svm.c revision 284894
11541Srgrimes/*- 21541Srgrimes * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com) 31541Srgrimes * All rights reserved. 41541Srgrimes * 551138Salfred * Redistribution and use in source and binary forms, with or without 61541Srgrimes * modification, are permitted provided that the following conditions 71541Srgrimes * are met: 8106149Sdwmalone * 1. Redistributions of source code must retain the above copyright 91541Srgrimes * notice unmodified, this list of conditions, and the following 1064002Speter * disclaimer. 111541Srgrimes * 2. Redistributions in binary form must reproduce the above copyright 121541Srgrimes * notice, this list of conditions and the following disclaimer in the 131541Srgrimes * documentation and/or other materials provided with the distribution. 141541Srgrimes * 151541Srgrimes * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 161541Srgrimes * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17171210Speter * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 181541Srgrimes * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 191541Srgrimes * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 201541Srgrimes * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 211541Srgrimes * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 221541Srgrimes * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 231541Srgrimes * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 241541Srgrimes * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 251541Srgrimes */ 261541Srgrimes 27194392Sjhb#include <sys/cdefs.h> 28171210Speter__FBSDID("$FreeBSD: stable/10/sys/amd64/vmm/amd/svm.c 284894 2015-06-27 22:48:22Z neel $"); 291541Srgrimes 301541Srgrimes#include <sys/param.h> 311541Srgrimes#include <sys/systm.h> 321541Srgrimes#include <sys/smp.h> 331541Srgrimes#include <sys/kernel.h> 341541Srgrimes#include <sys/malloc.h> 351541Srgrimes#include <sys/pcpu.h> 361541Srgrimes#include <sys/proc.h> 371541Srgrimes#include <sys/sysctl.h> 381541Srgrimes 391541Srgrimes#include <vm/vm.h> 401541Srgrimes#include <vm/pmap.h> 411541Srgrimes 421541Srgrimes#include <machine/cpufunc.h> 431541Srgrimes#include <machine/psl.h> 441541Srgrimes#include <machine/pmap.h> 451541Srgrimes#include <machine/md_var.h> 461541Srgrimes#include <machine/specialreg.h> 47171210Speter#include <machine/smp.h> 481541Srgrimes#include <machine/vmm.h> 49171210Speter#include <machine/vmm_dev.h> 501541Srgrimes#include <machine/vmm_instruction_emul.h> 511541Srgrimes 521541Srgrimes#include "vmm_lapic.h" 531541Srgrimes#include "vmm_stat.h" 541541Srgrimes#include "vmm_ktr.h" 55171210Speter#include "vmm_ioport.h" 561541Srgrimes#include "vatpic.h" 57171210Speter#include "vlapic.h" 581541Srgrimes#include "vlapic_priv.h" 591541Srgrimes 601541Srgrimes#include "x86.h" 61171210Speter#include "vmcb.h" 621541Srgrimes#include "svm.h" 631541Srgrimes#include "svm_softc.h" 641541Srgrimes#include "svm_msr.h" 651541Srgrimes#include "npt.h" 661541Srgrimes 671541SrgrimesSYSCTL_DECL(_hw_vmm); 681541SrgrimesSYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL); 691541Srgrimes 701541Srgrimes/* 71171210Speter * SVM CPUID function 0x8000_000A, edx bit decoding. 72171210Speter */ 73171210Speter#define AMD_CPUID_SVM_NP BIT(0) /* Nested paging or RVI */ 741541Srgrimes#define AMD_CPUID_SVM_LBR BIT(1) /* Last branch virtualization */ 751541Srgrimes#define AMD_CPUID_SVM_SVML BIT(2) /* SVM lock */ 761541Srgrimes#define AMD_CPUID_SVM_NRIP_SAVE BIT(3) /* Next RIP is saved */ 771541Srgrimes#define AMD_CPUID_SVM_TSC_RATE BIT(4) /* TSC rate control. */ 781541Srgrimes#define AMD_CPUID_SVM_VMCB_CLEAN BIT(5) /* VMCB state caching */ 791541Srgrimes#define AMD_CPUID_SVM_FLUSH_BY_ASID BIT(6) /* Flush by ASID */ 80171210Speter#define AMD_CPUID_SVM_DECODE_ASSIST BIT(7) /* Decode assist */ 811541Srgrimes#define AMD_CPUID_SVM_PAUSE_INC BIT(10) /* Pause intercept filter. */ 821541Srgrimes#define AMD_CPUID_SVM_PAUSE_FTH BIT(12) /* Pause filter threshold */ 831541Srgrimes#define AMD_CPUID_SVM_AVIC BIT(13) /* AVIC present */ 841541Srgrimes 851541Srgrimes#define VMCB_CACHE_DEFAULT (VMCB_CACHE_ASID | \ 861541Srgrimes VMCB_CACHE_IOPM | \ 871541Srgrimes VMCB_CACHE_I | \ 881541Srgrimes VMCB_CACHE_TPR | \ 891541Srgrimes VMCB_CACHE_CR2 | \ 901541Srgrimes VMCB_CACHE_CR | \ 911541Srgrimes VMCB_CACHE_DT | \ 921541Srgrimes VMCB_CACHE_SEG | \ 93171210Speter VMCB_CACHE_NP) 941541Srgrimes 951541Srgrimesstatic uint32_t vmcb_clean = VMCB_CACHE_DEFAULT; 96171210SpeterSYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean, 97171210Speter 0, NULL); 981541Srgrimes 991541Srgrimesstatic MALLOC_DEFINE(M_SVM, "svm", "svm"); 1001541Srgrimesstatic MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic"); 1011541Srgrimes 1021541Srgrimes/* Per-CPU context area. */ 1031541Srgrimesextern struct pcpu __pcpu[]; 1041541Srgrimes 1051541Srgrimesstatic uint32_t svm_feature; /* AMD SVM features. */ 1061541SrgrimesSYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RD, &svm_feature, 0, 1071541Srgrimes "SVM features advertised by CPUID.8000000AH:EDX"); 108171210Speter 1091541Srgrimesstatic int disable_npf_assist; 110171210SpeterSYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN, 111171210Speter &disable_npf_assist, 0, NULL); 112171210Speter 1131541Srgrimes/* Maximum ASIDs supported by the processor */ 1141541Srgrimesstatic uint32_t nasid; 1151541SrgrimesSYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RD, &nasid, 0, 1161541Srgrimes "Number of ASIDs supported by this processor"); 117171210Speter 118171210Speter/* Current ASID generation for each host cpu */ 119171210Speterstatic struct asid asid[MAXCPU]; 120171210Speter 121171210Speter/* 122171210Speter * SVM host state saved area of size 4KB for each core. 123171210Speter */ 1241541Srgrimesstatic uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE); 1251541Srgrimes 1261541Srgrimesstatic VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery"); 1271541Srgrimesstatic VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry"); 12814220Speterstatic VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window"); 1291541Srgrimes 1301541Srgrimesstatic int svm_setreg(void *arg, int vcpu, int ident, uint64_t val); 1311541Srgrimes 1321541Srgrimesstatic __inline int 1331541Srgrimesflush_by_asid(void) 134171210Speter{ 1358019Sache 1368019Sache return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID); 1371541Srgrimes} 138171210Speter 139171210Speterstatic __inline int 1401541Srgrimesdecode_assist(void) 1411541Srgrimes{ 1421541Srgrimes 1431541Srgrimes return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST); 1441541Srgrimes} 1451541Srgrimes 1461541Srgrimesstatic void 1471541Srgrimessvm_disable(void *arg __unused) 1481541Srgrimes{ 1491541Srgrimes uint64_t efer; 150171210Speter 151171210Speter efer = rdmsr(MSR_EFER); 152171210Speter efer &= ~EFER_SVM; 153171210Speter wrmsr(MSR_EFER, efer); 154171210Speter} 155171210Speter 1561541Srgrimes/* 1571541Srgrimes * Disable SVM on all CPUs. 158171210Speter */ 159171210Speterstatic int 16014220Spetersvm_cleanup(void) 16114220Speter{ 16214220Speter 163177634Sdfr smp_rendezvous(NULL, svm_disable, NULL, NULL); 1641541Srgrimes return (0); 165171210Speter} 166194392Sjhb 167194392Sjhb/* 1681541Srgrimes * Verify that all the features required by bhyve are available. 169127891Sdfr */ 1701541Srgrimesstatic int 171194392Sjhbcheck_svm_features(void) 172194392Sjhb{ 173194392Sjhb u_int regs[4]; 1741549Srgrimes 1752442Sdg /* CPUID Fn8000_000A is for SVM */ 1761541Srgrimes do_cpuid(0x8000000A, regs); 1771541Srgrimes svm_feature = regs[3]; 1782729Sdfr 1792729Sdfr nasid = regs[1]; 1801541Srgrimes KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid)); 1811541Srgrimes 182171210Speter /* bhyve requires the Nested Paging feature */ 183171210Speter if (!(svm_feature & AMD_CPUID_SVM_NP)) { 184178888Sjulian printf("SVM: Nested Paging feature not available.\n"); 1852297Swollman return (ENXIO); 18614220Speter } 18714220Speter 18814220Speter /* bhyve requires the NRIP Save feature */ 1891541Srgrimes if (!(svm_feature & AMD_CPUID_SVM_NRIP_SAVE)) { 1901541Srgrimes printf("SVM: NRIP Save feature not available.\n"); 1911541Srgrimes return (ENXIO); 1921541Srgrimes } 19332889Sphk 19432889Sphk return (0); 19532889Sphk} 19632889Sphk 1971541Srgrimesstatic void 1981541Srgrimessvm_enable(void *arg __unused) 1991541Srgrimes{ 2001541Srgrimes uint64_t efer; 2011541Srgrimes 2021541Srgrimes efer = rdmsr(MSR_EFER); 2031541Srgrimes efer |= EFER_SVM; 2041541Srgrimes wrmsr(MSR_EFER, efer); 2051541Srgrimes 206171210Speter wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave[curcpu])); 2071541Srgrimes} 208171210Speter 209171210Speter/* 210171210Speter * Return 1 if SVM is enabled on this processor and 0 otherwise. 