Because the emulation code is for both split-lock and uc-lock, Changed
these API names:
vcpu_kick_splitlock_emulation() -> vcpu_kick_lock_instr_emulation()
vcpu_complete_splitlock_emulation() -> vcpu_complete_lock_instr_emulation()
emulate_splitlock() -> emulate_lock_instr()
Tracked-On: #6299
Signed-off-by: Tao Yuhong <yuhong.tao@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Split-lock emulation can be re-used for uc-lock. In emulate_splitlock(),
it only work if this vmexit is for #AC trap and guest do not handle
split-lock and HV enable #AC for splitlock.
Add another condition to let emulate_splitlock() also work for #GP trap
and guest do not handle uc-lock and HV enable #GP for uc-lock.
Tracked-On: #6299
Signed-off-by: Tao Yuhong <yuhong.tao@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
When ACRN uses decode_instruction to emulate split-lock/uc-lock
instruction, It is actually a try-decode to see if it is XCHG.
If the instruction is XCHG instruction, ACRN must emulate it
(inject #PF if it is triggered) with peer VCPUs paused, and advance
the guest IP. If the instruction is a LOCK prefixed instruction
with accessing the UC memory, ACRN Halted the peer VCPUs, and
advance the IP to skip the LOCK prefix, and then let the VCPU
Executes one instruction by enabling IRQ Windows vm-exit. For
other cases, ACRN injects the exception back to VCPU without
emulating it.
So change the API to decode_instruction(vcpu, bool full_decode),
when full_decode is true, the API does same thing as before. When
full_decode is false, the different is if decode_instruction() meet unknown
instruction, will keep return = -1 and do not inject #UD. We can use
this to distinguish that an #UD has been skipped, and need inject #AC/#GP back.
Tracked-On: #6299
Signed-off-by: Tao Yuhong <yuhong.tao@intel.com>
Guest may not use INVEPT instruction after enabling any of bits 2:0 from
0 to 1 of a present EPT entry, then the shadow EPT entry has no chance
to sync guest EPT entry. According to the SDM,
"""
Software may use the INVEPT instruction after modifying a present EPT
paging-structure entry (see Section 28.2.2) to change any of the
privilege bits 2:0 from 0 to 1.1 Failure to do so may cause an EPT
violation that would not otherwise occur. Because an EPT violation
invalidates any mappings that would be used by the access that caused
the EPT violation (see Section 28.3.3.1), an EPT violation will not
recur if the original access is performed again, even if the INVEPT
instruction is not executed.
"""
Sync the afterthought of privilege bits from guest EPT entry to shadow
EPT entry to cover above case.
Tracked-On: #5923
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
MSR_IA32_VMX_EPT_VPID_CAP is 64 bits. Using 32 bits MACROs with it may
cause the bit expression wrong.
Unify the MSR_IA32_VMX_EPT_VPID_CAP operation with 64 bits definition.
Tracked-On: #5923
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
When hypervisor boots, the multiboot modules have been loaded to host space
by bootloader already. The space range of pre-launched VM modules is also
exposed to SOS VM, so SOS VM kernel might pick this range to extract kernel
when KASLR enabled. This would corrupt pre-launched VM modules and result in
pre-launched VM boot fail.
This patch will try to fix this issue. The SOS VM will not be loaded to guest
space until all pre-launched VMs are loaded successfully.
Tracked-On: #5879
Signed-off-by: Victor Sun <victor.sun@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Previously the load GPA of LaaG boot params like zeropage/cmdline and
initgdt are all hard-coded, this would bring potential LaaG boot issues.
The patch will try to fix this issue by finding a 32KB load_params memory
block for LaaG to store these guest boot params.
For other guest with raw image, in general only vgdt need to be cared of so
the load_params will be put at 0x800 since it is a common place that most
guests won't touch for entering protected mode.
Tracked-On: #5626
Signed-off-by: Victor Sun <victor.sun@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
The API would search ve820 table and return a valid GPA when the requested
size of memory is available in the specified memory range, or return
INVALID_GPA if the requested memory slot is not available;
Tracked-On: #5626
Signed-off-by: Victor Sun <victor.sun@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
The memory range of [0xA0000, 0xFFFFF] is a known reserved area for BIOS,
actually Linux kernel would enforce this area to be reserved during its
boot stage. Set this area to usable would cause potential compatibility
issues.
The patch set the range to reserved type to make it consistent with the
real world.
BTW, There should be a EBDA(Entended BIOS DATA Area) with reserved type
exist right before 0xA0000 in real world for non-EFI boot. But given ACRN
has no legacy BIOS emulation, we simply skipped the EBDA in vE820.
Tracked-On: #5626
Signed-off-by: Victor Sun <victor.sun@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Given the structure in multiboot.h could be used for any boot protocol,
use a more generic name "boot.h" instead;
Tracked-On: #5661
Signed-off-by: Victor Sun <victor.sun@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
vmptrld_vmexit_handler() has a same code snippet with
vmclear_vmexit_handler(). Wrap the same code snippet as a static
function clear_vmcs02().
There is only a small logic change that add
nested->current_vmcs12_ptr = INVALID_GPA
in vmptrld_vmexit_handler() for the old VMCS. That's reasonable.
Tracked-On: #5923
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
get_ept_entry() actually returns the EPTP of a VM. So rename it to
get_eptp() for readability.
Tracked-On: #5923
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
We need to deny accesses from SOS to the HV owned UART device, otherwise
SOS could have direct access to this physical device and mess up the HV
console.
If ACRN debug UART is configured as PIO based, For example,
CONFIG_SERIAL_PIO_BASE is generated from acrn-config tool, or the UART
config is overwritten by hypervisor parameter "uart=port@<port address>",
it could run into problem if ACRN doesn't emulate this UART PIO port
to SOS. For example:
- none of the ACRN emulated vUART devices has same PIO port with the
port of the debug UART device.
- ACRN emulates PCI vUART for SOS (configure "console_vuart" with
PCI_VUART in the scenario configuration)
This patch fixes the above issue by masking PIO accesses from SOS.
deny_hv_owned_devices() is moved after setup_io_bitmap() where
vm->arch_vm.io_bitmap is initialized.
Commit 50d852561 ("HV: deny HV owned PCI bar access from SOS") handles
the case that ACRN debug UART is configured as a PCI device. e.g.,
hypervisor parameter "uart=bdf@<BDF value>" is appended.
If the hypervisor debug UART is MMIO based, need to configured it as
a PCI type device, so that it can be hidden from SOS.
Tracked-On: #5923
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Malicious input 'index' may trigger buffer
overflow on array 'irte_alloc_bitmap[]'.
This patch validate that 'index' shall be
less than 'CONFIG_MAX_IR_ENTRIES' and also
remove unnecessary check on 'index' in
'ptirq_free_irte()' function with this fix.
Tracked-On: #6132
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
vlapic_write handle 'offset' that is valid and ignore
all other invalid 'offset'. so ASSERT on this 'offset'
input is unnecessary.
This patch removes above ASSERT to avoid potential
hypervisor crash by guest malicious input when debug
build is used.
Tracked-On: #6131
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
generate_shadow_ept_entry() didn't verify the correctness of the requested
guest EPT mapping. That might leak host memory access to L2 VM.
To simplify the implementation of the guest EPT audit, hide capabilities
'map 2-Mbyte page' and 'map 1-Gbyte page' from L1 VM. In addition,
minimize the attribute bits of EPT entry when create a shadow EPT entry.
Also, for invalid requested mapping address, reflect the EPT_VIOLATION to
L1 VM.
Here, we have some TODOs:
1) Enable large page support in generate_shadow_ept_entry()
2) Evaluate if need to emulate the invalid GPA access of L2 in HV directly.
3) Minimize EPT entry attributes.
