.. _hv-cpu-virt: CPU Virtualization ################## .. figure:: images/hld-image47.png :align: center :name: hv-cpu-virt-components ACRN Hypervisor CPU Virtualization Components The following sections discuss the major modules (shown in blue) in the CPU virtualization overview shown in :numref:`hv-cpu-virt-components`. Based on Intel VT-x virtualization technology, ACRN emulates a virtual CPU (vCPU) with the following methods: - **core partition**: one vCPU is dedicated and associated with one physical CPU (pCPU), making much of hardware register emulation simply pass-through and provides good isolation for physical interrupt and guest execution. (See `Static CPU Partitioning`_ for more information.) - **core sharing** (to be added): two or more vCPUs are sharing one physical CPU (pCPU); a more complicated context switch is needed between different vCPUs' switching, and provides flexible computing resources sharing for low performance demand vCPU tasks. (See `Flexible CPU Sharing`_ for more information.) - **simple schedule**: a well-designed scheduler framework allows ACRN to adopt different scheduling policy, for example - noop & round-robin: noop scheduler - only two thread loops are maintained for a CPU - vCPU thread and default idle thread. A CPU runs most of the time in the vCPU thread for emulating a guest CPU, switching between VMX root mode and non-root mode. A CPU schedules out to default idle when an operation needs it to stay in VMX root mode, such as when waiting for an I/O request from DM or ready to destroy. round-robin scheduler (to be added) - allow more vcpu thread loops running on a CPU. A CPU switches among different vCPU thread and default idle thread, upon running out corresponding timeslice or necessary scheduling out such as waiting for an I/O request. A vCPU could yield itself as well, for example when it executes "PAUSE" instruction. Static CPU Partitioning *********************** CPU partitioning is a policy for mapping a virtual CPU (VCPU) to a physical CPU. To enable this, the ACRN hypervisor could configure "noop scheduler" as the schedule policy for this physical CPU. ACRN then forces a fixed 1:1 mapping between a VCPU and this physical CPU when creating a VCPU for the guest Operating System. This makes the VCPU management code much simpler. ``vcpu_affinity`` in ``vm config`` help to decide which physical CPU a VCPU in a VM affine to, then finalize the fixed mapping. Flexible CPU Sharing ******************** This is a TODO feature. To enable CPU sharing, the ACRN hypervisor could configure "round-robin scheduler" as the schedule policy for corresponding physical CPU. ``vcpu_affinity`` in ``vm config`` help to decide which physical CPU two or more VCPUs from different VMs are sharing. CPU management in the Service VM under static CPU partitioning ============================================================== With ACRN, all ACPI table entries are pass-thru to the Service VM, including the Multiple Interrupt Controller Table (MADT). The Service VM sees all physical CPUs by parsing the MADT when the Service VM kernel boots. All physical CPUs are initially assigned to the Service VM by creating the same number of virtual CPUs. When the Service VM boot is finished, it releases the physical CPUs intended for User VM use. Here is an example flow of CPU allocation on a multi-core platform. .. figure:: images/static-core-image2.png :width: 600px :align: center :name: static-core-cpu-allocation CPU allocation on a multi-core platform CPU management in the Service VM under flexing CPU sharing ========================================================== As all Service VM CPUs could share with different User VMs, ACRN can still pass-thru MADT to Service VM, and the Service VM is still able to see all physical CPUs. But as under CPU sharing, the Service VM does not need offline/release the physical CPUs intended for User VM use. CPU management in User VM ========================= From the User VM point of view, CPU management is very simple - when DM does hypercalls to create VMs, the hypervisor will create its virtual CPUs based on the configuration in this User VM's ``vm config``. As mentioned in previous description, ``vcpu_affinity`` in ``vm config`` tells which physical CPUs a VM's VCPU will use, and the scheduler policy associated with corresponding physical CPU decide this VCPU will