2111541Srgrimes */ 2121541Srgrimesstatic int 2131541Srgrimessvm_available(void) 21435938Sdyson{ 21535938Sdyson uint64_t msr; 21628400Speter 21725582Speter /* Section 15.4 Enabling SVM from APM2. */ 21829349Speter if ((amd_feature2 & AMDID2_SVM) == 0) { 2192124Sdg printf("SVM: not available.\n"); 2202124Sdg return (0); 2212124Sdg } 2222124Sdg 2232124Sdg msr = rdmsr(MSR_VM_CR); 2242124Sdg if ((msr & VM_CR_SVMDIS) != 0) { 2252124Sdg printf("SVM: disabled by BIOS.\n"); 2262124Sdg return (0); 2272124Sdg } 2282124Sdg 229194919Sjhb return (1); 23012865Speter} 23112865Speter 23259829Speterstatic int 233194919Sjhbsvm_init(int ipinum) 23412865Speter{ 23512865Speter int error, cpu; 23612865Speter 23712865Speter if (!svm_available()) 238194919Sjhb return (ENXIO); 23912865Speter 24012865Speter error = check_svm_features(); 24125582Speter if (error) 24225582Speter return (error); 24325582Speter 244156138Sdavidxu vmcb_clean &= VMCB_CACHE_DEFAULT; 245156138Sdavidxu 246156138Sdavidxu for (cpu = 0; cpu < MAXCPU; cpu++) { 247156138Sdavidxu /* 248156138Sdavidxu * Initialize the host ASIDs to their "highest" valid values. 24925582Speter * 250227776Slstewart * The next ASID allocation will rollover both 'gen' and 'num' 251227776Slstewart * and start off the sequence at {1,1}. 252227776Slstewart */ 25314220Speter asid[cpu].gen = ~0UL; 25414220Speter asid[cpu].num = nasid - 1; 25514220Speter } 256239347Sdavidxu 257137875Smarks svm_msr_init(); 25814220Speter svm_npt_init(ipinum); 25914220Speter 26014220Speter /* Enable SVM on all CPUs */ 26129349Speter smp_rendezvous(NULL, svm_enable, NULL, NULL); 26224452Speter 26324440Speter return (0); 264151868Sdavidxu} 265151868Sdavidxu 266151868Sdavidxustatic void 267152846Sdavidxusvm_restore(void) 268152846Sdavidxu{ 269152846Sdavidxu 270152846Sdavidxu svm_enable(NULL); 271152846Sdavidxu} 272152846Sdavidxu 27325537Sdfr/* Pentium compatible MSRs */ 27425537Sdfr#define MSR_PENTIUM_START 0 27525537Sdfr#define MSR_PENTIUM_END 0x1FFF 27625537Sdfr/* AMD 6th generation and Intel compatible MSRs */ 27725537Sdfr#define MSR_AMD6TH_START 0xC0000000UL 27825537Sdfr#define MSR_AMD6TH_END 0xC0001FFFUL 27925537Sdfr/* AMD 7th and 8th generation compatible MSRs */ 28025537Sdfr#define MSR_AMD7TH_START 0xC0010000UL 28135938Sdyson#define MSR_AMD7TH_END 0xC0011FFFUL 28225537Sdfr 28335938Sdyson/* 28435938Sdyson * Get the index and bit position for a MSR in permission bitmap. 28535938Sdyson * Two bits are used for each MSR: lower bit for read and higher bit for write. 28635938Sdyson */ 28735938Sdysonstatic int 28835938Sdysonsvm_msr_index(uint64_t msr, int *index, int *bit) 28935938Sdyson{ 29025537Sdfr uint32_t base, off; 29125537Sdfr 29225537Sdfr *index = -1; 29325537Sdfr *bit = (msr % 4) * 2; 29425537Sdfr base = 0; 29525537Sdfr 29625537Sdfr if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) { 29725537Sdfr *index = msr / 4; 298147814Sjhb return (0); 299147814Sjhb } 30025537Sdfr 30125537Sdfr base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1); 30225537Sdfr if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) { 30325537Sdfr off = (msr - MSR_AMD6TH_START); 30425537Sdfr *index = (off + base) / 4; 30525537Sdfr return (0); 306194392Sjhb } 30751138Salfred 30851138Salfred base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1); 30925537Sdfr if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) { 31025537Sdfr off = (msr - MSR_AMD7TH_START); 31125537Sdfr *index = (off + base) / 4; 31225537Sdfr return (0); 31325537Sdfr } 31425537Sdfr 31525537Sdfr return (EINVAL); 31625537Sdfr} 31725537Sdfr 31825537Sdfr/* 31928400Speter * Allow vcpu to read or write the 'msr' without trapping into the hypervisor. 32056115Speter */ 32156115Speterstatic void 32236034Spetersvm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write) 32326671Sdyson{ 32426671Sdyson int index, bit, error; 32526671Sdyson 32626671Sdyson error = svm_msr_index(msr, &index, &bit); 327151868Sdavidxu KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr)); 328151868Sdavidxu KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE, 329151868Sdavidxu ("%s: invalid index %d for msr %#lx", __func__, index, msr)); 33026671Sdyson KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d " 33169514Sjake "msr %#lx", __func__, bit, msr)); 33269514Sjake 33326671Sdyson if (read) 33426671Sdyson perm_bitmap[index] &= ~(1UL << bit); 33529391Sphk 33634925Sdufault if (write) 33734925Sdufault perm_bitmap[index] &= ~(2UL << bit); 33834925Sdufault} 33934925Sdufault 34034925Sdufaultstatic void 34134925Sdufaultsvm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr) 34234925Sdufault{ 34334925Sdufault 34435938Sdyson svm_msr_perm(perm_bitmap, msr, true, true); 345194392Sjhb} 34641089Speter 34746155Sphkstatic void 348211999Skibsvm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr) 34951791Smarcel{ 35051791Smarcel 351194392Sjhb svm_msr_perm(perm_bitmap, msr, true, false); 35251791Smarcel} 353194392Sjhb 354112895Sjeffstatic __inline int 355112895Sjeffsvm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask) 35656271Srwatson{ 35756271Srwatson struct vmcb_ctrl *ctrl; 35856271Srwatson 35956271Srwatson KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx)); 36056271Srwatson 36156271Srwatson ctrl = svm_get_vmcb_ctrl(sc, vcpu); 36256271Srwatson return (ctrl->intercept[idx] & bitmask ? 1 : 0); 36356271Srwatson} 36454803Srwatson 36554803Srwatsonstatic __inline void 36654803Srwatsonsvm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask, 36754803Srwatson int enabled) 36855943Sjasone{ 36956115Speter struct vmcb_ctrl *ctrl; 37056115Speter uint32_t oldval; 37159288Sjlemon 37259288Sjlemon KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx)); 37398198Srwatson 37498198Srwatson ctrl = svm_get_vmcb_ctrl(sc, vcpu); 37598198Srwatson oldval = ctrl->intercept[idx]; 37698198Srwatson 37798198Srwatson if (enabled) 37898198Srwatson ctrl->intercept[idx] |= bitmask; 379183362Sjhb else 38075039Srwatson ctrl->intercept[idx] &= ~bitmask; 38175039Srwatson 38275039Srwatson if (ctrl->intercept[idx] != oldval) { 38375427Srwatson svm_set_dirty(sc, vcpu, VMCB_CACHE_I); 384194392Sjhb VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified " 38583796Srwatson "from %#x to %#x", idx, oldval, ctrl->intercept[idx]); 386211999Skib } 38785891Sphk} 388177091Sjeff 389177091Sjeffstatic __inline void 390177091Sjeffsvm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask) 391177091Sjeff{ 392177091Sjeff 393100897Srwatson svm_set_intercept(sc, vcpu, off, bitmask, 0); 394100897Srwatson} 395100897Srwatson 396100897Srwatsonstatic __inline void 397100897Srwatsonsvm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask) 398100897Srwatson{ 39994936Smux 40096084Smux svm_set_intercept(sc, vcpu, off, bitmask, 1); 40197372Smarcel} 40299856Salfred 403101426Srwatsonstatic void 404122540Smckusickvmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa, 405122540Smckusick uint64_t msrpm_base_pa, uint64_t np_pml4) 406122540Smckusick{ 407122540Smckusick struct vmcb_ctrl *ctrl; 408103575Salfred struct vmcb_state *state; 409103575Salfred uint32_t mask; 410103575Salfred int n; 411103575Salfred 412103575Salfred ctrl = svm_get_vmcb_ctrl(sc, vcpu); 413103575Salfred state = svm_get_vmcb_state(sc, vcpu); 414103575Salfred 415103575Salfred ctrl->iopm_base_pa = iopm_base_pa; 416103575Salfred ctrl->msrpm_base_pa = msrpm_base_pa; 417103575Salfred 418105692Srwatson /* Enable nested paging */ 419105692Srwatson ctrl->np_enable = 1; 420105692Srwatson ctrl->n_cr3 = np_pml4; 421104731Srwatson 422104731Srwatson /* 423104731Srwatson * Intercept accesses to the control registers that are not shadowed 424106467Srwatson * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8. 425105950Speter */ 426105950Speter for (n = 0; n < 16; n++) { 427105692Srwatson mask = (BIT(n) << 16) | BIT(n); 428105692Srwatson if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8) 429105692Srwatson svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask); 430106978Sdeischen else 431106978Sdeischen svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask); 432106978Sdeischen } 433107914Sdillon 434108406Srwatson 435108406Srwatson /* 436108406Srwatson * Intercept everything when tracing guest exceptions otherwise 437108406Srwatson * just intercept machine check exception. 438112895Sjeff */ 439112902Sjeff if (vcpu_trace_exceptions(sc->vm, vcpu)) { 440112902Sjeff for (n = 0; n < 32; n++) { 441112902Sjeff /* 442112902Sjeff * Skip unimplemented vectors in the exception bitmap. 