Tracked-On: #5923
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
L1 VM changes the guest EPT and do INVEPT to invalidate the previous
TLB cache of EPT entries. The shadow EPT replies on INVEPT instruction
to do the update.
The target shadow EPTs can be found according to the 'type' of INVEPT.
Here are two types and their target shadow EPT,
1) Single-context invalidation
Get the EPTP from the INVEPT descriptor. Then find the target
shadow EPT.
2) Global invalidation
All shadow EPTs of the L1 VM.
The INVEPT emulation handler invalidate all the EPT entries of the
target shadow EPTs.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
When a shadow EPT is not used anymore, its resources need to be
released.
free_sept_table() is introduced to walk the whole shadow EPT table and
free the pagetable pages.
Please note, the PML4E page of shadow EPT is not freed by
free_sept_table() as it still be used to present a shadow EPT pointer.
Tracked-On: #5923
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
With shadow EPT, the hypervisor walks through guest EPT table:
* If the entry is not present in guest EPT, ACRN injects EPT_VIOLATION
to L1 VM and resumes to L1 VM.
* If the entry is present in guest EPT, do the EPT_MISCONFIG check.
Inject EPT_MISCONFIG to L1 VM if the check failed.
* If the entry is present in guest EPT, do permission check.
Reflect EPT_VIOLATION to L1 VM if the check failed.
* If the entry is present in guest EPT but shadow EPT entry is not
present, create the shadow entry and resumes to L2 VM.
* If the entry is present in guest EPT but the GPA in the entry is
invalid, injects EPT_VIOLATION to L1 VM and resumes L1 VM.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
* Hide 5 level EPT capability, let L1 guest stick to 4 level EPT.
* Access/Dirty bits are not support currently, hide corresponding EPT
capability bits.
* "Mode-based execute control for EPT" is also not support well
currently, hide its capability bit from MSR_IA32_VMX_PROCBASED_CTLS2.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
'struct nept_desc' is used to associate guest EPTP with a shadow EPTP.
It's created in the first reference and be freed while no reference.
The life cycle seems like,
While guest VMCS VMX_EPT_POINTER_FULL is changed, the 'struct nept_desc'
of the new guest EPTP is referenced; the 'struct nept_desc' of the old
guest EPTP is dereferenced.
While guest VMCS be cleared(by VMCLEAR in L1 VM), the 'struct nept_desc'
of the old guest EPTP is dereferenced.
While a new guest VMCS be loaded(by VMPTRLD in L1 VM), the 'struct
nept_desc' of the new guest EPTP is referenced. The 'struct nept_desc'
of the old guest EPTP is dereferenced.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
To shadow guest EPT, the hypervisor needs construct a shadow EPT for each
guest EPT. The key to associate a shadow EPT and a guest EPT is the EPTP
(EPT pointer). This patch provides following structure to do the association.
struct nept_desc {
/*
* A shadow EPTP.
* The format is same with 'EPT pointer' in VMCS.
* Its PML4 address field is a HVA of the hypervisor.
*/
uint64_t shadow_eptp;
/*
* An guest EPTP configured by L1 VM.
* The format is same with 'EPT pointer' in VMCS.
* Its PML4 address field is a GPA of the L1 VM.
*/
uint64_t guest_eptp;
uint32_t ref_count;
};
Due to lack of dynamic memory allocation of the hypervisor, a array
nept_bucket of type 'struct nept_desc' is introduced to store those
association information. A guest EPT might be shared between different
L2 vCPUs, so this patch provides several functions to handle the
reference of the structure.
Interface get_shadow_eptp() also is introduced. To find the shadow EPTP
of a specified guest EPTP.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Shadow EPT uses lots of pages to construct the shadow page table. To
utilize the memory more efficient, a page poll sept_page_pool is
introduced.
For simplicity, total platform RAM size is considered to calculate the
memory needed for shadow page tables. This is not an accurate upper
bound. This can satisfy typical use-cases where there is not a lot
of overcommitment and sharing of memory between L2 VMs.
Memory of the pool is marked as reserved from E820 table in early stage.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Nested VM exits happen when vCPU is in guest mode (VMCS02 is current).
Initially we reflect all nested VM exits to L1 hypervisor. To prepare
the environment to run L1 guest:
- restore some VMCS fields to the value as what L1 hypervisor programmed.
- VMCLEAR VMCS02, VMPTRLD VMCS01 and enable VMCS shadowing.
- load the non-shadowing host states from VMCS12 to VMCS01 guest states.
- VMRESUME to L1 guest with this modified VMCS01.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Alexander Merritt <alex.merritt@intel.com>
Since L2 guest vCPU mode and VPID are managed by L1 hypervisor, so we
can skip these handling in run_vcpu().
And be careful that we can't cache L2 registers in struct acrn_vcpu.
Tracked-On: #5923
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
invvpid and invept instructions cause VM exits unconditionally.
For initial support, we pass all the instruction operands as is
to the pCPU.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Implement the VMLAUNCH and VMRESUME instructions, allowing a L1
hypervisor to run nested guests.
- merge VMCS control fields and VMCS guest fields to VMCS02
- clear shadow VMCS indicator on VMCS02 and load VMCS02 as current
- set VMCS12 launch state to "launched" in VMLAUNCH handler
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Alex Merritt <alex.merritt@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
In physical destination mode, the destination processor is specified by its
local APIC ID. When a CPU switch xAPIC Mode to x2APIC Mode or vice versa,
the local APIC ID is not changed. So a vcpu in x2APIC Mode could use physical
Destination Mode to send an IPI to another vcpu in xAPIC Mode by writing ICR.
This patch adds support for a vCPU A could write ICR to send IPI to another
vCPU B which is in different APIC mode.
Tracked-On: #5923
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Using physical APIC IDs as vLAPIC IDs for pre-Launched and post-launched VMs
is not sufficient to replicate the host CPU and cache topologies in guest VMs,
we also need to passthrough host CPUID leaf.0BH to guest VMs, otherwise,
guest VMs may see weird CPU topology.
Note that in current code, ACRN has already passthroughed host cache CPUID
leaf 04H to guest VMs
Tracked-On: #6020
Reviewed-by: Wang, Yu1 <yu1.wang@intel.com>
Signed-off-by: dongshen <dongsheng.x.zhang@intel.com>
In current code, ACRN uses physical APIC IDs as vLAPIC IDs for SOS,
and vCPU ids (contiguous) as vLAPIC IDs for pre-Launched and post-Launched VMs.
Using vCPU ids as vLAPIC IDs for pre-Launched and post-Launched VMs
would result in wrong CPU and cache topologies showing in the guest VMs,
and could adversely affect performance if the guest VM chooses to detect
CPU and cache topologies and optimize its behavior accordingly.
Uses physical APIC IDs as vLAPIC IDs (and related CPU/cache topology enumeration
CPUIDs passthrough) will replicate the host CPU and cache topologies in pre-Launched
and post-Launched VMs.
Tracked-On: #6020
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Signed-off-by: dongshen <dongsheng.x.zhang@intel.com>
Remove the direct calls to exec_vmptrld() or exec_vmclear(), and replace
with the wrapper APIs load_va_vmcs() and clear_va_vmcs().
Tracked-On: #5923
Signed-off-by: Zide Chen <zide.chen@intel.com>
This patch implements the VMREAD and VMWRITE instructions.
When L1 guest is running with an active VMCS12, the “VMCS shadowing”
VM-execution control is always set to 1 in VMCS01. Thus the possible
behavior of VMREAD or VMWRITE from L1 could be:
- It causes a VM exit to L0 if the bit corresponds to the target VMCS
field in the VMREAD bitmap or VMWRITE bitmap is set to 1.
- It accesses the VMCS referenced by VMCS01 link pointer (VMCS02 in
our case) if the above mentioned bit is set to 0.
