xref: /freebsd/sys/amd64/vmm/amd/svm.c (revision b9e5884ef786a6c0fd203064baa07ef308025707)
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 
510 	/*
511 	 * From section "Canonicalization and Consistency Checks" in APMv2
512 	 * the VMRUN intercept bit must be set to pass the consistency check.
513 	 */
514 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
515 
516 	/*
517 	 * The ASID will be set to a non-zero value just before VMRUN.
518 	 */
519 	ctrl->asid = 0;
520 
521 	/*
522 	 * Section 15.21.1, Interrupt Masking in EFLAGS
523 	 * Section 15.21.2, Virtualizing APIC.TPR
524 	 *
525 	 * This must be set for %rflag and %cr8 isolation of guest and host.
526 	 */
527 	ctrl->v_intr_masking = 1;
528 
529 	/* Enable Last Branch Record aka LBR for debugging */
530 	ctrl->lbr_virt_en = 1;
531 	state->dbgctl = BIT(0);
532 
533 	/* EFER_SVM must always be set when the guest is executing */
534 	state->efer = EFER_SVM;
535 
536 	/* Set up the PAT to power-on state */
537 	state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK)	|
538 	    PAT_VALUE(1, PAT_WRITE_THROUGH)	|
539 	    PAT_VALUE(2, PAT_UNCACHED)		|
540 	    PAT_VALUE(3, PAT_UNCACHEABLE)	|
541 	    PAT_VALUE(4, PAT_WRITE_BACK)	|
542 	    PAT_VALUE(5, PAT_WRITE_THROUGH)	|
543 	    PAT_VALUE(6, PAT_UNCACHED)		|
544 	    PAT_VALUE(7, PAT_UNCACHEABLE);
545 
546 	/* Set up DR6/7 to power-on state */
547 	state->dr6 = DBREG_DR6_RESERVED1;
548 	state->dr7 = DBREG_DR7_RESERVED1;
549 }
550 
551 /*
552  * Initialize a virtual machine.
553  */
554 static void *
555 svm_init(struct vm *vm, pmap_t pmap)
556 {
557 	struct svm_softc *svm_sc;
558 	struct svm_vcpu *vcpu;
559 	vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
560 	int i;
561 	uint16_t maxcpus;
562 
563 	svm_sc = malloc(sizeof (*svm_sc), M_SVM, M_WAITOK | M_ZERO);
564 	if (((uintptr_t)svm_sc & PAGE_MASK) != 0)
565 		panic("malloc of svm_softc not aligned on page boundary");
566 
567 	svm_sc->msr_bitmap = contigmalloc(SVM_MSR_BITMAP_SIZE, M_SVM,
568 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
569 	if (svm_sc->msr_bitmap == NULL)
570 		panic("contigmalloc of SVM MSR bitmap failed");
571 	svm_sc->iopm_bitmap = contigmalloc(SVM_IO_BITMAP_SIZE, M_SVM,
572 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
573 	if (svm_sc->iopm_bitmap == NULL)
574 		panic("contigmalloc of SVM IO bitmap failed");
575 
576 	svm_sc->vm = vm;
577 	svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pmltop);
578 
579 	/*
580 	 * Intercept read and write accesses to all MSRs.
581 	 */
582 	memset(svm_sc->msr_bitmap, 0xFF, SVM_MSR_BITMAP_SIZE);
583 
584 	/*
585 	 * Access to the following MSRs is redirected to the VMCB when the
586 	 * guest is executing. Therefore it is safe to allow the guest to
587 	 * read/write these MSRs directly without hypervisor involvement.
588 	 */
589 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
590 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
591 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
592 
593 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
594 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
595 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
596 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
597 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
598 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
599 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
600 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
601 
602 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
603 
604 	/*
605 	 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
606 	 */
607 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
608 
609 	/* Intercept access to all I/O ports. */
610 	memset(svm_sc->iopm_bitmap, 0xFF, SVM_IO_BITMAP_SIZE);
611 
612 	iopm_pa = vtophys(svm_sc->iopm_bitmap);
613 	msrpm_pa = vtophys(svm_sc->msr_bitmap);
614 	pml4_pa = svm_sc->nptp;
615 	maxcpus = vm_get_maxcpus(svm_sc->vm);
616 	for (i = 0; i < maxcpus; i++) {
617 		vcpu = svm_get_vcpu(svm_sc, i);
618 		vcpu->nextrip = ~0;
619 		vcpu->lastcpu = NOCPU;
620 		vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
621 		vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
622 		svm_msr_guest_init(svm_sc, i);
623 	}
624 	return (svm_sc);
625 }
626 
627 /*
628  * Collateral for a generic SVM VM-exit.
629  */
630 static void
631 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
632 {
633 
634 	vme->exitcode = VM_EXITCODE_SVM;
635 	vme->u.svm.exitcode = code;
636 	vme->u.svm.exitinfo1 = info1;
637 	vme->u.svm.exitinfo2 = info2;
638 }
639 
640 static int
641 svm_cpl(struct vmcb_state *state)
642 {
643 
644 	/*
645 	 * From APMv2:
646 	 *   "Retrieve the CPL from the CPL field in the VMCB, not
647 	 *    from any segment DPL"
648 	 */
649 	return (state->cpl);
650 }
651 
652 static enum vm_cpu_mode
653 svm_vcpu_mode(struct vmcb *vmcb)
654 {
655 	struct vmcb_segment seg;
656 	struct vmcb_state *state;
657 	int error __diagused;
658 
659 	state = &vmcb->state;
660 
661 	if (state->efer & EFER_LMA) {
662 		error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
663 		KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
664 		    error));
665 
666 		/*
667 		 * Section 4.8.1 for APM2, check if Code Segment has
668 		 * Long attribute set in descriptor.
669 		 */
670 		if (seg.attrib & VMCB_CS_ATTRIB_L)
671 			return (CPU_MODE_64BIT);
672 		else
673 			return (CPU_MODE_COMPATIBILITY);
674 	} else  if (state->cr0 & CR0_PE) {
675 		return (CPU_MODE_PROTECTED);
676 	} else {
677 		return (CPU_MODE_REAL);
678 	}
679 }
680 
681 static enum vm_paging_mode
682 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
683 {
684 
685 	if ((cr0 & CR0_PG) == 0)
686 		return (PAGING_MODE_FLAT);
687 	if ((cr4 & CR4_PAE) == 0)
688 		return (PAGING_MODE_32);
689 	if (efer & EFER_LME)
690 		return (PAGING_MODE_64);
691 	else
692 		return (PAGING_MODE_PAE);
693 }
694 
695 /*
696  * ins/outs utility routines
697  */
698 static uint64_t
699 svm_inout_str_index(struct svm_regctx *regs, int in)
700 {
701 	uint64_t val;
702 
703 	val = in ? regs->sctx_rdi : regs->sctx_rsi;
704 
705 	return (val);
706 }
707 
708 static uint64_t
709 svm_inout_str_count(struct svm_regctx *regs, int rep)
710 {
711 	uint64_t val;
712 
713 	val = rep ? regs->sctx_rcx : 1;
714 
715 	return (val);
716 }
717 
718 static void
719 svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
720     int in, struct vm_inout_str *vis)
721 {
722 	int error __diagused, s;
723 
724 	if (in) {
725 		vis->seg_name = VM_REG_GUEST_ES;
726 	} else {
727 		/* The segment field has standard encoding */
728 		s = (info1 >> 10) & 0x7;
729 		vis->seg_name = vm_segment_name(s);
730 	}
731 
732 	error = svm_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
733 	KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
734 }
735 
736 static int
737 svm_inout_str_addrsize(uint64_t info1)
738 {
739         uint32_t size;
740 
741         size = (info1 >> 7) & 0x7;
742         switch (size) {
743         case 1:
744                 return (2);     /* 16 bit */
745         case 2:
746                 return (4);     /* 32 bit */
747         case 4:
748                 return (8);     /* 64 bit */
749         default:
750                 panic("%s: invalid size encoding %d", __func__, size);
751         }
752 }
753 
754 static void
755 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
756 {
757 	struct vmcb_state *state;
758 
759 	state = &vmcb->state;
760 	paging->cr3 = state->cr3;
761 	paging->cpl = svm_cpl(state);
762 	paging->cpu_mode = svm_vcpu_mode(vmcb);
763 	paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
764 	    state->efer);
765 }
766 
767 #define	UNHANDLED 0
768 
769 /*
770  * Handle guest I/O intercept.