443112909Sjeff */ 444112909Sjeff if (n == 2 || n == 9) { 445113276Smike continue; 446115800Srwatson } 447115800Srwatson svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(n)); 448115800Srwatson } 449177091Sjeff } else { 450125369Sdeischen svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC)); 451127484Smtm } 452127484Smtm 453132117Sphk /* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */ 454136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO); 455136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR); 456136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID); 457136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR); 458136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT); 459136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI); 460136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI); 461136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN); 462136831Srwatson svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, 463139013Sdavidxu VMCB_INTCPT_FERR_FREEZE); 464145435Sdavidxu 465151317Sdavidxu svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR); 466156138Sdavidxu svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT); 467156138Sdavidxu 468156138Sdavidxu /* 469156138Sdavidxu * From section "Canonicalization and Consistency Checks" in APMv2 470156138Sdavidxu * the VMRUN intercept bit must be set to pass the consistency check. 471156138Sdavidxu */ 472153681Sphk svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN); 473155328Sdavidxu 474157039Sdavidxu /* 475162498Sdavidxu * The ASID will be set to a non-zero value just before VMRUN. 476162498Sdavidxu */ 477162498Sdavidxu ctrl->asid = 0; 478161679Sdavidxu 479161679Sdavidxu /* 480163953Srrs * Section 15.21.1, Interrupt Masking in EFLAGS 481163953Srrs * Section 15.21.2, Virtualizing APIC.TPR 482163953Srrs * 483163953Srrs * This must be set for %rflag and %cr8 isolation of guest and host. 484171210Speter */ 485171210Speter ctrl->v_intr_masking = 1; 486171210Speter 487171210Speter /* Enable Last Branch Record aka LBR for debugging */ 488171210Speter ctrl->lbr_virt_en = 1; 489171210Speter state->dbgctl = BIT(0); 490171861Sdavidxu 491175165Sjhb /* EFER_SVM must always be set when the guest is executing */ 492175165Sjhb state->efer = EFER_SVM; 493176731Sjeff 494176731Sjeff /* Set up the PAT to power-on state */ 495176731Sjeff state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK) | 496176731Sjeff PAT_VALUE(1, PAT_WRITE_THROUGH) | 497176731Sjeff PAT_VALUE(2, PAT_UNCACHED) | 498177790Skib PAT_VALUE(3, PAT_UNCACHEABLE) | 499177790Skib PAT_VALUE(4, PAT_WRITE_BACK) | 500177790Skib PAT_VALUE(5, PAT_WRITE_THROUGH) | 501177790Skib PAT_VALUE(6, PAT_UNCACHED) | 502177790Skib PAT_VALUE(7, PAT_UNCACHEABLE); 503177790Skib} 504177790Skib 505177790Skib/* 506177790Skib * Initialize a virtual machine. 507177790Skib */ 508177790Skibstatic void * 509177790Skibsvm_vminit(struct vm *vm, pmap_t pmap) 510177790Skib{ 511177790Skib struct svm_softc *svm_sc; 512177790Skib struct svm_vcpu *vcpu; 513181905Sed vm_paddr_t msrpm_pa, iopm_pa, pml4_pa; 514184589Sdfr int i; 515191675Sjamie 516191675Sjamie svm_sc = malloc(sizeof (struct svm_softc), M_SVM, M_WAITOK | M_ZERO); 517191675Sjamie svm_sc->vm = vm; 518194263Sjhb svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4); 519194919Sjhb 520194919Sjhb /* 521194919Sjhb * Intercept read and write accesses to all MSRs. 522195459Strasz */ 523255220Spjd memset(svm_sc->msr_bitmap, 0xFF, sizeof(svm_sc->msr_bitmap)); 524255220Spjd 525219132Srwatson /* 526219132Srwatson * Access to the following MSRs is redirected to the VMCB when the 527224988Sjonathan * guest is executing. Therefore it is safe to allow the guest to 528224988Sjonathan * read/write these MSRs directly without hypervisor involvement. 529224988Sjonathan */ 530224988Sjonathan svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE); 531198510Skib svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE); 532219305Strasz svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE); 533219305Strasz 534220164Strasz svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR); 535220164Strasz svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR); 536220164Strasz svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR); 537220164Strasz svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK); 538220164Strasz svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR); 539220792Smdf svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR); 540227071Sjhb svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR); 541242959Skib svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT); 542247604Spjd 543247604Spjd svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC); 544247604Spjd 545247604Spjd /* 546247604Spjd * Intercept writes to make sure that the EFER_SVM bit is not cleared. 547247668Spjd */ 548247668Spjd svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER); 549248600Spjd 550250155Sjilles /* Intercept access to all I/O ports. */ 551250160Sjilles memset(svm_sc->iopm_bitmap, 0xFF, sizeof(svm_sc->iopm_bitmap)); 552251527Sglebius 553255709Sjhb iopm_pa = vtophys(svm_sc->iopm_bitmap); 554275987Sdchagin msrpm_pa = vtophys(svm_sc->msr_bitmap); 555293475Sdchagin pml4_pa = svm_sc->nptp; 556293475Sdchagin for (i = 0; i < VM_MAXCPU; i++) { 5571541Srgrimes vcpu = svm_get_vcpu(svm_sc, i); 558 vcpu->nextrip = ~0; 559 vcpu->lastcpu = NOCPU; 560 vcpu->vmcb_pa = vtophys(&vcpu->vmcb); 561 vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa); 562 svm_msr_guest_init(svm_sc, i); 563 } 564 return (svm_sc); 565} 566 567static int 568svm_cpl(struct vmcb_state *state) 569{ 570 571 /* 572 * From APMv2: 573 * "Retrieve the CPL from the CPL field in the VMCB, not 574 * from any segment DPL" 575 */ 576 return (state->cpl); 577} 578 579static enum vm_cpu_mode 580svm_vcpu_mode(struct vmcb *vmcb) 581{ 582 struct vmcb_segment seg; 583 struct vmcb_state *state; 584 int error; 585 586 state = &vmcb->state; 587 588 if (state->efer & EFER_LMA) { 589 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg); 590 KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__, 591 error)); 592 593 /* 594 * Section 4.8.1 for APM2, check if Code Segment has 595 * Long attribute set in descriptor. 596 */ 597 if (seg.attrib & VMCB_CS_ATTRIB_L) 598 return (CPU_MODE_64BIT); 599 else 600 return (CPU_MODE_COMPATIBILITY); 601 } else if (state->cr0 & CR0_PE) { 602 return (CPU_MODE_PROTECTED); 603 } else { 604 return (CPU_MODE_REAL); 605 } 606} 607 608static enum vm_paging_mode 609svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer) 610{ 611 612 if ((cr0 & CR0_PG) == 0) 613 return (PAGING_MODE_FLAT); 614 if ((cr4 & CR4_PAE) == 0) 615 return (PAGING_MODE_32); 616 if (efer & EFER_LME) 617 return (PAGING_MODE_64); 618 else 619 return (PAGING_MODE_PAE); 620} 621 622/* 623 * ins/outs utility routines 624 */ 625static uint64_t 626svm_inout_str_index(struct svm_regctx *regs, int in) 627{ 628 uint64_t val; 629 630 val = in ? regs->sctx_rdi : regs->sctx_rsi; 631 632 return (val); 633} 634 635static uint64_t 636svm_inout_str_count(struct svm_regctx *regs, int rep) 637{ 638 uint64_t val; 639 640 val = rep ? regs->sctx_rcx : 1; 641 642 return (val); 643} 644 645static void 646svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1, 647 int in, struct vm_inout_str *vis) 648{ 649 int error, s; 650 651 if (in) { 652 vis->seg_name = VM_REG_GUEST_ES; 653 } else { 654 /* The segment field has standard encoding */ 655 s = (info1 >> 10) & 0x7; 656 vis->seg_name = vm_segment_name(s); 657 } 658 659 error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc); 660 KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error)); 661} 662 663static int 664svm_inout_str_addrsize(uint64_t info1) 665{ 666 uint32_t size; 667 668 size = (info1 >> 7) & 0x7; 669 switch (size) { 670 case 1: 671 return (2); /* 16 bit */ 672 case 2: 673 return (4); /* 32 bit */ 674 case 4: 675 return (8); /* 64 bit */ 676 default: 677 panic("%s: invalid size encoding %d", __func__, size); 678 } 679} 680 681static void 682svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging) 683{ 684 struct vmcb_state *state; 685 686 state = &vmcb->state; 687 paging->cr3 = state->cr3; 688 paging->cpl = svm_cpl(state); 689 paging->cpu_mode = svm_vcpu_mode(vmcb); 690 paging->paging_mode = svm_paging_mode(state->cr0, state->cr4, 691 state->efer); 692} 693 694#define UNHANDLED 0 695 696/* 697 * Handle guest I/O intercept. 698 */ 699static int 700svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit) 701{ 702 struct vmcb_ctrl *ctrl; 703 struct vmcb_state *state; 704 struct svm_regctx *regs; 705 struct vm_inout_str *vis; 706 uint64_t info1; 707 int inout_string; 708 709 state = svm_get_vmcb_state(svm_sc, vcpu); 710 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); 711 regs = svm_get_guest_regctx(svm_sc, vcpu); 712 713 info1 = ctrl->exitinfo1; 714 inout_string = info1 & BIT(2) ? 