xref: /freebsd/sys/amd64/vmm/amd/svm.c (revision 276da39af92f48350aa01091a2b8b3e735217eea)
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("$FreeBSD$");
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/specialreg.h>
47 #include <machine/smp.h>
48 #include <machine/vmm.h>
49 #include <machine/vmm_dev.h>
50 #include <machine/vmm_instruction_emul.h>
51 
52 #include "vmm_lapic.h"
53 #include "vmm_stat.h"
54 #include "vmm_ktr.h"
55 #include "vmm_ioport.h"
56 #include "vatpic.h"
57 #include "vlapic.h"
58 #include "vlapic_priv.h"
59 
60 #include "x86.h"
61 #include "vmcb.h"
62 #include "svm.h"
63 #include "svm_softc.h"
64 #include "svm_msr.h"
65 #include "npt.h"
66 
67 SYSCTL_DECL(_hw_vmm);
68 SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL);
69 
70 /*
71  * SVM CPUID function 0x8000_000A, edx bit decoding.
72  */
73 #define AMD_CPUID_SVM_NP		BIT(0)  /* Nested paging or RVI */
74 #define AMD_CPUID_SVM_LBR		BIT(1)  /* Last branch virtualization */
75 #define AMD_CPUID_SVM_SVML		BIT(2)  /* SVM lock */
76 #define AMD_CPUID_SVM_NRIP_SAVE		BIT(3)  /* Next RIP is saved */
77 #define AMD_CPUID_SVM_TSC_RATE		BIT(4)  /* TSC rate control. */
78 #define AMD_CPUID_SVM_VMCB_CLEAN	BIT(5)  /* VMCB state caching */
79 #define AMD_CPUID_SVM_FLUSH_BY_ASID	BIT(6)  /* Flush by ASID */
80 #define AMD_CPUID_SVM_DECODE_ASSIST	BIT(7)  /* Decode assist */
81 #define AMD_CPUID_SVM_PAUSE_INC		BIT(10) /* Pause intercept filter. */
82 #define AMD_CPUID_SVM_PAUSE_FTH		BIT(12) /* Pause filter threshold */
83 #define	AMD_CPUID_SVM_AVIC		BIT(13)	/* AVIC present */
84 
85 #define	VMCB_CACHE_DEFAULT	(VMCB_CACHE_ASID 	|	\
86 				VMCB_CACHE_IOPM		|	\
87 				VMCB_CACHE_I		|	\
88 				VMCB_CACHE_TPR		|	\
89 				VMCB_CACHE_CR2		|	\
90 				VMCB_CACHE_CR		|	\
91 				VMCB_CACHE_DT		|	\
92 				VMCB_CACHE_SEG		|	\
93 				VMCB_CACHE_NP)
94 
95 static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT;
96 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean,
97     0, NULL);
98 
99 static MALLOC_DEFINE(M_SVM, "svm", "svm");
100 static MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic");
101 
102 /* Per-CPU context area. */
103 extern struct pcpu __pcpu[];
104 
105 static uint32_t svm_feature = ~0U;	/* AMD SVM features. */
106 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RDTUN, &svm_feature, 0,
107     "SVM features advertised by CPUID.8000000AH:EDX");
108 
109 static int disable_npf_assist;
110 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN,
111     &disable_npf_assist, 0, NULL);
112 
113 /* Maximum ASIDs supported by the processor */
114 static uint32_t nasid;
115 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RDTUN, &nasid, 0,
116     "Number of ASIDs supported by this processor");
117 
118 /* Current ASID generation for each host cpu */
119 static struct asid asid[MAXCPU];
120 
121 /*
122  * SVM host state saved area of size 4KB for each core.
123  */
124 static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE);
125 
126 static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery");
127 static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry");
128 static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window");
129 
130 static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val);
131 
132 static __inline int
133 flush_by_asid(void)
134 {
135 
136 	return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID);
137 }
138 
139 static __inline int
140 decode_assist(void)
141 {
142 
143 	return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST);
144 }
145 
146 static void
147 svm_disable(void *arg __unused)
148 {
149 	uint64_t efer;
150 
151 	efer = rdmsr(MSR_EFER);
152 	efer &= ~EFER_SVM;
153 	wrmsr(MSR_EFER, efer);
154 }
155 
156 /*
157  * Disable SVM on all CPUs.
158  */
159 static int
160 svm_cleanup(void)
161 {
162 
163 	smp_rendezvous(NULL, svm_disable, NULL, NULL);
164 	return (0);
165 }
166 
167 /*
168  * Verify that all the features required by bhyve are available.
169  */
170 static int
171 check_svm_features(void)
172 {
173 	u_int regs[4];
174 
175 	/* CPUID Fn8000_000A is for SVM */
176 	do_cpuid(0x8000000A, regs);
177 	svm_feature &= regs[3];
178 
179 	/*
180 	 * The number of ASIDs can be configured to be less than what is
181 	 * supported by the hardware but not more.
182 	 */
183 	if (nasid == 0 || nasid > regs[1])
184 		nasid = regs[1];
185 	KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid));
186 
187 	/* bhyve requires the Nested Paging feature */
188 	if (!(svm_feature & AMD_CPUID_SVM_NP)) {
189 		printf("SVM: Nested Paging feature not available.\n");
190 		return (ENXIO);
191 	}
192 
193 	/* bhyve requires the NRIP Save feature */
194 	if (!(svm_feature & AMD_CPUID_SVM_NRIP_SAVE)) {
195 		printf("SVM: NRIP Save feature not available.\n");
196 		return (ENXIO);
197 	}
198 
199 	return (0);
200 }
201 
202 static void
203 svm_enable(void *arg __unused)
204 {
205 	uint64_t efer;
206 
207 	efer = rdmsr(MSR_EFER);
208 	efer |= EFER_SVM;
209 	wrmsr(MSR_EFER, efer);
210 
211 	wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave[curcpu]));
212 }
213 
214 /*
215  * Return 1 if SVM is enabled on this processor and 0 otherwise.
216  */
217 static int
218 svm_available(void)
219 {
220 	uint64_t msr;
221 
222 	/* Section 15.4 Enabling SVM from APM2. */
223 	if ((amd_feature2 & AMDID2_SVM) == 0) {
224 		printf("SVM: not available.\n");
225 		return (0);
226 	}
227 
228 	msr = rdmsr(MSR_VM_CR);
229 	if ((msr & VM_CR_SVMDIS) != 0) {
230 		printf("SVM: disabled by BIOS.\n");
231 		return (0);
232 	}
233 
234 	return (1);
235 }
236 
237 static int
238 svm_init(int ipinum)
239 {
240 	int error, cpu;
241 
242 	if (!svm_available())
243 		return (ENXIO);
244 
245 	error = check_svm_features();
246 	if (error)
247 		return (error);
248 
249 	vmcb_clean &= VMCB_CACHE_DEFAULT;
250 
251 	for (cpu = 0; cpu < MAXCPU; cpu++) {
252 		/*
253 		 * Initialize the host ASIDs to their "highest" valid values.
254 		 *
255 		 * The next ASID allocation will rollover both 'gen' and 'num'
256 		 * and start off the sequence at {1,1}.
257 		 */
258 		asid[cpu].gen = ~0UL;
259 		asid[cpu].num = nasid - 1;
260 	}
261 
262 	svm_msr_init();
263 	svm_npt_init(ipinum);
264 
265 	/* Enable SVM on all CPUs */
266 	smp_rendezvous(NULL, svm_enable, NULL, NULL);
267 
268 	return (0);
269 }
270 
271 static void
272 svm_restore(void)
273 {
274 
275 	svm_enable(NULL);
276 }
277 
278 /* Pentium compatible MSRs */
279 #define MSR_PENTIUM_START 	0
280 #define MSR_PENTIUM_END 	0x1FFF
281 /* AMD 6th generation and Intel compatible MSRs */
282 #define MSR_AMD6TH_START 	0xC0000000UL
283 #define MSR_AMD6TH_END 		0xC0001FFFUL
284 /* AMD 7th and 8th generation compatible MSRs */
285 #define MSR_AMD7TH_START 	0xC0010000UL
286 #define MSR_AMD7TH_END 		0xC0011FFFUL
287 
288 /*
289  * Get the index and bit position for a MSR in permission bitmap.
290  * Two bits are used for each MSR: lower bit for read and higher bit for write.
291  */
292 static int
293 svm_msr_index(uint64_t msr, int *index, int *bit)
294 {
295 	uint32_t base, off;
296 
297 	*index = -1;
298 	*bit = (msr % 4) * 2;
299 	base = 0;
300 
301 	if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) {
302 		*index = msr / 4;
303 		return (0);
304 	}
305 
306 	base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1);
307 	if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) {
308 		off = (msr - MSR_AMD6TH_START);
309 		*index = (off + base) / 4;
310 		return (0);
311 	}
312 
313 	base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1);
314 	if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) {
315 		off = (msr - MSR_AMD7TH_START);
316 		*index = (off + base) / 4;
317 		return (0);
318 	}
319 
320 	return (EINVAL);
321 }
322 
323 /*
324  * Allow vcpu to read or write the 'msr' without trapping into the hypervisor.
