[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] [PATCH v5 rebased 0/4] nested vmx: enable VMCS shadowing feature



Changes from v5 to v5 rebased:
 - Rebase the patches to latest staging tree.

Changes from v4 to v5:
 - Change the size for vvmcs buffer from 500 to 100.
 - Store page_info in per-pCPU structure instead of MFN in per-vCPU structure 
for vmread/vmwrite bitmaps.
 - Define correct data type to avoid force type conversion.

Changes from v3 to v4:
 - Allocate a buffer in initialization time for data exchange instead of 
runtime allocation (reflect Jun's comment).
 - Move the vmread/vmwrite bitmap logic into nvmx_vcpu_initialise(), also track 
MFN instead of virtual address in the vcpu global data. (reflect Jan's comment).

Changes from v2 to v3:
 - Use pfn_to_paddr() to get the address from frame number instead of doing 
shift directly.
 - Remove some unnecessary initialization code and add "static" to 
vmentry_fields and gpdptr_fields.
 - Enable the VMREAD/VMWRITE bitmap only if nested hvm is enabled.
 - Use clear_page() to set all 0 to the page instead of memset().
 - Use domheap to allocate the VMREAD/VMWRITE bitmap instead of xenheap.

Changes from v1 to v2:
 - Use domain_page_map_to_mfn() instead of virt_to_mfn() to get the machine 
frame.
 - Address other comments from Jan, including some programming format, return 
value type, etc.

Latest Intel SDM introduced a new feature "VMCS shadowing" at bit 14 in
Secondary Processor-Based VM-Execution Controls for nested virtualization.

The main purpose of this feature is to reduce or eliminate the number of VM exit
for non-root VMREAD and VMWRITE. It provides the capability to link a
"virtual VMCS" with the current running VMCS, so that after VM entry, the
non-root VMREAD and VMWRITE can get/set related data directly from/to the
"virtual VMCS" without trap and emulation.

A separate bitmap is introduced for VMREAD and VMWRITE, from which hypervisor
can control whether VMREAD/VMWRITE from/to certain VMCS field will trigger
VM exit or directly get/set data by hardware.

With the new feature introduced, all the in "virtual VMCS" need to be operated
by VMREAD and VMWRITE because this VMCS will also be loaded into hardware.
This requires the capability to VMWRITE all the VMCS fields, including those
readonly ones. Intel SDM introduces this functionality at bit 29 in
IA32_VMX_MISC MSR.

For details, please refer to:
http://www.intel.com/content/www/us/en/processors/architectures-software-developer-manuals.html

Thanks,
Dongxiao

Dongxiao Xu (4):
  nested vmx: Use a list to store the launched vvmcs for L1 VMM
  nested vmx: use VMREAD/VMWRITE to construct vVMCS if enabled VMCS
    shadowing
  nested vmx: optimize for bulk access of virtual VMCS
  nested vmx: enable VMCS shadowing feature

 xen/arch/x86/hvm/vmx/vmcs.c        |   48 ++++++
 xen/arch/x86/hvm/vmx/vvmx.c        |  324 ++++++++++++++++++++++++++++++++----
 xen/include/asm-x86/hvm/vmx/vmcs.h |   23 +++-
 xen/include/asm-x86/hvm/vmx/vvmx.h |   24 +++-
 4 files changed, 383 insertions(+), 36 deletions(-)


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.