1/*-
2 * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com)
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice unmodified, this list of conditions, and the following
10 * disclaimer.
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.
14 *
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.
25 */
26
27#include <sys/cdefs.h>
28__FBSDID("$FreeBSD: projects/bhyve_svm/sys/amd64/vmm/amd/svm.c 271939 2014-09-21 23:42:54Z neel $");
29
30#include <sys/param.h>
31#include <sys/systm.h>
32#include <sys/smp.h>
33#include <sys/kernel.h>
34#include <sys/malloc.h>
35#include <sys/pcpu.h>
36#include <sys/proc.h>
37#include <sys/sysctl.h>
38
39#include <vm/vm.h>
40#include <vm/pmap.h>
41
42#include <machine/cpufunc.h>
43#include <machine/psl.h>
44#include <machine/pmap.h>
45#include <machine/md_var.h>
46#include <machine/vmparam.h>
47#include <machine/specialreg.h>
48#include <machine/segments.h>
49#include <machine/smp.h>
50#include <machine/vmm.h>
51#include <machine/vmm_dev.h>
52#include <machine/vmm_instruction_emul.h>
53
54#include <x86/apicreg.h>
55
56#include "vmm_lapic.h"
57#include "vmm_stat.h"
58#include "vmm_ktr.h"
59#include "vmm_ioport.h"
60#include "vatpic.h"
61#include "vlapic.h"
62#include "vlapic_priv.h"
63
64#include "x86.h"
65#include "vmcb.h"
66#include "svm.h"
67#include "svm_softc.h"
68#include "svm_msr.h"
69#include "npt.h"
70
71SYSCTL_DECL(_hw_vmm);
72SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL);
73
74/*
75 * SVM CPUID function 0x8000_000A, edx bit decoding.
76 */
77#define AMD_CPUID_SVM_NP BIT(0) /* Nested paging or RVI */
78#define AMD_CPUID_SVM_LBR BIT(1) /* Last branch virtualization */
79#define AMD_CPUID_SVM_SVML BIT(2) /* SVM lock */
80#define AMD_CPUID_SVM_NRIP_SAVE BIT(3) /* Next RIP is saved */
81#define AMD_CPUID_SVM_TSC_RATE BIT(4) /* TSC rate control. */
82#define AMD_CPUID_SVM_VMCB_CLEAN BIT(5) /* VMCB state caching */
83#define AMD_CPUID_SVM_FLUSH_BY_ASID BIT(6) /* Flush by ASID */
84#define AMD_CPUID_SVM_DECODE_ASSIST BIT(7) /* Decode assist */
85#define AMD_CPUID_SVM_PAUSE_INC BIT(10) /* Pause intercept filter. */
86#define AMD_CPUID_SVM_PAUSE_FTH BIT(12) /* Pause filter threshold */
87
88#define VMCB_CACHE_DEFAULT (VMCB_CACHE_ASID | \
89 VMCB_CACHE_IOPM | \
90 VMCB_CACHE_I | \
91 VMCB_CACHE_TPR | \
92 VMCB_CACHE_CR2 | \
93 VMCB_CACHE_CR | \
94 VMCB_CACHE_DT | \
95 VMCB_CACHE_SEG | \
96 VMCB_CACHE_NP)
97
98static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT;
99SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean,
100 0, NULL);
101
102MALLOC_DEFINE(M_SVM, "svm", "svm");
103MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic");
104
105/* Per-CPU context area. */
106extern struct pcpu __pcpu[];
107
108static uint32_t svm_feature; /* AMD SVM features. */
109SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RD, &svm_feature, 0,
110 "SVM features advertised by CPUID.8000000AH:EDX");
111
112static int disable_npf_assist;
113SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN,
114 &disable_npf_assist, 0, NULL);
115
116/* Maximum ASIDs supported by the processor */
117static uint32_t nasid;
118SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RD, &nasid, 0,
119 "Number of ASIDs supported by this processor");
120
121/* Current ASID generation for each host cpu */
122static struct asid asid[MAXCPU];
123
124/*
125 * SVM host state saved area of size 4KB for each core.
126 */
127static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE);
128
129/*
130 * S/w saved host context.
131 */
132static struct svm_regctx host_ctx[MAXCPU];
133
134static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery");
135static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry");
136static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window");
137
138static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val);
139
140/*
141 * Common function to enable or disabled SVM for a CPU.
142 */
143static int
144cpu_svm_enable_disable(boolean_t enable)
145{
146 uint64_t efer_msr;
147
148 efer_msr = rdmsr(MSR_EFER);
149
150 if (enable)
151 efer_msr |= EFER_SVM;
152 else
153 efer_msr &= ~EFER_SVM;
154
155 wrmsr(MSR_EFER, efer_msr);
156
157 return(0);
158}
159
160/*
161 * Disable SVM on a CPU.
162 */
163static void
164svm_disable(void *arg __unused)
165{
166
167 (void)cpu_svm_enable_disable(FALSE);
168}
169
170/*
171 * Disable SVM for all CPUs.
172 */
173static int
174svm_cleanup(void)
175{
176
177 smp_rendezvous(NULL, svm_disable, NULL, NULL);
178 return (0);
179}
180
181/*
182 * Check for required BHyVe SVM features in a CPU.
183 */
184static int
185svm_cpuid_features(void)
186{
187 u_int regs[4];
188
189 /* CPUID Fn8000_000A is for SVM */
190 do_cpuid(0x8000000A, regs);
191 svm_feature = regs[3];
192
193 printf("SVM rev: 0x%x NASID:0x%x\n", regs[0] & 0xFF, regs[1]);
194 nasid = regs[1];
195 KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid));
196
197 printf("SVM Features:0x%b\n", svm_feature,
198 "\020"
199 "\001NP" /* Nested paging */
200 "\002LbrVirt" /* LBR virtualization */
201 "\003SVML" /* SVM lock */
202 "\004NRIPS" /* NRIP save */
203 "\005TscRateMsr" /* MSR based TSC rate control */
204 "\006VmcbClean" /* VMCB clean bits */
205 "\007FlushByAsid" /* Flush by ASID */
206 "\010DecodeAssist" /* Decode assist */
207 "\011<b20>"
208 "\012<b20>"
209 "\013PauseFilter"
210 "\014<b20>"
211 "\015PauseFilterThreshold"
212 "\016AVIC"
213 );
214
215 /* SVM Lock */
216 if (!(svm_feature & AMD_CPUID_SVM_SVML)) {
217 printf("SVM is disabled by BIOS, please enable in BIOS.\n");
218 return (ENXIO);
219 }
220
221 /*
222 * bhyve need RVI to work.
223 */
224 if (!(svm_feature & AMD_CPUID_SVM_NP)) {
225 printf("Missing Nested paging or RVI SVM support in processor.\n");
226 return (EIO);
227 }
228
229 if (svm_feature & AMD_CPUID_SVM_NRIP_SAVE)
230 return (0);
231
232 return (EIO);
233}
234
235static __inline int
236flush_by_asid(void)
237{
238
239 return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID);
240}
241
242static __inline int
243decode_assist(void)
244{
245
246 return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST);
247}
248
249/*
250 * Enable SVM for a CPU.
251 */
252static void
253svm_enable(void *arg __unused)
254{
255 uint64_t hsave_pa;
256
257 (void)cpu_svm_enable_disable(TRUE);
258
259 hsave_pa = vtophys(hsave[curcpu]);
260 wrmsr(MSR_VM_HSAVE_PA, hsave_pa);
261
262 if (rdmsr(MSR_VM_HSAVE_PA) != hsave_pa) {
263 panic("VM_HSAVE_PA is wrong on CPU%d\n", curcpu);
264 }
265}
266
267/*
268 * Check if a processor support SVM.
269 */
270static int
271is_svm_enabled(void)
272{
273 uint64_t msr;
274
275 /* Section 15.4 Enabling SVM from APM2. */
276 if ((amd_feature2 & AMDID2_SVM) == 0) {
277 printf("SVM is not supported on this processor.\n");
278 return (ENXIO);
279 }
280
281 msr = rdmsr(MSR_VM_CR);
282 /* Make sure SVM is not disabled by BIOS. */
283 if ((msr & VM_CR_SVMDIS) == 0) {
284 return svm_cpuid_features();
285 }
286
287 printf("SVM disabled by Key, consult TPM/BIOS manual.\n");
288 return (ENXIO);
289}
290
291/*
292 * Enable SVM on CPU and initialize nested page table h/w.
293 */
294static int
295svm_init(int ipinum)
296{
297 int err, cpu;
298
299 err = is_svm_enabled();
300 if (err)
301 return (err);
302
303 vmcb_clean &= VMCB_CACHE_DEFAULT;
304
305 for (cpu = 0; cpu < MAXCPU; cpu++) {
306 /*
307 * Initialize the host ASIDs to their "highest" valid values.
308 *
309 * The next ASID allocation will rollover both 'gen' and 'num'
310 * and start off the sequence at {1,1}.
311 */
312 asid[cpu].gen = ~0UL;
313 asid[cpu].num = nasid - 1;
314 }
315
316 svm_msr_init();
317 svm_npt_init(ipinum);
318
319 /* Start SVM on all CPUs */
320 smp_rendezvous(NULL, svm_enable, NULL, NULL);
321
322 return (0);
323}
324
325static void
326svm_restore(void)
327{
328 svm_enable(NULL);
329}
330
331/*
332 * Get index and bit position for a MSR in MSR permission
333 * bitmap. Two bits are used for each MSR, lower bit is
334 * for read and higher bit is for write.
335 */
336static int
337svm_msr_index(uint64_t msr, int *index, int *bit)
338{
339 uint32_t base, off;
340
341/* Pentium compatible MSRs */
342#define MSR_PENTIUM_START 0
343#define MSR_PENTIUM_END 0x1FFF
344/* AMD 6th generation and Intel compatible MSRs */
345#define MSR_AMD6TH_START 0xC0000000UL
346#define MSR_AMD6TH_END 0xC0001FFFUL
347/* AMD 7th and 8th generation compatible MSRs */
348#define MSR_AMD7TH_START 0xC0010000UL
349#define MSR_AMD7TH_END 0xC0011FFFUL
350
351 *index = -1;
352 *bit = (msr % 4) * 2;
353 base = 0;
354
355 if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) {
356 *index = msr / 4;
357 return (0);
358 }
359
360 base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1);
361 if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) {
362 off = (msr - MSR_AMD6TH_START);
363 *index = (off + base) / 4;
364 return (0);
365 }
366
367 base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1);
368 if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) {
369 off = (msr - MSR_AMD7TH_START);
370 *index = (off + base) / 4;
371 return (0);
372 }
373
374 return (EIO);
375}
376
377/*
378 * Give virtual cpu the complete access to MSR(read & write).
379 */
380static int
381svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write)
382{
383 int index, bit, err;
384
385 err = svm_msr_index(msr, &index, &bit);
386 if (err) {
387 ERR("MSR 0x%lx is not writeable by guest.\n", msr);
388 return (err);
389 }
390
391 if (index < 0 || index > (SVM_MSR_BITMAP_SIZE)) {
392 ERR("MSR 0x%lx index out of range(%d).\n", msr, index);
393 return (EINVAL);
394 }
395 if (bit < 0 || bit > 8) {
396 ERR("MSR 0x%lx bit out of range(%d).\n", msr, bit);
397 return (EINVAL);
398 }
399
400 /* Disable intercept for read and write. */
401 if (read)
402 perm_bitmap[index] &= ~(1UL << bit);
403 if (write)
404 perm_bitmap[index] &= ~(2UL << bit);
405 CTR2(KTR_VMM, "Guest has control:0x%x on SVM:MSR(0x%lx).\n",
406 (perm_bitmap[index] >> bit) & 0x3, msr);
407
408 return (0);
409}
410
411static int
412svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr)
413{
414 return svm_msr_perm(perm_bitmap, msr, true, true);
415}
416
417static int
418svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr)
419{
420 return svm_msr_perm(perm_bitmap, msr, true, false);
421}
422
423static __inline int
424svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask)
425{
426 struct vmcb_ctrl *ctrl;
427
428 KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
429
430 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
431 return (ctrl->intercept[idx] & bitmask ? 1 : 0);
432}
433
434static __inline void
435svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask,
436 int enabled)
437{
438 struct vmcb_ctrl *ctrl;
439 uint32_t oldval;
440
441 KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
442
443 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
444 oldval = ctrl->intercept[idx];
445
446 if (enabled)
447 ctrl->intercept[idx] |= bitmask;
448 else
449 ctrl->intercept[idx] &= ~bitmask;
450
451 if (ctrl->intercept[idx] != oldval) {
452 svm_set_dirty(sc, vcpu, VMCB_CACHE_I);
453 VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified "
454 "from %#x to %#x", idx, oldval, ctrl->intercept[idx]);
455 }
456}
457
458static __inline void
459svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
460{
461 svm_set_intercept(sc, vcpu, off, bitmask, 0);
462}
463
464static __inline void
465svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
466{
467 svm_set_intercept(sc, vcpu, off, bitmask, 1);
468}
469
470static void
471vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa,
472 uint64_t msrpm_base_pa, uint64_t np_pml4)
473{
474 struct vmcb_ctrl *ctrl;
475 struct vmcb_state *state;
476 uint32_t mask;
477 int n;
478
479 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
480 state = svm_get_vmcb_state(sc, vcpu);
481
482 ctrl->iopm_base_pa = iopm_base_pa;
483 ctrl->msrpm_base_pa = msrpm_base_pa;
484
485 /* Enable nested paging */
486 ctrl->np_enable = 1;
487 ctrl->n_cr3 = np_pml4;
488
489 /*
490 * Intercept accesses to the control registers that are not shadowed
491 * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8.
492 */
493 for (n = 0; n < 16; n++) {
494 mask = (BIT(n) << 16) | BIT(n);
495 if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8)
496 svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
497 else
498 svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
499 }
500
501 /* Intercept Machine Check exceptions. */
502 svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC));
503
504 /* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */
505 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO);
506 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR);
507 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID);
508 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR);
509 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT);
510 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI);
511 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI);
512 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN);
513 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
514 VMCB_INTCPT_FERR_FREEZE);
515
516 /*
517 * From section "Canonicalization and Consistency Checks" in APMv2
518 * the VMRUN intercept bit must be set to pass the consistency check.
519 */
520 svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
521
522 /*
523 * The ASID will be set to a non-zero value just before VMRUN.
524 */
525 ctrl->asid = 0;
526
527 /*
528 * Section 15.21.1, Interrupt Masking in EFLAGS
529 * Section 15.21.2, Virtualizing APIC.TPR
530 *
531 * This must be set for %rflag and %cr8 isolation of guest and host.
532 */
533 ctrl->v_intr_masking = 1;
534
535 /* Enable Last Branch Record aka LBR for debugging */
536 ctrl->lbr_virt_en = 1;
537 state->dbgctl = BIT(0);
538
539 /* EFER_SVM must always be set when the guest is executing */
540 state->efer = EFER_SVM;
541
542 /* Set up the PAT to power-on state */
543 state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK) |
544 PAT_VALUE(1, PAT_WRITE_THROUGH) |
545 PAT_VALUE(2, PAT_UNCACHED) |
546 PAT_VALUE(3, PAT_UNCACHEABLE) |
547 PAT_VALUE(4, PAT_WRITE_BACK) |
548 PAT_VALUE(5, PAT_WRITE_THROUGH) |
549 PAT_VALUE(6, PAT_UNCACHED) |
550 PAT_VALUE(7, PAT_UNCACHEABLE);
551}
552
553/*
554 * Initialise a virtual machine.