325  */
326 static void
327 svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write)
328 {
329 	int index, bit, error;
330 
331 	error = svm_msr_index(msr, &index, &bit);
332 	KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr));
333 	KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE,
334 	    ("%s: invalid index %d for msr %#lx", __func__, index, msr));
335 	KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d "
336 	    "msr %#lx", __func__, bit, msr));
337 
338 	if (read)
339 		perm_bitmap[index] &= ~(1UL << bit);
340 
341 	if (write)
342 		perm_bitmap[index] &= ~(2UL << bit);
343 }
344 
345 static void
346 svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr)
347 {
348 
349 	svm_msr_perm(perm_bitmap, msr, true, true);
350 }
351 
352 static void
353 svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr)
354 {
355 
356 	svm_msr_perm(perm_bitmap, msr, true, false);
357 }
358 
359 static __inline int
360 svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask)
361 {
362 	struct vmcb_ctrl *ctrl;
363 
364 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
365 
366 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
367 	return (ctrl->intercept[idx] & bitmask ? 1 : 0);
368 }
369 
370 static __inline void
371 svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask,
372     int enabled)
373 {
374 	struct vmcb_ctrl *ctrl;
375 	uint32_t oldval;
376 
377 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
378 
379 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
380 	oldval = ctrl->intercept[idx];
381 
382 	if (enabled)
383 		ctrl->intercept[idx] |= bitmask;
384 	else
385 		ctrl->intercept[idx] &= ~bitmask;
386 
387 	if (ctrl->intercept[idx] != oldval) {
388 		svm_set_dirty(sc, vcpu, VMCB_CACHE_I);
389 		VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified "
390 		    "from %#x to %#x", idx, oldval, ctrl->intercept[idx]);
391 	}
392 }
393 
394 static __inline void
395 svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
396 {
397 
398 	svm_set_intercept(sc, vcpu, off, bitmask, 0);
399 }
400 
401 static __inline void
402 svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
403 {
404 
405 	svm_set_intercept(sc, vcpu, off, bitmask, 1);
406 }
407 
408 static void
409 vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa,
410     uint64_t msrpm_base_pa, uint64_t np_pml4)
411 {
412 	struct vmcb_ctrl *ctrl;
413 	struct vmcb_state *state;
414 	uint32_t mask;
415 	int n;
416 
417 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
418 	state = svm_get_vmcb_state(sc, vcpu);
419 
420 	ctrl->iopm_base_pa = iopm_base_pa;
421 	ctrl->msrpm_base_pa = msrpm_base_pa;
422 
423 	/* Enable nested paging */
424 	ctrl->np_enable = 1;
425 	ctrl->n_cr3 = np_pml4;
426 
427 	/*
428 	 * Intercept accesses to the control registers that are not shadowed
429 	 * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8.
430 	 */
431 	for (n = 0; n < 16; n++) {
432 		mask = (BIT(n) << 16) | BIT(n);
433 		if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8)
434 			svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
435 		else
436 			svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
437 	}
438 
439 
440 	/*
441 	 * Intercept everything when tracing guest exceptions otherwise
442 	 * just intercept machine check exception.
443 	 */
444 	if (vcpu_trace_exceptions(sc->vm, vcpu)) {
445 		for (n = 0; n < 32; n++) {
446 			/*
447 			 * Skip unimplemented vectors in the exception bitmap.
448 			 */
449 			if (n == 2 || n == 9) {
450 				continue;
451 			}
452 			svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(n));
453 		}
454 	} else {
455 		svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC));
456 	}
457 
458 	/* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */
459 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO);
460 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR);
461 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID);
462 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR);
463 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT);
464 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI);
465 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI);
466 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN);
467 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
468 	    VMCB_INTCPT_FERR_FREEZE);
469 
470 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR);
471 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT);
472 
473 	/*
474 	 * From section "Canonicalization and Consistency Checks" in APMv2
475 	 * the VMRUN intercept bit must be set to pass the consistency check.
476 	 */
477 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
478 
479 	/*
480 	 * The ASID will be set to a non-zero value just before VMRUN.
481 	 */
482 	ctrl->asid = 0;
483 
484 	/*
485 	 * Section 15.21.1, Interrupt Masking in EFLAGS
486 	 * Section 15.21.2, Virtualizing APIC.TPR
487 	 *
488 	 * This must be set for %rflag and %cr8 isolation of guest and host.
489 	 */
490 	ctrl->v_intr_masking = 1;
491 
492 	/* Enable Last Branch Record aka LBR for debugging */
493 	ctrl->lbr_virt_en = 1;
494 	state->dbgctl = BIT(0);
495 
496 	/* EFER_SVM must always be set when the guest is executing */
497 	state->efer = EFER_SVM;
498 
499 	/* Set up the PAT to power-on state */
500 	state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK)	|
501 	    PAT_VALUE(1, PAT_WRITE_THROUGH)	|
502 	    PAT_VALUE(2, PAT_UNCACHED)		|
503 	    PAT_VALUE(3, PAT_UNCACHEABLE)	|
504 	    PAT_VALUE(4, PAT_WRITE_BACK)	|
505 	    PAT_VALUE(5, PAT_WRITE_THROUGH)	|
506 	    PAT_VALUE(6, PAT_UNCACHED)		|
507 	    PAT_VALUE(7, PAT_UNCACHEABLE);
508 }
509 
510 /*
511  * Initialize a virtual machine.
512  */
513 static void *
514 svm_vminit(struct vm *vm, pmap_t pmap)
515 {
516 	struct svm_softc *svm_sc;
517 	struct svm_vcpu *vcpu;
518 	vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
519 	int i;
520 
521 	svm_sc = malloc(sizeof (struct svm_softc), M_SVM, M_WAITOK | M_ZERO);
522 	svm_sc->vm = vm;
523 	svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4);
524 
525 	/*
526 	 * Intercept read and write accesses to all MSRs.
527 	 */
528 	memset(svm_sc->msr_bitmap, 0xFF, sizeof(svm_sc->msr_bitmap));
529 
530 	/*
531 	 * Access to the following MSRs is redirected to the VMCB when the
532 	 * guest is executing. Therefore it is safe to allow the guest to
533 	 * read/write these MSRs directly without hypervisor involvement.
534 	 */
535 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
536 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
537 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
538 
539 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
540 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
541 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
542 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
543 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
544 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
545 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
546 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
547 
548 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
549 
550 	/*
551 	 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
552 	 */
553 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
554 
555 	/* Intercept access to all I/O ports. */
556 	memset(svm_sc->iopm_bitmap, 0xFF, sizeof(svm_sc->iopm_bitmap));
557 
558 	iopm_pa = vtophys(svm_sc->iopm_bitmap);
559 	msrpm_pa = vtophys(svm_sc->msr_bitmap);
560 	pml4_pa = svm_sc->nptp;
561 	for (i = 0; i < VM_MAXCPU; i++) {
562 		vcpu = svm_get_vcpu(svm_sc, i);
563 		vcpu->nextrip = ~0;
564 		vcpu->lastcpu = NOCPU;
565 		vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
566 		vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
567 		svm_msr_guest_init(svm_sc, i);
568 	}
569 	return (svm_sc);
570 }
571 
572 /*
573  * Collateral for a generic SVM VM-exit.
574  */
575 static void
576 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
577 {
578 
579 	vme->exitcode = VM_EXITCODE_SVM;
580 	vme->u.svm.exitcode = code;
581 	vme->u.svm.exitinfo1 = info1;
582 	vme->u.svm.exitinfo2 = info2;
583 }
584 
585 static int
586 svm_cpl(struct vmcb_state *state)
587 {
588 
589 	/*
590 	 * From APMv2:
591 	 *   "Retrieve the CPL from the CPL field in the VMCB, not
592 	 *    from any segment DPL"
593 	 */
594 	return (state->cpl);
595 }
596 
597 static enum vm_cpu_mode
598 svm_vcpu_mode(struct vmcb *vmcb)
599 {
600 	struct vmcb_segment seg;
601 	struct vmcb_state *state;
602 	int error;
603 
604 	state = &vmcb->state;
605 
606 	if (state->efer & EFER_LMA) {
607 		error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
608 		KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
609 		    error));
610 
611 		/*
612 		 * Section 4.8.1 for APM2, check if Code Segment has
613 		 * Long attribute set in descriptor.
