xref: /freebsd/sys/amd64/vmm/amd/svm.c (revision d0ba1baed3f6e4936a0c1b89c25f6c59168ef6de)
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/md_var.h>
45 #include <machine/specialreg.h>
46 #include <machine/smp.h>
47 #include <machine/vmm.h>
48 #include <machine/vmm_dev.h>
49 #include <machine/vmm_instruction_emul.h>
50 
51 #include "vmm_lapic.h"
52 #include "vmm_stat.h"
53 #include "vmm_ktr.h"
54 #include "vmm_ioport.h"
55 #include "vatpic.h"
56 #include "vlapic.h"
57 #include "vlapic_priv.h"
58 
59 #include "x86.h"
60 #include "vmcb.h"
61 #include "svm.h"
62 #include "svm_softc.h"
63 #include "svm_msr.h"
64 #include "npt.h"
65 
66 SYSCTL_DECL(_hw_vmm);
67 SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL);
68 
69 /*
70  * SVM CPUID function 0x8000_000A, edx bit decoding.
71  */
72 #define AMD_CPUID_SVM_NP		BIT(0)  /* Nested paging or RVI */
73 #define AMD_CPUID_SVM_LBR		BIT(1)  /* Last branch virtualization */
74 #define AMD_CPUID_SVM_SVML		BIT(2)  /* SVM lock */
75 #define AMD_CPUID_SVM_NRIP_SAVE		BIT(3)  /* Next RIP is saved */
76 #define AMD_CPUID_SVM_TSC_RATE		BIT(4)  /* TSC rate control. */
77 #define AMD_CPUID_SVM_VMCB_CLEAN	BIT(5)  /* VMCB state caching */
78 #define AMD_CPUID_SVM_FLUSH_BY_ASID	BIT(6)  /* Flush by ASID */
79 #define AMD_CPUID_SVM_DECODE_ASSIST	BIT(7)  /* Decode assist */
80 #define AMD_CPUID_SVM_PAUSE_INC		BIT(10) /* Pause intercept filter. */
81 #define AMD_CPUID_SVM_PAUSE_FTH		BIT(12) /* Pause filter threshold */
82 #define	AMD_CPUID_SVM_AVIC		BIT(13)	/* AVIC present */
83 
84 #define	VMCB_CACHE_DEFAULT	(VMCB_CACHE_ASID 	|	\
85 				VMCB_CACHE_IOPM		|	\
86 				VMCB_CACHE_I		|	\
87 				VMCB_CACHE_TPR		|	\
88 				VMCB_CACHE_CR2		|	\
89 				VMCB_CACHE_CR		|	\
90 				VMCB_CACHE_DR		|	\
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 	/* Set up DR6/7 to power-on state */
510 	state->dr6 = 0xffff0ff0;
511 	state->dr7 = 0x400;
512 }
513 
514 /*
515  * Initialize a virtual machine.
516  */
517 static void *
518 svm_vminit(struct vm *vm, pmap_t pmap)
519 {
520 	struct svm_softc *svm_sc;
521 	struct svm_vcpu *vcpu;
522 	vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
523 	int i;
524 
525 	svm_sc = malloc(sizeof (*svm_sc), M_SVM, M_WAITOK | M_ZERO);
526 	if (((uintptr_t)svm_sc & PAGE_MASK) != 0)
527 		panic("malloc of svm_softc not aligned on page boundary");
528 
529 	svm_sc->msr_bitmap = contigmalloc(SVM_MSR_BITMAP_SIZE, M_SVM,
530 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
531 	if (svm_sc->msr_bitmap == NULL)
532 		panic("contigmalloc of SVM MSR bitmap failed");
533 	svm_sc->iopm_bitmap = contigmalloc(SVM_IO_BITMAP_SIZE, M_SVM,
534 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
535 	if (svm_sc->iopm_bitmap == NULL)
536 		panic("contigmalloc of SVM IO bitmap failed");
537 
538 	svm_sc->vm = vm;
539 	svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4);
540 
541 	/*
542 	 * Intercept read and write accesses to all MSRs.
543 	 */
544 	memset(svm_sc->msr_bitmap, 0xFF, SVM_MSR_BITMAP_SIZE);
545 
546 	/*
547 	 * Access to the following MSRs is redirected to the VMCB when the
548 	 * guest is executing. Therefore it is safe to allow the guest to
549 	 * read/write these MSRs directly without hypervisor involvement.
550 	 */
551 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
552 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
553 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
554 
555 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
556 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
557 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
558 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
559 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
560 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
561 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
562 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
563 
564 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
565 
566 	/*
567 	 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
568 	 */
569 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
570 
571 	/* Intercept access to all I/O ports. */
572 	memset(svm_sc->iopm_bitmap, 0xFF, SVM_IO_BITMAP_SIZE);
573 
574 	iopm_pa = vtophys(svm_sc->iopm_bitmap);
575 	msrpm_pa = vtophys(svm_sc->msr_bitmap);
576 	pml4_pa = svm_sc->nptp;
577 	for (i = 0; i < VM_MAXCPU; i++) {
578 		vcpu = svm_get_vcpu(svm_sc, i);
579 		vcpu->nextrip = ~0;
580 		vcpu->lastcpu = NOCPU;
581 		vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
582 		vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
583 		svm_msr_guest_init(svm_sc, i);
584 	}
585 	return (svm_sc);
586 }
587 
588 /*
589  * Collateral for a generic SVM VM-exit.
590  */
591 static void
592 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
593 {
594 
595 	vme->exitcode = VM_EXITCODE_SVM;
596 	vme->u.svm.exitcode = code;
597 	vme->u.svm.exitinfo1 = info1;
598 	vme->u.svm.exitinfo2 = info2;
599 }
600 
601 static int
602 svm_cpl(struct vmcb_state *state)
603 {
604 
605 	/*
606 	 * From APMv2:
607 	 *   "Retrieve the CPL from the CPL field in the VMCB, not
608 	 *    from any segment DPL"
609 	 */
610 	return (state->cpl);
611 }
612 
613 static enum vm_cpu_mode
614 svm_vcpu_mode(struct vmcb *vmcb)
615 {
616 	struct vmcb_segment seg;
617 	struct vmcb_state *state;
618 	int error;
619 
620 	state = &vmcb->state;
621 
622 	if (state->efer & EFER_LMA) {
623 		error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
624 		KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
625 		    error));
626 
627 		/*
628 		 * Section 4.8.1 for APM2, check if Code Segment has
629 		 * Long attribute set in descriptor.