555 */
556static void *
557svm_vminit(struct vm *vm, pmap_t pmap)
558{
559 struct svm_softc *svm_sc;
560 struct svm_vcpu *vcpu;
561 vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
562 int i;
563
564 svm_sc = (struct svm_softc *)malloc(sizeof (struct svm_softc),
565 M_SVM, M_WAITOK | M_ZERO);
566
567 svm_sc->vm = vm;
568 svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4);
569
570 /*
571 * Intercept MSR access to all MSRs except GSBASE, FSBASE,... etc.
572 */
573 memset(svm_sc->msr_bitmap, 0xFF, sizeof(svm_sc->msr_bitmap));
574
575 /*
576 * Following MSR can be completely controlled by virtual machines
577 * since access to following are translated to access to VMCB.
578 */
579 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
580 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
581 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
582
583 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
584 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
585 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
586 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
587 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
588 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
589 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
590
591 /* For Nested Paging/RVI only. */
592 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
593
594 svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
595
596 /*
597 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
598 */
599 svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
600
601 /* Intercept access to all I/O ports. */
602 memset(svm_sc->iopm_bitmap, 0xFF, sizeof(svm_sc->iopm_bitmap));
603
604 /* Cache physical address for multiple vcpus. */
605 iopm_pa = vtophys(svm_sc->iopm_bitmap);
606 msrpm_pa = vtophys(svm_sc->msr_bitmap);
607 pml4_pa = svm_sc->nptp;
608
609 for (i = 0; i < VM_MAXCPU; i++) {
610 vcpu = svm_get_vcpu(svm_sc, i);
611 vcpu->lastcpu = NOCPU;
612 vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
613 vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
614 svm_msr_guest_init(svm_sc, i);
615 }
616 return (svm_sc);
617}
618
619static int
620svm_cpl(struct vmcb_state *state)
621{
622
623 /*
624 * From APMv2:
625 * "Retrieve the CPL from the CPL field in the VMCB, not
626 * from any segment DPL"
627 */
628 return (state->cpl);
629}
630
631static enum vm_cpu_mode
632svm_vcpu_mode(struct vmcb *vmcb)
633{
634 struct vmcb_segment seg;
635 struct vmcb_state *state;
636 int error;
637
638 state = &vmcb->state;
639
640 if (state->efer & EFER_LMA) {
641 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
642 KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
643 error));
644
645 /*
646 * Section 4.8.1 for APM2, check if Code Segment has
647 * Long attribute set in descriptor.
648 */
649 if (seg.attrib & VMCB_CS_ATTRIB_L)
650 return (CPU_MODE_64BIT);
651 else
652 return (CPU_MODE_COMPATIBILITY);
653 } else if (state->cr0 & CR0_PE) {
654 return (CPU_MODE_PROTECTED);
655 } else {
656 return (CPU_MODE_REAL);
657 }
658}
659
660static enum vm_paging_mode
661svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
662{
663
664 if ((cr0 & CR0_PG) == 0)
665 return (PAGING_MODE_FLAT);
666 if ((cr4 & CR4_PAE) == 0)
667 return (PAGING_MODE_32);
668 if (efer & EFER_LME)
669 return (PAGING_MODE_64);
670 else
671 return (PAGING_MODE_PAE);
672}
673
674/*
675 * ins/outs utility routines
676 */
677static uint64_t
678svm_inout_str_index(struct svm_regctx *regs, int in)
679{
680 uint64_t val;
681
682 val = in ? regs->e.g.sctx_rdi : regs->e.g.sctx_rsi;
683
684 return (val);
685}
686
687static uint64_t
688svm_inout_str_count(struct svm_regctx *regs, int rep)
689{
690 uint64_t val;
691
692 val = rep ? regs->sctx_rcx : 1;
693
694 return (val);
695}
696
697static void
698svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
699 int in, struct vm_inout_str *vis)
700{
701 int error, s;
702
703 if (in) {
704 vis->seg_name = VM_REG_GUEST_ES;
705 } else {
706 /* The segment field has standard encoding */
707 s = (info1 >> 10) & 0x7;
708 vis->seg_name = vm_segment_name(s);
709 }
710
711 error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
712 KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
713}
714
715static int
716svm_inout_str_addrsize(uint64_t info1)
717{
718 uint32_t size;
719
720 size = (info1 >> 7) & 0x7;
721 switch (size) {
722 case 1:
723 return (2); /* 16 bit */
724 case 2:
725 return (4); /* 32 bit */
726 case 4:
727 return (8); /* 64 bit */
728 default:
729 panic("%s: invalid size encoding %d", __func__, size);
730 }
731}
732
733static void
734svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
735{
736 struct vmcb_state *state;
737
738 state = &vmcb->state;
739 paging->cr3 = state->cr3;
740 paging->cpl = svm_cpl(state);
741 paging->cpu_mode = svm_vcpu_mode(vmcb);
742 paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
743 state->efer);
744}
745
746#define UNHANDLED 0
747
748/*
749 * Handle guest I/O intercept.
750 */
751static int
752svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
753{
754 struct vmcb_ctrl *ctrl;
755 struct vmcb_state *state;
756 struct svm_regctx *regs;
757 struct vm_inout_str *vis;
758 uint64_t info1;
759 int inout_string;
760
761 state = svm_get_vmcb_state(svm_sc, vcpu);
762 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
763 regs = svm_get_guest_regctx(svm_sc, vcpu);
764
765 info1 = ctrl->exitinfo1;
766 inout_string = info1 & BIT(2) ? 1 : 0;
767
768 /*
769 * The effective segment number in EXITINFO1[12:10] is populated
770 * only if the processor has the DecodeAssist capability.
771 *
772 * XXX this is not specified explicitly in APMv2 but can be verified
773 * empirically.
774 */
775 if (inout_string && !decode_assist())
776 return (UNHANDLED);
777
778 vmexit->exitcode = VM_EXITCODE_INOUT;
779 vmexit->u.inout.in = (info1 & BIT(0)) ? 1 : 0;
780 vmexit->u.inout.string = inout_string;
781 vmexit->u.inout.rep = (info1 & BIT(3)) ? 1 : 0;
782 vmexit->u.inout.bytes = (info1 >> 4) & 0x7;
783 vmexit->u.inout.port = (uint16_t)(info1 >> 16);
784 vmexit->u.inout.eax = (uint32_t)(state->rax);
785
786 if (inout_string) {
787 vmexit->exitcode = VM_EXITCODE_INOUT_STR;
788 vis = &vmexit->u.inout_str;
789 svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
790 vis->rflags = state->rflags;
791 vis->cr0 = state->cr0;
792 vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
793 vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
794 vis->addrsize = svm_inout_str_addrsize(info1);
795 svm_inout_str_seginfo(svm_sc, vcpu, info1,
796 vmexit->u.inout.in, vis);
797 }
798
799 return (UNHANDLED);
800}
801
802static int
803svm_npf_paging(uint64_t exitinfo1)
804{
805
806 if (exitinfo1 & VMCB_NPF_INFO1_W)
807 return (VM_PROT_WRITE);
808
809 return (VM_PROT_READ);
810}
811
812static bool
813svm_npf_emul_fault(uint64_t exitinfo1)
814{
815
816 if (exitinfo1 & VMCB_NPF_INFO1_ID) {
817 return (false);
818 }
819
820 if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
821 return (false);
822 }
823
824 if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
825 return (false);
826 }
827
828 return (true);
829}
830
831static void
832svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
833{
834 struct vm_guest_paging *paging;
835 struct vmcb_segment seg;
836 struct vmcb_ctrl *ctrl;
837 char *inst_bytes;
838 int error, inst_len;
839
840 ctrl = &vmcb->ctrl;
841 paging = &vmexit->u.inst_emul.paging;
842
843 vmexit->exitcode = VM_EXITCODE_INST_EMUL;
844 vmexit->u.inst_emul.gpa = gpa;
845 vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
846 svm_paging_info(vmcb, paging);
847
848 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
849 KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
850
851 switch(paging->cpu_mode) {
852 case CPU_MODE_PROTECTED:
853 case CPU_MODE_COMPATIBILITY:
854 /*
855 * Section 4.8.1 of APM2, Default Operand Size or D bit.
856 */
857 vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
858 1 : 0;
859 break;
860 default:
861 vmexit->u.inst_emul.cs_d = 0;
862 break;
863 }
864
865 /*
866 * Copy the instruction bytes into 'vie' if available.
867 */
868 if (decode_assist() && !disable_npf_assist) {
869 inst_len = ctrl->inst_len;
870 inst_bytes = ctrl->inst_bytes;
871 } else {
872 inst_len = 0;
873 inst_bytes = NULL;
874 }
875 vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
876}
877
878#ifdef KTR
879static const char *
880intrtype_to_str(int intr_type)
881{
882 switch (intr_type) {
883 case VMCB_EVENTINJ_TYPE_INTR:
884 return ("hwintr");
885 case VMCB_EVENTINJ_TYPE_NMI:
886 return ("nmi");
887 case VMCB_EVENTINJ_TYPE_INTn:
888 return ("swintr");
889 case VMCB_EVENTINJ_TYPE_EXCEPTION:
890 return ("exception");
891 default:
892 panic("%s: unknown intr_type %d", __func__, intr_type);
893 }
894}
895#endif
896
897/*
898 * Inject an event to vcpu as described in section 15.20, "Event injection".
899 */
900static void
901svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
902 uint32_t error, bool ec_valid)
903{
904 struct vmcb_ctrl *ctrl;
905
906 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
907
908 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
909 ("%s: event already pending %#lx", __func__, ctrl->eventinj));
910
911 KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
912 __func__, vector));
913
914 switch (intr_type) {
915 case VMCB_EVENTINJ_TYPE_INTR:
916 case VMCB_EVENTINJ_TYPE_NMI:
917 case VMCB_EVENTINJ_TYPE_INTn:
918 break;
919 case VMCB_EVENTINJ_TYPE_EXCEPTION:
920 if (vector >= 0 && vector <= 31 && vector != 2)
921 break;
922 /* FALLTHROUGH */
923 default:
924 panic("%s: invalid intr_type/vector: %d/%d", __func__,
925 intr_type, vector);
926 }
927 ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
928 if (ec_valid) {
929 ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
930 ctrl->eventinj |= (uint64_t)error << 32;
931 VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
932 intrtype_to_str(intr_type), vector, error);
933 } else {
934 VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
935 intrtype_to_str(intr_type), vector);
936 }
937}
938
939static void
940svm_update_virqinfo(struct svm_softc *sc, int vcpu)
941{
942 struct vm *vm;
943 struct vlapic *vlapic;
944 struct vmcb_ctrl *ctrl;
945 int pending;
946
947 vm = sc->vm;
948 vlapic = vm_lapic(vm, vcpu);
949 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
950
951 /* Update %cr8 in the emulated vlapic */
952 vlapic_set_cr8(vlapic, ctrl->v_tpr);
953
954 /*
955 * If V_IRQ indicates that the interrupt injection attempted on then
956 * last VMRUN was successful then update the vlapic accordingly.
957 */
958 if (ctrl->v_intr_vector != 0) {
959 pending = ctrl->v_irq;
960 KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid "
961 "v_intr_vector %d", __func__, ctrl->v_intr_vector));
962 KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
963 VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector,
964 pending ? "pending" : "accepted");
965 if (!pending)
966 vlapic_intr_accepted(vlapic, ctrl->v_intr_vector);
967 }
968}
969
970static void
971svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
972{
973 struct vmcb_ctrl *ctrl;
974 uint64_t intinfo;
975
976 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
977 intinfo = ctrl->exitintinfo;
978 if (!VMCB_EXITINTINFO_VALID(intinfo))
979 return;
980
981 /*
982 * From APMv2, Section "Intercepts during IDT interrupt delivery"
983 *
984 * If a #VMEXIT happened during event delivery then record the event
985 * that was being delivered.
986 */
987 VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
988 intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
989 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
990 vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
991}
992
993static __inline int
994vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
995{
996
997 return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
998 VMCB_INTCPT_VINTR));
999}
1000
1001static __inline void
1002enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1003{
1004 struct vmcb_ctrl *ctrl;
1005
1006 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1007
1008 if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
1009 KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
1010 KASSERT(vintr_intercept_enabled(sc, vcpu),
1011 ("%s: vintr intercept should be enabled", __func__));
1012 return;
1013 }
1014
1015 VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
1016 ctrl->v_irq = 1;
1017 ctrl->v_ign_tpr = 1;
1018 ctrl->v_intr_vector = 0;
1019 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1020 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1021}
1022
1023static __inline void
1024disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1025{
1026 struct vmcb_ctrl *ctrl;
1027
1028 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1029
1030 if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1031 KASSERT(!vintr_intercept_enabled(sc, vcpu),
1032 ("%s: vintr intercept should be disabled", __func__));
1033 return;
1034 }
1035
1036#ifdef KTR
1037 if (ctrl->v_intr_vector == 0)
1038 VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
1039 else
1040 VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection");
1041#endif
1042 ctrl->v_irq = 0;
1043 ctrl->v_intr_vector = 0;
1044 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1045 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1046}
1047
1048static int
1049svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
1050{
1051 struct vmcb_ctrl *ctrl;
1052 int oldval, newval;
1053
1054 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1055 oldval = ctrl->intr_shadow;
1056 newval = val ? 1 : 0;
1057 if (newval != oldval) {
1058 ctrl->intr_shadow = newval;
1059 VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1060 }
1061 return (0);
1062}
1063
1064static int
1065svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1066{
1067 struct vmcb_ctrl *ctrl;
1068
1069 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1070 *val = ctrl->intr_shadow;
1071 return (0);
1072}
1073
1074/*
1075 * Once an NMI is injected it blocks delivery of further NMIs until the handler
1076 * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1077 * to track when the vcpu is done handling the NMI.
1078 */
1079static int
1080nmi_blocked(struct svm_softc *sc, int vcpu)
1081{
1082 int blocked;
1083
1084 blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1085 VMCB_INTCPT_IRET);
1086 return (blocked);
1087}
1088
1089static void
1090enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1091{
1092
1093 KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1094 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1095 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1096}
1097
1098static void
1099clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1100{
1101 int error;
1102
1103 KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1104 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1105 /*
1106 * When the IRET intercept is cleared the vcpu will attempt to execute
1107 * the "iret" when it runs next. However, it is possible to inject
1108 * another NMI into the vcpu before the "iret" has actually executed.
1109 *
1110 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1111 * it will trap back into the hypervisor. If an NMI is pending for
1112 * the vcpu it will be injected into the guest.
1113 *
1114 * XXX this needs to be fixed
1115 */
1116 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1117
1118 /*
1119 * Set 'intr_shadow' to prevent an NMI from being injected on the
1120 * immediate VMRUN.