614 		 */
615 		if (seg.attrib & VMCB_CS_ATTRIB_L)
616 			return (CPU_MODE_64BIT);
617 		else
618 			return (CPU_MODE_COMPATIBILITY);
619 	} else  if (state->cr0 & CR0_PE) {
620 		return (CPU_MODE_PROTECTED);
621 	} else {
622 		return (CPU_MODE_REAL);
623 	}
624 }
625 
626 static enum vm_paging_mode
627 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
628 {
629 
630 	if ((cr0 & CR0_PG) == 0)
631 		return (PAGING_MODE_FLAT);
632 	if ((cr4 & CR4_PAE) == 0)
633 		return (PAGING_MODE_32);
634 	if (efer & EFER_LME)
635 		return (PAGING_MODE_64);
636 	else
637 		return (PAGING_MODE_PAE);
638 }
639 
640 /*
641  * ins/outs utility routines
642  */
643 static uint64_t
644 svm_inout_str_index(struct svm_regctx *regs, int in)
645 {
646 	uint64_t val;
647 
648 	val = in ? regs->sctx_rdi : regs->sctx_rsi;
649 
650 	return (val);
651 }
652 
653 static uint64_t
654 svm_inout_str_count(struct svm_regctx *regs, int rep)
655 {
656 	uint64_t val;
657 
658 	val = rep ? regs->sctx_rcx : 1;
659 
660 	return (val);
661 }
662 
663 static void
664 svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
665     int in, struct vm_inout_str *vis)
666 {
667 	int error, s;
668 
669 	if (in) {
670 		vis->seg_name = VM_REG_GUEST_ES;
671 	} else {
672 		/* The segment field has standard encoding */
673 		s = (info1 >> 10) & 0x7;
674 		vis->seg_name = vm_segment_name(s);
675 	}
676 
677 	error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
678 	KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
679 }
680 
681 static int
682 svm_inout_str_addrsize(uint64_t info1)
683 {
684         uint32_t size;
685 
686         size = (info1 >> 7) & 0x7;
687         switch (size) {
688         case 1:
689                 return (2);     /* 16 bit */
690         case 2:
691                 return (4);     /* 32 bit */
692         case 4:
693                 return (8);     /* 64 bit */
694         default:
695                 panic("%s: invalid size encoding %d", __func__, size);
696         }
697 }
698 
699 static void
700 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
701 {
702 	struct vmcb_state *state;
703 
704 	state = &vmcb->state;
705 	paging->cr3 = state->cr3;
706 	paging->cpl = svm_cpl(state);
707 	paging->cpu_mode = svm_vcpu_mode(vmcb);
708 	paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
709 	    state->efer);
710 }
711 
712 #define	UNHANDLED 0
713 
714 /*
715  * Handle guest I/O intercept.
716  */
717 static int
718 svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
719 {
720 	struct vmcb_ctrl *ctrl;
721 	struct vmcb_state *state;
722 	struct svm_regctx *regs;
723 	struct vm_inout_str *vis;
724 	uint64_t info1;
725 	int inout_string;
726 
727 	state = svm_get_vmcb_state(svm_sc, vcpu);
728 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
729 	regs  = svm_get_guest_regctx(svm_sc, vcpu);
730 
731 	info1 = ctrl->exitinfo1;
732 	inout_string = info1 & BIT(2) ? 1 : 0;
733 
734 	/*
735 	 * The effective segment number in EXITINFO1[12:10] is populated
736 	 * only if the processor has the DecodeAssist capability.
737 	 *
738 	 * XXX this is not specified explicitly in APMv2 but can be verified
739 	 * empirically.
740 	 */
741 	if (inout_string && !decode_assist())
742 		return (UNHANDLED);
743 
744 	vmexit->exitcode 	= VM_EXITCODE_INOUT;
745 	vmexit->u.inout.in 	= (info1 & BIT(0)) ? 1 : 0;
746 	vmexit->u.inout.string 	= inout_string;
747 	vmexit->u.inout.rep 	= (info1 & BIT(3)) ? 1 : 0;
748 	vmexit->u.inout.bytes 	= (info1 >> 4) & 0x7;
749 	vmexit->u.inout.port 	= (uint16_t)(info1 >> 16);
750 	vmexit->u.inout.eax 	= (uint32_t)(state->rax);
751 
752 	if (inout_string) {
753 		vmexit->exitcode = VM_EXITCODE_INOUT_STR;
754 		vis = &vmexit->u.inout_str;
755 		svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
756 		vis->rflags = state->rflags;
757 		vis->cr0 = state->cr0;
758 		vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
759 		vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
760 		vis->addrsize = svm_inout_str_addrsize(info1);
761 		svm_inout_str_seginfo(svm_sc, vcpu, info1,
762 		    vmexit->u.inout.in, vis);
763 	}
764 
765 	return (UNHANDLED);
766 }
767 
768 static int
769 npf_fault_type(uint64_t exitinfo1)
770 {
771 
772 	if (exitinfo1 & VMCB_NPF_INFO1_W)
773 		return (VM_PROT_WRITE);
774 	else if (exitinfo1 & VMCB_NPF_INFO1_ID)
775 		return (VM_PROT_EXECUTE);
776 	else
777 		return (VM_PROT_READ);
778 }
779 
780 static bool
781 svm_npf_emul_fault(uint64_t exitinfo1)
782 {
783 
784 	if (exitinfo1 & VMCB_NPF_INFO1_ID) {
785 		return (false);
786 	}
787 
788 	if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
789 		return (false);
790 	}
791 
792 	if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
793 		return (false);
794 	}
795 
796 	return (true);
797 }
798 
799 static void
800 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
801 {
802 	struct vm_guest_paging *paging;
803 	struct vmcb_segment seg;
804 	struct vmcb_ctrl *ctrl;
805 	char *inst_bytes;
806 	int error, inst_len;
807 
808 	ctrl = &vmcb->ctrl;
809 	paging = &vmexit->u.inst_emul.paging;
810 
811 	vmexit->exitcode = VM_EXITCODE_INST_EMUL;
812 	vmexit->u.inst_emul.gpa = gpa;
813 	vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
814 	svm_paging_info(vmcb, paging);
815 
816 	error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
817 	KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
818 
819 	switch(paging->cpu_mode) {
820 	case CPU_MODE_REAL:
821 		vmexit->u.inst_emul.cs_base = seg.base;
822 		vmexit->u.inst_emul.cs_d = 0;
823 		break;
824 	case CPU_MODE_PROTECTED:
825 	case CPU_MODE_COMPATIBILITY:
826 		vmexit->u.inst_emul.cs_base = seg.base;
827 
828 		/*
829 		 * Section 4.8.1 of APM2, Default Operand Size or D bit.
830 		 */
831 		vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
832 		    1 : 0;
833 		break;
834 	default:
835 		vmexit->u.inst_emul.cs_base = 0;
836 		vmexit->u.inst_emul.cs_d = 0;
837 		break;
838 	}
839 
840 	/*
841 	 * Copy the instruction bytes into 'vie' if available.
842 	 */
843 	if (decode_assist() && !disable_npf_assist) {
844 		inst_len = ctrl->inst_len;
845 		inst_bytes = ctrl->inst_bytes;
846 	} else {
847 		inst_len = 0;
848 		inst_bytes = NULL;
849 	}
850 	vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
851 }
852 
853 #ifdef KTR
854 static const char *
855 intrtype_to_str(int intr_type)
856 {
857 	switch (intr_type) {
858 	case VMCB_EVENTINJ_TYPE_INTR:
859 		return ("hwintr");
860 	case VMCB_EVENTINJ_TYPE_NMI:
861 		return ("nmi");
862 	case VMCB_EVENTINJ_TYPE_INTn:
863 		return ("swintr");
864 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
865 		return ("exception");
866 	default:
867 		panic("%s: unknown intr_type %d", __func__, intr_type);
868 	}
869 }
870 #endif
871 
872 /*
873  * Inject an event to vcpu as described in section 15.20, "Event injection".
874  */
875 static void
876 svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
877 		 uint32_t error, bool ec_valid)
878 {
879 	struct vmcb_ctrl *ctrl;
880 
881 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
882 
883 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
884 	    ("%s: event already pending %#lx", __func__, ctrl->eventinj));
885 
886 	KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
887 	    __func__, vector));
888 
889 	switch (intr_type) {
890 	case VMCB_EVENTINJ_TYPE_INTR:
891 	case VMCB_EVENTINJ_TYPE_NMI:
892 	case VMCB_EVENTINJ_TYPE_INTn:
893 		break;
894 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
895 		if (vector >= 0 && vector <= 31 && vector != 2)
896 			break;
897 		/* FALLTHROUGH */
898 	default:
899 		panic("%s: invalid intr_type/vector: %d/%d", __func__,
900 		    intr_type, vector);
901 	}
902 	ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
903 	if (ec_valid) {
904 		ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
905 		ctrl->eventinj |= (uint64_t)error << 32;
906 		VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
907 		    intrtype_to_str(intr_type), vector, error);
908 	} else {
909 		VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
910 		    intrtype_to_str(intr_type), vector);
911 	}
912 }
913 
914 static void
915 svm_update_virqinfo(struct svm_softc *sc, int vcpu)
916 {
917 	struct vm *vm;
918 	struct vlapic *vlapic;
919 	struct vmcb_ctrl *ctrl;
920 	int pending;
921 
922 	vm = sc->vm;
923 	vlapic = vm_lapic(vm, vcpu);
924 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
925 
926 	/* Update %cr8 in the emulated vlapic */
927 	vlapic_set_cr8(vlapic, ctrl->v_tpr);
928 
929 	/*
930 	 * If V_IRQ indicates that the interrupt injection attempted on then
931 	 * last VMRUN was successful then update the vlapic accordingly.