This patch handles the VMREAD and VMWRITE VM exits in this way:
- on VMWRITE, it writes the desired VMCS value to the respective field
in the cached VMCS12. For VMCS fields that need to be synced to VMCS02,
sets the corresponding dirty flag.
- on VMREAD, it reads the desired VMCS value from the cached VMCS12.
Tracked-On: #5923
Signed-off-by: Alex Merritt <alex.merritt@intel.com>
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@Intel.com>
This patch is to emulate VMCLEAR instruction.
L1 hypervisor issues VMCLEAR on a VMCS12 whose state could be any of
these: active and current, active but not current, not yet VMPTRLDed.
To emulate the VMCLEAR instruction, ACRN sets the VMCS12 launch state to
"clear", and if L0 already cached this VMCS12, need to sync it back to
guest memory:
- sync shadow fields from shadow VMCS VMCS to cache VMCS12
- copy cache VMCS12 to L1 guest memory
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Enable VMCS shadowing for most of the VMCS fields, so that execution of
the VMREAD or VMWRITE on these shadow VMCS fields from L1 hypervisor
won't cause VM exits, but read from or write to the shadow VMCS.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Alexander Merritt <alex.merritt@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Software layout of VMCS12 data is a contract between L1 guest and L0
hypervisor to run a L2 guest.
ACRN hypervisor caches the VMCS12 which is passed down from L1 hypervisor
by the VMPTRLD instructin. At the time of VMCLEAR, ACRN syncs the cached
VMCS12 back to L1 guest memory.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@Intel.com>
This patch emulates the VMPTRLD instruction. L0 hypervisor (ACRN) caches
the VMCS12 that is passed down from the VMPTRLD instruction, and merges it
with VMCS01 to create VMCS02 to run the nested VM.
- Currently ACRN can't cache multiple VMCS12 on one vCPU, so it needs to
flushes active but not current VMCS12s to L1 guest.
- ACRN creates VMCS02 to run nested VM based on VMCS12:
1) copy VMCS12 from guest memory to the per vCPU cache VMCS12
2) initialize VMCS02 revision ID and host-state area
3) load shadow fields from cache VMCS12 to VMCS02
4) enable VMCS shadowing before L1 Vm entry
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
This patch implements the VMXOFF instruction. By issuing VMXOFF,
L1 guest Leaves VMX Operation.
- cleanup VCPU nested virtualization context states in VMXOFF handler.
- implement check_vmx_permission() to check permission for VMX operation
for VMXOFF and other VMX instructions.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@Intel.com>
According to VMXON Instruction Reference, do the following checks in the
virtual hardware environment: vCPU CPL, guest CR0, CR4, revision ID
in VMXON region, etc.
Currently ACRN doesn't support 32-bit L1 hypervisor, and injects an #UD
exception if L1 hypervisor is not running in 64-bit mode.
Tracked-On: #5923
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@Intel.com>
This patch emulates VMXON instruction. Basically checks some
prerequisites to enable VMX operation on L1 guest (next patch), and
prepares some virtual hardware environment in L0.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@Intel.com>
Now guest would use `Destination Shorthand` to broadcast IPIs if there're more
than one destination. However, it is not supported when the guest is in LAPIC
passthru situation, and all active VCPUs are working in X2APIC mode. As a result,
the guest would not work properly since this kind broadcast IPIs was ignored
by ACRN. What's worse, ACRN Hypervisor would inject GP to the guest in this case.
This patch extend vlapic_x2apic_pt_icr_access to support more destination modes
(both `Physical` and `Logical`) and destination shorthand (`No Shorthand`, `Self`,
`All Including Self` and `All Excluding Self`).
Tracked-On: #5923
Signed-off-by: Zide Chen <zide.chen@intel.com>
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
1. do not allow external modules to touch internal field of a timer.
2. make timer mode internal, period_in_ticks will decide the mode.
API wise:
1. the "mode" parameter was taken out of initialize_timer().
2. a new function update_timer() was added to update the timeout and
period fields.
3. the timer_expired() function was extended with an output parameter
to return the remaining cycles before expiration.
Also, the "fire_tsc" field name of hv_timer was renamed to "timeout".
With the new API, however, this change should not concern user code.
Tracked-On: #5920
Signed-off-by: Jason Chen CJ <jason.cj.chen@intel.com>
Generalize and split basic cpu cycle/tick routines from x86/timer:
- Instead of rdstc(), use cpu_ticks() in generic code.
- Instead of get_tsc_khz(), use cpu_tickrate() in generic code.
- Include "common/ticks.h" instead of "x86/timer.h" in generic code.
- CYCLES_PER_MS is renamed to TICKS_PER_MS.
The x86 specific API rdstc() and get_tsc_khz(), as well as TSC_PER_MS
are still available in arch/x86/tsc.h but only for x86 specific usage.
Tracked-On: #5920
Signed-off-by: Rong Liu <rong2.liu@intel.com>
Signed-off-by: Yi Liang <yi.liang@intel.com>
Define LIST_OF_VMX_MSRS which includes a list of MSRs that are visible to
L1 guests if nested virtualization is enabled.
- If CONFIG_NVMX_ENABLED is set, these MSRs are included in
emulated_guest_msrs[].
- otherwise, they are included in unsupported_msrs[].
In this way we can take advantage of the existing infrastructure to
emulate these MSRs.
Tracked-On: #5923
Spick igned-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
In order to support nested virtualization, need to expose the "Enable VMX
outside SMX operation" bit to L1 hypervisor.
Tracked-On: #5923
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Allow guest set CR4_VMXE if CONFIG_NVMX_ENABLED is set:
- move CR4_VMXE from CR4_EMULATED_RESERVE_BITS to CR4_TRAP_AND_EMULATE_BITS
so that CR4_VMXE is removed from cr4_reserved_bits_mask.
- force CR4_VMXE to be removed from cr4_rsv_bits_guest_value so that CR4_VMXE
is able to be set.
Expose VMX feature (CPUID01.01H:ECX[5]) to L1 guests whose GUEST_FLAG_NVMX_ENABLED
is set.
Assuming guest hypervisor (L1) is KVM, and KVM uses EPT for L2 guests.
Constraints on ACRN VM.
- LAPIC passthrough should be enabled.
- use SCHED_NOOP scheduler.
Tracked-On: #5923
Signed-off-by: Sainath Grandhi <sainath.grandhi@intel.com>
Signed-off-by: Zide Chen <zide.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
TPAUSE, UMONITOR or UMWAIT instructions execution in guest VM cause
a #UD if "enable user wait and pause" (bit 26) of VMX_PROCBASED_CTLS2
is not set. To fix this issue, set the bit 26 of VMX_PROCBASED_CTLS2.
Besides, these WAITPKG instructions uses MSR_IA32_UMWAIT_CONTROL. So
load corresponding vMSR value during context switch in of a vCPU.
Please note, the TPAUSE or UMWAIT instruction causes a VM exit if the
"RDTSC exiting" and "enable user wait and pause" are both 1. In ACRN
hypervisor, "RDTSC exiting" is always 0. So TPAUSE or UMWAIT doesn't
cause a VM exit.
Performance impact:
MSR_IA32_UMWAIT_CONTROL read costs ~19 cycles;
MSR_IA32_UMWAIT_CONTROL write costs ~63 cycles.
Tracked-On: #6006
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
The current permission-checking and dispatching mechanism of hypercalls is
not unified because:
1. Some hypercalls require the exact vCPU initiating the call, while the
others only need to know the VM.
2. Different hypercalls have different permission requirements: the
trusty-related ones are enabled by a guest flag, while the others
require the initiating VM to be the Service OS.
Without a unified logic it could be hard to scale when more kinds of
hypercalls are added later.
The objectives of this patch are as follows.
1. All hypercalls have the same prototype and are dispatched by a unified
logic.
2. Permissions are checked by a unified logic without consulting the
hypercall ID.