630 		 */
631 		if (seg.attrib & VMCB_CS_ATTRIB_L)
632 			return (CPU_MODE_64BIT);
633 		else
634 			return (CPU_MODE_COMPATIBILITY);
635 	} else  if (state->cr0 & CR0_PE) {
636 		return (CPU_MODE_PROTECTED);
637 	} else {
638 		return (CPU_MODE_REAL);
639 	}
640 }
641 
642 static enum vm_paging_mode
643 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
644 {
645 
646 	if ((cr0 & CR0_PG) == 0)
647 		return (PAGING_MODE_FLAT);
648 	if ((cr4 & CR4_PAE) == 0)
649 		return (PAGING_MODE_32);
650 	if (efer & EFER_LME)
651 		return (PAGING_MODE_64);
652 	else
653 		return (PAGING_MODE_PAE);
654 }
655 
656 /*
657  * ins/outs utility routines
658  */
659 static uint64_t
660 svm_inout_str_index(struct svm_regctx *regs, int in)
661 {
662 	uint64_t val;
663 
664 	val = in ? regs->sctx_rdi : regs->sctx_rsi;
665 
666 	return (val);
667 }
668 
669 static uint64_t
670 svm_inout_str_count(struct svm_regctx *regs, int rep)
671 {
672 	uint64_t val;
673 
674 	val = rep ? regs->sctx_rcx : 1;
675 
676 	return (val);
677 }
678 
679 static void
680 svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
681     int in, struct vm_inout_str *vis)
682 {
683 	int error, s;
684 
685 	if (in) {
686 		vis->seg_name = VM_REG_GUEST_ES;
687 	} else {
688 		/* The segment field has standard encoding */
689 		s = (info1 >> 10) & 0x7;
690 		vis->seg_name = vm_segment_name(s);
691 	}
692 
693 	error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
694 	KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
695 }
696 
697 static int
698 svm_inout_str_addrsize(uint64_t info1)
699 {
700         uint32_t size;
701 
702         size = (info1 >> 7) & 0x7;
703         switch (size) {
704         case 1:
705                 return (2);     /* 16 bit */
706         case 2:
707                 return (4);     /* 32 bit */
708         case 4:
709                 return (8);     /* 64 bit */
710         default:
711                 panic("%s: invalid size encoding %d", __func__, size);
712         }
713 }
714 
715 static void
716 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
717 {
718 	struct vmcb_state *state;
719 
720 	state = &vmcb->state;
721 	paging->cr3 = state->cr3;
722 	paging->cpl = svm_cpl(state);
723 	paging->cpu_mode = svm_vcpu_mode(vmcb);
724 	paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
725 	    state->efer);
726 }
727 
728 #define	UNHANDLED 0
729 
730 /*
731  * Handle guest I/O intercept.
732  */
733 static int
734 svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
735 {
736 	struct vmcb_ctrl *ctrl;
737 	struct vmcb_state *state;
738 	struct svm_regctx *regs;
739 	struct vm_inout_str *vis;
740 	uint64_t info1;
741 	int inout_string;
742 
743 	state = svm_get_vmcb_state(svm_sc, vcpu);
744 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
745 	regs  = svm_get_guest_regctx(svm_sc, vcpu);
746 
747 	info1 = ctrl->exitinfo1;
748 	inout_string = info1 & BIT(2) ? 1 : 0;
749 
750 	/*
751 	 * The effective segment number in EXITINFO1[12:10] is populated
752 	 * only if the processor has the DecodeAssist capability.
753 	 *
754 	 * XXX this is not specified explicitly in APMv2 but can be verified
755 	 * empirically.
756 	 */
757 	if (inout_string && !decode_assist())
758 		return (UNHANDLED);
759 
760 	vmexit->exitcode 	= VM_EXITCODE_INOUT;
761 	vmexit->u.inout.in 	= (info1 & BIT(0)) ? 1 : 0;
762 	vmexit->u.inout.string 	= inout_string;
763 	vmexit->u.inout.rep 	= (info1 & BIT(3)) ? 1 : 0;
764 	vmexit->u.inout.bytes 	= (info1 >> 4) & 0x7;
765 	vmexit->u.inout.port 	= (uint16_t)(info1 >> 16);
766 	vmexit->u.inout.eax 	= (uint32_t)(state->rax);
767 
768 	if (inout_string) {
769 		vmexit->exitcode = VM_EXITCODE_INOUT_STR;
770 		vis = &vmexit->u.inout_str;
771 		svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
772 		vis->rflags = state->rflags;
773 		vis->cr0 = state->cr0;
774 		vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
775 		vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
776 		vis->addrsize = svm_inout_str_addrsize(info1);
777 		svm_inout_str_seginfo(svm_sc, vcpu, info1,
778 		    vmexit->u.inout.in, vis);
779 	}
780 
781 	return (UNHANDLED);
782 }
783 
784 static int
785 npf_fault_type(uint64_t exitinfo1)
786 {
787 
788 	if (exitinfo1 & VMCB_NPF_INFO1_W)
789 		return (VM_PROT_WRITE);
790 	else if (exitinfo1 & VMCB_NPF_INFO1_ID)
791 		return (VM_PROT_EXECUTE);
792 	else
793 		return (VM_PROT_READ);
794 }
795 
796 static bool
797 svm_npf_emul_fault(uint64_t exitinfo1)
798 {
799 
800 	if (exitinfo1 & VMCB_NPF_INFO1_ID) {
801 		return (false);
802 	}
803 
804 	if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
805 		return (false);
806 	}
807 
808 	if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
809 		return (false);
810 	}
811 
812 	return (true);
813 }
814 
815 static void
816 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
817 {
818 	struct vm_guest_paging *paging;
819 	struct vmcb_segment seg;
820 	struct vmcb_ctrl *ctrl;
821 	char *inst_bytes;
822 	int error, inst_len;
823 
824 	ctrl = &vmcb->ctrl;
825 	paging = &vmexit->u.inst_emul.paging;
826 
827 	vmexit->exitcode = VM_EXITCODE_INST_EMUL;
828 	vmexit->u.inst_emul.gpa = gpa;
829 	vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
830 	svm_paging_info(vmcb, paging);
831 
832 	error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
833 	KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
834 
835 	switch(paging->cpu_mode) {
836 	case CPU_MODE_REAL:
837 		vmexit->u.inst_emul.cs_base = seg.base;
838 		vmexit->u.inst_emul.cs_d = 0;
839 		break;
840 	case CPU_MODE_PROTECTED:
841 	case CPU_MODE_COMPATIBILITY:
842 		vmexit->u.inst_emul.cs_base = seg.base;
843 
844 		/*
845 		 * Section 4.8.1 of APM2, Default Operand Size or D bit.
846 		 */
847 		vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
848 		    1 : 0;
849 		break;
850 	default:
851 		vmexit->u.inst_emul.cs_base = 0;
852 		vmexit->u.inst_emul.cs_d = 0;
853 		break;
854 	}
855 
856 	/*
857 	 * Copy the instruction bytes into 'vie' if available.
858 	 */
859 	if (decode_assist() && !disable_npf_assist) {
860 		inst_len = ctrl->inst_len;
861 		inst_bytes = ctrl->inst_bytes;
862 	} else {
863 		inst_len = 0;
864 		inst_bytes = NULL;
865 	}
866 	vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
867 }
868 
869 #ifdef KTR
870 static const char *
871 intrtype_to_str(int intr_type)
872 {
873 	switch (intr_type) {
874 	case VMCB_EVENTINJ_TYPE_INTR:
875 		return ("hwintr");
876 	case VMCB_EVENTINJ_TYPE_NMI:
877 		return ("nmi");
878 	case VMCB_EVENTINJ_TYPE_INTn:
879 		return ("swintr");
880 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
881 		return ("exception");
882 	default:
883 		panic("%s: unknown intr_type %d", __func__, intr_type);
884 	}
885 }
886 #endif
887 
888 /*
889  * Inject an event to vcpu as described in section 15.20, "Event injection".