932 	 */
933 	if (ctrl->v_intr_vector != 0) {
934 		pending = ctrl->v_irq;
935 		KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid "
936 		    "v_intr_vector %d", __func__, ctrl->v_intr_vector));
937 		KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
938 		VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector,
939 		    pending ? "pending" : "accepted");
940 		if (!pending)
941 			vlapic_intr_accepted(vlapic, ctrl->v_intr_vector);
942 	}
943 }
944 
945 static void
946 svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
947 {
948 	struct vmcb_ctrl *ctrl;
949 	uint64_t intinfo;
950 
951 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
952 	intinfo = ctrl->exitintinfo;
953 	if (!VMCB_EXITINTINFO_VALID(intinfo))
954 		return;
955 
956 	/*
957 	 * From APMv2, Section "Intercepts during IDT interrupt delivery"
958 	 *
959 	 * If a #VMEXIT happened during event delivery then record the event
960 	 * that was being delivered.
961 	 */
962 	VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
963 		intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
964 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
965 	vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
966 }
967 
968 static __inline int
969 vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
970 {
971 
972 	return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
973 	    VMCB_INTCPT_VINTR));
974 }
975 
976 static __inline void
977 enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
978 {
979 	struct vmcb_ctrl *ctrl;
980 
981 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
982 
983 	if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
984 		KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
985 		KASSERT(vintr_intercept_enabled(sc, vcpu),
986 		    ("%s: vintr intercept should be enabled", __func__));
987 		return;
988 	}
989 
990 	VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
991 	ctrl->v_irq = 1;
992 	ctrl->v_ign_tpr = 1;
993 	ctrl->v_intr_vector = 0;
994 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
995 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
996 }
997 
998 static __inline void
999 disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1000 {
1001 	struct vmcb_ctrl *ctrl;
1002 
1003 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1004 
1005 	if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1006 		KASSERT(!vintr_intercept_enabled(sc, vcpu),
1007 		    ("%s: vintr intercept should be disabled", __func__));
1008 		return;
1009 	}
1010 
1011 #ifdef KTR
1012 	if (ctrl->v_intr_vector == 0)
1013 		VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
1014 	else
1015 		VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection");
1016 #endif
1017 	ctrl->v_irq = 0;
1018 	ctrl->v_intr_vector = 0;
1019 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1020 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1021 }
1022 
1023 static int
1024 svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
1025 {
1026 	struct vmcb_ctrl *ctrl;
1027 	int oldval, newval;
1028 
1029 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1030 	oldval = ctrl->intr_shadow;
1031 	newval = val ? 1 : 0;
1032 	if (newval != oldval) {
1033 		ctrl->intr_shadow = newval;
1034 		VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1035 	}
1036 	return (0);
1037 }
1038 
1039 static int
1040 svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1041 {
1042 	struct vmcb_ctrl *ctrl;
1043 
1044 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1045 	*val = ctrl->intr_shadow;
1046 	return (0);
1047 }
1048 
1049 /*
1050  * Once an NMI is injected it blocks delivery of further NMIs until the handler
1051  * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1052  * to track when the vcpu is done handling the NMI.
1053  */
1054 static int
1055 nmi_blocked(struct svm_softc *sc, int vcpu)
1056 {
1057 	int blocked;
1058 
1059 	blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1060 	    VMCB_INTCPT_IRET);
1061 	return (blocked);
1062 }
1063 
1064 static void
1065 enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1066 {
1067 
1068 	KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1069 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1070 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1071 }
1072 
1073 static void
1074 clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1075 {
1076 	int error;
1077 
1078 	KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1079 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1080 	/*
1081 	 * When the IRET intercept is cleared the vcpu will attempt to execute
1082 	 * the "iret" when it runs next. However, it is possible to inject
1083 	 * another NMI into the vcpu before the "iret" has actually executed.
1084 	 *
1085 	 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1086 	 * it will trap back into the hypervisor. If an NMI is pending for
1087 	 * the vcpu it will be injected into the guest.
1088 	 *
1089 	 * XXX this needs to be fixed
1090 	 */
1091 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1092 
1093 	/*
1094 	 * Set 'intr_shadow' to prevent an NMI from being injected on the
1095 	 * immediate VMRUN.
1096 	 */
1097 	error = svm_modify_intr_shadow(sc, vcpu, 1);
1098 	KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1099 }
1100 
1101 #define	EFER_MBZ_BITS	0xFFFFFFFFFFFF0200UL
1102 
1103 static int
1104 svm_write_efer(struct svm_softc *sc, int vcpu, uint64_t newval, bool *retu)
1105 {
1106 	struct vm_exit *vme;
1107 	struct vmcb_state *state;
1108 	uint64_t changed, lma, oldval;
1109 	int error;
1110 
1111 	state = svm_get_vmcb_state(sc, vcpu);
1112 
1113 	oldval = state->efer;
1114 	VCPU_CTR2(sc->vm, vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval);
1115 
1116 	newval &= ~0xFE;		/* clear the Read-As-Zero (RAZ) bits */
1117 	changed = oldval ^ newval;
1118 
1119 	if (newval & EFER_MBZ_BITS)
1120 		goto gpf;
1121 
1122 	/* APMv2 Table 14-5 "Long-Mode Consistency Checks" */
1123 	if (changed & EFER_LME) {
1124 		if (state->cr0 & CR0_PG)
1125 			goto gpf;
1126 	}
1127 
1128 	/* EFER.LMA = EFER.LME & CR0.PG */
1129 	if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0)
1130 		lma = EFER_LMA;
1131 	else
1132 		lma = 0;
1133 
1134 	if ((newval & EFER_LMA) != lma)
1135 		goto gpf;
1136 
1137 	if (newval & EFER_NXE) {
1138 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_NO_EXECUTE))
1139 			goto gpf;
1140 	}
1141 
1142 	/*
1143 	 * XXX bhyve does not enforce segment limits in 64-bit mode. Until
1144 	 * this is fixed flag guest attempt to set EFER_LMSLE as an error.
1145 	 */
1146 	if (newval & EFER_LMSLE) {
1147 		vme = vm_exitinfo(sc->vm, vcpu);
1148 		vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0);
1149 		*retu = true;
1150 		return (0);
1151 	}
1152 
1153 	if (newval & EFER_FFXSR) {
1154 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_FFXSR))
1155 			goto gpf;
1156 	}
1157 
1158 	if (newval & EFER_TCE) {
1159 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_TCE))
1160 			goto gpf;
1161 	}
1162 
1163 	error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, newval);
1164 	KASSERT(error == 0, ("%s: error %d updating efer", __func__, error));
1165 	return (0);
1166 gpf:
1167 	vm_inject_gp(sc->vm, vcpu);
1168 	return (0);
1169 }
1170 
1171 static int
1172 emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1173     bool *retu)
1174 {
1175 	int error;
1176 
1177 	if (lapic_msr(num))
1178 		error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1179 	else if (num == MSR_EFER)
1180 		error = svm_write_efer(sc, vcpu, val, retu);
1181 	else
1182 		error = svm_wrmsr(sc, vcpu, num, val, retu);
1183 
1184 	return (error);
1185 }
1186 
1187 static int
1188 emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1189 {
1190 	struct vmcb_state *state;
1191 	struct svm_regctx *ctx;
1192 	uint64_t result;
1193 	int error;
1194 
1195 	if (lapic_msr(num))
1196 		error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1197 	else
1198 		error = svm_rdmsr(sc, vcpu, num, &result, retu);
1199 
1200 	if (error == 0) {
1201 		state = svm_get_vmcb_state(sc, vcpu);
1202 		ctx = svm_get_guest_regctx(sc, vcpu);
1203 		state->rax = result & 0xffffffff;
1204 		ctx->sctx_rdx = result >> 32;
1205 	}
1206 
1207 	return (error);
1208 }
1209 
1210 #ifdef KTR
1211 static const char *
1212 exit_reason_to_str(uint64_t reason)
1213 {
1214 	static char reasonbuf[32];
1215 
1216 	switch (reason) {
1217 	case VMCB_EXIT_INVALID:
1218 		return ("invalvmcb");
1219 	case VMCB_EXIT_SHUTDOWN:
1220 		return ("shutdown");
1221 	case VMCB_EXIT_NPF:
1222 		return ("nptfault");
1223 	case VMCB_EXIT_PAUSE:
1224 		return ("pause");
1225 	case VMCB_EXIT_HLT:
1226 		return ("hlt");
1227 	case VMCB_EXIT_CPUID:
1228 		return ("cpuid");
1229 	case VMCB_EXIT_IO:
1230 		return ("inout");
1231 	case VMCB_EXIT_MC:
1232 		return ("mchk");
1233 	case VMCB_EXIT_INTR:
1234 		return ("extintr");
1235 	case VMCB_EXIT_NMI:
1236 		return ("nmi");
1237 	case VMCB_EXIT_VINTR:
1238 		return ("vintr");
1239 	case VMCB_EXIT_MSR:
1240 		return ("msr");
1241 	case VMCB_EXIT_IRET:
1242 		return ("iret");
1243 	case VMCB_EXIT_MONITOR:
1244 		return ("monitor");
1245 	case VMCB_EXIT_MWAIT:
1246 		return ("mwait");
1247 	default:
1248 		snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1249 		return (reasonbuf);
1250 	}
1251 }
1252 #endif	/* KTR */
1253 
1254 /*
1255  * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1256  * that are due to instruction intercepts as well as MSR and IOIO intercepts
1257  * and exceptions caused by INT3, INTO and BOUND instructions.