To achieve the first objective, this patch modifies the type of the first
parameter of hcall_* functions (which are the callbacks implementing the
hypercalls) from `struct acrn_vm *` to `struct acrn_vcpu *`. The
doxygen-style documentations are updated accordingly.
To achieve the second objective, this patch adds to `struct hc_dispatch` a
`permission_flags` field which specifies the guest flags that must ALL be
set for a VM to be able to invoke the hypercall. The default value (which
is 0UL) indicates that this hypercall is for SOS only. Currently only the
`permission_flag` of trusty-related hypercalls have the non-zero value
GUEST_FLAG_SECURE_WORLD_ENABLED.
With `permission_flag`, the permission checking logic of hypercalls is
unified as follows.
1. General checks
i. If the VM is neither SOS nor having any guest flag that allows
certain hypercalls, it gets #UD upon executing the `vmcall`
instruction.
ii. If the VM is allowed to execute the `vmcall` instruction, but
attempts to execute it in ring 1, 2 or 3, the VM gets #GP(0).
2. Hypercall-specific checks
i. If the hypercall is for SOS (i.e. `permission_flag` is 0), the
initiating VM must be SOS and the specified target VM cannot be a
pre-launched VM. Otherwise the hypercall returns -EINVAL without
further actions.
ii. If the hypercall requires certain guest flags, the initiating VM
must have all the required flags. Otherwise the hypercall returns
-EINVAL without further actions.
iii. A hypercall with an unknown hypercall ID makes the hypercall
returns -EINVAL without further actions.
The logic above is different from the current implementation in the
following aspects.
1. A pre-launched VM now gets #UD (rather than #GP(0)) when it attempts
to execute `vmcall` in ring 1, 2 or 3.
2. A pre-launched VM now gets #UD (rather than the return value -EPERM)
when it attempts to execute a trusty hypercall in ring 0.
3. The SOS now gets the return value -EINVAL (rather than -EPERM) when it
attempts to invoke a trusty hypercall.
4. A post-launched VM with trusty support now gets the return value
-EINVAL (rather than #UD) when it attempts to invoke a non-trusty
hypercall or an invalid hypercall.
v1 -> v2:
- Update documentation that describe hypercall behavior.
- Fix Doxygen warnings
Tracked-On: #5924
Signed-off-by: Junjie Mao <junjie.mao@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Instead of "#include <x86/foo.h>", use "#include <asm/foo.h>".
In other words, we are adopting the same practice in Linux kernel.
Tracked-On: #5920
Signed-off-by: Liang Yi <yi.liang@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Both Windows guest and Linux guest use the MSR MSR_IA32_CSTAR, while
Linux uses it rarely. Now vcpu context switch doesn't save/restore it.
Windows detects the change of the MSR and rises a exception.
Do the save/resotre MSR_IA32_CSTAR during context switch.
Tracked-On: #5899
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
TLFS spec defines that when a VM is created, the value of
HV_X64_MSR_TIME_REF_COUNT is set to zero. Now tsc_offset is not
supported properly, so guest get a drifted reference time.
This patch implements tsc_offset. tsc_scale and tsc_offset
are calculated when a VM is launched and are saved in
struct acrn_hyperv of struct acrn_vm.
Tracked-On: #5956
Signed-off-by: Jian Jun Chen <jian.jun.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
TLFS spec defines that HV_X64_MSR_VP_INDEX and HV_X64_MSR_TIME_REF_COUNT
are read-only MSRs. Any attempt to write to them results in a #GP fault.
Fix the issue by returning error in handler hyperv_wrmsr() of MSRs
HV_X64_MSR_VP_INDEX/HV_X64_MSR_TIME_REF_COUNT emulation.
Tracked-On: #5956
Signed-off-by: Jian Jun Chen <jian.jun.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
TLFS spec defines different hypercall ABIs for X86 and x64. Currently
x64 hypercall interface is not supported well.
Setup the hypercall interface page according to the vcpu mode.
Tracked-On: #5956
Signed-off-by: Jian Jun Chen <jian.jun.chen@intel.com>
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
In order to support platform (such as Ander Lake) which physical address width
bits is 46, the current code need to reserve 2^16 PD page ((2^46) / (2^30)).
This is a complete waste of memory.
This patch would reserve PD page by three parts:
1. DRAM - may take PD_PAGE_NUM(CONFIG_PLATFORM_RAM_SIZE) PD pages at most;
2. low MMIO - may take PD_PAGE_NUM(MEM_1G << 2U) PD pages at most;
3. high MMIO - may takes (CONFIG_MAX_PCI_DEV_NUM * 6U) PD pages (may plus
PDPT entries if its size is larger than 1GB ) at most for:
(a) MMIO BAR size must be a power of 2 from 16 bytes;
(b) MMIO BAR base address must be power of two in size and are aligned with
its size.
Tracked-On: #5929
Signed-off-by: Li Fei1 <fei1.li@intel.com>
We used get_mem_range_info to get the top memory address and then use this address
as the high 64 bits max memory address of SOS. This assumes the platform must have
high memory space.
This patch removes the assumption. It will set high 64 bits max memory address of
SOS to 4G by default (Which means there's no 64 bits high memory), then update
the high 64 bits max memory address if the SOS really has high memory space.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: eddie Dong <eddie.dong@intel.com>
SOS's memory size could be calculated by its vE820 Tables easily.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: eddie Dong <eddie.dong@intel.com>
For platform with HLAT (Hypervisor-managed Linear Address Translation)
capability, the hypervisor shall hide this feature to its guest.
This patch adds MSR_IA32_VMX_PROCBASED_CTLS3 MSR to unsupported MSR
list.
The presence of this MSR is determined by 1-setting of bit 49 of MSR
MSR_IA32_VMX_PROCBASED_CTLS. which is already in unsupported MSR list. [2]
Related documentations:
[1] Intel Architecture Instruction Set Extensions, version Feb 16, 2021,
Ch 6.12
[2] Intel KeyLocker Specification, Sept 2020, Ch 7.2
Tracked-On: #5895
Signed-off-by: Yifan Liu <yifan1.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
sanitize_pte is used to set page table entry to map to an sanitized page to
mitigate l1tf. It should belongs to pgtable module. So move it to pagetable.c
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
lookup_address is used to lookup a pagetable entry by an address. So rename it
to pgtable_lookup_entry to indicate this clearly.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
alloc_page/free_page should been called in pagetable module. In order to do this,
we add pgtable_create_root and pgtable_create_trusty_root to create PML4 page table
page for normal world and secure world.
After this done, no one uses alloc_ept_page. So remove it.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Add pgtable_create_trusty_root to allocate a page for trusty PML4 page table page.
This function also copy PDPT entries from Normal world to Secure world.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Add pgtable_create_root to allocate a page for PMl4 page table page.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Rename mmu_add to pgtable_add_map;
Rename mmu_modify_or_del to pgtable_modify_or_del_map.
And move these functions declaration into pgtable.h
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Requires explicit arch path name in the include directive.
The config scripts was also updated to reflect this change.
Tracked-On: #5825
Signed-off-by: Peter Fang <peter.fang@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Each .c file includes the arch specific irq header file (with full
path) by itself if required.
Tracked-On: #5825
Signed-off-by: Jason Chen CJ <jason.cj.chen@intel.com>
A new x86/guest/virq.h head file now contains all guest
related interrupt handling API.
Tracked-On: #5825
Signed-off-by: Jason Chen CJ <jason.cj.chen@intel.com>
Move the EPT page table related APIs to ept.c. page module only provides APIs to
allocate/free page for page table page. pagetabl module only provides APIs to
add/modify/delete/lookup page table entry. The page pool and the page table
related APIs for EPT should defined in EPT module.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
post_uos_sworld_memory are used for post-launched VM which support trusty.