771  */
772 static int
773 svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
774 {
775 	struct vmcb_ctrl *ctrl;
776 	struct vmcb_state *state;
777 	struct svm_regctx *regs;
778 	struct vm_inout_str *vis;
779 	uint64_t info1;
780 	int inout_string;
781 
782 	state = svm_get_vmcb_state(svm_sc, vcpu);
783 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
784 	regs  = svm_get_guest_regctx(svm_sc, vcpu);
785 
786 	info1 = ctrl->exitinfo1;
787 	inout_string = info1 & BIT(2) ? 1 : 0;
788 
789 	/*
790 	 * The effective segment number in EXITINFO1[12:10] is populated
791 	 * only if the processor has the DecodeAssist capability.
792 	 *
793 	 * XXX this is not specified explicitly in APMv2 but can be verified
794 	 * empirically.
795 	 */
796 	if (inout_string && !decode_assist())
797 		return (UNHANDLED);
798 
799 	vmexit->exitcode 	= VM_EXITCODE_INOUT;
800 	vmexit->u.inout.in 	= (info1 & BIT(0)) ? 1 : 0;
801 	vmexit->u.inout.string 	= inout_string;
802 	vmexit->u.inout.rep 	= (info1 & BIT(3)) ? 1 : 0;
803 	vmexit->u.inout.bytes 	= (info1 >> 4) & 0x7;
804 	vmexit->u.inout.port 	= (uint16_t)(info1 >> 16);
805 	vmexit->u.inout.eax 	= (uint32_t)(state->rax);
806 
807 	if (inout_string) {
808 		vmexit->exitcode = VM_EXITCODE_INOUT_STR;
809 		vis = &vmexit->u.inout_str;
810 		svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
811 		vis->rflags = state->rflags;
812 		vis->cr0 = state->cr0;
813 		vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
814 		vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
815 		vis->addrsize = svm_inout_str_addrsize(info1);
816 		svm_inout_str_seginfo(svm_sc, vcpu, info1,
817 		    vmexit->u.inout.in, vis);
818 	}
819 
820 	return (UNHANDLED);
821 }
822 
823 static int
824 npf_fault_type(uint64_t exitinfo1)
825 {
826 
827 	if (exitinfo1 & VMCB_NPF_INFO1_W)
828 		return (VM_PROT_WRITE);
829 	else if (exitinfo1 & VMCB_NPF_INFO1_ID)
830 		return (VM_PROT_EXECUTE);
831 	else
832 		return (VM_PROT_READ);
833 }
834 
835 static bool
836 svm_npf_emul_fault(uint64_t exitinfo1)
837 {
838 
839 	if (exitinfo1 & VMCB_NPF_INFO1_ID) {
840 		return (false);
841 	}
842 
843 	if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
844 		return (false);
845 	}
846 
847 	if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
848 		return (false);
849 	}
850 
851 	return (true);
852 }
853 
854 static void
855 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
856 {
857 	struct vm_guest_paging *paging;
858 	struct vmcb_segment seg;
859 	struct vmcb_ctrl *ctrl;
860 	char *inst_bytes;
861 	int error __diagused, inst_len;
862 
863 	ctrl = &vmcb->ctrl;
864 	paging = &vmexit->u.inst_emul.paging;
865 
866 	vmexit->exitcode = VM_EXITCODE_INST_EMUL;
867 	vmexit->u.inst_emul.gpa = gpa;
868 	vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
869 	svm_paging_info(vmcb, paging);
870 
871 	error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
872 	KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
873 
874 	switch(paging->cpu_mode) {
875 	case CPU_MODE_REAL:
876 		vmexit->u.inst_emul.cs_base = seg.base;
877 		vmexit->u.inst_emul.cs_d = 0;
878 		break;
879 	case CPU_MODE_PROTECTED:
880 	case CPU_MODE_COMPATIBILITY:
881 		vmexit->u.inst_emul.cs_base = seg.base;
882 
883 		/*
884 		 * Section 4.8.1 of APM2, Default Operand Size or D bit.
885 		 */
886 		vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
887 		    1 : 0;
888 		break;
889 	default:
890 		vmexit->u.inst_emul.cs_base = 0;
891 		vmexit->u.inst_emul.cs_d = 0;
892 		break;
893 	}
894 
895 	/*
896 	 * Copy the instruction bytes into 'vie' if available.
897 	 */
898 	if (decode_assist() && !disable_npf_assist) {
899 		inst_len = ctrl->inst_len;
900 		inst_bytes = ctrl->inst_bytes;
901 	} else {
902 		inst_len = 0;
903 		inst_bytes = NULL;
904 	}
905 	vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
906 }
907 
908 #ifdef KTR
909 static const char *
910 intrtype_to_str(int intr_type)
911 {
912 	switch (intr_type) {
913 	case VMCB_EVENTINJ_TYPE_INTR:
914 		return ("hwintr");
915 	case VMCB_EVENTINJ_TYPE_NMI:
916 		return ("nmi");
917 	case VMCB_EVENTINJ_TYPE_INTn:
918 		return ("swintr");
919 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
920 		return ("exception");
921 	default:
922 		panic("%s: unknown intr_type %d", __func__, intr_type);
923 	}
924 }
925 #endif
926 
927 /*
928  * Inject an event to vcpu as described in section 15.20, "Event injection".
929  */
930 static void
931 svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
932 		 uint32_t error, bool ec_valid)
933 {
934 	struct vmcb_ctrl *ctrl;
935 
936 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
937 
938 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
939 	    ("%s: event already pending %#lx", __func__, ctrl->eventinj));
940 
941 	KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
942 	    __func__, vector));
943 
944 	switch (intr_type) {
945 	case VMCB_EVENTINJ_TYPE_INTR:
946 	case VMCB_EVENTINJ_TYPE_NMI:
947 	case VMCB_EVENTINJ_TYPE_INTn:
948 		break;
949 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
950 		if (vector >= 0 && vector <= 31 && vector != 2)
951 			break;
952 		/* FALLTHROUGH */
953 	default:
954 		panic("%s: invalid intr_type/vector: %d/%d", __func__,
955 		    intr_type, vector);
956 	}
957 	ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
958 	if (ec_valid) {
959 		ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
960 		ctrl->eventinj |= (uint64_t)error << 32;
961 		VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
962 		    intrtype_to_str(intr_type), vector, error);
963 	} else {
964 		VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
965 		    intrtype_to_str(intr_type), vector);
966 	}
967 }
968 
969 static void
970 svm_update_virqinfo(struct svm_softc *sc, int vcpu)
971 {
972 	struct vm *vm;
973 	struct vlapic *vlapic;
974 	struct vmcb_ctrl *ctrl;
975 
976 	vm = sc->vm;
977 	vlapic = vm_lapic(vm, vcpu);
978 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
979 
980 	/* Update %cr8 in the emulated vlapic */
981 	vlapic_set_cr8(vlapic, ctrl->v_tpr);
982 
983 	/* Virtual interrupt injection is not used. */
984 	KASSERT(ctrl->v_intr_vector == 0, ("%s: invalid "
985 	    "v_intr_vector %d", __func__, ctrl->v_intr_vector));
986 }
987 
988 static void
989 svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
990 {
991 	struct vmcb_ctrl *ctrl;
992 	uint64_t intinfo;
993 
994 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
995 	intinfo = ctrl->exitintinfo;
996 	if (!VMCB_EXITINTINFO_VALID(intinfo))
997 		return;
998 
999 	/*
1000 	 * From APMv2, Section "Intercepts during IDT interrupt delivery"
1001 	 *
1002 	 * If a #VMEXIT happened during event delivery then record the event
1003 	 * that was being delivered.
