svm: implement new DecodeAssist feature
cover svm: implement new DecodeAssist feature
Hi,
newer AMD SVM implementations (starting with Bulldozer) have some additions to SVM called DecodeAssist. These bits improve and simplify the handling of some intercepts, now more information is provided in the VMCB already and saves us from decoding this information "manually". These new features are described in recent editions of the APM Vol.2 manual, chapter 15.33.
The first patch moves some code from VMX into HVM to be later used by SVM. The second patch adds the newly defined bits to VMCB and CPUID. Patch 3 to 5 implement three parts of the feature set, more details in the specific emails.
Please comment.
All patches:
Regards,
Andre.
Signature · 2 lines
Andre Przywara AMD-Operating System Research Center (OSRC), Dresden, Germany
PATCH 1/5 vmx/hvm: move mov-cr handling functions to generic HVM code
Currently the handling of CR accesses intercepts is done much differently in SVM and VMX. For future usage move the VMX part into the generic HVM path and use the exported functions.
commit 48aec1e4591449768193b0221c7304f56a3dbe5c
Author: Andre Przywara <[email protected]> Date: Sun Jul 25 13:56:25 2010 +0200
vmx/hvm: move mov-cr handling functions to generic HVM code
Currently the handling of CR accesses intercepts is done much differently in SVM and VMX. For future usage move the VMX part into the generic HVM path and use the exported functions.
xen/arch/x86/hvm/hvm.c +147 -0
@@ -1409,6 +1409,74 @@ static void hvm_set_uc_mode(struct vcpu *v, bool_t is_in_uc_mode)return hvm_funcs.set_uc_mode (v); } +#ifdef __i386__ +#define GPRNAME (suffix) e ## suffix +#else +#define GPRNAME (suffix) r ## suffix +#endif + +#define GPR_GET_CASE (num, name, val) case num: val = regs->GPRNAME(name);break +#define GPR_SET_CASE (num, name, val) case num: regs->GPRNAME(name) = val;break + +#define GPR_CASES (GS , value) \ + GPR_ ## GS ## ET_CASE(0, ax, value); \ + GPR_ ## GS ## ET_CASE(1, cx, value); \ + GPR_ ## GS ## ET_CASE(2, dx, value); \ + GPR_ ## GS ## ET_CASE(3, bx, value); \ + GPR_ ## GS ## ET_CASE(4, sp, value); \ + GPR_ ## GS ## ET_CASE(5, bp, value); \ + GPR_ ## GS ## ET_CASE(6, si, value); \ + GPR_ ## GS ## ET_CASE(7, di, value); + +#define GPR_64_CASES (GS , value) \ + GPR_ ## GS ## ET_CASE(8, 8, value); \ + GPR_ ## GS ## ET_CASE(9, 9, value); \ + GPR_ ## GS ## ET_CASE(10, 10, value); \ + GPR_ ## GS ## ET_CASE(11, 11, value); \ + GPR_ ## GS ## ET_CASE(12, 12, value); \ + GPR_ ## GS ## ET_CASE(13, 13, value); \ + GPR_ ## GS ## ET_CASE(14, 14, value); \ + GPR_ ## GS ## ET_CASE(15, 15, value); + +int hvm_get_gpr (struct cpu_user_regs *regs,int gpr, +unsigned long *valueptr) +{ +uint64_t mask; +unsigned long value; +int width =0 ; + +if (width ==0 ) { + mask = (unsigned long )-1 ; + }else { +if (width >=64 ) + mask = (uint64_t )-1 ; +else + mask = (1ULL << width) -1 ; + } + +switch (gpr) { + GPR_CASES(G, value) +#ifndef __i386__ + GPR_64_CASES(G, value) +#endif +default :return 1 ; + } + *valueptr = value & mask; +return 0 ; +} + +int hvm_set_gpr (struct cpu_user_regs *regs,int gpr,unsigned long value) +{ +switch (gpr) { + GPR_CASES(S, value) +#ifndef __i386__ + GPR_64_CASES(S, value) +#endif +default :return 1 ; + } +return 0 ; +} +int hvm_set_cr0 (unsigned long value) {struct vcpu *v = current;