1258  *
1259  * Return 1 if the nRIP is valid and 0 otherwise.
1260  */
1261 static int
1262 nrip_valid(uint64_t exitcode)
1263 {
1264 	switch (exitcode) {
1265 	case 0x00 ... 0x0F:	/* read of CR0 through CR15 */
1266 	case 0x10 ... 0x1F:	/* write of CR0 through CR15 */
1267 	case 0x20 ... 0x2F:	/* read of DR0 through DR15 */
1268 	case 0x30 ... 0x3F:	/* write of DR0 through DR15 */
1269 	case 0x43:		/* INT3 */
1270 	case 0x44:		/* INTO */
1271 	case 0x45:		/* BOUND */
1272 	case 0x65 ... 0x7C:	/* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1273 	case 0x80 ... 0x8D:	/* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1274 		return (1);
1275 	default:
1276 		return (0);
1277 	}
1278 }
1279 
1280 static int
1281 svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1282 {
1283 	struct vmcb *vmcb;
1284 	struct vmcb_state *state;
1285 	struct vmcb_ctrl *ctrl;
1286 	struct svm_regctx *ctx;
1287 	uint64_t code, info1, info2, val;
1288 	uint32_t eax, ecx, edx;
1289 	int error, errcode_valid, handled, idtvec, reflect;
1290 	bool retu;
1291 
1292 	ctx = svm_get_guest_regctx(svm_sc, vcpu);
1293 	vmcb = svm_get_vmcb(svm_sc, vcpu);
1294 	state = &vmcb->state;
1295 	ctrl = &vmcb->ctrl;
1296 
1297 	handled = 0;
1298 	code = ctrl->exitcode;
1299 	info1 = ctrl->exitinfo1;
1300 	info2 = ctrl->exitinfo2;
1301 
1302 	vmexit->exitcode = VM_EXITCODE_BOGUS;
1303 	vmexit->rip = state->rip;
1304 	vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1305 
1306 	vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1307 
1308 	/*
1309 	 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1310 	 * in an inconsistent state and can trigger assertions that would
1311 	 * never happen otherwise.
1312 	 */
1313 	if (code == VMCB_EXIT_INVALID) {
1314 		vm_exit_svm(vmexit, code, info1, info2);
1315 		return (0);
1316 	}
1317 
1318 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1319 	    "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1320 
1321 	KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1322 	    ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1323 	    vmexit->inst_length, code, info1, info2));
1324 
1325 	svm_update_virqinfo(svm_sc, vcpu);
1326 	svm_save_intinfo(svm_sc, vcpu);
1327 
1328 	switch (code) {
1329 	case VMCB_EXIT_IRET:
1330 		/*
1331 		 * Restart execution at "iret" but with the intercept cleared.
1332 		 */
1333 		vmexit->inst_length = 0;
1334 		clear_nmi_blocking(svm_sc, vcpu);
1335 		handled = 1;
1336 		break;
1337 	case VMCB_EXIT_VINTR:	/* interrupt window exiting */
1338 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1339 		handled = 1;
1340 		break;
1341 	case VMCB_EXIT_INTR:	/* external interrupt */
1342 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1343 		handled = 1;
1344 		break;
1345 	case VMCB_EXIT_NMI:	/* external NMI */
1346 		handled = 1;
1347 		break;
1348 	case 0x40 ... 0x5F:
1349 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1350 		reflect = 1;
1351 		idtvec = code - 0x40;
1352 		switch (idtvec) {
1353 		case IDT_MC:
1354 			/*
1355 			 * Call the machine check handler by hand. Also don't
1356 			 * reflect the machine check back into the guest.
1357 			 */
1358 			reflect = 0;
1359 			VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler");
1360 			__asm __volatile("int $18");
1361 			break;
1362 		case IDT_PF:
1363 			error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2,
1364 			    info2);
1365 			KASSERT(error == 0, ("%s: error %d updating cr2",
1366 			    __func__, error));
1367 			/* fallthru */
1368 		case IDT_NP:
1369 		case IDT_SS:
1370 		case IDT_GP:
1371 		case IDT_AC:
1372 		case IDT_TS:
1373 			errcode_valid = 1;
1374 			break;
1375 
1376 		case IDT_DF:
1377 			errcode_valid = 1;
1378 			info1 = 0;
1379 			break;
1380 
1381 		case IDT_BP:
1382 		case IDT_OF:
1383 		case IDT_BR:
1384 			/*
1385 			 * The 'nrip' field is populated for INT3, INTO and
1386 			 * BOUND exceptions and this also implies that
1387 			 * 'inst_length' is non-zero.
1388 			 *
1389 			 * Reset 'inst_length' to zero so the guest %rip at
1390 			 * event injection is identical to what it was when
1391 			 * the exception originally happened.
1392 			 */
1393 			VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d "
1394 			    "to zero before injecting exception %d",
1395 			    vmexit->inst_length, idtvec);
1396 			vmexit->inst_length = 0;
1397 			/* fallthru */
1398 		default:
1399 			errcode_valid = 0;
1400 			info1 = 0;
1401 			break;
1402 		}
1403 		KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1404 		    "when reflecting exception %d into guest",
1405 		    vmexit->inst_length, idtvec));
1406 
1407 		if (reflect) {
1408 			/* Reflect the exception back into the guest */
1409 			VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception "
1410 			    "%d/%#x into the guest", idtvec, (int)info1);
1411 			error = vm_inject_exception(svm_sc->vm, vcpu, idtvec,
1412 			    errcode_valid, info1, 0);
1413 			KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1414 			    __func__, error));
1415 		}
1416 		handled = 1;
1417 		break;
1418 	case VMCB_EXIT_MSR:	/* MSR access. */
1419 		eax = state->rax;
1420 		ecx = ctx->sctx_rcx;
1421 		edx = ctx->sctx_rdx;
1422 		retu = false;
1423 
1424 		if (info1) {
1425 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1426 			val = (uint64_t)edx << 32 | eax;
1427 			VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1428 			    ecx, val);
1429 			if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1430 				vmexit->exitcode = VM_EXITCODE_WRMSR;
1431 				vmexit->u.msr.code = ecx;
1432 				vmexit->u.msr.wval = val;
1433 			} else if (!retu) {
1434 				handled = 1;
1435 			} else {
1436 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1437 				    ("emulate_wrmsr retu with bogus exitcode"));
1438 			}
1439 		} else {
1440 			VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1441 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1442 			if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1443 				vmexit->exitcode = VM_EXITCODE_RDMSR;
1444 				vmexit->u.msr.code = ecx;
1445 			} else if (!retu) {
1446 				handled = 1;
1447 			} else {
1448 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1449 				    ("emulate_rdmsr retu with bogus exitcode"));
1450 			}
1451 		}
1452 		break;
1453 	case VMCB_EXIT_IO:
1454 		handled = svm_handle_io(svm_sc, vcpu, vmexit);
1455 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1456 		break;
1457 	case VMCB_EXIT_CPUID:
1458 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1459 		handled = x86_emulate_cpuid(svm_sc->vm, vcpu,
1460 		    (uint32_t *)&state->rax,
1461 		    (uint32_t *)&ctx->sctx_rbx,
1462 		    (uint32_t *)&ctx->sctx_rcx,
1463 		    (uint32_t *)&ctx->sctx_rdx);
1464 		break;
1465 	case VMCB_EXIT_HLT:
1466 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1467 		vmexit->exitcode = VM_EXITCODE_HLT;
1468 		vmexit->u.hlt.rflags = state->rflags;
1469 		break;
1470 	case VMCB_EXIT_PAUSE:
1471 		vmexit->exitcode = VM_EXITCODE_PAUSE;
1472 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1473 		break;
1474 	case VMCB_EXIT_NPF:
1475 		/* EXITINFO2 contains the faulting guest physical address */
1476 		if (info1 & VMCB_NPF_INFO1_RSV) {
1477 			VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1478 			    "reserved bits set: info1(%#lx) info2(%#lx)",
1479 			    info1, info2);
1480 		} else if (vm_mem_allocated(svm_sc->vm, vcpu, info2)) {
1481 			vmexit->exitcode = VM_EXITCODE_PAGING;
1482 			vmexit->u.paging.gpa = info2;
1483 			vmexit->u.paging.fault_type = npf_fault_type(info1);
1484 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1485 			VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1486 			    "on gpa %#lx/%#lx at rip %#lx",
1487 			    info2, info1, state->rip);
1488 		} else if (svm_npf_emul_fault(info1)) {
1489 			svm_handle_inst_emul(vmcb, info2, vmexit);
1490 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1491 			VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1492 			    "for gpa %#lx/%#lx at rip %#lx",
1493 			    info2, info1, state->rip);
1494 		}
1495 		break;
1496 	case VMCB_EXIT_MONITOR:
1497 		vmexit->exitcode = VM_EXITCODE_MONITOR;
1498 		break;
1499 	case VMCB_EXIT_MWAIT:
1500 		vmexit->exitcode = VM_EXITCODE_MWAIT;
1501 		break;
1502 	default:
1503 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1504 		break;
1505 	}
1506 
1507 	VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1508 	    handled ? "handled" : "unhandled", exit_reason_to_str(code),
1509 	    vmexit->rip, vmexit->inst_length);
1510 
1511 	if (handled) {
1512 		vmexit->rip += vmexit->inst_length;
1513 		vmexit->inst_length = 0;
1514 		state->rip = vmexit->rip;
1515 	} else {
1516 		if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1517 			/*
1518 			 * If this VM exit was not claimed by anybody then
1519 			 * treat it as a generic SVM exit.
