xref: /freebsd/sys/amd64/vmm/amd/svm.c (revision a50d73d5782a351ad83e8d1f84d11720a12e70d3)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com)
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_bhyve_snapshot.h"
33 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/smp.h>
37 #include <sys/kernel.h>
38 #include <sys/malloc.h>
39 #include <sys/pcpu.h>
40 #include <sys/proc.h>
41 #include <sys/reg.h>
42 #include <sys/smr.h>
43 #include <sys/sysctl.h>
44 
45 #include <vm/vm.h>
46 #include <vm/pmap.h>
47 
48 #include <machine/cpufunc.h>
49 #include <machine/psl.h>
50 #include <machine/md_var.h>
51 #include <machine/specialreg.h>
52 #include <machine/smp.h>
53 #include <machine/vmm.h>
54 #include <machine/vmm_dev.h>
55 #include <machine/vmm_instruction_emul.h>
56 #include <machine/vmm_snapshot.h>
57 
58 #include "vmm_lapic.h"
59 #include "vmm_stat.h"
60 #include "vmm_ktr.h"
61 #include "vmm_ioport.h"
62 #include "vatpic.h"
63 #include "vlapic.h"
64 #include "vlapic_priv.h"
65 
66 #include "x86.h"
67 #include "vmcb.h"
68 #include "svm.h"
69 #include "svm_softc.h"
70 #include "svm_msr.h"
71 #include "npt.h"
72 
73 SYSCTL_DECL(_hw_vmm);
74 SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
75     NULL);
76 
77 /*
78  * SVM CPUID function 0x8000_000A, edx bit decoding.
79  */
80 #define AMD_CPUID_SVM_NP		BIT(0)  /* Nested paging or RVI */
81 #define AMD_CPUID_SVM_LBR		BIT(1)  /* Last branch virtualization */
82 #define AMD_CPUID_SVM_SVML		BIT(2)  /* SVM lock */
83 #define AMD_CPUID_SVM_NRIP_SAVE		BIT(3)  /* Next RIP is saved */
84 #define AMD_CPUID_SVM_TSC_RATE		BIT(4)  /* TSC rate control. */
85 #define AMD_CPUID_SVM_VMCB_CLEAN	BIT(5)  /* VMCB state caching */
86 #define AMD_CPUID_SVM_FLUSH_BY_ASID	BIT(6)  /* Flush by ASID */
87 #define AMD_CPUID_SVM_DECODE_ASSIST	BIT(7)  /* Decode assist */
88 #define AMD_CPUID_SVM_PAUSE_INC		BIT(10) /* Pause intercept filter. */
89 #define AMD_CPUID_SVM_PAUSE_FTH		BIT(12) /* Pause filter threshold */
90 #define	AMD_CPUID_SVM_AVIC		BIT(13)	/* AVIC present */
91 
92 #define	VMCB_CACHE_DEFAULT	(VMCB_CACHE_ASID 	|	\
93 				VMCB_CACHE_IOPM		|	\
94 				VMCB_CACHE_I		|	\
95 				VMCB_CACHE_TPR		|	\
96 				VMCB_CACHE_CR2		|	\
97 				VMCB_CACHE_CR		|	\
98 				VMCB_CACHE_DR		|	\
99 				VMCB_CACHE_DT		|	\
100 				VMCB_CACHE_SEG		|	\
101 				VMCB_CACHE_NP)
102 
103 static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT;
104 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean,
105     0, NULL);
106 
107 static MALLOC_DEFINE(M_SVM, "svm", "svm");
108 static MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic");
109 
110 static uint32_t svm_feature = ~0U;	/* AMD SVM features. */
111 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RDTUN, &svm_feature, 0,
112     "SVM features advertised by CPUID.8000000AH:EDX");
113 
114 static int disable_npf_assist;
115 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN,
116     &disable_npf_assist, 0, NULL);
117 
118 /* Maximum ASIDs supported by the processor */
119 static uint32_t nasid;
120 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RDTUN, &nasid, 0,
121     "Number of ASIDs supported by this processor");
122 
123 /* Current ASID generation for each host cpu */
124 static struct asid asid[MAXCPU];
125 
126 /*
127  * SVM host state saved area of size 4KB for each core.
128  */
129 static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE);
130 
131 static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery");
132 static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry");
133 static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window");
134 
135 static int svm_getdesc(void *vcpui, int reg, struct seg_desc *desc);
136 static int svm_setreg(void *vcpui, int ident, uint64_t val);
137 
138 static __inline int
139 flush_by_asid(void)
140 {
141 
142 	return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID);
143 }
144 
145 static __inline int
146 decode_assist(void)
147 {
148 
149 	return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST);
150 }
151 
152 static void
153 svm_disable(void *arg __unused)
154 {
155 	uint64_t efer;
156 
157 	efer = rdmsr(MSR_EFER);
158 	efer &= ~EFER_SVM;
159 	wrmsr(MSR_EFER, efer);
160 }
161 
162 /*
163  * Disable SVM on all CPUs.
164  */
165 static int
166 svm_modcleanup(void)
167 {
168 
169 	smp_rendezvous(NULL, svm_disable, NULL, NULL);
170 	return (0);
171 }
172 
173 /*
174  * Verify that all the features required by bhyve are available.
175  */
176 static int
177 check_svm_features(void)
178 {
179 	u_int regs[4];
180 
181 	/* CPUID Fn8000_000A is for SVM */
182 	do_cpuid(0x8000000A, regs);
183 	svm_feature &= regs[3];
184 
185 	/*
186 	 * The number of ASIDs can be configured to be less than what is
187 	 * supported by the hardware but not more.
188 	 */
189 	if (nasid == 0 || nasid > regs[1])
190 		nasid = regs[1];
191 	KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid));
192 
193 	/* bhyve requires the Nested Paging feature */
194 	if (!(svm_feature & AMD_CPUID_SVM_NP)) {
195 		printf("SVM: Nested Paging feature not available.\n");
196 		return (ENXIO);
197 	}
198 
199 	/* bhyve requires the NRIP Save feature */
200 	if (!(svm_feature & AMD_CPUID_SVM_NRIP_SAVE)) {
201 		printf("SVM: NRIP Save feature not available.\n");
202 		return (ENXIO);
203 	}
204 
205 	return (0);
206 }
207 
208 static void
209 svm_enable(void *arg __unused)
210 {
211 	uint64_t efer;
212 
213 	efer = rdmsr(MSR_EFER);
214 	efer |= EFER_SVM;
215 	wrmsr(MSR_EFER, efer);
216 
217 	wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave[curcpu]));
218 }
219 
220 /*
221  * Return 1 if SVM is enabled on this processor and 0 otherwise.
222  */
223 static int
224 svm_available(void)
225 {
226 	uint64_t msr;
227 
228 	/* Section 15.4 Enabling SVM from APM2. */
229 	if ((amd_feature2 & AMDID2_SVM) == 0) {
230 		printf("SVM: not available.\n");
231 		return (0);
232 	}
233 
234 	msr = rdmsr(MSR_VM_CR);
235 	if ((msr & VM_CR_SVMDIS) != 0) {
236 		printf("SVM: disabled by BIOS.\n");
237 		return (0);
238 	}
239 
240 	return (1);
241 }
242 
243 static int
244 svm_modinit(int ipinum)
245 {
246 	int error, cpu;
247 
248 	if (!svm_available())
249 		return (ENXIO);
250 
251 	error = check_svm_features();
252 	if (error)
253 		return (error);
254 
255 	vmcb_clean &= VMCB_CACHE_DEFAULT;
256 
257 	for (cpu = 0; cpu < MAXCPU; cpu++) {
258 		/*
259 		 * Initialize the host ASIDs to their "highest" valid values.
260 		 *
261 		 * The next ASID allocation will rollover both 'gen' and 'num'
262 		 * and start off the sequence at {1,1}.
263 		 */
264 		asid[cpu].gen = ~0UL;
265 		asid[cpu].num = nasid - 1;
266 	}
267 
268 	svm_msr_init();
269 	svm_npt_init(ipinum);
270 
271 	/* Enable SVM on all CPUs */
272 	smp_rendezvous(NULL, svm_enable, NULL, NULL);
273 
274 	return (0);
275 }
276 
277 static void
278 svm_modresume(void)
279 {
280 
281 	svm_enable(NULL);
282 }
283 
284 #ifdef BHYVE_SNAPSHOT
285 void
286 svm_set_tsc_offset(struct svm_vcpu *vcpu, uint64_t offset)
287 {
288 	struct vmcb_ctrl *ctrl;
289 
290 	ctrl = svm_get_vmcb_ctrl(vcpu);
291 	ctrl->tsc_offset = offset;
292 
293 	svm_set_dirty(vcpu, VMCB_CACHE_I);
294 	SVM_CTR1(vcpu, "tsc offset changed to %#lx", offset);
295 
296 	vm_set_tsc_offset(vcpu->vcpu, offset);
297 }
298 #endif
299 
300 /* Pentium compatible MSRs */
301 #define MSR_PENTIUM_START 	0
302 #define MSR_PENTIUM_END 	0x1FFF
303 /* AMD 6th generation and Intel compatible MSRs */
304 #define MSR_AMD6TH_START 	0xC0000000UL
305 #define MSR_AMD6TH_END 		0xC0001FFFUL
306 /* AMD 7th and 8th generation compatible MSRs */
307 #define MSR_AMD7TH_START 	0xC0010000UL
308 #define MSR_AMD7TH_END 		0xC0011FFFUL
309 
310 /*
311  * Get the index and bit position for a MSR in permission bitmap.
312  * Two bits are used for each MSR: lower bit for read and higher bit for write.
313  */
314 static int
315 svm_msr_index(uint64_t msr, int *index, int *bit)
316 {
317 	uint32_t base, off;
318 
319 	*index = -1;
320 	*bit = (msr % 4) * 2;
321 	base = 0;
322 
323 	if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) {
324 		*index = msr / 4;
325 		return (0);
326 	}
327 
328 	base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1);
329 	if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) {
330 		off = (msr - MSR_AMD6TH_START);
331 		*index = (off + base) / 4;
332 		return (0);
333 	}
334 
335 	base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1);
336 	if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) {
337 		off = (msr - MSR_AMD7TH_START);
338 		*index = (off + base) / 4;
339 		return (0);
340 	}
341 
342 	return (EINVAL);
343 }
344 
345 /*
346  * Allow vcpu to read or write the 'msr' without trapping into the hypervisor.
347  */
348 static void
349 svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write)
350 {
351 	int index, bit, error __diagused;
352 
353 	error = svm_msr_index(msr, &index, &bit);
354 	KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr));
355 	KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE,
356 	    ("%s: invalid index %d for msr %#lx", __func__, index, msr));
357 	KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d "
358 	    "msr %#lx", __func__, bit, msr));
359 
360 	if (read)
361 		perm_bitmap[index] &= ~(1UL << bit);
362 
363 	if (write)
364 		perm_bitmap[index] &= ~(2UL << bit);
365 }
366 
367 static void
368 svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr)
369 {
370 
371 	svm_msr_perm(perm_bitmap, msr, true, true);
372 }
373 
374 static void
375 svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr)
376 {
377 
378 	svm_msr_perm(perm_bitmap, msr, true, false);
379 }
380 
381 static __inline int
382 svm_get_intercept(struct svm_vcpu *vcpu, int idx, uint32_t bitmask)
383 {
384 	struct vmcb_ctrl *ctrl;
385 
386 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
387 
388 	ctrl = svm_get_vmcb_ctrl(vcpu);
389 	return (ctrl->intercept[idx] & bitmask ? 1 : 0);
390 }
391 
392 static __inline void
393 svm_set_intercept(struct svm_vcpu *vcpu, int idx, uint32_t bitmask, int enabled)
394 {
395 	struct vmcb_ctrl *ctrl;
396 	uint32_t oldval;
397 
398 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
399 
400 	ctrl = svm_get_vmcb_ctrl(vcpu);
401 	oldval = ctrl->intercept[idx];
402 
403 	if (enabled)
404 		ctrl->intercept[idx] |= bitmask;
405 	else
406 		ctrl->intercept[idx] &= ~bitmask;
407 
408 	if (ctrl->intercept[idx] != oldval) {
409 		svm_set_dirty(vcpu, VMCB_CACHE_I);
410 		SVM_CTR3(vcpu, "intercept[%d] modified from %#x to %#x", idx,
411 		    oldval, ctrl->intercept[idx]);
412 	}
413 }
414 
415 static __inline void
416 svm_disable_intercept(struct svm_vcpu *vcpu, int off, uint32_t bitmask)
417 {
418 
419 	svm_set_intercept(vcpu, off, bitmask, 0);
420 }
421 
422 static __inline void
423 svm_enable_intercept(struct svm_vcpu *vcpu, int off, uint32_t bitmask)
424 {
425 
426 	svm_set_intercept(vcpu, off, bitmask, 1);
427 }
428 
429 static void
430 vmcb_init(struct svm_softc *sc, struct svm_vcpu *vcpu, uint64_t iopm_base_pa,
431     uint64_t msrpm_base_pa, uint64_t np_pml4)
432 {
433 	struct vmcb_ctrl *ctrl;
434 	struct vmcb_state *state;
435 	uint32_t mask;
436 	int n;
437 
438 	ctrl = svm_get_vmcb_ctrl(vcpu);
439 	state = svm_get_vmcb_state(vcpu);
440 
441 	ctrl->iopm_base_pa = iopm_base_pa;
442 	ctrl->msrpm_base_pa = msrpm_base_pa;
443 
444 	/* Enable nested paging */
445 	ctrl->np_enable = 1;
446 	ctrl->n_cr3 = np_pml4;
447 
448 	/*
449 	 * Intercept accesses to the control registers that are not shadowed
450 	 * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8.