@@ -4082,6 +4150,85 @@ enum hvm_intblk nhvm_interrupt_blocked(struct vcpu *v)return hvm_funcs.nhvm_intr_blocked (v); } +int hvm_mov_to_cr (int gp,int cr,struct cpu_user_regs *regs) +{ +unsigned long value; +struct vcpu *v = current; +struct vlapic *vlapic =vcpu_vlapic (v); + +if (hvm_get_gpr (regs, gp, &value)) { +gdprintk (XENLOG_ERR ,"invalid gp: %d\n" , gp); + goto exit_and_crash; + } + +HVMTRACE_LONG_2D (CR_WRITE , cr,TRC_PAR_LONG (value)); + +HVM_DBG_LOG (DBG_LEVEL_1 ,"CR%d, value = %lx" , cr, value); + +switch ( cr ) + { +case 0 : +return !hvm_set_cr0 (value); + +case 3 : +return !hvm_set_cr3 (value); + +case 4 : +return !hvm_set_cr4 (value); + +case 8 : +vlapic_set_reg (vlapic,APIC_TASKPRI , ((value &0x0F ) <<4 )); +break ; + +default : +gdprintk (XENLOG_ERR ,"invalid cr: %d\n" , cr); + goto exit_and_crash; + } + +return 1 ; + + exit_and_crash: +domain_crash (v->domain); +return 0 ; +} + +/* +* Read from control registers. CR0 and CR4 are handled before. +*/ +void hvm_mov_from_cr (int cr,int gp,struct cpu_user_regs *regs) +{ +unsigned long value =0 ; +struct vcpu *v = current; +struct vlapic *vlapic =vcpu_vlapic (v); + +switch ( cr ) + { +case 0 : +case 2 : +case 3 : +case 4 : + value = (unsigned long )v->arch.hvm_vcpu.guest_cr[cr]; +break ; +case 8 : + value = (unsigned long )vlapic_get_reg (vlapic,APIC_TASKPRI ); + value = (value &0xF0 ) >>4 ; +break ; +default : +gdprintk (XENLOG_ERR ,"invalid cr: %d\n" , cr); +domain_crash (v->domain); +break ; + } + +if (hvm_set_gpr (regs, gp, value) !=0 ) { +printk ("invalid gp: %d\n" , gp); +domain_crash (v->domain); + } + +HVMTRACE_LONG_2D (CR_READ , cr,TRC_PAR_LONG (value)); + +HVM_DBG_LOG (DBG_LEVEL_VMMU ,"CR%d, value = %lx" , cr, value); +} +/* * Local variables: * mode: C
xen/arch/x86/hvm/vmx/vmx.c +2 -138
155 lines not on this page · raw message
xen/include/asm-x86/hvm/hvm.h +19 -0
@@ -311,6 +311,23 @@ static inline int hvm_do_pmu_interrupt(struct cpu_user_regs *regs) X86_CR4_OSFXSR | X86_CR4_OSXMMEXCPT | \ (xsave_enabled(_v) ? X86_CR4_OSXSAVE : 0)))) +#define HVM_GPR_NR_EAX 0 +#define HVM_GPR_NR_ECX 1 +#define HVM_GPR_NR_EDX 2 +#define HVM_GPR_NR_EBX 3 +#define HVM_GPR_NR_ESP 4 +#define HVM_GPR_NR_EBP 5 +#define HVM_GPR_NR_ESI 6 +#define HVM_GPR_NR_EDI 7 +#define HVM_GPR_NR_R8 8 +#define HVM_GPR_NR_R9 9 +#define HVM_GPR_NR_R10 10 +#define HVM_GPR_NR_R11 11 +#define HVM_GPR_NR_R12 12 +#define HVM_GPR_NR_R13 13 +#define HVM_GPR_NR_R14 14 +#define HVM_GPR_NR_R15 15 +/* These exceptions must always be intercepted. */ #define HVM_TRAP_MASK ((1U << TRAP_machine_check) | (1U << TRAP_invalid_op))
@@ -389,6 +406,8 @@ int hvm_hap_nested_page_fault(unsigned long gpa,int hvm_x2apic_msr_read (struct vcpu *v,unsigned int msr,uint64_t *msr_content);int hvm_x2apic_msr_write (struct vcpu *v,unsigned int msr,uint64_t msr_content); +int hvm_mov_to_cr (int gp,int cr,struct cpu_user_regs *regs); +void hvm_mov_from_cr (int cr,int gp,struct cpu_user_regs *regs);#ifdef __x86_64__/* Called for current VCPU on crX changes by guest */
xen/include/asm-x86/hvm/support.h +2 -0
@@ -129,6 +129,8 @@ void hvm_triple_fault(void);void hvm_rdtsc_intercept (struct cpu_user_regs *regs);int hvm_handle_xsetbv (u64 new_bv); +int hvm_get_gpr (struct cpu_user_regs *regs,int gpr,unsigned long *valueptr); +int hvm_set_gpr (struct cpu_user_regs *regs,int gpr,unsigned long value);/* These functions all return X86EMUL return codes. */ int hvm_set_efer (uint64_t value);
xen/include/asm-x86/hvm/vmx/vmx.h +0 -16