1520 			 */
1521 			vm_exit_svm(vmexit, code, info1, info2);
1522 		} else {
1523 			/*
1524 			 * The exitcode and collateral have been populated.
1525 			 * The VM exit will be processed further in userland.
1526 			 */
1527 		}
1528 	}
1529 	return (handled);
1530 }
1531 
1532 static void
1533 svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1534 {
1535 	uint64_t intinfo;
1536 
1537 	if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1538 		return;
1539 
1540 	KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1541 	    "valid: %#lx", __func__, intinfo));
1542 
1543 	svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1544 		VMCB_EXITINTINFO_VECTOR(intinfo),
1545 		VMCB_EXITINTINFO_EC(intinfo),
1546 		VMCB_EXITINTINFO_EC_VALID(intinfo));
1547 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1548 	VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1549 }
1550 
1551 /*
1552  * Inject event to virtual cpu.
1553  */
1554 static void
1555 svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1556 {
1557 	struct vmcb_ctrl *ctrl;
1558 	struct vmcb_state *state;
1559 	struct svm_vcpu *vcpustate;
1560 	uint8_t v_tpr;
1561 	int vector, need_intr_window, pending_apic_vector;
1562 
1563 	state = svm_get_vmcb_state(sc, vcpu);
1564 	ctrl  = svm_get_vmcb_ctrl(sc, vcpu);
1565 	vcpustate = svm_get_vcpu(sc, vcpu);
1566 
1567 	need_intr_window = 0;
1568 	pending_apic_vector = 0;
1569 
1570 	if (vcpustate->nextrip != state->rip) {
1571 		ctrl->intr_shadow = 0;
1572 		VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking "
1573 		    "cleared due to rip change: %#lx/%#lx",
1574 		    vcpustate->nextrip, state->rip);
1575 	}
1576 
1577 	/*
1578 	 * Inject pending events or exceptions for this vcpu.
1579 	 *
1580 	 * An event might be pending because the previous #VMEXIT happened
1581 	 * during event delivery (i.e. ctrl->exitintinfo).
1582 	 *
1583 	 * An event might also be pending because an exception was injected
1584 	 * by the hypervisor (e.g. #PF during instruction emulation).
1585 	 */
1586 	svm_inj_intinfo(sc, vcpu);
1587 
1588 	/* NMI event has priority over interrupts. */
1589 	if (vm_nmi_pending(sc->vm, vcpu)) {
1590 		if (nmi_blocked(sc, vcpu)) {
1591 			/*
1592 			 * Can't inject another NMI if the guest has not
1593 			 * yet executed an "iret" after the last NMI.
1594 			 */
1595 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1596 			    "to NMI-blocking");
1597 		} else if (ctrl->intr_shadow) {
1598 			/*
1599 			 * Can't inject an NMI if the vcpu is in an intr_shadow.
1600 			 */
1601 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1602 			    "interrupt shadow");
1603 			need_intr_window = 1;
1604 			goto done;
1605 		} else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1606 			/*
1607 			 * If there is already an exception/interrupt pending
1608 			 * then defer the NMI until after that.
1609 			 */
1610 			VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1611 			    "eventinj %#lx", ctrl->eventinj);
1612 
1613 			/*
1614 			 * Use self-IPI to trigger a VM-exit as soon as
1615 			 * possible after the event injection is completed.
1616 			 *
1617 			 * This works only if the external interrupt exiting
1618 			 * is at a lower priority than the event injection.
1619 			 *
1620 			 * Although not explicitly specified in APMv2 the
1621 			 * relative priorities were verified empirically.
1622 			 */
1623 			ipi_cpu(curcpu, IPI_AST);	/* XXX vmm_ipinum? */
1624 		} else {
1625 			vm_nmi_clear(sc->vm, vcpu);
1626 
1627 			/* Inject NMI, vector number is not used */
1628 			svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1629 			    IDT_NMI, 0, false);
1630 
1631 			/* virtual NMI blocking is now in effect */
1632 			enable_nmi_blocking(sc, vcpu);
1633 
1634 			VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1635 		}
1636 	}
1637 
1638 	if (!vm_extint_pending(sc->vm, vcpu)) {
1639 		/*
1640 		 * APIC interrupts are delivered using the V_IRQ offload.
1641 		 *
1642 		 * The primary benefit is that the hypervisor doesn't need to
1643 		 * deal with the various conditions that inhibit interrupts.
1644 		 * It also means that TPR changes via CR8 will be handled
1645 		 * without any hypervisor involvement.
1646 		 *
1647 		 * Note that the APIC vector must remain pending in the vIRR
1648 		 * until it is confirmed that it was delivered to the guest.
1649 		 * This can be confirmed based on the value of V_IRQ at the
1650 		 * next #VMEXIT (1 = pending, 0 = delivered).
1651 		 *
1652 		 * Also note that it is possible that another higher priority
1653 		 * vector can become pending before this vector is delivered
1654 		 * to the guest. This is alright because vcpu_notify_event()
1655 		 * will send an IPI and force the vcpu to trap back into the
1656 		 * hypervisor. The higher priority vector will be injected on
1657 		 * the next VMRUN.
1658 		 */
1659 		if (vlapic_pending_intr(vlapic, &vector)) {
1660 			KASSERT(vector >= 16 && vector <= 255,
1661 			    ("invalid vector %d from local APIC", vector));
1662 			pending_apic_vector = vector;
1663 		}
1664 		goto done;
1665 	}
1666 
1667 	/* Ask the legacy pic for a vector to inject */
1668 	vatpic_pending_intr(sc->vm, &vector);
1669 	KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR",
1670 	    vector));
1671 
1672 	/*
1673 	 * If the guest has disabled interrupts or is in an interrupt shadow
1674 	 * then we cannot inject the pending interrupt.
1675 	 */
1676 	if ((state->rflags & PSL_I) == 0) {
1677 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1678 		    "rflags %#lx", vector, state->rflags);
1679 		need_intr_window = 1;
1680 		goto done;
1681 	}
1682 
1683 	if (ctrl->intr_shadow) {
1684 		VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1685 		    "interrupt shadow", vector);
1686 		need_intr_window = 1;
1687 		goto done;
1688 	}
1689 
1690 	if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1691 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1692 		    "eventinj %#lx", vector, ctrl->eventinj);
1693 		need_intr_window = 1;
1694 		goto done;
1695 	}
1696 
1697 	/*
1698 	 * Legacy PIC interrupts are delivered via the event injection
1699 	 * mechanism.
1700 	 */
1701 	svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1702 
1703 	vm_extint_clear(sc->vm, vcpu);
1704 	vatpic_intr_accepted(sc->vm, vector);
1705 
1706 	/*
1707 	 * Force a VM-exit as soon as the vcpu is ready to accept another
1708 	 * interrupt. This is done because the PIC might have another vector
1709 	 * that it wants to inject. Also, if the APIC has a pending interrupt
1710 	 * that was preempted by the ExtInt then it allows us to inject the
1711 	 * APIC vector as soon as possible.
1712 	 */
1713 	need_intr_window = 1;
1714 done:
1715 	/*
1716 	 * The guest can modify the TPR by writing to %CR8. In guest mode
1717 	 * the processor reflects this write to V_TPR without hypervisor
1718 	 * intervention.