1 : 0; 715 716 /* 717 * The effective segment number in EXITINFO1[12:10] is populated 718 * only if the processor has the DecodeAssist capability. 719 * 720 * XXX this is not specified explicitly in APMv2 but can be verified 721 * empirically. 722 */ 723 if (inout_string && !decode_assist()) 724 return (UNHANDLED); 725 726 vmexit->exitcode = VM_EXITCODE_INOUT; 727 vmexit->u.inout.in = (info1 & BIT(0)) ? 1 : 0; 728 vmexit->u.inout.string = inout_string; 729 vmexit->u.inout.rep = (info1 & BIT(3)) ? 1 : 0; 730 vmexit->u.inout.bytes = (info1 >> 4) & 0x7; 731 vmexit->u.inout.port = (uint16_t)(info1 >> 16); 732 vmexit->u.inout.eax = (uint32_t)(state->rax); 733 734 if (inout_string) { 735 vmexit->exitcode = VM_EXITCODE_INOUT_STR; 736 vis = &vmexit->u.inout_str; 737 svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging); 738 vis->rflags = state->rflags; 739 vis->cr0 = state->cr0; 740 vis->index = svm_inout_str_index(regs, vmexit->u.inout.in); 741 vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep); 742 vis->addrsize = svm_inout_str_addrsize(info1); 743 svm_inout_str_seginfo(svm_sc, vcpu, info1, 744 vmexit->u.inout.in, vis); 745 } 746 747 return (UNHANDLED); 748} 749 750static int 751npf_fault_type(uint64_t exitinfo1) 752{ 753 754 if (exitinfo1 & VMCB_NPF_INFO1_W) 755 return (VM_PROT_WRITE); 756 else if (exitinfo1 & VMCB_NPF_INFO1_ID) 757 return (VM_PROT_EXECUTE); 758 else 759 return (VM_PROT_READ); 760} 761 762static bool 763svm_npf_emul_fault(uint64_t exitinfo1) 764{ 765 766 if (exitinfo1 & VMCB_NPF_INFO1_ID) { 767 return (false); 768 } 769 770 if (exitinfo1 & VMCB_NPF_INFO1_GPT) { 771 return (false); 772 } 773 774 if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) { 775 return (false); 776 } 777 778 return (true); 779} 780 781static void 782svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit) 783{ 784 struct vm_guest_paging *paging; 785 struct vmcb_segment seg; 786 struct vmcb_ctrl *ctrl; 787 char *inst_bytes; 788 int error, inst_len; 789 790 ctrl = &vmcb->ctrl; 791 paging = &vmexit->u.inst_emul.paging; 792 793 vmexit->exitcode = VM_EXITCODE_INST_EMUL; 794 vmexit->u.inst_emul.gpa = gpa; 795 vmexit->u.inst_emul.gla = VIE_INVALID_GLA; 796 svm_paging_info(vmcb, paging); 797 798 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg); 799 KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error)); 800 801 switch(paging->cpu_mode) { 802 case CPU_MODE_PROTECTED: 803 case CPU_MODE_COMPATIBILITY: 804 /* 805 * Section 4.8.1 of APM2, Default Operand Size or D bit. 806 */ 807 vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ? 808 1 : 0; 809 break; 810 default: 811 vmexit->u.inst_emul.cs_d = 0; 812 break; 813 } 814 815 /* 816 * Copy the instruction bytes into 'vie' if available. 817 */ 818 if (decode_assist() && !disable_npf_assist) { 819 inst_len = ctrl->inst_len; 820 inst_bytes = ctrl->inst_bytes; 821 } else { 822 inst_len = 0; 823 inst_bytes = NULL; 824 } 825 vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len); 826} 827 828#ifdef KTR 829static const char * 830intrtype_to_str(int intr_type) 831{ 832 switch (intr_type) { 833 case VMCB_EVENTINJ_TYPE_INTR: 834 return ("hwintr"); 835 case VMCB_EVENTINJ_TYPE_NMI: 836 return ("nmi"); 837 case VMCB_EVENTINJ_TYPE_INTn: 838 return ("swintr"); 839 case VMCB_EVENTINJ_TYPE_EXCEPTION: 840 return ("exception"); 841 default: 842 panic("%s: unknown intr_type %d", __func__, intr_type); 843 } 844} 845#endif 846 847/* 848 * Inject an event to vcpu as described in section 15.20, "Event injection". 849 */ 850static void 851svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector, 852 uint32_t error, bool ec_valid) 853{ 854 struct vmcb_ctrl *ctrl; 855 856 ctrl = svm_get_vmcb_ctrl(sc, vcpu); 857 858 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, 859 ("%s: event already pending %#lx", __func__, ctrl->eventinj)); 860 861 KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d", 862 __func__, vector)); 863 864 switch (intr_type) { 865 case VMCB_EVENTINJ_TYPE_INTR: 866 case VMCB_EVENTINJ_TYPE_NMI: 867 case VMCB_EVENTINJ_TYPE_INTn: 868 break; 869 case VMCB_EVENTINJ_TYPE_EXCEPTION: 870 if (vector >= 0 && vector <= 31 && vector != 2) 871 break; 872 /* FALLTHROUGH */ 873 default: 874 panic("%s: invalid intr_type/vector: %d/%d", __func__, 875 intr_type, vector); 876 } 877 ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID; 878 if (ec_valid) { 879 ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID; 880 ctrl->eventinj |= (uint64_t)error << 32; 881 VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x", 882 intrtype_to_str(intr_type), vector, error); 883 } else { 884 VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d", 885 intrtype_to_str(intr_type), vector); 886 } 887} 888 889static void 890svm_update_virqinfo(struct svm_softc *sc, int vcpu) 891{ 892 struct vm *vm; 893 struct vlapic *vlapic; 894 struct vmcb_ctrl *ctrl; 895 int pending; 896 897 vm = sc->vm; 898 vlapic = vm_lapic(vm, vcpu); 899 ctrl = svm_get_vmcb_ctrl(sc, vcpu); 900 901 /* Update %cr8 in the emulated vlapic */ 902 vlapic_set_cr8(vlapic, ctrl->v_tpr); 903 904 /* 905 * If V_IRQ indicates that the interrupt injection attempted on then 906 * last VMRUN was successful then update the vlapic accordingly. 907 */ 908 if (ctrl->v_intr_vector != 0) { 909 pending = ctrl->v_irq; 910 KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid " 911 "v_intr_vector %d", __func__, ctrl->v_intr_vector)); 912 KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__)); 913 VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector, 914 pending ? "pending" : "accepted"); 915 if (!pending) 916 vlapic_intr_accepted(vlapic, ctrl->v_intr_vector); 917 } 918} 919 920static void 921svm_save_intinfo(struct svm_softc *svm_sc, int vcpu) 922{ 923 struct vmcb_ctrl *ctrl; 924 uint64_t intinfo; 925 926 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); 927 intinfo = ctrl->exitintinfo; 928 if (!VMCB_EXITINTINFO_VALID(intinfo)) 929 return; 930 931 /* 932 * From APMv2, Section "Intercepts during IDT interrupt delivery" 933 * 934 * If a #VMEXIT happened during event delivery then record the event 935 * that was being delivered. 936 */ 937 VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n", 938 intinfo, VMCB_EXITINTINFO_VECTOR(intinfo)); 939 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1); 940 vm_exit_intinfo(svm_sc->vm, vcpu, intinfo); 941} 942 943static __inline int 944vintr_intercept_enabled(struct svm_softc *sc, int vcpu) 945{ 946 947 return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, 948 VMCB_INTCPT_VINTR)); 949} 950 951static __inline void 952enable_intr_window_exiting(struct svm_softc *sc, int vcpu) 953{ 954 struct vmcb_ctrl *ctrl; 955 956 ctrl = svm_get_vmcb_ctrl(sc, vcpu); 957 958 if (ctrl->v_irq && ctrl->v_intr_vector == 0) { 959 KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__)); 960 KASSERT(vintr_intercept_enabled(sc, vcpu), 961 ("%s: vintr intercept should be enabled", __func__)); 962 return; 963 } 964 965 VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting"); 966 ctrl->v_irq = 1; 967 ctrl->v_ign_tpr = 1; 968 ctrl->v_intr_vector = 0; 969 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); 970 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); 971} 972 973static __inline void 974disable_intr_window_exiting(struct svm_softc *sc, int vcpu) 975{ 976 struct vmcb_ctrl *ctrl; 977 978 ctrl = svm_get_vmcb_ctrl(sc, vcpu); 979 980 if (!ctrl->v_irq && ctrl->v_intr_vector == 0) { 981 KASSERT(!vintr_intercept_enabled(sc, vcpu), 982 ("%s: vintr intercept should be disabled", __func__)); 983 return; 984 } 985 986#ifdef KTR 987 if (ctrl->v_intr_vector == 0) 988 VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting"); 989 else 990 VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection"); 991#endif 992 ctrl->v_irq = 0; 993 ctrl->v_intr_vector = 0; 994 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); 995 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR); 996} 997 998static int 999svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val) 1000{ 1001 struct vmcb_ctrl *ctrl; 1002 int oldval, newval; 1003 1004 ctrl = svm_get_vmcb_ctrl(sc, vcpu); 1005 oldval = ctrl->intr_shadow; 1006 newval = val ? 