1121 */
1122 error = svm_modify_intr_shadow(sc, vcpu, 1);
1123 KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1124}
1125
1126static int
1127emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1128 bool *retu)
1129{
1130 int error;
1131
1132 if (lapic_msr(num))
1133 error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1134 else if (num == MSR_EFER)
1135 error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, val);
1136 else
1137 error = svm_wrmsr(sc, vcpu, num, val, retu);
1138
1139 return (error);
1140}
1141
1142static int
1143emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1144{
1145 struct vmcb_state *state;
1146 struct svm_regctx *ctx;
1147 uint64_t result;
1148 int error;
1149
1150 if (lapic_msr(num))
1151 error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1152 else
1153 error = svm_rdmsr(sc, vcpu, num, &result, retu);
1154
1155 if (error == 0) {
1156 state = svm_get_vmcb_state(sc, vcpu);
1157 ctx = svm_get_guest_regctx(sc, vcpu);
1158 state->rax = result & 0xffffffff;
1159 ctx->e.g.sctx_rdx = result >> 32;
1160 }
1161
1162 return (error);
1163}
1164
1165#ifdef KTR
1166static const char *
1167exit_reason_to_str(uint64_t reason)
1168{
1169 static char reasonbuf[32];
1170
1171 switch (reason) {
1172 case VMCB_EXIT_INVALID:
1173 return ("invalvmcb");
1174 case VMCB_EXIT_SHUTDOWN:
1175 return ("shutdown");
1176 case VMCB_EXIT_NPF:
1177 return ("nptfault");
1178 case VMCB_EXIT_PAUSE:
1179 return ("pause");
1180 case VMCB_EXIT_HLT:
1181 return ("hlt");
1182 case VMCB_EXIT_CPUID:
1183 return ("cpuid");
1184 case VMCB_EXIT_IO:
1185 return ("inout");
1186 case VMCB_EXIT_MC:
1187 return ("mchk");
1188 case VMCB_EXIT_INTR:
1189 return ("extintr");
1190 case VMCB_EXIT_NMI:
1191 return ("nmi");
1192 case VMCB_EXIT_VINTR:
1193 return ("vintr");
1194 case VMCB_EXIT_MSR:
1195 return ("msr");
1196 case VMCB_EXIT_IRET:
1197 return ("iret");
1198 default:
1199 snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1200 return (reasonbuf);
1201 }
1202}
1203#endif /* KTR */
1204
1205/*
1206 * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1207 * that are due to instruction intercepts as well as MSR and IOIO intercepts
1208 * and exceptions caused by INT3, INTO and BOUND instructions.
1209 *
1210 * Return 1 if the nRIP is valid and 0 otherwise.
1211 */
1212static int
1213nrip_valid(uint64_t exitcode)
1214{
1215 switch (exitcode) {
1216 case 0x00 ... 0x0F: /* read of CR0 through CR15 */
1217 case 0x10 ... 0x1F: /* write of CR0 through CR15 */
1218 case 0x20 ... 0x2F: /* read of DR0 through DR15 */
1219 case 0x30 ... 0x3F: /* write of DR0 through DR15 */
1220 case 0x43: /* INT3 */
1221 case 0x44: /* INTO */
1222 case 0x45: /* BOUND */
1223 case 0x65 ... 0x7C: /* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1224 case 0x80 ... 0x8D: /* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1225 return (1);
1226 default:
1227 return (0);
1228 }
1229}
1230
1231/*
1232 * Collateral for a generic SVM VM-exit.
1233 */
1234static void
1235vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
1236{
1237
1238 vme->exitcode = VM_EXITCODE_SVM;
1239 vme->u.svm.exitcode = code;
1240 vme->u.svm.exitinfo1 = info1;
1241 vme->u.svm.exitinfo2 = info2;
1242}
1243
1244static int
1245svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1246{
1247 struct vmcb *vmcb;
1248 struct vmcb_state *state;
1249 struct vmcb_ctrl *ctrl;
1250 struct svm_regctx *ctx;
1251 uint64_t code, info1, info2, val;
1252 uint32_t eax, ecx, edx;
1253 int handled;
1254 bool retu;
1255
1256 ctx = svm_get_guest_regctx(svm_sc, vcpu);
1257 vmcb = svm_get_vmcb(svm_sc, vcpu);
1258 state = &vmcb->state;
1259 ctrl = &vmcb->ctrl;
1260
1261 handled = 0;
1262 code = ctrl->exitcode;
1263 info1 = ctrl->exitinfo1;
1264 info2 = ctrl->exitinfo2;
1265
1266 vmexit->exitcode = VM_EXITCODE_BOGUS;
1267 vmexit->rip = state->rip;
1268 vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1269
1270 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1271
1272 /*
1273 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1274 * in an inconsistent state and can trigger assertions that would
1275 * never happen otherwise.
1276 */
1277 if (code == VMCB_EXIT_INVALID) {
1278 vm_exit_svm(vmexit, code, info1, info2);
1279 return (0);
1280 }
1281
1282 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1283 "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1284
1285 KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1286 ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1287 vmexit->inst_length, code, info1, info2));
1288
1289 svm_update_virqinfo(svm_sc, vcpu);
1290 svm_save_intinfo(svm_sc, vcpu);
1291
1292 switch (code) {
1293 case VMCB_EXIT_IRET:
1294 /*
1295 * Restart execution at "iret" but with the intercept cleared.
1296 */
1297 vmexit->inst_length = 0;
1298 clear_nmi_blocking(svm_sc, vcpu);
1299 handled = 1;
1300 break;
1301 case VMCB_EXIT_VINTR: /* interrupt window exiting */
1302 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1303 handled = 1;
1304 break;
1305 case VMCB_EXIT_INTR: /* external interrupt */
1306 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1307 handled = 1;
1308 break;
1309 case VMCB_EXIT_NMI: /* external NMI */
1310 handled = 1;
1311 break;
1312 case VMCB_EXIT_MC: /* machine check */
1313 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1314 break;
1315 case VMCB_EXIT_MSR: /* MSR access. */
1316 eax = state->rax;
1317 ecx = ctx->sctx_rcx;
1318 edx = ctx->e.g.sctx_rdx;
1319 retu = false;
1320
1321 if (info1) {
1322 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1323 val = (uint64_t)edx << 32 | eax;
1324 VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1325 ecx, val);
1326 if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1327 vmexit->exitcode = VM_EXITCODE_WRMSR;
1328 vmexit->u.msr.code = ecx;
1329 vmexit->u.msr.wval = val;
1330 } else if (!retu) {
1331 handled = 1;
1332 } else {
1333 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1334 ("emulate_wrmsr retu with bogus exitcode"));
1335 }
1336 } else {
1337 VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1338 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1339 if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1340 vmexit->exitcode = VM_EXITCODE_RDMSR;
1341 vmexit->u.msr.code = ecx;
1342 } else if (!retu) {
1343 handled = 1;
1344 } else {
1345 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1346 ("emulate_rdmsr retu with bogus exitcode"));
1347 }
1348 }
1349 break;
1350 case VMCB_EXIT_IO:
1351 handled = svm_handle_io(svm_sc, vcpu, vmexit);
1352 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1353 break;
1354 case VMCB_EXIT_CPUID:
1355 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1356 handled = x86_emulate_cpuid(svm_sc->vm, vcpu,
1357 (uint32_t *)&state->rax,
1358 (uint32_t *)&ctx->sctx_rbx,
1359 (uint32_t *)&ctx->sctx_rcx,
1360 (uint32_t *)&ctx->e.g.sctx_rdx);
1361 break;
1362 case VMCB_EXIT_HLT:
1363 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1364 vmexit->exitcode = VM_EXITCODE_HLT;
1365 vmexit->u.hlt.rflags = state->rflags;
1366 break;
1367 case VMCB_EXIT_PAUSE:
1368 vmexit->exitcode = VM_EXITCODE_PAUSE;
1369 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1370 break;
1371 case VMCB_EXIT_NPF:
1372 /* EXITINFO2 contains the faulting guest physical address */
1373 if (info1 & VMCB_NPF_INFO1_RSV) {
1374 VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1375 "reserved bits set: info1(%#lx) info2(%#lx)",
1376 info1, info2);
1377 } else if (vm_mem_allocated(svm_sc->vm, info2)) {
1378 vmexit->exitcode = VM_EXITCODE_PAGING;
1379 vmexit->u.paging.gpa = info2;
1380 vmexit->u.paging.fault_type = svm_npf_paging(info1);
1381 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1382 VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1383 "on gpa %#lx/%#lx at rip %#lx",
1384 info2, info1, state->rip);
1385 } else if (svm_npf_emul_fault(info1)) {
1386 svm_handle_inst_emul(vmcb, info2, vmexit);
1387 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1388 VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1389 "for gpa %#lx/%#lx at rip %#lx",
1390 info2, info1, state->rip);
1391 }
1392 break;
1393 default:
1394 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1395 break;
1396 }
1397
1398 VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1399 handled ? "handled" : "unhandled", exit_reason_to_str(code),
1400 vmexit->rip, vmexit->inst_length);
1401
1402 if (handled) {
1403 vmexit->rip += vmexit->inst_length;
1404 vmexit->inst_length = 0;
1405 state->rip = vmexit->rip;
1406 } else {
1407 if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1408 /*
1409 * If this VM exit was not claimed by anybody then
1410 * treat it as a generic SVM exit.
1411 */
1412 vm_exit_svm(vmexit, code, info1, info2);
1413 } else {
1414 /*
1415 * The exitcode and collateral have been populated.
1416 * The VM exit will be processed further in userland.
1417 */
1418 }
1419 }
1420 return (handled);
1421}
1422
1423static void
1424svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1425{
1426 uint64_t intinfo;
1427
1428 if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1429 return;
1430
1431 KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1432 "valid: %#lx", __func__, intinfo));
1433
1434 svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1435 VMCB_EXITINTINFO_VECTOR(intinfo),
1436 VMCB_EXITINTINFO_EC(intinfo),
1437 VMCB_EXITINTINFO_EC_VALID(intinfo));
1438 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1439 VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1440}
1441
1442/*
1443 * Inject event to virtual cpu.
1444 */
1445static void
1446svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1447{
1448 struct vmcb_ctrl *ctrl;
1449 struct vmcb_state *state;
1450 uint8_t v_tpr;
1451 int vector, need_intr_window, pending_apic_vector;
1452
1453 state = svm_get_vmcb_state(sc, vcpu);
1454 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1455
1456 need_intr_window = 0;
1457 pending_apic_vector = 0;
1458
1459 /*
1460 * Inject pending events or exceptions for this vcpu.
1461 *
1462 * An event might be pending because the previous #VMEXIT happened
1463 * during event delivery (i.e. ctrl->exitintinfo).
1464 *
1465 * An event might also be pending because an exception was injected
1466 * by the hypervisor (e.g. #PF during instruction emulation).
1467 */
1468 svm_inj_intinfo(sc, vcpu);
1469
1470 /* NMI event has priority over interrupts. */
1471 if (vm_nmi_pending(sc->vm, vcpu)) {
1472 if (nmi_blocked(sc, vcpu)) {
1473 /*
1474 * Can't inject another NMI if the guest has not
1475 * yet executed an "iret" after the last NMI.
1476 */
1477 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1478 "to NMI-blocking");
1479 } else if (ctrl->intr_shadow) {
1480 /*
1481 * Can't inject an NMI if the vcpu is in an intr_shadow.
1482 */
1483 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1484 "interrupt shadow");
1485 need_intr_window = 1;
1486 goto done;
1487 } else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1488 /*
1489 * If there is already an exception/interrupt pending
1490 * then defer the NMI until after that.
1491 */
1492 VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1493 "eventinj %#lx", ctrl->eventinj);
1494
1495 /*
1496 * Use self-IPI to trigger a VM-exit as soon as
1497 * possible after the event injection is completed.
1498 *
1499 * This works only if the external interrupt exiting
1500 * is at a lower priority than the event injection.
1501 *
1502 * Although not explicitly specified in APMv2 the
1503 * relative priorities were verified empirically.
1504 */
1505 ipi_cpu(curcpu, IPI_AST); /* XXX vmm_ipinum? */
1506 } else {
1507 vm_nmi_clear(sc->vm, vcpu);
1508
1509 /* Inject NMI, vector number is not used */
1510 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1511 IDT_NMI, 0, false);
1512
1513 /* virtual NMI blocking is now in effect */
1514 enable_nmi_blocking(sc, vcpu);
1515
1516 VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1517 }
1518 }
1519
1520 if (!vm_extint_pending(sc->vm, vcpu)) {
1521 /*
1522 * APIC interrupts are delivered using the V_IRQ offload.
1523 *
1524 * The primary benefit is that the hypervisor doesn't need to
1525 * deal with the various conditions that inhibit interrupts.
1526 * It also means that TPR changes via CR8 will be handled
1527 * without any hypervisor involvement.
1528 *
1529 * Note that the APIC vector must remain pending in the vIRR
1530 * until it is confirmed that it was delivered to the guest.
1531 * This can be confirmed based on the value of V_IRQ at the
1532 * next #VMEXIT (1 = pending, 0 = delivered).
1533 *
1534 * Also note that it is possible that another higher priority
1535 * vector can become pending before this vector is delivered
1536 * to the guest. This is alright because vcpu_notify_event()
1537 * will send an IPI and force the vcpu to trap back into the
1538 * hypervisor. The higher priority vector will be injected on
1539 * the next VMRUN.
1540 */
1541 if (vlapic_pending_intr(vlapic, &vector)) {
1542 KASSERT(vector >= 16 && vector <= 255,
1543 ("invalid vector %d from local APIC", vector));
1544 pending_apic_vector = vector;
1545 }
1546 goto done;
1547 }
1548
1549 /* Ask the legacy pic for a vector to inject */
1550 vatpic_pending_intr(sc->vm, &vector);
1551 KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR",
1552 vector));
1553
1554 /*
1555 * If the guest has disabled interrupts or is in an interrupt shadow
1556 * then we cannot inject the pending interrupt.
1557 */
1558 if ((state->rflags & PSL_I) == 0) {
1559 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1560 "rflags %#lx", vector, state->rflags);
1561 need_intr_window = 1;
1562 goto done;
1563 }
1564
1565 if (ctrl->intr_shadow) {
1566 VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1567 "interrupt shadow", vector);
1568 need_intr_window = 1;
1569 goto done;
1570 }
1571
1572 if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1573 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1574 "eventinj %#lx", vector, ctrl->eventinj);
1575 need_intr_window = 1;
1576 goto done;
1577 }
1578
1579 /*
1580 * Legacy PIC interrupts are delivered via the event injection
1581 * mechanism.
1582 */
1583 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1584
1585 vm_extint_clear(sc->vm, vcpu);
1586 vatpic_intr_accepted(sc->vm, vector);
1587
1588 /*
1589 * Force a VM-exit as soon as the vcpu is ready to accept another
1590 * interrupt. This is done because the PIC might have another vector
1591 * that it wants to inject. Also, if the APIC has a pending interrupt
1592 * that was preempted by the ExtInt then it allows us to inject the
1593 * APIC vector as soon as possible.
1594 */
1595 need_intr_window = 1;
1596done:
1597 /*
1598 * The guest can modify the TPR by writing to %CR8. In guest mode
1599 * the processor reflects this write to V_TPR without hypervisor
1600 * intervention.
1601 *
1602 * The guest can also modify the TPR by writing to it via the memory
1603 * mapped APIC page. In this case, the write will be emulated by the
1604 * hypervisor. For this reason V_TPR must be updated before every
1605 * VMRUN.
1606 */
1607 v_tpr = vlapic_get_cr8(vlapic);
1608 KASSERT(v_tpr >= 0 && v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1609 if (ctrl->v_tpr != v_tpr) {
1610 VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1611 ctrl->v_tpr, v_tpr);
1612 ctrl->v_tpr = v_tpr;
1613 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1614 }
1615
1616 if (pending_apic_vector) {
1617 /*
1618 * If an APIC vector is being injected then interrupt window
1619 * exiting is not possible on this VMRUN.
1620 */
1621 KASSERT(!need_intr_window, ("intr_window exiting impossible"));
1622 VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ",
1623 pending_apic_vector);
1624
1625 ctrl->v_irq = 1;
1626 ctrl->v_ign_tpr = 0;
1627 ctrl->v_intr_vector = pending_apic_vector;
1628 ctrl->v_intr_prio = pending_apic_vector >> 4;
1629 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1630 } else if (need_intr_window) {
1631 /*
1632 * We use V_IRQ in conjunction with the VINTR intercept to
1633 * trap into the hypervisor as soon as a virtual interrupt
1634 * can be delivered.
1635 *
1636 * Since injected events are not subject to intercept checks
1637 * we need to ensure that the V_IRQ is not actually going to
1638 * be delivered on VM entry. The KASSERT below enforces this.