890  */
891 static void
892 svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
893 		 uint32_t error, bool ec_valid)
894 {
895 	struct vmcb_ctrl *ctrl;
896 
897 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
898 
899 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
900 	    ("%s: event already pending %#lx", __func__, ctrl->eventinj));
901 
902 	KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
903 	    __func__, vector));
904 
905 	switch (intr_type) {
906 	case VMCB_EVENTINJ_TYPE_INTR:
907 	case VMCB_EVENTINJ_TYPE_NMI:
908 	case VMCB_EVENTINJ_TYPE_INTn:
909 		break;
910 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
911 		if (vector >= 0 && vector <= 31 && vector != 2)
912 			break;
913 		/* FALLTHROUGH */
914 	default:
915 		panic("%s: invalid intr_type/vector: %d/%d", __func__,
916 		    intr_type, vector);
917 	}
918 	ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
919 	if (ec_valid) {
920 		ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
921 		ctrl->eventinj |= (uint64_t)error << 32;
922 		VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
923 		    intrtype_to_str(intr_type), vector, error);
924 	} else {
925 		VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
926 		    intrtype_to_str(intr_type), vector);
927 	}
928 }
929 
930 static void
931 svm_update_virqinfo(struct svm_softc *sc, int vcpu)
932 {
933 	struct vm *vm;
934 	struct vlapic *vlapic;
935 	struct vmcb_ctrl *ctrl;
936 
937 	vm = sc->vm;
938 	vlapic = vm_lapic(vm, vcpu);
939 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
940 
941 	/* Update %cr8 in the emulated vlapic */
942 	vlapic_set_cr8(vlapic, ctrl->v_tpr);
943 
944 	/* Virtual interrupt injection is not used. */
945 	KASSERT(ctrl->v_intr_vector == 0, ("%s: invalid "
946 	    "v_intr_vector %d", __func__, ctrl->v_intr_vector));
947 }
948 
949 static void
950 svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
951 {
952 	struct vmcb_ctrl *ctrl;
953 	uint64_t intinfo;
954 
955 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
956 	intinfo = ctrl->exitintinfo;
957 	if (!VMCB_EXITINTINFO_VALID(intinfo))
958 		return;
959 
960 	/*
961 	 * From APMv2, Section "Intercepts during IDT interrupt delivery"
962 	 *
963 	 * If a #VMEXIT happened during event delivery then record the event
964 	 * that was being delivered.
965 	 */
966 	VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
967 		intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
968 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
969 	vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
970 }
971 
972 #ifdef INVARIANTS
973 static __inline int
974 vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
975 {
976 
977 	return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
978 	    VMCB_INTCPT_VINTR));
979 }
980 #endif
981 
982 static __inline void
983 enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
984 {
985 	struct vmcb_ctrl *ctrl;
986 
987 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
988 
989 	if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
990 		KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
991 		KASSERT(vintr_intercept_enabled(sc, vcpu),
992 		    ("%s: vintr intercept should be enabled", __func__));
993 		return;
994 	}
995 
996 	VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
997 	ctrl->v_irq = 1;
998 	ctrl->v_ign_tpr = 1;
999 	ctrl->v_intr_vector = 0;
1000 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1001 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1002 }
1003 
1004 static __inline void
1005 disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1006 {
1007 	struct vmcb_ctrl *ctrl;
1008 
1009 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1010 
1011 	if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1012 		KASSERT(!vintr_intercept_enabled(sc, vcpu),
1013 		    ("%s: vintr intercept should be disabled", __func__));
1014 		return;
1015 	}
1016 
1017 	VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
1018 	ctrl->v_irq = 0;
1019 	ctrl->v_intr_vector = 0;
1020 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1021 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1022 }
1023 
1024 static int
1025 svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
1026 {
1027 	struct vmcb_ctrl *ctrl;
1028 	int oldval, newval;
1029 
1030 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1031 	oldval = ctrl->intr_shadow;
1032 	newval = val ? 1 : 0;
1033 	if (newval != oldval) {
1034 		ctrl->intr_shadow = newval;
1035 		VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1036 	}
1037 	return (0);
1038 }
1039 
1040 static int
1041 svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1042 {
1043 	struct vmcb_ctrl *ctrl;
1044 
1045 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1046 	*val = ctrl->intr_shadow;
1047 	return (0);
1048 }
1049 
1050 /*
1051  * Once an NMI is injected it blocks delivery of further NMIs until the handler
1052  * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1053  * to track when the vcpu is done handling the NMI.
1054  */
1055 static int
1056 nmi_blocked(struct svm_softc *sc, int vcpu)
1057 {
1058 	int blocked;
1059 
1060 	blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1061 	    VMCB_INTCPT_IRET);
1062 	return (blocked);
1063 }
1064 
1065 static void
1066 enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1067 {
1068 
1069 	KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1070 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1071 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1072 }
1073 
1074 static void
1075 clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1076 {
1077 	int error;
1078 
1079 	KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1080 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1081 	/*
1082 	 * When the IRET intercept is cleared the vcpu will attempt to execute
1083 	 * the "iret" when it runs next. However, it is possible to inject
1084 	 * another NMI into the vcpu before the "iret" has actually executed.
1085 	 *
1086 	 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1087 	 * it will trap back into the hypervisor. If an NMI is pending for
1088 	 * the vcpu it will be injected into the guest.
1089 	 *
1090 	 * XXX this needs to be fixed
1091 	 */
1092 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1093 
1094 	/*
1095 	 * Set 'intr_shadow' to prevent an NMI from being injected on the
1096 	 * immediate VMRUN.
1097 	 */
1098 	error = svm_modify_intr_shadow(sc, vcpu, 1);
1099 	KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1100 }
1101 
1102 #define	EFER_MBZ_BITS	0xFFFFFFFFFFFF0200UL
1103 
1104 static int
1105 svm_write_efer(struct svm_softc *sc, int vcpu, uint64_t newval, bool *retu)
1106 {
1107 	struct vm_exit *vme;
1108 	struct vmcb_state *state;
1109 	uint64_t changed, lma, oldval;
1110 	int error;
1111 
1112 	state = svm_get_vmcb_state(sc, vcpu);
1113 
1114 	oldval = state->efer;
1115 	VCPU_CTR2(sc->vm, vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval);
1116 
1117 	newval &= ~0xFE;		/* clear the Read-As-Zero (RAZ) bits */
1118 	changed = oldval ^ newval;
1119 
1120 	if (newval & EFER_MBZ_BITS)
1121 		goto gpf;
1122 
1123 	/* APMv2 Table 14-5 "Long-Mode Consistency Checks" */
1124 	if (changed & EFER_LME) {
1125 		if (state->cr0 & CR0_PG)
1126 			goto gpf;
1127 	}
1128 
1129 	/* EFER.LMA = EFER.LME & CR0.PG */
1130 	if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0)
1131 		lma = EFER_LMA;
1132 	else
1133 		lma = 0;
1134 
1135 	if ((newval & EFER_LMA) != lma)
1136 		goto gpf;
1137 
1138 	if (newval & EFER_NXE) {
1139 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_NO_EXECUTE))
1140 			goto gpf;
1141 	}
1142 
1143 	/*
1144 	 * XXX bhyve does not enforce segment limits in 64-bit mode. Until
1145 	 * this is fixed flag guest attempt to set EFER_LMSLE as an error.