1 : 0; 1007 if (newval != oldval) { 1008 ctrl->intr_shadow = newval; 1009 VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval); 1010 } 1011 return (0); 1012} 1013 1014static int 1015svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val) 1016{ 1017 struct vmcb_ctrl *ctrl; 1018 1019 ctrl = svm_get_vmcb_ctrl(sc, vcpu); 1020 *val = ctrl->intr_shadow; 1021 return (0); 1022} 1023 1024/* 1025 * Once an NMI is injected it blocks delivery of further NMIs until the handler 1026 * executes an IRET. The IRET intercept is enabled when an NMI is injected to 1027 * to track when the vcpu is done handling the NMI. 1028 */ 1029static int 1030nmi_blocked(struct svm_softc *sc, int vcpu) 1031{ 1032 int blocked; 1033 1034 blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, 1035 VMCB_INTCPT_IRET); 1036 return (blocked); 1037} 1038 1039static void 1040enable_nmi_blocking(struct svm_softc *sc, int vcpu) 1041{ 1042 1043 KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked")); 1044 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled"); 1045 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); 1046} 1047 1048static void 1049clear_nmi_blocking(struct svm_softc *sc, int vcpu) 1050{ 1051 int error; 1052 1053 KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked")); 1054 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared"); 1055 /* 1056 * When the IRET intercept is cleared the vcpu will attempt to execute 1057 * the "iret" when it runs next. However, it is possible to inject 1058 * another NMI into the vcpu before the "iret" has actually executed. 1059 * 1060 * For e.g. if the "iret" encounters a #NPF when accessing the stack 1061 * it will trap back into the hypervisor. If an NMI is pending for 1062 * the vcpu it will be injected into the guest. 1063 * 1064 * XXX this needs to be fixed 1065 */ 1066 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET); 1067 1068 /* 1069 * Set 'intr_shadow' to prevent an NMI from being injected on the 1070 * immediate VMRUN. 1071 */ 1072 error = svm_modify_intr_shadow(sc, vcpu, 1); 1073 KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error)); 1074} 1075 1076static int 1077emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val, 1078 bool *retu) 1079{ 1080 int error; 1081 1082 if (lapic_msr(num)) 1083 error = lapic_wrmsr(sc->vm, vcpu, num, val, retu); 1084 else if (num == MSR_EFER) 1085 error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, val); 1086 else 1087 error = svm_wrmsr(sc, vcpu, num, val, retu); 1088 1089 return (error); 1090} 1091 1092static int 1093emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu) 1094{ 1095 struct vmcb_state *state; 1096 struct svm_regctx *ctx; 1097 uint64_t result; 1098 int error; 1099 1100 if (lapic_msr(num)) 1101 error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu); 1102 else 1103 error = svm_rdmsr(sc, vcpu, num, &result, retu); 1104 1105 if (error == 0) { 1106 state = svm_get_vmcb_state(sc, vcpu); 1107 ctx = svm_get_guest_regctx(sc, vcpu); 1108 state->rax = result & 0xffffffff; 1109 ctx->sctx_rdx = result >> 32; 1110 } 1111 1112 return (error); 1113} 1114 1115#ifdef KTR 1116static const char * 1117exit_reason_to_str(uint64_t reason) 1118{ 1119 static char reasonbuf[32]; 1120 1121 switch (reason) { 1122 case VMCB_EXIT_INVALID: 1123 return ("invalvmcb"); 1124 case VMCB_EXIT_SHUTDOWN: 1125 return ("shutdown"); 1126 case VMCB_EXIT_NPF: 1127 return ("nptfault"); 1128 case VMCB_EXIT_PAUSE: 1129 return ("pause"); 1130 case VMCB_EXIT_HLT: 1131 return ("hlt"); 1132 case VMCB_EXIT_CPUID: 1133 return ("cpuid"); 1134 case VMCB_EXIT_IO: 1135 return ("inout"); 1136 case VMCB_EXIT_MC: 1137 return ("mchk"); 1138 case VMCB_EXIT_INTR: 1139 return ("extintr"); 1140 case VMCB_EXIT_NMI: 1141 return ("nmi"); 1142 case VMCB_EXIT_VINTR: 1143 return ("vintr"); 1144 case VMCB_EXIT_MSR: 1145 return ("msr"); 1146 case VMCB_EXIT_IRET: 1147 return ("iret"); 1148 case VMCB_EXIT_MONITOR: 1149 return ("monitor"); 1150 case VMCB_EXIT_MWAIT: 1151 return ("mwait"); 1152 default: 1153 snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason); 1154 return (reasonbuf); 1155 } 1156} 1157#endif /* KTR */ 1158 1159/* 1160 * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs 1161 * that are due to instruction intercepts as well as MSR and IOIO intercepts 1162 * and exceptions caused by INT3, INTO and BOUND instructions. 1163 * 1164 * Return 1 if the nRIP is valid and 0 otherwise. 1165 */ 1166static int 1167nrip_valid(uint64_t exitcode) 1168{ 1169 switch (exitcode) { 1170 case 0x00 ... 0x0F: /* read of CR0 through CR15 */ 1171 case 0x10 ... 0x1F: /* write of CR0 through CR15 */ 1172 case 0x20 ... 0x2F: /* read of DR0 through DR15 */ 1173 case 0x30 ... 0x3F: /* write of DR0 through DR15 */ 1174 case 0x43: /* INT3 */ 1175 case 0x44: /* INTO */ 1176 case 0x45: /* BOUND */ 1177 case 0x65 ... 0x7C: /* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */ 1178 case 0x80 ... 0x8D: /* VMEXIT_VMRUN ... VMEXIT_XSETBV */ 1179 return (1); 1180 default: 1181 return (0); 1182 } 1183} 1184 1185/* 1186 * Collateral for a generic SVM VM-exit. 1187 */ 1188static void 1189vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2) 1190{ 1191 1192 vme->exitcode = VM_EXITCODE_SVM; 1193 vme->u.svm.exitcode = code; 1194 vme->u.svm.exitinfo1 = info1; 1195 vme->u.svm.exitinfo2 = info2; 1196} 1197 1198static int 1199svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit) 1200{ 1201 struct vmcb *vmcb; 1202 struct vmcb_state *state; 1203 struct vmcb_ctrl *ctrl; 1204 struct svm_regctx *ctx; 1205 uint64_t code, info1, info2, val; 1206 uint32_t eax, ecx, edx; 1207 int error, errcode_valid, handled, idtvec, reflect; 1208 bool retu; 1209 1210 ctx = svm_get_guest_regctx(svm_sc, vcpu); 1211 vmcb = svm_get_vmcb(svm_sc, vcpu); 1212 state = &vmcb->state; 1213 ctrl = &vmcb->ctrl; 1214 1215 handled = 0; 1216 code = ctrl->exitcode; 1217 info1 = ctrl->exitinfo1; 1218 info2 = ctrl->exitinfo2; 1219 1220 vmexit->exitcode = VM_EXITCODE_BOGUS; 1221 vmexit->rip = state->rip; 1222 vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0; 1223 1224 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1); 1225 1226 /* 1227 * #VMEXIT(INVALID) needs to be handled early because the VMCB is 1228 * in an inconsistent state and can trigger assertions that would 1229 * never happen otherwise. 1230 */ 1231 if (code == VMCB_EXIT_INVALID) { 1232 vm_exit_svm(vmexit, code, info1, info2); 1233 return (0); 1234 } 1235 1236 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event " 1237 "injection valid bit is set %#lx", __func__, ctrl->eventinj)); 1238 1239 KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15, 1240 ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)", 1241 vmexit->inst_length, code, info1, info2)); 1242 1243 svm_update_virqinfo(svm_sc, vcpu); 1244 svm_save_intinfo(svm_sc, vcpu); 1245 1246 switch (code) { 1247 case VMCB_EXIT_IRET: 1248 /* 1249 * Restart execution at "iret" but with the intercept cleared. 1250 */ 1251 vmexit->inst_length = 0; 1252 clear_nmi_blocking(svm_sc, vcpu); 1253 handled = 1; 1254 break; 1255 case VMCB_EXIT_VINTR: /* interrupt window exiting */ 1256 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1); 1257 handled = 1; 1258 break; 1259 case VMCB_EXIT_INTR: /* external interrupt */ 1260 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1); 1261 handled = 1; 1262 break; 1263 case VMCB_EXIT_NMI: /* external NMI */ 1264 handled = 1; 1265 break; 1266 case 0x40 ... 0x5F: 1267 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1); 1268 reflect = 1; 1269 idtvec = code - 0x40; 1270 switch (idtvec) { 1271 case IDT_MC: 1272 /* 1273 * Call the machine check handler by hand. Also don't 1274 * reflect the machine check back into the guest. 1275 */ 1276 reflect = 0; 1277 VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler"); 1278 __asm __volatile("int $18"); 1279 break; 1280 case IDT_PF: 1281 error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2, 1282 info2); 1283 KASSERT(error == 0, ("%s: error %d updating cr2", 1284 __func__, error)); 1285 /* fallthru */ 1286 case IDT_NP: 1287 case IDT_SS: 1288 case IDT_GP: 1289 case IDT_AC: 1290 case IDT_TS: 1291 errcode_valid = 1; 1292 break; 1293 1294 case IDT_DF: 1295 errcode_valid = 1; 1296 info1 = 0; 1297 break; 1298 1299 case IDT_BP: 1300 case IDT_OF: 1301 case IDT_BR: 1302 /* 1303 * The 'nrip' field is populated for INT3, INTO and 1304 * BOUND exceptions and this also implies that 1305 * 'inst_length' is non-zero. 1306 * 1307 * Reset 'inst_length' to zero so the guest %rip at 1308 * event injection is identical to what it was when 1309 * the exception originally happened. 