1639 */
1640 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1641 (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1642 ("Bogus intr_window_exiting: eventinj (%#lx), "
1643 "intr_shadow (%u), rflags (%#lx)",
1644 ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1645 enable_intr_window_exiting(sc, vcpu);
1646 } else {
1647 disable_intr_window_exiting(sc, vcpu);
1648 }
1649}
1650
1651static __inline void
1652restore_host_tss(void)
1653{
1654 struct system_segment_descriptor *tss_sd;
1655
1656 /*
1657 * The TSS descriptor was in use prior to launching the guest so it
1658 * has been marked busy.
1659 *
1660 * 'ltr' requires the descriptor to be marked available so change the
1661 * type to "64-bit available TSS".
1662 */
1663 tss_sd = PCPU_GET(tss);
1664 tss_sd->sd_type = SDT_SYSTSS;
1665 ltr(GSEL(GPROC0_SEL, SEL_KPL));
1666}
1667
1668static void
1669check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu)
1670{
1671 struct svm_vcpu *vcpustate;
1672 struct vmcb_ctrl *ctrl;
1673 long eptgen;
1674 bool alloc_asid;
1675
1676 KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not "
1677 "active on cpu %u", __func__, thiscpu));
1678
1679 vcpustate = svm_get_vcpu(sc, vcpuid);
1680 ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1681
1682 /*
1683 * The TLB entries associated with the vcpu's ASID are not valid
1684 * if either of the following conditions is true:
1685 *
1686 * 1. The vcpu's ASID generation is different than the host cpu's
1687 * ASID generation. This happens when the vcpu migrates to a new
1688 * host cpu. It can also happen when the number of vcpus executing
1689 * on a host cpu is greater than the number of ASIDs available.
1690 *
1691 * 2. The pmap generation number is different than the value cached in
1692 * the 'vcpustate'. This happens when the host invalidates pages
1693 * belonging to the guest.
1694 *
1695 * asidgen eptgen Action
1696 * mismatch mismatch
1697 * 0 0 (a)
1698 * 0 1 (b1) or (b2)
1699 * 1 0 (c)
1700 * 1 1 (d)
1701 *
1702 * (a) There is no mismatch in eptgen or ASID generation and therefore
1703 * no further action is needed.
1704 *
1705 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1706 * retained and the TLB entries associated with this ASID
1707 * are flushed by VMRUN.
1708 *
1709 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1710 * allocated.
1711 *
1712 * (c) A new ASID is allocated.
1713 *
1714 * (d) A new ASID is allocated.
1715 */
1716
1717 alloc_asid = false;
1718 eptgen = pmap->pm_eptgen;
1719 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1720
1721 if (vcpustate->asid.gen != asid[thiscpu].gen) {
1722 alloc_asid = true; /* (c) and (d) */
1723 } else if (vcpustate->eptgen != eptgen) {
1724 if (flush_by_asid())
1725 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; /* (b1) */
1726 else
1727 alloc_asid = true; /* (b2) */
1728 } else {
1729 /*
1730 * This is the common case (a).
1731 */
1732 KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1733 KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1734 ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1735 }
1736
1737 if (alloc_asid) {
1738 if (++asid[thiscpu].num >= nasid) {
1739 asid[thiscpu].num = 1;
1740 if (++asid[thiscpu].gen == 0)
1741 asid[thiscpu].gen = 1;
1742 /*
1743 * If this cpu does not support "flush-by-asid"
1744 * then flush the entire TLB on a generation
1745 * bump. Subsequent ASID allocation in this
1746 * generation can be done without a TLB flush.
1747 */
1748 if (!flush_by_asid())
1749 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1750 }
1751 vcpustate->asid.gen = asid[thiscpu].gen;
1752 vcpustate->asid.num = asid[thiscpu].num;
1753
1754 ctrl->asid = vcpustate->asid.num;
1755 svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1756 /*
1757 * If this cpu supports "flush-by-asid" then the TLB
1758 * was not flushed after the generation bump. The TLB
1759 * is flushed selectively after every new ASID allocation.
1760 */
1761 if (flush_by_asid())
1762 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1763 }
1764 vcpustate->eptgen = eptgen;
1765
1766 KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1767 KASSERT(ctrl->asid == vcpustate->asid.num,
1768 ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1769}
1770
1771/*
1772 * Start vcpu with specified RIP.
1773 */
1774static int
1775svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap,
1776 void *rend_cookie, void *suspended_cookie)
1777{
1778 struct svm_regctx *hctx, *gctx;
1779 struct svm_softc *svm_sc;
1780 struct svm_vcpu *vcpustate;
1781 struct vmcb_state *state;
1782 struct vmcb_ctrl *ctrl;
1783 struct vm_exit *vmexit;
1784 struct vlapic *vlapic;
1785 struct vm *vm;
1786 uint64_t vmcb_pa;
1787 u_int thiscpu;
1788 int handled;
1789
1790 svm_sc = arg;
1791 vm = svm_sc->vm;
1792
1793 vcpustate = svm_get_vcpu(svm_sc, vcpu);
1794 state = svm_get_vmcb_state(svm_sc, vcpu);
1795 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
1796 vmexit = vm_exitinfo(vm, vcpu);
1797 vlapic = vm_lapic(vm, vcpu);
1798
1799 /*
1800 * Stash 'curcpu' on the stack as 'thiscpu'.
1801 *
1802 * The per-cpu data area is not accessible until MSR_GSBASE is restored
1803 * after the #VMEXIT. Since VMRUN is executed inside a critical section
1804 * 'curcpu' and 'thiscpu' are guaranteed to identical.
1805 */
1806 thiscpu = curcpu;
1807
1808 gctx = svm_get_guest_regctx(svm_sc, vcpu);
1809 hctx = &host_ctx[thiscpu];
1810 vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
1811
1812 if (vcpustate->lastcpu != thiscpu) {
1813 /*
1814 * Force new ASID allocation by invalidating the generation.
1815 */
1816 vcpustate->asid.gen = 0;
1817
1818 /*
1819 * Invalidate the VMCB state cache by marking all fields dirty.
1820 */
1821 svm_set_dirty(svm_sc, vcpu, 0xffffffff);
1822
1823 /*
1824 * XXX
1825 * Setting 'vcpustate->lastcpu' here is bit premature because
1826 * we may return from this function without actually executing
1827 * the VMRUN instruction. This could happen if a rendezvous
1828 * or an AST is pending on the first time through the loop.
1829 *
1830 * This works for now but any new side-effects of vcpu
1831 * migration should take this case into account.
1832 */
1833 vcpustate->lastcpu = thiscpu;
1834 vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
1835 }
1836
1837 svm_msr_guest_enter(svm_sc, vcpu);
1838
1839 /* Update Guest RIP */
1840 state->rip = rip;
1841
1842 do {
1843 /*
1844 * Disable global interrupts to guarantee atomicity during
1845 * loading of guest state. This includes not only the state
1846 * loaded by the "vmrun" instruction but also software state
1847 * maintained by the hypervisor: suspended and rendezvous
1848 * state, NPT generation number, vlapic interrupts etc.
1849 */
1850 disable_gintr();
1851
1852 if (vcpu_suspended(suspended_cookie)) {
1853 enable_gintr();
1854 vm_exit_suspended(vm, vcpu, state->rip);
1855 break;
1856 }
1857
1858 if (vcpu_rendezvous_pending(rend_cookie)) {
1859 enable_gintr();
1860 vm_exit_rendezvous(vm, vcpu, state->rip);
1861 break;
1862 }
1863
1864 /* We are asked to give the cpu by scheduler. */
1865 if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) {
1866 enable_gintr();
1867 vm_exit_astpending(vm, vcpu, state->rip);
1868 break;
1869 }
1870
1871 svm_inj_interrupts(svm_sc, vcpu, vlapic);
1872
1873 /* Activate the nested pmap on 'thiscpu' */
1874 CPU_SET_ATOMIC_ACQ(thiscpu, &pmap->pm_active);
1875
1876 /*
1877 * Check the pmap generation and the ASID generation to
1878 * ensure that the vcpu does not use stale TLB mappings.
1879 */
1880 check_asid(svm_sc, vcpu, pmap, thiscpu);
1881
1882 ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
1883 vcpustate->dirty = 0;
1884 VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
1885
1886 /* Launch Virtual Machine. */
1887 VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
1888 svm_launch(vmcb_pa, gctx, hctx);
1889
1890 CPU_CLR_ATOMIC(thiscpu, &pmap->pm_active);
1891
1892 /*
1893 * Restore MSR_GSBASE to point to the pcpu data area.
1894 *
1895 * Note that accesses done via PCPU_GET/PCPU_SET will work
1896 * only after MSR_GSBASE is restored.
1897 *
1898 * Also note that we don't bother restoring MSR_KGSBASE
1899 * since it is not used in the kernel and will be restored
1900 * when the VMRUN ioctl returns to userspace.
1901 */
1902 wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[thiscpu]);
1903 KASSERT(curcpu == thiscpu, ("thiscpu/curcpu (%u/%u) mismatch",
1904 thiscpu, curcpu));
1905
1906 /*
1907 * The host GDTR and IDTR is saved by VMRUN and restored
1908 * automatically on #VMEXIT. However, the host TSS needs
1909 * to be restored explicitly.
1910 */
1911 restore_host_tss();
1912
1913 /* #VMEXIT disables interrupts so re-enable them here. */
1914 enable_gintr();
1915
1916 /* Handle #VMEXIT and if required return to user space. */
1917 handled = svm_vmexit(svm_sc, vcpu, vmexit);
1918 } while (handled);
1919
1920 svm_msr_guest_exit(svm_sc, vcpu);
1921
1922 return (0);
1923}
1924
1925/*
1926 * Cleanup for virtual machine.
1927 */
1928static void
1929svm_vmcleanup(void *arg)
1930{
1931 struct svm_softc *svm_sc;
1932
1933 svm_sc = arg;
1934
1935 VCPU_CTR0(svm_sc->vm, 0, "SVM:cleanup\n");
1936
1937 free(svm_sc, M_SVM);
1938}
1939
1940/*
1941 * Return pointer to hypervisor saved register state.
1942 */
1943static register_t *
1944swctx_regptr(struct svm_regctx *regctx, int reg)
1945{
1946
1947 switch (reg) {
1948 case VM_REG_GUEST_RBX:
1949 return (®ctx->sctx_rbx);
1950 case VM_REG_GUEST_RCX:
1951 return (®ctx->sctx_rcx);
1952 case VM_REG_GUEST_RDX:
1953 return (®ctx->e.g.sctx_rdx);
1954 case VM_REG_GUEST_RDI:
1955 return (®ctx->e.g.sctx_rdi);
1956 case VM_REG_GUEST_RSI:
1957 return (®ctx->e.g.sctx_rsi);
1958 case VM_REG_GUEST_RBP:
1959 return (®ctx->sctx_rbp);
1960 case VM_REG_GUEST_R8:
1961 return (®ctx->sctx_r8);
1962 case VM_REG_GUEST_R9:
1963 return (®ctx->sctx_r9);
1964 case VM_REG_GUEST_R10:
1965 return (®ctx->sctx_r10);
1966 case VM_REG_GUEST_R11:
1967 return (®ctx->sctx_r11);
1968 case VM_REG_GUEST_R12:
1969 return (®ctx->sctx_r12);
1970 case VM_REG_GUEST_R13:
1971 return (®ctx->sctx_r13);
1972 case VM_REG_GUEST_R14:
1973 return (®ctx->sctx_r14);
1974 case VM_REG_GUEST_R15:
1975 return (®ctx->sctx_r15);
1976 default:
1977 ERR("Unknown register requested, reg=%d.\n", reg);
1978 break;
1979 }
1980
1981 return (NULL);
1982}
1983
1984/*
1985 * Interface to read guest registers.
1986 * This can be SVM h/w saved or hypervisor saved register.
1987 */
1988static int
1989svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
1990{
1991 struct svm_softc *svm_sc;
1992 register_t *reg;
1993
1994 svm_sc = arg;
1995
1996 if (ident == VM_REG_GUEST_INTR_SHADOW) {
1997 return (svm_get_intr_shadow(svm_sc, vcpu, val));
1998 }
1999
2000 if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2001 return (0);
2002 }
2003
2004 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2005
2006 if (reg != NULL) {
2007 *val = *reg;
2008 return (0);
2009 }
2010
2011 ERR("SVM_ERR:reg type %x is not saved in VMCB.\n", ident);
2012 return (EINVAL);
2013}
2014
2015/*
2016 * Interface to write to guest registers.
2017 * This can be SVM h/w saved or hypervisor saved register.
2018 */
2019static int
2020svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2021{
2022 struct svm_softc *svm_sc;
2023 register_t *reg;
2024
2025 svm_sc = arg;
2026
2027 if (ident == VM_REG_GUEST_INTR_SHADOW) {
2028 return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2029 }
2030
2031 if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2032 return (0);
2033 }
2034
2035 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2036
2037 if (reg != NULL) {
2038 *reg = val;
2039 return (0);
2040 }
2041
2042 /*
2043 * XXX deal with CR3 and invalidate TLB entries tagged with the
2044 * vcpu's ASID. This needs to be treated differently depending on
2045 * whether 'running' is true/false.
2046 */
2047
2048 ERR("SVM_ERR:reg type %x is not saved in VMCB.\n", ident);
2049 return (EINVAL);
2050}
2051
2052static int
2053svm_setcap(void *arg, int vcpu, int type, int val)
2054{
2055 struct svm_softc *sc;
2056 int error;
2057
2058 sc = arg;
2059 error = 0;
2060 switch (type) {
2061 case VM_CAP_HALT_EXIT:
2062 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2063 VMCB_INTCPT_HLT, val);
2064 break;
2065 case VM_CAP_PAUSE_EXIT:
2066 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2067 VMCB_INTCPT_PAUSE, val);
2068 break;
2069 case VM_CAP_UNRESTRICTED_GUEST:
2070 /* Unrestricted guest execution cannot be disabled in SVM */
2071 if (val == 0)
2072 error = EINVAL;
2073 break;
2074 default:
2075 error = ENOENT;
2076 break;
2077 }
2078 return (error);
2079}
2080
2081static int
2082svm_getcap(void *arg, int vcpu, int type, int *retval)
2083{
2084 struct svm_softc *sc;
2085 int error;
2086
2087 sc = arg;
2088 error = 0;
2089
2090 switch (type) {
2091 case VM_CAP_HALT_EXIT:
2092 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2093 VMCB_INTCPT_HLT);
2094 break;
2095 case VM_CAP_PAUSE_EXIT:
2096 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2097 VMCB_INTCPT_PAUSE);
2098 break;
2099 case VM_CAP_UNRESTRICTED_GUEST:
2100 *retval = 1; /* unrestricted guest is always enabled */
2101 break;
2102 default:
2103 error = ENOENT;
2104 break;
2105 }
2106 return (error);
2107}
2108
2109static struct vlapic *
2110svm_vlapic_init(void *arg, int vcpuid)
2111{
2112 struct svm_softc *svm_sc;
2113 struct vlapic *vlapic;
2114
2115 svm_sc = arg;
2116 vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2117 vlapic->vm = svm_sc->vm;
2118 vlapic->vcpuid = vcpuid;
2119 vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2120
2121 vlapic_init(vlapic);
2122
2123 return (vlapic);
2124}
2125
2126static void
2127svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2128{
2129
2130 vlapic_cleanup(vlapic);
2131 free(vlapic, M_SVM_VLAPIC);
2132}
2133
2134struct vmm_ops vmm_ops_amd = {
2135 svm_init,
2136 svm_cleanup,
2137 svm_restore,
2138 svm_vminit,
2139 svm_vmrun,
2140 svm_vmcleanup,
2141 svm_getreg,
2142 svm_setreg,
2143 vmcb_getdesc,
2144 vmcb_setdesc,
2145 svm_getcap,
2146 svm_setcap,
2147 svm_npt_alloc,
2148 svm_npt_free,
2149 svm_vlapic_init,
2150 svm_vlapic_cleanup
2151};
2 * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com)
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice unmodified, this list of conditions, and the following
10 * disclaimer.