1004 	 */
1005 	VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
1006 		intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
1007 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
1008 	vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
1009 }
1010 
1011 #ifdef INVARIANTS
1012 static __inline int
1013 vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
1014 {
1015 
1016 	return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1017 	    VMCB_INTCPT_VINTR));
1018 }
1019 #endif
1020 
1021 static __inline void
1022 enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1023 {
1024 	struct vmcb_ctrl *ctrl;
1025 
1026 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1027 
1028 	if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
1029 		KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
1030 		KASSERT(vintr_intercept_enabled(sc, vcpu),
1031 		    ("%s: vintr intercept should be enabled", __func__));
1032 		return;
1033 	}
1034 
1035 	VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
1036 	ctrl->v_irq = 1;
1037 	ctrl->v_ign_tpr = 1;
1038 	ctrl->v_intr_vector = 0;
1039 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1040 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1041 }
1042 
1043 static __inline void
1044 disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
1045 {
1046 	struct vmcb_ctrl *ctrl;
1047 
1048 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1049 
1050 	if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
1051 		KASSERT(!vintr_intercept_enabled(sc, vcpu),
1052 		    ("%s: vintr intercept should be disabled", __func__));
1053 		return;
1054 	}
1055 
1056 	VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
1057 	ctrl->v_irq = 0;
1058 	ctrl->v_intr_vector = 0;
1059 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1060 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
1061 }
1062 
1063 static int
1064 svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
1065 {
1066 	struct vmcb_ctrl *ctrl;
1067 	int oldval, newval;
1068 
1069 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1070 	oldval = ctrl->intr_shadow;
1071 	newval = val ? 1 : 0;
1072 	if (newval != oldval) {
1073 		ctrl->intr_shadow = newval;
1074 		VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1075 	}
1076 	return (0);
1077 }
1078 
1079 static int
1080 svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1081 {
1082 	struct vmcb_ctrl *ctrl;
1083 
1084 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1085 	*val = ctrl->intr_shadow;
1086 	return (0);
1087 }
1088 
1089 /*
1090  * Once an NMI is injected it blocks delivery of further NMIs until the handler
1091  * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1092  * to track when the vcpu is done handling the NMI.
1093  */
1094 static int
1095 nmi_blocked(struct svm_softc *sc, int vcpu)
1096 {
1097 	int blocked;
1098 
1099 	blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1100 	    VMCB_INTCPT_IRET);
1101 	return (blocked);
1102 }
1103 
1104 static void
1105 enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1106 {
1107 
1108 	KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1109 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1110 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1111 }
1112 
1113 static void
1114 clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1115 {
1116 	int error __diagused;
1117 
1118 	KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1119 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1120 	/*
1121 	 * When the IRET intercept is cleared the vcpu will attempt to execute
1122 	 * the "iret" when it runs next. However, it is possible to inject
1123 	 * another NMI into the vcpu before the "iret" has actually executed.
1124 	 *
1125 	 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1126 	 * it will trap back into the hypervisor. If an NMI is pending for
1127 	 * the vcpu it will be injected into the guest.
1128 	 *
1129 	 * XXX this needs to be fixed
1130 	 */
1131 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1132 
1133 	/*
1134 	 * Set 'intr_shadow' to prevent an NMI from being injected on the
1135 	 * immediate VMRUN.
1136 	 */
1137 	error = svm_modify_intr_shadow(sc, vcpu, 1);
1138 	KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1139 }
1140 
1141 #define	EFER_MBZ_BITS	0xFFFFFFFFFFFF0200UL
1142 
1143 static int
1144 svm_write_efer(struct svm_softc *sc, int vcpu, uint64_t newval, bool *retu)
1145 {
1146 	struct vm_exit *vme;
1147 	struct vmcb_state *state;
1148 	uint64_t changed, lma, oldval;
1149 	int error __diagused;
1150 
1151 	state = svm_get_vmcb_state(sc, vcpu);
1152 
1153 	oldval = state->efer;
1154 	VCPU_CTR2(sc->vm, vcpu, "wrmsr(efer) %#lx/%#lx", oldval, newval);
1155 
1156 	newval &= ~0xFE;		/* clear the Read-As-Zero (RAZ) bits */
1157 	changed = oldval ^ newval;
1158 
1159 	if (newval & EFER_MBZ_BITS)
1160 		goto gpf;
1161 
1162 	/* APMv2 Table 14-5 "Long-Mode Consistency Checks" */
1163 	if (changed & EFER_LME) {
1164 		if (state->cr0 & CR0_PG)
1165 			goto gpf;
1166 	}
1167 
1168 	/* EFER.LMA = EFER.LME & CR0.PG */
1169 	if ((newval & EFER_LME) != 0 && (state->cr0 & CR0_PG) != 0)
1170 		lma = EFER_LMA;
1171 	else
1172 		lma = 0;
1173 
1174 	if ((newval & EFER_LMA) != lma)
1175 		goto gpf;
1176 
1177 	if (newval & EFER_NXE) {
1178 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_NO_EXECUTE))
1179 			goto gpf;
1180 	}
1181 
1182 	/*
1183 	 * XXX bhyve does not enforce segment limits in 64-bit mode. Until
1184 	 * this is fixed flag guest attempt to set EFER_LMSLE as an error.
1185 	 */
1186 	if (newval & EFER_LMSLE) {
1187 		vme = vm_exitinfo(sc->vm, vcpu);
1188 		vm_exit_svm(vme, VMCB_EXIT_MSR, 1, 0);
1189 		*retu = true;
1190 		return (0);
1191 	}
1192 
1193 	if (newval & EFER_FFXSR) {
1194 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_FFXSR))
1195 			goto gpf;
1196 	}
1197 
1198 	if (newval & EFER_TCE) {
1199 		if (!vm_cpuid_capability(sc->vm, vcpu, VCC_TCE))
1200 			goto gpf;
1201 	}
1202 
1203 	error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, newval);
1204 	KASSERT(error == 0, ("%s: error %d updating efer", __func__, error));
1205 	return (0);
1206 gpf:
1207 	vm_inject_gp(sc->vm, vcpu);
1208 	return (0);
1209 }
1210 
1211 static int
1212 emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1213     bool *retu)
1214 {
1215 	int error;
1216 
1217 	if (lapic_msr(num))
1218 		error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1219 	else if (num == MSR_EFER)
1220 		error = svm_write_efer(sc, vcpu, val, retu);
1221 	else
1222 		error = svm_wrmsr(sc, vcpu, num, val, retu);
1223 
1224 	return (error);
1225 }
1226 
1227 static int
1228 emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1229 {
1230 	struct vmcb_state *state;
1231 	struct svm_regctx *ctx;
1232 	uint64_t result;
1233 	int error;
1234 
1235 	if (lapic_msr(num))
1236 		error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1237 	else
1238 		error = svm_rdmsr(sc, vcpu, num, &result, retu);
1239 
1240 	if (error == 0) {
1241 		state = svm_get_vmcb_state(sc, vcpu);
1242 		ctx = svm_get_guest_regctx(sc, vcpu);
1243 		state->rax = result & 0xffffffff;
1244 		ctx->sctx_rdx = result >> 32;
1245 	}
1246 
1247 	return (error);
1248 }
1249 
1250 #ifdef KTR
1251 static const char *
1252 exit_reason_to_str(uint64_t reason)
1253 {
1254 	int i;
1255 	static char reasonbuf[32];
1256 	static const struct {
1257 		int reason;
1258 		const char *str;
1259 	} reasons[] = {
1260 		{ .reason = VMCB_EXIT_INVALID,	.str = "invalvmcb" },
1261 		{ .reason = VMCB_EXIT_SHUTDOWN,	.str = "shutdown" },
1262 		{ .reason = VMCB_EXIT_NPF, 	.str = "nptfault" },
1263 		{ .reason = VMCB_EXIT_PAUSE,	.str = "pause" },
1264 		{ .reason = VMCB_EXIT_HLT,	.str = "hlt" },
1265 		{ .reason = VMCB_EXIT_CPUID,	.str = "cpuid" },
1266 		{ .reason = VMCB_EXIT_IO,	.str = "inout" },
1267 		{ .reason = VMCB_EXIT_MC,	.str = "mchk" },
1268 		{ .reason = VMCB_EXIT_INTR,	.str = "extintr" },
1269 		{ .reason = VMCB_EXIT_NMI,	.str = "nmi" },
1270 		{ .reason = VMCB_EXIT_VINTR,	.str = "vintr" },
1271 		{ .reason = VMCB_EXIT_MSR,	.str = "msr" },
1272 		{ .reason = VMCB_EXIT_IRET,	.str = "iret" },
1273 		{ .reason = VMCB_EXIT_MONITOR,	.str = "monitor" },
1274 		{ .reason = VMCB_EXIT_MWAIT,	.str = "mwait" },
1275 		{ .reason = VMCB_EXIT_VMRUN,	.str = "vmrun" },
1276 		{ .reason = VMCB_EXIT_VMMCALL,	.str = "vmmcall" },
1277 		{ .reason = VMCB_EXIT_VMLOAD,	.str = "vmload" },
1278 		{ .reason = VMCB_EXIT_VMSAVE,	.str = "vmsave" },
1279 		{ .reason = VMCB_EXIT_STGI,	.str = "stgi" },
1280 		{ .reason = VMCB_EXIT_CLGI,	.str = "clgi" },
1281 		{ .reason = VMCB_EXIT_SKINIT,	.str = "skinit" },
1282 		{ .reason = VMCB_EXIT_ICEBP,	.str = "icebp" },
1283 		{ .reason = VMCB_EXIT_INVD,	.str = "invd" },
1284 		{ .reason = VMCB_EXIT_INVLPGA,	.str = "invlpga" },
1285 	};
1286 
1287 	for (i = 0; i < nitems(reasons); i++) {
1288 		if (reasons[i].reason == reason)
1289 			return (reasons[i].str);
1290 	}
1291 	snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1292 	return (reasonbuf);
1293 }
1294 #endif	/* KTR */
1295 
1296 /*
1297  * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1298  * that are due to instruction intercepts as well as MSR and IOIO intercepts
1299  * and exceptions caused by INT3, INTO and BOUND instructions.