1146 	 */
1147 	if (newval & EFER_LMSLE) {
1148 		vme = vm_exitinfo(sc->vm, vcpu);
1149 		vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0);
1150 		*retu = true;
1151 		return (0);
1152 	}
1153 
1154 	if (newval & EFER_FFXSR) {
1155 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_FFXSR))
1156 			goto gpf;
1157 	}
1158 
1159 	if (newval & EFER_TCE) {
1160 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_TCE))
1161 			goto gpf;
1162 	}
1163 
1164 	error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, newval);
1165 	KASSERT(error == 0, ("%s: error %d updating efer", __func__, error));
1166 	return (0);
1167 gpf:
1168 	vm_inject_gp(sc->vm, vcpu);
1169 	return (0);
1170 }
1171 
1172 static int
1173 emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1174     bool *retu)
1175 {
1176 	int error;
1177 
1178 	if (lapic_msr(num))
1179 		error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1180 	else if (num == MSR_EFER)
1181 		error = svm_write_efer(sc, vcpu, val, retu);
1182 	else
1183 		error = svm_wrmsr(sc, vcpu, num, val, retu);
1184 
1185 	return (error);
1186 }
1187 
1188 static int
1189 emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1190 {
1191 	struct vmcb_state *state;
1192 	struct svm_regctx *ctx;
1193 	uint64_t result;
1194 	int error;
1195 
1196 	if (lapic_msr(num))
1197 		error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1198 	else
1199 		error = svm_rdmsr(sc, vcpu, num, &result, retu);
1200 
1201 	if (error == 0) {
1202 		state = svm_get_vmcb_state(sc, vcpu);
1203 		ctx = svm_get_guest_regctx(sc, vcpu);
1204 		state->rax = result & 0xffffffff;
1205 		ctx->sctx_rdx = result >> 32;
1206 	}
1207 
1208 	return (error);
1209 }
1210 
1211 #ifdef KTR
1212 static const char *
1213 exit_reason_to_str(uint64_t reason)
1214 {
1215 	static char reasonbuf[32];
1216 
1217 	switch (reason) {
1218 	case VMCB_EXIT_INVALID:
1219 		return ("invalvmcb");
1220 	case VMCB_EXIT_SHUTDOWN:
1221 		return ("shutdown");
1222 	case VMCB_EXIT_NPF:
1223 		return ("nptfault");
1224 	case VMCB_EXIT_PAUSE:
1225 		return ("pause");
1226 	case VMCB_EXIT_HLT:
1227 		return ("hlt");
1228 	case VMCB_EXIT_CPUID:
1229 		return ("cpuid");
1230 	case VMCB_EXIT_IO:
1231 		return ("inout");
1232 	case VMCB_EXIT_MC:
1233 		return ("mchk");
1234 	case VMCB_EXIT_INTR:
1235 		return ("extintr");
1236 	case VMCB_EXIT_NMI:
1237 		return ("nmi");
1238 	case VMCB_EXIT_VINTR:
1239 		return ("vintr");
1240 	case VMCB_EXIT_MSR:
1241 		return ("msr");
1242 	case VMCB_EXIT_IRET:
1243 		return ("iret");
1244 	case VMCB_EXIT_MONITOR:
1245 		return ("monitor");
1246 	case VMCB_EXIT_MWAIT:
1247 		return ("mwait");
1248 	default:
1249 		snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1250 		return (reasonbuf);
1251 	}
1252 }
1253 #endif	/* KTR */
1254 
1255 /*
1256  * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1257  * that are due to instruction intercepts as well as MSR and IOIO intercepts
1258  * and exceptions caused by INT3, INTO and BOUND instructions.
1259  *
1260  * Return 1 if the nRIP is valid and 0 otherwise.
1261  */
1262 static int
1263 nrip_valid(uint64_t exitcode)
1264 {
1265 	switch (exitcode) {
1266 	case 0x00 ... 0x0F:	/* read of CR0 through CR15 */
1267 	case 0x10 ... 0x1F:	/* write of CR0 through CR15 */
1268 	case 0x20 ... 0x2F:	/* read of DR0 through DR15 */
1269 	case 0x30 ... 0x3F:	/* write of DR0 through DR15 */
1270 	case 0x43:		/* INT3 */
1271 	case 0x44:		/* INTO */
1272 	case 0x45:		/* BOUND */
1273 	case 0x65 ... 0x7C:	/* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1274 	case 0x80 ... 0x8D:	/* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1275 		return (1);
1276 	default:
1277 		return (0);
1278 	}
1279 }
1280 
1281 static int
1282 svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1283 {
1284 	struct vmcb *vmcb;
1285 	struct vmcb_state *state;
1286 	struct vmcb_ctrl *ctrl;
1287 	struct svm_regctx *ctx;
1288 	uint64_t code, info1, info2, val;
1289 	uint32_t eax, ecx, edx;
1290 	int error, errcode_valid, handled, idtvec, reflect;
1291 	bool retu;
1292 
1293 	ctx = svm_get_guest_regctx(svm_sc, vcpu);
1294 	vmcb = svm_get_vmcb(svm_sc, vcpu);
1295 	state = &vmcb->state;
1296 	ctrl = &vmcb->ctrl;
1297 
1298 	handled = 0;
1299 	code = ctrl->exitcode;
1300 	info1 = ctrl->exitinfo1;
1301 	info2 = ctrl->exitinfo2;
1302 
1303 	vmexit->exitcode = VM_EXITCODE_BOGUS;
1304 	vmexit->rip = state->rip;
1305 	vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1306 
1307 	vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1308 
1309 	/*
1310 	 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1311 	 * in an inconsistent state and can trigger assertions that would
1312 	 * never happen otherwise.
1313 	 */
1314 	if (code == VMCB_EXIT_INVALID) {
1315 		vm_exit_svm(vmexit, code, info1, info2);
1316 		return (0);
1317 	}
1318 
1319 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1320 	    "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1321 
1322 	KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1323 	    ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1324 	    vmexit->inst_length, code, info1, info2));
1325 
1326 	svm_update_virqinfo(svm_sc, vcpu);
1327 	svm_save_intinfo(svm_sc, vcpu);
1328 
1329 	switch (code) {
1330 	case VMCB_EXIT_IRET:
1331 		/*
1332 		 * Restart execution at "iret" but with the intercept cleared.
1333 		 */
1334 		vmexit->inst_length = 0;
1335 		clear_nmi_blocking(svm_sc, vcpu);
1336 		handled = 1;
1337 		break;
1338 	case VMCB_EXIT_VINTR:	/* interrupt window exiting */
1339 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1340 		handled = 1;
1341 		break;
1342 	case VMCB_EXIT_INTR:	/* external interrupt */
1343 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1344 		handled = 1;
1345 		break;
1346 	case VMCB_EXIT_NMI:	/* external NMI */
1347 		handled = 1;
1348 		break;
1349 	case 0x40 ... 0x5F:
1350 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1351 		reflect = 1;
1352 		idtvec = code - 0x40;
1353 		switch (idtvec) {
1354 		case IDT_MC:
1355 			/*
1356 			 * Call the machine check handler by hand. Also don't
1357 			 * reflect the machine check back into the guest.
1358 			 */
1359 			reflect = 0;
1360 			VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler");
1361 			__asm __volatile("int $18");
1362 			break;
1363 		case IDT_PF:
1364 			error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2,
1365 			    info2);
1366 			KASSERT(error == 0, ("%s: error %d updating cr2",
1367 			    __func__, error));
1368 			/* fallthru */
1369 		case IDT_NP:
1370 		case IDT_SS:
1371 		case IDT_GP:
1372 		case IDT_AC:
1373 		case IDT_TS:
1374 			errcode_valid = 1;
1375 			break;
1376 
1377 		case IDT_DF:
1378 			errcode_valid = 1;
1379 			info1 = 0;
1380 			break;
1381 
1382 		case IDT_BP:
1383 		case IDT_OF:
1384 		case IDT_BR:
1385 			/*
1386 			 * The 'nrip' field is populated for INT3, INTO and
1387 			 * BOUND exceptions and this also implies that
1388 			 * 'inst_length' is non-zero.
1389 			 *
1390 			 * Reset 'inst_length' to zero so the guest %rip at
1391 			 * event injection is identical to what it was when
1392 			 * the exception originally happened.