It's more VM related. So move it definition into vm.c
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Per-core software SRAM L2 cache may be flushed by 'mwait'
extension instruction, which guest VM may execute to enter
core deep sleep. Such kind of flushing is not expected when
software SRAM is enabled for RTVM.
Hypervisor disables MONITOR-WAIT support on both hypervisor
and VMs sides to protect above software SRAM from being flushed.
This patch disable ACRN guest MONITOR-WAIT support if software
SRAM is configured.
Tracked-On: #5649
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Reviewed-by: Fei Li <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Below boolean function are defined in this patch:
- is_software_sram_enabled() to check if SW SRAM
feature is enabled or not.
- set global variable 'is_sw_sram_initialized'
to file static.
Tracked-On: #5649
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Reviewed-by: Fei Li <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
The fields and APIs in old 'struct memory_ops' are used to add/modify/delete
page table (page or entry). So rename 'struct memory_ops' to 'struct pgtable'.
Tracked-On: #5830
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
RTVM is enforced to use 4KB pages to mitigate CVE-2018-12207 and performance jitter,
which may be introduced by splitting large page into 4KB pages on demand. It works
fine in previous hardware platform where the size of address space for the RTVM is
relatively small. However, this is a problem when the platforms support 64 bits
high MMIO space, which could be super large and therefore consumes large # of
EPT page table pages.
This patch optimize it by using large page for purely data pages, such as MMIO spaces,
even for the RTVM.
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Tracked-On: #5788
For FuSa's case, we remove all dynamic memory allocation use in ACRN HV. Instead,
we use static memory allocation or embedded data structure. For pagetable page,
we prefer to use an index (hva for MMU, gpa for EPT) to get a page from a special
page pool. The special page pool should be big enougn for each possible index.
This is not a big problem when we don't support 64 bits MMIO. Without 64 bits MMIO
support, we could use the index to search addrss not larger than DRAM_SIZE + 4G.
However, if ACRN plan to support 64 bits MMIO in SOS, we could not use the static
memory alocation any more. This is because there's a very huge hole between the
top DRAM address and the bottom 64 bits MMIO address. We could not reserve such
many pages for pagetable mapping as the CPU physical address bits may very large.
This patch will use dynamic page allocation for pagetable mapping. We also need
reserve a big enough page pool at first. For HV MMU, we don't use 4K granularity
page table mapping, we need reserve PML4, PDPT and PD pages according the maximum
physical address space (PPT va and pa are identical mapping); For each VM EPT,
we reserve PML4, PDPT and PD pages according to the maximum physical address space
too, (the EPT address sapce can't beyond the physical address space), and we reserve
PT pages by real use cases of DRAM, low MMIO and high MMIO.
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Tracked-On: #5788
memory_ops structure will be changed to store page table related fields.
However, secure world memory base address is not one of them, it's VM
related. So save sworld_memory_base_hva in vm_arch structure directly.
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Tracked-On: #5788
Accessing to software SRAM region is not allowed when
software SRAM is pass-thru to prelaunch RTVM.
This patch removes software SRAM region from service VM
EPT if it is enabled for prelaunch RTVM.
Tracked-On: #5649
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
This patch denies Service VM the access permission to device resources
owned by hypervisor.
HV may own these devices: (1) debug uart pci device for debug version
(2) type 1 pci device if have pre-launched VMs.
Current implementation exposes the mmio/pio resource of HV owned devices
to SOS, should remove them from SOS.
Tracked-On: #5615
Signed-off-by: Tao Yuhong <yuhong.tao@intel.com>
This patch denies Service VM the access permission to device
resources owned by pre-launched VMs.
Rationale:
* Pre-launched VMs in ACRN are independent of service VM,
and should be immune to attacks from service VM. However,
current implementation exposes the bar resource of passthru
devices to service VM for some reason. This makes it possible
for service VM to crash or attack pre-launched VMs.
* It is same for hypervisor owned devices.
NOTE:
* The MMIO spaces pre-allocated to VFs are still presented to
Service VM. The SR-IOV capable devices assigned to pre-launched
VMs doesn't have the SR-IOV capability. So the MMIO address spaces
pre-allocated by BIOS for VFs are not decoded by hardware and
couldn't be enabled by guest. SOS may live with seeing the address
space or not. We will revisit later.
Tracked-On: #5615
Signed-off-by: Tao Yuhong <yuhong.tao@intel.com>
Reviewed-by: Fei Li <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
The logical processor scoped IWKey can be copied to or from a
platform-scope storage copy called IWKeyBackup. Copying IWKey to
IWKeyBackup is called ‘backing up IWKey’ and copying from IWKeyBackup to
IWKey is called ‘restoring IWKey’.
IWKeyBackup and the path between it and IWKey are protected against
software and simple hardware attacks. This means that IWKeyBackup can be
used to distribute an IWKey within the logical processors in a platform
in a protected manner.
Linux keylocker implementation uses this feature, so they are
introduced by this patch.
Tracked-On: #5695
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Different vCPU may have different IWKeys. Hypervisor need do the iwkey
context switch.
This patch introduce a load_iwkey() function to do that. Switches the
host iwkey when the switch_in vCPU satisfies:
1) keylocker feature enabled
2) Different from the current loaded one.
Two opportunities to do the load_iwkey():
1) Guest enables CR4.KL bit.
2) vCPU thread context switch.
load_iwkey() costs ~600 cycles when do the load IWKey action.
Tracked-On: #5695
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
KeyLocker is a new security feature available in new Intel CPUs that
protects data-encryption keys for the Advanced Encryption Standard (AES)
algorithm. These keys are more valuable than what they guard. If stolen
once, the key can be repeatedly used even on another system and even
after vulnerability closed.
It also introduces a CPU-internal wrapping key (IWKey), which is a key-
encryption key to wrap AES keys into handles. While the IWKey is
inaccessible to software, randomizing the value during the boot-time
helps its value unpredictable.
Keylocker usage:
- New “ENCODEKEY” instructions take original key input and returns HANDLE
crypted by an internal wrap key (IWKey, init by “LOADIWKEY” instruction)
- Software can then delete the original key from memory
- Early in boot/software, less likely to have vulnerability that allows
stealing original key
- Later encrypt/decrypt can use the HANDLE through new AES KeyLocker
instructions
- Note:
* Software can use original key without knowing it (use HANDLE)
* HANDLE cannot be used on other systems or after warm/cold reset
* IWKey cannot be read from CPU after it's loaded (this is the
nature of this feature) and only 1 copy of IWKey inside CPU.
The virtualization implementation of Key Locker on ACRN is:
- Each vCPU has a 'struct iwkey' to store its IWKey in struct
acrn_vcpu_arch.
- At initilization, every vCPU is created with a random IWKey.
- Hypervisor traps the execution of LOADIWKEY (by 'LOADIWKEY exiting'
VM-exectuion control) of vCPU to capture and save the IWKey if guest
set a new IWKey. Don't support randomization (emulate CPUID to
disable) of the LOADIWKEY as hypervisor cannot capture and save the
random IWKey. From keylocker spec:
"Note that a VMM may wish to enumerate no support for HW random IWKeys
to the guest (i.e. enumerate CPUID.19H:ECX[1] as 0) as such IWKeys
cannot be easily context switched. A guest ENCODEKEY will return the
type of IWKey used (IWKey.KeySource) and thus will notice if a VMM
virtualized a HW random IWKey with a SW specified IWKey."
- In context_switch_in() of each vCPU, hypervisor loads that vCPU's
IWKey into pCPU by LOADIWKEY instruction.
- There is an assumption that ACRN hypervisor will never use the
KeyLocker feature itself.
This patch implements the vCPU's IWKey management and the next patch
implements host context save/restore IWKey logic.
Tracked-On: #5695
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
In order for a VMM to capture the IWKey values of guests, processors
that support Key Locker also support a new "LOADIWKEY exiting"
VM-execution control in bit 0 of the tertiary processor-based
VM-execution controls.