1300  *
1301  * Return 1 if the nRIP is valid and 0 otherwise.
1302  */
1303 static int
1304 nrip_valid(uint64_t exitcode)
1305 {
1306 	switch (exitcode) {
1307 	case 0x00 ... 0x0F:	/* read of CR0 through CR15 */
1308 	case 0x10 ... 0x1F:	/* write of CR0 through CR15 */
1309 	case 0x20 ... 0x2F:	/* read of DR0 through DR15 */
1310 	case 0x30 ... 0x3F:	/* write of DR0 through DR15 */
1311 	case 0x43:		/* INT3 */
1312 	case 0x44:		/* INTO */
1313 	case 0x45:		/* BOUND */
1314 	case 0x65 ... 0x7C:	/* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1315 	case 0x80 ... 0x8D:	/* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1316 		return (1);
1317 	default:
1318 		return (0);
1319 	}
1320 }
1321 
1322 static int
1323 svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1324 {
1325 	struct vmcb *vmcb;
1326 	struct vmcb_state *state;
1327 	struct vmcb_ctrl *ctrl;
1328 	struct svm_regctx *ctx;
1329 	uint64_t code, info1, info2, val;
1330 	uint32_t eax, ecx, edx;
1331 	int error __diagused, errcode_valid, handled, idtvec, reflect;
1332 	bool retu;
1333 
1334 	ctx = svm_get_guest_regctx(svm_sc, vcpu);
1335 	vmcb = svm_get_vmcb(svm_sc, vcpu);
1336 	state = &vmcb->state;
1337 	ctrl = &vmcb->ctrl;
1338 
1339 	handled = 0;
1340 	code = ctrl->exitcode;
1341 	info1 = ctrl->exitinfo1;
1342 	info2 = ctrl->exitinfo2;
1343 
1344 	vmexit->exitcode = VM_EXITCODE_BOGUS;
1345 	vmexit->rip = state->rip;
1346 	vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1347 
1348 	vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1349 
1350 	/*
1351 	 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1352 	 * in an inconsistent state and can trigger assertions that would
1353 	 * never happen otherwise.
1354 	 */
1355 	if (code == VMCB_EXIT_INVALID) {
1356 		vm_exit_svm(vmexit, code, info1, info2);
1357 		return (0);
1358 	}
1359 
1360 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1361 	    "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1362 
1363 	KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1364 	    ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1365 	    vmexit->inst_length, code, info1, info2));
1366 
1367 	svm_update_virqinfo(svm_sc, vcpu);
1368 	svm_save_intinfo(svm_sc, vcpu);
1369 
1370 	switch (code) {
1371 	case VMCB_EXIT_IRET:
1372 		/*
1373 		 * Restart execution at "iret" but with the intercept cleared.
1374 		 */
1375 		vmexit->inst_length = 0;
1376 		clear_nmi_blocking(svm_sc, vcpu);
1377 		handled = 1;
1378 		break;
1379 	case VMCB_EXIT_VINTR:	/* interrupt window exiting */
1380 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1381 		handled = 1;
1382 		break;
1383 	case VMCB_EXIT_INTR:	/* external interrupt */
1384 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1385 		handled = 1;
1386 		break;
1387 	case VMCB_EXIT_NMI:	/* external NMI */
1388 		handled = 1;
1389 		break;
1390 	case 0x40 ... 0x5F:
1391 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1392 		reflect = 1;
1393 		idtvec = code - 0x40;
1394 		switch (idtvec) {
1395 		case IDT_MC:
1396 			/*
1397 			 * Call the machine check handler by hand. Also don't
1398 			 * reflect the machine check back into the guest.
1399 			 */
1400 			reflect = 0;
1401 			VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler");
1402 			__asm __volatile("int $18");
1403 			break;
1404 		case IDT_PF:
1405 			error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2,
1406 			    info2);
1407 			KASSERT(error == 0, ("%s: error %d updating cr2",
1408 			    __func__, error));
1409 			/* fallthru */
1410 		case IDT_NP:
1411 		case IDT_SS:
1412 		case IDT_GP:
1413 		case IDT_AC:
1414 		case IDT_TS:
1415 			errcode_valid = 1;
1416 			break;
1417 
1418 		case IDT_DF:
1419 			errcode_valid = 1;
1420 			info1 = 0;
1421 			break;
1422 
1423 		case IDT_BP:
1424 		case IDT_OF:
1425 		case IDT_BR:
1426 			/*
1427 			 * The 'nrip' field is populated for INT3, INTO and
1428 			 * BOUND exceptions and this also implies that
1429 			 * 'inst_length' is non-zero.
1430 			 *
1431 			 * Reset 'inst_length' to zero so the guest %rip at
1432 			 * event injection is identical to what it was when
1433 			 * the exception originally happened.