1393 			 */
1394 			VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d "
1395 			    "to zero before injecting exception %d",
1396 			    vmexit->inst_length, idtvec);
1397 			vmexit->inst_length = 0;
1398 			/* fallthru */
1399 		default:
1400 			errcode_valid = 0;
1401 			info1 = 0;
1402 			break;
1403 		}
1404 		KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1405 		    "when reflecting exception %d into guest",
1406 		    vmexit->inst_length, idtvec));
1407 
1408 		if (reflect) {
1409 			/* Reflect the exception back into the guest */
1410 			VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception "
1411 			    "%d/%#x into the guest", idtvec, (int)info1);
1412 			error = vm_inject_exception(svm_sc->vm, vcpu, idtvec,
1413 			    errcode_valid, info1, 0);
1414 			KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1415 			    __func__, error));
1416 		}
1417 		handled = 1;
1418 		break;
1419 	case VMCB_EXIT_MSR:	/* MSR access. */
1420 		eax = state->rax;
1421 		ecx = ctx->sctx_rcx;
1422 		edx = ctx->sctx_rdx;
1423 		retu = false;
1424 
1425 		if (info1) {
1426 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1427 			val = (uint64_t)edx << 32 | eax;
1428 			VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1429 			    ecx, val);
1430 			if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1431 				vmexit->exitcode = VM_EXITCODE_WRMSR;
1432 				vmexit->u.msr.code = ecx;
1433 				vmexit->u.msr.wval = val;
1434 			} else if (!retu) {
1435 				handled = 1;
1436 			} else {
1437 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1438 				    ("emulate_wrmsr retu with bogus exitcode"));
1439 			}
1440 		} else {
1441 			VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1442 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1443 			if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1444 				vmexit->exitcode = VM_EXITCODE_RDMSR;
1445 				vmexit->u.msr.code = ecx;
1446 			} else if (!retu) {
1447 				handled = 1;
1448 			} else {
1449 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1450 				    ("emulate_rdmsr retu with bogus exitcode"));
1451 			}
1452 		}
1453 		break;
1454 	case VMCB_EXIT_IO:
1455 		handled = svm_handle_io(svm_sc, vcpu, vmexit);
1456 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1457 		break;
1458 	case VMCB_EXIT_CPUID:
1459 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1460 		handled = x86_emulate_cpuid(svm_sc->vm, vcpu,
1461 		    (uint32_t *)&state->rax,
1462 		    (uint32_t *)&ctx->sctx_rbx,
1463 		    (uint32_t *)&ctx->sctx_rcx,
1464 		    (uint32_t *)&ctx->sctx_rdx);
1465 		break;
1466 	case VMCB_EXIT_HLT:
1467 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1468 		vmexit->exitcode = VM_EXITCODE_HLT;
1469 		vmexit->u.hlt.rflags = state->rflags;
1470 		break;
1471 	case VMCB_EXIT_PAUSE:
1472 		vmexit->exitcode = VM_EXITCODE_PAUSE;
1473 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1474 		break;
1475 	case VMCB_EXIT_NPF:
1476 		/* EXITINFO2 contains the faulting guest physical address */
1477 		if (info1 & VMCB_NPF_INFO1_RSV) {
1478 			VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1479 			    "reserved bits set: info1(%#lx) info2(%#lx)",
1480 			    info1, info2);
1481 		} else if (vm_mem_allocated(svm_sc->vm, vcpu, info2)) {
1482 			vmexit->exitcode = VM_EXITCODE_PAGING;
1483 			vmexit->u.paging.gpa = info2;
1484 			vmexit->u.paging.fault_type = npf_fault_type(info1);
1485 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1486 			VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1487 			    "on gpa %#lx/%#lx at rip %#lx",
1488 			    info2, info1, state->rip);
1489 		} else if (svm_npf_emul_fault(info1)) {
1490 			svm_handle_inst_emul(vmcb, info2, vmexit);
1491 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1492 			VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1493 			    "for gpa %#lx/%#lx at rip %#lx",
1494 			    info2, info1, state->rip);
1495 		}
1496 		break;
1497 	case VMCB_EXIT_MONITOR:
1498 		vmexit->exitcode = VM_EXITCODE_MONITOR;
1499 		break;
1500 	case VMCB_EXIT_MWAIT:
1501 		vmexit->exitcode = VM_EXITCODE_MWAIT;
1502 		break;
1503 	default:
1504 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1505 		break;
1506 	}
1507 
1508 	VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1509 	    handled ? "handled" : "unhandled", exit_reason_to_str(code),
1510 	    vmexit->rip, vmexit->inst_length);
1511 
1512 	if (handled) {
1513 		vmexit->rip += vmexit->inst_length;
1514 		vmexit->inst_length = 0;
1515 		state->rip = vmexit->rip;
1516 	} else {
1517 		if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1518 			/*
1519 			 * If this VM exit was not claimed by anybody then
1520 			 * treat it as a generic SVM exit.
1521 			 */
1522 			vm_exit_svm(vmexit, code, info1, info2);
1523 		} else {
1524 			/*
1525 			 * The exitcode and collateral have been populated.
1526 			 * The VM exit will be processed further in userland.
1527 			 */
1528 		}
1529 	}
1530 	return (handled);
1531 }
1532 
1533 static void
1534 svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1535 {
1536 	uint64_t intinfo;
1537 
1538 	if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1539 		return;
1540 
1541 	KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1542 	    "valid: %#lx", __func__, intinfo));
1543 
1544 	svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1545 		VMCB_EXITINTINFO_VECTOR(intinfo),
1546 		VMCB_EXITINTINFO_EC(intinfo),
1547 		VMCB_EXITINTINFO_EC_VALID(intinfo));
1548 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1549 	VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1550 }
1551 
1552 /*
1553  * Inject event to virtual cpu.
1554  */
1555 static void
1556 svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1557 {
1558 	struct vmcb_ctrl *ctrl;
1559 	struct vmcb_state *state;
1560 	struct svm_vcpu *vcpustate;
1561 	uint8_t v_tpr;
1562 	int vector, need_intr_window;
1563 	int extint_pending;
1564 
1565 	state = svm_get_vmcb_state(sc, vcpu);
1566 	ctrl  = svm_get_vmcb_ctrl(sc, vcpu);
1567 	vcpustate = svm_get_vcpu(sc, vcpu);
1568 
1569 	need_intr_window = 0;
1570 
1571 	if (vcpustate->nextrip != state->rip) {
1572 		ctrl->intr_shadow = 0;
1573 		VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking "
1574 		    "cleared due to rip change: %#lx/%#lx",
1575 		    vcpustate->nextrip, state->rip);
1576 	}
1577 
1578 	/*
1579 	 * Inject pending events or exceptions for this vcpu.
1580 	 *
1581 	 * An event might be pending because the previous #VMEXIT happened
1582 	 * during event delivery (i.e. ctrl->exitintinfo).
1583 	 *
1584 	 * An event might also be pending because an exception was injected
1585 	 * by the hypervisor (e.g. #PF during instruction emulation).
1586 	 */
1587 	svm_inj_intinfo(sc, vcpu);
1588 
1589 	/* NMI event has priority over interrupts. */
1590 	if (vm_nmi_pending(sc->vm, vcpu)) {
1591 		if (nmi_blocked(sc, vcpu)) {
1592 			/*
1593 			 * Can't inject another NMI if the guest has not
1594 			 * yet executed an "iret" after the last NMI.
1595 			 */
1596 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1597 			    "to NMI-blocking");
1598 		} else if (ctrl->intr_shadow) {
1599 			/*
1600 			 * Can't inject an NMI if the vcpu is in an intr_shadow.
1601 			 */
1602 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1603 			    "interrupt shadow");
1604 			need_intr_window = 1;
1605 			goto done;
1606 		} else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1607 			/*
1608 			 * If there is already an exception/interrupt pending
1609 			 * then defer the NMI until after that.
1610 			 */
1611 			VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1612 			    "eventinj %#lx", ctrl->eventinj);
1613 
1614 			/*
1615 			 * Use self-IPI to trigger a VM-exit as soon as
1616 			 * possible after the event injection is completed.
1617 			 *
1618 			 * This works only if the external interrupt exiting
1619 			 * is at a lower priority than the event injection.
1620 			 *
1621 			 * Although not explicitly specified in APMv2 the
1622 			 * relative priorities were verified empirically.