run in partition or sharing mode. CPU assignment management in HV =============================== The physical CPU assignment is pre-defined by ``vcpu_affinity`` in ``vm config``, necessary sanitize check should be done to ensure - in one VM, each VCPU will have only one prefer physical CPU - in one VM, its VCPUs will not share same physical CPU - in one VM, if a VCPU is using "noop scheduler", corresponding physical CPU will not be shared with any other VM's VCPU Currently, the ACRN hypervisor does not support virtual CPU migration to different physical CPUs. This means no changes to the virtual CPU to physical CPU can happen without first calling offline_vcpu. .. _vCPU_lifecycle: vCPU Lifecycle ************** A vCPU lifecycle is shown in :numref:`hv-vcpu-transitions` below, where the major states are: - **VCPU_INIT**: vCPU is in an initialized state, and its vCPU thread is not ready to run on its associated CPU - **VCPU_RUNNING**: vCPU is running, and its vCPU thread is ready (in the queue) or running on its associated CPU - **VCPU_PAUSED**: vCPU is paused, and its vCPU thread is not running on its associated CPU - **VPCU_ZOMBIE**: vCPU is being offline, and its vCPU thread is not running on its associated CPU - **VPCU_OFFLINE**: vCPU is offline .. figure:: images/hld-image17.png :align: center :name: hv-vcpu-transitions ACRN vCPU state transitions Following functions are used to drive the state machine of the vCPU lifecycle: .. doxygenfunction:: create_vcpu :project: Project ACRN .. doxygenfunction:: pause_vcpu :project: Project ACRN .. doxygenfunction:: resume_vcpu :project: Project ACRN .. doxygenfunction:: reset_vcpu :project: Project ACRN .. doxygenfunction:: offline_vcpu :project: Project ACRN vCPU Scheduling under static CPU partitioning ********************************************* .. figure:: images/hld-image35.png :align: center :name: hv-vcpu-schedule ACRN vCPU scheduling flow under static CPU partitioning As describes in the CPU virtualization overview, if under static CPU partitioning, ACRN implements a simple scheduling mechanism based on two threads: vcpu_thread and default_idle. A vCPU with VCPU_RUNNING state always runs in a vcpu_thread loop, meanwhile a vCPU with VCPU_PAUSED or VCPU_ZOMBIE state runs in default_idle loop. The detail behaviors in vcpu_thread and default_idle threads are illustrated in :numref:`hv-vcpu-schedule`: - The **vcpu_thread** loop will do the loop of handling vm exits, and pending requests around the VM entry/exit. It will also check the reschedule request then schedule out to default_idle if necessary. See `vCPU Thread`_ for more details of vcpu_thread. - The **default_idle** loop simply does do_cpu_idle while also checking for need-offline and reschedule requests. If a CPU is marked as need-offline, it will go to cpu_dead. If a reschedule request is made for this CPU, it will schedule out to vcpu_thread if necessary. - The function ``make_reschedule_request`` drives the thread switch between vcpu_thread and default_idle. Some example scenario flows are shown here: .. figure:: images/hld-image7.png :align: center ACRN vCPU scheduling scenarios - **During starting a VM**: after create a vCPU, BSP calls *launch_vcpu* through *start_vm*, AP calls *launch_vcpu* through vlapic INIT-SIPI emulation, finally this vCPU runs in a *vcpu_thread* loop. - **During shutting down a VM**: *pause_vm* function call makes a vCPU running in *vcpu_thread* to schedule out to *default_idle*. The following *reset_vcpu* and *offline_vcpu* de-init and then offline this vCPU instance. - **During IOReq handling**: after an IOReq is sent to DM for emulation, a vCPU running in *vcpu_thread* schedules out to *default_idle* through *acrn_insert_request_wait->pause_vcpu*. After DM complete the emulation for this IOReq, it calls *hcall_notify_ioreq_finish->resume_vcpu* and makes the vCPU schedule back to *vcpu_thread* to continue its guest execution. vCPU Scheduling under flexible CPU sharing ****************************************** To be added. vCPU Thread *********** The vCPU thread flow is a loop as shown and described below: .. figure:: images/hld-image68.png :align: center ACRN vCPU thread 1. Check if *vcpu_thread* needs to schedule out to *default_idle* or other *vcpu_thread* by reschedule request. If needed, then schedule out to *default_idle* or other *vcpu_thread*. 