1310 */ 1311 VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d " 1312 "to zero before injecting exception %d", 1313 vmexit->inst_length, idtvec); 1314 vmexit->inst_length = 0; 1315 /* fallthru */ 1316 default: 1317 errcode_valid = 0; 1318 info1 = 0; 1319 break; 1320 } 1321 KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) " 1322 "when reflecting exception %d into guest", 1323 vmexit->inst_length, idtvec)); 1324 1325 if (reflect) { 1326 /* Reflect the exception back into the guest */ 1327 VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception " 1328 "%d/%#x into the guest", idtvec, (int)info1); 1329 error = vm_inject_exception(svm_sc->vm, vcpu, idtvec, 1330 errcode_valid, info1, 0); 1331 KASSERT(error == 0, ("%s: vm_inject_exception error %d", 1332 __func__, error)); 1333 } 1334 handled = 1; 1335 break; 1336 case VMCB_EXIT_MSR: /* MSR access. */ 1337 eax = state->rax; 1338 ecx = ctx->sctx_rcx; 1339 edx = ctx->sctx_rdx; 1340 retu = false; 1341 1342 if (info1) { 1343 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1); 1344 val = (uint64_t)edx << 32 | eax; 1345 VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx", 1346 ecx, val); 1347 if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) { 1348 vmexit->exitcode = VM_EXITCODE_WRMSR; 1349 vmexit->u.msr.code = ecx; 1350 vmexit->u.msr.wval = val; 1351 } else if (!retu) { 1352 handled = 1; 1353 } else { 1354 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, 1355 ("emulate_wrmsr retu with bogus exitcode")); 1356 } 1357 } else { 1358 VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx); 1359 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1); 1360 if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) { 1361 vmexit->exitcode = VM_EXITCODE_RDMSR; 1362 vmexit->u.msr.code = ecx; 1363 } else if (!retu) { 1364 handled = 1; 1365 } else { 1366 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS, 1367 ("emulate_rdmsr retu with bogus exitcode")); 1368 } 1369 } 1370 break; 1371 case VMCB_EXIT_IO: 1372 handled = svm_handle_io(svm_sc, vcpu, vmexit); 1373 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1); 1374 break; 1375 case VMCB_EXIT_CPUID: 1376 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1); 1377 handled = x86_emulate_cpuid(svm_sc->vm, vcpu, 1378 (uint32_t *)&state->rax, 1379 (uint32_t *)&ctx->sctx_rbx, 1380 (uint32_t *)&ctx->sctx_rcx, 1381 (uint32_t *)&ctx->sctx_rdx); 1382 break; 1383 case VMCB_EXIT_HLT: 1384 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1); 1385 vmexit->exitcode = VM_EXITCODE_HLT; 1386 vmexit->u.hlt.rflags = state->rflags; 1387 break; 1388 case VMCB_EXIT_PAUSE: 1389 vmexit->exitcode = VM_EXITCODE_PAUSE; 1390 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1); 1391 break; 1392 case VMCB_EXIT_NPF: 1393 /* EXITINFO2 contains the faulting guest physical address */ 1394 if (info1 & VMCB_NPF_INFO1_RSV) { 1395 VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with " 1396 "reserved bits set: info1(%#lx) info2(%#lx)", 1397 info1, info2); 1398 } else if (vm_mem_allocated(svm_sc->vm, info2)) { 1399 vmexit->exitcode = VM_EXITCODE_PAGING; 1400 vmexit->u.paging.gpa = info2; 1401 vmexit->u.paging.fault_type = npf_fault_type(info1); 1402 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1); 1403 VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault " 1404 "on gpa %#lx/%#lx at rip %#lx", 1405 info2, info1, state->rip); 1406 } else if (svm_npf_emul_fault(info1)) { 1407 svm_handle_inst_emul(vmcb, info2, vmexit); 1408 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1); 1409 VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault " 1410 "for gpa %#lx/%#lx at rip %#lx", 1411 info2, info1, state->rip); 1412 } 1413 break; 1414 case VMCB_EXIT_MONITOR: 1415 vmexit->exitcode = VM_EXITCODE_MONITOR; 1416 break; 1417 case VMCB_EXIT_MWAIT: 1418 vmexit->exitcode = VM_EXITCODE_MWAIT; 1419 break; 1420 default: 1421 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1); 1422 break; 1423 } 1424 1425 VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d", 1426 handled ? "handled" : "unhandled", exit_reason_to_str(code), 1427 vmexit->rip, vmexit->inst_length); 1428 1429 if (handled) { 1430 vmexit->rip += vmexit->inst_length; 1431 vmexit->inst_length = 0; 1432 state->rip = vmexit->rip; 1433 } else { 1434 if (vmexit->exitcode == VM_EXITCODE_BOGUS) { 1435 /* 1436 * If this VM exit was not claimed by anybody then 1437 * treat it as a generic SVM exit. 1438 */ 1439 vm_exit_svm(vmexit, code, info1, info2); 1440 } else { 1441 /* 1442 * The exitcode and collateral have been populated. 1443 * The VM exit will be processed further in userland. 1444 */ 1445 } 1446 } 1447 return (handled); 1448} 1449 1450static void 1451svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu) 1452{ 1453 uint64_t intinfo; 1454 1455 if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo)) 1456 return; 1457 1458 KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not " 1459 "valid: %#lx", __func__, intinfo)); 1460 1461 svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo), 1462 VMCB_EXITINTINFO_VECTOR(intinfo), 1463 VMCB_EXITINTINFO_EC(intinfo), 1464 VMCB_EXITINTINFO_EC_VALID(intinfo)); 1465 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1); 1466 VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo); 1467} 1468 1469/* 1470 * Inject event to virtual cpu. 1471 */ 1472static void 1473svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic) 1474{ 1475 struct vmcb_ctrl *ctrl; 1476 struct vmcb_state *state; 1477 struct svm_vcpu *vcpustate; 1478 uint8_t v_tpr; 1479 int vector, need_intr_window, pending_apic_vector; 1480 1481 state = svm_get_vmcb_state(sc, vcpu); 1482 ctrl = svm_get_vmcb_ctrl(sc, vcpu); 1483 vcpustate = svm_get_vcpu(sc, vcpu); 1484 1485 need_intr_window = 0; 1486 pending_apic_vector = 0; 1487 1488 if (vcpustate->nextrip != state->rip) { 1489 ctrl->intr_shadow = 0; 1490 VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking " 1491 "cleared due to rip change: %#lx/%#lx", 1492 vcpustate->nextrip, state->rip); 1493 } 1494 1495 /* 1496 * Inject pending events or exceptions for this vcpu. 1497 * 1498 * An event might be pending because the previous #VMEXIT happened 1499 * during event delivery (i.e. ctrl->exitintinfo). 1500 * 1501 * An event might also be pending because an exception was injected 1502 * by the hypervisor (e.g. #PF during instruction emulation). 1503 */ 1504 svm_inj_intinfo(sc, vcpu); 1505 1506 /* NMI event has priority over interrupts. */ 1507 if (vm_nmi_pending(sc->vm, vcpu)) { 1508 if (nmi_blocked(sc, vcpu)) { 1509 /* 1510 * Can't inject another NMI if the guest has not 1511 * yet executed an "iret" after the last NMI. 1512 */ 1513 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due " 1514 "to NMI-blocking"); 1515 } else if (ctrl->intr_shadow) { 1516 /* 1517 * Can't inject an NMI if the vcpu is in an intr_shadow. 1518 */ 1519 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to " 1520 "interrupt shadow"); 1521 need_intr_window = 1; 1522 goto done; 1523 } else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) { 1524 /* 1525 * If there is already an exception/interrupt pending 1526 * then defer the NMI until after that. 1527 */ 1528 VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to " 1529 "eventinj %#lx", ctrl->eventinj); 1530 1531 /* 1532 * Use self-IPI to trigger a VM-exit as soon as 1533 * possible after the event injection is completed. 1534 * 1535 * This works only if the external interrupt exiting 1536 * is at a lower priority than the event injection. 1537 * 1538 * Although not explicitly specified in APMv2 the 1539 * relative priorities were verified empirically. 1540 */ 1541 ipi_cpu(curcpu, IPI_AST); /* XXX vmm_ipinum? */ 1542 } else { 1543 vm_nmi_clear(sc->vm, vcpu); 1544 1545 /* Inject NMI, vector number is not used */ 1546 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI, 1547 IDT_NMI, 0, false); 1548 1549 /* virtual NMI blocking is now in effect */ 1550 enable_nmi_blocking(sc, vcpu); 1551 1552 VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI"); 1553 } 1554 } 1555 1556 if (!vm_extint_pending(sc->vm, vcpu)) { 1557 /* 1558 * APIC interrupts are delivered using the V_IRQ offload. 1559 * 1560 * The primary benefit is that the hypervisor doesn't need to 1561 * deal with the various conditions that inhibit interrupts. 1562 * It also means that TPR changes via CR8 will be handled 1563 * without any hypervisor involvement. 1564 * 1565 * Note that the APIC vector must remain pending in the vIRR 1566 * until it is confirmed that it was delivered to the guest. 1567 * This can be confirmed based on the value of V_IRQ at the 1568 * next #VMEXIT (1 = pending, 0 = delivered). 1569 * 1570 * Also note that it is possible that another higher priority 1571 * vector can become pending before this vector is delivered 1572 * to the guest. This is alright because vcpu_notify_event() 1573 * will send an IPI and force the vcpu to trap back into the 1574 * hypervisor. The higher priority vector will be injected on 1575 * the next VMRUN. 