1623 			 */
1624 			ipi_cpu(curcpu, IPI_AST);	/* XXX vmm_ipinum? */
1625 		} else {
1626 			vm_nmi_clear(sc->vm, vcpu);
1627 
1628 			/* Inject NMI, vector number is not used */
1629 			svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1630 			    IDT_NMI, 0, false);
1631 
1632 			/* virtual NMI blocking is now in effect */
1633 			enable_nmi_blocking(sc, vcpu);
1634 
1635 			VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1636 		}
1637 	}
1638 
1639 	extint_pending = vm_extint_pending(sc->vm, vcpu);
1640 	if (!extint_pending) {
1641 		if (!vlapic_pending_intr(vlapic, &vector))
1642 			goto done;
1643 		KASSERT(vector >= 16 && vector <= 255,
1644 		    ("invalid vector %d from local APIC", vector));
1645 	} else {
1646 		/* Ask the legacy pic for a vector to inject */
1647 		vatpic_pending_intr(sc->vm, &vector);
1648 		KASSERT(vector >= 0 && vector <= 255,
1649 		    ("invalid vector %d from INTR", vector));
1650 	}
1651 
1652 	/*
1653 	 * If the guest has disabled interrupts or is in an interrupt shadow
1654 	 * then we cannot inject the pending interrupt.
1655 	 */
1656 	if ((state->rflags & PSL_I) == 0) {
1657 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1658 		    "rflags %#lx", vector, state->rflags);
1659 		need_intr_window = 1;
1660 		goto done;
1661 	}
1662 
1663 	if (ctrl->intr_shadow) {
1664 		VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1665 		    "interrupt shadow", vector);
1666 		need_intr_window = 1;
1667 		goto done;
1668 	}
1669 
1670 	if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1671 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1672 		    "eventinj %#lx", vector, ctrl->eventinj);
1673 		need_intr_window = 1;
1674 		goto done;
1675 	}
1676 
1677 	svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1678 
1679 	if (!extint_pending) {
1680 		vlapic_intr_accepted(vlapic, vector);
1681 	} else {
1682 		vm_extint_clear(sc->vm, vcpu);
1683 		vatpic_intr_accepted(sc->vm, vector);
1684 	}
1685 
1686 	/*
1687 	 * Force a VM-exit as soon as the vcpu is ready to accept another
1688 	 * interrupt. This is done because the PIC might have another vector
1689 	 * that it wants to inject. Also, if the APIC has a pending interrupt
1690 	 * that was preempted by the ExtInt then it allows us to inject the
1691 	 * APIC vector as soon as possible.
1692 	 */
1693 	need_intr_window = 1;
1694 done:
1695 	/*
1696 	 * The guest can modify the TPR by writing to %CR8. In guest mode
1697 	 * the processor reflects this write to V_TPR without hypervisor
1698 	 * intervention.
1699 	 *
1700 	 * The guest can also modify the TPR by writing to it via the memory
1701 	 * mapped APIC page. In this case, the write will be emulated by the
1702 	 * hypervisor. For this reason V_TPR must be updated before every
1703 	 * VMRUN.
1704 	 */
1705 	v_tpr = vlapic_get_cr8(vlapic);
1706 	KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1707 	if (ctrl->v_tpr != v_tpr) {
1708 		VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1709 		    ctrl->v_tpr, v_tpr);
1710 		ctrl->v_tpr = v_tpr;
1711 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1712 	}
1713 
1714 	if (need_intr_window) {
1715 		/*
1716 		 * We use V_IRQ in conjunction with the VINTR intercept to
1717 		 * trap into the hypervisor as soon as a virtual interrupt
1718 		 * can be delivered.
1719 		 *
1720 		 * Since injected events are not subject to intercept checks
1721 		 * we need to ensure that the V_IRQ is not actually going to
1722 		 * be delivered on VM entry. The KASSERT below enforces this.
1723 		 */
1724 		KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1725 		    (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1726 		    ("Bogus intr_window_exiting: eventinj (%#lx), "
1727 		    "intr_shadow (%u), rflags (%#lx)",
1728 		    ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1729 		enable_intr_window_exiting(sc, vcpu);
1730 	} else {
1731 		disable_intr_window_exiting(sc, vcpu);
1732 	}
1733 }
1734 
1735 static __inline void
1736 restore_host_tss(void)
1737 {
1738 	struct system_segment_descriptor *tss_sd;
1739 
1740 	/*
1741 	 * The TSS descriptor was in use prior to launching the guest so it
1742 	 * has been marked busy.
1743 	 *
1744 	 * 'ltr' requires the descriptor to be marked available so change the
1745 	 * type to "64-bit available TSS".
1746 	 */
1747 	tss_sd = PCPU_GET(tss);
1748 	tss_sd->sd_type = SDT_SYSTSS;
1749 	ltr(GSEL(GPROC0_SEL, SEL_KPL));
1750 }
1751 
1752 static void
1753 check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu)
1754 {
1755 	struct svm_vcpu *vcpustate;
1756 	struct vmcb_ctrl *ctrl;
1757 	long eptgen;
1758 	bool alloc_asid;
1759 
1760 	KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not "
1761 	    "active on cpu %u", __func__, thiscpu));
1762 
1763 	vcpustate = svm_get_vcpu(sc, vcpuid);
1764 	ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1765 
1766 	/*
1767 	 * The TLB entries associated with the vcpu's ASID are not valid
1768 	 * if either of the following conditions is true:
1769 	 *
1770 	 * 1. The vcpu's ASID generation is different than the host cpu's
1771 	 *    ASID generation. This happens when the vcpu migrates to a new
1772 	 *    host cpu. It can also happen when the number of vcpus executing
1773 	 *    on a host cpu is greater than the number of ASIDs available.
1774 	 *
1775 	 * 2. The pmap generation number is different than the value cached in
1776 	 *    the 'vcpustate'. This happens when the host invalidates pages
1777 	 *    belonging to the guest.
1778 	 *
1779 	 *	asidgen		eptgen	      Action
1780 	 *	mismatch	mismatch
1781 	 *	   0		   0		(a)
1782 	 *	   0		   1		(b1) or (b2)
1783 	 *	   1		   0		(c)
1784 	 *	   1		   1		(d)
1785 	 *
1786 	 * (a) There is no mismatch in eptgen or ASID generation and therefore
1787 	 *     no further action is needed.
1788 	 *
1789 	 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1790 	 *      retained and the TLB entries associated with this ASID
1791 	 *      are flushed by VMRUN.
1792 	 *
1793 	 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1794 	 *      allocated.
1795 	 *
1796 	 * (c) A new ASID is allocated.
1797 	 *
1798 	 * (d) A new ASID is allocated.
1799 	 */
1800 
1801 	alloc_asid = false;
1802 	eptgen = pmap->pm_eptgen;
1803 	ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1804 
1805 	if (vcpustate->asid.gen != asid[thiscpu].gen) {
1806 		alloc_asid = true;	/* (c) and (d) */
1807 	} else if (vcpustate->eptgen != eptgen) {
1808 		if (flush_by_asid())
1809 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;	/* (b1) */
1810 		else
1811 			alloc_asid = true;			/* (b2) */
1812 	} else {
1813 		/*
1814 		 * This is the common case (a).
1815 		 */
1816 		KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1817 		KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1818 		    ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1819 	}
1820 
1821 	if (alloc_asid) {
1822 		if (++asid[thiscpu].num >= nasid) {
1823 			asid[thiscpu].num = 1;
1824 			if (++asid[thiscpu].gen == 0)
1825 				asid[thiscpu].gen = 1;
1826 			/*
1827 			 * If this cpu does not support "flush-by-asid"
1828 			 * then flush the entire TLB on a generation
1829 			 * bump. Subsequent ASID allocation in this
1830 			 * generation can be done without a TLB flush.