2. Handle pending request by calling *acrn_handle_pending_request*. (See `Pending Request Handlers`_.) 3. VM Enter by calling *start/run_vcpu*, then enter non-root mode to do guest execution. 4. VM Exit from *start/run_vcpu* when guest trigger vm exit reason in non-root mode. 5. Handle vm exit based on specific reason. 6. Loop back to step 1. vCPU Run Context ================ During a vCPU switch between root and non-root mode, the run context of the vCPU is saved and restored using this structure: .. doxygenstruct:: run_context :project: Project ACRN The vCPU handles runtime context saving by three different categories: - Always save/restore during vm exit/entry: - These registers must be saved every time vm exit, and restored every time vm entry - Registers include: general purpose registers, CR2, and IA32_SPEC_CTRL - Definition in *vcpu->run_context* - Get/Set them through *vcpu_get/set_xxx* - On-demand cache/update during vm exit/entry: - These registers are used frequently. They should be cached from VMCS on first time access after a VM exit, and updated to VMCS on VM entry if marked dirty - Registers include: RSP, RIP, EFER, RFLAGS, CR0, and CR4 - Definition in *vcpu->run_context* - Get/Set them through *vcpu_get/set_xxx* - Always read/write from/to VMCS: - These registers are rarely used. Access to them is always from/to VMCS. - Registers are in VMCS but not list in the two cases above. - No definition in *vcpu->run_context* - Get/Set them through VMCS API For the first two categories above, ACRN provides these get/set APIs: .. doxygenfunction:: vcpu_get_gpreg :project: Project ACRN .. doxygenfunction:: vcpu_set_gpreg :project: Project ACRN .. doxygenfunction:: vcpu_get_rip :project: Project ACRN .. doxygenfunction:: vcpu_set_rip :project: Project ACRN .. doxygenfunction:: vcpu_get_rsp :project: Project ACRN .. doxygenfunction:: vcpu_set_rsp :project: Project ACRN .. doxygenfunction:: vcpu_get_efer :project: Project ACRN .. doxygenfunction:: vcpu_set_efer :project: Project ACRN .. doxygenfunction:: vcpu_get_rflags :project: Project ACRN .. doxygenfunction:: vcpu_set_rflags :project: Project ACRN .. doxygenfunction:: vcpu_get_cr0 :project: Project ACRN .. doxygenfunction:: vcpu_set_cr0 :project: Project ACRN .. doxygenfunction:: vcpu_get_cr2 :project: Project ACRN .. doxygenfunction:: vcpu_set_cr2 :project: Project ACRN .. doxygenfunction:: vcpu_get_cr4 :project: Project ACRN .. doxygenfunction:: vcpu_set_cr4 :project: Project ACRN VM Exit Handlers ================ ACRN implements its VM exit handlers with a static table. Except for the exit reasons listed below, a default *unhandled_vmexit_handler* is used that will trigger an error message and return without handling: .. list-table:: :widths: 33 33 33 :header-rows: 1 * - **VM Exit Reason** - **Handler** - **Desc** * - VMX_EXIT_REASON_EXCEPTION_OR_NMI - exception_vmexit_handler - Only trap #MC, print error then inject back to guest * - VMX_EXIT_REASON_EXTERNAL_INTERRUPT - external_interrupt_vmexit_handler - External interrupt handler for physical interrupt happening in non-root mode * - VMX_EXIT_REASON_TRIPLE_FAULT - triple_fault_vmexit_handler - Handle triple fault from vcpu * - VMX_EXIT_REASON_INIT_SIGNAL - init_signal_vmexit_handler - Handle INIT signal from vcpu * - VMX_EXIT_REASON_INTERRUPT_WINDOW - interrupt_window_vmexit_handler - To support interrupt window if VID is disabled * - VMX_EXIT_REASON_CPUID - cpuid_vmexit_handler - Handle CPUID access from guest * - VMX_EXIT_REASON_VMCALL - vmcall_vmexit_handler - Handle hypercall from guest * - VMX_EXIT_REASON_CR_ACCESS - cr_access_vmexit_handler - Handle CR registers access from guest * - VMX_EXIT_REASON_IO_INSTRUCTION - pio_instr_vmexit_handler - Emulate I/O access with range in IO_BITMAP, which may have a handler in hypervisor (such as vuart or vpic), or need to create an I/O request to DM * - VMX_EXIT_REASON_RDMSR - rdmsr_vmexit_handler - Read MSR from guest in MSR_BITMAP * - VMX_EXIT_REASON_WRMSR - wrmsr_vmexit_handler - Write MSR from guest in MSR_BITMAP * - VMX_EXIT_REASON_APIC_ACCESS - apic_access_vmexit_handler - APIC access for APICv * - VMX_EXIT_REASON_VIRTUALIZED_EOI - veoi_vmexit_handler - Trap vLAPIC EOI for specific vector with level trigger mode in vIOAPIC, required for supporting PTdev * - VMX_EXIT_REASON_EPT_VIOLATION - ept_violation_vmexit_handler - MMIO emulation, which may have handler in hypervisor (such as vLAPIC or vIOAPIC), or need to create an I/O request to DM * - VMX_EXIT_REASON_XSETBV - xsetbv_vmexit_handler - Set host owned XCR0 for supporting xsave * - VMX_EXIT_REASON_APIC_WRITE - apic_write_vmexit_handler - APIC write for APICv Details of each vm exit reason handler are described in other sections. .. _pending-request-handlers: Pending Request Handlers ======================== ACRN uses the function *acrn_handle_pending_request* to handle requests before VM entry in *vcpu_thread*. A bitmap in the vCPU structure lists the different requests: .. code-block:: c #define ACRN_REQUEST_EXCP 0U #define ACRN_REQUEST_EVENT 1U #define ACRN_REQUEST_EXTINT 2U #define ACRN_REQUEST_NMI 3U #define ACRN_REQUEST_EOI_EXIT_BITMAP_UPDATE 4U #define ACRN_REQUEST_EPT_FLUSH 5U #define ACRN_REQUEST_TRP_FAULT 6U #define ACRN_REQUEST_VPID_FLUSH 7U /* flush vpid tlb */ ACRN provides the function *vcpu_make_request* to make different requests, set the bitmap of the corresponding request, and notify the target vCPU through the IPI if necessary (when the target vCPU is not currently running). See :ref:`vcpu-request-interrupt-injection` for details. .. code-block:: c void vcpu_make_request(struct vcpu *vcpu, uint16_t eventid) { uint16_t pcpu_id = pcpuid_from_vcpu(vcpu); bitmap_set_lock(eventid, &vcpu->arch_vcpu.pending_req); /* * if current hostcpu is not the target vcpu's hostcpu, we need * to invoke IPI to wake up target vcpu * * TODO: Here we just compare with cpuid, since cpuid currently is * global under pCPU / vCPU 1:1 mapping. If later we enabled vcpu * scheduling, we need change here to determine it target vcpu is * VMX non-root or root mode */ if (get_cpu_id() != pcpu_id) { send_single_ipi(pcpu_id, VECTOR_NOTIFY_VCPU); } } For each request, function *acrn_handle_pending_request* handles each request as shown below. .. list-table:: :widths: 25 25 25 25 :header-rows: 1 * - **Request** - **Desc** - **Request Maker** - **Request Handler** * - ACRN_REQUEST_EXCP - Request for exception injection - vcpu_inject_gp, vcpu_inject_pf, vcpu_inject_ud, vcpu_inject_ac, or vcpu_inject_ss and then queue corresponding exception by vcpu_queue_exception - vcpu_inject_hi_exception, vcpu_inject_lo_exception based on exception priority * - ACRN_REQUEST_EVENT - Request for vlapic interrupt vector injection - vlapic_fire_lvt or vlapic_set_intr, which could be triggered by vlapic lvt, vioapic, or vmsi - vcpu_do_pending_event * - ACRN_REQUEST_EXTINT - Request for extint vector injection - vcpu_inject_extint, triggered by vpic - vcpu_do_pending_extint * - ACRN_REQUEST_NMI - Request for nmi injection - vcpu_inject_nmi - program VMX_ENTRY_INT_INFO_FIELD directly * - ACRN_REQUEST_EOI_EXIT_BITMAP_UPDATE - Request for update VEOI bitmap update for level triggered vector - vlapic_reset_tmr or vlapic_set_tmr change trigger mode in RTC - vcpu_set_vmcs_eoi_exit * - ACRN_REQUEST_EPT_FLUSH - Request for EPT flush - ept_add_mr, ept_modify_mr, ept_del_mr, or vmx_write_cr0 disable cache - invept * - ACRN_REQUEST_TRP_FAULT - Request for handling triple fault - vcpu_queue_exception meet triple fault - fatal error * - ACRN_REQUEST_VPID_FLUSH - Request for VPID flush - None - flush_vpid_single .. note:: Refer to the interrupt management chapter for request handling order for exception, nmi, and interrupts. For other requests such as tmr update, or EPT flush, there is no mandatory order. VMX Initialization ****************** ACRN will attempt to initialize the vCPU's VMCS before its first launch with the host state, execution control, guest state, entry control and exit control, as shown in the table below. The table briefly shows how each field got configured. The guest state field is critical for a guest CPU start to run based on different CPU modes. For a guest vCPU's state initialization: - If it's BSP, the guest state configuration is done in SW load, which could be initialized by different objects: - The Service VM BSP: hypervisor will do context initialization in different SW load based on different boot mode - User VM BSP: DM context initialization through hypercall - If it's AP, then it will always start from real mode, and the start vector will always come from vlapic INIT-SIPI emulation. .. doxygenstruct:: acrn_vcpu_regs :project: Project ACRN .. list-table:: :widths: 20 40 10 30 :header-rows: 1 * - **VMX Domain** - **Fields** - **Bits** - **Description** * - **host