451 	 */
452 	for (n = 0; n < 16; n++) {
453 		mask = (BIT(n) << 16) | BIT(n);
454 		if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8)
455 			svm_disable_intercept(vcpu, VMCB_CR_INTCPT, mask);
456 		else
457 			svm_enable_intercept(vcpu, VMCB_CR_INTCPT, mask);
458 	}
459 
460 	/*
461 	 * Intercept everything when tracing guest exceptions otherwise
462 	 * just intercept machine check exception.
463 	 */
464 	if (vcpu_trace_exceptions(vcpu->vcpu)) {
465 		for (n = 0; n < 32; n++) {
466 			/*
467 			 * Skip unimplemented vectors in the exception bitmap.
468 			 */
469 			if (n == 2 || n == 9) {
470 				continue;
471 			}
472 			svm_enable_intercept(vcpu, VMCB_EXC_INTCPT, BIT(n));
473 		}
474 	} else {
475 		svm_enable_intercept(vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC));
476 	}
477 
478 	/* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */
479 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO);
480 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR);
481 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID);
482 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR);
483 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT);
484 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI);
485 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI);
486 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN);
487 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_FERR_FREEZE);
488 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INVD);
489 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INVLPGA);
490 
491 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR);
492 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT);
493 
494 	/*
495 	 * Intercept SVM instructions since AMD enables them in guests otherwise.
496 	 * Non-intercepted VMMCALL causes #UD, skip it.
497 	 */
498 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMLOAD);
499 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMSAVE);
500 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_STGI);
501 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_CLGI);
502 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_SKINIT);
503 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_ICEBP);
504 	if (vcpu_trap_wbinvd(vcpu->vcpu)) {
505 		svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT,
506 		    VMCB_INTCPT_WBINVD);
507 	}
508 
509 	/*
510 	 * From section "Canonicalization and Consistency Checks" in APMv2
511 	 * the VMRUN intercept bit must be set to pass the consistency check.
512 	 */
513 	svm_enable_intercept(vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
514 
515 	/*
516 	 * The ASID will be set to a non-zero value just before VMRUN.
517 	 */
518 	ctrl->asid = 0;
519 
520 	/*
521 	 * Section 15.21.1, Interrupt Masking in EFLAGS
522 	 * Section 15.21.2, Virtualizing APIC.TPR
523 	 *
524 	 * This must be set for %rflag and %cr8 isolation of guest and host.
525 	 */
526 	ctrl->v_intr_masking = 1;
527 
528 	/* Enable Last Branch Record aka LBR for debugging */
529 	ctrl->lbr_virt_en = 1;
530 	state->dbgctl = BIT(0);
531 
532 	/* EFER_SVM must always be set when the guest is executing */
533 	state->efer = EFER_SVM;
534 
535 	/* Set up the PAT to power-on state */
536 	state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK)	|
537 	    PAT_VALUE(1, PAT_WRITE_THROUGH)	|
538 	    PAT_VALUE(2, PAT_UNCACHED)		|
539 	    PAT_VALUE(3, PAT_UNCACHEABLE)	|
540 	    PAT_VALUE(4, PAT_WRITE_BACK)	|
541 	    PAT_VALUE(5, PAT_WRITE_THROUGH)	|
542 	    PAT_VALUE(6, PAT_UNCACHED)		|
543 	    PAT_VALUE(7, PAT_UNCACHEABLE);
544 
545 	/* Set up DR6/7 to power-on state */
546 	state->dr6 = DBREG_DR6_RESERVED1;
547 	state->dr7 = DBREG_DR7_RESERVED1;
548 }
549 
550 /*
551  * Initialize a virtual machine.
552  */
553 static void *
554 svm_init(struct vm *vm, pmap_t pmap)
555 {
556 	struct svm_softc *svm_sc;
557 
558 	svm_sc = malloc(sizeof (*svm_sc), M_SVM, M_WAITOK | M_ZERO);
559 
560 	svm_sc->msr_bitmap = contigmalloc(SVM_MSR_BITMAP_SIZE, M_SVM,
561 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
562 	if (svm_sc->msr_bitmap == NULL)
563 		panic("contigmalloc of SVM MSR bitmap failed");
564 	svm_sc->iopm_bitmap = contigmalloc(SVM_IO_BITMAP_SIZE, M_SVM,
565 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
566 	if (svm_sc->iopm_bitmap == NULL)
567 		panic("contigmalloc of SVM IO bitmap failed");
568 
569 	svm_sc->vm = vm;
570 	svm_sc->nptp = vtophys(pmap->pm_pmltop);
571 
572 	/*
573 	 * Intercept read and write accesses to all MSRs.
574 	 */
575 	memset(svm_sc->msr_bitmap, 0xFF, SVM_MSR_BITMAP_SIZE);
576 
577 	/*
578 	 * Access to the following MSRs is redirected to the VMCB when the
579 	 * guest is executing. Therefore it is safe to allow the guest to
580 	 * read/write these MSRs directly without hypervisor involvement.
581 	 */
582 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
583 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
584 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
585 
586 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
587 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
588 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
589 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
590 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
591 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
592 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
593 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
594 
595 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
596 
597 	/*
598 	 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
599 	 */
600 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
601 
602 	/* Intercept access to all I/O ports. */
603 	memset(svm_sc->iopm_bitmap, 0xFF, SVM_IO_BITMAP_SIZE);
604 
605 	return (svm_sc);
606 }
607 
608 static void *
609 svm_vcpu_init(void *vmi, struct vcpu *vcpu1, int vcpuid)
610 {
611 	struct svm_softc *sc = vmi;
612 	struct svm_vcpu *vcpu;
613 
614 	vcpu = malloc(sizeof(*vcpu), M_SVM, M_WAITOK | M_ZERO);
615 	vcpu->sc = sc;
616 	vcpu->vcpu = vcpu1;
617 	vcpu->vcpuid = vcpuid;
618 	vcpu->vmcb = malloc_aligned(sizeof(struct vmcb), PAGE_SIZE, M_SVM,
619 	    M_WAITOK | M_ZERO);
620 	vcpu->nextrip = ~0;
621 	vcpu->lastcpu = NOCPU;
622 	vcpu->vmcb_pa = vtophys(vcpu->vmcb);
623 	vmcb_init(sc, vcpu, vtophys(sc->iopm_bitmap), vtophys(sc->msr_bitmap),
624 	    sc->nptp);
625 	svm_msr_guest_init(sc, vcpu);
626 	return (vcpu);
627 }
628 
629 /*
630  * Collateral for a generic SVM VM-exit.
631  */
632 static void
633 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
634 {
635 
636 	vme->exitcode = VM_EXITCODE_SVM;
637 	vme->u.svm.exitcode = code;
638 	vme->u.svm.exitinfo1 = info1;
639 	vme->u.svm.exitinfo2 = info2;
640 }
641 
642 static int
643 svm_cpl(struct vmcb_state *state)
644 {
645 
646 	/*
647 	 * From APMv2:
648 	 *   "Retrieve the CPL from the CPL field in the VMCB, not
649 	 *    from any segment DPL"
650 	 */
651 	return (state->cpl);
652 }
653 
654 static enum vm_cpu_mode
655 svm_vcpu_mode(struct vmcb *vmcb)
656 {
657 	struct vmcb_segment seg;
658 	struct vmcb_state *state;
659 	int error __diagused;
660 
661 	state = &vmcb->state;
662 
663 	if (state->efer & EFER_LMA) {
664 		error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
665 		KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
666 		    error));
667 
668 		/*
669 		 * Section 4.8.1 for APM2, check if Code Segment has
670 		 * Long attribute set in descriptor.
671 		 */
672 		if (seg.attrib & VMCB_CS_ATTRIB_L)
673 			return (CPU_MODE_64BIT);
674 		else
675 			return (CPU_MODE_COMPATIBILITY);
676 	} else  if (state->cr0 & CR0_PE) {
677 		return (CPU_MODE_PROTECTED);
678 	} else {
679 		return (CPU_MODE_REAL);
680 	}
681 }
682 
683 static enum vm_paging_mode
684 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
685 {
686 
687 	if ((cr0 & CR0_PG) == 0)
688 		return (PAGING_MODE_FLAT);
689 	if ((cr4 & CR4_PAE) == 0)
690 		return (PAGING_MODE_32);
691 	if (efer & EFER_LME)
692 		return (PAGING_MODE_64);
693 	else
694 		return (PAGING_MODE_PAE);
695 }
696 
697 /*
698  * ins/outs utility routines
699  */
700 static uint64_t
701 svm_inout_str_index(struct svm_regctx *regs, int in)
702 {
703 	uint64_t val;
704 
705 	val = in ? regs->sctx_rdi : regs->sctx_rsi;
706 
707 	return (val);
708 }
709 
710 static uint64_t
711 svm_inout_str_count(struct svm_regctx *regs, int rep)
712 {
713 	uint64_t val;
714 
715 	val = rep ? regs->sctx_rcx : 1;
716 
717 	return (val);
718 }
719 
720 static void
721 svm_inout_str_seginfo(struct svm_vcpu *vcpu, int64_t info1, int in,
722     struct vm_inout_str *vis)
723 {
724 	int error __diagused, s;
725 
726 	if (in) {
727 		vis->seg_name = VM_REG_GUEST_ES;
728 	} else {
729 		/* The segment field has standard encoding */
730 		s = (info1 >> 10) & 0x7;
731 		vis->seg_name = vm_segment_name(s);
732 	}
733 
734 	error = svm_getdesc(vcpu, vis->seg_name, &vis->seg_desc);
735 	KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
736 }
737 
738 static int
739 svm_inout_str_addrsize(uint64_t info1)
740 {
741         uint32_t size;
742 
743         size = (info1 >> 7) & 0x7;
744         switch (size) {
745         case 1:
746                 return (2);     /* 16 bit */
747         case 2:
748                 return (4);     /* 32 bit */
749         case 4:
750                 return (8);     /* 64 bit */
751         default:
752                 panic("%s: invalid size encoding %d", __func__, size);
753         }
754 }
755 
756 static void
757 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
758 {
759 	struct vmcb_state *state;
760 
761 	state = &vmcb->state;
762 	paging->cr3 = state->cr3;
763 	paging->cpl = svm_cpl(state);
764 	paging->cpu_mode = svm_vcpu_mode(vmcb);
765 	paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
766 	    state->efer);
767 }
768 
769 #define	UNHANDLED 0
770 
771 /*
772  * Handle guest I/O intercept.
773  */
774 static int
775 svm_handle_io(struct svm_vcpu *vcpu, struct vm_exit *vmexit)
776 {
777 	struct vmcb_ctrl *ctrl;
778 	struct vmcb_state *state;
779 	struct svm_regctx *regs;
780 	struct vm_inout_str *vis;
781 	uint64_t info1;
782 	int inout_string;
783 
784 	state = svm_get_vmcb_state(vcpu);
785 	ctrl  = svm_get_vmcb_ctrl(vcpu);
786 	regs  = svm_get_guest_regctx(vcpu);
787 
788 	info1 = ctrl->exitinfo1;
789 	inout_string = info1 & BIT(2) ? 1 : 0;
790 
791 	/*
792 	 * The effective segment number in EXITINFO1[12:10] is populated
793 	 * only if the processor has the DecodeAssist capability.