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.
14 *
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.
25 */
26
27#include <sys/cdefs.h>
28__FBSDID("$FreeBSD: projects/bhyve_svm/sys/amd64/vmm/amd/svm.c 271939 2014-09-21 23:42:54Z neel $");
29
30#include <sys/param.h>
31#include <sys/systm.h>
32#include <sys/smp.h>
33#include <sys/kernel.h>
34#include <sys/malloc.h>
35#include <sys/pcpu.h>
36#include <sys/proc.h>
37#include <sys/sysctl.h>
38
39#include <vm/vm.h>
40#include <vm/pmap.h>
41
42#include <machine/cpufunc.h>
43#include <machine/psl.h>
44#include <machine/pmap.h>
45#include <machine/md_var.h>
46#include <machine/vmparam.h>
47#include <machine/specialreg.h>
48#include <machine/segments.h>
49#include <machine/smp.h>
50#include <machine/vmm.h>
51#include <machine/vmm_dev.h>
52#include <machine/vmm_instruction_emul.h>
53
54#include <x86/apicreg.h>
55
56#include "vmm_lapic.h"
57#include "vmm_stat.h"
58#include "vmm_ktr.h"
59#include "vmm_ioport.h"
60#include "vatpic.h"
61#include "vlapic.h"
62#include "vlapic_priv.h"
63
64#include "x86.h"
65#include "vmcb.h"
66#include "svm.h"
67#include "svm_softc.h"
68#include "svm_msr.h"
69#include "npt.h"
70
71SYSCTL_DECL(_hw_vmm);
72SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL);
73
74/*
75 * SVM CPUID function 0x8000_000A, edx bit decoding.
76 */
77#define AMD_CPUID_SVM_NP BIT(0) /* Nested paging or RVI */
78#define AMD_CPUID_SVM_LBR BIT(1) /* Last branch virtualization */
79#define AMD_CPUID_SVM_SVML BIT(2) /* SVM lock */
80#define AMD_CPUID_SVM_NRIP_SAVE BIT(3) /* Next RIP is saved */
81#define AMD_CPUID_SVM_TSC_RATE BIT(4) /* TSC rate control. */
82#define AMD_CPUID_SVM_VMCB_CLEAN BIT(5) /* VMCB state caching */
83#define AMD_CPUID_SVM_FLUSH_BY_ASID BIT(6) /* Flush by ASID */
84#define AMD_CPUID_SVM_DECODE_ASSIST BIT(7) /* Decode assist */
85#define AMD_CPUID_SVM_PAUSE_INC BIT(10) /* Pause intercept filter. */
86#define AMD_CPUID_SVM_PAUSE_FTH BIT(12) /* Pause filter threshold */
87
88#define VMCB_CACHE_DEFAULT (VMCB_CACHE_ASID | \
89 VMCB_CACHE_IOPM | \
90 VMCB_CACHE_I | \
91 VMCB_CACHE_TPR | \
92 VMCB_CACHE_CR2 | \
93 VMCB_CACHE_CR | \
94 VMCB_CACHE_DT | \
95 VMCB_CACHE_SEG | \
96 VMCB_CACHE_NP)
97
98static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT;
99SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean,
100 0, NULL);
101
102MALLOC_DEFINE(M_SVM, "svm", "svm");
103MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic");
104
105/* Per-CPU context area. */
106extern struct pcpu __pcpu[];
107
108static uint32_t svm_feature; /* AMD SVM features. */
109SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RD, &svm_feature, 0,
110 "SVM features advertised by CPUID.8000000AH:EDX");
111
112static int disable_npf_assist;
113SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN,
114 &disable_npf_assist, 0, NULL);
115
116/* Maximum ASIDs supported by the processor */
117static uint32_t nasid;
118SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RD, &nasid, 0,
119 "Number of ASIDs supported by this processor");
120
121/* Current ASID generation for each host cpu */
122static struct asid asid[MAXCPU];
123
124/*
125 * SVM host state saved area of size 4KB for each core.
126 */
127static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE);
128
129/*
130 * S/w saved host context.
131 */
132static struct svm_regctx host_ctx[MAXCPU];
133
134static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery");
135static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry");
136static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window");
137
138static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val);
139
140/*
141 * Common function to enable or disabled SVM for a CPU.
142 */
143static int
144cpu_svm_enable_disable(boolean_t enable)
145{
146 uint64_t efer_msr;
147
148 efer_msr = rdmsr(MSR_EFER);
149
150 if (enable)
151 efer_msr |= EFER_SVM;
152 else
153 efer_msr &= ~EFER_SVM;
154
155 wrmsr(MSR_EFER, efer_msr);
156
157 return(0);
158}
159
160/*
161 * Disable SVM on a CPU.
162 */
163static void
164svm_disable(void *arg __unused)
165{
166
167 (void)cpu_svm_enable_disable(FALSE);
168}
169
170/*
171 * Disable SVM for all CPUs.
172 */
173static int
174svm_cleanup(void)
175{
176
177 smp_rendezvous(NULL, svm_disable, NULL, NULL);
178 return (0);
179}
180
181/*
182 * Check for required BHyVe SVM features in a CPU.
183 */
184static int
185svm_cpuid_features(void)
186{
187 u_int regs[4];
188
189 /* CPUID Fn8000_000A is for SVM */
190 do_cpuid(0x8000000A, regs);
191 svm_feature = regs[3];
192
193 printf("SVM rev: 0x%x NASID:0x%x\n", regs[0] & 0xFF, regs[1]);
194 nasid = regs[1];
195 KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid));
196
197 printf("SVM Features:0x%b\n", svm_feature,
198 "\020"
199 "\001NP" /* Nested paging */
200 "\002LbrVirt" /* LBR virtualization */
201 "\003SVML" /* SVM lock */
202 "\004NRIPS" /* NRIP save */
203 "\005TscRateMsr" /* MSR based TSC rate control */
204 "\006VmcbClean" /* VMCB clean bits */
205 "\007FlushByAsid" /* Flush by ASID */
206 "\010DecodeAssist" /* Decode assist */
207 "\011<b20>"
208 "\012<b20>"
209 "\013PauseFilter"
210 "\014<b20>"
211 "\015PauseFilterThreshold"
212 "\016AVIC"
213 );
214
215 /* SVM Lock */
216 if (!(svm_feature & AMD_CPUID_SVM_SVML)) {
217 printf("SVM is disabled by BIOS, please enable in BIOS.\n");
218 return (ENXIO);
219 }
220
221 /*
222 * bhyve need RVI to work.
223 */
224 if (!(svm_feature & AMD_CPUID_SVM_NP)) {
225 printf("Missing Nested paging or RVI SVM support in processor.\n");
226 return (EIO);
227 }
228
229 if (svm_feature & AMD_CPUID_SVM_NRIP_SAVE)
230 return (0);
231
232 return (EIO);
233}
234
235static __inline int
236flush_by_asid(void)
237{
238
239 return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID);
240}
241
242static __inline int
243decode_assist(void)
244{
245
246 return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST);
247}
248
249/*
250 * Enable SVM for a CPU.
251 */
252static void
253svm_enable(void *arg __unused)
254{
255 uint64_t hsave_pa;
256
257 (void)cpu_svm_enable_disable(TRUE);
258
259 hsave_pa = vtophys(hsave[curcpu]);
260 wrmsr(MSR_VM_HSAVE_PA, hsave_pa);
261
262 if (rdmsr(MSR_VM_HSAVE_PA) != hsave_pa) {
263 panic("VM_HSAVE_PA is wrong on CPU%d\n", curcpu);
264 }
265}
266
267/*
268 * Check if a processor support SVM.
269 */
270static int
271is_svm_enabled(void)
272{
273 uint64_t msr;
274
275 /* Section 15.4 Enabling SVM from APM2. */
276 if ((amd_feature2 & AMDID2_SVM) == 0) {
277 printf("SVM is not supported on this processor.\n");
278 return (ENXIO);
279 }
280
281 msr = rdmsr(MSR_VM_CR);
282 /* Make sure SVM is not disabled by BIOS. */
283 if ((msr & VM_CR_SVMDIS) == 0) {
284 return svm_cpuid_features();
285 }
286
287 printf("SVM disabled by Key, consult TPM/BIOS manual.\n");
288 return (ENXIO);
289}
290
291/*
292 * Enable SVM on CPU and initialize nested page table h/w.
293 */
294static int
295svm_init(int ipinum)
296{
297 int err, cpu;
298
299 err = is_svm_enabled();
300 if (err)
301 return (err);
302
303 vmcb_clean &= VMCB_CACHE_DEFAULT;
304
305 for (cpu = 0; cpu < MAXCPU; cpu++) {
306 /*
307 * Initialize the host ASIDs to their "highest" valid values.
308 *
309 * The next ASID allocation will rollover both 'gen' and 'num'
310 * and start off the sequence at {1,1}.
311 */
312 asid[cpu].gen = ~0UL;
313 asid[cpu].num = nasid - 1;
314 }
315
316 svm_msr_init();
317 svm_npt_init(ipinum);
318
319 /* Start SVM on all CPUs */
320 smp_rendezvous(NULL, svm_enable, NULL, NULL);
321
322 return (0);
323}
324
325static void
326svm_restore(void)
327{
328 svm_enable(NULL);
329}
330
331/*
332 * Get index and bit position for a MSR in MSR permission
333 * bitmap. Two bits are used for each MSR, lower bit is
334 * for read and higher bit is for write.
335 */
336static int
337svm_msr_index(uint64_t msr, int *index, int *bit)
338{
339 uint32_t base, off;
340
341/* Pentium compatible MSRs */
342#define MSR_PENTIUM_START 0
343#define MSR_PENTIUM_END 0x1FFF
344/* AMD 6th generation and Intel compatible MSRs */
345#define MSR_AMD6TH_START 0xC0000000UL
346#define MSR_AMD6TH_END 0xC0001FFFUL
347/* AMD 7th and 8th generation compatible MSRs */
348#define MSR_AMD7TH_START 0xC0010000UL
349#define MSR_AMD7TH_END 0xC0011FFFUL
350
351 *index = -1;
352 *bit = (msr % 4) * 2;
353 base = 0;
354
355 if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) {
356 *index = msr / 4;
357 return (0);
358 }
359
360 base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1);
361 if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) {
362 off = (msr - MSR_AMD6TH_START);
363 *index = (off + base) / 4;
364 return (0);
365 }
366
367 base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1);
368 if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) {
369 off = (msr - MSR_AMD7TH_START);
370 *index = (off + base) / 4;
371 return (0);
372 }
373
374 return (EIO);
375}
376
377/*
378 * Give virtual cpu the complete access to MSR(read & write).
379 */
380static int
381svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write)
382{
383 int index, bit, err;
384
385 err = svm_msr_index(msr, &index, &bit);
386 if (err) {
387 ERR("MSR 0x%lx is not writeable by guest.\n", msr);
388 return (err);
389 }
390
391 if (index < 0 || index > (SVM_MSR_BITMAP_SIZE)) {
392 ERR("MSR 0x%lx index out of range(%d).\n", msr, index);
393 return (EINVAL);
394 }
395 if (bit < 0 || bit > 8) {
396 ERR("MSR 0x%lx bit out of range(%d).\n", msr, bit);
397 return (EINVAL);
398 }
399
400 /* Disable intercept for read and write. */
401 if (read)
402 perm_bitmap[index] &= ~(1UL << bit);
403 if (write)
404 perm_bitmap[index] &= ~(2UL << bit);
405 CTR2(KTR_VMM, "Guest has control:0x%x on SVM:MSR(0x%lx).\n",
406 (perm_bitmap[index] >> bit) & 0x3, msr);
407
408 return (0);
409}
410
411static int
412svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr)
413{
414 return svm_msr_perm(perm_bitmap, msr, true, true);
415}
416
417static int
418svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr)
419{
420 return svm_msr_perm(perm_bitmap, msr, true, false);
421}
422
423static __inline int
424svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask)
425{
426 struct vmcb_ctrl *ctrl;
427
428 KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
429
430 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
431 return (ctrl->intercept[idx] & bitmask ? 1 : 0);
432}
433
434static __inline void
435svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask,
436 int enabled)
437{
438 struct vmcb_ctrl *ctrl;
439 uint32_t oldval;
440
441 KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
442
443 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
444 oldval = ctrl->intercept[idx];
445
446 if (enabled)
447 ctrl->intercept[idx] |= bitmask;
448 else
449 ctrl->intercept[idx] &= ~bitmask;
450
451 if (ctrl->intercept[idx] != oldval) {
452 svm_set_dirty(sc, vcpu, VMCB_CACHE_I);
453 VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified "
454 "from %#x to %#x", idx, oldval, ctrl->intercept[idx]);
455 }
456}
457
458static __inline void
459svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
460{
461 svm_set_intercept(sc, vcpu, off, bitmask, 0);
462}
463
464static __inline void
465svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
466{
467 svm_set_intercept(sc, vcpu, off, bitmask, 1);
468}
469
470static void
471vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa,
472 uint64_t msrpm_base_pa, uint64_t np_pml4)
473{
474 struct vmcb_ctrl *ctrl;
475 struct vmcb_state *state;
476 uint32_t mask;
477 int n;
478
479 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
480 state = svm_get_vmcb_state(sc, vcpu);
481
482 ctrl->iopm_base_pa = iopm_base_pa;
483 ctrl->msrpm_base_pa = msrpm_base_pa;
484
485 /* Enable nested paging */
486 ctrl->np_enable = 1;
487 ctrl->n_cr3 = np_pml4;
488
489 /*
490 * Intercept accesses to the control registers that are not shadowed
491 * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8.
492 */
493 for (n = 0; n < 16; n++) {
494 mask = (BIT(n) << 16) | BIT(n);
495 if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8)
496 svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
497 else
498 svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
499 }
500
501 /* Intercept Machine Check exceptions. */
502 svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC));
503
504 /* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */
505 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO);
506 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR);
507 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID);
508 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR);
509 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT);
510 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI);
511 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI);
512 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN);
513 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
514 VMCB_INTCPT_FERR_FREEZE);
515
516 /*
517 * From section "Canonicalization and Consistency Checks" in APMv2
518 * the VMRUN intercept bit must be set to pass the consistency check.
519 */
520 svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
521
522 /*
523 * The ASID will be set to a non-zero value just before VMRUN.
524 */
525 ctrl->asid = 0;
526
527 /*
528 * Section 15.21.1, Interrupt Masking in EFLAGS
529 * Section 15.21.2, Virtualizing APIC.TPR
530 *
531 * This must be set for %rflag and %cr8 isolation of guest and host.
532 */
533 ctrl->v_intr_masking = 1;
534
535 /* Enable Last Branch Record aka LBR for debugging */
536 ctrl->lbr_virt_en = 1;
537 state->dbgctl = BIT(0);
538
539 /* EFER_SVM must always be set when the guest is executing */
540 state->efer = EFER_SVM;
541
542 /* Set up the PAT to power-on state */
543 state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK) |
544 PAT_VALUE(1, PAT_WRITE_THROUGH) |
545 PAT_VALUE(2, PAT_UNCACHED) |
546 PAT_VALUE(3, PAT_UNCACHEABLE) |
547 PAT_VALUE(4, PAT_WRITE_BACK) |
548 PAT_VALUE(5, PAT_WRITE_THROUGH) |
549 PAT_VALUE(6, PAT_UNCACHED) |
550 PAT_VALUE(7, PAT_UNCACHEABLE);
551}
552
553/*
554 * Initialise a virtual machine.
