x86/hvm: accelerate IO intercept handling
PATCH 1/1 x86/hvm: accelerate IO intercept handling
Hi,
currently we go through the emulator every time a HVM guest does an I/O port access (in/out). This is unnecessary most of the times, as both VMX and SVM provide all the necessary information already in the VMCS/VMCB. String instructions are not covered by this shortcut, but they are quite rare and we would need to access the guest memory anyway. This patch decodes the information from VMCB/VMCS and calls a simple handle_mmio wrapper. In handle_mmio() itself the emulation part will simply be skipped, this approach avoids code duplication.
Since the vendor specific part is quite trivial, I implemented both the VMX and SVM part, please check the VMX part for sanity.
I simply boot tested both versions and ran some simple benchmarks.
A micro benchmark (hammering an I/O port in a tight loop) shows a significant performance improvement (down to 66% of the time needed to handle the intercept on a K8, measured in the guest with TSC). Even with reading a 1GB file from an emulated IDE harddisk (Dom0 cached) I could get a 4-5% improvement.
We found some guests (e.g. the TCP stack in some Windows version) which exercise the PM-Timer I/O port (0x1F48) very often (multiple 10,000 times per second), these workloads also benefit from this patch.
Since this is not a regression, but only a performance improvement, I'd suggest to apply to the post 4.0 tree.
Regards,
Andre.
Signature · 9 lines
Andre Przywara AMD-Operating System Research Center (OSRC), Dresden, Germany Tel: +49 351 448 3567 12 ----to satisfy European Law for business letters: Advanced Micro Devices GmbH Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen Geschaeftsfuehrer: Andrew Bowd; Thomas M. McCoy; Giuliano Meroni Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen Registergericht Muenchen, HRB Nr. 43632
xen/arch/x86/hvm/emulate.c +1 -1
@@ -48,7 +48,7 @@trace_var (event,0 /*!cycles*/ , size, buffer); } -static int hvmemul_do_io ( +int hvmemul_do_io (int is_mmio,paddr_t addr,unsigned long *reps,int size,paddr_t ram_gpa,int dir,int df,void *p_data) {
xen/arch/x86/hvm/io.c +46 -13
85 lines not on this page · raw message
xen/arch/x86/hvm/svm/emulate.c +3 -0
@@ -149,6 +149,9 @@if ( (inst_len =svm_nextrip_insn_length (v)) !=0 )return inst_len; +if ( vmcb->exitcode ==VMEXIT_IOIO ) +return vmcb->exitinfo2 - vmcb->rip; +/* Fetch up to the next page break; we'll fetch from the next page * later if we have to. */ fetch_addr =svm_rip2pointer (v);
xen/arch/x86/hvm/svm/svm.c +25 -1
@@ -42,6 +42,7 @@#include <asm/hvm/hvm.h> #include <asm/hvm/support.h> #include <asm/hvm/io.h> +#include <asm/hvm/emulate.h> #include <asm/hvm/svm/asid.h> #include <asm/hvm/svm/svm.h> #include <asm/hvm/svm/vmcb.h>
@@ -1244,6 +1245,23 @@__update_guest_eip (regs, inst_len); } +static int svm_vmexit_do_io (struct vmcb_struct *vmcb, +struct cpu_user_regs *regs) +{ +uint16_t port; +int bytes, dir; +int rc; + + port = (vmcb->exitinfo1 >>16 ) &0xFFFF ; + bytes = ((vmcb->exitinfo1 >>4 ) &0x07 ); + dir = (vmcb->exitinfo1 &1 ) ?IOREQ_READ :IOREQ_WRITE ; + + rc =handle_mmio_decoded (port, bytes, dir); +if ( rc !=X86EMUL_RETRY ) +__update_guest_eip (regs, vmcb->exitinfo2 - vmcb->rip); +return rc; +} +static void svm_invlpg_intercept (unsigned long vaddr) {struct vcpu *curr = current;
@@ -1450,11 +1468,17 @@svm_vmexit_do_hlt (vmcb, regs);break ; +case VMEXIT_IOIO : +if ( ( vmcb->exitinfo1 & (1 <<2 ) ) ==0 ) { +if ( !svm_vmexit_do_io (vmcb, regs) ) +hvm_inject_exception (TRAP_gp_fault,0 ,0 ); +break ; + } +/* fallthrough to emulation if a string instruction */ case VMEXIT_CR0_READ ... VMEXIT_CR15_READ: case VMEXIT_CR0_WRITE ... VMEXIT_CR15_WRITE:case VMEXIT_INVLPG :case VMEXIT_INVLPGA : -case VMEXIT_IOIO :if ( !handle_mmio () )hvm_inject_exception (TRAP_gp_fault,0 ,0 );break ;
xen/arch/x86/hvm/vmx/vmx.c +27 -2
@@ -1719,6 +1719,25 @@return 1 ; } +static int vmx_io_intercept (unsigned long exit_qualification, +struct cpu_user_regs *regs) +{ +uint16_t port; +int bytes, dir; +int rc; +int inst_len; + + port = (exit_qualification >>16 ) &0xFFFF ; + bytes = (exit_qualification &0x07 ) +1 ; + dir = (exit_qualification &0x08 ) ?IOREQ_READ :IOREQ_WRITE ; + + inst_len =__get_instruction_length (); + rc =handle_mmio_decoded (port, bytes, dir); +if ( rc !=X86EMUL_RETRY ) +__update_guest_eip (inst_len); +return rc; +} +static const struct lbr_info {u32 base, count; } p4_lbr[] = {