1434 			 */
1435 			VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d "
1436 			    "to zero before injecting exception %d",
1437 			    vmexit->inst_length, idtvec);
1438 			vmexit->inst_length = 0;
1439 			/* fallthru */
1440 		default:
1441 			errcode_valid = 0;
1442 			info1 = 0;
1443 			break;
1444 		}
1445 		KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1446 		    "when reflecting exception %d into guest",
1447 		    vmexit->inst_length, idtvec));
1448 
1449 		if (reflect) {
1450 			/* Reflect the exception back into the guest */
1451 			VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception "
1452 			    "%d/%#x into the guest", idtvec, (int)info1);
1453 			error = vm_inject_exception(svm_sc->vm, vcpu, idtvec,
1454 			    errcode_valid, info1, 0);
1455 			KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1456 			    __func__, error));
1457 		}
1458 		handled = 1;
1459 		break;
1460 	case VMCB_EXIT_MSR:	/* MSR access. */
1461 		eax = state->rax;
1462 		ecx = ctx->sctx_rcx;
1463 		edx = ctx->sctx_rdx;
1464 		retu = false;
1465 
1466 		if (info1) {
1467 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1468 			val = (uint64_t)edx << 32 | eax;
1469 			VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1470 			    ecx, val);
1471 			if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1472 				vmexit->exitcode = VM_EXITCODE_WRMSR;
1473 				vmexit->u.msr.code = ecx;
1474 				vmexit->u.msr.wval = val;
1475 			} else if (!retu) {
1476 				handled = 1;
1477 			} else {
1478 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1479 				    ("emulate_wrmsr retu with bogus exitcode"));
1480 			}
1481 		} else {
1482 			VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1483 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1484 			if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1485 				vmexit->exitcode = VM_EXITCODE_RDMSR;
1486 				vmexit->u.msr.code = ecx;
1487 			} else if (!retu) {
1488 				handled = 1;
1489 			} else {
1490 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1491 				    ("emulate_rdmsr retu with bogus exitcode"));
1492 			}
1493 		}
1494 		break;
1495 	case VMCB_EXIT_IO:
1496 		handled = svm_handle_io(svm_sc, vcpu, vmexit);
1497 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1498 		break;
1499 	case VMCB_EXIT_CPUID:
1500 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1501 		handled = x86_emulate_cpuid(svm_sc->vm, vcpu, &state->rax,
1502 		    &ctx->sctx_rbx, &ctx->sctx_rcx, &ctx->sctx_rdx);
1503 		break;
1504 	case VMCB_EXIT_HLT:
1505 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1506 		vmexit->exitcode = VM_EXITCODE_HLT;
1507 		vmexit->u.hlt.rflags = state->rflags;
1508 		break;
1509 	case VMCB_EXIT_PAUSE:
1510 		vmexit->exitcode = VM_EXITCODE_PAUSE;
1511 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1512 		break;
1513 	case VMCB_EXIT_NPF:
1514 		/* EXITINFO2 contains the faulting guest physical address */
1515 		if (info1 & VMCB_NPF_INFO1_RSV) {
1516 			VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1517 			    "reserved bits set: info1(%#lx) info2(%#lx)",
1518 			    info1, info2);
1519 		} else if (vm_mem_allocated(svm_sc->vm, vcpu, info2)) {
1520 			vmexit->exitcode = VM_EXITCODE_PAGING;
1521 			vmexit->u.paging.gpa = info2;
1522 			vmexit->u.paging.fault_type = npf_fault_type(info1);
1523 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1524 			VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1525 			    "on gpa %#lx/%#lx at rip %#lx",
1526 			    info2, info1, state->rip);
1527 		} else if (svm_npf_emul_fault(info1)) {
1528 			svm_handle_inst_emul(vmcb, info2, vmexit);
1529 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1530 			VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1531 			    "for gpa %#lx/%#lx at rip %#lx",
1532 			    info2, info1, state->rip);
1533 		}
1534 		break;
1535 	case VMCB_EXIT_MONITOR:
1536 		vmexit->exitcode = VM_EXITCODE_MONITOR;
1537 		break;
1538 	case VMCB_EXIT_MWAIT:
1539 		vmexit->exitcode = VM_EXITCODE_MWAIT;
1540 		break;
1541 	case VMCB_EXIT_SHUTDOWN:
1542 	case VMCB_EXIT_VMRUN:
1543 	case VMCB_EXIT_VMMCALL:
1544 	case VMCB_EXIT_VMLOAD:
1545 	case VMCB_EXIT_VMSAVE:
1546 	case VMCB_EXIT_STGI:
1547 	case VMCB_EXIT_CLGI:
1548 	case VMCB_EXIT_SKINIT:
1549 	case VMCB_EXIT_ICEBP:
1550 	case VMCB_EXIT_INVD:
1551 	case VMCB_EXIT_INVLPGA:
1552 		vm_inject_ud(svm_sc->vm, vcpu);
1553 		handled = 1;
1554 		break;
1555 	default:
1556 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1557 		break;
1558 	}
1559 
1560 	VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1561 	    handled ? "handled" : "unhandled", exit_reason_to_str(code),
1562 	    vmexit->rip, vmexit->inst_length);
1563 
1564 	if (handled) {
1565 		vmexit->rip += vmexit->inst_length;
1566 		vmexit->inst_length = 0;
1567 		state->rip = vmexit->rip;
1568 	} else {
1569 		if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1570 			/*
1571 			 * If this VM exit was not claimed by anybody then
1572 			 * treat it as a generic SVM exit.
1573 			 */
1574 			vm_exit_svm(vmexit, code, info1, info2);
1575 		} else {
1576 			/*
1577 			 * The exitcode and collateral have been populated.
1578 			 * The VM exit will be processed further in userland.
1579 			 */
1580 		}
1581 	}
1582 	return (handled);
1583 }
1584 
1585 static void
1586 svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1587 {
1588 	uint64_t intinfo;
1589 
1590 	if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1591 		return;
1592 
1593 	KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1594 	    "valid: %#lx", __func__, intinfo));
1595 
1596 	svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1597 		VMCB_EXITINTINFO_VECTOR(intinfo),
1598 		VMCB_EXITINTINFO_EC(intinfo),
1599 		VMCB_EXITINTINFO_EC_VALID(intinfo));
1600 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1601 	VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1602 }
1603 
1604 /*
1605  * Inject event to virtual cpu.
1606  */
1607 static void
1608 svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1609 {
1610 	struct vmcb_ctrl *ctrl;
1611 	struct vmcb_state *state;
1612 	struct svm_vcpu *vcpustate;
1613 	uint8_t v_tpr;
1614 	int vector, need_intr_window;
1615 	int extint_pending;
1616 
1617 	state = svm_get_vmcb_state(sc, vcpu);
1618 	ctrl  = svm_get_vmcb_ctrl(sc, vcpu);
1619 	vcpustate = svm_get_vcpu(sc, vcpu);
1620 
1621 	need_intr_window = 0;
1622 
1623 	if (vcpustate->nextrip != state->rip) {
1624 		ctrl->intr_shadow = 0;
1625 		VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking "
1626 		    "cleared due to rip change: %#lx/%#lx",
1627 		    vcpustate->nextrip, state->rip);
1628 	}
1629 
1630 	/*
1631 	 * Inject pending events or exceptions for this vcpu.
1632 	 *
1633 	 * An event might be pending because the previous #VMEXIT happened
1634 	 * during event delivery (i.e. ctrl->exitintinfo).
1635 	 *
1636 	 * An event might also be pending because an exception was injected
1637 	 * by the hypervisor (e.g. #PF during instruction emulation).
1638 	 */
1639 	svm_inj_intinfo(sc, vcpu);
1640 
1641 	/* NMI event has priority over interrupts. */
1642 	if (vm_nmi_pending(sc->vm, vcpu)) {
1643 		if (nmi_blocked(sc, vcpu)) {
1644 			/*
1645 			 * Can't inject another NMI if the guest has not
1646 			 * yet executed an "iret" after the last NMI.
1647 			 */
1648 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1649 			    "to NMI-blocking");
1650 		} else if (ctrl->intr_shadow) {
1651 			/*
1652 			 * Can't inject an NMI if the vcpu is in an intr_shadow.
1653 			 */
1654 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1655 			    "interrupt shadow");
1656 			need_intr_window = 1;
1657 			goto done;
1658 		} else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1659 			/*
1660 			 * If there is already an exception/interrupt pending
1661 			 * then defer the NMI until after that.
1662 			 */
1663 			VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1664 			    "eventinj %#lx", ctrl->eventinj);
1665 
1666 			/*
1667 			 * Use self-IPI to trigger a VM-exit as soon as
1668 			 * possible after the event injection is completed.
1669 			 *
1670 			 * This works only if the external interrupt exiting
1671 			 * is at a lower priority than the event injection.
1672 			 *
1673 			 * Although not explicitly specified in APMv2 the
1674 			 * relative priorities were verified empirically.
1675 			 */
1676 			ipi_cpu(curcpu, IPI_AST);	/* XXX vmm_ipinum? */
1677 		} else {
1678 			vm_nmi_clear(sc->vm, vcpu);
1679 
1680 			/* Inject NMI, vector number is not used */
1681 			svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1682 			    IDT_NMI, 0, false);
1683 
1684 			/* virtual NMI blocking is now in effect */
1685 			enable_nmi_blocking(sc, vcpu);
1686 
1687 			VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1688 		}
1689 	}
1690 
1691 	extint_pending = vm_extint_pending(sc->vm, vcpu);
1692 	if (!extint_pending) {
1693 		if (!vlapic_pending_intr(vlapic, &vector))
1694 			goto done;
1695 		KASSERT(vector >= 16 && vector <= 255,
1696 		    ("invalid vector %d from local APIC", vector));
1697 	} else {
1698 		/* Ask the legacy pic for a vector to inject */
1699 		vatpic_pending_intr(sc->vm, &vector);
1700 		KASSERT(vector >= 0 && vector <= 255,
1701 		    ("invalid vector %d from INTR", vector));
1702 	}
1703 
1704 	/*
1705 	 * If the guest has disabled interrupts or is in an interrupt shadow
1706 	 * then we cannot inject the pending interrupt.