@@ -153,22 +153,6 @@ void vmx_update_cpu_exec_control(struct vcpu *v);#define VMX_CONTROL_REG_ACCESS_TYPE_MOV_FROM_CR (1 << 4)#define VMX_CONTROL_REG_ACCESS_TYPE_CLTS (2 << 4)#define VMX_CONTROL_REG_ACCESS_TYPE_LMSW (3 << 4) -#define VMX_CONTROL_REG_ACCESS_GPR_EAX (0 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_ECX (1 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_EDX (2 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_EBX (3 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_ESP (4 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_EBP (5 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_ESI (6 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_EDI (7 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R8 (8 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R9 (9 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R10 (10 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R11 (11 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R12 (12 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R13 (13 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R14 (14 << 8) -#define VMX_CONTROL_REG_ACCESS_GPR_R15 (15 << 8)/* * Access Rights
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
PATCH 2/5 svm: add bit definitions for SVM DecodeAssist
In reply to Andre PrzywaraChapter 15.33 of recent APM Vol.2 manuals describe some additions to SVM called DecodeAssist. Add the newly added fields to the VMCB structure and name the associated CPUID bit.
commit c3a6d77443d2d074b5b553ac4167f4b59f9b6408
Author: Andre Przywara <[email protected]> Date: Sun Jul 25 14:09:35 2010 +0200
svm: add bit definitions for SVM DecodeAssist
Chapter 15.33 of recent APM Vol.2 manuals describe some additions to SVM called DecodeAssist. Add the newly added fields to the VMCB structure and name the associated CPUID bit.
xen/arch/x86/hvm/svm/svm.c +1 -0
@@ -1000,6 +1000,7 @@ struct hvm_function_table * __init start_svm(void)P (cpu_has_svm_lbrv,"Last Branch Record (LBR) Virtualisation" );P (cpu_has_svm_nrips,"Next-RIP Saved on #VMEXIT" );P (cpu_has_svm_cleanbits,"VMCB Clean Bits" ); +P (cpu_has_svm_decode,"DecodeAssists" );P (cpu_has_pause_filter,"Pause-Intercept Filter" );#undef P
xen/include/asm-x86/hvm/svm/svm.h +1 -0
@@ -90,6 +90,7 @@ extern u32 svm_feature_flags;#define cpu_has_svm_svml cpu_has_svm_feature(SVM_FEATURE_SVML)#define cpu_has_svm_nrips cpu_has_svm_feature(SVM_FEATURE_NRIPS)#define cpu_has_svm_cleanbits cpu_has_svm_feature(SVM_FEATURE_VMCBCLEAN) +#define cpu_has_svm_decode cpu_has_svm_feature(SVM_FEATURE_DECODEASSISTS)#define cpu_has_pause_filter cpu_has_svm_feature(SVM_FEATURE_PAUSEFILTER)#define SVM_PAUSEFILTER_INIT 3000
xen/include/asm-x86/hvm/svm/vmcb.h +3 -1
@@ -435,7 +435,9 @@ struct vmcb_struct {vmcbcleanbits_t cleanbits;/* offset 0xC0 */ u32 res09;/* offset 0xC4 */ u64 nextrip;/* offset 0xC8 */ -u64 res10a[102 ];/* offset 0xD0 pad to save area */ +u8 guest_ins_len;/* offset 0xD0 */ +u8 guest_ins[15 ];/* offset 0xD1 */ +u64 res10a[100 ];/* offset 0xE0 pad to save area */ svm_segment_register_t es;/* offset 1024 - cleanbit 8 */ svm_segment_register_t cs;/* cleanbit 8 */
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
PATCH 3/5 svm: implement instruction fetch part of DecodeAssist
In reply to Andre PrzywaraNewer SVM implementations (Bulldozer) copy up to 15 bytes from the instruction stream into the VMCB when a #PF or #NPF exception is intercepted. This patch makes use of this information if available. This saves us from a) traversing the guest's page tables, b) mapping the guest's memory and c) copy the instructions from there into the hypervisor's address space.