state** - CS, DS, ES, FS, GS, TR, LDTR, GDTR, IDTR - n/a - According to host * - - MSR_IA32_PAT, MSR_IA32_EFER - n/a - According to host * - - CR0, CR3, CR4 - n/a - According to host * - - RIP - n/a - Set to vm_exit pointer * - - IA32_SYSENTER_CS/ESP/EIP - n/a - Set to 0 * - **exec control** - VMX_PIN_VM_EXEC_CONTROLS - 0 - Enable external-interrupt exiting * - - - 7 - Enable posted interrupts * - - VMX_PROC_VM_EXEC_CONTROLS - 3 - Use TSC offsetting * - - - 21 - Use TPR shadow * - - - 25 - Use I/O bitmaps * - - - 28 - Use MSR bitmaps * - - - 31 - Activate secondary controls * - - VMX_PROC_VM_EXEC_CONTROLS2 - 0 - Virtualize APIC accesses * - - - 1 - Enable EPT * - - - 3 - Enable RDTSCP * - - - 5 - Enable VPID * - - - 7 - Unrestricted guest * - - - 8 - APIC-register virtualization * - - - 9 - Virtual-interrupt delivery * - - - 20 - Enable XSAVES/XRSTORS * - **guest state** - CS, DS, ES, FS, GS, TR, LDTR, GDTR, IDTR - n/a - According to vCPU mode and init_ctx * - - RIP, RSP - n/a - According to vCPU mode and init_ctx * - - CR0, CR3, CR4 - n/a - According to vCPU mode and init_ctx * - - GUEST_IA32_SYSENTER_CS/ESP/EIP - n/a - Set to 0 * - - GUEST_IA32_PAT - n/a - Set to PAT_POWER_ON_VALUE * - **entry control** - VMX_ENTRY_CONTROLS - 2 - Load debug controls * - - - 14 - Load IA32_PAT * - - - 15 - Load IA23_EFER * - **exit control** - VMX_EXIT_CONTROLS - 2 - Save debug controls * - - - 9 - Host address space size * - - - 15 - Acknowledge Interrupt on exit * - - - 18 - Save IA32_PAT * - - - 19 - Load IA32_PAT * - - - 20 - Save IA32_EFER * - - - 21 - Load IA32_EFER CPUID Virtualization ******************** CPUID access from guest would cause VM exits unconditionally if executed as a VMX non-root operation. ACRN must return the emulated processor identification and feature information in the EAX, EBX, ECX, and EDX registers. To simplify, ACRN returns the same values from the physical CPU for most of the CPUID, and specially handle a few CPUID features which are APIC ID related such as CPUID.01H. ACRN emulates some extra CPUID features for the hypervisor as well. There is a per-vm *vcpuid_entries* array, initialized during VM creation and used to cache most of the CPUID entries for each VM. During guest CPUID emulation, ACRN will read the cached value from this array, except some APIC ID-related CPUID data emulated at runtime. This table describes details for CPUID emulation: .. list-table:: :widths: 20 80 :header-rows: 1 * - **CPUID** - **Emulation Description** * - 01H - - Get original value from physical CPUID - Fill APIC ID from vLAPIC - Disable x2APIC - Disable PCID - Disable VMX - Disable XSAVE if host not enabled * - 0BH - - Fill according to X2APIC feature support (default is disabled) - If not supported, fill all registers with 0 - If supported, get from physical CPUID * - 0DH - - Fill according to XSAVE feature support - If not supported, fill all registers with 0 - If supported, get from physical CPUID * - 07H - - Get from per-vm CPUID entries cache - For subleaf 0, disabled INVPCID, Intel RDT * - 16H - - Get from per-vm CPUID entries cache - If physical CPU support CPUID.16H, read from physical CPUID - If physical CPU does not support it, emulate with tsc freq * - 40000000H - - Get from per-vm CPUID entries cache - EAX: the maximum input value for CPUID supported by ACRN (40000010) - EBX, ECX, EDX: hypervisor vendor ID signature - "ACRNACRNACRN" * - 40000010H - - Get from per-vm CPUID entries cache - EAX: virtual TSC frequency in KHz - EBX, ECX, EDX: reserved to 0 * - 0AH - - PMU Currently disabled * - 0FH, 10H - - Intel RDT Currently disabled * - 12H - - Fill according to SGX virtualization * - 14H - - Intel Processor Trace Currently disabled * - Others - - Get from per-vm CPUID entries cache .. note:: ACRN needs to take care of some CPUID values that can change at runtime, for example, XD feature in CPUID.80000001H may be cleared by MISC_ENABLE MSR. MSR Virtualization ****************** ACRN always enables MSR bitmap in *VMX_PROC_VM_EXEC_CONTROLS* VMX execution control field. This bitmap marks the MSRs to cause a VM exit upon guest access for both read and write. The VM exit reason for reading or writing these MSRs is respectively *VMX_EXIT_REASON_RDMSR* or *VMX_EXIT_REASON_WRMSR* and the vm exit handler is *rdmsr_vmexit_handler* or *wrmsr_vmexit_handler*. This table shows the