This patch enables the tertiary VM-execution controls.
Tracked-On: #5695
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
KeyLocker is a new security feature available in new Intel CPUs that
protects data-encryption keys for the Advanced Encryption Standard (AES)
algorithm.
This patch emulates Keylocker CPUID leaf 19H to support Keylocker
feature for guest VM.
To make the hypervisor being able to manage the IWKey correctly, this
patch doesn't expose hardware random IWKey capability
(CPUID.0x19.ECX[1]) to guest VM.
Tracked-On: #5695
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@Intel.com>
Bit19 (CR4_KL) of CR4 is CPU KeyLocker feature enable bit. Hypervisor
traps the bit's writing to track the keylocker feature on/off of guest.
While the bit is set by guest,
- set cr4_kl_enabled to indicate the vcpu's keylocker feature enabled status
- load vcpu's IWKey in host (will add in later patch)
While the bit is clear by guest,
- clear cr4_kl_enabled
This patch trap and passthru the CR4_KL bit to guest for operation.
Tracked-On: #5695
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Current implementation, SOS may allocate the memory region belonging to
hypervisor/pre-launched VM to a post-launched VM. Because it only verifies
the start address rather than the entire memory region.
This patch verifies the validity of the entire memory region before
allocating to a post-launched VM so that the specified memory can only
be allocated to a post-launched VM if the entire memory region is mapped
in SOS’s EPT.
Tracked-On: #5555
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Reviewed-by: Yonghua Huang <yonghua.huang@intel.com>
Currently, we hardcode the GPA base of Software SRAM
to an address that is derived from TGL platform,
as this GPA is identical with HPA for Pre-launch VM,
This hardcoded address may not work on other platforms
if the HPA bases of Software SRAM are different.
Now, Offline tool configures above GPA based on the
detection of Software SRAM on specific platform.
This patch removes the hardcoding GPA of Software SRAM,
and also renames MACRO 'SOFTWARE_SRAM_BASE_GPA' to
'PRE_RTVM_SW_SRAM_BASE_GPA' to avoid confusing, as it
is for Prelaunch VM only.
Tracked-On: #5649
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Physical address to SW SRAM region maybe different
on different platforms, this hardcoded address may
result in address mismatch for SW SRAM operations.
This patch removes above hardcoded address and uses
the physical address parsed from native RTCT.
Tracked-On: #5649
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Reviewed-by: Fei Li <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
'ptcm' and 'ptct' are legacy name according
to the latest TCC spec, hence rename below files
to avoid confusing:
ptcm.c -> rtcm.c
ptcm.h -> rtcm.h
ptct.h -> rtct.h
Tracked-On: #5649
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
When "signal_event" is called, "wait_event" will actually not block.
So it is ok to remove this line.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
Now, we use hash table to maintain intx irq mapping by using
the key generated from sid. So once the entry is added,we can
not update source ide any more. Otherwise, we can't locate the
entry with the key generated from new source ide.
For source id change, remove_remapping/add_remapping is used
instead of update source id directly if entry was added already.
Tracked-On: #5640
Signed-off-by: Yin Fengwei <fengwei.yin@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
This patch move the split-lock logic into dedicated file
to reduce LOC. This may make the logic more clear.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
This patch adds a cache register for VMX_PROC_VM_EXEC_CONTROLS
to avoid the frequent VMCS access.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
The TF is visible to guest which may be modified by
the guest, so it is not a safe method to emulate the
split-lock. While MTF is specifically designed for
single-stepping in x86/Intel hardware virtualization
VT-x technology which is invisible to the guest. Use MTF
to single step the VCPU during the emulation of split lock.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
For a SMP guest, split-lock check may happen on
multiple vCPUs simultaneously. In this case, one
vCPU at most can be allowed running in the
split-lock emulation window. And if the vCPU is
doing the emulation, it should never be blocked
in the hypervisor, it should go back to the guest
to execute the lock instruction immediately and
trap back to the hypervisor with #DB to complete the
split-lock emulation.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
We have trapped the #DB for split-lock emulation.
Only fault exception need RIP being retained.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
xchg may also cause the #AC for split-lock check.
This patch adds this emulation.
1. Kick other vcpus of the guest to stop execution
if the guest has more than one vcpu.
2. Emulate the xchg instruction.
3. Notify other vcpus (if any) to restart execution.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
This patch adds the split-lock emulation.
If a #AC is caused by instruction with LOCK prefix then
emulate it, otherwise, inject it back as it used to be.
1. Kick other vcpus of the guest to stop execution
and set the TF flag to have #DB if the guest has more
than one vcpu.
2. Skip over the LOCK prefix and resume the current
vcpu back to guest for execution.
3. Notify other vcpus to restart exception at the end
of handling the #DB since we have completed
the LOCK prefix instruction emulation.
Tracked-On: #5605
Signed-off-by: Jie Deng <jie.deng@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Check hardware support for all features in CR4,
and hide bits from guest by vcpuid if they're not supported
for guests OS.
Tracked-On: #5586
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
- The current code to virtualize CR0/CR4 is not
well designed, and hard to read.
This patch reshuffle the logic to make it clear
and classify those bits into PASSTHRU,
TRAP_AND_PASSTHRU, TRAP_AND_EMULATE & reserved bits.
Tracked-On: #5586
Signed-off-by: Eddie Dong <eddie.dong@intel.com>
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
While following two styles are both correct, the 2nd one is simpler.
bool is_level_triggered;
1. if (is_level_triggered == true) {...}
2. if (is_level_triggered) {...}
This patch cleans up the style in hypervisor.
Tracked-On: #861
Signed-off-by: Shiqing Gao <shiqing.gao@intel.com>
From SDM Vol.2C - XSETBV instruction description,
If CR4.OSXSAVE[bit 18] = 0,
execute "XSETBV" instruction will generate #UD exception.
From SDM Vol.3C 25.1.1,#UD exception has priority over VM exits,
So if vCPU execute "XSETBV" instruction when CR4.OSXSAVE[bit 18] = 0,
VM exits won't happen.
While hv inject #GP if vCPU execute "XSETBV" instruction
when CR4.OSXSAVE[bit 18] = 0.
It's a wrong behavior, this patch will fix the bug.
Tracked-On: #4020
Signed-off-by: Junming Liu <junming.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
It is possible for more than one vCPUs to trigger shutdown on an RTVM.
We need to avoid entering VM_READY_TO_POWEROFF state again after the
RTVM has been paused or shut down.
Also, make sure an RTVM enters VM_READY_TO_POWEROFF state before it can
be paused.
v1 -> v2:
- rename to poweroff_if_rt_vm for better clarity
Tracked-On: #5411
Signed-off-by: Peter Fang <peter.fang@intel.com>
Currently, ACRN only support shutdown when triple fault happens, because ACRN
doesn't present/emulate a virtual HW, i.e. port IO, to support shutdown. This
patch emulate a virtual shutdown component, and the vACPI method for guest OS
to use.
Pre-launched VM uses ACPI reduced HW mode, intercept the virtual sleep control/status
registers for pre-launched VMs shutdown
Tracked-On: #5411
Signed-off-by: dongshen <dongsheng.x.zhang@intel.com>
Like post-launched VMs, for pre-launched VMs, the ACPI reset register
is also fixed at 0xcf9 and the reset value is 0xE, so pre-launched VMs
now also use ACPI reset register for rebooting.
Tracked-On: #5411
Signed-off-by: dongshen <dongsheng.x.zhang@intel.com>
More than one VM may request shutdown on the same pCPU before
shutdown_vm_from_idle() is called in the idle thread when pCPUs are
shared among VMs.