794 	 *
795 	 * XXX this is not specified explicitly in APMv2 but can be verified
796 	 * empirically.
797 	 */
798 	if (inout_string && !decode_assist())
799 		return (UNHANDLED);
800 
801 	vmexit->exitcode 	= VM_EXITCODE_INOUT;
802 	vmexit->u.inout.in 	= (info1 & BIT(0)) ? 1 : 0;
803 	vmexit->u.inout.string 	= inout_string;
804 	vmexit->u.inout.rep 	= (info1 & BIT(3)) ? 1 : 0;
805 	vmexit->u.inout.bytes 	= (info1 >> 4) & 0x7;
806 	vmexit->u.inout.port 	= (uint16_t)(info1 >> 16);
807 	vmexit->u.inout.eax 	= (uint32_t)(state->rax);
808 
809 	if (inout_string) {
810 		vmexit->exitcode = VM_EXITCODE_INOUT_STR;
811 		vis = &vmexit->u.inout_str;
812 		svm_paging_info(svm_get_vmcb(vcpu), &vis->paging);
813 		vis->rflags = state->rflags;
814 		vis->cr0 = state->cr0;
815 		vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
816 		vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
817 		vis->addrsize = svm_inout_str_addrsize(info1);
818 		svm_inout_str_seginfo(vcpu, info1, vmexit->u.inout.in, vis);
819 	}
820 
821 	return (UNHANDLED);
822 }
823 
824 static int
825 npf_fault_type(uint64_t exitinfo1)
826 {
827 
828 	if (exitinfo1 & VMCB_NPF_INFO1_W)
829 		return (VM_PROT_WRITE);
830 	else if (exitinfo1 & VMCB_NPF_INFO1_ID)
831 		return (VM_PROT_EXECUTE);
832 	else
833 		return (VM_PROT_READ);
834 }
835 
836 static bool
837 svm_npf_emul_fault(uint64_t exitinfo1)
838 {
839 
840 	if (exitinfo1 & VMCB_NPF_INFO1_ID) {
841 		return (false);
842 	}
843 
844 	if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
845 		return (false);
846 	}
847 
848 	if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
849 		return (false);
850 	}
851 
852 	return (true);
853 }
854 
855 static void
856 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
857 {
858 	struct vm_guest_paging *paging;
859 	struct vmcb_segment seg;
860 	struct vmcb_ctrl *ctrl;
861 	char *inst_bytes;
862 	int error __diagused, inst_len;
863 
864 	ctrl = &vmcb->ctrl;
865 	paging = &vmexit->u.inst_emul.paging;
866 
867 	vmexit->exitcode = VM_EXITCODE_INST_EMUL;
868 	vmexit->u.inst_emul.gpa = gpa;
869 	vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
870 	svm_paging_info(vmcb, paging);
871 
872 	error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
873 	KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
874 
875 	switch(paging->cpu_mode) {
876 	case CPU_MODE_REAL:
877 		vmexit->u.inst_emul.cs_base = seg.base;
878 		vmexit->u.inst_emul.cs_d = 0;
879 		break;
880 	case CPU_MODE_PROTECTED:
881 	case CPU_MODE_COMPATIBILITY:
882 		vmexit->u.inst_emul.cs_base = seg.base;
883 
884 		/*
885 		 * Section 4.8.1 of APM2, Default Operand Size or D bit.
886 		 */
887 		vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
888 		    1 : 0;
889 		break;
890 	default:
891 		vmexit->u.inst_emul.cs_base = 0;
892 		vmexit->u.inst_emul.cs_d = 0;
893 		break;
894 	}
895 
896 	/*
897 	 * Copy the instruction bytes into 'vie' if available.
898 	 */
899 	if (decode_assist() && !disable_npf_assist) {
900 		inst_len = ctrl->inst_len;
901 		inst_bytes = ctrl->inst_bytes;
902 	} else {
903 		inst_len = 0;
904 		inst_bytes = NULL;
905 	}
906 	vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
907 }
908 
909 #ifdef KTR
910 static const char *
911 intrtype_to_str(int intr_type)
912 {
913 	switch (intr_type) {
914 	case VMCB_EVENTINJ_TYPE_INTR:
915 		return ("hwintr");
916 	case VMCB_EVENTINJ_TYPE_NMI:
917 		return ("nmi");
918 	case VMCB_EVENTINJ_TYPE_INTn:
919 		return ("swintr");
920 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
921 		return ("exception");
922 	default:
923 		panic("%s: unknown intr_type %d", __func__, intr_type);
924 	}
925 }
926 #endif
927 
928 /*
929  * Inject an event to vcpu as described in section 15.20, "Event injection".
930  */
931 static void
932 svm_eventinject(struct svm_vcpu *vcpu, int intr_type, int vector,
933     uint32_t error, bool ec_valid)
934 {
935 	struct vmcb_ctrl *ctrl;
936 
937 	ctrl = svm_get_vmcb_ctrl(vcpu);
938 
939 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
940 	    ("%s: event already pending %#lx", __func__, ctrl->eventinj));
941 
942 	KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
943 	    __func__, vector));
944 
945 	switch (intr_type) {
946 	case VMCB_EVENTINJ_TYPE_INTR:
947 	case VMCB_EVENTINJ_TYPE_NMI:
948 	case VMCB_EVENTINJ_TYPE_INTn:
949 		break;
950 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
951 		if (vector >= 0 && vector <= 31 && vector != 2)
952 			break;
953 		/* FALLTHROUGH */
954 	default:
955 		panic("%s: invalid intr_type/vector: %d/%d", __func__,
956 		    intr_type, vector);
957 	}
958 	ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
959 	if (ec_valid) {
960 		ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
961 		ctrl->eventinj |= (uint64_t)error << 32;
962 		SVM_CTR3(vcpu, "Injecting %s at vector %d errcode %#x",
963 		    intrtype_to_str(intr_type), vector, error);
964 	} else {
965 		SVM_CTR2(vcpu, "Injecting %s at vector %d",
966 		    intrtype_to_str(intr_type), vector);
967 	}
968 }
969 
970 static void
971 svm_update_virqinfo(struct svm_vcpu *vcpu)
972 {
973 	struct vlapic *vlapic;
974 	struct vmcb_ctrl *ctrl;
975 
976 	vlapic = vm_lapic(vcpu->vcpu);
977 	ctrl = svm_get_vmcb_ctrl(vcpu);
978 
979 	/* Update %cr8 in the emulated vlapic */
980 	vlapic_set_cr8(vlapic, ctrl->v_tpr);
981 
982 	/* Virtual interrupt injection is not used. */
983 	KASSERT(ctrl->v_intr_vector == 0, ("%s: invalid "
984 	    "v_intr_vector %d", __func__, ctrl->v_intr_vector));
985 }
986 
987 static void
988 svm_save_intinfo(struct svm_softc *svm_sc, struct svm_vcpu *vcpu)
989 {
990 	struct vmcb_ctrl *ctrl;
991 	uint64_t intinfo;
992 
993 	ctrl = svm_get_vmcb_ctrl(vcpu);
994 	intinfo = ctrl->exitintinfo;
995 	if (!VMCB_EXITINTINFO_VALID(intinfo))
996 		return;
997 
998 	/*
999 	 * From APMv2, Section "Intercepts during IDT interrupt delivery"
1000 	 *
1001 	 * If a #VMEXIT happened during event delivery then record the event
1002 	 * that was being delivered.
1003 	 */
1004 	SVM_CTR2(vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n", intinfo,
1005 	    VMCB_EXITINTINFO_VECTOR(intinfo));
1006 	vmm_stat_incr(vcpu->vcpu, VCPU_EXITINTINFO, 1);
1007 	vm_exit_intinfo(vcpu->vcpu, intinfo);
1008 }
1009 
1010 #ifdef INVARIANTS
1011 static __inline int
1012 vintr_intercept_enabled(struct svm_vcpu *vcpu)
1013 {
1014 
1015 	return (svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR));
1016 }
1017 #endif
1018 
1019 static __inline void
1020 enable_intr_window_exiting(struct svm_vcpu *vcpu)
1021 {
1022 	struct vmcb_ctrl *ctrl;
1023 
1024 	ctrl = svm_get_vmcb_ctrl(vcpu);
1025 
1026 	if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
1027 		KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
1028 		KASSERT(vintr_intercept_enabled(vcpu),
1029 		    ("%s: vintr intercept should be enabled", __func__));
1030 		return;
1031 	}
1032 
1033 	SVM_CTR0(vcpu, "Enable intr window exiting");
1034 	ctrl->v_irq = 1;
1035 	ctrl->v_ign_tpr = 1;
1036 	ctrl->v_intr_vector = 0;
1037 	svm_set_dirty(vcpu, VMCB_CACHE_TPR);
1038 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1039 }
1040 
1041 static __inline void
1042 disable_intr_window_exiting(struct svm_vcpu *vcpu)
1043 {
1044 	struct vmcb_ctrl *ctrl;
1045 
1046 	ctrl = svm_get_vmcb_ctrl(vcpu);
1047 
1048 	if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1049 		KASSERT(!vintr_intercept_enabled(vcpu),
1050 		    ("%s: vintr intercept should be disabled", __func__));
1051 		return;
1052 	}
1053 
1054 	SVM_CTR0(vcpu, "Disable intr window exiting");
1055 	ctrl->v_irq = 0;
1056 	ctrl->v_intr_vector = 0;
1057 	svm_set_dirty(vcpu, VMCB_CACHE_TPR);
1058 	svm_disable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1059 }
1060 
1061 static int
1062 svm_modify_intr_shadow(struct svm_vcpu *vcpu, uint64_t val)
1063 {
1064 	struct vmcb_ctrl *ctrl;
1065 	int oldval, newval;
1066 
1067 	ctrl = svm_get_vmcb_ctrl(vcpu);
1068 	oldval = ctrl->intr_shadow;
1069 	newval = val ? 1 : 0;
1070 	if (newval != oldval) {
1071 		ctrl->intr_shadow = newval;
1072 		SVM_CTR1(vcpu, "Setting intr_shadow to %d", newval);
1073 	}
1074 	return (0);
1075 }
1076 
1077 static int
1078 svm_get_intr_shadow(struct svm_vcpu *vcpu, uint64_t *val)
1079 {
1080 	struct vmcb_ctrl *ctrl;
1081 
1082 	ctrl = svm_get_vmcb_ctrl(vcpu);
1083 	*val = ctrl->intr_shadow;
1084 	return (0);
1085 }
1086 
1087 /*
1088  * Once an NMI is injected it blocks delivery of further NMIs until the handler
1089  * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1090  * to track when the vcpu is done handling the NMI.
1091  */
1092 static int
1093 nmi_blocked(struct svm_vcpu *vcpu)
1094 {
1095 	int blocked;
1096 
1097 	blocked = svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1098 	return (blocked);
1099 }
1100 
1101 static void
1102 enable_nmi_blocking(struct svm_vcpu *vcpu)
1103 {
1104 
1105 	KASSERT(!nmi_blocked(vcpu), ("vNMI already blocked"));
1106 	SVM_CTR0(vcpu, "vNMI blocking enabled");
1107 	svm_enable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1108 }
1109 
1110 static void
1111 clear_nmi_blocking(struct svm_vcpu *vcpu)
1112 {
1113 	int error __diagused;
1114 
1115 	KASSERT(nmi_blocked(vcpu), ("vNMI already unblocked"));
1116 	SVM_CTR0(vcpu, "vNMI blocking cleared");
1117 	/*
1118 	 * When the IRET intercept is cleared the vcpu will attempt to execute
1119 	 * the "iret" when it runs next. However, it is possible to inject
1120 	 * another NMI into the vcpu before the "iret" has actually executed.
1121 	 *
1122 	 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1123 	 * it will trap back into the hypervisor. If an NMI is pending for
1124 	 * the vcpu it will be injected into the guest.
1125 	 *
1126 	 * XXX this needs to be fixed
1127 	 */
1128 	svm_disable_intercept(vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1129 
1130 	/*
1131 	 * Set 'intr_shadow' to prevent an NMI from being injected on the
1132 	 * immediate VMRUN.