555 */
556static void *
557svm_vminit(struct vm *vm, pmap_t pmap)
558{
559 struct svm_softc *svm_sc;
560 struct svm_vcpu *vcpu;
561 vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
562 int i;
563
564 svm_sc = (struct svm_softc *)malloc(sizeof (struct svm_softc),
565 M_SVM, M_WAITOK | M_ZERO);
566
567 svm_sc->vm = vm;
568 svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4);
569
570 /*
571 * Intercept MSR access to all MSRs except GSBASE, FSBASE,... etc.
572 */
573 memset(svm_sc->msr_bitmap, 0xFF, sizeof(svm_sc->msr_bitmap));
574
575 /*
576 * Following MSR can be completely controlled by virtual machines
577 * since access to following are translated to access to VMCB.
578 */
579 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
580 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
581 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
582
583 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
584 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
585 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
586 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
587 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
588 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
589 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
590
591 /* For Nested Paging/RVI only. */
592 svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
593
594 svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
595
596 /*
597 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
598 */
599 svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
600
601 /* Intercept access to all I/O ports. */
602 memset(svm_sc->iopm_bitmap, 0xFF, sizeof(svm_sc->iopm_bitmap));
603
604 /* Cache physical address for multiple vcpus. */
605 iopm_pa = vtophys(svm_sc->iopm_bitmap);
606 msrpm_pa = vtophys(svm_sc->msr_bitmap);
607 pml4_pa = svm_sc->nptp;
608
609 for (i = 0; i < VM_MAXCPU; i++) {
610 vcpu = svm_get_vcpu(svm_sc, i);
611 vcpu->lastcpu = NOCPU;
612 vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
613 vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
614 svm_msr_guest_init(svm_sc, i);
615 }
616 return (svm_sc);
617}
618
619static int
620svm_cpl(struct vmcb_state *state)
621{
622
623 /*
624 * From APMv2:
625 * "Retrieve the CPL from the CPL field in the VMCB, not
626 * from any segment DPL"
627 */
628 return (state->cpl);
629}
630
631static enum vm_cpu_mode
632svm_vcpu_mode(struct vmcb *vmcb)
633{
634 struct vmcb_segment seg;
635 struct vmcb_state *state;
636 int error;
637
638 state = &vmcb->state;
639
640 if (state->efer & EFER_LMA) {
641 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
642 KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
643 error));
644
645 /*
646 * Section 4.8.1 for APM2, check if Code Segment has
647 * Long attribute set in descriptor.
648 */
649 if (seg.attrib & VMCB_CS_ATTRIB_L)
650 return (CPU_MODE_64BIT);
651 else
652 return (CPU_MODE_COMPATIBILITY);
653 } else if (state->cr0 & CR0_PE) {
654 return (CPU_MODE_PROTECTED);
655 } else {
656 return (CPU_MODE_REAL);
657 }
658}
659
660static enum vm_paging_mode
661svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
662{
663
664 if ((cr0 & CR0_PG) == 0)
665 return (PAGING_MODE_FLAT);
666 if ((cr4 & CR4_PAE) == 0)
667 return (PAGING_MODE_32);
668 if (efer & EFER_LME)
669 return (PAGING_MODE_64);
670 else
671 return (PAGING_MODE_PAE);
672}
673
674/*
675 * ins/outs utility routines
676 */
677static uint64_t
678svm_inout_str_index(struct svm_regctx *regs, int in)
679{
680 uint64_t val;
681
682 val = in ? regs->e.g.sctx_rdi : regs->e.g.sctx_rsi;
683
684 return (val);
685}
686
687static uint64_t
688svm_inout_str_count(struct svm_regctx *regs, int rep)
689{
690 uint64_t val;
691
692 val = rep ? regs->sctx_rcx : 1;
693
694 return (val);
695}
696
697static void
698svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
699 int in, struct vm_inout_str *vis)
700{
701 int error, s;
702
703 if (in) {
704 vis->seg_name = VM_REG_GUEST_ES;
705 } else {
706 /* The segment field has standard encoding */
707 s = (info1 >> 10) & 0x7;
708 vis->seg_name = vm_segment_name(s);
709 }
710
711 error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
712 KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
713}
714
715static int
716svm_inout_str_addrsize(uint64_t info1)
717{
718 uint32_t size;
719
720 size = (info1 >> 7) & 0x7;
721 switch (size) {
722 case 1:
723 return (2); /* 16 bit */
724 case 2:
725 return (4); /* 32 bit */
726 case 4:
727 return (8); /* 64 bit */
728 default:
729 panic("%s: invalid size encoding %d", __func__, size);
730 }
731}
732
733static void
734svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
735{
736 struct vmcb_state *state;
737
738 state = &vmcb->state;
739 paging->cr3 = state->cr3;
740 paging->cpl = svm_cpl(state);
741 paging->cpu_mode = svm_vcpu_mode(vmcb);
742 paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
743 state->efer);
744}
745
746#define UNHANDLED 0
747
748/*
749 * Handle guest I/O intercept.
750 */
751static int
752svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
753{
754 struct vmcb_ctrl *ctrl;
755 struct vmcb_state *state;
756 struct svm_regctx *regs;
757 struct vm_inout_str *vis;
758 uint64_t info1;
759 int inout_string;
760
761 state = svm_get_vmcb_state(svm_sc, vcpu);
762 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
763 regs = svm_get_guest_regctx(svm_sc, vcpu);
764
765 info1 = ctrl->exitinfo1;
766 inout_string = info1 & BIT(2) ? 1 : 0;
767
768 /*
769 * The effective segment number in EXITINFO1[12:10] is populated
770 * only if the processor has the DecodeAssist capability.
771 *
772 * XXX this is not specified explicitly in APMv2 but can be verified
773 * empirically.
774 */
775 if (inout_string && !decode_assist())
776 return (UNHANDLED);
777
778 vmexit->exitcode = VM_EXITCODE_INOUT;
779 vmexit->u.inout.in = (info1 & BIT(0)) ? 1 : 0;
780 vmexit->u.inout.string = inout_string;
781 vmexit->u.inout.rep = (info1 & BIT(3)) ? 1 : 0;
782 vmexit->u.inout.bytes = (info1 >> 4) & 0x7;
783 vmexit->u.inout.port = (uint16_t)(info1 >> 16);
784 vmexit->u.inout.eax = (uint32_t)(state->rax);
785
786 if (inout_string) {
787 vmexit->exitcode = VM_EXITCODE_INOUT_STR;
788 vis = &vmexit->u.inout_str;
789 svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
790 vis->rflags = state->rflags;
791 vis->cr0 = state->cr0;
792 vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
793 vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
794 vis->addrsize = svm_inout_str_addrsize(info1);
795 svm_inout_str_seginfo(svm_sc, vcpu, info1,
796 vmexit->u.inout.in, vis);
797 }
798
799 return (UNHANDLED);
800}
801
802static int
803svm_npf_paging(uint64_t exitinfo1)
804{
805
806 if (exitinfo1 & VMCB_NPF_INFO1_W)
807 return (VM_PROT_WRITE);
808
809 return (VM_PROT_READ);
810}
811
812static bool
813svm_npf_emul_fault(uint64_t exitinfo1)
814{
815
816 if (exitinfo1 & VMCB_NPF_INFO1_ID) {
817 return (false);
818 }
819
820 if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
821 return (false);
822 }
823
824 if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
825 return (false);
826 }
827
828 return (true);
829}
830
831static void
832svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
833{
834 struct vm_guest_paging *paging;
835 struct vmcb_segment seg;
836 struct vmcb_ctrl *ctrl;
837 char *inst_bytes;
838 int error, inst_len;
839
840 ctrl = &vmcb->ctrl;
841 paging = &vmexit->u.inst_emul.paging;
842
843 vmexit->exitcode = VM_EXITCODE_INST_EMUL;
844 vmexit->u.inst_emul.gpa = gpa;
845 vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
846 svm_paging_info(vmcb, paging);
847
848 error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
849 KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
850
851 switch(paging->cpu_mode) {
852 case CPU_MODE_PROTECTED:
853 case CPU_MODE_COMPATIBILITY:
854 /*
855 * Section 4.8.1 of APM2, Default Operand Size or D bit.
856 */
857 vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
858 1 : 0;
859 break;
860 default:
861 vmexit->u.inst_emul.cs_d = 0;
862 break;
863 }
864
865 /*
866 * Copy the instruction bytes into 'vie' if available.
867 */
868 if (decode_assist() && !disable_npf_assist) {
869 inst_len = ctrl->inst_len;
870 inst_bytes = ctrl->inst_bytes;
871 } else {
872 inst_len = 0;
873 inst_bytes = NULL;
874 }
875 vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
876}
877
878#ifdef KTR
879static const char *
880intrtype_to_str(int intr_type)
881{
882 switch (intr_type) {
883 case VMCB_EVENTINJ_TYPE_INTR:
884 return ("hwintr");
885 case VMCB_EVENTINJ_TYPE_NMI:
886 return ("nmi");
887 case VMCB_EVENTINJ_TYPE_INTn:
888 return ("swintr");
889 case VMCB_EVENTINJ_TYPE_EXCEPTION:
890 return ("exception");
891 default:
892 panic("%s: unknown intr_type %d", __func__, intr_type);
893 }
894}
895#endif
896
897/*
898 * Inject an event to vcpu as described in section 15.20, "Event injection".
899 */
900static void
901svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
902 uint32_t error, bool ec_valid)
903{
904 struct vmcb_ctrl *ctrl;
905
906 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
907
908 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
909 ("%s: event already pending %#lx", __func__, ctrl->eventinj));
910
911 KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
912 __func__, vector));
913
914 switch (intr_type) {
915 case VMCB_EVENTINJ_TYPE_INTR:
916 case VMCB_EVENTINJ_TYPE_NMI:
917 case VMCB_EVENTINJ_TYPE_INTn:
918 break;
919 case VMCB_EVENTINJ_TYPE_EXCEPTION:
920 if (vector >= 0 && vector <= 31 && vector != 2)
921 break;
922 /* FALLTHROUGH */
923 default:
924 panic("%s: invalid intr_type/vector: %d/%d", __func__,
925 intr_type, vector);
926 }
927 ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
928 if (ec_valid) {
929 ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
930 ctrl->eventinj |= (uint64_t)error << 32;
931 VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
932 intrtype_to_str(intr_type), vector, error);
933 } else {
934 VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
935 intrtype_to_str(intr_type), vector);
936 }
937}
938
939static void
940svm_update_virqinfo(struct svm_softc *sc, int vcpu)
941{
942 struct vm *vm;
943 struct vlapic *vlapic;
944 struct vmcb_ctrl *ctrl;
945 int pending;
946
947 vm = sc->vm;
948 vlapic = vm_lapic(vm, vcpu);
949 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
950
951 /* Update %cr8 in the emulated vlapic */
952 vlapic_set_cr8(vlapic, ctrl->v_tpr);
953
954 /*
955 * If V_IRQ indicates that the interrupt injection attempted on then
956 * last VMRUN was successful then update the vlapic accordingly.
957 */
958 if (ctrl->v_intr_vector != 0) {
959 pending = ctrl->v_irq;
960 KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid "
961 "v_intr_vector %d", __func__, ctrl->v_intr_vector));
962 KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
963 VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector,
964 pending ? "pending" : "accepted");
965 if (!pending)
966 vlapic_intr_accepted(vlapic, ctrl->v_intr_vector);
967 }
968}
969
970static void
971svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
972{
973 struct vmcb_ctrl *ctrl;
974 uint64_t intinfo;
975
976 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
977 intinfo = ctrl->exitintinfo;
978 if (!VMCB_EXITINTINFO_VALID(intinfo))
979 return;
980
981 /*
982 * From APMv2, Section "Intercepts during IDT interrupt delivery"
983 *
984 * If a #VMEXIT happened during event delivery then record the event
985 * that was being delivered.
986 */
987 VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
988 intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
989 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
990 vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
991}
992
993static __inline int
994vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
995{
996
997 return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
998 VMCB_INTCPT_VINTR));
999}
1000
1001static __inline void
1002enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1003{
1004 struct vmcb_ctrl *ctrl;
1005
1006 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1007
1008 if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
1009 KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
1010 KASSERT(vintr_intercept_enabled(sc, vcpu),
1011 ("%s: vintr intercept should be enabled", __func__));
1012 return;
1013 }
1014
1015 VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
1016 ctrl->v_irq = 1;
1017 ctrl->v_ign_tpr = 1;
1018 ctrl->v_intr_vector = 0;
1019 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1020 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1021}
1022
1023static __inline void
1024disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1025{
1026 struct vmcb_ctrl *ctrl;
1027
1028 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1029
1030 if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1031 KASSERT(!vintr_intercept_enabled(sc, vcpu),
1032 ("%s: vintr intercept should be disabled", __func__));
1033 return;
1034 }
1035
1036#ifdef KTR
1037 if (ctrl->v_intr_vector == 0)
1038 VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
1039 else
1040 VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection");
1041#endif
1042 ctrl->v_irq = 0;
1043 ctrl->v_intr_vector = 0;
1044 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1045 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1046}
1047
1048static int
1049svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
1050{
1051 struct vmcb_ctrl *ctrl;
1052 int oldval, newval;
1053
1054 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1055 oldval = ctrl->intr_shadow;
1056 newval = val ? 1 : 0;
1057 if (newval != oldval) {
1058 ctrl->intr_shadow = newval;
1059 VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1060 }
1061 return (0);
1062}
1063
1064static int
1065svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1066{
1067 struct vmcb_ctrl *ctrl;
1068
1069 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1070 *val = ctrl->intr_shadow;
1071 return (0);
1072}
1073
1074/*
1075 * Once an NMI is injected it blocks delivery of further NMIs until the handler
1076 * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1077 * to track when the vcpu is done handling the NMI.
1078 */
1079static int
1080nmi_blocked(struct svm_softc *sc, int vcpu)
1081{
1082 int blocked;
1083
1084 blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1085 VMCB_INTCPT_IRET);
1086 return (blocked);
1087}
1088
1089static void
1090enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1091{
1092
1093 KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1094 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1095 svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1096}
1097
1098static void
1099clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1100{
1101 int error;
1102
1103 KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1104 VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1105 /*
1106 * When the IRET intercept is cleared the vcpu will attempt to execute
1107 * the "iret" when it runs next. However, it is possible to inject
1108 * another NMI into the vcpu before the "iret" has actually executed.
1109 *
1110 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1111 * it will trap back into the hypervisor. If an NMI is pending for
1112 * the vcpu it will be injected into the guest.
1113 *
1114 * XXX this needs to be fixed
1115 */
1116 svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1117
1118 /*
1119 * Set 'intr_shadow' to prevent an NMI from being injected on the
1120 * immediate VMRUN.
1121 */
1122 error = svm_modify_intr_shadow(sc, vcpu, 1);
1123 KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1124}
1125
1126static int
1127emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1128 bool *retu)
1129{
1130 int error;
1131
1132 if (lapic_msr(num))
1133 error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1134 else if (num == MSR_EFER)
1135 error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, val);
1136 else
1137 error = svm_wrmsr(sc, vcpu, num, val, retu);
1138
1139 return (error);
1140}
1141
1142static int
1143emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1144{
1145 struct vmcb_state *state;
1146 struct svm_regctx *ctx;
1147 uint64_t result;
1148 int error;
1149
1150 if (lapic_msr(num))
1151 error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1152 else
1153 error = svm_rdmsr(sc, vcpu, num, &result, retu);
1154
1155 if (error == 0) {
1156 state = svm_get_vmcb_state(sc, vcpu);
1157 ctx = svm_get_guest_regctx(sc, vcpu);
1158 state->rax = result & 0xffffffff;
1159 ctx->e.g.sctx_rdx = result >> 32;
1160 }
1161
1162 return (error);
1163}
1164
1165#ifdef KTR
1166static const char *
1167exit_reason_to_str(uint64_t reason)
1168{
1169 static char reasonbuf[32];
1170
1171 switch (reason) {
1172 case VMCB_EXIT_INVALID:
1173 return ("invalvmcb");
1174 case VMCB_EXIT_SHUTDOWN:
1175 return ("shutdown");
1176 case VMCB_EXIT_NPF:
1177 return ("nptfault");
1178 case VMCB_EXIT_PAUSE:
1179 return ("pause");
1180 case VMCB_EXIT_HLT:
1181 return ("hlt");
1182 case VMCB_EXIT_CPUID:
1183 return ("cpuid");
1184 case VMCB_EXIT_IO:
1185 return ("inout");
1186 case VMCB_EXIT_MC:
1187 return ("mchk");
1188 case VMCB_EXIT_INTR:
1189 return ("extintr");
1190 case VMCB_EXIT_NMI:
1191 return ("nmi");
1192 case VMCB_EXIT_VINTR:
1193 return ("vintr");
1194 case VMCB_EXIT_MSR:
1195 return ("msr");
1196 case VMCB_EXIT_IRET:
1197 return ("iret");
1198 default:
1199 snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1200 return (reasonbuf);
1201 }
1202}
1203#endif /* KTR */
1204
1205/*
1206 * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1207 * that are due to instruction intercepts as well as MSR and IOIO intercepts
1208 * and exceptions caused by INT3, INTO and BOUND instructions.