1719 	 *
1720 	 * The guest can also modify the TPR by writing to it via the memory
1721 	 * mapped APIC page. In this case, the write will be emulated by the
1722 	 * hypervisor. For this reason V_TPR must be updated before every
1723 	 * VMRUN.
1724 	 */
1725 	v_tpr = vlapic_get_cr8(vlapic);
1726 	KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1727 	if (ctrl->v_tpr != v_tpr) {
1728 		VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1729 		    ctrl->v_tpr, v_tpr);
1730 		ctrl->v_tpr = v_tpr;
1731 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1732 	}
1733 
1734 	if (pending_apic_vector) {
1735 		/*
1736 		 * If an APIC vector is being injected then interrupt window
1737 		 * exiting is not possible on this VMRUN.
1738 		 */
1739 		KASSERT(!need_intr_window, ("intr_window exiting impossible"));
1740 		VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ",
1741 		    pending_apic_vector);
1742 
1743 		ctrl->v_irq = 1;
1744 		ctrl->v_ign_tpr = 0;
1745 		ctrl->v_intr_vector = pending_apic_vector;
1746 		ctrl->v_intr_prio = pending_apic_vector >> 4;
1747 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1748 	} else if (need_intr_window) {
1749 		/*
1750 		 * We use V_IRQ in conjunction with the VINTR intercept to
1751 		 * trap into the hypervisor as soon as a virtual interrupt
1752 		 * can be delivered.
1753 		 *
1754 		 * Since injected events are not subject to intercept checks
1755 		 * we need to ensure that the V_IRQ is not actually going to
1756 		 * be delivered on VM entry. The KASSERT below enforces this.
1757 		 */
1758 		KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1759 		    (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1760 		    ("Bogus intr_window_exiting: eventinj (%#lx), "
1761 		    "intr_shadow (%u), rflags (%#lx)",
1762 		    ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1763 		enable_intr_window_exiting(sc, vcpu);
1764 	} else {
1765 		disable_intr_window_exiting(sc, vcpu);
1766 	}
1767 }
1768 
1769 static __inline void
1770 restore_host_tss(void)
1771 {
1772 	struct system_segment_descriptor *tss_sd;
1773 
1774 	/*
1775 	 * The TSS descriptor was in use prior to launching the guest so it
1776 	 * has been marked busy.
1777 	 *
1778 	 * 'ltr' requires the descriptor to be marked available so change the
1779 	 * type to "64-bit available TSS".
1780 	 */
1781 	tss_sd = PCPU_GET(tss);
1782 	tss_sd->sd_type = SDT_SYSTSS;
1783 	ltr(GSEL(GPROC0_SEL, SEL_KPL));
1784 }
1785 
1786 static void
1787 check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu)
1788 {
1789 	struct svm_vcpu *vcpustate;
1790 	struct vmcb_ctrl *ctrl;
1791 	long eptgen;
1792 	bool alloc_asid;
1793 
1794 	KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not "
1795 	    "active on cpu %u", __func__, thiscpu));
1796 
1797 	vcpustate = svm_get_vcpu(sc, vcpuid);
1798 	ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1799 
1800 	/*
1801 	 * The TLB entries associated with the vcpu's ASID are not valid
1802 	 * if either of the following conditions is true:
1803 	 *
1804 	 * 1. The vcpu's ASID generation is different than the host cpu's
1805 	 *    ASID generation. This happens when the vcpu migrates to a new
1806 	 *    host cpu. It can also happen when the number of vcpus executing
1807 	 *    on a host cpu is greater than the number of ASIDs available.
1808 	 *
1809 	 * 2. The pmap generation number is different than the value cached in
1810 	 *    the 'vcpustate'. This happens when the host invalidates pages
1811 	 *    belonging to the guest.
1812 	 *
1813 	 *	asidgen		eptgen	      Action
1814 	 *	mismatch	mismatch
1815 	 *	   0		   0		(a)
1816 	 *	   0		   1		(b1) or (b2)
1817 	 *	   1		   0		(c)
1818 	 *	   1		   1		(d)
1819 	 *
1820 	 * (a) There is no mismatch in eptgen or ASID generation and therefore
1821 	 *     no further action is needed.
1822 	 *
1823 	 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1824 	 *      retained and the TLB entries associated with this ASID
1825 	 *      are flushed by VMRUN.
1826 	 *
1827 	 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1828 	 *      allocated.
1829 	 *
1830 	 * (c) A new ASID is allocated.
1831 	 *
1832 	 * (d) A new ASID is allocated.
1833 	 */
1834 
1835 	alloc_asid = false;
1836 	eptgen = pmap->pm_eptgen;
1837 	ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1838 
1839 	if (vcpustate->asid.gen != asid[thiscpu].gen) {
1840 		alloc_asid = true;	/* (c) and (d) */
1841 	} else if (vcpustate->eptgen != eptgen) {
1842 		if (flush_by_asid())
1843 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;	/* (b1) */
1844 		else
1845 			alloc_asid = true;			/* (b2) */
1846 	} else {
1847 		/*
1848 		 * This is the common case (a).
1849 		 */
1850 		KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1851 		KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1852 		    ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1853 	}
1854 
1855 	if (alloc_asid) {
1856 		if (++asid[thiscpu].num >= nasid) {
1857 			asid[thiscpu].num = 1;
1858 			if (++asid[thiscpu].gen == 0)
1859 				asid[thiscpu].gen = 1;
1860 			/*
1861 			 * If this cpu does not support "flush-by-asid"
1862 			 * then flush the entire TLB on a generation
1863 			 * bump. Subsequent ASID allocation in this
1864 			 * generation can be done without a TLB flush.
1865 			 */
1866 			if (!flush_by_asid())
1867 				ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1868 		}
1869 		vcpustate->asid.gen = asid[thiscpu].gen;
1870 		vcpustate->asid.num = asid[thiscpu].num;
1871 
1872 		ctrl->asid = vcpustate->asid.num;
1873 		svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1874 		/*
1875 		 * If this cpu supports "flush-by-asid" then the TLB
1876 		 * was not flushed after the generation bump. The TLB
1877 		 * is flushed selectively after every new ASID allocation.
1878 		 */
1879 		if (flush_by_asid())
1880 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1881 	}
1882 	vcpustate->eptgen = eptgen;
1883 
1884 	KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1885 	KASSERT(ctrl->asid == vcpustate->asid.num,
1886 	    ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1887 }
1888 
1889 static __inline void
1890 disable_gintr(void)
1891 {
1892 
1893 	__asm __volatile("clgi");
1894 }
1895 
1896 static __inline void
1897 enable_gintr(void)
1898 {
1899 
1900         __asm __volatile("stgi");
1901 }
1902 
1903 /*
1904  * Start vcpu with specified RIP.
1905  */
1906 static int
1907 svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap,
1908 	struct vm_eventinfo *evinfo)
1909 {
1910 	struct svm_regctx *gctx;
1911 	struct svm_softc *svm_sc;
1912 	struct svm_vcpu *vcpustate;
1913 	struct vmcb_state *state;
1914 	struct vmcb_ctrl *ctrl;
1915 	struct vm_exit *vmexit;
1916 	struct vlapic *vlapic;
1917 	struct vm *vm;
1918 	uint64_t vmcb_pa;
1919 	int handled;
1920 
1921 	svm_sc = arg;
1922 	vm = svm_sc->vm;
1923 
1924 	vcpustate = svm_get_vcpu(svm_sc, vcpu);
1925 	state = svm_get_vmcb_state(svm_sc, vcpu);
1926 	ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
1927 	vmexit = vm_exitinfo(vm, vcpu);
1928 	vlapic = vm_lapic(vm, vcpu);
1929 
1930 	gctx = svm_get_guest_regctx(svm_sc, vcpu);
1931 	vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
1932 
1933 	if (vcpustate->lastcpu != curcpu) {
1934 		/*
1935 		 * Force new ASID allocation by invalidating the generation.
1936 		 */
1937 		vcpustate->asid.gen = 0;
1938 
1939 		/*
1940 		 * Invalidate the VMCB state cache by marking all fields dirty.
1941 		 */
1942 		svm_set_dirty(svm_sc, vcpu, 0xffffffff);
1943 
1944 		/*
1945 		 * XXX
1946 		 * Setting 'vcpustate->lastcpu' here is bit premature because
1947 		 * we may return from this function without actually executing
1948 		 * the VMRUN  instruction. This could happen if a rendezvous
1949 		 * or an AST is pending on the first time through the loop.
1950 		 *
1951 		 * This works for now but any new side-effects of vcpu
1952 		 * migration should take this case into account.
1953 		 */
1954 		vcpustate->lastcpu = curcpu;
1955 		vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
1956 	}
1957 
1958 	svm_msr_guest_enter(svm_sc, vcpu);
1959 
1960 	/* Update Guest RIP */
1961 	state->rip = rip;
1962 
1963 	do {
1964 		/*
1965 		 * Disable global interrupts to guarantee atomicity during
1966 		 * loading of guest state. This includes not only the state
1967 		 * loaded by the "vmrun" instruction but also software state
1968 		 * maintained by the hypervisor: suspended and rendezvous
1969 		 * state, NPT generation number, vlapic interrupts etc.