Use a per-pCPU bitmap to store all the VMIDs requesting shutdown.
v1 -> v2:
- use vm_lock to avoid a race on shutdown
Tracked-On: #5411
Signed-off-by: Peter Fang <peter.fang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Add two Kconfig pSRAM config:
one for whether to enable the pSRAM on the platfrom or not;
another for if the pSRAM is enabled on the platform whether to enable
the pSRAM in the pre-launched RTVM.
If we enable the pSRAM on the platform, we should remove the pSRAM EPT
mapping from the SOS to prevent it could flush the pSRAM cache.
Tracked-On: #5330
Signed-off-by: Qian Wang <qian1.wang@intel.com>
1.Modified the virtual e820 table for pre-launched VM. We added a
segment for pSRAM, and thus lowmem RAM is split into two parts.
Logics are added to deal with the split.
2.Added EPT mapping of pSRAM segment for pre-launched RTVM if it
uses pSRAM.
Tracked-On: #5330
Signed-off-by: Qian Wang <qian1.wang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
pSRAM memory should be cachable. However, it's not a RAM or a normal MMIO,
so we can't use the an exist API to do the EPT mapping and set the EPT cache
attribute to WB for it. Now we assume that SOS must assign the PSRAM area as
a whole and as a separate memory region whose base address is PSRAM_BASE_HPA.
If the hpa of the EPT mapping region is equal to PSRAM_BASE_HPA, we think this
EPT mapping is for pSRAM, we change the EPT mapping cache attribute to WB.
And fix a minor bug when SOS trap out to emulate wbinvd when pSRAM is enabled.
Tracked-On: #5330
Signed-off-by: Qian Wang <qian1.wang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Use ept_flush_leaf_page to emulate guest WBINVD when PTCM is enabled and skip
the pSRAM in ept_flush_leaf_page.
TODO: do we need to emulate WBINVD in HV side.
Tracked-On: #5330
Signed-off-by: Qian Wang <qian1.wang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
According 11.5.1 Cache Control Registers and Bits, Intel SDM Vol 3,
change CR0.CD will not flush cache to insure memory coherency. So
it's not needed to call wbinvd to flush cache in ACRN Hypervisor.
That's what the guest should do.
Tracked-On: #5330
Signed-off-by: Qian Wang <qian1.wang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
support pci-vuart type, and refine:
1.Rename init_vuart() to init_legacy_vuarts(), only init PIO type.
2.Rename deinit_vuart() to deinit_legacy_vuarts(), only deinit PIO type.
3.Move io handler code out of setup_vuart(), into init_legacy_vuarts()
4.add init_pci_vuart(), deinit_pci_vuart, for one pci vuart vdev.
and some change from requirement:
1.Increase MAX_VUART_NUM_PER_VM to 8.
Tracked-On: #5394
Signed-off-by: Tao Yuhong <yuhong.tao@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Reviewed-by: Wang, Yu1 <yu1.wang@intel.com>
This function can be used by other modules instead of hypercall
handling only, hence move it to vlapic.c
Tracked-On: #5407
Signed-off-by: Yonghua Huang <yonghua.huang@intel.com>
Reviewed-by: Li, Fei <fei1.li@intel.com>
Reviewed-by: Wang, Yu1 <yu1.wang@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Now ACRN supports direct boot mode, which could be SBL/ABL, or GRUB boot.
Thus the vboot wrapper layer can be removed and the direct boot functions
don't need to be wrapped in direct_boot.c:
- remove call to init_vboot(), and call e820_alloc_memory() directly at the
time when the trampoline buffer is actually needed.
- Similarly, call CPU_IRQ_ENABLE() instead of the wrapper init_vboot_irq().
- remove get_ap_trampoline_buf(), since the existing function
get_trampoline_start16_paddr() returns the exact same value.
- merge init_general_vm_boot_info() into init_vm_boot_info().
- remove vm_sw_loader pointer, and call direct_boot_sw_loader() directly.
- move get_rsdp_ptr() from vboot_wrapper.c to multiboot.c, and remove the
wrapper over two boot modes.
Tracked-On: #5197
Signed-off-by: Zide Chen <zide.chen@intel.com>
The commit of da81a0041d
"HV: add e820 ACPI entry for pre-launched VM" introduced a issue that the
base_hpa and remaining_hpa_size are also calculated on the entry of 32bit
PCI hole which from 0x80000000 to 0xffffffff, which is incorrect;
Tracked-On: #5266
Signed-off-by: Victor Sun <victor.sun@intel.com>
Per PCI Firmware Specification Revision 3.0, 4.1.2. MCFG Table Description:
Memory Mapped Enhanced Configuration Space Base Address Allocation Structure
assign the Start Bus Number and the End Bus Number which could decoded by the
Host Bridge. We should not access the PCI device which bus number outside of
the range of [Start Bus Number, End Bus Number).
For ACRN, we should:
1. Don't detect PCI device which bus number outside the range of
[Start Bus Number, End Bus Number) of MCFG ACPI Table.
2. Only trap the ECAM MMIO size: [MMCFG_BASE_ADDRESS, MMCFG_BASE_ADDRESS +
(End Bus Number - Start Bus Number + 1) * 0x100000) for SOS.
Tracked-On: #5233
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Previously we use a pre-defined structure as vACPI table for pre-launched
VM, the structure is initialized by HV code. Now change the method to use a
pre-loaded multiboot module instead. The module file will be generated by
acrn-config tool and loaded to GPA 0x7ff00000, a hardcoded RSDP table at
GPA 0x000f2400 will point to the XSDT table which at GPA 0x7ff00080;
Tracked-On: #5266
Signed-off-by: Victor Sun <victor.sun@intel.com>
Signed-off-by: Shuang Zheng <shuang.zheng@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Previously the ACPI table was stored in F segment which might not be big
enough for a customized ACPI table, hence reserve 1MB space in pre-launched
VM e820 table to store the ACPI related data:
0x7ff00000 ~ 0x7ffeffff : ACPI Reclaim memory
0x7fff0000 ~ 0x7fffffff : ACPI NVS memory
Tracked-On: #5266
Signed-off-by: Victor Sun <victor.sun@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
When HV pass through the P2SB MMIO device to pre-launched VM, vgpio
device model traps MMIO access to the GPIO registers within P2SB so
that it can expose virtual IOAPIC pins to the VM in accordance with
the programmed mappings between gsi and vgsi.
Tracked-On: #5246
Signed-off-by: Toshiki Nishioka <toshiki.nishioka@intel.com>
Reviewed-by: Junjie Mao <junjie.mao@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Add the capability of forwarding specified physical IOAPIC interrupt
lines to pre-launched VMs as virtual IOAPIC interrupts. This is for the
sake of the certain MMIO pass-thru devices on EHL CRB which can support
only INTx interrupts.
Tracked-On: #5245
Signed-off-by: Toshiki Nishioka <toshiki.nishioka@intel.com>
Reviewed-by: Junjie Mao <junjie.mao@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Add HC_CREATE_VDEV and HC_DESTROY_VDEV two hypercalls that are used to
create and destroy an emulated device(PCI device or legacy device) in hypervisor
v3: 1) change HC_CREATE_DEVICE and HC_DESTROY_DEVICE to HC_CREATE_VDEV
and HC_DESTROY_VDEV
2) refine code style
v4: 1) remove unnecessary parameter
2) add VM state check for HC_CREATE_VDEV and HC_DESTROY hypercalls
Tracked-On: #4853
Reviewed-by: Wang, Yu1 <yu1.wang@intel.com>
Signed-off-by: Yuan Liu <yuan1.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
-- use an array to fast locate the hypercall handler
to replace switch case.
-- uniform hypercall handler as below:
int32_t (*handler)(sos_vm, target_vm, param1, param2)
Tracked-On: #4958
Signed-off-by: Mingqiang Chi <mingqiang.chi@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Reviewed-by: Eddie Dong <eddie.dong@intel.com>
Fix the bug for "is_apl_platform" func.
"monitor_cap_buggy" is identical to "is_apl_platform", so remove it.