predefined MSRs ACRN will trap for all the guests. For the MSRs whose bitmap are not set in the MSR bitmap, guest access will be pass-through directly: .. list-table:: :widths: 33 33 33 :header-rows: 1 * - **MSR** - **Description** - **Handler** * - MSR_IA32_TSC_ADJUST - TSC adjustment of local APIC's TSC deadline mode - emulates with vlapic * - MSR_IA32_TSC_DEADLINE - TSC target of local APIC's TSC deadline mode - emulates with vlapic * - MSR_IA32_BIOS_UPDT_TRIG - BIOS update trigger - work for update microcode from the Service VM, the signature ID read is from physical MSR, and a BIOS update trigger from the Service VM will trigger a physical microcode update. * - MSR_IA32_BIOS_SIGN_ID - BIOS update signature ID - " * - MSR_IA32_TIME_STAMP_COUNTER - Time-stamp counter - work with VMX_TSC_OFFSET_FULL to emulate virtual TSC * - MSR_IA32_APIC_BASE - APIC base address - emulates with vlapic * - MSR_IA32_PAT - Page-attribute table - save/restore in vCPU, write to VMX_GUEST_IA32_PAT_FULL if cr0.cd is 0 * - MSR_IA32_PERF_CTL - Performance control - Trigger real p-state change if p-state is valid when writing, fetch physical MSR when reading * - MSR_IA32_FEATURE_CONTROL - Feature control bits that configure operation of VMX and SMX - disabled, locked * - MSR_IA32_MCG_CAP/STATUS - Machine-Check global control/status - emulates with vMCE * - MSR_IA32_MISC_ENABLE - Miscellaneous feature control - readonly, except MONITOR/MWAIT enable bit * - MSR_IA32_SGXLEPUBKEYHASH0/1/2/3 - SHA256 digest of the authorized launch enclaves - emulates with vSGX * - MSR_IA32_SGX_SVN_STATUS - Status and SVN threshold of SGX support for ACM - readonly, emulates with vSGX * - MSR_IA32_MTRR_CAP - Memory type range register related - Handled by MTRR emulation. * - MSR_IA32_MTRR_DEF_TYPE - " - " * - MSR_IA32_MTRR_PHYSBASE_0~9 - " - " * - MSR_IA32_MTRR_FIX64K_00000 - " - " * - MSR_IA32_MTRR_FIX16K_80000/A0000 - " - " * - MSR_IA32_MTRR_FIX4K_C0000~F8000 - " - " * - MSR_IA32_X2APIC_* - x2APIC related MSRs (offset from 0x800 to 0x900) - emulates with vlapic * - MSR_IA32_L2_MASK_BASE~n - L2 CAT mask for CLOSn - disabled for guest access * - MSR_IA32_L3_MASK_BASE~n - L3 CAT mask for CLOSn - disabled for guest access * - MSR_IA32_MBA_MASK_BASE~n - MBA delay mask for CLOSn - disabled for guest access * - MSR_IA32_VMX_BASIC~VMX_TRUE_ENTRY_CTLS - VMX related MSRs - not support, access will inject #GP CR Virtualization ***************** ACRN emulates ``mov to cr0``, ``mov to cr4``, ``mov to cr8``, and ``mov from cr8`` through *cr_access_vmexit_handler* based on *VMX_EXIT_REASON_CR_ACCESS*. .. note:: Currently ``mov to cr8`` and ``mov from cr8`` are actually not valid as ``CR8-load/store exiting`` bits are set as 0 in *VMX_PROC_VM_EXEC_CONTROLS*. A VM can ``mov from cr0`` and ``mov from cr4`` without triggering a VM exit. The value read are the read shadows of the corresponding register in VMCS. The shadows are updated by the hypervisor on CR writes. .. list-table:: :widths: 30 70 :header-rows: 1 * - **Operation** - **Handler** * - mov to cr0 - Based on vCPU set context API: vcpu_set_cr0 -> vmx_write_cr0 * - mov to cr4 - Based on vCPU set context API: vcpu_set_cr4 ->vmx_write_cr4 * - mov to cr8 - Based on vlapic tpr API: vlapic_set_cr8->vlapic_set_tpr * - mov from cr8 - Based on vlapic tpr API: vlapic_get_cr8->vlapic_get_tpr For ``mov to cr0`` and ``mov to cr4``, ACRN sets *cr0_host_mask/cr4_host_mask* into *VMX_CR0_MASK/VMX_CR4_MASK* for the bitmask causing vm exit. As ACRN always enables ``unrestricted guest`` in *VMX_PROC_VM_EXEC_CONTROLS2*, *CR0.PE* and *CR0.PG* can be controlled by guest. .. list-table:: :widths: 20 40 40 :header-rows: 1 * - **CR0 MASK** - **Value** - **Comments** * - cr0_always_on_mask - fixed0 & (~(CR0_PE | CR0_PG)) - where fixed0 is gotten from MSR_IA32_VMX_CR0_FIXED0, means these bits are fixed to be 1 under VMX operation * - cr0_always_off_mask - ~fixed1 - where ~fixed1 is gotten from MSR_IA32_VMX_CR0_FIXED1, means these bits are fixed to be 0 under VMX operation * - CR0_TRAP_MASK - CR0_PE | CR0_PG | CR0_WP | CR0_CD | CR0_NW - ACRN will also trap PE, PG, WP, CD, and NW bits * - cr0_host_mask - ~(fixed0 ^ fixed1) | CR0_TRAP_MASK - ACRN will finally trap bits under VMX root mode control plus additionally added bits For ``mov to cr0`` emulation, ACRN will handle a paging mode change based on PG bit change, and a cache mode change based