1209 *
1210 * Return 1 if the nRIP is valid and 0 otherwise.
1211 */
1212static int
1213nrip_valid(uint64_t exitcode)
1214{
1215 switch (exitcode) {
1216 case 0x00 ... 0x0F: /* read of CR0 through CR15 */
1217 case 0x10 ... 0x1F: /* write of CR0 through CR15 */
1218 case 0x20 ... 0x2F: /* read of DR0 through DR15 */
1219 case 0x30 ... 0x3F: /* write of DR0 through DR15 */
1220 case 0x43: /* INT3 */
1221 case 0x44: /* INTO */
1222 case 0x45: /* BOUND */
1223 case 0x65 ... 0x7C: /* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1224 case 0x80 ... 0x8D: /* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1225 return (1);
1226 default:
1227 return (0);
1228 }
1229}
1230
1231/*
1232 * Collateral for a generic SVM VM-exit.
1233 */
1234static void
1235vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
1236{
1237
1238 vme->exitcode = VM_EXITCODE_SVM;
1239 vme->u.svm.exitcode = code;
1240 vme->u.svm.exitinfo1 = info1;
1241 vme->u.svm.exitinfo2 = info2;
1242}
1243
1244static int
1245svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1246{
1247 struct vmcb *vmcb;
1248 struct vmcb_state *state;
1249 struct vmcb_ctrl *ctrl;
1250 struct svm_regctx *ctx;
1251 uint64_t code, info1, info2, val;
1252 uint32_t eax, ecx, edx;
1253 int handled;
1254 bool retu;
1255
1256 ctx = svm_get_guest_regctx(svm_sc, vcpu);
1257 vmcb = svm_get_vmcb(svm_sc, vcpu);
1258 state = &vmcb->state;
1259 ctrl = &vmcb->ctrl;
1260
1261 handled = 0;
1262 code = ctrl->exitcode;
1263 info1 = ctrl->exitinfo1;
1264 info2 = ctrl->exitinfo2;
1265
1266 vmexit->exitcode = VM_EXITCODE_BOGUS;
1267 vmexit->rip = state->rip;
1268 vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1269
1270 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1271
1272 /*
1273 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1274 * in an inconsistent state and can trigger assertions that would
1275 * never happen otherwise.
1276 */
1277 if (code == VMCB_EXIT_INVALID) {
1278 vm_exit_svm(vmexit, code, info1, info2);
1279 return (0);
1280 }
1281
1282 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1283 "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1284
1285 KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1286 ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1287 vmexit->inst_length, code, info1, info2));
1288
1289 svm_update_virqinfo(svm_sc, vcpu);
1290 svm_save_intinfo(svm_sc, vcpu);
1291
1292 switch (code) {
1293 case VMCB_EXIT_IRET:
1294 /*
1295 * Restart execution at "iret" but with the intercept cleared.
1296 */
1297 vmexit->inst_length = 0;
1298 clear_nmi_blocking(svm_sc, vcpu);
1299 handled = 1;
1300 break;
1301 case VMCB_EXIT_VINTR: /* interrupt window exiting */
1302 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1303 handled = 1;
1304 break;
1305 case VMCB_EXIT_INTR: /* external interrupt */
1306 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1307 handled = 1;
1308 break;
1309 case VMCB_EXIT_NMI: /* external NMI */
1310 handled = 1;
1311 break;
1312 case VMCB_EXIT_MC: /* machine check */
1313 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1314 break;
1315 case VMCB_EXIT_MSR: /* MSR access. */
1316 eax = state->rax;
1317 ecx = ctx->sctx_rcx;
1318 edx = ctx->e.g.sctx_rdx;
1319 retu = false;
1320
1321 if (info1) {
1322 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1323 val = (uint64_t)edx << 32 | eax;
1324 VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1325 ecx, val);
1326 if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1327 vmexit->exitcode = VM_EXITCODE_WRMSR;
1328 vmexit->u.msr.code = ecx;
1329 vmexit->u.msr.wval = val;
1330 } else if (!retu) {
1331 handled = 1;
1332 } else {
1333 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1334 ("emulate_wrmsr retu with bogus exitcode"));
1335 }
1336 } else {
1337 VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1338 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1339 if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1340 vmexit->exitcode = VM_EXITCODE_RDMSR;
1341 vmexit->u.msr.code = ecx;
1342 } else if (!retu) {
1343 handled = 1;
1344 } else {
1345 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1346 ("emulate_rdmsr retu with bogus exitcode"));
1347 }
1348 }
1349 break;
1350 case VMCB_EXIT_IO:
1351 handled = svm_handle_io(svm_sc, vcpu, vmexit);
1352 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1353 break;
1354 case VMCB_EXIT_CPUID:
1355 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1356 handled = x86_emulate_cpuid(svm_sc->vm, vcpu,
1357 (uint32_t *)&state->rax,
1358 (uint32_t *)&ctx->sctx_rbx,
1359 (uint32_t *)&ctx->sctx_rcx,
1360 (uint32_t *)&ctx->e.g.sctx_rdx);
1361 break;
1362 case VMCB_EXIT_HLT:
1363 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1364 vmexit->exitcode = VM_EXITCODE_HLT;
1365 vmexit->u.hlt.rflags = state->rflags;
1366 break;
1367 case VMCB_EXIT_PAUSE:
1368 vmexit->exitcode = VM_EXITCODE_PAUSE;
1369 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1370 break;
1371 case VMCB_EXIT_NPF:
1372 /* EXITINFO2 contains the faulting guest physical address */
1373 if (info1 & VMCB_NPF_INFO1_RSV) {
1374 VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1375 "reserved bits set: info1(%#lx) info2(%#lx)",
1376 info1, info2);
1377 } else if (vm_mem_allocated(svm_sc->vm, info2)) {
1378 vmexit->exitcode = VM_EXITCODE_PAGING;
1379 vmexit->u.paging.gpa = info2;
1380 vmexit->u.paging.fault_type = svm_npf_paging(info1);
1381 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1382 VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1383 "on gpa %#lx/%#lx at rip %#lx",
1384 info2, info1, state->rip);
1385 } else if (svm_npf_emul_fault(info1)) {
1386 svm_handle_inst_emul(vmcb, info2, vmexit);
1387 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1388 VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1389 "for gpa %#lx/%#lx at rip %#lx",
1390 info2, info1, state->rip);
1391 }
1392 break;
1393 default:
1394 vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1395 break;
1396 }
1397
1398 VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1399 handled ? "handled" : "unhandled", exit_reason_to_str(code),
1400 vmexit->rip, vmexit->inst_length);
1401
1402 if (handled) {
1403 vmexit->rip += vmexit->inst_length;
1404 vmexit->inst_length = 0;
1405 state->rip = vmexit->rip;
1406 } else {
1407 if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1408 /*
1409 * If this VM exit was not claimed by anybody then
1410 * treat it as a generic SVM exit.
1411 */
1412 vm_exit_svm(vmexit, code, info1, info2);
1413 } else {
1414 /*
1415 * The exitcode and collateral have been populated.
1416 * The VM exit will be processed further in userland.
1417 */
1418 }
1419 }
1420 return (handled);
1421}
1422
1423static void
1424svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1425{
1426 uint64_t intinfo;
1427
1428 if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1429 return;
1430
1431 KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1432 "valid: %#lx", __func__, intinfo));
1433
1434 svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1435 VMCB_EXITINTINFO_VECTOR(intinfo),
1436 VMCB_EXITINTINFO_EC(intinfo),
1437 VMCB_EXITINTINFO_EC_VALID(intinfo));
1438 vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1439 VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1440}
1441
1442/*
1443 * Inject event to virtual cpu.
1444 */
1445static void
1446svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1447{
1448 struct vmcb_ctrl *ctrl;
1449 struct vmcb_state *state;
1450 uint8_t v_tpr;
1451 int vector, need_intr_window, pending_apic_vector;
1452
1453 state = svm_get_vmcb_state(sc, vcpu);
1454 ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1455
1456 need_intr_window = 0;
1457 pending_apic_vector = 0;
1458
1459 /*
1460 * Inject pending events or exceptions for this vcpu.
1461 *
1462 * An event might be pending because the previous #VMEXIT happened
1463 * during event delivery (i.e. ctrl->exitintinfo).
1464 *
1465 * An event might also be pending because an exception was injected
1466 * by the hypervisor (e.g. #PF during instruction emulation).
1467 */
1468 svm_inj_intinfo(sc, vcpu);
1469
1470 /* NMI event has priority over interrupts. */
1471 if (vm_nmi_pending(sc->vm, vcpu)) {
1472 if (nmi_blocked(sc, vcpu)) {
1473 /*
1474 * Can't inject another NMI if the guest has not
1475 * yet executed an "iret" after the last NMI.
1476 */
1477 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1478 "to NMI-blocking");
1479 } else if (ctrl->intr_shadow) {
1480 /*
1481 * Can't inject an NMI if the vcpu is in an intr_shadow.
1482 */
1483 VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1484 "interrupt shadow");
1485 need_intr_window = 1;
1486 goto done;
1487 } else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1488 /*
1489 * If there is already an exception/interrupt pending
1490 * then defer the NMI until after that.
1491 */
1492 VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1493 "eventinj %#lx", ctrl->eventinj);
1494
1495 /*
1496 * Use self-IPI to trigger a VM-exit as soon as
1497 * possible after the event injection is completed.
1498 *
1499 * This works only if the external interrupt exiting
1500 * is at a lower priority than the event injection.
1501 *
1502 * Although not explicitly specified in APMv2 the
1503 * relative priorities were verified empirically.
1504 */
1505 ipi_cpu(curcpu, IPI_AST); /* XXX vmm_ipinum? */
1506 } else {
1507 vm_nmi_clear(sc->vm, vcpu);
1508
1509 /* Inject NMI, vector number is not used */
1510 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1511 IDT_NMI, 0, false);
1512
1513 /* virtual NMI blocking is now in effect */
1514 enable_nmi_blocking(sc, vcpu);
1515
1516 VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1517 }
1518 }
1519
1520 if (!vm_extint_pending(sc->vm, vcpu)) {
1521 /*
1522 * APIC interrupts are delivered using the V_IRQ offload.
1523 *
1524 * The primary benefit is that the hypervisor doesn't need to
1525 * deal with the various conditions that inhibit interrupts.
1526 * It also means that TPR changes via CR8 will be handled
1527 * without any hypervisor involvement.
1528 *
1529 * Note that the APIC vector must remain pending in the vIRR
1530 * until it is confirmed that it was delivered to the guest.
1531 * This can be confirmed based on the value of V_IRQ at the
1532 * next #VMEXIT (1 = pending, 0 = delivered).
1533 *
1534 * Also note that it is possible that another higher priority
1535 * vector can become pending before this vector is delivered
1536 * to the guest. This is alright because vcpu_notify_event()
1537 * will send an IPI and force the vcpu to trap back into the
1538 * hypervisor. The higher priority vector will be injected on
1539 * the next VMRUN.
1540 */
1541 if (vlapic_pending_intr(vlapic, &vector)) {
1542 KASSERT(vector >= 16 && vector <= 255,
1543 ("invalid vector %d from local APIC", vector));
1544 pending_apic_vector = vector;
1545 }
1546 goto done;
1547 }
1548
1549 /* Ask the legacy pic for a vector to inject */
1550 vatpic_pending_intr(sc->vm, &vector);
1551 KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR",
1552 vector));
1553
1554 /*
1555 * If the guest has disabled interrupts or is in an interrupt shadow
1556 * then we cannot inject the pending interrupt.
1557 */
1558 if ((state->rflags & PSL_I) == 0) {
1559 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1560 "rflags %#lx", vector, state->rflags);
1561 need_intr_window = 1;
1562 goto done;
1563 }
1564
1565 if (ctrl->intr_shadow) {
1566 VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1567 "interrupt shadow", vector);
1568 need_intr_window = 1;
1569 goto done;
1570 }
1571
1572 if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1573 VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1574 "eventinj %#lx", vector, ctrl->eventinj);
1575 need_intr_window = 1;
1576 goto done;
1577 }
1578
1579 /*
1580 * Legacy PIC interrupts are delivered via the event injection
1581 * mechanism.
1582 */
1583 svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1584
1585 vm_extint_clear(sc->vm, vcpu);
1586 vatpic_intr_accepted(sc->vm, vector);
1587
1588 /*
1589 * Force a VM-exit as soon as the vcpu is ready to accept another
1590 * interrupt. This is done because the PIC might have another vector
1591 * that it wants to inject. Also, if the APIC has a pending interrupt
1592 * that was preempted by the ExtInt then it allows us to inject the
1593 * APIC vector as soon as possible.
1594 */
1595 need_intr_window = 1;
1596done:
1597 /*
1598 * The guest can modify the TPR by writing to %CR8. In guest mode
1599 * the processor reflects this write to V_TPR without hypervisor
1600 * intervention.
1601 *
1602 * The guest can also modify the TPR by writing to it via the memory
1603 * mapped APIC page. In this case, the write will be emulated by the
1604 * hypervisor. For this reason V_TPR must be updated before every
1605 * VMRUN.
1606 */
1607 v_tpr = vlapic_get_cr8(vlapic);
1608 KASSERT(v_tpr >= 0 && v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1609 if (ctrl->v_tpr != v_tpr) {
1610 VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1611 ctrl->v_tpr, v_tpr);
1612 ctrl->v_tpr = v_tpr;
1613 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1614 }
1615
1616 if (pending_apic_vector) {
1617 /*
1618 * If an APIC vector is being injected then interrupt window
1619 * exiting is not possible on this VMRUN.
1620 */
1621 KASSERT(!need_intr_window, ("intr_window exiting impossible"));
1622 VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ",
1623 pending_apic_vector);
1624
1625 ctrl->v_irq = 1;
1626 ctrl->v_ign_tpr = 0;
1627 ctrl->v_intr_vector = pending_apic_vector;
1628 ctrl->v_intr_prio = pending_apic_vector >> 4;
1629 svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1630 } else if (need_intr_window) {
1631 /*
1632 * We use V_IRQ in conjunction with the VINTR intercept to
1633 * trap into the hypervisor as soon as a virtual interrupt
1634 * can be delivered.
1635 *
1636 * Since injected events are not subject to intercept checks
1637 * we need to ensure that the V_IRQ is not actually going to
1638 * be delivered on VM entry. The KASSERT below enforces this.
