'
* rename 'vmx_wrmsr_pat' with 'write_pat_msr'
Tracked-On: #1842
Signed-off-by: Shiqing Gao <shiqing.gao@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
The policy of vART is that software in native can run in
VM too. And in native side, the relationship between the
ART hardware and TSC is:
pTSC = (pART * M) / N + pAdjust
The vART solution is:
- Present the ART capability to guest through CPUID leaf
15H for M/N which identical to the physical values.
- PT devices see the pART (vART = pART).
- Guest expect: vTSC = vART * M / N + vAdjust.
- VMCS.OFFSET = vTSC - pTSC = vAdjust - pAdjust.
So to support vART, we should do the following:
1. if vAdjust and vTSC are changed by guest, we should change
VMCS.OFFSET accordingly.
2. Make the assumption that the pAjust is never touched by ACRN.
For #1, commit "a958fea hv: emulate IA32_TSC_ADJUST MSR" has implementation
it. And for #2, acrn never touch pAdjust.
--
v2 -> v3:
- Add comment when handle guest TSC_ADJUST and TSC accessing.
- Initialize the VMCS.OFFSET = vAdjust - pAdjust.
v1 -> v2
Refine commit message to describe the whole vART solution.
Tracked-On: #3501
Signed-off-by: Kaige Fu <kaige.fu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
ACRN supports LAPIC emulation for guests using x86 APICv. When guest OS/BIOS
switches from xAPIC to x2APIC mode of operation, ACRN also supports switching
froom LAPIC emulation to LAPIC passthrough to guest. User/developer needs to
configure GUEST_FLAG_LAPIC_PASSTHROUGH for guest_flags in the corresponding
VM's config for ACRN to enable LAPIC passthrough.
This patch does the following
1)Fixes a bug in the abovementioned feature. For a guest that is
configured with GUEST_FLAG_LAPIC_PASSTHROUGH, during the time period guest is
using xAPIC mode of LAPIC, virtual interrupts are not delivered. This can be
manifested as guest hang when it does not receive virtual timer interrupts.
2)ACRN exposes physical topology via CPUID leaf 0xb to LAPIC PT VMs. This patch
removes that condition and exposes virtual topology via CPUID leaf 0xb.
Tracked-On: #3136
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Reviewed-by: Eddie Dong <eddie.dong@intel.com>
Add MSR_IA32_MISC_ENABLE to emulated_guest_msrs to enable the emulation.
Init MSR_IA32_MISC_ENABLE for guest.
Tracked-On: #2834
Signed-off-by: Binbin Wu <binbin.wu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
This patch adds prefix 'p' before 'cpu' to physical cpu related functions.
And there is no code logic change.
Tracked-On: #2991
Signed-off-by: Kaige Fu <kaige.fu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
1) Shouldn't try to set APIC-register virtualization if the physical doesn't
support APICV advanced mode.
2) Remove all APICv features VMCS setting when LAPIC is passed through to guest.
Tracked-On: #1842
Signed-off-by: Li, Fei1 <fei1.li@intel.com>
Add two functions to combine constraint for APICv:
is_apicv_basic_feature_supported: check the physical platform whether support
"Use TPR shadow", "Virtualize APIC accesses" and "Virtualize x2APIC mode"
is_apicv_advanced_feature_supported: check the physical platform whether support
"APIC-register virtualization", "Virtual-interrupt delivery" and
"Process posted interrupts".
If the physical platform only support APICv basic feature, enable "Use TPR shadow"
and "Virtualize APIC accesses" for xAPIC mode; enable "Use TPR shadow" and
"Virtualize x2APIC mode" for x2APIC. Otherwise, if the physical platform support
APICv advanced feature, enable APICv feature for xAPIC mode and x2APIC mode.
Tracked-On: #1842
Signed-off-by: Li, Fei1 <fei1.li@intel.com>
According to SDM 24.7.2, these two MSRs should be
configured with physical address.
Tracked-On: #861
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
With lapic_pt based on vlapic, guest always see vitual apic_id.
We need to convert the virtual apic_id from guest to physical apic_id
before writing ICR.
SMP for VM with lapic_pt is supported with this fix.
Tracked-On: #2351
Signed-off-by: Yan, Like <like.yan@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