1831 			 */
1832 			if (!flush_by_asid())
1833 				ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1834 		}
1835 		vcpustate->asid.gen = asid[thiscpu].gen;
1836 		vcpustate->asid.num = asid[thiscpu].num;
1837 
1838 		ctrl->asid = vcpustate->asid.num;
1839 		svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1840 		/*
1841 		 * If this cpu supports "flush-by-asid" then the TLB
1842 		 * was not flushed after the generation bump. The TLB
1843 		 * is flushed selectively after every new ASID allocation.
1844 		 */
1845 		if (flush_by_asid())
1846 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1847 	}
1848 	vcpustate->eptgen = eptgen;
1849 
1850 	KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1851 	KASSERT(ctrl->asid == vcpustate->asid.num,
1852 	    ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1853 }
1854 
1855 static __inline void
1856 disable_gintr(void)
1857 {
1858 
1859 	__asm __volatile("clgi");
1860 }
1861 
1862 static __inline void
1863 enable_gintr(void)
1864 {
1865 
1866         __asm __volatile("stgi");
1867 }
1868 
1869 static __inline void
1870 svm_dr_enter_guest(struct svm_regctx *gctx)
1871 {
1872 
1873 	/* Save host control debug registers. */
1874 	gctx->host_dr7 = rdr7();
1875 	gctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR);
1876 
1877 	/*
1878 	 * Disable debugging in DR7 and DEBUGCTL to avoid triggering
1879 	 * exceptions in the host based on the guest DRx values.  The
1880 	 * guest DR6, DR7, and DEBUGCTL are saved/restored in the
1881 	 * VMCB.
1882 	 */
1883 	load_dr7(0);
1884 	wrmsr(MSR_DEBUGCTLMSR, 0);
1885 
1886 	/* Save host debug registers. */
1887 	gctx->host_dr0 = rdr0();
1888 	gctx->host_dr1 = rdr1();
1889 	gctx->host_dr2 = rdr2();
1890 	gctx->host_dr3 = rdr3();
1891 	gctx->host_dr6 = rdr6();
1892 
1893 	/* Restore guest debug registers. */
1894 	load_dr0(gctx->sctx_dr0);
1895 	load_dr1(gctx->sctx_dr1);
1896 	load_dr2(gctx->sctx_dr2);
1897 	load_dr3(gctx->sctx_dr3);
1898 }
1899 
1900 static __inline void
1901 svm_dr_leave_guest(struct svm_regctx *gctx)
1902 {
1903 
1904 	/* Save guest debug registers. */
1905 	gctx->sctx_dr0 = rdr0();
1906 	gctx->sctx_dr1 = rdr1();
1907 	gctx->sctx_dr2 = rdr2();
1908 	gctx->sctx_dr3 = rdr3();
1909 
1910 	/*
1911 	 * Restore host debug registers.  Restore DR7 and DEBUGCTL
1912 	 * last.
1913 	 */
1914 	load_dr0(gctx->host_dr0);
1915 	load_dr1(gctx->host_dr1);
1916 	load_dr2(gctx->host_dr2);
1917 	load_dr3(gctx->host_dr3);
1918 	load_dr6(gctx->host_dr6);
1919 	wrmsr(MSR_DEBUGCTLMSR, gctx->host_debugctl);
1920 	load_dr7(gctx->host_dr7);
1921 }
1922 
1923 /*
1924  * Start vcpu with specified RIP.
1925  */
1926 static int
1927 svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap,
1928 	struct vm_eventinfo *evinfo)
1929 {
1930 	struct svm_regctx *gctx;
1931 	struct svm_softc *svm_sc;
1932 	struct svm_vcpu *vcpustate;
1933 	struct vmcb_state *state;
1934 	struct vmcb_ctrl *ctrl;
1935 	struct vm_exit *vmexit;
1936 	struct vlapic *vlapic;
1937 	struct vm *vm;
1938 	uint64_t vmcb_pa;
1939 	int handled;
1940 
1941 	svm_sc = arg;
1942 	vm = svm_sc->vm;
1943 
1944 	vcpustate = svm_get_vcpu(svm_sc, vcpu);
1945 	state = svm_get_vmcb_state(svm_sc, vcpu);
1946 	ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
1947 	vmexit = vm_exitinfo(vm, vcpu);
1948 	vlapic = vm_lapic(vm, vcpu);
1949 
1950 	gctx = svm_get_guest_regctx(svm_sc, vcpu);
1951 	vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
1952 
1953 	if (vcpustate->lastcpu != curcpu) {
1954 		/*
1955 		 * Force new ASID allocation by invalidating the generation.
1956 		 */
1957 		vcpustate->asid.gen = 0;
1958 
1959 		/*
1960 		 * Invalidate the VMCB state cache by marking all fields dirty.
1961 		 */
1962 		svm_set_dirty(svm_sc, vcpu, 0xffffffff);
1963 
1964 		/*
1965 		 * XXX
1966 		 * Setting 'vcpustate->lastcpu' here is bit premature because
1967 		 * we may return from this function without actually executing
1968 		 * the VMRUN  instruction. This could happen if a rendezvous
1969 		 * or an AST is pending on the first time through the loop.
1970 		 *
1971 		 * This works for now but any new side-effects of vcpu
1972 		 * migration should take this case into account.
1973 		 */
1974 		vcpustate->lastcpu = curcpu;
1975 		vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
1976 	}
1977 
1978 	svm_msr_guest_enter(svm_sc, vcpu);
1979 
1980 	/* Update Guest RIP */
1981 	state->rip = rip;
1982 
1983 	do {
1984 		/*
1985 		 * Disable global interrupts to guarantee atomicity during
1986 		 * loading of guest state. This includes not only the state
1987 		 * loaded by the "vmrun" instruction but also software state
1988 		 * maintained by the hypervisor: suspended and rendezvous
1989 		 * state, NPT generation number, vlapic interrupts etc.
1990 		 */
1991 		disable_gintr();
1992 
1993 		if (vcpu_suspended(evinfo)) {
1994 			enable_gintr();
1995 			vm_exit_suspended(vm, vcpu, state->rip);
1996 			break;
1997 		}
1998 
1999 		if (vcpu_rendezvous_pending(evinfo)) {
2000 			enable_gintr();
2001 			vm_exit_rendezvous(vm, vcpu, state->rip);
2002 			break;
2003 		}
2004 
2005 		if (vcpu_reqidle(evinfo)) {
2006 			enable_gintr();
2007 			vm_exit_reqidle(vm, vcpu, state->rip);
2008 			break;
2009 		}
2010 
2011 		/* We are asked to give the cpu by scheduler. */
2012 		if (vcpu_should_yield(vm, vcpu)) {
2013 			enable_gintr();
2014 			vm_exit_astpending(vm, vcpu, state->rip);
2015 			break;
2016 		}
2017 
2018 		if (vcpu_debugged(vm, vcpu)) {
2019 			enable_gintr();
2020 			vm_exit_debug(vm, vcpu, state->rip);
2021 			break;
2022 		}
2023 
2024 		svm_inj_interrupts(svm_sc, vcpu, vlapic);
2025 
2026 		/* Activate the nested pmap on 'curcpu' */
2027 		CPU_SET_ATOMIC_ACQ(curcpu, &pmap->pm_active);
2028 
2029 		/*
2030 		 * Check the pmap generation and the ASID generation to
2031 		 * ensure that the vcpu does not use stale TLB mappings.
