xref: /freebsd/sys/amd64/vmm/amd/svm.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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 *arg, int vcpu, int reg, struct seg_desc *desc);
136 static int svm_setreg(void *arg, int vcpu, 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 int
286 svm_set_tsc_offset(struct svm_softc *sc, int vcpu, uint64_t offset)
287 {
288 	int error;
289 	struct vmcb_ctrl *ctrl;
290 
291 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
292 	ctrl->tsc_offset = offset;
293 
294 	svm_set_dirty(sc, vcpu, VMCB_CACHE_I);
295 	VCPU_CTR1(sc->vm, vcpu, "tsc offset changed to %#lx", offset);
296 
297 	error = vm_set_tsc_offset(sc->vm, vcpu, offset);
298 
299 	return (error);
300 }
301 #endif
302 
303 /* Pentium compatible MSRs */
304 #define MSR_PENTIUM_START 	0
305 #define MSR_PENTIUM_END 	0x1FFF
306 /* AMD 6th generation and Intel compatible MSRs */
307 #define MSR_AMD6TH_START 	0xC0000000UL
308 #define MSR_AMD6TH_END 		0xC0001FFFUL
309 /* AMD 7th and 8th generation compatible MSRs */
310 #define MSR_AMD7TH_START 	0xC0010000UL
311 #define MSR_AMD7TH_END 		0xC0011FFFUL
312 
313 /*
314  * Get the index and bit position for a MSR in permission bitmap.
315  * Two bits are used for each MSR: lower bit for read and higher bit for write.
316  */
317 static int
318 svm_msr_index(uint64_t msr, int *index, int *bit)
319 {
320 	uint32_t base, off;
321 
322 	*index = -1;
323 	*bit = (msr % 4) * 2;
324 	base = 0;
325 
326 	if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) {
327 		*index = msr / 4;
328 		return (0);
329 	}
330 
331 	base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1);
332 	if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) {
333 		off = (msr - MSR_AMD6TH_START);
334 		*index = (off + base) / 4;
335 		return (0);
336 	}
337 
338 	base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1);
339 	if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) {
340 		off = (msr - MSR_AMD7TH_START);
341 		*index = (off + base) / 4;
342 		return (0);
343 	}
344 
345 	return (EINVAL);
346 }
347 
348 /*
349  * Allow vcpu to read or write the 'msr' without trapping into the hypervisor.
350  */
351 static void
352 svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write)
353 {
354 	int index, bit, error __diagused;
355 
356 	error = svm_msr_index(msr, &index, &bit);
357 	KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr));
358 	KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE,
359 	    ("%s: invalid index %d for msr %#lx", __func__, index, msr));
360 	KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d "
361 	    "msr %#lx", __func__, bit, msr));
362 
363 	if (read)
364 		perm_bitmap[index] &= ~(1UL << bit);
365 
366 	if (write)
367 		perm_bitmap[index] &= ~(2UL << bit);
368 }
369 
370 static void
371 svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr)
372 {
373 
374 	svm_msr_perm(perm_bitmap, msr, true, true);
375 }
376 
377 static void
378 svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr)
379 {
380 
381 	svm_msr_perm(perm_bitmap, msr, true, false);
382 }
383 
384 static __inline int
385 svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask)
386 {
387 	struct vmcb_ctrl *ctrl;
388 
389 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
390 
391 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
392 	return (ctrl->intercept[idx] & bitmask ? 1 : 0);
393 }
394 
395 static __inline void
396 svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask,
397     int enabled)
398 {
399 	struct vmcb_ctrl *ctrl;
400 	uint32_t oldval;
401 
402 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
403 
404 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
405 	oldval = ctrl->intercept[idx];
406 
407 	if (enabled)
408 		ctrl->intercept[idx] |= bitmask;
409 	else
410 		ctrl->intercept[idx] &= ~bitmask;
411 
412 	if (ctrl->intercept[idx] != oldval) {
413 		svm_set_dirty(sc, vcpu, VMCB_CACHE_I);
414 		VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified "
415 		    "from %#x to %#x", idx, oldval, ctrl->intercept[idx]);
416 	}
417 }
418 
419 static __inline void
420 svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
421 {
422 
423 	svm_set_intercept(sc, vcpu, off, bitmask, 0);
424 }
425 
426 static __inline void
427 svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
428 {
429 
430 	svm_set_intercept(sc, vcpu, off, bitmask, 1);
431 }
432 
433 static void
434 vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa,
435     uint64_t msrpm_base_pa, uint64_t np_pml4)
436 {
437 	struct vmcb_ctrl *ctrl;
438 	struct vmcb_state *state;
439 	uint32_t mask;
440 	int n;
441 
442 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
443 	state = svm_get_vmcb_state(sc, vcpu);
444 
445 	ctrl->iopm_base_pa = iopm_base_pa;
446 	ctrl->msrpm_base_pa = msrpm_base_pa;
447 
448 	/* Enable nested paging */
449 	ctrl->np_enable = 1;
450 	ctrl->n_cr3 = np_pml4;
451 
452 	/*
453 	 * Intercept accesses to the control registers that are not shadowed
454 	 * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8.
455 	 */
456 	for (n = 0; n < 16; n++) {
457 		mask = (BIT(n) << 16) | BIT(n);
458 		if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8)
459 			svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
460 		else
461 			svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
462 	}
463 
464 	/*
465 	 * Intercept everything when tracing guest exceptions otherwise
466 	 * just intercept machine check exception.
467 	 */
468 	if (vcpu_trace_exceptions(sc->vm, vcpu)) {
469 		for (n = 0; n < 32; n++) {
470 			/*
471 			 * Skip unimplemented vectors in the exception bitmap.
472 			 */
473 			if (n == 2 || n == 9) {
474 				continue;
475 			}
476 			svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(n));
477 		}
478 	} else {
479 		svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC));
480 	}
481 
482 	/* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */
483 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO);
484 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR);
485 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID);
486 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR);
487 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT);
488 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI);
489 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI);
490 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN);
491 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
492 	    VMCB_INTCPT_FERR_FREEZE);
493 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INVD);
494 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INVLPGA);
495 
496 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR);
497 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT);
498 
499 	/*
500 	 * Intercept SVM instructions since AMD enables them in guests otherwise.
501 	 * Non-intercepted VMMCALL causes #UD, skip it.
502 	 */
503 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMLOAD);
504 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMSAVE);
505 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_STGI);
506 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_CLGI);
507 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_SKINIT);
508 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_ICEBP);
509 	if (vcpu_trap_wbinvd(sc->vm, vcpu)) {
510 		svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT,
511 		    VMCB_INTCPT_WBINVD);
512 	}
513 
514 	/*
515 	 * From section "Canonicalization and Consistency Checks" in APMv2
516 	 * the VMRUN intercept bit must be set to pass the consistency check.
517 	 */
518 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
519 
520 	/*
521 	 * The ASID will be set to a non-zero value just before VMRUN.
522 	 */
523 	ctrl->asid = 0;
524 
525 	/*
526 	 * Section 15.21.1, Interrupt Masking in EFLAGS
527 	 * Section 15.21.2, Virtualizing APIC.TPR
528 	 *
529 	 * This must be set for %rflag and %cr8 isolation of guest and host.
530 	 */
531 	ctrl->v_intr_masking = 1;
532 
533 	/* Enable Last Branch Record aka LBR for debugging */
534 	ctrl->lbr_virt_en = 1;
535 	state->dbgctl = BIT(0);
536 
537 	/* EFER_SVM must always be set when the guest is executing */
538 	state->efer = EFER_SVM;
539 
540 	/* Set up the PAT to power-on state */
541 	state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK)	|
542 	    PAT_VALUE(1, PAT_WRITE_THROUGH)	|
543 	    PAT_VALUE(2, PAT_UNCACHED)		|
544 	    PAT_VALUE(3, PAT_UNCACHEABLE)	|
545 	    PAT_VALUE(4, PAT_WRITE_BACK)	|
546 	    PAT_VALUE(5, PAT_WRITE_THROUGH)	|
547 	    PAT_VALUE(6, PAT_UNCACHED)		|
548 	    PAT_VALUE(7, PAT_UNCACHEABLE);
549 
550 	/* Set up DR6/7 to power-on state */
551 	state->dr6 = DBREG_DR6_RESERVED1;
552 	state->dr7 = DBREG_DR7_RESERVED1;
553 }
554 
555 /*
556  * Initialize a virtual machine.
557  */
558 static void *
559 svm_init(struct vm *vm, pmap_t pmap)
560 {
561 	struct svm_softc *svm_sc;
562 	struct svm_vcpu *vcpu;
563 	vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
564 	int i;
565 	uint16_t maxcpus;
566 
567 	svm_sc = malloc(sizeof (*svm_sc), M_SVM, M_WAITOK | M_ZERO);
568 	if (((uintptr_t)svm_sc & PAGE_MASK) != 0)
569 		panic("malloc of svm_softc not aligned on page boundary");
570 
571 	svm_sc->msr_bitmap = contigmalloc(SVM_MSR_BITMAP_SIZE, M_SVM,
572 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
573 	if (svm_sc->msr_bitmap == NULL)
574 		panic("contigmalloc of SVM MSR bitmap failed");
575 	svm_sc->iopm_bitmap = contigmalloc(SVM_IO_BITMAP_SIZE, M_SVM,
576 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
577 	if (svm_sc->iopm_bitmap == NULL)
578 		panic("contigmalloc of SVM IO bitmap failed");
579 
580 	svm_sc->vm = vm;
581 	svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pmltop);
582 
583 	/*
584 	 * Intercept read and write accesses to all MSRs.
585 	 */
586 	memset(svm_sc->msr_bitmap, 0xFF, SVM_MSR_BITMAP_SIZE);
587 
588 	/*
589 	 * Access to the following MSRs is redirected to the VMCB when the
590 	 * guest is executing. Therefore it is safe to allow the guest to
591 	 * read/write these MSRs directly without hypervisor involvement.