1707 	 */
1708 	if ((state->rflags & PSL_I) == 0) {
1709 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1710 		    "rflags %#lx", vector, state->rflags);
1711 		need_intr_window = 1;
1712 		goto done;
1713 	}
1714 
1715 	if (ctrl->intr_shadow) {
1716 		VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1717 		    "interrupt shadow", vector);
1718 		need_intr_window = 1;
1719 		goto done;
1720 	}
1721 
1722 	if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1723 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1724 		    "eventinj %#lx", vector, ctrl->eventinj);
1725 		need_intr_window = 1;
1726 		goto done;
1727 	}
1728 
1729 	svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1730 
1731 	if (!extint_pending) {
1732 		vlapic_intr_accepted(vlapic, vector);
1733 	} else {
1734 		vm_extint_clear(sc->vm, vcpu);
1735 		vatpic_intr_accepted(sc->vm, vector);
1736 	}
1737 
1738 	/*
1739 	 * Force a VM-exit as soon as the vcpu is ready to accept another
1740 	 * interrupt. This is done because the PIC might have another vector
1741 	 * that it wants to inject. Also, if the APIC has a pending interrupt
1742 	 * that was preempted by the ExtInt then it allows us to inject the
1743 	 * APIC vector as soon as possible.
1744 	 */
1745 	need_intr_window = 1;
1746 done:
1747 	/*
1748 	 * The guest can modify the TPR by writing to %CR8. In guest mode
1749 	 * the processor reflects this write to V_TPR without hypervisor
1750 	 * intervention.
1751 	 *
1752 	 * The guest can also modify the TPR by writing to it via the memory
1753 	 * mapped APIC page. In this case, the write will be emulated by the
1754 	 * hypervisor. For this reason V_TPR must be updated before every
1755 	 * VMRUN.
1756 	 */
1757 	v_tpr = vlapic_get_cr8(vlapic);
1758 	KASSERT(v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1759 	if (ctrl->v_tpr != v_tpr) {
1760 		VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1761 		    ctrl->v_tpr, v_tpr);
1762 		ctrl->v_tpr = v_tpr;
1763 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1764 	}
1765 
1766 	if (need_intr_window) {
1767 		/*
1768 		 * We use V_IRQ in conjunction with the VINTR intercept to
1769 		 * trap into the hypervisor as soon as a virtual interrupt
1770 		 * can be delivered.
1771 		 *
1772 		 * Since injected events are not subject to intercept checks
1773 		 * we need to ensure that the V_IRQ is not actually going to
1774 		 * be delivered on VM entry. The KASSERT below enforces this.
1775 		 */
1776 		KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1777 		    (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1778 		    ("Bogus intr_window_exiting: eventinj (%#lx), "
1779 		    "intr_shadow (%u), rflags (%#lx)",
1780 		    ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1781 		enable_intr_window_exiting(sc, vcpu);
1782 	} else {
1783 		disable_intr_window_exiting(sc, vcpu);
1784 	}
1785 }
1786 
1787 static __inline void
1788 restore_host_tss(void)
1789 {
1790 	struct system_segment_descriptor *tss_sd;
1791 
1792 	/*
1793 	 * The TSS descriptor was in use prior to launching the guest so it
1794 	 * has been marked busy.
1795 	 *
1796 	 * 'ltr' requires the descriptor to be marked available so change the
1797 	 * type to "64-bit available TSS".
1798 	 */
1799 	tss_sd = PCPU_GET(tss);
1800 	tss_sd->sd_type = SDT_SYSTSS;
1801 	ltr(GSEL(GPROC0_SEL, SEL_KPL));
1802 }
1803 
1804 static void
1805 svm_pmap_activate(struct svm_softc *sc, int vcpuid, pmap_t pmap)
1806 {
1807 	struct svm_vcpu *vcpustate;
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 	vcpustate = svm_get_vcpu(sc, vcpuid);
1818 	ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1819 
1820 	/*
1821 	 * The TLB entries associated with the vcpu's ASID are not valid
1822 	 * if either of the following conditions is true:
1823 	 *
1824 	 * 1. The vcpu's ASID generation is different than the host cpu's
1825 	 *    ASID generation. This happens when the vcpu migrates to a new
1826 	 *    host cpu. It can also happen when the number of vcpus executing
1827 	 *    on a host cpu is greater than the number of ASIDs available.
1828 	 *
1829 	 * 2. The pmap generation number is different than the value cached in
1830 	 *    the 'vcpustate'. This happens when the host invalidates pages
1831 	 *    belonging to the guest.
1832 	 *
1833 	 *	asidgen		eptgen	      Action
1834 	 *	mismatch	mismatch
1835 	 *	   0		   0		(a)
1836 	 *	   0		   1		(b1) or (b2)
1837 	 *	   1		   0		(c)
1838 	 *	   1		   1		(d)
1839 	 *
1840 	 * (a) There is no mismatch in eptgen or ASID generation and therefore
1841 	 *     no further action is needed.
1842 	 *
1843 	 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1844 	 *      retained and the TLB entries associated with this ASID
1845 	 *      are flushed by VMRUN.
1846 	 *
1847 	 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1848 	 *      allocated.
1849 	 *
1850 	 * (c) A new ASID is allocated.
1851 	 *
1852 	 * (d) A new ASID is allocated.
1853 	 */
1854 
1855 	alloc_asid = false;
1856 	eptgen = atomic_load_long(&pmap->pm_eptgen);
1857 	ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1858 
1859 	if (vcpustate->asid.gen != asid[cpu].gen) {
1860 		alloc_asid = true;	/* (c) and (d) */
1861 	} else if (vcpustate->eptgen != eptgen) {
1862 		if (flush_by_asid())
1863 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;	/* (b1) */
1864 		else
1865 			alloc_asid = true;			/* (b2) */
1866 	} else {
1867 		/*
1868 		 * This is the common case (a).
1869 		 */
1870 		KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1871 		KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1872 		    ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1873 	}
1874 
1875 	if (alloc_asid) {
1876 		if (++asid[cpu].num >= nasid) {
1877 			asid[cpu].num = 1;
1878 			if (++asid[cpu].gen == 0)
1879 				asid[cpu].gen = 1;
1880 			/*
1881 			 * If this cpu does not support "flush-by-asid"
1882 			 * then flush the entire TLB on a generation
1883 			 * bump. Subsequent ASID allocation in this
1884 			 * generation can be done without a TLB flush.
1885 			 */
1886 			if (!flush_by_asid())
1887 				ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1888 		}
1889 		vcpustate->asid.gen = asid[cpu].gen;
1890 		vcpustate->asid.num = asid[cpu].num;
1891 
1892 		ctrl->asid = vcpustate->asid.num;
1893 		svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1894 		/*
1895 		 * If this cpu supports "flush-by-asid" then the TLB
1896 		 * was not flushed after the generation bump. The TLB
1897 		 * is flushed selectively after every new ASID allocation.
1898 		 */
1899 		if (flush_by_asid())
1900 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1901 	}
1902 	vcpustate->eptgen = eptgen;
1903 
1904 	KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1905 	KASSERT(ctrl->asid == vcpustate->asid.num,
1906 	    ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1907 }
1908 
1909 static void
1910 svm_pmap_deactivate(pmap_t pmap)
1911 {
1912 	smr_exit(pmap->pm_eptsmr);
1913 	CPU_CLR_ATOMIC(curcpu, &pmap->pm_active);
1914 }
1915 
1916 static __inline void
1917 disable_gintr(void)
1918 {
1919 
1920 	__asm __volatile("clgi");
1921 }
1922 
1923 static __inline void
1924 enable_gintr(void)
1925 {
1926 
1927         __asm __volatile("stgi");
1928 }
1929 
1930 static __inline void
1931 svm_dr_enter_guest(struct svm_regctx *gctx)
1932 {
1933 
1934 	/* Save host control debug registers. */
1935 	gctx->host_dr7 = rdr7();
1936 	gctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR);
1937 
1938 	/*
1939 	 * Disable debugging in DR7 and DEBUGCTL to avoid triggering
1940 	 * exceptions in the host based on the guest DRx values.  The
1941 	 * guest DR6, DR7, and DEBUGCTL are saved/restored in the
1942 	 * VMCB.