1639 */
1640 KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1641 (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1642 ("Bogus intr_window_exiting: eventinj (%#lx), "
1643 "intr_shadow (%u), rflags (%#lx)",
1644 ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1645 enable_intr_window_exiting(sc, vcpu);
1646 } else {
1647 disable_intr_window_exiting(sc, vcpu);
1648 }
1649}
1650
1651static __inline void
1652restore_host_tss(void)
1653{
1654 struct system_segment_descriptor *tss_sd;
1655
1656 /*
1657 * The TSS descriptor was in use prior to launching the guest so it
1658 * has been marked busy.
1659 *
1660 * 'ltr' requires the descriptor to be marked available so change the
1661 * type to "64-bit available TSS".
1662 */
1663 tss_sd = PCPU_GET(tss);
1664 tss_sd->sd_type = SDT_SYSTSS;
1665 ltr(GSEL(GPROC0_SEL, SEL_KPL));
1666}
1667
1668static void
1669check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu)
1670{
1671 struct svm_vcpu *vcpustate;
1672 struct vmcb_ctrl *ctrl;
1673 long eptgen;
1674 bool alloc_asid;
1675
1676 KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not "
1677 "active on cpu %u", __func__, thiscpu));
1678
1679 vcpustate = svm_get_vcpu(sc, vcpuid);
1680 ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1681
1682 /*
1683 * The TLB entries associated with the vcpu's ASID are not valid
1684 * if either of the following conditions is true:
1685 *
1686 * 1. The vcpu's ASID generation is different than the host cpu's
1687 * ASID generation. This happens when the vcpu migrates to a new
1688 * host cpu. It can also happen when the number of vcpus executing
1689 * on a host cpu is greater than the number of ASIDs available.
1690 *
1691 * 2. The pmap generation number is different than the value cached in
1692 * the 'vcpustate'. This happens when the host invalidates pages
1693 * belonging to the guest.
1694 *
1695 * asidgen eptgen Action
1696 * mismatch mismatch
1697 * 0 0 (a)
1698 * 0 1 (b1) or (b2)
1699 * 1 0 (c)
1700 * 1 1 (d)
1701 *
1702 * (a) There is no mismatch in eptgen or ASID generation and therefore
1703 * no further action is needed.
1704 *
1705 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1706 * retained and the TLB entries associated with this ASID
1707 * are flushed by VMRUN.
1708 *
1709 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1710 * allocated.
1711 *
1712 * (c) A new ASID is allocated.
1713 *
1714 * (d) A new ASID is allocated.
1715 */
1716
1717 alloc_asid = false;
1718 eptgen = pmap->pm_eptgen;
1719 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1720
1721 if (vcpustate->asid.gen != asid[thiscpu].gen) {
1722 alloc_asid = true; /* (c) and (d) */
1723 } else if (vcpustate->eptgen != eptgen) {
1724 if (flush_by_asid())
1725 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST; /* (b1) */
1726 else
1727 alloc_asid = true; /* (b2) */
1728 } else {
1729 /*
1730 * This is the common case (a).
1731 */
1732 KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1733 KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1734 ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1735 }
1736
1737 if (alloc_asid) {
1738 if (++asid[thiscpu].num >= nasid) {
1739 asid[thiscpu].num = 1;
1740 if (++asid[thiscpu].gen == 0)
1741 asid[thiscpu].gen = 1;
1742 /*
1743 * If this cpu does not support "flush-by-asid"
1744 * then flush the entire TLB on a generation
1745 * bump. Subsequent ASID allocation in this
1746 * generation can be done without a TLB flush.
1747 */
1748 if (!flush_by_asid())
1749 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1750 }
1751 vcpustate->asid.gen = asid[thiscpu].gen;
1752 vcpustate->asid.num = asid[thiscpu].num;
1753
1754 ctrl->asid = vcpustate->asid.num;
1755 svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1756 /*
1757 * If this cpu supports "flush-by-asid" then the TLB
1758 * was not flushed after the generation bump. The TLB
1759 * is flushed selectively after every new ASID allocation.
1760 */
1761 if (flush_by_asid())
1762 ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1763 }
1764 vcpustate->eptgen = eptgen;
1765
1766 KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1767 KASSERT(ctrl->asid == vcpustate->asid.num,
1768 ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1769}
1770
1771/*
1772 * Start vcpu with specified RIP.
1773 */
1774static int
1775svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap,
1776 void *rend_cookie, void *suspended_cookie)
1777{
1778 struct svm_regctx *hctx, *gctx;
1779 struct svm_softc *svm_sc;
1780 struct svm_vcpu *vcpustate;
1781 struct vmcb_state *state;
1782 struct vmcb_ctrl *ctrl;
1783 struct vm_exit *vmexit;
1784 struct vlapic *vlapic;
1785 struct vm *vm;
1786 uint64_t vmcb_pa;
1787 u_int thiscpu;
1788 int handled;
1789
1790 svm_sc = arg;
1791 vm = svm_sc->vm;
1792
1793 vcpustate = svm_get_vcpu(svm_sc, vcpu);
1794 state = svm_get_vmcb_state(svm_sc, vcpu);
1795 ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
1796 vmexit = vm_exitinfo(vm, vcpu);
1797 vlapic = vm_lapic(vm, vcpu);
1798
1799 /*
1800 * Stash 'curcpu' on the stack as 'thiscpu'.
1801 *
1802 * The per-cpu data area is not accessible until MSR_GSBASE is restored
1803 * after the #VMEXIT. Since VMRUN is executed inside a critical section
1804 * 'curcpu' and 'thiscpu' are guaranteed to identical.
1805 */
1806 thiscpu = curcpu;
1807
1808 gctx = svm_get_guest_regctx(svm_sc, vcpu);
1809 hctx = &host_ctx[thiscpu];
1810 vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
1811
1812 if (vcpustate->lastcpu != thiscpu) {
1813 /*
1814 * Force new ASID allocation by invalidating the generation.
1815 */
1816 vcpustate->asid.gen = 0;
1817
1818 /*
1819 * Invalidate the VMCB state cache by marking all fields dirty.
1820 */
1821 svm_set_dirty(svm_sc, vcpu, 0xffffffff);
1822
1823 /*
1824 * XXX
1825 * Setting 'vcpustate->lastcpu' here is bit premature because
1826 * we may return from this function without actually executing
1827 * the VMRUN instruction. This could happen if a rendezvous
1828 * or an AST is pending on the first time through the loop.
1829 *
1830 * This works for now but any new side-effects of vcpu
1831 * migration should take this case into account.
1832 */
1833 vcpustate->lastcpu = thiscpu;
1834 vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
1835 }
1836
1837 svm_msr_guest_enter(svm_sc, vcpu);
1838
1839 /* Update Guest RIP */
1840 state->rip = rip;
1841
1842 do {
1843 /*
1844 * Disable global interrupts to guarantee atomicity during
1845 * loading of guest state. This includes not only the state
1846 * loaded by the "vmrun" instruction but also software state
1847 * maintained by the hypervisor: suspended and rendezvous
1848 * state, NPT generation number, vlapic interrupts etc.
1849 */
1850 disable_gintr();
1851
1852 if (vcpu_suspended(suspended_cookie)) {
1853 enable_gintr();
1854 vm_exit_suspended(vm, vcpu, state->rip);
1855 break;
1856 }
1857
1858 if (vcpu_rendezvous_pending(rend_cookie)) {
1859 enable_gintr();
1860 vm_exit_rendezvous(vm, vcpu, state->rip);
1861 break;
1862 }
1863
1864 /* We are asked to give the cpu by scheduler. */
1865 if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) {
1866 enable_gintr();
1867 vm_exit_astpending(vm, vcpu, state->rip);
1868 break;
1869 }
1870
1871 svm_inj_interrupts(svm_sc, vcpu, vlapic);
1872
1873 /* Activate the nested pmap on 'thiscpu' */
1874 CPU_SET_ATOMIC_ACQ(thiscpu, &pmap->pm_active);
1875
1876 /*
1877 * Check the pmap generation and the ASID generation to
1878 * ensure that the vcpu does not use stale TLB mappings.
1879 */
1880 check_asid(svm_sc, vcpu, pmap, thiscpu);
1881
1882 ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
1883 vcpustate->dirty = 0;
1884 VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
1885
1886 /* Launch Virtual Machine. */
1887 VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
1888 svm_launch(vmcb_pa, gctx, hctx);
1889
1890 CPU_CLR_ATOMIC(thiscpu, &pmap->pm_active);
1891
1892 /*
1893 * Restore MSR_GSBASE to point to the pcpu data area.
1894 *
1895 * Note that accesses done via PCPU_GET/PCPU_SET will work
1896 * only after MSR_GSBASE is restored.
1897 *
1898 * Also note that we don't bother restoring MSR_KGSBASE
1899 * since it is not used in the kernel and will be restored
1900 * when the VMRUN ioctl returns to userspace.
1901 */
1902 wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[thiscpu]);
1903 KASSERT(curcpu == thiscpu, ("thiscpu/curcpu (%u/%u) mismatch",
1904 thiscpu, curcpu));
1905
1906 /*
1907 * The host GDTR and IDTR is saved by VMRUN and restored
1908 * automatically on #VMEXIT. However, the host TSS needs
1909 * to be restored explicitly.
1910 */
1911 restore_host_tss();
1912
1913 /* #VMEXIT disables interrupts so re-enable them here. */
1914 enable_gintr();
1915
1916 /* Handle #VMEXIT and if required return to user space. */
1917 handled = svm_vmexit(svm_sc, vcpu, vmexit);
1918 } while (handled);
1919
1920 svm_msr_guest_exit(svm_sc, vcpu);
1921
1922 return (0);
1923}
1924
1925/*
1926 * Cleanup for virtual machine.
1927 */
1928static void
1929svm_vmcleanup(void *arg)
1930{
1931 struct svm_softc *svm_sc;
1932
1933 svm_sc = arg;
1934
1935 VCPU_CTR0(svm_sc->vm, 0, "SVM:cleanup\n");
1936
1937 free(svm_sc, M_SVM);
1938}
1939
1940/*
1941 * Return pointer to hypervisor saved register state.
1942 */
1943static register_t *
1944swctx_regptr(struct svm_regctx *regctx, int reg)
1945{
1946
1947 switch (reg) {
1948 case VM_REG_GUEST_RBX:
1949 return (®ctx->sctx_rbx);
1950 case VM_REG_GUEST_RCX:
1951 return (®ctx->sctx_rcx);
1952 case VM_REG_GUEST_RDX:
1953 return (®ctx->e.g.sctx_rdx);
1954 case VM_REG_GUEST_RDI:
1955 return (®ctx->e.g.sctx_rdi);
1956 case VM_REG_GUEST_RSI:
1957 return (®ctx->e.g.sctx_rsi);
1958 case VM_REG_GUEST_RBP:
1959 return (®ctx->sctx_rbp);
1960 case VM_REG_GUEST_R8:
1961 return (®ctx->sctx_r8);
1962 case VM_REG_GUEST_R9:
1963 return (®ctx->sctx_r9);
1964 case VM_REG_GUEST_R10:
1965 return (®ctx->sctx_r10);
1966 case VM_REG_GUEST_R11:
1967 return (®ctx->sctx_r11);
1968 case VM_REG_GUEST_R12:
1969 return (®ctx->sctx_r12);
1970 case VM_REG_GUEST_R13:
1971 return (®ctx->sctx_r13);
1972 case VM_REG_GUEST_R14:
1973 return (®ctx->sctx_r14);
1974 case VM_REG_GUEST_R15:
1975 return (®ctx->sctx_r15);
1976 default:
1977 ERR("Unknown register requested, reg=%d.\n", reg);
1978 break;
1979 }
1980
1981 return (NULL);
1982}
1983
1984/*
1985 * Interface to read guest registers.
1986 * This can be SVM h/w saved or hypervisor saved register.
1987 */
1988static int
1989svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
1990{
1991 struct svm_softc *svm_sc;
1992 register_t *reg;
1993
1994 svm_sc = arg;
1995
1996 if (ident == VM_REG_GUEST_INTR_SHADOW) {
1997 return (svm_get_intr_shadow(svm_sc, vcpu, val));
1998 }
1999
2000 if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2001 return (0);
2002 }
2003
2004 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2005
2006 if (reg != NULL) {
2007 *val = *reg;
2008 return (0);
2009 }
2010
2011 ERR("SVM_ERR:reg type %x is not saved in VMCB.\n", ident);
2012 return (EINVAL);
2013}
2014
2015/*
2016 * Interface to write to guest registers.
2017 * This can be SVM h/w saved or hypervisor saved register.
2018 */
2019static int
2020svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2021{
2022 struct svm_softc *svm_sc;
2023 register_t *reg;
2024
2025 svm_sc = arg;
2026
2027 if (ident == VM_REG_GUEST_INTR_SHADOW) {
2028 return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2029 }
2030
2031 if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2032 return (0);
2033 }
2034
2035 reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2036
2037 if (reg != NULL) {
2038 *reg = val;
2039 return (0);
2040 }
2041
2042 /*
2043 * XXX deal with CR3 and invalidate TLB entries tagged with the
2044 * vcpu's ASID. This needs to be treated differently depending on
2045 * whether 'running' is true/false.
2046 */
2047
2048 ERR("SVM_ERR:reg type %x is not saved in VMCB.\n", ident);
2049 return (EINVAL);
2050}
2051
2052static int
2053svm_setcap(void *arg, int vcpu, int type, int val)
2054{
2055 struct svm_softc *sc;
2056 int error;
2057
2058 sc = arg;
2059 error = 0;
2060 switch (type) {
2061 case VM_CAP_HALT_EXIT:
2062 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2063 VMCB_INTCPT_HLT, val);
2064 break;
2065 case VM_CAP_PAUSE_EXIT:
2066 svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2067 VMCB_INTCPT_PAUSE, val);
2068 break;
2069 case VM_CAP_UNRESTRICTED_GUEST:
2070 /* Unrestricted guest execution cannot be disabled in SVM */
2071 if (val == 0)
2072 error = EINVAL;
2073 break;
2074 default:
2075 error = ENOENT;
2076 break;
2077 }
2078 return (error);
2079}
2080
2081static int
2082svm_getcap(void *arg, int vcpu, int type, int *retval)
2083{
2084 struct svm_softc *sc;
2085 int error;
2086
2087 sc = arg;
2088 error = 0;
2089
2090 switch (type) {
2091 case VM_CAP_HALT_EXIT:
2092 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2093 VMCB_INTCPT_HLT);
2094 break;
2095 case VM_CAP_PAUSE_EXIT:
2096 *retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2097 VMCB_INTCPT_PAUSE);
2098 break;
2099 case VM_CAP_UNRESTRICTED_GUEST:
2100 *retval = 1; /* unrestricted guest is always enabled */
2101 break;
2102 default:
2103 error = ENOENT;
2104 break;
2105 }
2106 return (error);
2107}
2108
2109static struct vlapic *
2110svm_vlapic_init(void *arg, int vcpuid)
2111{
2112 struct svm_softc *svm_sc;
2113 struct vlapic *vlapic;
2114
2115 svm_sc = arg;
2116 vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2117 vlapic->vm = svm_sc->vm;
2118 vlapic->vcpuid = vcpuid;
2119 vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2120
2121 vlapic_init(vlapic);
2122
2123 return (vlapic);
2124}
2125
2126static void
2127svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2128{
2129
2130 vlapic_cleanup(vlapic);
2131 free(vlapic, M_SVM_VLAPIC);
2132}
2133
2134struct vmm_ops vmm_ops_amd = {
2135 svm_init,
2136 svm_cleanup,
2137 svm_restore,
2138 svm_vminit,
2139 svm_vmrun,
2140 svm_vmcleanup,
2141 svm_getreg,
2142 svm_setreg,
2143 vmcb_getdesc,
2144 vmcb_setdesc,
2145 svm_getcap,
2146 svm_setcap,
2147 svm_npt_alloc,
2148 svm_npt_free,
2149 svm_vlapic_init,
2150 svm_vlapic_cleanup
2151};