This speeds up #NPF intercepts quite a lot and avoids cache and TLB trashing.
commit 242318e2ae4229ba3ac6a5253f3f4722133348a6
Author: Andre Przywara <[email protected]> Date: Sun Jul 25 14:33:21 2010 +0200
svm: implement instruction fetch part of DecodeAssist
Newer SVM implementations (Bulldozer) copy up to 15 bytes from the instruction stream into the VMCB when a #PF or #NPF exception is intercepted. This patch makes use of this information if available. This saves us from a) traversing the guest's page tables, b) mapping the guest's memory and c) copy the instructions from there into the hypervisor's address space.
This speeds up #NPF intercepts quite a lot and avoids cache and TLB trashing.
xen/arch/x86/hvm/emulate.c +16 -9
@@ -995,15 +995,22 @@ int hvm_emulate_one( pfec |= PFEC_user_mode; hvmemul_ctxt->insn_buf_eip = regs->eip; - hvmemul_ctxt->insn_buf_bytes = - (hvm_virtual_to_linear_addr ( - x86_seg_cs, &hvmemul_ctxt->seg_reg[x86_seg_cs], - regs->eip,sizeof (hvmemul_ctxt->insn_buf), - hvm_access_insn_fetch, hvmemul_ctxt->ctxt.addr_size, &addr) && - !hvm_fetch_from_guest_virt_nofault ( - hvmemul_ctxt->insn_buf, addr, -sizeof (hvmemul_ctxt->insn_buf), pfec)) - ?sizeof (hvmemul_ctxt->insn_buf) :0 ; +if (curr->arch.hvm_vcpu.guest_ins_len ==0 ) { + hvmemul_ctxt->insn_buf_bytes = + (hvm_virtual_to_linear_addr ( + x86_seg_cs, &hvmemul_ctxt->seg_reg[x86_seg_cs], + regs->eip,sizeof (hvmemul_ctxt->insn_buf), + hvm_access_insn_fetch, hvmemul_ctxt->ctxt.addr_size, &addr) && + !hvm_fetch_from_guest_virt_nofault ( + hvmemul_ctxt->insn_buf, addr, +sizeof (hvmemul_ctxt->insn_buf), pfec)) + ?sizeof (hvmemul_ctxt->insn_buf) :0 ; + }else { + hvmemul_ctxt->insn_buf_bytes = curr->arch.hvm_vcpu.guest_ins_len; +memcpy (hvmemul_ctxt->insn_buf, curr->arch.hvm_vcpu.guest_ins, + hvmemul_ctxt->insn_buf_bytes); + curr->arch.hvm_vcpu.guest_ins_len =0 ; + } hvmemul_ctxt->exn_pending =0 ;
xen/arch/x86/hvm/svm/svm.c +16 -0
@@ -1589,6 +1589,18 @@ static void svm_invlpg_intercept(unsigned long vaddr)svm_asid_g_invlpg (curr, vaddr); } +static void svm_set_instruction_bytes (struct vmcb_struct *vmcb) +{ +struct vcpu *curr = current; +int len = vmcb->guest_ins_len &0x0F ; + +if (len ==0 ) +return ; + curr->arch.hvm_vcpu.guest_ins_len = len; + curr->arch.hvm_vcpu.guest_ins = vmcb->guest_ins; +return ; +} +static struct hvm_function_table __read_mostly svm_function_table = { .name ="SVM" , .cpu_up_prepare = svm_cpu_up_prepare,
@@ -1801,6 +1813,7 @@ asmlinkage void svm_vmexit_handler(struct cpu_user_regs *regs)unsigned long va; va = vmcb->exitinfo2; regs->error_code = vmcb->exitinfo1; +svm_set_instruction_bytes (vmcb); HVM_DBG_LOG(DBG_LEVEL_VMMU, "eax=%lx, ebx=%lx, ecx=%lx, edx=%lx, esi=%lx, edi=%lx", (unsigned long)regs->eax, (unsigned long)regs->ebx,
@@ -1809,6 +1822,7 @@ asmlinkage void svm_vmexit_handler(struct cpu_user_regs *regs)if (paging_fault (va, regs) ) { + v->arch.hvm_vcpu.guest_ins_len =0 ;if (trace_will_trace_event (TRC_SHADOW ) )break ;if (hvm_long_mode_enabled (v) )