1133 	 */
1134 	error = svm_modify_intr_shadow(vcpu, 1);
1135 	KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1136 }
1137 
1138 #define	EFER_MBZ_BITS	0xFFFFFFFFFFFF0200UL
1139 
1140 static int
1141 svm_write_efer(struct svm_softc *sc, struct svm_vcpu *vcpu, uint64_t newval,
1142     bool *retu)
1143 {
1144 	struct vm_exit *vme;
1145 	struct vmcb_state *state;
1146 	uint64_t changed, lma, oldval;
1147 	int error __diagused;
1148 
1149 	state = svm_get_vmcb_state(vcpu);
1150 
1151 	oldval = state->efer;
1152 	SVM_CTR2(vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval);
1153 
1154 	newval &= ~0xFE;		/* clear the Read-As-Zero (RAZ) bits */
1155 	changed = oldval ^ newval;
1156 
1157 	if (newval & EFER_MBZ_BITS)
1158 		goto gpf;
1159 
1160 	/* APMv2 Table 14-5 "Long-Mode Consistency Checks" */
1161 	if (changed & EFER_LME) {
1162 		if (state->cr0 & CR0_PG)
1163 			goto gpf;
1164 	}
1165 
1166 	/* EFER.LMA = EFER.LME & CR0.PG */
1167 	if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0)
1168 		lma = EFER_LMA;
1169 	else
1170 		lma = 0;
1171 
1172 	if ((newval & EFER_LMA) != lma)
1173 		goto gpf;
1174 
1175 	if (newval & EFER_NXE) {
1176 		if (!vm_cpuid_capability(vcpu->vcpu, VCC_NO_EXECUTE))
1177 			goto gpf;
1178 	}
1179 
1180 	/*
1181 	 * XXX bhyve does not enforce segment limits in 64-bit mode. Until
1182 	 * this is fixed flag guest attempt to set EFER_LMSLE as an error.
1183 	 */
1184 	if (newval & EFER_LMSLE) {
1185 		vme = vm_exitinfo(vcpu->vcpu);
1186 		vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0);
1187 		*retu = true;
1188 		return (0);
1189 	}
1190 
1191 	if (newval & EFER_FFXSR) {
1192 		if (!vm_cpuid_capability(vcpu->vcpu, VCC_FFXSR))
1193 			goto gpf;
1194 	}
1195 
1196 	if (newval & EFER_TCE) {
1197 		if (!vm_cpuid_capability(vcpu->vcpu, VCC_TCE))
1198 			goto gpf;
1199 	}
1200 
1201 	error = svm_setreg(vcpu, VM_REG_GUEST_EFER, newval);
1202 	KASSERT(error == 0, ("%s: error %d updating efer", __func__, error));
1203 	return (0);
1204 gpf:
1205 	vm_inject_gp(vcpu->vcpu);
1206 	return (0);
1207 }
1208 
1209 static int
1210 emulate_wrmsr(struct svm_softc *sc, struct svm_vcpu *vcpu, u_int num,
1211     uint64_t val, bool *retu)
1212 {
1213 	int error;
1214 
1215 	if (lapic_msr(num))
1216 		error = lapic_wrmsr(vcpu->vcpu, num, val, retu);
1217 	else if (num == MSR_EFER)
1218 		error = svm_write_efer(sc, vcpu, val, retu);
1219 	else
1220 		error = svm_wrmsr(vcpu, num, val, retu);
1221 
1222 	return (error);
1223 }
1224 
1225 static int
1226 emulate_rdmsr(struct svm_vcpu *vcpu, u_int num, bool *retu)
1227 {
1228 	struct vmcb_state *state;
1229 	struct svm_regctx *ctx;
1230 	uint64_t result;
1231 	int error;
1232 
1233 	if (lapic_msr(num))
1234 		error = lapic_rdmsr(vcpu->vcpu, num, &result, retu);
1235 	else
1236 		error = svm_rdmsr(vcpu, num, &result, retu);
1237 
1238 	if (error == 0) {
1239 		state = svm_get_vmcb_state(vcpu);
1240 		ctx = svm_get_guest_regctx(vcpu);
1241 		state->rax = result & 0xffffffff;
1242 		ctx->sctx_rdx = result >> 32;
1243 	}
1244 
1245 	return (error);
1246 }
1247 
1248 #ifdef KTR
1249 static const char *
1250 exit_reason_to_str(uint64_t reason)
1251 {
1252 	int i;
1253 	static char reasonbuf[32];
1254 	static const struct {
1255 		int reason;
1256 		const char *str;
1257 	} reasons[] = {
1258 		{ .reason = VMCB_EXIT_INVALID,	.str = "invalvmcb" },
1259 		{ .reason = VMCB_EXIT_SHUTDOWN,	.str = "shutdown" },
1260 		{ .reason = VMCB_EXIT_NPF, 	.str = "nptfault" },
1261 		{ .reason = VMCB_EXIT_PAUSE,	.str = "pause" },
1262 		{ .reason = VMCB_EXIT_HLT,	.str = "hlt" },
1263 		{ .reason = VMCB_EXIT_CPUID,	.str = "cpuid" },
1264 		{ .reason = VMCB_EXIT_IO,	.str = "inout" },
1265 		{ .reason = VMCB_EXIT_MC,	.str = "mchk" },
1266 		{ .reason = VMCB_EXIT_INTR,	.str = "extintr" },
1267 		{ .reason = VMCB_EXIT_NMI,	.str = "nmi" },
1268 		{ .reason = VMCB_EXIT_VINTR,	.str = "vintr" },
1269 		{ .reason = VMCB_EXIT_MSR,	.str = "msr" },
1270 		{ .reason = VMCB_EXIT_IRET,	.str = "iret" },
1271 		{ .reason = VMCB_EXIT_MONITOR,	.str = "monitor" },
1272 		{ .reason = VMCB_EXIT_MWAIT,	.str = "mwait" },
1273 		{ .reason = VMCB_EXIT_VMRUN,	.str = "vmrun" },
1274 		{ .reason = VMCB_EXIT_VMMCALL,	.str = "vmmcall" },
1275 		{ .reason = VMCB_EXIT_VMLOAD,	.str = "vmload" },
1276 		{ .reason = VMCB_EXIT_VMSAVE,	.str = "vmsave" },
1277 		{ .reason = VMCB_EXIT_STGI,	.str = "stgi" },
1278 		{ .reason = VMCB_EXIT_CLGI,	.str = "clgi" },
1279 		{ .reason = VMCB_EXIT_SKINIT,	.str = "skinit" },
1280 		{ .reason = VMCB_EXIT_ICEBP,	.str = "icebp" },
1281 		{ .reason = VMCB_EXIT_INVD,	.str = "invd" },
1282 		{ .reason = VMCB_EXIT_INVLPGA,	.str = "invlpga" },
1283 	};
1284 
1285 	for (i = 0; i < nitems(reasons); i++) {
1286 		if (reasons[i].reason == reason)
1287 			return (reasons[i].str);
1288 	}
1289 	snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1290 	return (reasonbuf);
1291 }
1292 #endif	/* KTR */
1293 
1294 /*
1295  * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1296  * that are due to instruction intercepts as well as MSR and IOIO intercepts
1297  * and exceptions caused by INT3, INTO and BOUND instructions.
1298  *
1299  * Return 1 if the nRIP is valid and 0 otherwise.
1300  */
1301 static int
1302 nrip_valid(uint64_t exitcode)
1303 {
1304 	switch (exitcode) {
1305 	case 0x00 ... 0x0F:	/* read of CR0 through CR15 */
1306 	case 0x10 ... 0x1F:	/* write of CR0 through CR15 */
1307 	case 0x20 ... 0x2F:	/* read of DR0 through DR15 */
1308 	case 0x30 ... 0x3F:	/* write of DR0 through DR15 */
1309 	case 0x43:		/* INT3 */
1310 	case 0x44:		/* INTO */
1311 	case 0x45:		/* BOUND */
1312 	case 0x65 ... 0x7C:	/* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1313 	case 0x80 ... 0x8D:	/* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1314 		return (1);
1315 	default:
1316 		return (0);
1317 	}
1318 }
1319 
1320 static int
1321 svm_vmexit(struct svm_softc *svm_sc, struct svm_vcpu *vcpu,
1322     struct vm_exit *vmexit)
1323 {
1324 	struct vmcb *vmcb;
1325 	struct vmcb_state *state;
1326 	struct vmcb_ctrl *ctrl;
1327 	struct svm_regctx *ctx;
1328 	uint64_t code, info1, info2, val;
1329 	uint32_t eax, ecx, edx;
1330 	int error __diagused, errcode_valid, handled, idtvec, reflect;
1331 	bool retu;
1332 
1333 	ctx = svm_get_guest_regctx(vcpu);
1334 	vmcb = svm_get_vmcb(vcpu);
1335 	state = &vmcb->state;
1336 	ctrl = &vmcb->ctrl;
1337 
1338 	handled = 0;
1339 	code = ctrl->exitcode;
1340 	info1 = ctrl->exitinfo1;
1341 	info2 = ctrl->exitinfo2;
1342 
1343 	vmexit->exitcode = VM_EXITCODE_BOGUS;
1344 	vmexit->rip = state->rip;
1345 	vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1346 
1347 	vmm_stat_incr(vcpu->vcpu, VMEXIT_COUNT, 1);
1348 
1349 	/*
1350 	 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1351 	 * in an inconsistent state and can trigger assertions that would
1352 	 * never happen otherwise.
1353 	 */
1354 	if (code == VMCB_EXIT_INVALID) {
1355 		vm_exit_svm(vmexit, code, info1, info2);
1356 		return (0);
1357 	}
1358 
1359 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1360 	    "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1361 
1362 	KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1363 	    ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1364 	    vmexit->inst_length, code, info1, info2));
1365 
1366 	svm_update_virqinfo(vcpu);
1367 	svm_save_intinfo(svm_sc, vcpu);
1368 
1369 	switch (code) {
1370 	case VMCB_EXIT_IRET:
1371 		/*
1372 		 * Restart execution at "iret" but with the intercept cleared.
1373 		 */
1374 		vmexit->inst_length = 0;
1375 		clear_nmi_blocking(vcpu);
1376 		handled = 1;
1377 		break;
1378 	case VMCB_EXIT_VINTR:	/* interrupt window exiting */
1379 		vmm_stat_incr(vcpu->vcpu, VMEXIT_VINTR, 1);
1380 		handled = 1;
1381 		break;
1382 	case VMCB_EXIT_INTR:	/* external interrupt */
1383 		vmm_stat_incr(vcpu->vcpu, VMEXIT_EXTINT, 1);
1384 		handled = 1;
1385 		break;
1386 	case VMCB_EXIT_NMI:	/* external NMI */
1387 		handled = 1;
1388 		break;
1389 	case 0x40 ... 0x5F:
1390 		vmm_stat_incr(vcpu->vcpu, VMEXIT_EXCEPTION, 1);
1391 		reflect = 1;
1392 		idtvec = code - 0x40;
1393 		switch (idtvec) {
1394 		case IDT_MC:
1395 			/*
1396 			 * Call the machine check handler by hand. Also don't
1397 			 * reflect the machine check back into the guest.
1398 			 */
1399 			reflect = 0;
1400 			SVM_CTR0(vcpu, "Vectoring to MCE handler");
1401 			__asm __volatile("int $18");
1402 			break;
1403 		case IDT_PF:
1404 			error = svm_setreg(vcpu, VM_REG_GUEST_CR2, info2);
1405 			KASSERT(error == 0, ("%s: error %d updating cr2",
1406 			    __func__, error));
1407 			/* fallthru */
1408 		case IDT_NP:
1409 		case IDT_SS:
1410 		case IDT_GP:
1411 		case IDT_AC:
1412 		case IDT_TS:
1413 			errcode_valid = 1;
1414 			break;
1415 
1416 		case IDT_DF:
1417 			errcode_valid = 1;
1418 			info1 = 0;
1419 			break;
1420 
1421 		case IDT_BP:
1422 		case IDT_OF:
1423 		case IDT_BR:
1424 			/*
1425 			 * The 'nrip' field is populated for INT3, INTO and
1426 			 * BOUND exceptions and this also implies that
1427 			 * 'inst_length' is non-zero.
1428 			 *
1429 			 * Reset 'inst_length' to zero so the guest %rip at
1430 			 * event injection is identical to what it was when
1431 			 * the exception originally happened.
1432 			 */
1433 			SVM_CTR2(vcpu, "Reset inst_length from %d "
1434 			    "to zero before injecting exception %d",
1435 			    vmexit->inst_length, idtvec);
1436 			vmexit->inst_length = 0;
1437 			/* fallthru */
1438 		default:
1439 			errcode_valid = 0;
1440 			info1 = 0;
1441 			break;
1442 		}
1443 		KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1444 		    "when reflecting exception %d into guest",
1445 		    vmexit->inst_length, idtvec));
1446 
1447 		if (reflect) {
1448 			/* Reflect the exception back into the guest */
1449 			SVM_CTR2(vcpu, "Reflecting exception "
1450 			    "%d/%#x into the guest", idtvec, (int)info1);
1451 			error = vm_inject_exception(vcpu->vcpu, idtvec,
1452 			    errcode_valid, info1, 0);
1453 			KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1454 			    __func__, error));
1455 		}
1456 		handled = 1;
1457 		break;
1458 	case VMCB_EXIT_MSR:	/* MSR access. */
1459 		eax = state->rax;
1460 		ecx = ctx->sctx_rcx;
1461 		edx = ctx->sctx_rdx;
1462 		retu = false;
1463 
1464 		if (info1) {
1465 			vmm_stat_incr(vcpu->vcpu, VMEXIT_WRMSR, 1);
1466 			val = (uint64_t)edx << 32 | eax;
1467 			SVM_CTR2(vcpu, "wrmsr %#x val %#lx", ecx, val);
1468 			if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1469 				vmexit->exitcode = VM_EXITCODE_WRMSR;
1470 				vmexit->u.msr.code = ecx;
1471 				vmexit->u.msr.wval = val;
1472 			} else if (!retu) {
1473 				handled = 1;
1474 			} else {
1475 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1476 				    ("emulate_wrmsr retu with bogus exitcode"));
1477 			}
1478 		} else {
1479 			SVM_CTR1(vcpu, "rdmsr %#x", ecx);
1480 			vmm_stat_incr(vcpu->vcpu, VMEXIT_RDMSR, 1);
1481 			if (emulate_rdmsr(vcpu, ecx, &retu)) {
1482 				vmexit->exitcode = VM_EXITCODE_RDMSR;
1483 				vmexit->u.msr.code = ecx;
1484 			} else if (!retu) {
1485 				handled = 1;
1486 			} else {
1487 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1488 				    ("emulate_rdmsr retu with bogus exitcode"));
1489 			}
1490 		}
1491 		break;
1492 	case VMCB_EXIT_IO:
1493 		handled = svm_handle_io(vcpu, vmexit);
1494 		vmm_stat_incr(vcpu->vcpu, VMEXIT_INOUT, 1);
1495 		break;
1496 	case VMCB_EXIT_CPUID:
1497 		vmm_stat_incr(vcpu->vcpu, VMEXIT_CPUID, 1);
1498 		handled = x86_emulate_cpuid(vcpu->vcpu,
1499 		    &state->rax, &ctx->sctx_rbx, &ctx->sctx_rcx,
1500 		    &ctx->sctx_rdx);
1501 		break;
1502 	case VMCB_EXIT_HLT:
1503 		vmm_stat_incr(vcpu->vcpu, VMEXIT_HLT, 1);
1504 		vmexit->exitcode = VM_EXITCODE_HLT;
1505 		vmexit->u.hlt.rflags = state->rflags;
1506 		break;
1507 	case VMCB_EXIT_PAUSE:
1508 		vmexit->exitcode = VM_EXITCODE_PAUSE;
1509 		vmm_stat_incr(vcpu->vcpu, VMEXIT_PAUSE, 1);
1510 		break;
1511 	case VMCB_EXIT_NPF:
1512 		/* EXITINFO2 contains the faulting guest physical address */
1513 		if (info1 & VMCB_NPF_INFO1_RSV) {
1514 			SVM_CTR2(vcpu, "nested page fault with "
1515 			    "reserved bits set: info1(%#lx) info2(%#lx)",
1516 			    info1, info2);
1517 		} else if (vm_mem_allocated(vcpu->vcpu, info2)) {
1518 			vmexit->exitcode = VM_EXITCODE_PAGING;
1519 			vmexit->u.paging.gpa = info2;
1520 			vmexit->u.paging.fault_type = npf_fault_type(info1);
1521 			vmm_stat_incr(vcpu->vcpu, VMEXIT_NESTED_FAULT, 1);
1522 			SVM_CTR3(vcpu, "nested page fault "
1523 			    "on gpa %#lx/%#lx at rip %#lx",
1524 			    info2, info1, state->rip);
1525 		} else if (svm_npf_emul_fault(info1)) {
1526 			svm_handle_inst_emul(vmcb, info2, vmexit);
1527 			vmm_stat_incr(vcpu->vcpu, VMEXIT_INST_EMUL, 1);
1528 			SVM_CTR3(vcpu, "inst_emul fault "
1529 			    "for gpa %#lx/%#lx at rip %#lx",
1530 			    info2, info1, state->rip);
1531 		}
1532 		break;
1533 	case VMCB_EXIT_MONITOR:
1534 		vmexit->exitcode = VM_EXITCODE_MONITOR;
1535 		break;
1536 	case VMCB_EXIT_MWAIT:
1537 		vmexit->exitcode = VM_EXITCODE_MWAIT;
1538 		break;
1539 	case VMCB_EXIT_SHUTDOWN:
1540 	case VMCB_EXIT_VMRUN:
1541 	case VMCB_EXIT_VMMCALL:
1542 	case VMCB_EXIT_VMLOAD:
1543 	case VMCB_EXIT_VMSAVE:
1544 	case VMCB_EXIT_STGI:
1545 	case VMCB_EXIT_CLGI:
1546 	case VMCB_EXIT_SKINIT:
1547 	case VMCB_EXIT_ICEBP:
1548 	case VMCB_EXIT_INVLPGA:
1549 		vm_inject_ud(vcpu->vcpu);
1550 		handled = 1;
1551 		break;
1552 	case VMCB_EXIT_INVD:
1553 	case VMCB_EXIT_WBINVD:
1554 		/* ignore exit */
1555 		handled = 1;
1556 		break;
1557 	default:
1558 		vmm_stat_incr(vcpu->vcpu, VMEXIT_UNKNOWN, 1);
1559 		break;
1560 	}
1561 
1562 	SVM_CTR4(vcpu, "%s %s vmexit at %#lx/%d",
1563 	    handled ? "handled" : "unhandled", exit_reason_to_str(code),
1564 	    vmexit->rip, vmexit->inst_length);
1565 
1566 	if (handled) {
1567 		vmexit->rip += vmexit->inst_length;
1568 		vmexit->inst_length = 0;
1569 		state->rip = vmexit->rip;
1570 	} else {
1571 		if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1572 			/*
1573 			 * If this VM exit was not claimed by anybody then
1574 			 * treat it as a generic SVM exit.
1575 			 */
1576 			vm_exit_svm(vmexit, code, info1, info2);
1577 		} else {
1578 			/*
1579 			 * The exitcode and collateral have been populated.
1580 			 * The VM exit will be processed further in userland.
1581 			 */
1582 		}
1583 	}
1584 	return (handled);
1585 }
1586 
1587 static void
1588 svm_inj_intinfo(struct svm_softc *svm_sc, struct svm_vcpu *vcpu)
1589 {
1590 	uint64_t intinfo;
1591 
1592 	if (!vm_entry_intinfo(vcpu->vcpu, &intinfo))
1593 		return;
1594 
1595 	KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1596 	    "valid: %#lx", __func__, intinfo));
1597 
1598 	svm_eventinject(vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1599 		VMCB_EXITINTINFO_VECTOR(intinfo),
1600 		VMCB_EXITINTINFO_EC(intinfo),
1601 		VMCB_EXITINTINFO_EC_VALID(intinfo));
1602 	vmm_stat_incr(vcpu->vcpu, VCPU_INTINFO_INJECTED, 1);
1603 	SVM_CTR1(vcpu, "Injected entry intinfo: %#lx", intinfo);
1604 }
1605 
1606 /*
1607  * Inject event to virtual cpu.
1608  */
1609 static void
1610 svm_inj_interrupts(struct svm_softc *sc, struct svm_vcpu *vcpu,
1611     struct vlapic *vlapic)
1612 {
1613 	struct vmcb_ctrl *ctrl;
1614 	struct vmcb_state *state;
1615 	uint8_t v_tpr;
1616 	int vector, need_intr_window;
1617 	int extint_pending;
1618 
1619 	state = svm_get_vmcb_state(vcpu);
1620 	ctrl  = svm_get_vmcb_ctrl(vcpu);
1621 
1622 	need_intr_window = 0;
1623 
1624 	if (vcpu->nextrip != state->rip) {
1625 		ctrl->intr_shadow = 0;
1626 		SVM_CTR2(vcpu, "Guest interrupt blocking "
1627 		    "cleared due to rip change: %#lx/%#lx",
1628 		    vcpu->nextrip, state->rip);
1629 	}
1630 
1631 	/*
1632 	 * Inject pending events or exceptions for this vcpu.
1633 	 *
1634 	 * An event might be pending because the previous #VMEXIT happened
1635 	 * during event delivery (i.e. ctrl->exitintinfo).
1636 	 *
1637 	 * An event might also be pending because an exception was injected
1638 	 * by the hypervisor (e.g. #PF during instruction emulation).
1639 	 */
1640 	svm_inj_intinfo(sc, vcpu);
1641 
1642 	/* NMI event has priority over interrupts. */
1643 	if (vm_nmi_pending(vcpu->vcpu)) {
1644 		if (nmi_blocked(vcpu)) {
1645 			/*
1646 			 * Can't inject another NMI if the guest has not
1647 			 * yet executed an "iret" after the last NMI.
1648 			 */
1649 			SVM_CTR0(vcpu, "Cannot inject NMI due "
1650 			    "to NMI-blocking");
1651 		} else if (ctrl->intr_shadow) {
1652 			/*
1653 			 * Can't inject an NMI if the vcpu is in an intr_shadow.
1654 			 */
1655 			SVM_CTR0(vcpu, "Cannot inject NMI due to "
1656 			    "interrupt shadow");
1657 			need_intr_window = 1;
1658 			goto done;
1659 		} else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1660 			/*
1661 			 * If there is already an exception/interrupt pending
1662 			 * then defer the NMI until after that.
1663 			 */
1664 			SVM_CTR1(vcpu, "Cannot inject NMI due to "
1665 			    "eventinj %#lx", ctrl->eventinj);
1666 
1667 			/*
1668 			 * Use self-IPI to trigger a VM-exit as soon as
1669 			 * possible after the event injection is completed.
1670 			 *
1671 			 * This works only if the external interrupt exiting
1672 			 * is at a lower priority than the event injection.
1673 			 *
1674 			 * Although not explicitly specified in APMv2 the
1675 			 * relative priorities were verified empirically.
1676 			 */
1677 			ipi_cpu(curcpu, IPI_AST);	/* XXX vmm_ipinum? */
1678 		} else {
1679 			vm_nmi_clear(vcpu->vcpu);
1680 
1681 			/* Inject NMI, vector number is not used */
1682 			svm_eventinject(vcpu, VMCB_EVENTINJ_TYPE_NMI,
1683 			    IDT_NMI, 0, false);
1684 
1685 			/* virtual NMI blocking is now in effect */
1686 			enable_nmi_blocking(vcpu);
1687 
1688 			SVM_CTR0(vcpu, "Injecting vNMI");
1689 		}
1690 	}
1691 
1692 	extint_pending = vm_extint_pending(vcpu->vcpu);
1693 	if (!extint_pending) {
1694 		if (!vlapic_pending_intr(vlapic, &vector))
1695 			goto done;
1696 		KASSERT(vector >= 16 && vector <= 255,
1697 		    ("invalid vector %d from local APIC", vector));
1698 	} else {
1699 		/* Ask the legacy pic for a vector to inject */
1700 		vatpic_pending_intr(sc->vm, &vector);
1701 		KASSERT(vector >= 0 && vector <= 255,
1702 		    ("invalid vector %d from INTR", vector));
1703 	}
1704 
1705 	/*
1706 	 * If the guest has disabled interrupts or is in an interrupt shadow
1707 	 * then we cannot inject the pending interrupt.
1708 	 */
1709 	if ((state->rflags & PSL_I) == 0) {
1710 		SVM_CTR2(vcpu, "Cannot inject vector %d due to "
1711 		    "rflags %#lx", vector, state->rflags);
1712 		need_intr_window = 1;
1713 		goto done;
1714 	}
1715 
1716 	if (ctrl->intr_shadow) {
1717 		SVM_CTR1(vcpu, "Cannot inject vector %d due to "
1718 		    "interrupt shadow", vector);
1719 		need_intr_window = 1;
1720 		goto done;
1721 	}
1722 
1723 	if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1724 		SVM_CTR2(vcpu, "Cannot inject vector %d due to "
1725 		    "eventinj %#lx", vector, ctrl->eventinj);
1726 		need_intr_window = 1;
1727 		goto done;
1728 	}
1729 
1730 	svm_eventinject(vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1731 
1732 	if (!extint_pending) {
1733 		vlapic_intr_accepted(vlapic, vector);
1734 	} else {
1735 		vm_extint_clear(vcpu->vcpu);
1736 		vatpic_intr_accepted(sc->vm, vector);
1737 	}
1738 
1739 	/*
1740 	 * Force a VM-exit as soon as the vcpu is ready to accept another
1741 	 * interrupt. This is done because the PIC might have another vector
1742 	 * that it wants to inject. Also, if the APIC has a pending interrupt
1743 	 * that was preempted by the ExtInt then it allows us to inject the
1744 	 * APIC vector as soon as possible.
1745 	 */
1746 	need_intr_window = 1;
1747 done:
1748 	/*
1749 	 * The guest can modify the TPR by writing to %CR8. In guest mode
1750 	 * the processor reflects this write to V_TPR without hypervisor
1751 	 * intervention.
1752 	 *
1753 	 * The guest can also modify the TPR by writing to it via the memory
1754 	 * mapped APIC page. In this case, the write will be emulated by the
1755 	 * hypervisor. For this reason V_TPR must be updated before every
1756 	 * VMRUN.
1757 	 */
1758 	v_tpr = vlapic_get_cr8(vlapic);
1759 	KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1760 	if (ctrl->v_tpr != v_tpr) {
1761 		SVM_CTR2(vcpu, "VMCB V_TPR changed from %#x to %#x",
1762 		    ctrl->v_tpr, v_tpr);
1763 		ctrl->v_tpr = v_tpr;
1764 		svm_set_dirty(vcpu, VMCB_CACHE_TPR);
1765 	}
1766 
1767 	if (need_intr_window) {
1768 		/*
1769 		 * We use V_IRQ in conjunction with the VINTR intercept to
1770 		 * trap into the hypervisor as soon as a virtual interrupt
1771 		 * can be delivered.
1772 		 *
1773 		 * Since injected events are not subject to intercept checks
1774 		 * we need to ensure that the V_IRQ is not actually going to
1775 		 * be delivered on VM entry. The KASSERT below enforces this.
1776 		 */
1777 		KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1778 		    (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1779 		    ("Bogus intr_window_exiting: eventinj (%#lx), "
1780 		    "intr_shadow (%u), rflags (%#lx)",
1781 		    ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1782 		enable_intr_window_exiting(vcpu);
1783 	} else {
1784 		disable_intr_window_exiting(vcpu);
1785 	}
1786 }
1787 
1788 static __inline void
1789 restore_host_tss(void)
1790 {
1791 	struct system_segment_descriptor *tss_sd;
1792 
1793 	/*
1794 	 * The TSS descriptor was in use prior to launching the guest so it
1795 	 * has been marked busy.
1796 	 *
1797 	 * 'ltr' requires the descriptor to be marked available so change the
1798 	 * type to "64-bit available TSS".
1799 	 */
1800 	tss_sd = PCPU_GET(tss);
1801 	tss_sd->sd_type = SDT_SYSTSS;
1802 	ltr(GSEL(GPROC0_SEL, SEL_KPL));
1803 }
1804 
1805 static void
1806 svm_pmap_activate(struct svm_vcpu *vcpu, pmap_t pmap)
1807 {
1808 	struct vmcb_ctrl *ctrl;
1809 	long eptgen;
1810 	int cpu;
1811 	bool alloc_asid;
1812 
1813 	cpu = curcpu;
1814 	CPU_SET_ATOMIC(cpu, &pmap->pm_active);
1815 	smr_enter(pmap->pm_eptsmr);
1816 
1817 	ctrl = svm_get_vmcb_ctrl(vcpu);
1818 
1819 	/*
1820 	 * The TLB entries associated with the vcpu's ASID are not valid
1821 	 * if either of the following conditions is true:
1822 	 *
1823 	 * 1. The vcpu's ASID generation is different than the host cpu's
1824 	 *    ASID generation. This happens when the vcpu migrates to a new
1825 	 *    host cpu. It can also happen when the number of vcpus executing
1826 	 *    on a host cpu is greater than the number of ASIDs available.
1827 	 *
1828 	 * 2. The pmap generation number is different than the value cached in
1829 	 *    the 'vcpustate'. This happens when the host invalidates pages
1830 	 *    belonging to the guest.
1831 	 *
1832 	 *	asidgen		eptgen	      Action
1833 	 *	mismatch	mismatch
1834 	 *	   0		   0		(a)
1835 	 *	   0		   1		(b1) or (b2)
1836 	 *	   1		   0		(c)
1837 	 *	   1		   1		(d)
1838 	 *
1839 	 * (a) There is no mismatch in eptgen or ASID generation and therefore
1840 	 *     no further action is needed.
1841 	 *
1842 	 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1843 	 *      retained and the TLB entries associated with this ASID
1844 	 *      are flushed by VMRUN.
1845 	 *
1846 	 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1847 	 *      allocated.
1848 	 *
1849 	 * (c) A new ASID is allocated.
1850 	 *
1851 	 * (d) A new ASID is allocated.
1852 	 */
1853 
1854 	alloc_asid = false;
1855 	eptgen = atomic_load_long(&pmap->pm_eptgen);
1856 	ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1857 
1858 	if (vcpu->asid.gen != asid[cpu].gen) {
1859 		alloc_asid = true;	/* (c) and (d) */
1860 	} else if (vcpu->eptgen != eptgen) {
1861 		if (flush_by_asid())
1862 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;	/* (b1) */
1863 		else
1864 			alloc_asid = true;			/* (b2) */
1865 	} else {
1866 		/*
1867 		 * This is the common case (a).
1868 		 */
1869 		KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1870 		KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1871 		    ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1872 	}
1873 
1874 	if (alloc_asid) {
1875 		if (++asid[cpu].num >= nasid) {
1876 			asid[cpu].num = 1;
1877 			if (++asid[cpu].gen == 0)
1878 				asid[cpu].gen = 1;
1879 			/*
1880 			 * If this cpu does not support "flush-by-asid"
1881 			 * then flush the entire TLB on a generation
1882 			 * bump. Subsequent ASID allocation in this
1883 			 * generation can be done without a TLB flush.
1884 			 */
1885 			if (!flush_by_asid())
1886 				ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1887 		}
1888 		vcpu->asid.gen = asid[cpu].gen;
1889 		vcpu->asid.num = asid[cpu].num;
1890 
1891 		ctrl->asid = vcpu->asid.num;
1892 		svm_set_dirty(vcpu, VMCB_CACHE_ASID);
1893 		/*
1894 		 * If this cpu supports "flush-by-asid" then the TLB
1895 		 * was not flushed after the generation bump. The TLB
1896 		 * is flushed selectively after every new ASID allocation.
1897 		 */
1898 		if (flush_by_asid())
1899 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1900 	}
1901 	vcpu->eptgen = eptgen;
1902 
1903 	KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1904 	KASSERT(ctrl->asid == vcpu->asid.num,
1905 	    ("ASID mismatch: %u/%u", ctrl->asid, vcpu->asid.num));
1906 }
1907 
1908 static void
1909 svm_pmap_deactivate(pmap_t pmap)
1910 {
1911 	smr_exit(pmap->pm_eptsmr);
1912 	CPU_CLR_ATOMIC(curcpu, &pmap->pm_active);
1913 }
1914 
1915 static __inline void
1916 disable_gintr(void)
1917 {
1918 
1919 	__asm __volatile("clgi");
1920 }
1921 
1922 static __inline void
1923 enable_gintr(void)
1924 {
1925 
1926         __asm __volatile("stgi");
1927 }
1928 
1929 static __inline void
1930 svm_dr_enter_guest(struct svm_regctx *gctx)
1931 {
1932 
1933 	/* Save host control debug registers. */
1934 	gctx->host_dr7 = rdr7();
1935 	gctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR);
1936 
1937 	/*
1938 	 * Disable debugging in DR7 and DEBUGCTL to avoid triggering
1939 	 * exceptions in the host based on the guest DRx values.  The
1940 	 * guest DR6, DR7, and DEBUGCTL are saved/restored in the
1941 	 * VMCB.
1942 	 */
1943 	load_dr7(0);
1944 	wrmsr(MSR_DEBUGCTLMSR, 0);
1945 
1946 	/* Save host debug registers. */
1947 	gctx->host_dr0 = rdr0();
1948 	gctx->host_dr1 = rdr1();
1949 	gctx->host_dr2 = rdr2();
1950 	gctx->host_dr3 = rdr3();
1951 	gctx->host_dr6 = rdr6();
1952 
1953 	/* Restore guest debug registers. */
1954 	load_dr0(gctx->sctx_dr0);
1955 	load_dr1(gctx->sctx_dr1);
1956 	load_dr2(gctx->sctx_dr2);
1957 	load_dr3(gctx->sctx_dr3);
1958 }
1959 
1960 static __inline void
1961 svm_dr_leave_guest(struct svm_regctx *gctx)
1962 {
1963 
1964 	/* Save guest debug registers. */
1965 	gctx->sctx_dr0 = rdr0();
1966 	gctx->sctx_dr1 = rdr1();
1967 	gctx->sctx_dr2 = rdr2();
1968 	gctx->sctx_dr3 = rdr3();
1969 
1970 	/*
1971 	 * Restore host debug registers.  Restore DR7 and DEBUGCTL
1972 	 * last.
1973 	 */
1974 	load_dr0(gctx->host_dr0);
1975 	load_dr1(gctx->host_dr1);
1976 	load_dr2(gctx->host_dr2);
1977 	load_dr3(gctx->host_dr3);
1978 	load_dr6(gctx->host_dr6);
1979 	wrmsr(MSR_DEBUGCTLMSR, gctx->host_debugctl);
1980 	load_dr7(gctx->host_dr7);
1981 }
1982 
1983 /*
1984  * Start vcpu with specified RIP.
1985  */
1986 static int
1987 svm_run(void *vcpui, register_t rip, pmap_t pmap, struct vm_eventinfo *evinfo)
1988 {
1989 	struct svm_regctx *gctx;
1990 	struct svm_softc *svm_sc;
1991 	struct svm_vcpu *vcpu;
1992 	struct vmcb_state *state;
1993 	struct vmcb_ctrl *ctrl;
1994 	struct vm_exit *vmexit;
1995 	struct vlapic *vlapic;
1996 	uint64_t vmcb_pa;
1997 	int handled;
1998 	uint16_t ldt_sel;
1999 
2000 	vcpu = vcpui;
2001 	svm_sc = vcpu->sc;
2002 	state = svm_get_vmcb_state(vcpu);
2003 	ctrl = svm_get_vmcb_ctrl(vcpu);
2004 	vmexit = vm_exitinfo(vcpu->vcpu);
2005 	vlapic = vm_lapic(vcpu->vcpu);
2006 
2007 	gctx = svm_get_guest_regctx(vcpu);
2008 	vmcb_pa = vcpu->vmcb_pa;
2009 
2010 	if (vcpu->lastcpu != curcpu) {
2011 		/*
2012 		 * Force new ASID allocation by invalidating the generation.
2013 		 */
2014 		vcpu->asid.gen = 0;
2015 
2016 		/*
2017 		 * Invalidate the VMCB state cache by marking all fields dirty.
2018 		 */
2019 		svm_set_dirty(vcpu, 0xffffffff);
2020 
2021 		/*
2022 		 * XXX
2023 		 * Setting 'vcpu->lastcpu' here is bit premature because
2024 		 * we may return from this function without actually executing
2025 		 * the VMRUN  instruction. This could happen if a rendezvous
2026 		 * or an AST is pending on the first time through the loop.
2027 		 *
2028 		 * This works for now but any new side-effects of vcpu
2029 		 * migration should take this case into account.
2030 		 */
2031 		vcpu->lastcpu = curcpu;
2032 		vmm_stat_incr(vcpu->vcpu, VCPU_MIGRATIONS, 1);
2033 	}
2034 
2035 	svm_msr_guest_enter(vcpu);
2036 
2037 	/* Update Guest RIP */
2038 	state->rip = rip;
2039 
2040 	do {
2041 		/*
2042 		 * Disable global interrupts to guarantee atomicity during
2043 		 * loading of guest state. This includes not only the state
2044 		 * loaded by the "vmrun" instruction but also software state
2045 		 * maintained by the hypervisor: suspended and rendezvous
2046 		 * state, NPT generation number, vlapic interrupts etc.
2047 		 */
2048 		disable_gintr();
2049 
2050 		if (vcpu_suspended(evinfo)) {
2051 			enable_gintr();
2052 			vm_exit_suspended(vcpu->vcpu, state->rip);
2053 			break;
2054 		}
2055 
2056 		if (vcpu_rendezvous_pending(vcpu->vcpu, evinfo)) {
2057 			enable_gintr();
2058 			vm_exit_rendezvous(vcpu->vcpu, state->rip);
2059 			break;
2060 		}
2061 
2062 		if (vcpu_reqidle(evinfo)) {
2063 			enable_gintr();
2064 			vm_exit_reqidle(vcpu->vcpu, state->rip);
2065 			break;
2066 		}
2067 
2068 		/* We are asked to give the cpu by scheduler. */
2069 		if (vcpu_should_yield(vcpu->vcpu)) {
2070 			enable_gintr();
2071 			vm_exit_astpending(vcpu->vcpu, state->rip);
2072 			break;
2073 		}
2074 
2075 		if (vcpu_debugged(vcpu->vcpu)) {
2076 			enable_gintr();
2077 			vm_exit_debug(vcpu->vcpu, state->rip);
2078 			break;
2079 		}
2080 
2081 		/*
2082 		 * #VMEXIT resumes the host with the guest LDTR, so
2083 		 * save the current LDT selector so it can be restored
2084 		 * after an exit.  The userspace hypervisor probably
2085 		 * doesn't use a LDT, but save and restore it to be
2086 		 * safe.
2087 		 */
2088 		ldt_sel = sldt();
2089 
2090 		svm_inj_interrupts(svm_sc, vcpu, vlapic);
2091 
2092 		/*
2093 		 * Check the pmap generation and the ASID generation to
2094 		 * ensure that the vcpu does not use stale TLB mappings.
2095 		 */
2096 		svm_pmap_activate(vcpu, pmap);
2097 
2098 		ctrl->vmcb_clean = vmcb_clean & ~vcpu->dirty;
2099 		vcpu->dirty = 0;
2100 		SVM_CTR1(vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
2101 
2102 		/* Launch Virtual Machine. */
2103 		SVM_CTR1(vcpu, "Resume execution at %#lx", state->rip);
2104 		svm_dr_enter_guest(gctx);
2105 		svm_launch(vmcb_pa, gctx, get_pcpu());
2106 		svm_dr_leave_guest(gctx);
2107 
2108 		svm_pmap_deactivate(pmap);
2109 
2110 		/*
2111 		 * The host GDTR and IDTR is saved by VMRUN and restored
2112 		 * automatically on #VMEXIT. However, the host TSS needs
2113 		 * to be restored explicitly.
2114 		 */
2115 		restore_host_tss();
2116 
2117 		/* Restore host LDTR. */
2118 		lldt(ldt_sel);
2119 
2120 		/* #VMEXIT disables interrupts so re-enable them here. */
2121 		enable_gintr();
2122 
2123 		/* Update 'nextrip' */
2124 		vcpu->nextrip = state->rip;
2125 
2126 		/* Handle #VMEXIT and if required return to user space. */
2127 		handled = svm_vmexit(svm_sc, vcpu, vmexit);
2128 	} while (handled);
2129 
2130 	svm_msr_guest_exit(vcpu);
2131 
2132 	return (0);
2133 }
2134 
2135 static void
2136 svm_vcpu_cleanup(void *vcpui)
2137 {
2138 	struct svm_vcpu *vcpu = vcpui;
2139 
2140 	free(vcpu->vmcb, M_SVM);
2141 	free(vcpu, M_SVM);
2142 }
2143 
2144 static void
2145 svm_cleanup(void *vmi)
2146 {
2147 	struct svm_softc *sc = vmi;
2148 
2149 	contigfree(sc->iopm_bitmap, SVM_IO_BITMAP_SIZE, M_SVM);
2150 	contigfree(sc->msr_bitmap, SVM_MSR_BITMAP_SIZE, M_SVM);
2151 	free(sc, M_SVM);
2152 }
2153 
2154 static register_t *
2155 swctx_regptr(struct svm_regctx *regctx, int reg)
2156 {
2157 
2158 	switch (reg) {
2159 	case VM_REG_GUEST_RBX:
2160 		return (&regctx->sctx_rbx);
2161 	case VM_REG_GUEST_RCX:
2162 		return (&regctx->sctx_rcx);
2163 	case VM_REG_GUEST_RDX:
2164 		return (&regctx->sctx_rdx);
2165 	case VM_REG_GUEST_RDI:
2166 		return (&regctx->sctx_rdi);
2167 	case VM_REG_GUEST_RSI:
2168 		return (&regctx->sctx_rsi);
2169 	case VM_REG_GUEST_RBP:
2170 		return (&regctx->sctx_rbp);
2171 	case VM_REG_GUEST_R8:
2172 		return (&regctx->sctx_r8);
2173 	case VM_REG_GUEST_R9:
2174 		return (&regctx->sctx_r9);
2175 	case VM_REG_GUEST_R10:
2176 		return (&regctx->sctx_r10);
2177 	case VM_REG_GUEST_R11:
2178 		return (&regctx->sctx_r11);
2179 	case VM_REG_GUEST_R12:
2180 		return (&regctx->sctx_r12);
2181 	case VM_REG_GUEST_R13:
2182 		return (&regctx->sctx_r13);
2183 	case VM_REG_GUEST_R14:
2184 		return (&regctx->sctx_r14);
2185 	case VM_REG_GUEST_R15:
2186 		return (&regctx->sctx_r15);
2187 	case VM_REG_GUEST_DR0:
2188 		return (&regctx->sctx_dr0);
2189 	case VM_REG_GUEST_DR1:
2190 		return (&regctx->sctx_dr1);
2191 	case VM_REG_GUEST_DR2:
2192 		return (&regctx->sctx_dr2);
2193 	case VM_REG_GUEST_DR3:
2194 		return (&regctx->sctx_dr3);
2195 	default:
2196 		return (NULL);
2197 	}
2198 }
2199 
2200 static int
2201 svm_getreg(void *vcpui, int ident, uint64_t *val)
2202 {
2203 	struct svm_vcpu *vcpu;
2204 	register_t *reg;
2205 
2206 	vcpu = vcpui;
2207 
2208 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2209 		return (svm_get_intr_shadow(vcpu, val));
2210 	}
2211 
2212 	if (vmcb_read(vcpu, ident, val) == 0) {
2213 		return (0);
2214 	}
2215 
2216 	reg = swctx_regptr(svm_get_guest_regctx(vcpu), ident);
2217 
2218 	if (reg != NULL) {
2219 		*val = *reg;
2220 		return (0);
2221 	}
2222 
2223 	SVM_CTR1(vcpu, "svm_getreg: unknown register %#x", ident);
2224 	return (EINVAL);
2225 }
2226 
2227 static int
2228 svm_setreg(void *vcpui, int ident, uint64_t val)
2229 {
2230 	struct svm_vcpu *vcpu;
2231 	register_t *reg;
2232 
2233 	vcpu = vcpui;
2234 
2235 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2236 		return (svm_modify_intr_shadow(vcpu, val));
2237 	}
2238 
2239 	/* Do not permit user write access to VMCB fields by offset. */
2240 	if (!VMCB_ACCESS_OK(ident)) {
2241 		if (vmcb_write(vcpu, ident, val) == 0) {
2242 			return (0);
2243 		}
2244 	}
2245 
2246 	reg = swctx_regptr(svm_get_guest_regctx(vcpu), ident);
2247 
2248 	if (reg != NULL) {
2249 		*reg = val;
2250 		return (0);
2251 	}
2252 
2253 	if (ident == VM_REG_GUEST_ENTRY_INST_LENGTH) {
2254 		/* Ignore. */
2255 		return (0);
2256 	}
2257 
2258 	/*
2259 	 * XXX deal with CR3 and invalidate TLB entries tagged with the
2260 	 * vcpu's ASID. This needs to be treated differently depending on
2261 	 * whether 'running' is true/false.
2262 	 */
2263 
2264 	SVM_CTR1(vcpu, "svm_setreg: unknown register %#x", ident);
2265 	return (EINVAL);
2266 }
2267 
2268 static int
2269 svm_getdesc(void *vcpui, int reg, struct seg_desc *desc)
2270 {
2271 	return (vmcb_getdesc(vcpui, reg, desc));
2272 }
2273 
2274 static int
2275 svm_setdesc(void *vcpui, int reg, struct seg_desc *desc)
2276 {
2277 	return (vmcb_setdesc(vcpui, reg, desc));
2278 }
2279 
2280 #ifdef BHYVE_SNAPSHOT
2281 static int
2282 svm_snapshot_reg(void *vcpui, int ident, struct vm_snapshot_meta *meta)
2283 {
2284 	int ret;
2285 	uint64_t val;
2286 
2287 	if (meta->op == VM_SNAPSHOT_SAVE) {
2288 		ret = svm_getreg(vcpui, ident, &val);
2289 		if (ret != 0)
2290 			goto done;
2291 
2292 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2293 	} else if (meta->op == VM_SNAPSHOT_RESTORE) {
2294 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2295 
2296 		ret = svm_setreg(vcpui, ident, val);
2297 		if (ret != 0)
2298 			goto done;
2299 	} else {
2300 		ret = EINVAL;
2301 		goto done;
2302 	}
2303 
2304 done:
2305 	return (ret);
2306 }
2307 #endif
2308 
2309 static int
2310 svm_setcap(void *vcpui, int type, int val)
2311 {
2312 	struct svm_vcpu *vcpu;
2313 	struct vlapic *vlapic;
2314 	int error;
2315 
2316 	vcpu = vcpui;
2317 	error = 0;
2318 
2319 	switch (type) {
2320 	case VM_CAP_HALT_EXIT:
2321 		svm_set_intercept(vcpu, VMCB_CTRL1_INTCPT,
2322 		    VMCB_INTCPT_HLT, val);
2323 		break;
2324 	case VM_CAP_PAUSE_EXIT:
2325 		svm_set_intercept(vcpu, VMCB_CTRL1_INTCPT,
2326 		    VMCB_INTCPT_PAUSE, val);
2327 		break;
2328 	case VM_CAP_UNRESTRICTED_GUEST:
2329 		/* Unrestricted guest execution cannot be disabled in SVM */
2330 		if (val == 0)
2331 			error = EINVAL;
2332 		break;
2333 	case VM_CAP_IPI_EXIT:
2334 		vlapic = vm_lapic(vcpu->vcpu);
2335 		vlapic->ipi_exit = val;
2336 		break;
2337 	default:
2338 		error = ENOENT;
2339 		break;
2340 	}
2341 	return (error);
2342 }
2343 
2344 static int
2345 svm_getcap(void *vcpui, int type, int *retval)
2346 {
2347 	struct svm_vcpu *vcpu;
2348 	struct vlapic *vlapic;
2349 	int error;
2350 
2351 	vcpu = vcpui;
2352 	error = 0;
2353 
2354 	switch (type) {
2355 	case VM_CAP_HALT_EXIT:
2356 		*retval = svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT,
2357 		    VMCB_INTCPT_HLT);
2358 		break;
2359 	case VM_CAP_PAUSE_EXIT:
2360 		*retval = svm_get_intercept(vcpu, VMCB_CTRL1_INTCPT,
2361 		    VMCB_INTCPT_PAUSE);
2362 		break;
2363 	case VM_CAP_UNRESTRICTED_GUEST:
2364 		*retval = 1;	/* unrestricted guest is always enabled */
2365 		break;
2366 	case VM_CAP_IPI_EXIT:
2367 		vlapic = vm_lapic(vcpu->vcpu);
2368 		*retval = vlapic->ipi_exit;
2369 		break;
2370 	default:
2371 		error = ENOENT;
2372 		break;
2373 	}
2374 	return (error);
2375 }
2376 
2377 static struct vmspace *
2378 svm_vmspace_alloc(vm_offset_t min, vm_offset_t max)
2379 {
2380 	return (svm_npt_alloc(min, max));
2381 }
2382 
2383 static void
2384 svm_vmspace_free(struct vmspace *vmspace)
2385 {
2386 	svm_npt_free(vmspace);
2387 }
2388 
2389 static struct vlapic *
2390 svm_vlapic_init(void *vcpui)
2391 {
2392 	struct svm_vcpu *vcpu;
2393 	struct vlapic *vlapic;
2394 
2395 	vcpu = vcpui;
2396 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2397 	vlapic->vm = vcpu->sc->vm;
2398 	vlapic->vcpu = vcpu->vcpu;
2399 	vlapic->vcpuid = vcpu->vcpuid;
2400 	vlapic->apic_page = malloc_aligned(PAGE_SIZE, PAGE_SIZE, M_SVM_VLAPIC,
2401 	    M_WAITOK | M_ZERO);
2402 
2403 	vlapic_init(vlapic);
2404 
2405 	return (vlapic);
2406 }
2407 
2408 static void
2409 svm_vlapic_cleanup(struct vlapic *vlapic)
2410 {
2411 
2412         vlapic_cleanup(vlapic);
2413 	free(vlapic->apic_page, M_SVM_VLAPIC);
2414         free(vlapic, M_SVM_VLAPIC);
2415 }
2416 
2417 #ifdef BHYVE_SNAPSHOT
2418 static int
2419 svm_vcpu_snapshot(void *vcpui, struct vm_snapshot_meta *meta)
2420 {
2421 	struct svm_vcpu *vcpu;
2422 	int err, running, hostcpu;
2423 
2424 	vcpu = vcpui;
2425 	err = 0;
2426 
2427 	running = vcpu_is_running(vcpu->vcpu, &hostcpu);
2428 	if (running && hostcpu != curcpu) {
2429 		printf("%s: %s%d is running", __func__, vm_name(vcpu->sc->vm),
2430 		    vcpu->vcpuid);
2431 		return (EINVAL);
2432 	}
2433 
2434 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR0, meta);
2435 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR2, meta);
2436 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR3, meta);
2437 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CR4, meta);
2438 
2439 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_DR6, meta);
2440 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_DR7, meta);
2441 
2442 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RAX, meta);
2443 
2444 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RSP, meta);
2445 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RIP, meta);
2446 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_RFLAGS, meta);
2447 
2448 	/* Guest segments */
2449 	/* ES */
2450 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_ES, meta);
2451 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_ES, meta);
2452 
2453 	/* CS */
2454 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_CS, meta);
2455 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_CS, meta);
2456 
2457 	/* SS */
2458 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_SS, meta);
2459 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_SS, meta);
2460 
2461 	/* DS */
2462 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_DS, meta);
2463 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_DS, meta);
2464 
2465 	/* FS */
2466 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_FS, meta);
2467 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_FS, meta);
2468 
2469 	/* GS */
2470 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_GS, meta);
2471 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_GS, meta);
2472 
2473 	/* TR */
2474 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_TR, meta);
2475 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_TR, meta);
2476 
2477 	/* LDTR */
2478 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_LDTR, meta);
2479 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_LDTR, meta);
2480 
2481 	/* EFER */
2482 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_EFER, meta);
2483 
2484 	/* IDTR and GDTR */
2485 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_IDTR, meta);
2486 	err += vmcb_snapshot_desc(vcpu, VM_REG_GUEST_GDTR, meta);
2487 
2488 	/* Specific AMD registers */
2489 	err += svm_snapshot_reg(vcpu, VM_REG_GUEST_INTR_SHADOW, meta);
2490 
2491 	err += vmcb_snapshot_any(vcpu,
2492 				VMCB_ACCESS(VMCB_OFF_CR_INTERCEPT, 4), meta);
2493 	err += vmcb_snapshot_any(vcpu,
2494 				VMCB_ACCESS(VMCB_OFF_DR_INTERCEPT, 4), meta);
2495 	err += vmcb_snapshot_any(vcpu,
2496 				VMCB_ACCESS(VMCB_OFF_EXC_INTERCEPT, 4), meta);
2497 	err += vmcb_snapshot_any(vcpu,
2498 				VMCB_ACCESS(VMCB_OFF_INST1_INTERCEPT, 4), meta);
2499 	err += vmcb_snapshot_any(vcpu,
2500 				VMCB_ACCESS(VMCB_OFF_INST2_INTERCEPT, 4), meta);
2501 
2502 	err += vmcb_snapshot_any(vcpu,
2503 				VMCB_ACCESS(VMCB_OFF_PAUSE_FILTHRESH, 2), meta);
2504 	err += vmcb_snapshot_any(vcpu,
2505 				VMCB_ACCESS(VMCB_OFF_PAUSE_FILCNT, 2), meta);
2506 
2507 	err += vmcb_snapshot_any(vcpu,
2508 				VMCB_ACCESS(VMCB_OFF_ASID, 4), meta);
2509 
2510 	err += vmcb_snapshot_any(vcpu,
2511 				VMCB_ACCESS(VMCB_OFF_TLB_CTRL, 4), meta);
2512 
2513 	err += vmcb_snapshot_any(vcpu,
2514 				VMCB_ACCESS(VMCB_OFF_VIRQ, 8), meta);
2515 
2516 	err += vmcb_snapshot_any(vcpu,
2517 				VMCB_ACCESS(VMCB_OFF_EXIT_REASON, 8), meta);
2518 	err += vmcb_snapshot_any(vcpu,
2519 				VMCB_ACCESS(VMCB_OFF_EXITINFO1, 8), meta);
2520 	err += vmcb_snapshot_any(vcpu,
2521 				VMCB_ACCESS(VMCB_OFF_EXITINFO2, 8), meta);
2522 	err += vmcb_snapshot_any(vcpu,
2523 				VMCB_ACCESS(VMCB_OFF_EXITINTINFO, 8), meta);
2524 
2525 	err += vmcb_snapshot_any(vcpu,
2526 				VMCB_ACCESS(VMCB_OFF_NP_ENABLE, 1), meta);
2527 
2528 	err += vmcb_snapshot_any(vcpu,
2529 				VMCB_ACCESS(VMCB_OFF_AVIC_BAR, 8), meta);
2530 	err += vmcb_snapshot_any(vcpu,
2531 				VMCB_ACCESS(VMCB_OFF_AVIC_PAGE, 8), meta);
2532 	err += vmcb_snapshot_any(vcpu,
2533 				VMCB_ACCESS(VMCB_OFF_AVIC_LT, 8), meta);
2534 	err += vmcb_snapshot_any(vcpu,
2535 				VMCB_ACCESS(VMCB_OFF_AVIC_PT, 8), meta);
2536 
2537 	err += vmcb_snapshot_any(vcpu,
2538 				VMCB_ACCESS(VMCB_OFF_CPL, 1), meta);
2539 
2540 	err += vmcb_snapshot_any(vcpu,
2541 				VMCB_ACCESS(VMCB_OFF_STAR, 8), meta);
2542 	err += vmcb_snapshot_any(vcpu,
2543 				VMCB_ACCESS(VMCB_OFF_LSTAR, 8), meta);
2544 	err += vmcb_snapshot_any(vcpu,
2545 				VMCB_ACCESS(VMCB_OFF_CSTAR, 8), meta);
2546 
2547 	err += vmcb_snapshot_any(vcpu,
2548 				VMCB_ACCESS(VMCB_OFF_SFMASK, 8), meta);
2549 
2550 	err += vmcb_snapshot_any(vcpu,
2551 				VMCB_ACCESS(VMCB_OFF_KERNELGBASE, 8), meta);
2552 
2553 	err += vmcb_snapshot_any(vcpu,
2554 				VMCB_ACCESS(VMCB_OFF_SYSENTER_CS, 8), meta);
2555 	err += vmcb_snapshot_any(vcpu,
2556 				VMCB_ACCESS(VMCB_OFF_SYSENTER_ESP, 8), meta);
2557 	err += vmcb_snapshot_any(vcpu,
2558 				VMCB_ACCESS(VMCB_OFF_SYSENTER_EIP, 8), meta);
2559 
2560 	err += vmcb_snapshot_any(vcpu,
2561 				VMCB_ACCESS(VMCB_OFF_GUEST_PAT, 8), meta);
2562 
2563 	err += vmcb_snapshot_any(vcpu,
2564 				VMCB_ACCESS(VMCB_OFF_DBGCTL, 8), meta);
2565 	err += vmcb_snapshot_any(vcpu,
2566 				VMCB_ACCESS(VMCB_OFF_BR_FROM, 8), meta);
2567 	err += vmcb_snapshot_any(vcpu,
2568 				VMCB_ACCESS(VMCB_OFF_BR_TO, 8), meta);
2569 	err += vmcb_snapshot_any(vcpu,
2570 				VMCB_ACCESS(VMCB_OFF_INT_FROM, 8), meta);
2571 	err += vmcb_snapshot_any(vcpu,
2572 				VMCB_ACCESS(VMCB_OFF_INT_TO, 8), meta);
2573 	if (err != 0)
2574 		goto done;
2575 
2576 	/* Snapshot swctx for virtual cpu */
2577 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbp, meta, err, done);
2578 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbx, meta, err, done);
2579 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rcx, meta, err, done);
2580 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdx, meta, err, done);
2581 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdi, meta, err, done);
2582 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rsi, meta, err, done);
2583 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r8, meta, err, done);
2584 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r9, meta, err, done);
2585 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r10, meta, err, done);
2586 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r11, meta, err, done);
2587 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r12, meta, err, done);
2588 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r13, meta, err, done);
2589 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r14, meta, err, done);
2590 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r15, meta, err, done);
2591 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr0, meta, err, done);
2592 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr1, meta, err, done);
2593 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr2, meta, err, done);
2594 	SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr3, meta, err, done);
2595 
2596 	/* Restore other svm_vcpu struct fields */
2597 
2598 	/* Restore NEXTRIP field */
2599 	SNAPSHOT_VAR_OR_LEAVE(vcpu->nextrip, meta, err, done);
2600 
2601 	/* Restore lastcpu field */
2602 	SNAPSHOT_VAR_OR_LEAVE(vcpu->lastcpu, meta, err, done);
2603 	SNAPSHOT_VAR_OR_LEAVE(vcpu->dirty, meta, err, done);
2604 
2605 	/* Restore EPTGEN field - EPT is Extended Page Table */
2606 	SNAPSHOT_VAR_OR_LEAVE(vcpu->eptgen, meta, err, done);
2607 
2608 	SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.gen, meta, err, done);
2609 	SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.num, meta, err, done);
2610 
2611 	/* Set all caches dirty */
2612 	if (meta->op == VM_SNAPSHOT_RESTORE)
2613 		svm_set_dirty(vcpu, 0xffffffff);
2614 
2615 done:
2616 	return (err);
2617 }
2618 
2619 static int
2620 svm_restore_tsc(void *vcpui, uint64_t offset)
2621 {
2622 	struct svm_vcpu *vcpu = vcpui;
2623 
2624 	svm_set_tsc_offset(vcpu, offset);
2625 
2626 	return (0);
2627 }
2628 #endif
2629 
2630 const struct vmm_ops vmm_ops_amd = {
2631 	.modinit	= svm_modinit,
2632 	.modcleanup	= svm_modcleanup,
2633 	.modresume	= svm_modresume,
2634 	.init		= svm_init,
2635 	.run		= svm_run,
2636 	.cleanup	= svm_cleanup,
2637 	.vcpu_init	= svm_vcpu_init,
2638 	.vcpu_cleanup	= svm_vcpu_cleanup,
2639 	.getreg		= svm_getreg,
2640 	.setreg		= svm_setreg,
2641 	.getdesc	= svm_getdesc,
2642 	.setdesc	= svm_setdesc,
2643 	.getcap		= svm_getcap,
2644 	.setcap		= svm_setcap,
2645 	.vmspace_alloc	= svm_vmspace_alloc,
2646 	.vmspace_free	= svm_vmspace_free,
2647 	.vlapic_init	= svm_vlapic_init,
2648 	.vlapic_cleanup	= svm_vlapic_cleanup,
2649 #ifdef BHYVE_SNAPSHOT
2650 	.vcpu_snapshot	= svm_vcpu_snapshot,
2651 	.restore_tsc	= svm_restore_tsc,
2652 #endif
2653 };
2654