592 	 */
593 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
594 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
595 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
596 
597 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
598 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
599 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
600 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
601 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
602 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
603 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
604 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
605 
606 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
607 
608 	/*
609 	 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
610 	 */
611 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
612 
613 	/* Intercept access to all I/O ports. */
614 	memset(svm_sc->iopm_bitmap, 0xFF, SVM_IO_BITMAP_SIZE);
615 
616 	iopm_pa = vtophys(svm_sc->iopm_bitmap);
617 	msrpm_pa = vtophys(svm_sc->msr_bitmap);
618 	pml4_pa = svm_sc->nptp;
619 	maxcpus = vm_get_maxcpus(svm_sc->vm);
620 	for (i = 0; i < maxcpus; i++) {
621 		vcpu = svm_get_vcpu(svm_sc, i);
622 		vcpu->nextrip = ~0;
623 		vcpu->lastcpu = NOCPU;
624 		vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
625 		vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
626 		svm_msr_guest_init(svm_sc, i);
627 	}
628 	return (svm_sc);
629 }
630 
631 /*
632  * Collateral for a generic SVM VM-exit.
633  */
634 static void
635 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
636 {
637 
638 	vme->exitcode = VM_EXITCODE_SVM;
639 	vme->u.svm.exitcode = code;
640 	vme->u.svm.exitinfo1 = info1;
641 	vme->u.svm.exitinfo2 = info2;
642 }
643 
644 static int
645 svm_cpl(struct vmcb_state *state)
646 {
647 
648 	/*
649 	 * From APMv2:
650 	 *   "Retrieve the CPL from the CPL field in the VMCB, not
651 	 *    from any segment DPL"
652 	 */
653 	return (state->cpl);
654 }
655 
656 static enum vm_cpu_mode
657 svm_vcpu_mode(struct vmcb *vmcb)
658 {
659 	struct vmcb_segment seg;
660 	struct vmcb_state *state;
661 	int error __diagused;
662 
663 	state = &vmcb->state;
664 
665 	if (state->efer & EFER_LMA) {
666 		error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
667 		KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
668 		    error));
669 
670 		/*
671 		 * Section 4.8.1 for APM2, check if Code Segment has
672 		 * Long attribute set in descriptor.
673 		 */
674 		if (seg.attrib & VMCB_CS_ATTRIB_L)
675 			return (CPU_MODE_64BIT);
676 		else
677 			return (CPU_MODE_COMPATIBILITY);
678 	} else  if (state->cr0 & CR0_PE) {
679 		return (CPU_MODE_PROTECTED);
680 	} else {
681 		return (CPU_MODE_REAL);
682 	}
683 }
684 
685 static enum vm_paging_mode
686 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
687 {
688 
689 	if ((cr0 & CR0_PG) == 0)
690 		return (PAGING_MODE_FLAT);
691 	if ((cr4 & CR4_PAE) == 0)
692 		return (PAGING_MODE_32);
693 	if (efer & EFER_LME)
694 		return (PAGING_MODE_64);
695 	else
696 		return (PAGING_MODE_PAE);
697 }
698 
699 /*
700  * ins/outs utility routines
701  */
702 static uint64_t
703 svm_inout_str_index(struct svm_regctx *regs, int in)
704 {
705 	uint64_t val;
706 
707 	val = in ? regs->sctx_rdi : regs->sctx_rsi;
708 
709 	return (val);
710 }
711 
712 static uint64_t
713 svm_inout_str_count(struct svm_regctx *regs, int rep)
714 {
715 	uint64_t val;
716 
717 	val = rep ? regs->sctx_rcx : 1;
718 
719 	return (val);
720 }
721 
722 static void
723 svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
724     int in, struct vm_inout_str *vis)
725 {
726 	int error __diagused, s;
727 
728 	if (in) {
729 		vis->seg_name = VM_REG_GUEST_ES;
730 	} else {
731 		/* The segment field has standard encoding */
732 		s = (info1 >> 10) & 0x7;
733 		vis->seg_name = vm_segment_name(s);
734 	}
735 
736 	error = svm_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
737 	KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
738 }
739 
740 static int
741 svm_inout_str_addrsize(uint64_t info1)
742 {
743         uint32_t size;
744 
745         size = (info1 >> 7) & 0x7;
746         switch (size) {
747         case 1:
748                 return (2);     /* 16 bit */
749         case 2:
750                 return (4);     /* 32 bit */
751         case 4:
752                 return (8);     /* 64 bit */
753         default:
754                 panic("%s: invalid size encoding %d", __func__, size);
755         }
756 }
757 
758 static void
759 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
760 {
761 	struct vmcb_state *state;
762 
763 	state = &vmcb->state;
764 	paging->cr3 = state->cr3;
765 	paging->cpl = svm_cpl(state);
766 	paging->cpu_mode = svm_vcpu_mode(vmcb);
767 	paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
768 	    state->efer);
769 }
770 
771 #define	UNHANDLED 0
772 
773 /*
774  * Handle guest I/O intercept.
775  */
776 static int
777 svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
778 {
779 	struct vmcb_ctrl *ctrl;
780 	struct vmcb_state *state;
781 	struct svm_regctx *regs;
782 	struct vm_inout_str *vis;
783 	uint64_t info1;
784 	int inout_string;
785 
786 	state = svm_get_vmcb_state(svm_sc, vcpu);
787 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
788 	regs  = svm_get_guest_regctx(svm_sc, vcpu);
789 
790 	info1 = ctrl->exitinfo1;
791 	inout_string = info1 & BIT(2) ? 1 : 0;
792 
793 	/*
794 	 * The effective segment number in EXITINFO1[12:10] is populated
795 	 * only if the processor has the DecodeAssist capability.
796 	 *
797 	 * XXX this is not specified explicitly in APMv2 but can be verified
798 	 * empirically.
799 	 */
800 	if (inout_string && !decode_assist())
801 		return (UNHANDLED);
802 
803 	vmexit->exitcode 	= VM_EXITCODE_INOUT;
804 	vmexit->u.inout.in 	= (info1 & BIT(0)) ? 1 : 0;
805 	vmexit->u.inout.string 	= inout_string;
806 	vmexit->u.inout.rep 	= (info1 & BIT(3)) ? 1 : 0;
807 	vmexit->u.inout.bytes 	= (info1 >> 4) & 0x7;
808 	vmexit->u.inout.port 	= (uint16_t)(info1 >> 16);
809 	vmexit->u.inout.eax 	= (uint32_t)(state->rax);
810 
811 	if (inout_string) {
812 		vmexit->exitcode = VM_EXITCODE_INOUT_STR;
813 		vis = &vmexit->u.inout_str;
814 		svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
815 		vis->rflags = state->rflags;
816 		vis->cr0 = state->cr0;
817 		vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
818 		vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
819 		vis->addrsize = svm_inout_str_addrsize(info1);
820 		svm_inout_str_seginfo(svm_sc, vcpu, info1,
821 		    vmexit->u.inout.in, vis);
822 	}
823 
824 	return (UNHANDLED);
825 }
826 
827 static int
828 npf_fault_type(uint64_t exitinfo1)
829 {
830 
831 	if (exitinfo1 & VMCB_NPF_INFO1_W)
832 		return (VM_PROT_WRITE);
833 	else if (exitinfo1 & VMCB_NPF_INFO1_ID)
834 		return (VM_PROT_EXECUTE);
835 	else
836 		return (VM_PROT_READ);
837 }
838 
839 static bool
840 svm_npf_emul_fault(uint64_t exitinfo1)
841 {
842 
843 	if (exitinfo1 & VMCB_NPF_INFO1_ID) {
844 		return (false);
845 	}
846 
847 	if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
848 		return (false);
849 	}
850 
851 	if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
852 		return (false);
853 	}
854 
855 	return (true);
856 }
857 
858 static void
859 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
860 {
861 	struct vm_guest_paging *paging;
862 	struct vmcb_segment seg;
863 	struct vmcb_ctrl *ctrl;
864 	char *inst_bytes;
865 	int error __diagused, inst_len;
866 
867 	ctrl = &vmcb->ctrl;
868 	paging = &vmexit->u.inst_emul.paging;
869 
870 	vmexit->exitcode = VM_EXITCODE_INST_EMUL;
871 	vmexit->u.inst_emul.gpa = gpa;
872 	vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
873 	svm_paging_info(vmcb, paging);
874 
875 	error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
876 	KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
877 
878 	switch(paging->cpu_mode) {
879 	case CPU_MODE_REAL:
880 		vmexit->u.inst_emul.cs_base = seg.base;
881 		vmexit->u.inst_emul.cs_d = 0;
882 		break;
883 	case CPU_MODE_PROTECTED:
884 	case CPU_MODE_COMPATIBILITY:
885 		vmexit->u.inst_emul.cs_base = seg.base;
886 
887 		/*
888 		 * Section 4.8.1 of APM2, Default Operand Size or D bit.
889 		 */
890 		vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
891 		    1 : 0;
892 		break;
893 	default:
894 		vmexit->u.inst_emul.cs_base = 0;
895 		vmexit->u.inst_emul.cs_d = 0;
896 		break;
897 	}
898 
899 	/*
900 	 * Copy the instruction bytes into 'vie' if available.
901 	 */
902 	if (decode_assist() && !disable_npf_assist) {
903 		inst_len = ctrl->inst_len;
904 		inst_bytes = ctrl->inst_bytes;
905 	} else {
906 		inst_len = 0;
907 		inst_bytes = NULL;
908 	}
909 	vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
910 }
911 
912 #ifdef KTR
913 static const char *
914 intrtype_to_str(int intr_type)
915 {
916 	switch (intr_type) {
917 	case VMCB_EVENTINJ_TYPE_INTR:
918 		return ("hwintr");
919 	case VMCB_EVENTINJ_TYPE_NMI:
920 		return ("nmi");
921 	case VMCB_EVENTINJ_TYPE_INTn:
922 		return ("swintr");
923 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
924 		return ("exception");
925 	default:
926 		panic("%s: unknown intr_type %d", __func__, intr_type);
927 	}
928 }
929 #endif
930 
931 /*
932  * Inject an event to vcpu as described in section 15.20, "Event injection".
933  */
934 static void
935 svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
936 		 uint32_t error, bool ec_valid)
937 {
938 	struct vmcb_ctrl *ctrl;
939 
940 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
941 
942 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
943 	    ("%s: event already pending %#lx", __func__, ctrl->eventinj));
944 
945 	KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
946 	    __func__, vector));
947 
948 	switch (intr_type) {
949 	case VMCB_EVENTINJ_TYPE_INTR:
950 	case VMCB_EVENTINJ_TYPE_NMI:
951 	case VMCB_EVENTINJ_TYPE_INTn:
952 		break;
953 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
954 		if (vector >= 0 && vector <= 31 && vector != 2)
955 			break;
956 		/* FALLTHROUGH */
957 	default:
958 		panic("%s: invalid intr_type/vector: %d/%d", __func__,
959 		    intr_type, vector);
960 	}
961 	ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
962 	if (ec_valid) {
963 		ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
964 		ctrl->eventinj |= (uint64_t)error << 32;
965 		VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
966 		    intrtype_to_str(intr_type), vector, error);
967 	} else {
968 		VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
969 		    intrtype_to_str(intr_type), vector);
970 	}
971 }
972 
973 static void
974 svm_update_virqinfo(struct svm_softc *sc, int vcpu)
975 {
976 	struct vm *vm;
977 	struct vlapic *vlapic;
978 	struct vmcb_ctrl *ctrl;
979 
980 	vm = sc->vm;
981 	vlapic = vm_lapic(vm, vcpu);
982 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
983 
984 	/* Update %cr8 in the emulated vlapic */
985 	vlapic_set_cr8(vlapic, ctrl->v_tpr);
986 
987 	/* Virtual interrupt injection is not used. */
988 	KASSERT(ctrl->v_intr_vector == 0, ("%s: invalid "
989 	    "v_intr_vector %d", __func__, ctrl->v_intr_vector));
990 }
991 
992 static void
993 svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
994 {
995 	struct vmcb_ctrl *ctrl;
996 	uint64_t intinfo;
997 
998 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
999 	intinfo = ctrl->exitintinfo;
1000 	if (!VMCB_EXITINTINFO_VALID(intinfo))
1001 		return;
1002 
1003 	/*
1004 	 * From APMv2, Section "Intercepts during IDT interrupt delivery"
1005 	 *
1006 	 * If a #VMEXIT happened during event delivery then record the event
1007 	 * that was being delivered.
1008 	 */
1009 	VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
1010 		intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
1011 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
1012 	vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
1013 }
1014 
1015 #ifdef INVARIANTS
1016 static __inline int
1017 vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
1018 {
1019 
1020 	return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1021 	    VMCB_INTCPT_VINTR));
1022 }
1023 #endif
1024 
1025 static __inline void
1026 enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1027 {
1028 	struct vmcb_ctrl *ctrl;
1029 
1030 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1031 
1032 	if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
1033 		KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
1034 		KASSERT(vintr_intercept_enabled(sc, vcpu),
1035 		    ("%s: vintr intercept should be enabled", __func__));
1036 		return;
1037 	}
1038 
1039 	VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
1040 	ctrl->v_irq = 1;
1041 	ctrl->v_ign_tpr = 1;
1042 	ctrl->v_intr_vector = 0;
1043 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1044 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1045 }
1046 
1047 static __inline void
1048 disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1049 {
1050 	struct vmcb_ctrl *ctrl;
1051 
1052 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1053 
1054 	if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1055 		KASSERT(!vintr_intercept_enabled(sc, vcpu),
1056 		    ("%s: vintr intercept should be disabled", __func__));
1057 		return;
1058 	}
1059 
1060 	VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
1061 	ctrl->v_irq = 0;
1062 	ctrl->v_intr_vector = 0;
1063 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1064 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1065 }
1066 
1067 static int
1068 svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
1069 {
1070 	struct vmcb_ctrl *ctrl;
1071 	int oldval, newval;
1072 
1073 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1074 	oldval = ctrl->intr_shadow;
1075 	newval = val ? 1 : 0;
1076 	if (newval != oldval) {
1077 		ctrl->intr_shadow = newval;
1078 		VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1079 	}
1080 	return (0);
1081 }
1082 
1083 static int
1084 svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1085 {
1086 	struct vmcb_ctrl *ctrl;
1087 
1088 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1089 	*val = ctrl->intr_shadow;
1090 	return (0);
1091 }
1092 
1093 /*
1094  * Once an NMI is injected it blocks delivery of further NMIs until the handler
1095  * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1096  * to track when the vcpu is done handling the NMI.
1097  */
1098 static int
1099 nmi_blocked(struct svm_softc *sc, int vcpu)
1100 {
1101 	int blocked;
1102 
1103 	blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1104 	    VMCB_INTCPT_IRET);
1105 	return (blocked);
1106 }
1107 
1108 static void
1109 enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1110 {
1111 
1112 	KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1113 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1114 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1115 }
1116 
1117 static void
1118 clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1119 {
1120 	int error __diagused;
1121 
1122 	KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1123 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1124 	/*
1125 	 * When the IRET intercept is cleared the vcpu will attempt to execute
1126 	 * the "iret" when it runs next. However, it is possible to inject
1127 	 * another NMI into the vcpu before the "iret" has actually executed.
1128 	 *
1129 	 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1130 	 * it will trap back into the hypervisor. If an NMI is pending for
1131 	 * the vcpu it will be injected into the guest.
1132 	 *
1133 	 * XXX this needs to be fixed
1134 	 */
1135 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1136 
1137 	/*
1138 	 * Set 'intr_shadow' to prevent an NMI from being injected on the
1139 	 * immediate VMRUN.
1140 	 */
1141 	error = svm_modify_intr_shadow(sc, vcpu, 1);
1142 	KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1143 }
1144 
1145 #define	EFER_MBZ_BITS	0xFFFFFFFFFFFF0200UL
1146 
1147 static int
1148 svm_write_efer(struct svm_softc *sc, int vcpu, uint64_t newval, bool *retu)
1149 {
1150 	struct vm_exit *vme;
1151 	struct vmcb_state *state;
1152 	uint64_t changed, lma, oldval;
1153 	int error __diagused;
1154 
1155 	state = svm_get_vmcb_state(sc, vcpu);
1156 
1157 	oldval = state->efer;
1158 	VCPU_CTR2(sc->vm, vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval);
1159 
1160 	newval &= ~0xFE;		/* clear the Read-As-Zero (RAZ) bits */
1161 	changed = oldval ^ newval;
1162 
1163 	if (newval & EFER_MBZ_BITS)
1164 		goto gpf;
1165 
1166 	/* APMv2 Table 14-5 "Long-Mode Consistency Checks" */
1167 	if (changed & EFER_LME) {
1168 		if (state->cr0 & CR0_PG)
1169 			goto gpf;
1170 	}
1171 
1172 	/* EFER.LMA = EFER.LME & CR0.PG */
1173 	if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0)
1174 		lma = EFER_LMA;
1175 	else
1176 		lma = 0;
1177 
1178 	if ((newval & EFER_LMA) != lma)
1179 		goto gpf;
1180 
1181 	if (newval & EFER_NXE) {
1182 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_NO_EXECUTE))
1183 			goto gpf;
1184 	}
1185 
1186 	/*
1187 	 * XXX bhyve does not enforce segment limits in 64-bit mode. Until
1188 	 * this is fixed flag guest attempt to set EFER_LMSLE as an error.
1189 	 */
1190 	if (newval & EFER_LMSLE) {
1191 		vme = vm_exitinfo(sc->vm, vcpu);
1192 		vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0);
1193 		*retu = true;
1194 		return (0);
1195 	}
1196 
1197 	if (newval & EFER_FFXSR) {
1198 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_FFXSR))
1199 			goto gpf;
1200 	}
1201 
1202 	if (newval & EFER_TCE) {
1203 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_TCE))
1204 			goto gpf;
1205 	}
1206 
1207 	error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, newval);
1208 	KASSERT(error == 0, ("%s: error %d updating efer", __func__, error));
1209 	return (0);
1210 gpf:
1211 	vm_inject_gp(sc->vm, vcpu);
1212 	return (0);
1213 }
1214 
1215 static int
1216 emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1217     bool *retu)
1218 {
1219 	int error;
1220 
1221 	if (lapic_msr(num))
1222 		error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1223 	else if (num == MSR_EFER)
1224 		error = svm_write_efer(sc, vcpu, val, retu);
1225 	else
1226 		error = svm_wrmsr(sc, vcpu, num, val, retu);
1227 
1228 	return (error);
1229 }
1230 
1231 static int
1232 emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1233 {
1234 	struct vmcb_state *state;
1235 	struct svm_regctx *ctx;
1236 	uint64_t result;
1237 	int error;
1238 
1239 	if (lapic_msr(num))
1240 		error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1241 	else
1242 		error = svm_rdmsr(sc, vcpu, num, &result, retu);
1243 
1244 	if (error == 0) {
1245 		state = svm_get_vmcb_state(sc, vcpu);
1246 		ctx = svm_get_guest_regctx(sc, vcpu);
1247 		state->rax = result & 0xffffffff;
1248 		ctx->sctx_rdx = result >> 32;
1249 	}
1250 
1251 	return (error);
1252 }
1253 
1254 #ifdef KTR
1255 static const char *
1256 exit_reason_to_str(uint64_t reason)
1257 {
1258 	int i;
1259 	static char reasonbuf[32];
1260 	static const struct {
1261 		int reason;
1262 		const char *str;
1263 	} reasons[] = {
1264 		{ .reason = VMCB_EXIT_INVALID,	.str = "invalvmcb" },
1265 		{ .reason = VMCB_EXIT_SHUTDOWN,	.str = "shutdown" },
1266 		{ .reason = VMCB_EXIT_NPF, 	.str = "nptfault" },
1267 		{ .reason = VMCB_EXIT_PAUSE,	.str = "pause" },
1268 		{ .reason = VMCB_EXIT_HLT,	.str = "hlt" },
1269 		{ .reason = VMCB_EXIT_CPUID,	.str = "cpuid" },
1270 		{ .reason = VMCB_EXIT_IO,	.str = "inout" },
1271 		{ .reason = VMCB_EXIT_MC,	.str = "mchk" },
1272 		{ .reason = VMCB_EXIT_INTR,	.str = "extintr" },
1273 		{ .reason = VMCB_EXIT_NMI,	.str = "nmi" },
1274 		{ .reason = VMCB_EXIT_VINTR,	.str = "vintr" },
1275 		{ .reason = VMCB_EXIT_MSR,	.str = "msr" },
1276 		{ .reason = VMCB_EXIT_IRET,	.str = "iret" },
1277 		{ .reason = VMCB_EXIT_MONITOR,	.str = "monitor" },
1278 		{ .reason = VMCB_EXIT_MWAIT,	.str = "mwait" },
1279 		{ .reason = VMCB_EXIT_VMRUN,	.str = "vmrun" },
1280 		{ .reason = VMCB_EXIT_VMMCALL,	.str = "vmmcall" },
1281 		{ .reason = VMCB_EXIT_VMLOAD,	.str = "vmload" },
1282 		{ .reason = VMCB_EXIT_VMSAVE,	.str = "vmsave" },
1283 		{ .reason = VMCB_EXIT_STGI,	.str = "stgi" },
1284 		{ .reason = VMCB_EXIT_CLGI,	.str = "clgi" },
1285 		{ .reason = VMCB_EXIT_SKINIT,	.str = "skinit" },
1286 		{ .reason = VMCB_EXIT_ICEBP,	.str = "icebp" },
1287 		{ .reason = VMCB_EXIT_INVD,	.str = "invd" },
1288 		{ .reason = VMCB_EXIT_INVLPGA,	.str = "invlpga" },
1289 	};
1290 
1291 	for (i = 0; i < nitems(reasons); i++) {
1292 		if (reasons[i].reason == reason)
1293 			return (reasons[i].str);
1294 	}
1295 	snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1296 	return (reasonbuf);
1297 }
1298 #endif	/* KTR */
1299 
1300 /*
1301  * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1302  * that are due to instruction intercepts as well as MSR and IOIO intercepts
1303  * and exceptions caused by INT3, INTO and BOUND instructions.
1304  *
1305  * Return 1 if the nRIP is valid and 0 otherwise.
1306  */
1307 static int
1308 nrip_valid(uint64_t exitcode)
1309 {
1310 	switch (exitcode) {
1311 	case 0x00 ... 0x0F:	/* read of CR0 through CR15 */
1312 	case 0x10 ... 0x1F:	/* write of CR0 through CR15 */
1313 	case 0x20 ... 0x2F:	/* read of DR0 through DR15 */
1314 	case 0x30 ... 0x3F:	/* write of DR0 through DR15 */
1315 	case 0x43:		/* INT3 */
1316 	case 0x44:		/* INTO */
1317 	case 0x45:		/* BOUND */
1318 	case 0x65 ... 0x7C:	/* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1319 	case 0x80 ... 0x8D:	/* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1320 		return (1);
1321 	default:
1322 		return (0);
1323 	}
1324 }
1325 
1326 static int
1327 svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1328 {
1329 	struct vmcb *vmcb;
1330 	struct vmcb_state *state;
1331 	struct vmcb_ctrl *ctrl;
1332 	struct svm_regctx *ctx;
1333 	uint64_t code, info1, info2, val;
1334 	uint32_t eax, ecx, edx;
1335 	int error __diagused, errcode_valid, handled, idtvec, reflect;
1336 	bool retu;
1337 
1338 	ctx = svm_get_guest_regctx(svm_sc, vcpu);
1339 	vmcb = svm_get_vmcb(svm_sc, vcpu);
1340 	state = &vmcb->state;
1341 	ctrl = &vmcb->ctrl;
1342 
1343 	handled = 0;
1344 	code = ctrl->exitcode;
1345 	info1 = ctrl->exitinfo1;
1346 	info2 = ctrl->exitinfo2;
1347 
1348 	vmexit->exitcode = VM_EXITCODE_BOGUS;
1349 	vmexit->rip = state->rip;
1350 	vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1351 
1352 	vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1353 
1354 	/*
1355 	 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1356 	 * in an inconsistent state and can trigger assertions that would
1357 	 * never happen otherwise.
1358 	 */
1359 	if (code == VMCB_EXIT_INVALID) {
1360 		vm_exit_svm(vmexit, code, info1, info2);
1361 		return (0);
1362 	}
1363 
1364 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1365 	    "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1366 
1367 	KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1368 	    ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1369 	    vmexit->inst_length, code, info1, info2));
1370 
1371 	svm_update_virqinfo(svm_sc, vcpu);
1372 	svm_save_intinfo(svm_sc, vcpu);
1373 
1374 	switch (code) {
1375 	case VMCB_EXIT_IRET:
1376 		/*
1377 		 * Restart execution at "iret" but with the intercept cleared.
1378 		 */
1379 		vmexit->inst_length = 0;
1380 		clear_nmi_blocking(svm_sc, vcpu);
1381 		handled = 1;
1382 		break;
1383 	case VMCB_EXIT_VINTR:	/* interrupt window exiting */
1384 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1385 		handled = 1;
1386 		break;
1387 	case VMCB_EXIT_INTR:	/* external interrupt */
1388 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1389 		handled = 1;
1390 		break;
1391 	case VMCB_EXIT_NMI:	/* external NMI */
1392 		handled = 1;
1393 		break;
1394 	case 0x40 ... 0x5F:
1395 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1396 		reflect = 1;
1397 		idtvec = code - 0x40;
1398 		switch (idtvec) {
1399 		case IDT_MC:
1400 			/*
1401 			 * Call the machine check handler by hand. Also don't
1402 			 * reflect the machine check back into the guest.
1403 			 */
1404 			reflect = 0;
1405 			VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler");
1406 			__asm __volatile("int $18");
1407 			break;
1408 		case IDT_PF:
1409 			error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2,
1410 			    info2);
1411 			KASSERT(error == 0, ("%s: error %d updating cr2",
1412 			    __func__, error));
1413 			/* fallthru */
1414 		case IDT_NP:
1415 		case IDT_SS:
1416 		case IDT_GP:
1417 		case IDT_AC:
1418 		case IDT_TS:
1419 			errcode_valid = 1;
1420 			break;
1421 
1422 		case IDT_DF:
1423 			errcode_valid = 1;
1424 			info1 = 0;
1425 			break;
1426 
1427 		case IDT_BP:
1428 		case IDT_OF:
1429 		case IDT_BR:
1430 			/*
1431 			 * The 'nrip' field is populated for INT3, INTO and
1432 			 * BOUND exceptions and this also implies that
1433 			 * 'inst_length' is non-zero.
1434 			 *
1435 			 * Reset 'inst_length' to zero so the guest %rip at
1436 			 * event injection is identical to what it was when
1437 			 * the exception originally happened.
1438 			 */
1439 			VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d "
1440 			    "to zero before injecting exception %d",
1441 			    vmexit->inst_length, idtvec);
1442 			vmexit->inst_length = 0;
1443 			/* fallthru */
1444 		default:
1445 			errcode_valid = 0;
1446 			info1 = 0;
1447 			break;
1448 		}
1449 		KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1450 		    "when reflecting exception %d into guest",
1451 		    vmexit->inst_length, idtvec));
1452 
1453 		if (reflect) {
1454 			/* Reflect the exception back into the guest */
1455 			VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception "
1456 			    "%d/%#x into the guest", idtvec, (int)info1);
1457 			error = vm_inject_exception(svm_sc->vm, vcpu, idtvec,
1458 			    errcode_valid, info1, 0);
1459 			KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1460 			    __func__, error));
1461 		}
1462 		handled = 1;
1463 		break;
1464 	case VMCB_EXIT_MSR:	/* MSR access. */
1465 		eax = state->rax;
1466 		ecx = ctx->sctx_rcx;
1467 		edx = ctx->sctx_rdx;
1468 		retu = false;
1469 
1470 		if (info1) {
1471 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1472 			val = (uint64_t)edx << 32 | eax;
1473 			VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1474 			    ecx, val);
1475 			if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1476 				vmexit->exitcode = VM_EXITCODE_WRMSR;
1477 				vmexit->u.msr.code = ecx;
1478 				vmexit->u.msr.wval = val;
1479 			} else if (!retu) {
1480 				handled = 1;
1481 			} else {
1482 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1483 				    ("emulate_wrmsr retu with bogus exitcode"));
1484 			}
1485 		} else {
1486 			VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1487 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1488 			if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1489 				vmexit->exitcode = VM_EXITCODE_RDMSR;
1490 				vmexit->u.msr.code = ecx;
1491 			} else if (!retu) {
1492 				handled = 1;
1493 			} else {
1494 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1495 				    ("emulate_rdmsr retu with bogus exitcode"));
1496 			}
1497 		}
1498 		break;
1499 	case VMCB_EXIT_IO:
1500 		handled = svm_handle_io(svm_sc, vcpu, vmexit);
1501 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1502 		break;
1503 	case VMCB_EXIT_CPUID:
1504 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1505 		handled = x86_emulate_cpuid(svm_sc->vm, vcpu, &state->rax,
1506 		    &ctx->sctx_rbx, &ctx->sctx_rcx, &ctx->sctx_rdx);
1507 		break;
1508 	case VMCB_EXIT_HLT:
1509 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1510 		vmexit->exitcode = VM_EXITCODE_HLT;
1511 		vmexit->u.hlt.rflags = state->rflags;
1512 		break;
1513 	case VMCB_EXIT_PAUSE:
1514 		vmexit->exitcode = VM_EXITCODE_PAUSE;
1515 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1516 		break;
1517 	case VMCB_EXIT_NPF:
1518 		/* EXITINFO2 contains the faulting guest physical address */
1519 		if (info1 & VMCB_NPF_INFO1_RSV) {
1520 			VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1521 			    "reserved bits set: info1(%#lx) info2(%#lx)",
1522 			    info1, info2);
1523 		} else if (vm_mem_allocated(svm_sc->vm, vcpu, info2)) {
1524 			vmexit->exitcode = VM_EXITCODE_PAGING;
1525 			vmexit->u.paging.gpa = info2;
1526 			vmexit->u.paging.fault_type = npf_fault_type(info1);
1527 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1528 			VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1529 			    "on gpa %#lx/%#lx at rip %#lx",
1530 			    info2, info1, state->rip);
1531 		} else if (svm_npf_emul_fault(info1)) {
1532 			svm_handle_inst_emul(vmcb, info2, vmexit);
1533 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1534 			VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1535 			    "for gpa %#lx/%#lx at rip %#lx",
1536 			    info2, info1, state->rip);
1537 		}
1538 		break;
1539 	case VMCB_EXIT_MONITOR:
1540 		vmexit->exitcode = VM_EXITCODE_MONITOR;
1541 		break;
1542 	case VMCB_EXIT_MWAIT:
1543 		vmexit->exitcode = VM_EXITCODE_MWAIT;
1544 		break;
1545 	case VMCB_EXIT_SHUTDOWN:
1546 	case VMCB_EXIT_VMRUN:
1547 	case VMCB_EXIT_VMMCALL:
1548 	case VMCB_EXIT_VMLOAD:
1549 	case VMCB_EXIT_VMSAVE:
1550 	case VMCB_EXIT_STGI:
1551 	case VMCB_EXIT_CLGI:
1552 	case VMCB_EXIT_SKINIT:
1553 	case VMCB_EXIT_ICEBP:
1554 	case VMCB_EXIT_INVLPGA:
1555 		vm_inject_ud(svm_sc->vm, vcpu);
1556 		handled = 1;
1557 		break;
1558 	case VMCB_EXIT_INVD:
1559 	case VMCB_EXIT_WBINVD:
1560 		/* ignore exit */
1561 		handled = 1;
1562 		break;
1563 	default:
1564 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1565 		break;
1566 	}
1567 
1568 	VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1569 	    handled ? "handled" : "unhandled", exit_reason_to_str(code),
1570 	    vmexit->rip, vmexit->inst_length);
1571 
1572 	if (handled) {
1573 		vmexit->rip += vmexit->inst_length;
1574 		vmexit->inst_length = 0;
1575 		state->rip = vmexit->rip;
1576 	} else {
1577 		if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1578 			/*
1579 			 * If this VM exit was not claimed by anybody then
1580 			 * treat it as a generic SVM exit.
1581 			 */
1582 			vm_exit_svm(vmexit, code, info1, info2);
1583 		} else {
1584 			/*
1585 			 * The exitcode and collateral have been populated.
1586 			 * The VM exit will be processed further in userland.
1587 			 */
1588 		}
1589 	}
1590 	return (handled);
1591 }
1592 
1593 static void
1594 svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1595 {
1596 	uint64_t intinfo;
1597 
1598 	if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1599 		return;
1600 
1601 	KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1602 	    "valid: %#lx", __func__, intinfo));
1603 
1604 	svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1605 		VMCB_EXITINTINFO_VECTOR(intinfo),
1606 		VMCB_EXITINTINFO_EC(intinfo),
1607 		VMCB_EXITINTINFO_EC_VALID(intinfo));
1608 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1609 	VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1610 }
1611 
1612 /*
1613  * Inject event to virtual cpu.
1614  */
1615 static void
1616 svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1617 {
1618 	struct vmcb_ctrl *ctrl;
1619 	struct vmcb_state *state;
1620 	struct svm_vcpu *vcpustate;
1621 	uint8_t v_tpr;
1622 	int vector, need_intr_window;
1623 	int extint_pending;
1624 
1625 	state = svm_get_vmcb_state(sc, vcpu);
1626 	ctrl  = svm_get_vmcb_ctrl(sc, vcpu);
1627 	vcpustate = svm_get_vcpu(sc, vcpu);
1628 
1629 	need_intr_window = 0;
1630 
1631 	if (vcpustate->nextrip != state->rip) {
1632 		ctrl->intr_shadow = 0;
1633 		VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking "
1634 		    "cleared due to rip change: %#lx/%#lx",
1635 		    vcpustate->nextrip, state->rip);
1636 	}
1637 
1638 	/*
1639 	 * Inject pending events or exceptions for this vcpu.
1640 	 *
1641 	 * An event might be pending because the previous #VMEXIT happened
1642 	 * during event delivery (i.e. ctrl->exitintinfo).
1643 	 *
1644 	 * An event might also be pending because an exception was injected
1645 	 * by the hypervisor (e.g. #PF during instruction emulation).
1646 	 */
1647 	svm_inj_intinfo(sc, vcpu);
1648 
1649 	/* NMI event has priority over interrupts. */
1650 	if (vm_nmi_pending(sc->vm, vcpu)) {
1651 		if (nmi_blocked(sc, vcpu)) {
1652 			/*
1653 			 * Can't inject another NMI if the guest has not
1654 			 * yet executed an "iret" after the last NMI.
1655 			 */
1656 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1657 			    "to NMI-blocking");
1658 		} else if (ctrl->intr_shadow) {
1659 			/*
1660 			 * Can't inject an NMI if the vcpu is in an intr_shadow.
1661 			 */
1662 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1663 			    "interrupt shadow");
1664 			need_intr_window = 1;
1665 			goto done;
1666 		} else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1667 			/*
1668 			 * If there is already an exception/interrupt pending
1669 			 * then defer the NMI until after that.
1670 			 */
1671 			VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1672 			    "eventinj %#lx", ctrl->eventinj);
1673 
1674 			/*
1675 			 * Use self-IPI to trigger a VM-exit as soon as
1676 			 * possible after the event injection is completed.
1677 			 *
1678 			 * This works only if the external interrupt exiting
1679 			 * is at a lower priority than the event injection.
1680 			 *
1681 			 * Although not explicitly specified in APMv2 the
1682 			 * relative priorities were verified empirically.
1683 			 */
1684 			ipi_cpu(curcpu, IPI_AST);	/* XXX vmm_ipinum? */
1685 		} else {
1686 			vm_nmi_clear(sc->vm, vcpu);
1687 
1688 			/* Inject NMI, vector number is not used */
1689 			svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1690 			    IDT_NMI, 0, false);
1691 
1692 			/* virtual NMI blocking is now in effect */
1693 			enable_nmi_blocking(sc, vcpu);
1694 
1695 			VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1696 		}
1697 	}
1698 
1699 	extint_pending = vm_extint_pending(sc->vm, vcpu);
1700 	if (!extint_pending) {
1701 		if (!vlapic_pending_intr(vlapic, &vector))
1702 			goto done;
1703 		KASSERT(vector >= 16 && vector <= 255,
1704 		    ("invalid vector %d from local APIC", vector));
1705 	} else {
1706 		/* Ask the legacy pic for a vector to inject */
1707 		vatpic_pending_intr(sc->vm, &vector);
1708 		KASSERT(vector >= 0 && vector <= 255,
1709 		    ("invalid vector %d from INTR", vector));
1710 	}
1711 
1712 	/*
1713 	 * If the guest has disabled interrupts or is in an interrupt shadow
1714 	 * then we cannot inject the pending interrupt.
1715 	 */
1716 	if ((state->rflags & PSL_I) == 0) {
1717 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1718 		    "rflags %#lx", vector, state->rflags);
1719 		need_intr_window = 1;
1720 		goto done;
1721 	}
1722 
1723 	if (ctrl->intr_shadow) {
1724 		VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1725 		    "interrupt shadow", vector);
1726 		need_intr_window = 1;
1727 		goto done;
1728 	}
1729 
1730 	if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1731 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1732 		    "eventinj %#lx", vector, ctrl->eventinj);
1733 		need_intr_window = 1;
1734 		goto done;
1735 	}
1736 
1737 	svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1738 
1739 	if (!extint_pending) {
1740 		vlapic_intr_accepted(vlapic, vector);
1741 	} else {
1742 		vm_extint_clear(sc->vm, vcpu);
1743 		vatpic_intr_accepted(sc->vm, vector);
1744 	}
1745 
1746 	/*
1747 	 * Force a VM-exit as soon as the vcpu is ready to accept another
1748 	 * interrupt. This is done because the PIC might have another vector
1749 	 * that it wants to inject. Also, if the APIC has a pending interrupt
1750 	 * that was preempted by the ExtInt then it allows us to inject the
1751 	 * APIC vector as soon as possible.
1752 	 */
1753 	need_intr_window = 1;
1754 done:
1755 	/*
1756 	 * The guest can modify the TPR by writing to %CR8. In guest mode
1757 	 * the processor reflects this write to V_TPR without hypervisor
1758 	 * intervention.
1759 	 *
1760 	 * The guest can also modify the TPR by writing to it via the memory
1761 	 * mapped APIC page. In this case, the write will be emulated by the
1762 	 * hypervisor. For this reason V_TPR must be updated before every
1763 	 * VMRUN.
1764 	 */
1765 	v_tpr = vlapic_get_cr8(vlapic);
1766 	KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1767 	if (ctrl->v_tpr != v_tpr) {
1768 		VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1769 		    ctrl->v_tpr, v_tpr);
1770 		ctrl->v_tpr = v_tpr;
1771 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1772 	}
1773 
1774 	if (need_intr_window) {
1775 		/*
1776 		 * We use V_IRQ in conjunction with the VINTR intercept to
1777 		 * trap into the hypervisor as soon as a virtual interrupt
1778 		 * can be delivered.
1779 		 *
1780 		 * Since injected events are not subject to intercept checks
1781 		 * we need to ensure that the V_IRQ is not actually going to
1782 		 * be delivered on VM entry. The KASSERT below enforces this.
1783 		 */
1784 		KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1785 		    (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1786 		    ("Bogus intr_window_exiting: eventinj (%#lx), "
1787 		    "intr_shadow (%u), rflags (%#lx)",
1788 		    ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1789 		enable_intr_window_exiting(sc, vcpu);
1790 	} else {
1791 		disable_intr_window_exiting(sc, vcpu);
1792 	}
1793 }
1794 
1795 static __inline void
1796 restore_host_tss(void)
1797 {
1798 	struct system_segment_descriptor *tss_sd;
1799 
1800 	/*
1801 	 * The TSS descriptor was in use prior to launching the guest so it
1802 	 * has been marked busy.
1803 	 *
1804 	 * 'ltr' requires the descriptor to be marked available so change the
1805 	 * type to "64-bit available TSS".
1806 	 */
1807 	tss_sd = PCPU_GET(tss);
1808 	tss_sd->sd_type = SDT_SYSTSS;
1809 	ltr(GSEL(GPROC0_SEL, SEL_KPL));
1810 }
1811 
1812 static void
1813 svm_pmap_activate(struct svm_softc *sc, int vcpuid, pmap_t pmap)
1814 {
1815 	struct svm_vcpu *vcpustate;
1816 	struct vmcb_ctrl *ctrl;
1817 	long eptgen;
1818 	int cpu;
1819 	bool alloc_asid;
1820 
1821 	cpu = curcpu;
1822 	CPU_SET_ATOMIC(cpu, &pmap->pm_active);
1823 	smr_enter(pmap->pm_eptsmr);
1824 
1825 	vcpustate = svm_get_vcpu(sc, vcpuid);
1826 	ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1827 
1828 	/*
1829 	 * The TLB entries associated with the vcpu's ASID are not valid
1830 	 * if either of the following conditions is true:
1831 	 *
1832 	 * 1. The vcpu's ASID generation is different than the host cpu's
1833 	 *    ASID generation. This happens when the vcpu migrates to a new
1834 	 *    host cpu. It can also happen when the number of vcpus executing
1835 	 *    on a host cpu is greater than the number of ASIDs available.
1836 	 *
1837 	 * 2. The pmap generation number is different than the value cached in
1838 	 *    the 'vcpustate'. This happens when the host invalidates pages
1839 	 *    belonging to the guest.
1840 	 *
1841 	 *	asidgen		eptgen	      Action
1842 	 *	mismatch	mismatch
1843 	 *	   0		   0		(a)
1844 	 *	   0		   1		(b1) or (b2)
1845 	 *	   1		   0		(c)
1846 	 *	   1		   1		(d)
1847 	 *
1848 	 * (a) There is no mismatch in eptgen or ASID generation and therefore
1849 	 *     no further action is needed.
1850 	 *
1851 	 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1852 	 *      retained and the TLB entries associated with this ASID
1853 	 *      are flushed by VMRUN.
1854 	 *
1855 	 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1856 	 *      allocated.
1857 	 *
1858 	 * (c) A new ASID is allocated.
1859 	 *
1860 	 * (d) A new ASID is allocated.
1861 	 */
1862 
1863 	alloc_asid = false;
1864 	eptgen = atomic_load_long(&pmap->pm_eptgen);
1865 	ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1866 
1867 	if (vcpustate->asid.gen != asid[cpu].gen) {
1868 		alloc_asid = true;	/* (c) and (d) */
1869 	} else if (vcpustate->eptgen != eptgen) {
1870 		if (flush_by_asid())
1871 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;	/* (b1) */
1872 		else
1873 			alloc_asid = true;			/* (b2) */
1874 	} else {
1875 		/*
1876 		 * This is the common case (a).
1877 		 */
1878 		KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1879 		KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1880 		    ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1881 	}
1882 
1883 	if (alloc_asid) {
1884 		if (++asid[cpu].num >= nasid) {
1885 			asid[cpu].num = 1;
1886 			if (++asid[cpu].gen == 0)
1887 				asid[cpu].gen = 1;
1888 			/*
1889 			 * If this cpu does not support "flush-by-asid"
1890 			 * then flush the entire TLB on a generation
1891 			 * bump. Subsequent ASID allocation in this
1892 			 * generation can be done without a TLB flush.
1893 			 */
1894 			if (!flush_by_asid())
1895 				ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1896 		}
1897 		vcpustate->asid.gen = asid[cpu].gen;
1898 		vcpustate->asid.num = asid[cpu].num;
1899 
1900 		ctrl->asid = vcpustate->asid.num;
1901 		svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1902 		/*
1903 		 * If this cpu supports "flush-by-asid" then the TLB
1904 		 * was not flushed after the generation bump. The TLB
1905 		 * is flushed selectively after every new ASID allocation.
1906 		 */
1907 		if (flush_by_asid())
1908 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1909 	}
1910 	vcpustate->eptgen = eptgen;
1911 
1912 	KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1913 	KASSERT(ctrl->asid == vcpustate->asid.num,
1914 	    ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1915 }
1916 
1917 static void
1918 svm_pmap_deactivate(pmap_t pmap)
1919 {
1920 	smr_exit(pmap->pm_eptsmr);
1921 	CPU_CLR_ATOMIC(curcpu, &pmap->pm_active);
1922 }
1923 
1924 static __inline void
1925 disable_gintr(void)
1926 {
1927 
1928 	__asm __volatile("clgi");
1929 }
1930 
1931 static __inline void
1932 enable_gintr(void)
1933 {
1934 
1935         __asm __volatile("stgi");
1936 }
1937 
1938 static __inline void
1939 svm_dr_enter_guest(struct svm_regctx *gctx)
1940 {
1941 
1942 	/* Save host control debug registers. */
1943 	gctx->host_dr7 = rdr7();
1944 	gctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR);
1945 
1946 	/*
1947 	 * Disable debugging in DR7 and DEBUGCTL to avoid triggering
1948 	 * exceptions in the host based on the guest DRx values.  The
1949 	 * guest DR6, DR7, and DEBUGCTL are saved/restored in the
1950 	 * VMCB.
1951 	 */
1952 	load_dr7(0);
1953 	wrmsr(MSR_DEBUGCTLMSR, 0);
1954 
1955 	/* Save host debug registers. */
1956 	gctx->host_dr0 = rdr0();
1957 	gctx->host_dr1 = rdr1();
1958 	gctx->host_dr2 = rdr2();
1959 	gctx->host_dr3 = rdr3();
1960 	gctx->host_dr6 = rdr6();
1961 
1962 	/* Restore guest debug registers. */
1963 	load_dr0(gctx->sctx_dr0);
1964 	load_dr1(gctx->sctx_dr1);
1965 	load_dr2(gctx->sctx_dr2);
1966 	load_dr3(gctx->sctx_dr3);
1967 }
1968 
1969 static __inline void
1970 svm_dr_leave_guest(struct svm_regctx *gctx)
1971 {
1972 
1973 	/* Save guest debug registers. */
1974 	gctx->sctx_dr0 = rdr0();
1975 	gctx->sctx_dr1 = rdr1();
1976 	gctx->sctx_dr2 = rdr2();
1977 	gctx->sctx_dr3 = rdr3();
1978 
1979 	/*
1980 	 * Restore host debug registers.  Restore DR7 and DEBUGCTL
1981 	 * last.
1982 	 */
1983 	load_dr0(gctx->host_dr0);
1984 	load_dr1(gctx->host_dr1);
1985 	load_dr2(gctx->host_dr2);
1986 	load_dr3(gctx->host_dr3);
1987 	load_dr6(gctx->host_dr6);
1988 	wrmsr(MSR_DEBUGCTLMSR, gctx->host_debugctl);
1989 	load_dr7(gctx->host_dr7);
1990 }
1991 
1992 /*
1993  * Start vcpu with specified RIP.
1994  */
1995 static int
1996 svm_run(void *arg, int vcpu, register_t rip, pmap_t pmap,
1997 	struct vm_eventinfo *evinfo)
1998 {
1999 	struct svm_regctx *gctx;
2000 	struct svm_softc *svm_sc;
2001 	struct svm_vcpu *vcpustate;
2002 	struct vmcb_state *state;
2003 	struct vmcb_ctrl *ctrl;
2004 	struct vm_exit *vmexit;
2005 	struct vlapic *vlapic;
2006 	struct vm *vm;
2007 	uint64_t vmcb_pa;
2008 	int handled;
2009 	uint16_t ldt_sel;
2010 
2011 	svm_sc = arg;
2012 	vm = svm_sc->vm;
2013 
2014 	vcpustate = svm_get_vcpu(svm_sc, vcpu);
2015 	state = svm_get_vmcb_state(svm_sc, vcpu);
2016 	ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
2017 	vmexit = vm_exitinfo(vm, vcpu);
2018 	vlapic = vm_lapic(vm, vcpu);
2019 
2020 	gctx = svm_get_guest_regctx(svm_sc, vcpu);
2021 	vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
2022 
2023 	if (vcpustate->lastcpu != curcpu) {
2024 		/*
2025 		 * Force new ASID allocation by invalidating the generation.
2026 		 */
2027 		vcpustate->asid.gen = 0;
2028 
2029 		/*
2030 		 * Invalidate the VMCB state cache by marking all fields dirty.
2031 		 */
2032 		svm_set_dirty(svm_sc, vcpu, 0xffffffff);
2033 
2034 		/*
2035 		 * XXX
2036 		 * Setting 'vcpustate->lastcpu' here is bit premature because
2037 		 * we may return from this function without actually executing
2038 		 * the VMRUN  instruction. This could happen if a rendezvous
2039 		 * or an AST is pending on the first time through the loop.
2040 		 *
2041 		 * This works for now but any new side-effects of vcpu
2042 		 * migration should take this case into account.
2043 		 */
2044 		vcpustate->lastcpu = curcpu;
2045 		vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
2046 	}
2047 
2048 	svm_msr_guest_enter(svm_sc, vcpu);
2049 
2050 	/* Update Guest RIP */
2051 	state->rip = rip;
2052 
2053 	do {
2054 		/*
2055 		 * Disable global interrupts to guarantee atomicity during
2056 		 * loading of guest state. This includes not only the state
2057 		 * loaded by the "vmrun" instruction but also software state
2058 		 * maintained by the hypervisor: suspended and rendezvous
2059 		 * state, NPT generation number, vlapic interrupts etc.
2060 		 */
2061 		disable_gintr();
2062 
2063 		if (vcpu_suspended(evinfo)) {
2064 			enable_gintr();
2065 			vm_exit_suspended(vm, vcpu, state->rip);
2066 			break;
2067 		}
2068 
2069 		if (vcpu_rendezvous_pending(evinfo)) {
2070 			enable_gintr();
2071 			vm_exit_rendezvous(vm, vcpu, state->rip);
2072 			break;
2073 		}
2074 
2075 		if (vcpu_reqidle(evinfo)) {
2076 			enable_gintr();
2077 			vm_exit_reqidle(vm, vcpu, state->rip);
2078 			break;
2079 		}
2080 
2081 		/* We are asked to give the cpu by scheduler. */
2082 		if (vcpu_should_yield(vm, vcpu)) {
2083 			enable_gintr();
2084 			vm_exit_astpending(vm, vcpu, state->rip);
2085 			break;
2086 		}
2087 
2088 		if (vcpu_debugged(vm, vcpu)) {
2089 			enable_gintr();
2090 			vm_exit_debug(vm, vcpu, state->rip);
2091 			break;
2092 		}
2093 
2094 		/*
2095 		 * #VMEXIT resumes the host with the guest LDTR, so
2096 		 * save the current LDT selector so it can be restored
2097 		 * after an exit.  The userspace hypervisor probably
2098 		 * doesn't use a LDT, but save and restore it to be
2099 		 * safe.
2100 		 */
2101 		ldt_sel = sldt();
2102 
2103 		svm_inj_interrupts(svm_sc, vcpu, vlapic);
2104 
2105 		/*
2106 		 * Check the pmap generation and the ASID generation to
2107 		 * ensure that the vcpu does not use stale TLB mappings.
2108 		 */
2109 		svm_pmap_activate(svm_sc, vcpu, pmap);
2110 
2111 		ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
2112 		vcpustate->dirty = 0;
2113 		VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
2114 
2115 		/* Launch Virtual Machine. */
2116 		VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
2117 		svm_dr_enter_guest(gctx);
2118 		svm_launch(vmcb_pa, gctx, get_pcpu());
2119 		svm_dr_leave_guest(gctx);
2120 
2121 		svm_pmap_deactivate(pmap);
2122 
2123 		/*
2124 		 * The host GDTR and IDTR is saved by VMRUN and restored
2125 		 * automatically on #VMEXIT. However, the host TSS needs
2126 		 * to be restored explicitly.
2127 		 */
2128 		restore_host_tss();
2129 
2130 		/* Restore host LDTR. */
2131 		lldt(ldt_sel);
2132 
2133 		/* #VMEXIT disables interrupts so re-enable them here. */
2134 		enable_gintr();
2135 
2136 		/* Update 'nextrip' */
2137 		vcpustate->nextrip = state->rip;
2138 
2139 		/* Handle #VMEXIT and if required return to user space. */
2140 		handled = svm_vmexit(svm_sc, vcpu, vmexit);
2141 	} while (handled);
2142 
2143 	svm_msr_guest_exit(svm_sc, vcpu);
2144 
2145 	return (0);
2146 }
2147 
2148 static void
2149 svm_cleanup(void *arg)
2150 {
2151 	struct svm_softc *sc = arg;
2152 
2153 	contigfree(sc->iopm_bitmap, SVM_IO_BITMAP_SIZE, M_SVM);
2154 	contigfree(sc->msr_bitmap, SVM_MSR_BITMAP_SIZE, M_SVM);
2155 	free(sc, M_SVM);
2156 }
2157 
2158 static register_t *
2159 swctx_regptr(struct svm_regctx *regctx, int reg)
2160 {
2161 
2162 	switch (reg) {
2163 	case VM_REG_GUEST_RBX:
2164 		return (&regctx->sctx_rbx);
2165 	case VM_REG_GUEST_RCX:
2166 		return (&regctx->sctx_rcx);
2167 	case VM_REG_GUEST_RDX:
2168 		return (&regctx->sctx_rdx);
2169 	case VM_REG_GUEST_RDI:
2170 		return (&regctx->sctx_rdi);
2171 	case VM_REG_GUEST_RSI:
2172 		return (&regctx->sctx_rsi);
2173 	case VM_REG_GUEST_RBP:
2174 		return (&regctx->sctx_rbp);
2175 	case VM_REG_GUEST_R8:
2176 		return (&regctx->sctx_r8);
2177 	case VM_REG_GUEST_R9:
2178 		return (&regctx->sctx_r9);
2179 	case VM_REG_GUEST_R10:
2180 		return (&regctx->sctx_r10);
2181 	case VM_REG_GUEST_R11:
2182 		return (&regctx->sctx_r11);
2183 	case VM_REG_GUEST_R12:
2184 		return (&regctx->sctx_r12);
2185 	case VM_REG_GUEST_R13:
2186 		return (&regctx->sctx_r13);
2187 	case VM_REG_GUEST_R14:
2188 		return (&regctx->sctx_r14);
2189 	case VM_REG_GUEST_R15:
2190 		return (&regctx->sctx_r15);
2191 	case VM_REG_GUEST_DR0:
2192 		return (&regctx->sctx_dr0);
2193 	case VM_REG_GUEST_DR1:
2194 		return (&regctx->sctx_dr1);
2195 	case VM_REG_GUEST_DR2:
2196 		return (&regctx->sctx_dr2);
2197 	case VM_REG_GUEST_DR3:
2198 		return (&regctx->sctx_dr3);
2199 	default:
2200 		return (NULL);
2201 	}
2202 }
2203 
2204 static int
2205 svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
2206 {
2207 	struct svm_softc *svm_sc;
2208 	register_t *reg;
2209 
2210 	svm_sc = arg;
2211 
2212 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2213 		return (svm_get_intr_shadow(svm_sc, vcpu, val));
2214 	}
2215 
2216 	if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2217 		return (0);
2218 	}
2219 
2220 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2221 
2222 	if (reg != NULL) {
2223 		*val = *reg;
2224 		return (0);
2225 	}
2226 
2227 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident);
2228 	return (EINVAL);
2229 }
2230 
2231 static int
2232 svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2233 {
2234 	struct svm_softc *svm_sc;
2235 	register_t *reg;
2236 
2237 	svm_sc = arg;
2238 
2239 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2240 		return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2241 	}
2242 
2243 	/* Do not permit user write access to VMCB fields by offset. */
2244 	if (!VMCB_ACCESS_OK(ident)) {
2245 		if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2246 			return (0);
2247 		}
2248 	}
2249 
2250 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2251 
2252 	if (reg != NULL) {
2253 		*reg = val;
2254 		return (0);
2255 	}
2256 
2257 	if (ident == VM_REG_GUEST_ENTRY_INST_LENGTH) {
2258 		/* Ignore. */
2259 		return (0);
2260 	}
2261 
2262 	/*
2263 	 * XXX deal with CR3 and invalidate TLB entries tagged with the
2264 	 * vcpu's ASID. This needs to be treated differently depending on
2265 	 * whether 'running' is true/false.
2266 	 */
2267 
2268 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident);
2269 	return (EINVAL);
2270 }
2271 
2272 static int
2273 svm_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc)
2274 {
2275 	return (vmcb_getdesc(arg, vcpu, reg, desc));
2276 }
2277 
2278 static int
2279 svm_setdesc(void *arg, int vcpu, int reg, struct seg_desc *desc)
2280 {
2281 	return (vmcb_setdesc(arg, vcpu, reg, desc));
2282 }
2283 
2284 #ifdef BHYVE_SNAPSHOT
2285 static int
2286 svm_snapshot_reg(void *arg, int vcpu, int ident,
2287 		 struct vm_snapshot_meta *meta)
2288 {
2289 	int ret;
2290 	uint64_t val;
2291 
2292 	if (meta->op == VM_SNAPSHOT_SAVE) {
2293 		ret = svm_getreg(arg, vcpu, ident, &val);
2294 		if (ret != 0)
2295 			goto done;
2296 
2297 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2298 	} else if (meta->op == VM_SNAPSHOT_RESTORE) {
2299 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2300 
2301 		ret = svm_setreg(arg, vcpu, ident, val);
2302 		if (ret != 0)
2303 			goto done;
2304 	} else {
2305 		ret = EINVAL;
2306 		goto done;
2307 	}
2308 
2309 done:
2310 	return (ret);
2311 }
2312 #endif
2313 
2314 static int
2315 svm_setcap(void *arg, int vcpu, int type, int val)
2316 {
2317 	struct svm_softc *sc;
2318 	int error;
2319 
2320 	sc = arg;
2321 	error = 0;
2322 	switch (type) {
2323 	case VM_CAP_HALT_EXIT:
2324 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2325 		    VMCB_INTCPT_HLT, val);
2326 		break;
2327 	case VM_CAP_PAUSE_EXIT:
2328 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2329 		    VMCB_INTCPT_PAUSE, val);
2330 		break;
2331 	case VM_CAP_UNRESTRICTED_GUEST:
2332 		/* Unrestricted guest execution cannot be disabled in SVM */
2333 		if (val == 0)
2334 			error = EINVAL;
2335 		break;
2336 	default:
2337 		error = ENOENT;
2338 		break;
2339 	}
2340 	return (error);
2341 }
2342 
2343 static int
2344 svm_getcap(void *arg, int vcpu, int type, int *retval)
2345 {
2346 	struct svm_softc *sc;
2347 	int error;
2348 
2349 	sc = arg;
2350 	error = 0;
2351 
2352 	switch (type) {
2353 	case VM_CAP_HALT_EXIT:
2354 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2355 		    VMCB_INTCPT_HLT);
2356 		break;
2357 	case VM_CAP_PAUSE_EXIT:
2358 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2359 		    VMCB_INTCPT_PAUSE);
2360 		break;
2361 	case VM_CAP_UNRESTRICTED_GUEST:
2362 		*retval = 1;	/* unrestricted guest is always enabled */
2363 		break;
2364 	default:
2365 		error = ENOENT;
2366 		break;
2367 	}
2368 	return (error);
2369 }
2370 
2371 static struct vmspace *
2372 svm_vmspace_alloc(vm_offset_t min, vm_offset_t max)
2373 {
2374 	return (svm_npt_alloc(min, max));
2375 }
2376 
2377 static void
2378 svm_vmspace_free(struct vmspace *vmspace)
2379 {
2380 	svm_npt_free(vmspace);
2381 }
2382 
2383 static struct vlapic *
2384 svm_vlapic_init(void *arg, int vcpuid)
2385 {
2386 	struct svm_softc *svm_sc;
2387 	struct vlapic *vlapic;
2388 
2389 	svm_sc = arg;
2390 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2391 	vlapic->vm = svm_sc->vm;
2392 	vlapic->vcpuid = vcpuid;
2393 	vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2394 
2395 	vlapic_init(vlapic);
2396 
2397 	return (vlapic);
2398 }
2399 
2400 static void
2401 svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2402 {
2403 
2404         vlapic_cleanup(vlapic);
2405         free(vlapic, M_SVM_VLAPIC);
2406 }
2407 
2408 #ifdef BHYVE_SNAPSHOT
2409 static int
2410 svm_snapshot(void *arg, struct vm_snapshot_meta *meta)
2411 {
2412 	/* struct svm_softc is AMD's representation for SVM softc */
2413 	struct svm_softc *sc;
2414 	struct svm_vcpu *vcpu;
2415 	int i;
2416 	int ret;
2417 
2418 	sc = arg;
2419 
2420 	KASSERT(sc != NULL, ("%s: arg was NULL", __func__));
2421 
2422 	for (i = 0; i < VM_MAXCPU; i++) {
2423 		vcpu = &sc->vcpu[i];
2424 
2425 		/* Snapshot swctx for virtual cpu i */
2426 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbp, meta, ret, done);
2427 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbx, meta, ret, done);
2428 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rcx, meta, ret, done);
2429 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdx, meta, ret, done);
2430 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdi, meta, ret, done);
2431 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rsi, meta, ret, done);
2432 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r8, meta, ret, done);
2433 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r9, meta, ret, done);
2434 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r10, meta, ret, done);
2435 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r11, meta, ret, done);
2436 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r12, meta, ret, done);
2437 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r13, meta, ret, done);
2438 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r14, meta, ret, done);
2439 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r15, meta, ret, done);
2440 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr0, meta, ret, done);
2441 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr1, meta, ret, done);
2442 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr2, meta, ret, done);
2443 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr3, meta, ret, done);
2444 
2445 		/* Restore other svm_vcpu struct fields */
2446 
2447 		/* Restore NEXTRIP field */
2448 		SNAPSHOT_VAR_OR_LEAVE(vcpu->nextrip, meta, ret, done);
2449 
2450 		/* Restore lastcpu field */
2451 		SNAPSHOT_VAR_OR_LEAVE(vcpu->lastcpu, meta, ret, done);
2452 		SNAPSHOT_VAR_OR_LEAVE(vcpu->dirty, meta, ret, done);
2453 
2454 		/* Restore EPTGEN field - EPT is Extended Page Table */
2455 		SNAPSHOT_VAR_OR_LEAVE(vcpu->eptgen, meta, ret, done);
2456 
2457 		SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.gen, meta, ret, done);
2458 		SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.num, meta, ret, done);
2459 
2460 		/* Set all caches dirty */
2461 		if (meta->op == VM_SNAPSHOT_RESTORE) {
2462 			svm_set_dirty(sc, i, VMCB_CACHE_ASID);
2463 			svm_set_dirty(sc, i, VMCB_CACHE_IOPM);
2464 			svm_set_dirty(sc, i, VMCB_CACHE_I);
2465 			svm_set_dirty(sc, i, VMCB_CACHE_TPR);
2466 			svm_set_dirty(sc, i, VMCB_CACHE_CR2);
2467 			svm_set_dirty(sc, i, VMCB_CACHE_CR);
2468 			svm_set_dirty(sc, i, VMCB_CACHE_DT);
2469 			svm_set_dirty(sc, i, VMCB_CACHE_SEG);
2470 			svm_set_dirty(sc, i, VMCB_CACHE_NP);
2471 		}
2472 	}
2473 
2474 	if (meta->op == VM_SNAPSHOT_RESTORE)
2475 		flush_by_asid();
2476 
2477 done:
2478 	return (ret);
2479 }
2480 
2481 static int
2482 svm_vmcx_snapshot(void *arg, struct vm_snapshot_meta *meta, int vcpu)
2483 {
2484 	struct svm_softc *sc;
2485 	int err, running, hostcpu;
2486 
2487 	sc = (struct svm_softc *)arg;
2488 	err = 0;
2489 
2490 	KASSERT(arg != NULL, ("%s: arg was NULL", __func__));
2491 
2492 	running = vcpu_is_running(sc->vm, vcpu, &hostcpu);
2493 	if (running && hostcpu !=curcpu) {
2494 		printf("%s: %s%d is running", __func__, vm_name(sc->vm), vcpu);
2495 		return (EINVAL);
2496 	}
2497 
2498 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR0, meta);
2499 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR2, meta);
2500 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR3, meta);
2501 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR4, meta);
2502 
2503 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_DR6, meta);
2504 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_DR7, meta);
2505 
2506 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RAX, meta);
2507 
2508 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RSP, meta);
2509 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RIP, meta);
2510 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RFLAGS, meta);
2511 
2512 	/* Guest segments */
2513 	/* ES */
2514 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_ES, meta);
2515 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_ES, meta);
2516 
2517 	/* CS */
2518 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CS, meta);
2519 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_CS, meta);
2520 
2521 	/* SS */
2522 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_SS, meta);
2523 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_SS, meta);
2524 
2525 	/* DS */
2526 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_DS, meta);
2527 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_DS, meta);
2528 
2529 	/* FS */
2530 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_FS, meta);
2531 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_FS, meta);
2532 
2533 	/* GS */
2534 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_GS, meta);
2535 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_GS, meta);
2536 
2537 	/* TR */
2538 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_TR, meta);
2539 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_TR, meta);
2540 
2541 	/* LDTR */
2542 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_LDTR, meta);
2543 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_LDTR, meta);
2544 
2545 	/* EFER */
2546 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_EFER, meta);
2547 
2548 	/* IDTR and GDTR */
2549 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_IDTR, meta);
2550 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_GDTR, meta);
2551 
2552 	/* Specific AMD registers */
2553 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_INTR_SHADOW, meta);
2554 
2555 	err += vmcb_snapshot_any(sc, vcpu,
2556 				VMCB_ACCESS(VMCB_OFF_CR_INTERCEPT, 4), meta);
2557 	err += vmcb_snapshot_any(sc, vcpu,
2558 				VMCB_ACCESS(VMCB_OFF_DR_INTERCEPT, 4), meta);
2559 	err += vmcb_snapshot_any(sc, vcpu,
2560 				VMCB_ACCESS(VMCB_OFF_EXC_INTERCEPT, 4), meta);
2561 	err += vmcb_snapshot_any(sc, vcpu,
2562 				VMCB_ACCESS(VMCB_OFF_INST1_INTERCEPT, 4), meta);
2563 	err += vmcb_snapshot_any(sc, vcpu,
2564 				VMCB_ACCESS(VMCB_OFF_INST2_INTERCEPT, 4), meta);
2565 
2566 	err += vmcb_snapshot_any(sc, vcpu,
2567 				VMCB_ACCESS(VMCB_OFF_PAUSE_FILTHRESH, 2), meta);
2568 	err += vmcb_snapshot_any(sc, vcpu,
2569 				VMCB_ACCESS(VMCB_OFF_PAUSE_FILCNT, 2), meta);
2570 
2571 	err += vmcb_snapshot_any(sc, vcpu,
2572 				VMCB_ACCESS(VMCB_OFF_ASID, 4), meta);
2573 
2574 	err += vmcb_snapshot_any(sc, vcpu,
2575 				VMCB_ACCESS(VMCB_OFF_TLB_CTRL, 4), meta);
2576 
2577 	err += vmcb_snapshot_any(sc, vcpu,
2578 				VMCB_ACCESS(VMCB_OFF_VIRQ, 8), meta);
2579 
2580 	err += vmcb_snapshot_any(sc, vcpu,
2581 				VMCB_ACCESS(VMCB_OFF_EXIT_REASON, 8), meta);
2582 	err += vmcb_snapshot_any(sc, vcpu,
2583 				VMCB_ACCESS(VMCB_OFF_EXITINFO1, 8), meta);
2584 	err += vmcb_snapshot_any(sc, vcpu,
2585 				VMCB_ACCESS(VMCB_OFF_EXITINFO2, 8), meta);
2586 	err += vmcb_snapshot_any(sc, vcpu,
2587 				VMCB_ACCESS(VMCB_OFF_EXITINTINFO, 8), meta);
2588 
2589 	err += vmcb_snapshot_any(sc, vcpu,
2590 				VMCB_ACCESS(VMCB_OFF_NP_ENABLE, 1), meta);
2591 
2592 	err += vmcb_snapshot_any(sc, vcpu,
2593 				VMCB_ACCESS(VMCB_OFF_AVIC_BAR, 8), meta);
2594 	err += vmcb_snapshot_any(sc, vcpu,
2595 				VMCB_ACCESS(VMCB_OFF_AVIC_PAGE, 8), meta);
2596 	err += vmcb_snapshot_any(sc, vcpu,
2597 				VMCB_ACCESS(VMCB_OFF_AVIC_LT, 8), meta);
2598 	err += vmcb_snapshot_any(sc, vcpu,
2599 				VMCB_ACCESS(VMCB_OFF_AVIC_PT, 8), meta);
2600 
2601 	err += vmcb_snapshot_any(sc, vcpu,
2602 				VMCB_ACCESS(VMCB_OFF_CPL, 1), meta);
2603 
2604 	err += vmcb_snapshot_any(sc, vcpu,
2605 				VMCB_ACCESS(VMCB_OFF_STAR, 8), meta);
2606 	err += vmcb_snapshot_any(sc, vcpu,
2607 				VMCB_ACCESS(VMCB_OFF_LSTAR, 8), meta);
2608 	err += vmcb_snapshot_any(sc, vcpu,
2609 				VMCB_ACCESS(VMCB_OFF_CSTAR, 8), meta);
2610 
2611 	err += vmcb_snapshot_any(sc, vcpu,
2612 				VMCB_ACCESS(VMCB_OFF_SFMASK, 8), meta);
2613 
2614 	err += vmcb_snapshot_any(sc, vcpu,
2615 				VMCB_ACCESS(VMCB_OFF_KERNELGBASE, 8), meta);
2616 
2617 	err += vmcb_snapshot_any(sc, vcpu,
2618 				VMCB_ACCESS(VMCB_OFF_SYSENTER_CS, 8), meta);
2619 	err += vmcb_snapshot_any(sc, vcpu,
2620 				VMCB_ACCESS(VMCB_OFF_SYSENTER_ESP, 8), meta);
2621 	err += vmcb_snapshot_any(sc, vcpu,
2622 				VMCB_ACCESS(VMCB_OFF_SYSENTER_EIP, 8), meta);
2623 
2624 	err += vmcb_snapshot_any(sc, vcpu,
2625 				VMCB_ACCESS(VMCB_OFF_GUEST_PAT, 8), meta);
2626 
2627 	err += vmcb_snapshot_any(sc, vcpu,
2628 				VMCB_ACCESS(VMCB_OFF_DBGCTL, 8), meta);
2629 	err += vmcb_snapshot_any(sc, vcpu,
2630 				VMCB_ACCESS(VMCB_OFF_BR_FROM, 8), meta);
2631 	err += vmcb_snapshot_any(sc, vcpu,
2632 				VMCB_ACCESS(VMCB_OFF_BR_TO, 8), meta);
2633 	err += vmcb_snapshot_any(sc, vcpu,
2634 				VMCB_ACCESS(VMCB_OFF_INT_FROM, 8), meta);
2635 	err += vmcb_snapshot_any(sc, vcpu,
2636 				VMCB_ACCESS(VMCB_OFF_INT_TO, 8), meta);
2637 
2638 	return (err);
2639 }
2640 
2641 static int
2642 svm_restore_tsc(void *arg, int vcpu, uint64_t offset)
2643 {
2644 	int err;
2645 
2646 	err = svm_set_tsc_offset(arg, vcpu, offset);
2647 
2648 	return (err);
2649 }
2650 #endif
2651 
2652 const struct vmm_ops vmm_ops_amd = {
2653 	.modinit	= svm_modinit,
2654 	.modcleanup	= svm_modcleanup,
2655 	.modresume	= svm_modresume,
2656 	.init		= svm_init,
2657 	.run		= svm_run,
2658 	.cleanup	= svm_cleanup,
2659 	.getreg		= svm_getreg,
2660 	.setreg		= svm_setreg,
2661 	.getdesc	= svm_getdesc,
2662 	.setdesc	= svm_setdesc,
2663 	.getcap		= svm_getcap,
2664 	.setcap		= svm_setcap,
2665 	.vmspace_alloc	= svm_vmspace_alloc,
2666 	.vmspace_free	= svm_vmspace_free,
2667 	.vlapic_init	= svm_vlapic_init,
2668 	.vlapic_cleanup	= svm_vlapic_cleanup,
2669 #ifdef BHYVE_SNAPSHOT
2670 	.snapshot	= svm_snapshot,
2671 	.vmcx_snapshot	= svm_vmcx_snapshot,
2672 	.restore_tsc	= svm_restore_tsc,
2673 #endif
2674 };
2675