on CD and NW bits changes. ACRN also takes care of illegal writing from guest to invalid CR0 bits (for example, set PG while CR4.PAE = 0 and IA32_EFER.LME = 1), which will finally inject a #GP to guest. Finally, *VMX_CR0_READ_SHADOW* will be updated for guest reading of host controlled bits, and *VMX_GUEST_CR0* will be updated for real vmx cr0 setting. .. list-table:: :widths: 20 40 40 :header-rows: 1 * - **CR4 MASK** - **Value** - **Comments** * - cr4_always_on_mask - fixed0 - where fixed0 is gotten from MSR_IA32_VMX_CR4_FIXED0, means these bits are fixed to be 1 under VMX operation * - cr4_always_off_mask - ~fixed1 - where ~fixed1 is gotten from MSR_IA32_VMX_CR4_FIXED1, means these bits are fixed to be 0 under VMX operation * - CR4_TRAP_MASK - CR4_PSE | CR4_PAE | CR4_VMXE | CR4_PCIDE | CR4_SMEP | CR4_SMAP | CR4_PKE - ACRN will also trap PSE, PAE, VMXE, and PCIDE bits * - cr4_host_mask - ~(fixed0 ^ fixed1) | CR4_TRAP_MASK - ACRN will finally trap bits under VMX root mode control plus additionally added bits The ``mov to cr4`` emulation is similar to cr0 emulation noted above. .. _io-mmio-emulation: IO/MMIO Emulation ***************** ACRN always enables I/O bitmap in *VMX_PROC_VM_EXEC_CONTROLS* and EPT in *VMX_PROC_VM_EXEC_CONTROLS2*. Based on them, *pio_instr_vmexit_handler* and *ept_violation_vmexit_handler* are used for IO/MMIO emulation for a emulated device. The emulated device could locate in hypervisor or DM in the Service VM. Please refer to the "I/O Emulation" section for more details. For an emulated device done in the hypervisor, ACRN provide some basic APIs to register its IO/MMIO range: - For the Service VM, the default I/O bitmap are all set to 0, which means the Service VM will pass through all I/O port access by default. Adding an I/O handler for a hypervisor emulated device needs to first set its corresponding I/O bitmap to 1. - For User VM, the default I/O bitmap are all set to 1, which means User VM will trap all I/O port access by default. Adding an I/O handler for a hypervisor emulated device does not need change its I/O bitmap. If the trapped I/O port access does not fall into a hypervisor emulated device, it will create an I/O request and pass it to the Service VM DM. - For the Service VM, EPT maps all range of memory to the Service VM except for ACRN hypervisor area. This means the Service VM will pass through all MMIO access by default. Adding a MMIO handler for a hypervisor emulated device needs to first remove its MMIO range from EPT mapping. - For User VM, EPT only maps its system RAM to the User VM, which means User VM will trap all MMIO access by default. Adding a MMIO handler for a hypervisor emulated device does not need to change its EPT mapping. If the trapped MMIO access does not fall into a hypervisor emulated device, it will create an I/O request and pass it to the Service VM DM. .. list-table:: :widths: 30 70 :header-rows: 1 * - **API** - **Description** * - register_pio_emulation_handler - register an I/O emulation handler for a hypervisor emulated device by specific I/O range * - register_mmio_emulation_handler - register a MMIO emulation handler for a hypervisor emulated device by specific MMIO range .. _instruction-emulation: Instruction Emulation ********************* ACRN implements a simple instruction emulation infrastructure for MMIO (EPT) and APIC access emulation. When such a VM exit is triggered, the hypervisor needs to decode the instruction from RIP then attempt the corresponding emulation based on its instruction and read/write direction. ACRN currently supports emulating instructions for ``mov``, ``movx``, ``movs``, ``stos``, ``test``, ``and``, ``or``, ``cmp``, ``sub`` and ``bittest`` without support for lock prefix. Real mode emulation is not supported. .. figure:: images/hld-image82.png :align: center Instruction Emulation Work Flow In the handlers for EPT violation or APIC access VM exit, ACRN will: 1. Fetch the MMIO access request's address and size 2. Do *decode_instruction* for the instruction in current RIP with the following check: a. Is the instruction supported? If not, inject #UD to guest. b. Is GVA of RIP, dest, and src valid? If not, inject #PF to guest. c. Is stack valid? If not, inject #SS to guest. 3. If step 2 succeeds, check the access direction. If it's a write, then do *emulate_instruction* to fetch MMIO request's value from instruction operands. 