1970 		 */
1971 		disable_gintr();
1972 
1973 		if (vcpu_suspended(evinfo)) {
1974 			enable_gintr();
1975 			vm_exit_suspended(vm, vcpu, state->rip);
1976 			break;
1977 		}
1978 
1979 		if (vcpu_rendezvous_pending(evinfo)) {
1980 			enable_gintr();
1981 			vm_exit_rendezvous(vm, vcpu, state->rip);
1982 			break;
1983 		}
1984 
1985 		if (vcpu_reqidle(evinfo)) {
1986 			enable_gintr();
1987 			vm_exit_reqidle(vm, vcpu, state->rip);
1988 			break;
1989 		}
1990 
1991 		/* We are asked to give the cpu by scheduler. */
1992 		if (vcpu_should_yield(vm, vcpu)) {
1993 			enable_gintr();
1994 			vm_exit_astpending(vm, vcpu, state->rip);
1995 			break;
1996 		}
1997 
1998 		svm_inj_interrupts(svm_sc, vcpu, vlapic);
1999 
2000 		/* Activate the nested pmap on 'curcpu' */
2001 		CPU_SET_ATOMIC_ACQ(curcpu, &pmap->pm_active);
2002 
2003 		/*
2004 		 * Check the pmap generation and the ASID generation to
2005 		 * ensure that the vcpu does not use stale TLB mappings.
2006 		 */
2007 		check_asid(svm_sc, vcpu, pmap, curcpu);
2008 
2009 		ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
2010 		vcpustate->dirty = 0;
2011 		VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
2012 
2013 		/* Launch Virtual Machine. */
2014 		VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
2015 		svm_launch(vmcb_pa, gctx, &__pcpu[curcpu]);
2016 
2017 		CPU_CLR_ATOMIC(curcpu, &pmap->pm_active);
2018 
2019 		/*
2020 		 * The host GDTR and IDTR is saved by VMRUN and restored
2021 		 * automatically on #VMEXIT. However, the host TSS needs
2022 		 * to be restored explicitly.
2023 		 */
2024 		restore_host_tss();
2025 
2026 		/* #VMEXIT disables interrupts so re-enable them here. */
2027 		enable_gintr();
2028 
2029 		/* Update 'nextrip' */
2030 		vcpustate->nextrip = state->rip;
2031 
2032 		/* Handle #VMEXIT and if required return to user space. */
2033 		handled = svm_vmexit(svm_sc, vcpu, vmexit);
2034 	} while (handled);
2035 
2036 	svm_msr_guest_exit(svm_sc, vcpu);
2037 
2038 	return (0);
2039 }
2040 
2041 static void
2042 svm_vmcleanup(void *arg)
2043 {
2044 	struct svm_softc *sc = arg;
2045 
2046 	free(sc, M_SVM);
2047 }
2048 
2049 static register_t *
2050 swctx_regptr(struct svm_regctx *regctx, int reg)
2051 {
2052 
2053 	switch (reg) {
2054 	case VM_REG_GUEST_RBX:
2055 		return (&regctx->sctx_rbx);
2056 	case VM_REG_GUEST_RCX:
2057 		return (&regctx->sctx_rcx);
2058 	case VM_REG_GUEST_RDX:
2059 		return (&regctx->sctx_rdx);
2060 	case VM_REG_GUEST_RDI:
2061 		return (&regctx->sctx_rdi);
2062 	case VM_REG_GUEST_RSI:
2063 		return (&regctx->sctx_rsi);
2064 	case VM_REG_GUEST_RBP:
2065 		return (&regctx->sctx_rbp);
2066 	case VM_REG_GUEST_R8:
2067 		return (&regctx->sctx_r8);
2068 	case VM_REG_GUEST_R9:
2069 		return (&regctx->sctx_r9);
2070 	case VM_REG_GUEST_R10:
2071 		return (&regctx->sctx_r10);
2072 	case VM_REG_GUEST_R11:
2073 		return (&regctx->sctx_r11);
2074 	case VM_REG_GUEST_R12:
2075 		return (&regctx->sctx_r12);
2076 	case VM_REG_GUEST_R13:
2077 		return (&regctx->sctx_r13);
2078 	case VM_REG_GUEST_R14:
2079 		return (&regctx->sctx_r14);
2080 	case VM_REG_GUEST_R15:
2081 		return (&regctx->sctx_r15);
2082 	default:
2083 		return (NULL);
2084 	}
2085 }
2086 
2087 static int
2088 svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
2089 {
2090 	struct svm_softc *svm_sc;
2091 	register_t *reg;
2092 
2093 	svm_sc = arg;
2094 
2095 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2096 		return (svm_get_intr_shadow(svm_sc, vcpu, val));
2097 	}
2098 
2099 	if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2100 		return (0);
2101 	}
2102 
2103 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2104 
2105 	if (reg != NULL) {
2106 		*val = *reg;
2107 		return (0);
2108 	}
2109 
2110 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident);
2111 	return (EINVAL);
2112 }
2113 
2114 static int
2115 svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2116 {
2117 	struct svm_softc *svm_sc;
2118 	register_t *reg;
2119 
2120 	svm_sc = arg;
2121 
2122 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2123 		return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2124 	}
2125 
2126 	if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2127 		return (0);
2128 	}
2129 
2130 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2131 
2132 	if (reg != NULL) {
2133 		*reg = val;
2134 		return (0);
2135 	}
2136 
2137 	/*
2138 	 * XXX deal with CR3 and invalidate TLB entries tagged with the
2139 	 * vcpu's ASID. This needs to be treated differently depending on
2140 	 * whether 'running' is true/false.
2141 	 */
2142 
2143 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident);
2144 	return (EINVAL);
2145 }
2146 
2147 static int
2148 svm_setcap(void *arg, int vcpu, int type, int val)
2149 {
2150 	struct svm_softc *sc;
2151 	int error;
2152 
2153 	sc = arg;
2154 	error = 0;
2155 	switch (type) {
2156 	case VM_CAP_HALT_EXIT:
2157 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2158 		    VMCB_INTCPT_HLT, val);
2159 		break;
2160 	case VM_CAP_PAUSE_EXIT:
2161 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2162 		    VMCB_INTCPT_PAUSE, val);
2163 		break;
2164 	case VM_CAP_UNRESTRICTED_GUEST:
2165 		/* Unrestricted guest execution cannot be disabled in SVM */
2166 		if (val == 0)
2167 			error = EINVAL;
2168 		break;
2169 	default:
2170 		error = ENOENT;
2171 		break;
2172 	}
2173 	return (error);
2174 }
2175 
2176 static int
2177 svm_getcap(void *arg, int vcpu, int type, int *retval)
2178 {
2179 	struct svm_softc *sc;
2180 	int error;
2181 
2182 	sc = arg;
2183 	error = 0;
2184 
2185 	switch (type) {
2186 	case VM_CAP_HALT_EXIT:
2187 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2188 		    VMCB_INTCPT_HLT);
2189 		break;
2190 	case VM_CAP_PAUSE_EXIT:
2191 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2192 		    VMCB_INTCPT_PAUSE);
2193 		break;
2194 	case VM_CAP_UNRESTRICTED_GUEST:
2195 		*retval = 1;	/* unrestricted guest is always enabled */
2196 		break;
2197 	default:
2198 		error = ENOENT;
2199 		break;
2200 	}
2201 	return (error);
2202 }
2203 
2204 static struct vlapic *
2205 svm_vlapic_init(void *arg, int vcpuid)
2206 {
2207 	struct svm_softc *svm_sc;
2208 	struct vlapic *vlapic;
2209 
2210 	svm_sc = arg;
2211 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2212 	vlapic->vm = svm_sc->vm;
2213 	vlapic->vcpuid = vcpuid;
2214 	vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2215 
2216 	vlapic_init(vlapic);
2217 
2218 	return (vlapic);
2219 }
2220 
2221 static void
2222 svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2223 {
2224 
2225         vlapic_cleanup(vlapic);
2226         free(vlapic, M_SVM_VLAPIC);
2227 }
2228 
2229 struct vmm_ops vmm_ops_amd = {
2230 	svm_init,
2231 	svm_cleanup,
2232 	svm_restore,
2233 	svm_vminit,
2234 	svm_vmrun,
2235 	svm_vmcleanup,
2236 	svm_getreg,
2237 	svm_setreg,
2238 	vmcb_getdesc,
2239 	vmcb_setdesc,
2240 	svm_getcap,
2241 	svm_setcap,
2242 	svm_npt_alloc,
2243 	svm_npt_free,
2244 	svm_vlapic_init,
2245 	svm_vlapic_cleanup
2246 };
2247