1576 */ 1577 if (vlapic_pending_intr(vlapic, &vector)) { 1578 KASSERT(vector >= 16 && vector <= 255, 1579 ("invalid vector %d from local APIC", vector)); 1580 pending_apic_vector = vector; 1581 } 1582 goto done; 1583 } 1584 1585 /* Ask the legacy pic for a vector to inject */ 1586 vatpic_pending_intr(sc->vm, &vector); 1587 KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR", 1588 vector)); 1589 1590 /* 1591 * If the guest has disabled interrupts or is in an interrupt shadow 1592 * then we cannot inject the pending interrupt. 1593 */ 1594 if ((state->rflags & PSL_I) == 0) { 1595 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to " 1596 "rflags %#lx", vector, state->rflags); 1597 need_intr_window = 1; 1598 goto done; 1599 } 1600 1601 if (ctrl->intr_shadow) { 1602 VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to " 1603 "interrupt shadow", vector); 1604 need_intr_window = 1; 1605 goto done; 1606 } 1607 1608 if (ctrl->eventinj & VMCB_EVENTINJ_VALID) { 1609 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to " 1610 "eventinj %#lx", vector, ctrl->eventinj); 1611 need_intr_window = 1; 1612 goto done; 1613 } 1614 1615 /* 1616 * Legacy PIC interrupts are delivered via the event injection 1617 * mechanism. 1618 */ 1619 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false); 1620 1621 vm_extint_clear(sc->vm, vcpu); 1622 vatpic_intr_accepted(sc->vm, vector); 1623 1624 /* 1625 * Force a VM-exit as soon as the vcpu is ready to accept another 1626 * interrupt. This is done because the PIC might have another vector 1627 * that it wants to inject. Also, if the APIC has a pending interrupt 1628 * that was preempted by the ExtInt then it allows us to inject the 1629 * APIC vector as soon as possible. 1630 */ 1631 need_intr_window = 1; 1632done: 1633 /* 1634 * The guest can modify the TPR by writing to %CR8. In guest mode 1635 * the processor reflects this write to V_TPR without hypervisor 1636 * intervention. 1637 * 1638 * The guest can also modify the TPR by writing to it via the memory 1639 * mapped APIC page. In this case, the write will be emulated by the 1640 * hypervisor. For this reason V_TPR must be updated before every 1641 * VMRUN. 1642 */ 1643 v_tpr = vlapic_get_cr8(vlapic); 1644 KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr)); 1645 if (ctrl->v_tpr != v_tpr) { 1646 VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x", 1647 ctrl->v_tpr, v_tpr); 1648 ctrl->v_tpr = v_tpr; 1649 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); 1650 } 1651 1652 if (pending_apic_vector) { 1653 /* 1654 * If an APIC vector is being injected then interrupt window 1655 * exiting is not possible on this VMRUN. 1656 */ 1657 KASSERT(!need_intr_window, ("intr_window exiting impossible")); 1658 VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ", 1659 pending_apic_vector); 1660 1661 ctrl->v_irq = 1; 1662 ctrl->v_ign_tpr = 0; 1663 ctrl->v_intr_vector = pending_apic_vector; 1664 ctrl->v_intr_prio = pending_apic_vector >> 4; 1665 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR); 1666 } else if (need_intr_window) { 1667 /* 1668 * We use V_IRQ in conjunction with the VINTR intercept to 1669 * trap into the hypervisor as soon as a virtual interrupt 1670 * can be delivered. 1671 * 1672 * Since injected events are not subject to intercept checks 1673 * we need to ensure that the V_IRQ is not actually going to 1674 * be delivered on VM entry. The KASSERT below enforces this. 1675 */ 1676 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 || 1677 (state->rflags & PSL_I) == 0 || ctrl->intr_shadow, 1678 ("Bogus intr_window_exiting: eventinj (%#lx), " 1679 "intr_shadow (%u), rflags (%#lx)", 1680 ctrl->eventinj, ctrl->intr_shadow, state->rflags)); 1681 enable_intr_window_exiting(sc, vcpu); 1682 } else { 1683 disable_intr_window_exiting(sc, vcpu); 1684 } 1685} 1686 1687static __inline void 1688restore_host_tss(void) 1689{ 1690 struct system_segment_descriptor *tss_sd; 1691 1692 /* 1693 * The TSS descriptor was in use prior to launching the guest so it 1694 * has been marked busy. 1695 * 1696 * 'ltr' requires the descriptor to be marked available so change the 1697 * type to "64-bit available TSS". 1698 */ 1699 tss_sd = PCPU_GET(tss); 1700 tss_sd->sd_type = SDT_SYSTSS; 1701 ltr(GSEL(GPROC0_SEL, SEL_KPL)); 1702} 1703 1704static void 1705check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu) 1706{ 1707 struct svm_vcpu *vcpustate; 1708 struct vmcb_ctrl *ctrl; 1709 long eptgen; 1710 bool alloc_asid; 1711 1712 KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not " 1713 "active on cpu %u", __func__, thiscpu)); 1714 1715 vcpustate = svm_get_vcpu(sc, vcpuid); 1716 ctrl = svm_get_vmcb_ctrl(sc, vcpuid); 1717 1718 /* 1719 * The TLB entries associated with the vcpu's ASID are not valid 1720 * if either of the following conditions is true: 1721 * 1722 * 1. The vcpu's ASID generation is different than the host cpu's 1723 * ASID generation. This happens when the vcpu migrates to a new 1724 * host cpu. It can also happen when the number of vcpus executing 1725 * on a host cpu is greater than the number of ASIDs available. 1726 * 1727 * 2. The pmap generation number is different than the value cached in 1728 * the 'vcpustate'. This happens when the host invalidates pages 1729 * belonging to the guest. 1730 * 1731 * asidgen eptgen Action 1732 * mismatch mismatch 1733 * 0 0 (a) 1734 * 0 1 (b1) or (b2) 1735 * 1 0 (c) 1736 * 1 1 (d) 1737 * 1738 * (a) There is no mismatch in eptgen or ASID generation and therefore 1739 * no further action is needed. 1740 * 1741 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is 1742 * retained and the TLB entries associated with this ASID 1743 * are flushed by VMRUN. 1744 * 1745 * (b2) If the cpu does not support FlushByAsid then a new ASID is 1746 * allocated. 1747 * 1748 * (c) A new ASID is allocated. 1749 * 1750 * (d) A new ASID is allocated. 1751 */ 1752 1753 alloc_asid = false; 1754 eptgen = pmap->pm_eptgen; 1755 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING; 1756 1757 if (vcpustate->asid.gen != asid[thiscpu].gen) { 1758 alloc_asid = true; /* (c) and (d) */ 1759 } else if (vcpustate->eptgen != eptgen) { 1760 if (flush_by_asid()) 1761 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; /* (b1) */ 1762 else 1763 alloc_asid = true; /* (b2) */ 1764 } else { 1765 /* 1766 * This is the common case (a). 1767 */ 1768 KASSERT(!alloc_asid, ("ASID allocation not necessary")); 1769 KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING, 1770 ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl)); 1771 } 1772 1773 if (alloc_asid) { 1774 if (++asid[thiscpu].num >= nasid) { 1775 asid[thiscpu].num = 1; 1776 if (++asid[thiscpu].gen == 0) 1777 asid[thiscpu].gen = 1; 1778 /* 1779 * If this cpu does not support "flush-by-asid" 1780 * then flush the entire TLB on a generation 1781 * bump. Subsequent ASID allocation in this 1782 * generation can be done without a TLB flush. 1783 */ 1784 if (!flush_by_asid()) 1785 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL; 1786 } 1787 vcpustate->asid.gen = asid[thiscpu].gen; 1788 vcpustate->asid.num = asid[thiscpu].num; 1789 1790 ctrl->asid = vcpustate->asid.num; 1791 svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID); 1792 /* 1793 * If this cpu supports "flush-by-asid" then the TLB 1794 * was not flushed after the generation bump. The TLB 1795 * is flushed selectively after every new ASID allocation. 1796 */ 1797 if (flush_by_asid()) 1798 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; 1799 } 1800 vcpustate->eptgen = eptgen; 1801 1802 KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero")); 1803 KASSERT(ctrl->asid == vcpustate->asid.num, 1804 ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num)); 1805} 1806 1807static __inline void 1808disable_gintr(void) 1809{ 1810 1811 __asm __volatile("clgi"); 1812} 1813 1814static __inline void 1815enable_gintr(void) 1816{ 1817 1818 __asm __volatile("stgi"); 1819} 1820 1821/* 1822 * Start vcpu with specified RIP. 1823 */ 1824static int 1825svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap, 1826 void *rend_cookie, void *suspended_cookie) 1827{ 1828 struct svm_regctx *gctx; 1829 struct svm_softc *svm_sc; 1830 struct svm_vcpu *vcpustate; 1831 struct vmcb_state *state; 1832 struct vmcb_ctrl *ctrl; 1833 struct vm_exit *vmexit; 1834 struct vlapic *vlapic; 1835 struct vm *vm; 1836 uint64_t vmcb_pa; 1837 u_int thiscpu; 1838 int handled; 1839 1840 svm_sc = arg; 1841 vm = svm_sc->vm; 1842 1843 vcpustate = svm_get_vcpu(svm_sc, vcpu); 1844 state = svm_get_vmcb_state(svm_sc, vcpu); 1845 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu); 1846 vmexit = vm_exitinfo(vm, vcpu); 1847 vlapic = vm_lapic(vm, vcpu); 1848 1849 /* 1850 * Stash 'curcpu' on the stack as 'thiscpu'. 