@@ -1817,6 +1831,7 @@ asmlinkage void svm_vmexit_handler(struct cpu_user_regs *regs)HVMTRACE_2D (PF_XEN , regs->error_code, va);break ; } + v->arch.hvm_vcpu.guest_ins_len =0 ;hvm_inject_exception (TRAP_page_fault, regs->error_code, va);break ;
@@ -1968,6 +1983,7 @@ asmlinkage void svm_vmexit_handler(struct cpu_user_regs *regs)case VMEXIT_NPF :perfc_incra (svmexits,VMEXIT_NPF_PERFC ); regs->error_code = vmcb->exitinfo1; +svm_set_instruction_bytes (vmcb);svm_do_nested_pgfault (v, regs, vmcb->exitinfo2);break ;
xen/include/asm-x86/hvm/vcpu.h +3 -0
@@ -140,6 +140,9 @@ struct hvm_vcpu {unsigned long mmio_gva;unsigned long mmio_gpfn; +uint8_t guest_ins_len; +void *guest_ins; +/* Callback into x86_emulate when emulating FPU/MMX/XMM instructions. */ void (*fpu_exception_callback)(void *,struct cpu_user_regs *);void *fpu_exception_callback_arg;
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
PATCH 4/5 svm: implement CR access part of DecodeAssist
In reply to Andre PrzywaraNewer SVM implementations (Bulldozer) now give the used general purpose register on a MOV-CR intercept explictly. This avoids fetching and decoding the instruction from guest's memory and speeds up some Windows guest, which exercise CR8 quite often.
commit 11fcadf1400313770480d6d5c4edcc1472c77b0c
Author: Andre Przywara <[email protected]> Date: Sun Jul 25 15:07:35 2010 +0200
svm: implement CR access part of DecodeAssist
Newer SVM implementations (Bulldozer) now give the used general purpose register on a MOV-CR intercept explictly. This avoids fetching and decoding the instruction from guest's memory and speeds up some Windows guest, which exercise CR8 quite often.
xen/arch/x86/hvm/svm/svm.c +29 -1
@@ -1130,6 +1130,25 @@ static void svm_vmexit_do_cpuid(struct cpu_user_regs *regs)__update_guest_eip (regs, inst_len); } +static int svm_vmexit_do_cr_access (struct vmcb_struct *vmcb, +struct cpu_user_regs *regs) +{ +int gp, cr, dir; + + cr = vmcb->exitcode -VMEXIT_CR0_READ ; + dir = (cr >15 ); + cr = cr %16 ; + gp = vmcb->exitinfo1 &0x0F ; + +if (dir) +hvm_mov_to_cr (gp, cr, regs); +else +hvm_mov_from_cr (cr, gp, regs); + +__update_guest_eip (regs, vmcb->nextrip - vmcb->rip); +return 0 ; +} +static void svm_dr_access (struct vcpu *v,struct cpu_user_regs *regs) {HVMTRACE_0D (DR_WRITE );
@@ -1906,9 +1925,18 @@ asmlinkage void svm_vmexit_handler(struct cpu_user_regs *regs)__update_guest_eip (regs, vmcb->exitinfo2 - vmcb->rip);break ; } -/* fallthrough to emulation if a string instruction */ +if ( !handle_mmio () ) +hvm_inject_exception (TRAP_gp_fault,0 ,0 ); +break ; + case VMEXIT_CR0_READ ... VMEXIT_CR15_READ: case VMEXIT_CR0_WRITE ... VMEXIT_CR15_WRITE: +if ((vmcb->exitinfo1 & (1ULL <<63 ) && +svm_vmexit_do_cr_access (vmcb, regs)) || + !handle_mmio () ) +hvm_inject_exception (TRAP_gp_fault,0 ,0 ); +break ; +case VMEXIT_INVLPG :case VMEXIT_INVLPGA :if ( !handle_mmio () )
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
PATCH 5/5 svm: implement INVLPG part of DecodeAssist
In reply to Andre PrzywaraNewer SVM implementations (Bulldozer) give the desired address on a INVLPG intercept explicitly in the EXITINFO1 field of the VMCB. Use this address to avoid a costly instruction fetch and deocde cycle.