On apl platform:
1) ACRN doesn't use monitor/mwait instructions
2) ACRN disable GPU IOMMU
Tracked-On:#3675
Signed-off-by: Junming Liu <junming.liu@intel.com>
-- move vm_state_lock to other place in vm structure
to avoid the memory waste because of the page-aligned.
-- remove the memset from create_vm
-- explicitly set max_emul_mmio_regions and vcpuid_entry_nr to 0
inside create_vm to avoid use without initialization.
-- rename max_emul_mmio_regions to nr_emul_mmio_regions
v1->v2:
add deinit_emul_io in shutdown_vm
Tracked-On: #4958
Signed-off-by: Mingqiang Chi <mingqiang.chi@intel.com>
Reviewed-by: Grandhi, Sainath <sainath.grandhi@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Previously the CPU affinity of SOS VM is initialized at runtime during
sanitize_vm_config() stage, follow the policy that all physical CPUs
except ocuppied by Pre-launched VMs are all belong to SOS_VM. Now change
the process that SOS CPU affinity should be initialized at build time
and has the assumption that its validity is guarenteed before runtime.
Tracked-On: #5077
Signed-off-by: Victor Sun <victor.sun@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
To hide CET feature from guest VM completely, the MSR IA32_MSR_XSS also
need to be intercepted because it comprises CET_U and CET_S feature bits
of xsave/xstors operations. Mask these two bits in IA32_MSR_XSS writing.
With IA32_MSR_XSS interception, member 'xss' of 'struct ext_context' can
be removed because it is duplicated with the MSR store array
'vcpu->arch.guest_msrs[]'.
Tracked-On: #5074
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Return-oriented programming (ROP), and similarly CALL/JMP-oriented
programming (COP/JOP), have been the prevalent attack methodologies for
stealth exploit writers targeting vulnerabilities in programs.
CET (Control-flow Enforcement Technology) provides the following
capabilities to defend against ROP/COP/JOP style control-flow subversion
attacks:
* Shadow stack: Return address protection to defend against ROP.
* Indirect branch tracking: Free branch protection to defend against
COP/JOP
The full support of CET for Linux kernel has not been merged yet. As the
first stage, hide CET from guest VM.
Tracked-On: #5074
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Reviewed-by: Jason Chen CJ <jason.cj.chen@intel.com>
Add mmio device pass through support for pre-launched VM.
When we pass through a MMIO device to pre-launched VM, we would remove its
resource from the SOS. Now these resources only include the MMIO regions.
Tracked-On: #5053
Acked-by: Eddie Dong <eddie.dong@intel.com>
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Add two hypercalls to support MMIO device pass through for post-launched VM.
And when we support MMIO pass through for pre-launched VM, we could re-use
the code in mmio_dev.c
Tracked-On: #5053
Signed-off-by: Li Fei1 <fei1.li@intel.com>
During context switch in hypervisor, xsave/xrstore are used to
save/resotre the XSAVE area according to the XCR0 and XSS. The legacy
region in XSAVE area include FPU and SSE, we should make sure the
legacy region be saved during contex switch. FPU in XCR0 is always
enabled according to SDM.
For SSE, we enable it in XCR0 during context switch.
Tracked-On: #5062
Signed-off-by: Conghui Chen <conghui.chen@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
kick_thread function is only used by kick_vcpu to kick vcpu out of
non-root mode, the implementation in it is sending IPI to target CPU if
target obj is running and target PCPU is not current one; while for
runnable obj, it will just make reschedule request. So the kick_thread
is not actually belong to scheduler module, we can drop it and just do
the cpu notification in kick_vcpu.
Tracked-On: #5057
Signed-off-by: Conghui Chen <conghui.chen@intel.com>
Reviewed-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
vcpu->running is duplicated with THREAD_STS_RUNNING status of thread
object. Introduce an API sleep_thread_sync(), which can utilize the
inner status of thread object, to do the sync sleep for zombie_vcpu().
Tracked-On: #5057
Signed-off-by: Conghui Chen <conghui.chen@intel.com>
Reviewed-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
-- replace global hypercall lock with per-vm lock
-- add spinlock protection for vm & vcpu state change
v1-->v2:
change get_vm_lock/put_vm_lock parameter from vm_id to vm
move lock obtain before vm state check
move all lock from vmcall.c to hypercall.c
Tracked-On: #4958
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>
Some OSes assume the platform must have the IOAPIC. For example:
Linux Kernel allocates IRQ force from GSI (0 if there's no PIC and IOAPIC) on x86.
And it thinks IRQ 0 is an architecture special IRQ, not for device driver. As a
result, the device driver may goes wrong if the allocated IRQ is 0 for RTVM.
This patch expose vIOAPIC to RTVM with LAPIC passthru even though the RTVM can't
use IOAPIC, it servers as a place holder to fullfil the guest assumption.
After vIOAPIC has exposed to guest unconditionally, the 'ready' field could be
removed since we do vIOAPIC initialization for each guest.
Tracked-On: #4691
Signed-off-by: Li Fei1 <fei1.li@intel.com>
Acked-by: Eddie Dong <eddie.dong@intel.com>
Currently, not all platforms support posted interrupt processing of both
VT-x and VT-d. On EHL, VT-d doesn't support posted interrupt processing.
So in such scenario, is_pi_capable() in vcpu_handle_pi_notification()
will bypass the PIR pending bits check which might cause a self-NV-IPI
lost.
With commit "bf1ff8c98 (hv: Offload syncing PIR to vIRR to processor
hardware)", the syncing PIR to vIRR is postponed and it is handled by a
self-NV-IPI in the following VMEnter. The process looks like,
a) vcpu A accepts a virtual interrupt ->
1) ACRN_REQUEST_EVENT is set
2) corresponding bit in PIR is set
3) Posted Interrupt ON bit is set
b) vcpu A does virtual interrupt injection on resume path due to
the pending ACRN_REQUEST_EVENT ->
1) hypervisor disables host interrupt
2) ACRN_REQUEST_EVENT is cleared
3) a self-NV-IPI is sent via ICR of LAPIC.
4) IRR bit of the self-NV-IPI is set
c) (VM-ENTRY) vcpu A returns into non-root mode
1) host interrupt enable(by HW)
2) posted interrupt processing clears the ON bit, sync PIR to vIRR
3) deliver the virtual interrupt if guest rflags.IF=1
d) (VM-EXIT) vcpu A traps due to a instruction execution (e.g. HLT)
1) host interrupt disable(by HW)
2) hypervisor enable host interrupt
Above illustrates a normal process of the virtual interrupt injection
with cpu PI support. However, a failing case is observed. The failing
case is that the self-NV-IPI from b-3 is not accepted by the core until
a timing between d-1 and d-2. b-4 happening between d-1 and d-2 is
observed by debug trace. So the self-NV-IPI will be handled in root-mode
which cannot do the syncing PIR to vIRR processing. Due to the bug
described in the first paragraph, vcpu_handle_pi_notification() cannot
succeed the virtual interrupt injection request. This patch fix it by
removing the wrong check in vcpu_handle_pi_notification() because
vcpu_handle_pi_notification() only happens on platform with cpu PI
support.
Here are some cost data for sending IPI via LAPIC ICR regsiter.
Normally, the cycles between ICR write and IRR got set is 140~260,
which is not accurate due to the MSR read overhead.
And from b-3 to c is about 560 cycles. So b-4 happens during this
period. But in bad case, b-4 doesn't happen even c is triggered.
The worse case i captured is that ICR write and IRR got set costs more
than 1900 cycles. Now, the best GUESS of the huge cost of IPI via ICR is
the ACPI bus arbitration(refer to SDM 10.6.3, 10.7 and Figure 10-17).
Tracked-On: #4937
Signed-off-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by： Eddie Dong <eddie.dong@intel.com>
Tao Yuhong