1851 * 1852 * The per-cpu data area is not accessible until MSR_GSBASE is restored 1853 * after the #VMEXIT. Since VMRUN is executed inside a critical section 1854 * 'curcpu' and 'thiscpu' are guaranteed to identical. 1855 */ 1856 thiscpu = curcpu; 1857 1858 gctx = svm_get_guest_regctx(svm_sc, vcpu); 1859 vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa; 1860 1861 if (vcpustate->lastcpu != thiscpu) { 1862 /* 1863 * Force new ASID allocation by invalidating the generation. 1864 */ 1865 vcpustate->asid.gen = 0; 1866 1867 /* 1868 * Invalidate the VMCB state cache by marking all fields dirty. 1869 */ 1870 svm_set_dirty(svm_sc, vcpu, 0xffffffff); 1871 1872 /* 1873 * XXX 1874 * Setting 'vcpustate->lastcpu' here is bit premature because 1875 * we may return from this function without actually executing 1876 * the VMRUN instruction. This could happen if a rendezvous 1877 * or an AST is pending on the first time through the loop. 1878 * 1879 * This works for now but any new side-effects of vcpu 1880 * migration should take this case into account. 1881 */ 1882 vcpustate->lastcpu = thiscpu; 1883 vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1); 1884 } 1885 1886 svm_msr_guest_enter(svm_sc, vcpu); 1887 1888 /* Update Guest RIP */ 1889 state->rip = rip; 1890 1891 do { 1892 /* 1893 * Disable global interrupts to guarantee atomicity during 1894 * loading of guest state. This includes not only the state 1895 * loaded by the "vmrun" instruction but also software state 1896 * maintained by the hypervisor: suspended and rendezvous 1897 * state, NPT generation number, vlapic interrupts etc. 1898 */ 1899 disable_gintr(); 1900 1901 if (vcpu_suspended(suspended_cookie)) { 1902 enable_gintr(); 1903 vm_exit_suspended(vm, vcpu, state->rip); 1904 break; 1905 } 1906 1907 if (vcpu_rendezvous_pending(rend_cookie)) { 1908 enable_gintr(); 1909 vm_exit_rendezvous(vm, vcpu, state->rip); 1910 break; 1911 } 1912 1913 /* We are asked to give the cpu by scheduler. */ 1914 if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) { 1915 enable_gintr(); 1916 vm_exit_astpending(vm, vcpu, state->rip); 1917 break; 1918 } 1919 1920 svm_inj_interrupts(svm_sc, vcpu, vlapic); 1921 1922 /* Activate the nested pmap on 'thiscpu' */ 1923 CPU_SET_ATOMIC_ACQ(thiscpu, &pmap->pm_active); 1924 1925 /* 1926 * Check the pmap generation and the ASID generation to 1927 * ensure that the vcpu does not use stale TLB mappings. 1928 */ 1929 check_asid(svm_sc, vcpu, pmap, thiscpu); 1930 1931 ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty; 1932 vcpustate->dirty = 0; 1933 VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean); 1934 1935 /* Launch Virtual Machine. */ 1936 VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip); 1937 svm_launch(vmcb_pa, gctx); 1938 1939 CPU_CLR_ATOMIC(thiscpu, &pmap->pm_active); 1940 1941 /* 1942 * Restore MSR_GSBASE to point to the pcpu data area. 1943 * 1944 * Note that accesses done via PCPU_GET/PCPU_SET will work 1945 * only after MSR_GSBASE is restored. 1946 * 1947 * Also note that we don't bother restoring MSR_KGSBASE 1948 * since it is not used in the kernel and will be restored 1949 * when the VMRUN ioctl returns to userspace. 1950 */ 1951 wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[thiscpu]); 1952 KASSERT(curcpu == thiscpu, ("thiscpu/curcpu (%u/%u) mismatch", 1953 thiscpu, curcpu)); 1954 1955 /* 1956 * The host GDTR and IDTR is saved by VMRUN and restored 1957 * automatically on #VMEXIT. However, the host TSS needs 1958 * to be restored explicitly. 1959 */ 1960 restore_host_tss(); 1961 1962 /* #VMEXIT disables interrupts so re-enable them here. */ 1963 enable_gintr(); 1964 1965 /* Update 'nextrip' */ 1966 vcpustate->nextrip = state->rip; 1967 1968 /* Handle #VMEXIT and if required return to user space. */ 1969 handled = svm_vmexit(svm_sc, vcpu, vmexit); 1970 } while (handled); 1971 1972 svm_msr_guest_exit(svm_sc, vcpu); 1973 1974 return (0); 1975} 1976 1977static void 1978svm_vmcleanup(void *arg) 1979{ 1980 struct svm_softc *sc = arg; 1981 1982 free(sc, M_SVM); 1983} 1984 1985static register_t * 1986swctx_regptr(struct svm_regctx *regctx, int reg) 1987{ 1988 1989 switch (reg) { 1990 case VM_REG_GUEST_RBX: 1991 return (®ctx->sctx_rbx); 1992 case VM_REG_GUEST_RCX: 1993 return (®ctx->sctx_rcx); 1994 case VM_REG_GUEST_RDX: 1995 return (®ctx->sctx_rdx); 1996 case VM_REG_GUEST_RDI: 1997 return (®ctx->sctx_rdi); 1998 case VM_REG_GUEST_RSI: 1999 return (®ctx->sctx_rsi); 2000 case VM_REG_GUEST_RBP: 2001 return (®ctx->sctx_rbp); 2002 case VM_REG_GUEST_R8: 2003 return (®ctx->sctx_r8); 2004 case VM_REG_GUEST_R9: 2005 return (®ctx->sctx_r9); 2006 case VM_REG_GUEST_R10: 2007 return (®ctx->sctx_r10); 2008 case VM_REG_GUEST_R11: 2009 return (®ctx->sctx_r11); 2010 case VM_REG_GUEST_R12: 2011 return (®ctx->sctx_r12); 2012 case VM_REG_GUEST_R13: 2013 return (®ctx->sctx_r13); 2014 case VM_REG_GUEST_R14: 2015 return (®ctx->sctx_r14); 2016 case VM_REG_GUEST_R15: 2017 return (®ctx->sctx_r15); 2018 default: 2019 return (NULL); 2020 } 2021} 2022 2023static int 2024svm_getreg(void *arg, int vcpu, int ident, uint64_t *val) 2025{ 2026 struct svm_softc *svm_sc; 2027 register_t *reg; 2028 2029 svm_sc = arg; 2030 2031 if (ident == VM_REG_GUEST_INTR_SHADOW) { 2032 return (svm_get_intr_shadow(svm_sc, vcpu, val)); 2033 } 2034 2035 if (vmcb_read(svm_sc, vcpu, ident, val) == 0) { 2036 return (0); 2037 } 2038 2039 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident); 2040 2041 if (reg != NULL) { 2042 *val = *reg; 2043 return (0); 2044 } 2045 2046 VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident); 2047 return (EINVAL); 2048} 2049 2050static int 2051svm_setreg(void *arg, int vcpu, int ident, uint64_t val) 2052{ 2053 struct svm_softc *svm_sc; 2054 register_t *reg; 2055 2056 svm_sc = arg; 2057 2058 if (ident == VM_REG_GUEST_INTR_SHADOW) { 2059 return (svm_modify_intr_shadow(svm_sc, vcpu, val)); 2060 } 2061 2062 if (vmcb_write(svm_sc, vcpu, ident, val) == 0) { 2063 return (0); 2064 } 2065 2066 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident); 2067 2068 if (reg != NULL) { 2069 *reg = val; 2070 return (0); 2071 } 2072 2073 /* 2074 * XXX deal with CR3 and invalidate TLB entries tagged with the 2075 * vcpu's ASID. This needs to be treated differently depending on 2076 * whether 'running' is true/false. 2077 */ 2078 2079 VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident); 2080 return (EINVAL); 2081} 2082 2083static int 2084svm_setcap(void *arg, int vcpu, int type, int val) 2085{ 2086 struct svm_softc *sc; 2087 int error; 2088 2089 sc = arg; 2090 error = 0; 2091 switch (type) { 2092 case VM_CAP_HALT_EXIT: 2093 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, 2094 VMCB_INTCPT_HLT, val); 2095 break; 2096 case VM_CAP_PAUSE_EXIT: 2097 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, 2098 VMCB_INTCPT_PAUSE, val); 2099 break; 2100 case VM_CAP_UNRESTRICTED_GUEST: 2101 /* Unrestricted guest execution cannot be disabled in SVM */ 2102 if (val == 0) 2103 error = EINVAL; 2104 break; 2105 default: 2106 error = ENOENT; 2107 break; 2108 } 2109 return (error); 2110} 2111 2112static int 2113svm_getcap(void *arg, int vcpu, int type, int *retval) 2114{ 2115 struct svm_softc *sc; 2116 int error; 2117 2118 sc = arg; 2119 error = 0; 2120 2121 switch (type) { 2122 case VM_CAP_HALT_EXIT: 2123 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, 2124 VMCB_INTCPT_HLT); 2125 break; 2126 case VM_CAP_PAUSE_EXIT: 2127 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, 2128 VMCB_INTCPT_PAUSE); 2129 break; 2130 case VM_CAP_UNRESTRICTED_GUEST: 2131 *retval = 1; /* unrestricted guest is always enabled */ 2132 break; 2133 default: 2134 error = ENOENT; 2135 break; 2136 } 2137 return (error); 2138} 2139 2140static struct vlapic * 2141svm_vlapic_init(void *arg, int vcpuid) 2142{ 2143 struct svm_softc *svm_sc; 2144 struct vlapic *vlapic; 2145 2146 svm_sc = arg; 2147 vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO); 2148 vlapic->vm = svm_sc->vm; 2149 vlapic->vcpuid = vcpuid; 2150 vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid]; 2151 2152 vlapic_init(vlapic); 2153 2154 return (vlapic); 2155} 2156 2157static void 2158svm_vlapic_cleanup(void *arg, struct vlapic *vlapic) 2159{ 2160 2161 vlapic_cleanup(vlapic); 2162 free(vlapic, M_SVM_VLAPIC); 2163} 2164 2165struct vmm_ops vmm_ops_amd = { 2166 svm_init, 2167 svm_cleanup, 2168 svm_restore, 2169 svm_vminit, 2170 svm_vmrun, 2171 svm_vmcleanup, 2172 svm_getreg, 2173 svm_setreg, 2174 vmcb_getdesc, 2175 vmcb_setdesc, 2176 svm_getcap, 2177 svm_setcap, 2178 svm_npt_alloc, 2179 svm_npt_free, 2180 svm_vlapic_init, 2181 svm_vlapic_cleanup 2182}; 2183