4. Execute MMIO request handler, for EPT violation is *emulate_io* while APIC access is *vlapic_write/read* based on access direction. It will finally complete this MMIO request emulation by: a. putting req.val to req.addr for write operation b. getting req.val from req.addr for read operation 5. If the access direction is read, then do *emulate_instruction* to put MMIO request's value into instruction operands. 6. Return to guest. TSC Emulation ************* Guest vCPU execution of *RDTSC/RDTSCP* and access to *MSR_IA32_TSC_AUX* does not cause a VM Exit to the hypervisor. Hypervisor uses *MSR_IA32_TSC_AUX* to record CPU ID, thus the CPU ID provided by *MSR_IA32_TSC_AUX* might be changed via Guest. *RDTSCP* is widely used by hypervisor to identify current CPU ID. Due to no VM Exit for *MSR_IA32_TSC_AUX* MSR register, ACRN hypervisor saves/restores *MSR_IA32_TSC_AUX* value on every VM Exit/Enter. Before hypervisor restores host CPU ID, *rdtscp* should not be called as it could get vCPU ID instead of host CPU ID. The *MSR_IA32_TIME_STAMP_COUNTER* is emulated by ACRN hypervisor, with a simple implementation based on *TSC_OFFSET* (enabled in *VMX_PROC_VM_EXEC_CONTROLS*): - For read: ``val = rdtsc() + exec_vmread64(VMX_TSC_OFFSET_FULL)`` - For write: ``exec_vmwrite64(VMX_TSC_OFFSET_FULL, val - rdtsc())`` ART Virtualization ****************** The invariant TSC is based on the invariant timekeeping hardware (called Always Running Timer or ART), that runs at the core crystal clock frequency. The ratio defined by the CPUID leaf 15H expresses the frequency relationship between the ART hardware and the TSC. If CPUID.15H.EBX[31:0] != 0 and CPUID.80000007H:EDX[InvariantTSC] = 1, the following linearity relationship holds between the TSC and the ART hardware: ``TSC_Value = (ART_Value * CPUID.15H:EBX[31:0]) / CPUID.15H:EAX[31:0] + K`` Where `K` is an offset that can be adjusted by a privileged agent. When ART hardware is reset, both invariant TSC and K are also reset. The guideline of ART virtualization (vART) is that software in native can run in VM too. The vART solution is: - Present the ART capability to guest through CPUID leaf 15H for `CPUID.15H:EBX[31:0]` and `CPUID.15H:EAX[31:0]`. - Passthrough devices see the physical ART_Value (vART_Value = pART_Value) - Relationship between the ART and TSC in guest is: ``vTSC_Value = (vART_Value * CPUID.15H:EBX[31:0]) / CPUID.15H:EAX[31:0] + vK`` Where `vK = K + VMCS.TSC_OFFSET`. - If `vK` or `vTSC_Value` are changed by guest, we change the `VMCS.TSC_OFFSET` accordingly. - `K` should never be changed by hypervisor. XSAVE Emulation *************** The XSAVE feature set is comprised of eight instructions: - *XGETBV* and *XSETBV* allow software to read and write the extended control register *XCR0*, which controls the operation of the XSAVE feature set. - *XSAVE*, *XSAVEOPT*, *XSAVEC*, and *XSAVES* are four instructions that save processor state to memory. - *XRSTOR* and *XRSTORS* are corresponding instructions that load processor state from memory. - *XGETBV*, *XSAVE*, *XSAVEOPT*, *XSAVEC*, and *XRSTOR* can be executed at any privilege level; - *XSETBV*, *XSAVES*, and *XRSTORS* can be executed only if CPL = 0. Enabling the XSAVE feature set is controlled by XCR0 (through XSETBV) and IA32_XSS MSR. Refer to the `Intel SDM Volume 1`_ chapter 13 for more details. .. _Intel SDM Volume 1: https://www.intel.com/content/www/us/en/architecture-and-technology/64-ia-32-architectures-software-developer-vol-1-manual.html .. figure:: images/hld-image38.png :align: center ACRN Hypervisor XSAVE emulation By default, ACRN enables XSAVES/XRSTORS in *VMX_PROC_VM_EXEC_CONTROLS2*, so it allows the guest to use the XSAVE feature. Because guest execution of *XSETBV* will always trigger XSETBV VM exit, ACRN actually needs to take care of XCR0 access. ACRN emulates XSAVE features through the following rules: 1. Enumerate CPUID.01H for native XSAVE feature support 2. If yes for step 1, enable XSAVE in hypervisor by CR4.OSXSAVE 3. Emulates XSAVE related CPUID.01H & CPUID.0DH to guest 4. Emulates XCR0 access through *xsetbv_vmexit_handler* 5. ACRN pass-through the access of IA32_XSS MSR to guest 6. ACRN hypervisor does NOT use any feature of XSAVE 7. As ACRN emulate vCPU with partition mode, so based on above rules 5 and 6, a guest vCPU will fully control the XSAVE feature in non-root mode.