1943 	 */
1944 	load_dr7(0);
1945 	wrmsr(MSR_DEBUGCTLMSR, 0);
1946 
1947 	/* Save host debug registers. */
1948 	gctx->host_dr0 = rdr0();
1949 	gctx->host_dr1 = rdr1();
1950 	gctx->host_dr2 = rdr2();
1951 	gctx->host_dr3 = rdr3();
1952 	gctx->host_dr6 = rdr6();
1953 
1954 	/* Restore guest debug registers. */
1955 	load_dr0(gctx->sctx_dr0);
1956 	load_dr1(gctx->sctx_dr1);
1957 	load_dr2(gctx->sctx_dr2);
1958 	load_dr3(gctx->sctx_dr3);
1959 }
1960 
1961 static __inline void
1962 svm_dr_leave_guest(struct svm_regctx *gctx)
1963 {
1964 
1965 	/* Save guest debug registers. */
1966 	gctx->sctx_dr0 = rdr0();
1967 	gctx->sctx_dr1 = rdr1();
1968 	gctx->sctx_dr2 = rdr2();
1969 	gctx->sctx_dr3 = rdr3();
1970 
1971 	/*
1972 	 * Restore host debug registers.  Restore DR7 and DEBUGCTL
1973 	 * last.
1974 	 */
1975 	load_dr0(gctx->host_dr0);
1976 	load_dr1(gctx->host_dr1);
1977 	load_dr2(gctx->host_dr2);
1978 	load_dr3(gctx->host_dr3);
1979 	load_dr6(gctx->host_dr6);
1980 	wrmsr(MSR_DEBUGCTLMSR, gctx->host_debugctl);
1981 	load_dr7(gctx->host_dr7);
1982 }
1983 
1984 /*
1985  * Start vcpu with specified RIP.
1986  */
1987 static int
1988 svm_run(void *arg, int vcpu, register_t rip, pmap_t pmap,
1989 	struct vm_eventinfo *evinfo)
1990 {
1991 	struct svm_regctx *gctx;
1992 	struct svm_softc *svm_sc;
1993 	struct svm_vcpu *vcpustate;
1994 	struct vmcb_state *state;
1995 	struct vmcb_ctrl *ctrl;
1996 	struct vm_exit *vmexit;
1997 	struct vlapic *vlapic;
1998 	struct vm *vm;
1999 	uint64_t vmcb_pa;
2000 	int handled;
2001 	uint16_t ldt_sel;
2002 
2003 	svm_sc = arg;
2004 	vm = svm_sc->vm;
2005 
2006 	vcpustate = svm_get_vcpu(svm_sc, vcpu);
2007 	state = svm_get_vmcb_state(svm_sc, vcpu);
2008 	ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
2009 	vmexit = vm_exitinfo(vm, vcpu);
2010 	vlapic = vm_lapic(vm, vcpu);
2011 
2012 	gctx = svm_get_guest_regctx(svm_sc, vcpu);
2013 	vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
2014 
2015 	if (vcpustate->lastcpu != curcpu) {
2016 		/*
2017 		 * Force new ASID allocation by invalidating the generation.
2018 		 */
2019 		vcpustate->asid.gen = 0;
2020 
2021 		/*
2022 		 * Invalidate the VMCB state cache by marking all fields dirty.
2023 		 */
2024 		svm_set_dirty(svm_sc, vcpu, 0xffffffff);
2025 
2026 		/*
2027 		 * XXX
2028 		 * Setting 'vcpustate->lastcpu' here is bit premature because
2029 		 * we may return from this function without actually executing
2030 		 * the VMRUN  instruction. This could happen if a rendezvous
2031 		 * or an AST is pending on the first time through the loop.
2032 		 *
2033 		 * This works for now but any new side-effects of vcpu
2034 		 * migration should take this case into account.
2035 		 */
2036 		vcpustate->lastcpu = curcpu;
2037 		vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
2038 	}
2039 
2040 	svm_msr_guest_enter(svm_sc, vcpu);
2041 
2042 	/* Update Guest RIP */
2043 	state->rip = rip;
2044 
2045 	do {
2046 		/*
2047 		 * Disable global interrupts to guarantee atomicity during
2048 		 * loading of guest state. This includes not only the state
2049 		 * loaded by the "vmrun" instruction but also software state
2050 		 * maintained by the hypervisor: suspended and rendezvous
2051 		 * state, NPT generation number, vlapic interrupts etc.
2052 		 */
2053 		disable_gintr();
2054 
2055 		if (vcpu_suspended(evinfo)) {
2056 			enable_gintr();
2057 			vm_exit_suspended(vm, vcpu, state->rip);
2058 			break;
2059 		}
2060 
2061 		if (vcpu_rendezvous_pending(evinfo)) {
2062 			enable_gintr();
2063 			vm_exit_rendezvous(vm, vcpu, state->rip);
2064 			break;
2065 		}
2066 
2067 		if (vcpu_reqidle(evinfo)) {
2068 			enable_gintr();
2069 			vm_exit_reqidle(vm, vcpu, state->rip);
2070 			break;
2071 		}
2072 
2073 		/* We are asked to give the cpu by scheduler. */
2074 		if (vcpu_should_yield(vm, vcpu)) {
2075 			enable_gintr();
2076 			vm_exit_astpending(vm, vcpu, state->rip);
2077 			break;
2078 		}
2079 
2080 		if (vcpu_debugged(vm, vcpu)) {
2081 			enable_gintr();
2082 			vm_exit_debug(vm, vcpu, state->rip);
2083 			break;
2084 		}
2085 
2086 		/*
2087 		 * #VMEXIT resumes the host with the guest LDTR, so
2088 		 * save the current LDT selector so it can be restored
2089 		 * after an exit.  The userspace hypervisor probably
2090 		 * doesn't use a LDT, but save and restore it to be
2091 		 * safe.
2092 		 */
2093 		ldt_sel = sldt();
2094 
2095 		svm_inj_interrupts(svm_sc, vcpu, vlapic);
2096 
2097 		/*
2098 		 * Check the pmap generation and the ASID generation to
2099 		 * ensure that the vcpu does not use stale TLB mappings.
2100 		 */
2101 		svm_pmap_activate(svm_sc, vcpu, pmap);
2102 
2103 		ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
2104 		vcpustate->dirty = 0;
2105 		VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
2106 
2107 		/* Launch Virtual Machine. */
2108 		VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
2109 		svm_dr_enter_guest(gctx);
2110 		svm_launch(vmcb_pa, gctx, get_pcpu());
2111 		svm_dr_leave_guest(gctx);
2112 
2113 		svm_pmap_deactivate(pmap);
2114 
2115 		/*
2116 		 * The host GDTR and IDTR is saved by VMRUN and restored
2117 		 * automatically on #VMEXIT. However, the host TSS needs
2118 		 * to be restored explicitly.