Only useful with shadow paging.
commit c89a3f986870176b2c178a28fce82e868160ee61
Author: Andre Przywara <[email protected]> Date: Sun Jul 25 15:09:44 2010 +0200
svm: implement INVLPG part of DecodeAssist
Newer SVM implementations (Bulldozer) give the desired address on a INVLPG intercept explicitly in the EXITINFO1 field of the VMCB. Use this address to avoid a costly instruction fetch and deocde cycle.
Only useful with shadow paging.
xen/arch/x86/hvm/svm/svm.c +4 -1
@@ -1939,7 +1939,10 @@ asmlinkage void svm_vmexit_handler(struct cpu_user_regs *regs)case VMEXIT_INVLPG :case VMEXIT_INVLPGA : -if ( !handle_mmio () ) +if (cpu_has_svm_decode) { +svm_invlpg_intercept (vmcb->exitinfo1); +__update_guest_eip (regs, vmcb->nextrip - vmcb->rip); + }else if ( !handle_mmio () )hvm_inject_exception (TRAP_gp_fault,0 ,0 );break ;
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
Re: [PATCH 1/5] vmx/hvm: move mov-cr handling functions to generic HVM code
In reply to Andre PrzywaraQuoted 3 lines · Andre Przywara
Currently the handling of CR accesses intercepts is done much differently in SVM and VMX. For future usage move the VMX part into the generic HVM path and use the exported functions.
Cleaned up and applied as c/s 23233. In particular it broke VMX, by not preoperly decoding the GPR component of the exit qualification before passing to the new generic handler.
-- Keir
Quoted 6 lines
Signed-off-by Andre Przywara <[email protected]>_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
Keir Fraser
AppliedRe: [PATCH 5/5] svm: implement INVLPG part of DecodeAssist
In reply to Andre PrzywaraQuoted 5 lines · Andre Przywara
Newer SVM implementations (Bulldozer) give the desired address on a INVLPG intercept explicitly in the EXITINFO1 field of the VMCB. Use this address to avoid a costly instruction fetch and deocde cycle.
Only useful with shadow paging.
Applied as c/2 23236. Fixed to not handle the INVLPGA exit code.
-- Keir
Quoted 6 lines
Signed-off-by Andre Przywara <[email protected]>_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
Re: [PATCH 3/5] svm: implement instruction fetch part of DecodeAssist
In reply to Andre PrzywaraQuoted 8 lines · Andre Przywara
Newer SVM implementations (Bulldozer) copy up to 15 bytes from the instruction stream into the VMCB when a #PF or #NPF exception is intercepted. This patch makes use of this information if available. This saves us from a) traversing the guest's page tables, b) mapping the guest's memory and c) copy the instructions from there into the hypervisor's address space.
This speeds up #NPF intercepts quite a lot and avoids cache and TLB trashing.
This is a bit skanky. I'll have a think about a neater way of doing it.
-- Keir
Quoted 6 lines
Signed-off-by Andre Przywara <[email protected]>_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
Re: [PATCH 3/5] svm: implement instruction fetch part of DecodeAssist
In reply to Andre PrzywaraQuoted 8 lines · Andre Przywara
Newer SVM implementations (Bulldozer) copy up to 15 bytes from the instruction stream into the VMCB when a #PF or #NPF exception is intercepted. This patch makes use of this information if available. This saves us from a) traversing the guest's page tables, b) mapping the guest's memory and c) copy the instructions from there into the hypervisor's address space.
This speeds up #NPF intercepts quite a lot and avoids cache and TLB trashing.
Rewritten quite a bit and applied as c/s 23238. You probably should check it still works and with the performance improvement you expect. :-)
-- Keir
Quoted 6 lines
Signed-off-by Andre Przywara <[email protected]>_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel