2 * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com)
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice unmodified, this list of conditions, and the following
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
30 #include <sys/param.h>
31 #include <sys/systm.h>
33 #include <sys/kernel.h>
34 #include <sys/malloc.h>
37 #include <sys/sysctl.h>
42 #include <machine/cpufunc.h>
43 #include <machine/psl.h>
44 #include <machine/md_var.h>
45 #include <machine/specialreg.h>
46 #include <machine/smp.h>
47 #include <machine/vmm.h>
48 #include <machine/vmm_dev.h>
49 #include <machine/vmm_instruction_emul.h>
51 #include "vmm_lapic.h"
54 #include "vmm_ioport.h"
57 #include "vlapic_priv.h"
62 #include "svm_softc.h"
67 SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL);
70 * SVM CPUID function 0x8000_000A, edx bit decoding.
72 #define AMD_CPUID_SVM_NP BIT(0) /* Nested paging or RVI */
73 #define AMD_CPUID_SVM_LBR BIT(1) /* Last branch virtualization */
74 #define AMD_CPUID_SVM_SVML BIT(2) /* SVM lock */
75 #define AMD_CPUID_SVM_NRIP_SAVE BIT(3) /* Next RIP is saved */
76 #define AMD_CPUID_SVM_TSC_RATE BIT(4) /* TSC rate control. */
77 #define AMD_CPUID_SVM_VMCB_CLEAN BIT(5) /* VMCB state caching */
78 #define AMD_CPUID_SVM_FLUSH_BY_ASID BIT(6) /* Flush by ASID */
79 #define AMD_CPUID_SVM_DECODE_ASSIST BIT(7) /* Decode assist */
80 #define AMD_CPUID_SVM_PAUSE_INC BIT(10) /* Pause intercept filter. */
81 #define AMD_CPUID_SVM_PAUSE_FTH BIT(12) /* Pause filter threshold */
82 #define AMD_CPUID_SVM_AVIC BIT(13) /* AVIC present */
84 #define VMCB_CACHE_DEFAULT (VMCB_CACHE_ASID | \
94 static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT;
95 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean,
98 static MALLOC_DEFINE(M_SVM, "svm", "svm");
99 static MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic");
101 /* Per-CPU context area. */
102 extern struct pcpu __pcpu[];
104 static uint32_t svm_feature = ~0U; /* AMD SVM features. */
105 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RDTUN, &svm_feature, 0,
106 "SVM features advertised by CPUID.8000000AH:EDX");
108 static int disable_npf_assist;
109 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN,
110 &disable_npf_assist, 0, NULL);
112 /* Maximum ASIDs supported by the processor */
113 static uint32_t nasid;
114 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RDTUN, &nasid, 0,
115 "Number of ASIDs supported by this processor");
117 /* Current ASID generation for each host cpu */
118 static struct asid asid[MAXCPU];
121 * SVM host state saved area of size 4KB for each core.
123 static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE);
125 static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery");
126 static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry");
127 static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window");
129 static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val);
135 return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID);
142 return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST);
146 svm_disable(void *arg __unused)
150 efer = rdmsr(MSR_EFER);
152 wrmsr(MSR_EFER, efer);
156 * Disable SVM on all CPUs.
162 smp_rendezvous(NULL, svm_disable, NULL, NULL);
167 * Verify that all the features required by bhyve are available.
170 check_svm_features(void)
174 /* CPUID Fn8000_000A is for SVM */
175 do_cpuid(0x8000000A, regs);
176 svm_feature &= regs[3];
179 * The number of ASIDs can be configured to be less than what is
180 * supported by the hardware but not more.
182 if (nasid == 0 || nasid > regs[1])
184 KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid));
186 /* bhyve requires the Nested Paging feature */
187 if (!(svm_feature & AMD_CPUID_SVM_NP)) {
188 printf("SVM: Nested Paging feature not available.\n");
192 /* bhyve requires the NRIP Save feature */
193 if (!(svm_feature & AMD_CPUID_SVM_NRIP_SAVE)) {
194 printf("SVM: NRIP Save feature not available.\n");
202 svm_enable(void *arg __unused)
206 efer = rdmsr(MSR_EFER);
208 wrmsr(MSR_EFER, efer);
210 wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave[curcpu]));
214 * Return 1 if SVM is enabled on this processor and 0 otherwise.
221 /* Section 15.4 Enabling SVM from APM2. */
222 if ((amd_feature2 & AMDID2_SVM) == 0) {
223 printf("SVM: not available.\n");
227 msr = rdmsr(MSR_VM_CR);
228 if ((msr & VM_CR_SVMDIS) != 0) {
229 printf("SVM: disabled by BIOS.\n");
241 if (!svm_available())
244 error = check_svm_features();
248 vmcb_clean &= VMCB_CACHE_DEFAULT;
250 for (cpu = 0; cpu < MAXCPU; cpu++) {
252 * Initialize the host ASIDs to their "highest" valid values.
254 * The next ASID allocation will rollover both 'gen' and 'num'
255 * and start off the sequence at {1,1}.
257 asid[cpu].gen = ~0UL;
258 asid[cpu].num = nasid - 1;
262 svm_npt_init(ipinum);
264 /* Enable SVM on all CPUs */
265 smp_rendezvous(NULL, svm_enable, NULL, NULL);
277 /* Pentium compatible MSRs */
278 #define MSR_PENTIUM_START 0
279 #define MSR_PENTIUM_END 0x1FFF
280 /* AMD 6th generation and Intel compatible MSRs */
281 #define MSR_AMD6TH_START 0xC0000000UL
282 #define MSR_AMD6TH_END 0xC0001FFFUL
283 /* AMD 7th and 8th generation compatible MSRs */
284 #define MSR_AMD7TH_START 0xC0010000UL
285 #define MSR_AMD7TH_END 0xC0011FFFUL
288 * Get the index and bit position for a MSR in permission bitmap.
289 * Two bits are used for each MSR: lower bit for read and higher bit for write.
292 svm_msr_index(uint64_t msr, int *index, int *bit)
297 *bit = (msr % 4) * 2;
300 if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) {
305 base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1);
306 if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) {
307 off = (msr - MSR_AMD6TH_START);
308 *index = (off + base) / 4;
312 base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1);
313 if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) {
314 off = (msr - MSR_AMD7TH_START);
315 *index = (off + base) / 4;
323 * Allow vcpu to read or write the 'msr' without trapping into the hypervisor.
326 svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write)
328 int index, bit, error;
330 error = svm_msr_index(msr, &index, &bit);
331 KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr));
332 KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE,
333 ("%s: invalid index %d for msr %#lx", __func__, index, msr));
334 KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d "
335 "msr %#lx", __func__, bit, msr));
338 perm_bitmap[index] &= ~(1UL << bit);
341 perm_bitmap[index] &= ~(2UL << bit);
345 svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr)
348 svm_msr_perm(perm_bitmap, msr, true, true);
352 svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr)
355 svm_msr_perm(perm_bitmap, msr, true, false);
359 svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask)
361 struct vmcb_ctrl *ctrl;
363 KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
365 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
366 return (ctrl->intercept[idx] & bitmask ? 1 : 0);
370 svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask,
373 struct vmcb_ctrl *ctrl;
376 KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
378 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
379 oldval = ctrl->intercept[idx];
382 ctrl->intercept[idx] |= bitmask;
384 ctrl->intercept[idx] &= ~bitmask;
386 if (ctrl->intercept[idx] != oldval) {
387 svm_set_dirty(sc, vcpu, VMCB_CACHE_I);
388 VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified "
389 "from %#x to %#x", idx, oldval, ctrl->intercept[idx]);
394 svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
397 svm_set_intercept(sc, vcpu, off, bitmask, 0);
401 svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
404 svm_set_intercept(sc, vcpu, off, bitmask, 1);
408 vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa,
409 uint64_t msrpm_base_pa, uint64_t np_pml4)
411 struct vmcb_ctrl *ctrl;
412 struct vmcb_state *state;
416 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
417 state = svm_get_vmcb_state(sc, vcpu);
419 ctrl->iopm_base_pa = iopm_base_pa;
420 ctrl->msrpm_base_pa = msrpm_base_pa;
422 /* Enable nested paging */
424 ctrl->n_cr3 = np_pml4;
427 * Intercept accesses to the control registers that are not shadowed
428 * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8.
430 for (n = 0; n < 16; n++) {
431 mask = (BIT(n) << 16) | BIT(n);
432 if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8)
433 svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
435 svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
440 * Intercept everything when tracing guest exceptions otherwise
441 * just intercept machine check exception.
443 if (vcpu_trace_exceptions(sc->vm, vcpu)) {
444 for (n = 0; n < 32; n++) {
446 * Skip unimplemented vectors in the exception bitmap.
448 if (n == 2 || n == 9) {
451 svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(n));
454 svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC));
457 /* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */
458 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO);
459 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR);
460 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID);
461 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR);
462 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT);
463 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI);
464 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI);
465 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN);
466 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
467 VMCB_INTCPT_FERR_FREEZE);
469 svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR);
470 svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT);
473 * From section "Canonicalization and Consistency Checks" in APMv2
474 * the VMRUN intercept bit must be set to pass the consistency check.
476 svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
479 * The ASID will be set to a non-zero value just before VMRUN.
484 * Section 15.21.1, Interrupt Masking in EFLAGS
485 * Section 15.21.2, Virtualizing APIC.TPR
487 * This must be set for %rflag and %cr8 isolation of guest and host.
489 ctrl->v_intr_masking = 1;
491 /* Enable Last Branch Record aka LBR for debugging */
492 ctrl->lbr_virt_en = 1;
493 state->dbgctl = BIT(0);
495 /* EFER_SVM must always be set when the guest is executing */
496 state->efer = EFER_SVM;
498 /* Set up the PAT to power-on state */
499 state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK) |
500 PAT_VALUE(1, PAT_WRITE_THROUGH) |
501 PAT_VALUE(2, PAT_UNCACHED) |
502 PAT_VALUE(3, PAT_UNCACHEABLE) |
503 PAT_VALUE(4, PAT_WRITE_BACK) |
504 PAT_VALUE(5, PAT_WRITE_THROUGH) |
505 PAT_VALUE(6, PAT_UNCACHED) |
506 PAT_VALUE(7, PAT_UNCACHEABLE);
510 * Initialize a virtual machine.
513 svm_vminit(struct vm *vm, pmap_t pmap)
515 struct svm_softc *svm_sc;
516 struct svm_vcpu *vcpu;
517 vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
520 svm_sc = malloc(sizeof (struct svm_softc), M_SVM, M_WAITOK | M_ZERO);
522 svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4);
525 * Intercept read and write accesses to all MSRs.
527 memset(svm_sc->msr_bitmap, 0xFF, sizeof(svm_sc->msr_bitmap));
530 * Access to the following MSRs is redirected to the VMCB when the
531 * guest is executing. Therefore it is safe to allow the guest to
532 * read/write these MSRs directly without hypervisor involvement.
534 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
535 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
536 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
538 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
539 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
540 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
541 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
542 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
543 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
544 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
545 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
547 svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
550 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
552 svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
554 /* Intercept access to all I/O ports. */
555 memset(svm_sc->iopm_bitmap, 0xFF, sizeof(svm_sc->iopm_bitmap));
557 iopm_pa = vtophys(svm_sc->iopm_bitmap);
558 msrpm_pa = vtophys(svm_sc->msr_bitmap);
559 pml4_pa = svm_sc->nptp;
560 for (i = 0; i < VM_MAXCPU; i++) {
561 vcpu = svm_get_vcpu(svm_sc, i);
563 vcpu->lastcpu = NOCPU;
564 vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
565 vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
566 svm_msr_guest_init(svm_sc, i);
572 * Collateral for a generic SVM VM-exit.
575 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
578 vme->exitcode = VM_EXITCODE_SVM;
579 vme->u.svm.exitcode = code;
580 vme->u.svm.exitinfo1 = info1;
581 vme->u.svm.exitinfo2 = info2;
585 svm_cpl(struct vmcb_state *state)
590 * "Retrieve the CPL from the CPL field in the VMCB, not
591 * from any segment DPL"
596 static enum vm_cpu_mode
597 svm_vcpu_mode(struct vmcb *vmcb)
599 struct vmcb_segment seg;
600 struct vmcb_state *state;
603 state = &vmcb->state;
605 if (state->efer & EFER_LMA) {
606 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
607 KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
611 * Section 4.8.1 for APM2, check if Code Segment has
612 * Long attribute set in descriptor.
614 if (seg.attrib & VMCB_CS_ATTRIB_L)
615 return (CPU_MODE_64BIT);
617 return (CPU_MODE_COMPATIBILITY);
618 } else if (state->cr0 & CR0_PE) {
619 return (CPU_MODE_PROTECTED);
621 return (CPU_MODE_REAL);
625 static enum vm_paging_mode
626 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
629 if ((cr0 & CR0_PG) == 0)
630 return (PAGING_MODE_FLAT);
631 if ((cr4 & CR4_PAE) == 0)
632 return (PAGING_MODE_32);
634 return (PAGING_MODE_64);
636 return (PAGING_MODE_PAE);
640 * ins/outs utility routines
643 svm_inout_str_index(struct svm_regctx *regs, int in)
647 val = in ? regs->sctx_rdi : regs->sctx_rsi;
653 svm_inout_str_count(struct svm_regctx *regs, int rep)
657 val = rep ? regs->sctx_rcx : 1;
663 svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
664 int in, struct vm_inout_str *vis)
669 vis->seg_name = VM_REG_GUEST_ES;
671 /* The segment field has standard encoding */
672 s = (info1 >> 10) & 0x7;
673 vis->seg_name = vm_segment_name(s);
676 error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
677 KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
681 svm_inout_str_addrsize(uint64_t info1)
685 size = (info1 >> 7) & 0x7;
688 return (2); /* 16 bit */
690 return (4); /* 32 bit */
692 return (8); /* 64 bit */
694 panic("%s: invalid size encoding %d", __func__, size);
699 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
701 struct vmcb_state *state;
703 state = &vmcb->state;
704 paging->cr3 = state->cr3;
705 paging->cpl = svm_cpl(state);
706 paging->cpu_mode = svm_vcpu_mode(vmcb);
707 paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
714 * Handle guest I/O intercept.
717 svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
719 struct vmcb_ctrl *ctrl;
720 struct vmcb_state *state;
721 struct svm_regctx *regs;
722 struct vm_inout_str *vis;
726 state = svm_get_vmcb_state(svm_sc, vcpu);
727 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
728 regs = svm_get_guest_regctx(svm_sc, vcpu);
730 info1 = ctrl->exitinfo1;
731 inout_string = info1 & BIT(2) ? 1 : 0;
734 * The effective segment number in EXITINFO1[12:10] is populated
735 * only if the processor has the DecodeAssist capability.
737 * XXX this is not specified explicitly in APMv2 but can be verified
740 if (inout_string && !decode_assist())
743 vmexit->exitcode = VM_EXITCODE_INOUT;
744 vmexit->u.inout.in = (info1 & BIT(0)) ? 1 : 0;
745 vmexit->u.inout.string = inout_string;
746 vmexit->u.inout.rep = (info1 & BIT(3)) ? 1 : 0;
747 vmexit->u.inout.bytes = (info1 >> 4) & 0x7;
748 vmexit->u.inout.port = (uint16_t)(info1 >> 16);
749 vmexit->u.inout.eax = (uint32_t)(state->rax);
752 vmexit->exitcode = VM_EXITCODE_INOUT_STR;
753 vis = &vmexit->u.inout_str;
754 svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
755 vis->rflags = state->rflags;
756 vis->cr0 = state->cr0;
757 vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
758 vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
759 vis->addrsize = svm_inout_str_addrsize(info1);
760 svm_inout_str_seginfo(svm_sc, vcpu, info1,
761 vmexit->u.inout.in, vis);
768 npf_fault_type(uint64_t exitinfo1)
771 if (exitinfo1 & VMCB_NPF_INFO1_W)
772 return (VM_PROT_WRITE);
773 else if (exitinfo1 & VMCB_NPF_INFO1_ID)
774 return (VM_PROT_EXECUTE);
776 return (VM_PROT_READ);
780 svm_npf_emul_fault(uint64_t exitinfo1)
783 if (exitinfo1 & VMCB_NPF_INFO1_ID) {
787 if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
791 if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
799 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
801 struct vm_guest_paging *paging;
802 struct vmcb_segment seg;
803 struct vmcb_ctrl *ctrl;
808 paging = &vmexit->u.inst_emul.paging;
810 vmexit->exitcode = VM_EXITCODE_INST_EMUL;
811 vmexit->u.inst_emul.gpa = gpa;
812 vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
813 svm_paging_info(vmcb, paging);
815 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
816 KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
818 switch(paging->cpu_mode) {
820 vmexit->u.inst_emul.cs_base = seg.base;
821 vmexit->u.inst_emul.cs_d = 0;
823 case CPU_MODE_PROTECTED:
824 case CPU_MODE_COMPATIBILITY:
825 vmexit->u.inst_emul.cs_base = seg.base;
828 * Section 4.8.1 of APM2, Default Operand Size or D bit.
830 vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
834 vmexit->u.inst_emul.cs_base = 0;
835 vmexit->u.inst_emul.cs_d = 0;
840 * Copy the instruction bytes into 'vie' if available.
842 if (decode_assist() && !disable_npf_assist) {
843 inst_len = ctrl->inst_len;
844 inst_bytes = ctrl->inst_bytes;
849 vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
854 intrtype_to_str(int intr_type)
857 case VMCB_EVENTINJ_TYPE_INTR:
859 case VMCB_EVENTINJ_TYPE_NMI:
861 case VMCB_EVENTINJ_TYPE_INTn:
863 case VMCB_EVENTINJ_TYPE_EXCEPTION:
864 return ("exception");
866 panic("%s: unknown intr_type %d", __func__, intr_type);
872 * Inject an event to vcpu as described in section 15.20, "Event injection".
875 svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
876 uint32_t error, bool ec_valid)
878 struct vmcb_ctrl *ctrl;
880 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
882 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
883 ("%s: event already pending %#lx", __func__, ctrl->eventinj));
885 KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
889 case VMCB_EVENTINJ_TYPE_INTR:
890 case VMCB_EVENTINJ_TYPE_NMI:
891 case VMCB_EVENTINJ_TYPE_INTn:
893 case VMCB_EVENTINJ_TYPE_EXCEPTION:
894 if (vector >= 0 && vector <= 31 && vector != 2)
898 panic("%s: invalid intr_type/vector: %d/%d", __func__,
901 ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
903 ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
904 ctrl->eventinj |= (uint64_t)error << 32;
905 VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
906 intrtype_to_str(intr_type), vector, error);
908 VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
909 intrtype_to_str(intr_type), vector);
914 svm_update_virqinfo(struct svm_softc *sc, int vcpu)
917 struct vlapic *vlapic;
918 struct vmcb_ctrl *ctrl;
922 vlapic = vm_lapic(vm, vcpu);
923 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
925 /* Update %cr8 in the emulated vlapic */
926 vlapic_set_cr8(vlapic, ctrl->v_tpr);
929 * If V_IRQ indicates that the interrupt injection attempted on then
930 * last VMRUN was successful then update the vlapic accordingly.
932 if (ctrl->v_intr_vector != 0) {
933 pending = ctrl->v_irq;
934 KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid "
935 "v_intr_vector %d", __func__, ctrl->v_intr_vector));
936 KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
937 VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector,
938 pending ? "pending" : "accepted");
940 vlapic_intr_accepted(vlapic, ctrl->v_intr_vector);
945 svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
947 struct vmcb_ctrl *ctrl;
950 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
951 intinfo = ctrl->exitintinfo;
952 if (!VMCB_EXITINTINFO_VALID(intinfo))
956 * From APMv2, Section "Intercepts during IDT interrupt delivery"
958 * If a #VMEXIT happened during event delivery then record the event
959 * that was being delivered.
961 VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
962 intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
963 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
964 vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
968 vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
971 return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
976 enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
978 struct vmcb_ctrl *ctrl;
980 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
982 if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
983 KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
984 KASSERT(vintr_intercept_enabled(sc, vcpu),
985 ("%s: vintr intercept should be enabled", __func__));
989 VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
992 ctrl->v_intr_vector = 0;
993 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
994 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
998 disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1000 struct vmcb_ctrl *ctrl;
1002 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1004 if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1005 KASSERT(!vintr_intercept_enabled(sc, vcpu),
1006 ("%s: vintr intercept should be disabled", __func__));
1011 if (ctrl->v_intr_vector == 0)
1012 VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
1014 VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection");
1017 ctrl->v_intr_vector = 0;
1018 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1019 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1023 svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
1025 struct vmcb_ctrl *ctrl;
1028 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1029 oldval = ctrl->intr_shadow;
1030 newval = val ? 1 : 0;
1031 if (newval != oldval) {
1032 ctrl->intr_shadow = newval;
1033 VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1039 svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1041 struct vmcb_ctrl *ctrl;
1043 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1044 *val = ctrl->intr_shadow;
1049 * Once an NMI is injected it blocks delivery of further NMIs until the handler
1050 * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1051 * to track when the vcpu is done handling the NMI.
1054 nmi_blocked(struct svm_softc *sc, int vcpu)
1058 blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1064 enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1067 KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1068 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1069 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1073 clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1077 KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1078 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1080 * When the IRET intercept is cleared the vcpu will attempt to execute
1081 * the "iret" when it runs next. However, it is possible to inject
1082 * another NMI into the vcpu before the "iret" has actually executed.
1084 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1085 * it will trap back into the hypervisor. If an NMI is pending for
1086 * the vcpu it will be injected into the guest.
1088 * XXX this needs to be fixed
1090 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1093 * Set 'intr_shadow' to prevent an NMI from being injected on the
1096 error = svm_modify_intr_shadow(sc, vcpu, 1);
1097 KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1100 #define EFER_MBZ_BITS 0xFFFFFFFFFFFF0200UL
1103 svm_write_efer(struct svm_softc *sc, int vcpu, uint64_t newval, bool *retu)
1105 struct vm_exit *vme;
1106 struct vmcb_state *state;
1107 uint64_t changed, lma, oldval;
1110 state = svm_get_vmcb_state(sc, vcpu);
1112 oldval = state->efer;
1113 VCPU_CTR2(sc->vm, vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval);
1115 newval &= ~0xFE; /* clear the Read-As-Zero (RAZ) bits */
1116 changed = oldval ^ newval;
1118 if (newval & EFER_MBZ_BITS)
1121 /* APMv2 Table 14-5 "Long-Mode Consistency Checks" */
1122 if (changed & EFER_LME) {
1123 if (state->cr0 & CR0_PG)
1127 /* EFER.LMA = EFER.LME & CR0.PG */
1128 if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0)
1133 if ((newval & EFER_LMA) != lma)
1136 if (newval & EFER_NXE) {
1137 if (!vm_cpuid_capability(sc->vm, vcpu, VCC_NO_EXECUTE))
1142 * XXX bhyve does not enforce segment limits in 64-bit mode. Until
1143 * this is fixed flag guest attempt to set EFER_LMSLE as an error.
1145 if (newval & EFER_LMSLE) {
1146 vme = vm_exitinfo(sc->vm, vcpu);
1147 vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0);
1152 if (newval & EFER_FFXSR) {
1153 if (!vm_cpuid_capability(sc->vm, vcpu, VCC_FFXSR))
1157 if (newval & EFER_TCE) {
1158 if (!vm_cpuid_capability(sc->vm, vcpu, VCC_TCE))
1162 error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, newval);
1163 KASSERT(error == 0, ("%s: error %d updating efer", __func__, error));
1166 vm_inject_gp(sc->vm, vcpu);
1171 emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1177 error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1178 else if (num == MSR_EFER)
1179 error = svm_write_efer(sc, vcpu, val, retu);
1181 error = svm_wrmsr(sc, vcpu, num, val, retu);
1187 emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1189 struct vmcb_state *state;
1190 struct svm_regctx *ctx;
1195 error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1197 error = svm_rdmsr(sc, vcpu, num, &result, retu);
1200 state = svm_get_vmcb_state(sc, vcpu);
1201 ctx = svm_get_guest_regctx(sc, vcpu);
1202 state->rax = result & 0xffffffff;
1203 ctx->sctx_rdx = result >> 32;
1211 exit_reason_to_str(uint64_t reason)
1213 static char reasonbuf[32];
1216 case VMCB_EXIT_INVALID:
1217 return ("invalvmcb");
1218 case VMCB_EXIT_SHUTDOWN:
1219 return ("shutdown");
1221 return ("nptfault");
1222 case VMCB_EXIT_PAUSE:
1226 case VMCB_EXIT_CPUID:
1232 case VMCB_EXIT_INTR:
1236 case VMCB_EXIT_VINTR:
1240 case VMCB_EXIT_IRET:
1242 case VMCB_EXIT_MONITOR:
1244 case VMCB_EXIT_MWAIT:
1247 snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1254 * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1255 * that are due to instruction intercepts as well as MSR and IOIO intercepts
1256 * and exceptions caused by INT3, INTO and BOUND instructions.
1258 * Return 1 if the nRIP is valid and 0 otherwise.
1261 nrip_valid(uint64_t exitcode)
1264 case 0x00 ... 0x0F: /* read of CR0 through CR15 */
1265 case 0x10 ... 0x1F: /* write of CR0 through CR15 */
1266 case 0x20 ... 0x2F: /* read of DR0 through DR15 */
1267 case 0x30 ... 0x3F: /* write of DR0 through DR15 */
1268 case 0x43: /* INT3 */
1269 case 0x44: /* INTO */
1270 case 0x45: /* BOUND */
1271 case 0x65 ... 0x7C: /* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1272 case 0x80 ... 0x8D: /* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1280 svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1283 struct vmcb_state *state;
1284 struct vmcb_ctrl *ctrl;
1285 struct svm_regctx *ctx;
1286 uint64_t code, info1, info2, val;
1287 uint32_t eax, ecx, edx;
1288 int error, errcode_valid, handled, idtvec, reflect;
1291 ctx = svm_get_guest_regctx(svm_sc, vcpu);
1292 vmcb = svm_get_vmcb(svm_sc, vcpu);
1293 state = &vmcb->state;
1297 code = ctrl->exitcode;
1298 info1 = ctrl->exitinfo1;
1299 info2 = ctrl->exitinfo2;
1301 vmexit->exitcode = VM_EXITCODE_BOGUS;
1302 vmexit->rip = state->rip;
1303 vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1305 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1308 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1309 * in an inconsistent state and can trigger assertions that would
1310 * never happen otherwise.
1312 if (code == VMCB_EXIT_INVALID) {
1313 vm_exit_svm(vmexit, code, info1, info2);
1317 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1318 "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1320 KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1321 ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1322 vmexit->inst_length, code, info1, info2));
1324 svm_update_virqinfo(svm_sc, vcpu);
1325 svm_save_intinfo(svm_sc, vcpu);
1328 case VMCB_EXIT_IRET:
1330 * Restart execution at "iret" but with the intercept cleared.
1332 vmexit->inst_length = 0;
1333 clear_nmi_blocking(svm_sc, vcpu);
1336 case VMCB_EXIT_VINTR: /* interrupt window exiting */
1337 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1340 case VMCB_EXIT_INTR: /* external interrupt */
1341 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1344 case VMCB_EXIT_NMI: /* external NMI */
1348 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1350 idtvec = code - 0x40;
1354 * Call the machine check handler by hand. Also don't
1355 * reflect the machine check back into the guest.
1358 VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler");
1359 __asm __volatile("int $18");
1362 error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2,
1364 KASSERT(error == 0, ("%s: error %d updating cr2",
1384 * The 'nrip' field is populated for INT3, INTO and
1385 * BOUND exceptions and this also implies that
1386 * 'inst_length' is non-zero.
1388 * Reset 'inst_length' to zero so the guest %rip at
1389 * event injection is identical to what it was when
1390 * the exception originally happened.
1392 VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d "
1393 "to zero before injecting exception %d",
1394 vmexit->inst_length, idtvec);
1395 vmexit->inst_length = 0;
1402 KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1403 "when reflecting exception %d into guest",
1404 vmexit->inst_length, idtvec));
1407 /* Reflect the exception back into the guest */
1408 VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception "
1409 "%d/%#x into the guest", idtvec, (int)info1);
1410 error = vm_inject_exception(svm_sc->vm, vcpu, idtvec,
1411 errcode_valid, info1, 0);
1412 KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1417 case VMCB_EXIT_MSR: /* MSR access. */
1419 ecx = ctx->sctx_rcx;
1420 edx = ctx->sctx_rdx;
1424 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1425 val = (uint64_t)edx << 32 | eax;
1426 VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1428 if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1429 vmexit->exitcode = VM_EXITCODE_WRMSR;
1430 vmexit->u.msr.code = ecx;
1431 vmexit->u.msr.wval = val;
1435 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1436 ("emulate_wrmsr retu with bogus exitcode"));
1439 VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1440 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1441 if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1442 vmexit->exitcode = VM_EXITCODE_RDMSR;
1443 vmexit->u.msr.code = ecx;
1447 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1448 ("emulate_rdmsr retu with bogus exitcode"));
1453 handled = svm_handle_io(svm_sc, vcpu, vmexit);
1454 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1456 case VMCB_EXIT_CPUID:
1457 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1458 handled = x86_emulate_cpuid(svm_sc->vm, vcpu,
1459 (uint32_t *)&state->rax,
1460 (uint32_t *)&ctx->sctx_rbx,
1461 (uint32_t *)&ctx->sctx_rcx,
1462 (uint32_t *)&ctx->sctx_rdx);
1465 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1466 vmexit->exitcode = VM_EXITCODE_HLT;
1467 vmexit->u.hlt.rflags = state->rflags;
1469 case VMCB_EXIT_PAUSE:
1470 vmexit->exitcode = VM_EXITCODE_PAUSE;
1471 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1474 /* EXITINFO2 contains the faulting guest physical address */
1475 if (info1 & VMCB_NPF_INFO1_RSV) {
1476 VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1477 "reserved bits set: info1(%#lx) info2(%#lx)",
1479 } else if (vm_mem_allocated(svm_sc->vm, vcpu, info2)) {
1480 vmexit->exitcode = VM_EXITCODE_PAGING;
1481 vmexit->u.paging.gpa = info2;
1482 vmexit->u.paging.fault_type = npf_fault_type(info1);
1483 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1484 VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1485 "on gpa %#lx/%#lx at rip %#lx",
1486 info2, info1, state->rip);
1487 } else if (svm_npf_emul_fault(info1)) {
1488 svm_handle_inst_emul(vmcb, info2, vmexit);
1489 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1490 VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1491 "for gpa %#lx/%#lx at rip %#lx",
1492 info2, info1, state->rip);
1495 case VMCB_EXIT_MONITOR:
1496 vmexit->exitcode = VM_EXITCODE_MONITOR;
1498 case VMCB_EXIT_MWAIT:
1499 vmexit->exitcode = VM_EXITCODE_MWAIT;
1502 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1506 VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1507 handled ? "handled" : "unhandled", exit_reason_to_str(code),
1508 vmexit->rip, vmexit->inst_length);
1511 vmexit->rip += vmexit->inst_length;
1512 vmexit->inst_length = 0;
1513 state->rip = vmexit->rip;
1515 if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1517 * If this VM exit was not claimed by anybody then
1518 * treat it as a generic SVM exit.
1520 vm_exit_svm(vmexit, code, info1, info2);
1523 * The exitcode and collateral have been populated.
1524 * The VM exit will be processed further in userland.
1532 svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1536 if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1539 KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1540 "valid: %#lx", __func__, intinfo));
1542 svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1543 VMCB_EXITINTINFO_VECTOR(intinfo),
1544 VMCB_EXITINTINFO_EC(intinfo),
1545 VMCB_EXITINTINFO_EC_VALID(intinfo));
1546 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1547 VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1551 * Inject event to virtual cpu.
1554 svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1556 struct vmcb_ctrl *ctrl;
1557 struct vmcb_state *state;
1558 struct svm_vcpu *vcpustate;
1560 int vector, need_intr_window, pending_apic_vector;
1562 state = svm_get_vmcb_state(sc, vcpu);
1563 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1564 vcpustate = svm_get_vcpu(sc, vcpu);
1566 need_intr_window = 0;
1567 pending_apic_vector = 0;
1569 if (vcpustate->nextrip != state->rip) {
1570 ctrl->intr_shadow = 0;
1571 VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking "
1572 "cleared due to rip change: %#lx/%#lx",
1573 vcpustate->nextrip, state->rip);
1577 * Inject pending events or exceptions for this vcpu.
1579 * An event might be pending because the previous #VMEXIT happened
1580 * during event delivery (i.e. ctrl->exitintinfo).
1582 * An event might also be pending because an exception was injected
1583 * by the hypervisor (e.g. #PF during instruction emulation).
1585 svm_inj_intinfo(sc, vcpu);
1587 /* NMI event has priority over interrupts. */
1588 if (vm_nmi_pending(sc->vm, vcpu)) {
1589 if (nmi_blocked(sc, vcpu)) {
1591 * Can't inject another NMI if the guest has not
1592 * yet executed an "iret" after the last NMI.
1594 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1596 } else if (ctrl->intr_shadow) {
1598 * Can't inject an NMI if the vcpu is in an intr_shadow.
1600 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1601 "interrupt shadow");
1602 need_intr_window = 1;
1604 } else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1606 * If there is already an exception/interrupt pending
1607 * then defer the NMI until after that.
1609 VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1610 "eventinj %#lx", ctrl->eventinj);
1613 * Use self-IPI to trigger a VM-exit as soon as
1614 * possible after the event injection is completed.
1616 * This works only if the external interrupt exiting
1617 * is at a lower priority than the event injection.
1619 * Although not explicitly specified in APMv2 the
1620 * relative priorities were verified empirically.
1622 ipi_cpu(curcpu, IPI_AST); /* XXX vmm_ipinum? */
1624 vm_nmi_clear(sc->vm, vcpu);
1626 /* Inject NMI, vector number is not used */
1627 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1630 /* virtual NMI blocking is now in effect */
1631 enable_nmi_blocking(sc, vcpu);
1633 VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1637 if (!vm_extint_pending(sc->vm, vcpu)) {
1639 * APIC interrupts are delivered using the V_IRQ offload.
1641 * The primary benefit is that the hypervisor doesn't need to
1642 * deal with the various conditions that inhibit interrupts.
1643 * It also means that TPR changes via CR8 will be handled
1644 * without any hypervisor involvement.
1646 * Note that the APIC vector must remain pending in the vIRR
1647 * until it is confirmed that it was delivered to the guest.
1648 * This can be confirmed based on the value of V_IRQ at the
1649 * next #VMEXIT (1 = pending, 0 = delivered).
1651 * Also note that it is possible that another higher priority
1652 * vector can become pending before this vector is delivered
1653 * to the guest. This is alright because vcpu_notify_event()
1654 * will send an IPI and force the vcpu to trap back into the
1655 * hypervisor. The higher priority vector will be injected on
1658 if (vlapic_pending_intr(vlapic, &vector)) {
1659 KASSERT(vector >= 16 && vector <= 255,
1660 ("invalid vector %d from local APIC", vector));
1661 pending_apic_vector = vector;
1666 /* Ask the legacy pic for a vector to inject */
1667 vatpic_pending_intr(sc->vm, &vector);
1668 KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR",
1672 * If the guest has disabled interrupts or is in an interrupt shadow
1673 * then we cannot inject the pending interrupt.
1675 if ((state->rflags & PSL_I) == 0) {
1676 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1677 "rflags %#lx", vector, state->rflags);
1678 need_intr_window = 1;
1682 if (ctrl->intr_shadow) {
1683 VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1684 "interrupt shadow", vector);
1685 need_intr_window = 1;
1689 if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1690 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1691 "eventinj %#lx", vector, ctrl->eventinj);
1692 need_intr_window = 1;
1697 * Legacy PIC interrupts are delivered via the event injection
1700 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1702 vm_extint_clear(sc->vm, vcpu);
1703 vatpic_intr_accepted(sc->vm, vector);
1706 * Force a VM-exit as soon as the vcpu is ready to accept another
1707 * interrupt. This is done because the PIC might have another vector
1708 * that it wants to inject. Also, if the APIC has a pending interrupt
1709 * that was preempted by the ExtInt then it allows us to inject the
1710 * APIC vector as soon as possible.
1712 need_intr_window = 1;
1715 * The guest can modify the TPR by writing to %CR8. In guest mode
1716 * the processor reflects this write to V_TPR without hypervisor
1719 * The guest can also modify the TPR by writing to it via the memory
1720 * mapped APIC page. In this case, the write will be emulated by the
1721 * hypervisor. For this reason V_TPR must be updated before every
1724 v_tpr = vlapic_get_cr8(vlapic);
1725 KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1726 if (ctrl->v_tpr != v_tpr) {
1727 VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1728 ctrl->v_tpr, v_tpr);
1729 ctrl->v_tpr = v_tpr;
1730 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1733 if (pending_apic_vector) {
1735 * If an APIC vector is being injected then interrupt window
1736 * exiting is not possible on this VMRUN.
1738 KASSERT(!need_intr_window, ("intr_window exiting impossible"));
1739 VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ",
1740 pending_apic_vector);
1743 ctrl->v_ign_tpr = 0;
1744 ctrl->v_intr_vector = pending_apic_vector;
1745 ctrl->v_intr_prio = pending_apic_vector >> 4;
1746 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1747 } else if (need_intr_window) {
1749 * We use V_IRQ in conjunction with the VINTR intercept to
1750 * trap into the hypervisor as soon as a virtual interrupt
1753 * Since injected events are not subject to intercept checks
1754 * we need to ensure that the V_IRQ is not actually going to
1755 * be delivered on VM entry. The KASSERT below enforces this.
1757 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1758 (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1759 ("Bogus intr_window_exiting: eventinj (%#lx), "
1760 "intr_shadow (%u), rflags (%#lx)",
1761 ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1762 enable_intr_window_exiting(sc, vcpu);
1764 disable_intr_window_exiting(sc, vcpu);
1768 static __inline void
1769 restore_host_tss(void)
1771 struct system_segment_descriptor *tss_sd;
1774 * The TSS descriptor was in use prior to launching the guest so it
1775 * has been marked busy.
1777 * 'ltr' requires the descriptor to be marked available so change the
1778 * type to "64-bit available TSS".
1780 tss_sd = PCPU_GET(tss);
1781 tss_sd->sd_type = SDT_SYSTSS;
1782 ltr(GSEL(GPROC0_SEL, SEL_KPL));
1786 check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu)
1788 struct svm_vcpu *vcpustate;
1789 struct vmcb_ctrl *ctrl;
1793 KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not "
1794 "active on cpu %u", __func__, thiscpu));
1796 vcpustate = svm_get_vcpu(sc, vcpuid);
1797 ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1800 * The TLB entries associated with the vcpu's ASID are not valid
1801 * if either of the following conditions is true:
1803 * 1. The vcpu's ASID generation is different than the host cpu's
1804 * ASID generation. This happens when the vcpu migrates to a new
1805 * host cpu. It can also happen when the number of vcpus executing
1806 * on a host cpu is greater than the number of ASIDs available.
1808 * 2. The pmap generation number is different than the value cached in
1809 * the 'vcpustate'. This happens when the host invalidates pages
1810 * belonging to the guest.
1812 * asidgen eptgen Action
1819 * (a) There is no mismatch in eptgen or ASID generation and therefore
1820 * no further action is needed.
1822 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1823 * retained and the TLB entries associated with this ASID
1824 * are flushed by VMRUN.
1826 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1829 * (c) A new ASID is allocated.
1831 * (d) A new ASID is allocated.
1835 eptgen = pmap->pm_eptgen;
1836 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1838 if (vcpustate->asid.gen != asid[thiscpu].gen) {
1839 alloc_asid = true; /* (c) and (d) */
1840 } else if (vcpustate->eptgen != eptgen) {
1841 if (flush_by_asid())
1842 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; /* (b1) */
1844 alloc_asid = true; /* (b2) */
1847 * This is the common case (a).
1849 KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1850 KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1851 ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1855 if (++asid[thiscpu].num >= nasid) {
1856 asid[thiscpu].num = 1;
1857 if (++asid[thiscpu].gen == 0)
1858 asid[thiscpu].gen = 1;
1860 * If this cpu does not support "flush-by-asid"
1861 * then flush the entire TLB on a generation
1862 * bump. Subsequent ASID allocation in this
1863 * generation can be done without a TLB flush.
1865 if (!flush_by_asid())
1866 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1868 vcpustate->asid.gen = asid[thiscpu].gen;
1869 vcpustate->asid.num = asid[thiscpu].num;
1871 ctrl->asid = vcpustate->asid.num;
1872 svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1874 * If this cpu supports "flush-by-asid" then the TLB
1875 * was not flushed after the generation bump. The TLB
1876 * is flushed selectively after every new ASID allocation.
1878 if (flush_by_asid())
1879 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1881 vcpustate->eptgen = eptgen;
1883 KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1884 KASSERT(ctrl->asid == vcpustate->asid.num,
1885 ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1888 static __inline void
1892 __asm __volatile("clgi");
1895 static __inline void
1899 __asm __volatile("stgi");
1903 * Start vcpu with specified RIP.
1906 svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap,
1907 struct vm_eventinfo *evinfo)
1909 struct svm_regctx *gctx;
1910 struct svm_softc *svm_sc;
1911 struct svm_vcpu *vcpustate;
1912 struct vmcb_state *state;
1913 struct vmcb_ctrl *ctrl;
1914 struct vm_exit *vmexit;
1915 struct vlapic *vlapic;
1923 vcpustate = svm_get_vcpu(svm_sc, vcpu);
1924 state = svm_get_vmcb_state(svm_sc, vcpu);
1925 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
1926 vmexit = vm_exitinfo(vm, vcpu);
1927 vlapic = vm_lapic(vm, vcpu);
1929 gctx = svm_get_guest_regctx(svm_sc, vcpu);
1930 vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
1932 if (vcpustate->lastcpu != curcpu) {
1934 * Force new ASID allocation by invalidating the generation.
1936 vcpustate->asid.gen = 0;
1939 * Invalidate the VMCB state cache by marking all fields dirty.
1941 svm_set_dirty(svm_sc, vcpu, 0xffffffff);
1945 * Setting 'vcpustate->lastcpu' here is bit premature because
1946 * we may return from this function without actually executing
1947 * the VMRUN instruction. This could happen if a rendezvous
1948 * or an AST is pending on the first time through the loop.
1950 * This works for now but any new side-effects of vcpu
1951 * migration should take this case into account.
1953 vcpustate->lastcpu = curcpu;
1954 vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
1957 svm_msr_guest_enter(svm_sc, vcpu);
1959 /* Update Guest RIP */
1964 * Disable global interrupts to guarantee atomicity during
1965 * loading of guest state. This includes not only the state
1966 * loaded by the "vmrun" instruction but also software state
1967 * maintained by the hypervisor: suspended and rendezvous
1968 * state, NPT generation number, vlapic interrupts etc.
1972 if (vcpu_suspended(evinfo)) {
1974 vm_exit_suspended(vm, vcpu, state->rip);
1978 if (vcpu_rendezvous_pending(evinfo)) {
1980 vm_exit_rendezvous(vm, vcpu, state->rip);
1984 if (vcpu_reqidle(evinfo)) {
1986 vm_exit_reqidle(vm, vcpu, state->rip);
1990 /* We are asked to give the cpu by scheduler. */
1991 if (vcpu_should_yield(vm, vcpu)) {
1993 vm_exit_astpending(vm, vcpu, state->rip);
1997 svm_inj_interrupts(svm_sc, vcpu, vlapic);
1999 /* Activate the nested pmap on 'curcpu' */
2000 CPU_SET_ATOMIC_ACQ(curcpu, &pmap->pm_active);
2003 * Check the pmap generation and the ASID generation to
2004 * ensure that the vcpu does not use stale TLB mappings.
2006 check_asid(svm_sc, vcpu, pmap, curcpu);
2008 ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
2009 vcpustate->dirty = 0;
2010 VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
2012 /* Launch Virtual Machine. */
2013 VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
2014 svm_launch(vmcb_pa, gctx, &__pcpu[curcpu]);
2016 CPU_CLR_ATOMIC(curcpu, &pmap->pm_active);
2019 * The host GDTR and IDTR is saved by VMRUN and restored
2020 * automatically on #VMEXIT. However, the host TSS needs
2021 * to be restored explicitly.
2025 /* #VMEXIT disables interrupts so re-enable them here. */
2028 /* Update 'nextrip' */
2029 vcpustate->nextrip = state->rip;
2031 /* Handle #VMEXIT and if required return to user space. */
2032 handled = svm_vmexit(svm_sc, vcpu, vmexit);
2035 svm_msr_guest_exit(svm_sc, vcpu);
2041 svm_vmcleanup(void *arg)
2043 struct svm_softc *sc = arg;
2049 swctx_regptr(struct svm_regctx *regctx, int reg)
2053 case VM_REG_GUEST_RBX:
2054 return (®ctx->sctx_rbx);
2055 case VM_REG_GUEST_RCX:
2056 return (®ctx->sctx_rcx);
2057 case VM_REG_GUEST_RDX:
2058 return (®ctx->sctx_rdx);
2059 case VM_REG_GUEST_RDI:
2060 return (®ctx->sctx_rdi);
2061 case VM_REG_GUEST_RSI:
2062 return (®ctx->sctx_rsi);
2063 case VM_REG_GUEST_RBP:
2064 return (®ctx->sctx_rbp);
2065 case VM_REG_GUEST_R8:
2066 return (®ctx->sctx_r8);
2067 case VM_REG_GUEST_R9:
2068 return (®ctx->sctx_r9);
2069 case VM_REG_GUEST_R10:
2070 return (®ctx->sctx_r10);
2071 case VM_REG_GUEST_R11:
2072 return (®ctx->sctx_r11);
2073 case VM_REG_GUEST_R12:
2074 return (®ctx->sctx_r12);
2075 case VM_REG_GUEST_R13:
2076 return (®ctx->sctx_r13);
2077 case VM_REG_GUEST_R14:
2078 return (®ctx->sctx_r14);
2079 case VM_REG_GUEST_R15:
2080 return (®ctx->sctx_r15);
2087 svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
2089 struct svm_softc *svm_sc;
2094 if (ident == VM_REG_GUEST_INTR_SHADOW) {
2095 return (svm_get_intr_shadow(svm_sc, vcpu, val));
2098 if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2102 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2109 VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident);
2114 svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2116 struct svm_softc *svm_sc;
2121 if (ident == VM_REG_GUEST_INTR_SHADOW) {
2122 return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2125 if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2129 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2137 * XXX deal with CR3 and invalidate TLB entries tagged with the
2138 * vcpu's ASID. This needs to be treated differently depending on
2139 * whether 'running' is true/false.
2142 VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident);
2147 svm_setcap(void *arg, int vcpu, int type, int val)
2149 struct svm_softc *sc;
2155 case VM_CAP_HALT_EXIT:
2156 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2157 VMCB_INTCPT_HLT, val);
2159 case VM_CAP_PAUSE_EXIT:
2160 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2161 VMCB_INTCPT_PAUSE, val);
2163 case VM_CAP_UNRESTRICTED_GUEST:
2164 /* Unrestricted guest execution cannot be disabled in SVM */
2176 svm_getcap(void *arg, int vcpu, int type, int *retval)
2178 struct svm_softc *sc;
2185 case VM_CAP_HALT_EXIT:
2186 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2189 case VM_CAP_PAUSE_EXIT:
2190 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2193 case VM_CAP_UNRESTRICTED_GUEST:
2194 *retval = 1; /* unrestricted guest is always enabled */
2203 static struct vlapic *
2204 svm_vlapic_init(void *arg, int vcpuid)
2206 struct svm_softc *svm_sc;
2207 struct vlapic *vlapic;
2210 vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2211 vlapic->vm = svm_sc->vm;
2212 vlapic->vcpuid = vcpuid;
2213 vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2215 vlapic_init(vlapic);
2221 svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2224 vlapic_cleanup(vlapic);
2225 free(vlapic, M_SVM_VLAPIC);
2228 struct vmm_ops vmm_ops_amd = {