@@ -2587,8 +2606,14 @@break ;case EXIT_REASON_IO_INSTRUCTION : -if ( !handle_mmio () ) -vmx_inject_hw_exception (TRAP_gp_fault,0 ); + exit_qualification =__vmread (EXIT_QUALIFICATION ); +if (exit_qualification &0x10 ) { +if ( !handle_mmio () ) +vmx_inject_hw_exception (TRAP_gp_fault,0 ); + }else { +if ( !vmx_io_intercept (exit_qualification, regs) ) +vmx_inject_hw_exception (TRAP_gp_fault,0 ); + }break ;case EXIT_REASON_INVD :
xen/include/asm-x86/hvm/emulate.h +4 -0
@@ -46,4 +46,8 @@enum x86_segment seg, struct hvm_emulate_ctxt *hvmemul_ctxt); +int hvmemul_do_io ( +int is_mmio,paddr_t addr,unsigned long *reps,int size, +paddr_t ram_gpa,int dir,int df,void *p_data); +#endif /* __ASM_X86_HVM_EMULATE_H__ */
xen/include/asm-x86/hvm/io.h +1 -0
@@ -100,6 +100,7 @@void send_invalidate_req (void );int handle_mmio (void );int handle_mmio_with_translation (unsigned long gva,unsigned long gpfn); +int handle_mmio_decoded (uint16_t port,int size,int dir);void hvm_interrupt_post (struct vcpu *v,int vector,int type);void hvm_io_assist (void ); void hvm_dpci_eoi(struct domain *d, unsigned int guest_irq,
xen/include/asm-x86/hvm/vcpu.h +4 -0
@@ -103,6 +103,10 @@*/ unsigned long mmio_gva;unsigned long mmio_gpfn; +uint16_t io_port; +int io_size; +unsigned io_dir; +/* 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
Quoted 9 lines · Andre Przywara
I simply boot tested both versions and ran some simple benchmarks. A micro benchmark (hammering an I/O port in a tight loop) shows a significant performance improvement (down to 66% of the time needed to handle the intercept on a K8, measured in the guest with TSC). Even with reading a 1GB file from an emulated IDE harddisk (Dom0 cached) I could get a 4-5% improvement.
We found some guests (e.g. the TCP stack in some Windows version) which exercise the PM-Timer I/O port (0x1F48) very often (multiple 10,000 times per second), these workloads also benefit from this patch.
By how much? I mean, the microbenchmark and 5% speedup on our poor-man's IO path are not very interesting. Unless the speedup on the only possibly-interesting workload you mention is significant, this whole optimisation seems unnecessary.
-- Keir
Quoted 12 lines · Keir Fraser
Quoted 9 lines · Andre Przywara
I simply boot tested both versions and ran some simple benchmarks. A micro benchmark (hammering an I/O port in a tight loop) shows a significant performance improvement (down to 66% of the time needed to handle the intercept on a K8, measured in the guest with TSC). Even with reading a 1GB file from an emulated IDE harddisk (Dom0 cached) I could get a 4-5% improvement.
We found some guests (e.g. the TCP stack in some Windows version) which exercise the PM-Timer I/O port (0x1F48) very often (multiple 10,000 times per second), these workloads also benefit from this patch.
Quoted 2 lines
By how much? I mean, the microbenchmark and 5% speedup on our poor-man's IO path are not very interesting.
Educated estimation: Sysmark productivity should give about 0.5%, Passmark TCP localhost transfer on Windows 2008R2 should improve about 5%.
Quoted 3 lines
Unless the speedup on the only
possibly-interesting workload you mention is significant, this whole optimisation seems unnecessary.
Actually it is missing enablement. What is the purpose of going through the emulator (mapping and walking guest page tables, reading guest instruction memory, decoding x86 code) when you don't have to? KVM is implementing this for quite some time now.
And, after all, low hanging fruits are growing higher nowadays, so I'd consider an even modest performance improvement for _every_ machine by just a software patch a valuable thing.
Regards,
Andre.
Signature · 9 lines
Andre Przywara AMD-Operating System Research Center (OSRC), Dresden, Germany Tel: +49 351 448 3567 12 ----to satisfy European Law for business letters: Advanced Micro Devices GmbH Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen Geschaeftsfuehrer: Andrew Bowd; Thomas M. McCoy; Giuliano Meroni Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen Registergericht Muenchen, HRB Nr. 43632
Quoted 4 lines · Andre Przywara
Quoted 3 lines
Unless the speedup on the only
possibly-interesting workload you mention is significant, this whole optimisation seems unnecessary.Actually it is missing enablement.
Yes it's very important to use *every* CPU feature. ;-)
Well, the patch looks okay, and it may speed up things we care about a percent or two I guess. I will look at it properly after 4.0 and I'm sure we can get it in the tree.
-- Keir