2032 		 */
2033 		check_asid(svm_sc, vcpu, pmap, curcpu);
2034 
2035 		ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
2036 		vcpustate->dirty = 0;
2037 		VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
2038 
2039 		/* Launch Virtual Machine. */
2040 		VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
2041 		svm_dr_enter_guest(gctx);
2042 		svm_launch(vmcb_pa, gctx, &__pcpu[curcpu]);
2043 		svm_dr_leave_guest(gctx);
2044 
2045 		CPU_CLR_ATOMIC(curcpu, &pmap->pm_active);
2046 
2047 		/*
2048 		 * The host GDTR and IDTR is saved by VMRUN and restored
2049 		 * automatically on #VMEXIT. However, the host TSS needs
2050 		 * to be restored explicitly.
2051 		 */
2052 		restore_host_tss();
2053 
2054 		/* #VMEXIT disables interrupts so re-enable them here. */
2055 		enable_gintr();
2056 
2057 		/* Update 'nextrip' */
2058 		vcpustate->nextrip = state->rip;
2059 
2060 		/* Handle #VMEXIT and if required return to user space. */
2061 		handled = svm_vmexit(svm_sc, vcpu, vmexit);
2062 	} while (handled);
2063 
2064 	svm_msr_guest_exit(svm_sc, vcpu);
2065 
2066 	return (0);
2067 }
2068 
2069 static void
2070 svm_vmcleanup(void *arg)
2071 {
2072 	struct svm_softc *sc = arg;
2073 
2074 	contigfree(sc->iopm_bitmap, SVM_IO_BITMAP_SIZE, M_SVM);
2075 	contigfree(sc->msr_bitmap, SVM_MSR_BITMAP_SIZE, M_SVM);
2076 	free(sc, M_SVM);
2077 }
2078 
2079 static register_t *
2080 swctx_regptr(struct svm_regctx *regctx, int reg)
2081 {
2082 
2083 	switch (reg) {
2084 	case VM_REG_GUEST_RBX:
2085 		return (&regctx->sctx_rbx);
2086 	case VM_REG_GUEST_RCX:
2087 		return (&regctx->sctx_rcx);
2088 	case VM_REG_GUEST_RDX:
2089 		return (&regctx->sctx_rdx);
2090 	case VM_REG_GUEST_RDI:
2091 		return (&regctx->sctx_rdi);
2092 	case VM_REG_GUEST_RSI:
2093 		return (&regctx->sctx_rsi);
2094 	case VM_REG_GUEST_RBP:
2095 		return (&regctx->sctx_rbp);
2096 	case VM_REG_GUEST_R8:
2097 		return (&regctx->sctx_r8);
2098 	case VM_REG_GUEST_R9:
2099 		return (&regctx->sctx_r9);
2100 	case VM_REG_GUEST_R10:
2101 		return (&regctx->sctx_r10);
2102 	case VM_REG_GUEST_R11:
2103 		return (&regctx->sctx_r11);
2104 	case VM_REG_GUEST_R12:
2105 		return (&regctx->sctx_r12);
2106 	case VM_REG_GUEST_R13:
2107 		return (&regctx->sctx_r13);
2108 	case VM_REG_GUEST_R14:
2109 		return (&regctx->sctx_r14);
2110 	case VM_REG_GUEST_R15:
2111 		return (&regctx->sctx_r15);
2112 	case VM_REG_GUEST_DR0:
2113 		return (&regctx->sctx_dr0);
2114 	case VM_REG_GUEST_DR1:
2115 		return (&regctx->sctx_dr1);
2116 	case VM_REG_GUEST_DR2:
2117 		return (&regctx->sctx_dr2);
2118 	case VM_REG_GUEST_DR3:
2119 		return (&regctx->sctx_dr3);
2120 	default:
2121 		return (NULL);
2122 	}
2123 }
2124 
2125 static int
2126 svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
2127 {
2128 	struct svm_softc *svm_sc;
2129 	register_t *reg;
2130 
2131 	svm_sc = arg;
2132 
2133 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2134 		return (svm_get_intr_shadow(svm_sc, vcpu, val));
2135 	}
2136 
2137 	if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2138 		return (0);
2139 	}
2140 
2141 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2142 
2143 	if (reg != NULL) {
2144 		*val = *reg;
2145 		return (0);
2146 	}
2147 
2148 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident);
2149 	return (EINVAL);
2150 }
2151 
2152 static int
2153 svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2154 {
2155 	struct svm_softc *svm_sc;
2156 	register_t *reg;
2157 
2158 	svm_sc = arg;
2159 
2160 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2161 		return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2162 	}
2163 
2164 	if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2165 		return (0);
2166 	}
2167 
2168 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2169 
2170 	if (reg != NULL) {
2171 		*reg = val;
2172 		return (0);
2173 	}
2174 
2175 	/*
2176 	 * XXX deal with CR3 and invalidate TLB entries tagged with the
2177 	 * vcpu's ASID. This needs to be treated differently depending on
2178 	 * whether 'running' is true/false.
2179 	 */
2180 
2181 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident);
2182 	return (EINVAL);
2183 }
2184 
2185 static int
2186 svm_setcap(void *arg, int vcpu, int type, int val)
2187 {
2188 	struct svm_softc *sc;
2189 	int error;
2190 
2191 	sc = arg;
2192 	error = 0;
2193 	switch (type) {
2194 	case VM_CAP_HALT_EXIT:
2195 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2196 		    VMCB_INTCPT_HLT, val);
2197 		break;
2198 	case VM_CAP_PAUSE_EXIT:
2199 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2200 		    VMCB_INTCPT_PAUSE, val);
2201 		break;
2202 	case VM_CAP_UNRESTRICTED_GUEST:
2203 		/* Unrestricted guest execution cannot be disabled in SVM */
2204 		if (val == 0)
2205 			error = EINVAL;
2206 		break;
2207 	default:
2208 		error = ENOENT;
2209 		break;
2210 	}
2211 	return (error);
2212 }
2213 
2214 static int
2215 svm_getcap(void *arg, int vcpu, int type, int *retval)
2216 {
2217 	struct svm_softc *sc;
2218 	int error;
2219 
2220 	sc = arg;
2221 	error = 0;
2222 
2223 	switch (type) {
2224 	case VM_CAP_HALT_EXIT:
2225 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2226 		    VMCB_INTCPT_HLT);
2227 		break;
2228 	case VM_CAP_PAUSE_EXIT:
2229 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2230 		    VMCB_INTCPT_PAUSE);
2231 		break;
2232 	case VM_CAP_UNRESTRICTED_GUEST:
2233 		*retval = 1;	/* unrestricted guest is always enabled */
2234 		break;
2235 	default:
2236 		error = ENOENT;
2237 		break;
2238 	}
2239 	return (error);
2240 }
2241 
2242 static struct vlapic *
2243 svm_vlapic_init(void *arg, int vcpuid)
2244 {
2245 	struct svm_softc *svm_sc;
2246 	struct vlapic *vlapic;
2247 
2248 	svm_sc = arg;
2249 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2250 	vlapic->vm = svm_sc->vm;
2251 	vlapic->vcpuid = vcpuid;
2252 	vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2253 
2254 	vlapic_init(vlapic);
2255 
2256 	return (vlapic);
2257 }
2258 
2259 static void
2260 svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2261 {
2262 
2263         vlapic_cleanup(vlapic);
2264         free(vlapic, M_SVM_VLAPIC);
2265 }
2266 
2267 struct vmm_ops vmm_ops_amd = {
2268 	svm_init,
2269 	svm_cleanup,
2270 	svm_restore,
2271 	svm_vminit,
2272 	svm_vmrun,
2273 	svm_vmcleanup,
2274 	svm_getreg,
2275 	svm_setreg,
2276 	vmcb_getdesc,
2277 	vmcb_setdesc,
2278 	svm_getcap,
2279 	svm_setcap,
2280 	svm_npt_alloc,
2281 	svm_npt_free,
2282 	svm_vlapic_init,
2283 	svm_vlapic_cleanup
2284 };
2285