2119 		 */
2120 		restore_host_tss();
2121 
2122 		/* Restore host LDTR. */
2123 		lldt(ldt_sel);
2124 
2125 		/* #VMEXIT disables interrupts so re-enable them here. */
2126 		enable_gintr();
2127 
2128 		/* Update 'nextrip' */
2129 		vcpustate->nextrip = state->rip;
2130 
2131 		/* Handle #VMEXIT and if required return to user space. */
2132 		handled = svm_vmexit(svm_sc, vcpu, vmexit);
2133 	} while (handled);
2134 
2135 	svm_msr_guest_exit(svm_sc, vcpu);
2136 
2137 	return (0);
2138 }
2139 
2140 static void
2141 svm_cleanup(void *arg)
2142 {
2143 	struct svm_softc *sc = arg;
2144 
2145 	contigfree(sc->iopm_bitmap, SVM_IO_BITMAP_SIZE, M_SVM);
2146 	contigfree(sc->msr_bitmap, SVM_MSR_BITMAP_SIZE, M_SVM);
2147 	free(sc, M_SVM);
2148 }
2149 
2150 static register_t *
2151 swctx_regptr(struct svm_regctx *regctx, int reg)
2152 {
2153 
2154 	switch (reg) {
2155 	case VM_REG_GUEST_RBX:
2156 		return (&regctx->sctx_rbx);
2157 	case VM_REG_GUEST_RCX:
2158 		return (&regctx->sctx_rcx);
2159 	case VM_REG_GUEST_RDX:
2160 		return (&regctx->sctx_rdx);
2161 	case VM_REG_GUEST_RDI:
2162 		return (&regctx->sctx_rdi);
2163 	case VM_REG_GUEST_RSI:
2164 		return (&regctx->sctx_rsi);
2165 	case VM_REG_GUEST_RBP:
2166 		return (&regctx->sctx_rbp);
2167 	case VM_REG_GUEST_R8:
2168 		return (&regctx->sctx_r8);
2169 	case VM_REG_GUEST_R9:
2170 		return (&regctx->sctx_r9);
2171 	case VM_REG_GUEST_R10:
2172 		return (&regctx->sctx_r10);
2173 	case VM_REG_GUEST_R11:
2174 		return (&regctx->sctx_r11);
2175 	case VM_REG_GUEST_R12:
2176 		return (&regctx->sctx_r12);
2177 	case VM_REG_GUEST_R13:
2178 		return (&regctx->sctx_r13);
2179 	case VM_REG_GUEST_R14:
2180 		return (&regctx->sctx_r14);
2181 	case VM_REG_GUEST_R15:
2182 		return (&regctx->sctx_r15);
2183 	case VM_REG_GUEST_DR0:
2184 		return (&regctx->sctx_dr0);
2185 	case VM_REG_GUEST_DR1:
2186 		return (&regctx->sctx_dr1);
2187 	case VM_REG_GUEST_DR2:
2188 		return (&regctx->sctx_dr2);
2189 	case VM_REG_GUEST_DR3:
2190 		return (&regctx->sctx_dr3);
2191 	default:
2192 		return (NULL);
2193 	}
2194 }
2195 
2196 static int
2197 svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
2198 {
2199 	struct svm_softc *svm_sc;
2200 	register_t *reg;
2201 
2202 	svm_sc = arg;
2203 
2204 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2205 		return (svm_get_intr_shadow(svm_sc, vcpu, val));
2206 	}
2207 
2208 	if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2209 		return (0);
2210 	}
2211 
2212 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2213 
2214 	if (reg != NULL) {
2215 		*val = *reg;
2216 		return (0);
2217 	}
2218 
2219 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident);
2220 	return (EINVAL);
2221 }
2222 
2223 static int
2224 svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2225 {
2226 	struct svm_softc *svm_sc;
2227 	register_t *reg;
2228 
2229 	svm_sc = arg;
2230 
2231 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2232 		return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2233 	}
2234 
2235 	/* Do not permit user write access to VMCB fields by offset. */
2236 	if (!VMCB_ACCESS_OK(ident)) {
2237 		if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2238 			return (0);
2239 		}
2240 	}
2241 
2242 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2243 
2244 	if (reg != NULL) {
2245 		*reg = val;
2246 		return (0);
2247 	}
2248 
2249 	if (ident == VM_REG_GUEST_ENTRY_INST_LENGTH) {
2250 		/* Ignore. */
2251 		return (0);
2252 	}
2253 
2254 	/*
2255 	 * XXX deal with CR3 and invalidate TLB entries tagged with the
2256 	 * vcpu's ASID. This needs to be treated differently depending on
2257 	 * whether 'running' is true/false.
2258 	 */
2259 
2260 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident);
2261 	return (EINVAL);
2262 }
2263 
2264 static int
2265 svm_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc)
2266 {
2267 	return (vmcb_getdesc(arg, vcpu, reg, desc));
2268 }
2269 
2270 static int
2271 svm_setdesc(void *arg, int vcpu, int reg, struct seg_desc *desc)
2272 {
2273 	return (vmcb_setdesc(arg, vcpu, reg, desc));
2274 }
2275 
2276 #ifdef BHYVE_SNAPSHOT
2277 static int
2278 svm_snapshot_reg(void *arg, int vcpu, int ident,
2279 		 struct vm_snapshot_meta *meta)
2280 {
2281 	int ret;
2282 	uint64_t val;
2283 
2284 	if (meta->op == VM_SNAPSHOT_SAVE) {
2285 		ret = svm_getreg(arg, vcpu, ident, &val);
2286 		if (ret != 0)
2287 			goto done;
2288 
2289 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2290 	} else if (meta->op == VM_SNAPSHOT_RESTORE) {
2291 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2292 
2293 		ret = svm_setreg(arg, vcpu, ident, val);
2294 		if (ret != 0)
2295 			goto done;
2296 	} else {
2297 		ret = EINVAL;
2298 		goto done;
2299 	}
2300 
2301 done:
2302 	return (ret);
2303 }
2304 #endif
2305 
2306 static int
2307 svm_setcap(void *arg, int vcpu, int type, int val)
2308 {
2309 	struct svm_softc *sc;
2310 	int error;
2311 
2312 	sc = arg;
2313 	error = 0;
2314 	switch (type) {
2315 	case VM_CAP_HALT_EXIT:
2316 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2317 		    VMCB_INTCPT_HLT, val);
2318 		break;
2319 	case VM_CAP_PAUSE_EXIT:
2320 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2321 		    VMCB_INTCPT_PAUSE, val);
2322 		break;
2323 	case VM_CAP_UNRESTRICTED_GUEST:
2324 		/* Unrestricted guest execution cannot be disabled in SVM */
2325 		if (val == 0)
2326 			error = EINVAL;
2327 		break;
2328 	default:
2329 		error = ENOENT;
2330 		break;
2331 	}
2332 	return (error);
2333 }
2334 
2335 static int
2336 svm_getcap(void *arg, int vcpu, int type, int *retval)
2337 {
2338 	struct svm_softc *sc;
2339 	int error;
2340 
2341 	sc = arg;
2342 	error = 0;
2343 
2344 	switch (type) {
2345 	case VM_CAP_HALT_EXIT:
2346 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2347 		    VMCB_INTCPT_HLT);
2348 		break;
2349 	case VM_CAP_PAUSE_EXIT:
2350 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2351 		    VMCB_INTCPT_PAUSE);
2352 		break;
2353 	case VM_CAP_UNRESTRICTED_GUEST:
2354 		*retval = 1;	/* unrestricted guest is always enabled */
2355 		break;
2356 	default:
2357 		error = ENOENT;
2358 		break;
2359 	}
2360 	return (error);
2361 }
2362 
2363 static struct vmspace *
2364 svm_vmspace_alloc(vm_offset_t min, vm_offset_t max)
2365 {
2366 	return (svm_npt_alloc(min, max));
2367 }
2368 
2369 static void
2370 svm_vmspace_free(struct vmspace *vmspace)
2371 {
2372 	svm_npt_free(vmspace);
2373 }
2374 
2375 static struct vlapic *
2376 svm_vlapic_init(void *arg, int vcpuid)
2377 {
2378 	struct svm_softc *svm_sc;
2379 	struct vlapic *vlapic;
2380 
2381 	svm_sc = arg;
2382 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2383 	vlapic->vm = svm_sc->vm;
2384 	vlapic->vcpuid = vcpuid;
2385 	vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2386 
2387 	vlapic_init(vlapic);
2388 
2389 	return (vlapic);
2390 }
2391 
2392 static void
2393 svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2394 {
2395 
2396         vlapic_cleanup(vlapic);
2397         free(vlapic, M_SVM_VLAPIC);
2398 }
2399 
2400 #ifdef BHYVE_SNAPSHOT
2401 static int
2402 svm_snapshot(void *arg, struct vm_snapshot_meta *meta)
2403 {
2404 	/* struct svm_softc is AMD's representation for SVM softc */
2405 	struct svm_softc *sc;
2406 	struct svm_vcpu *vcpu;
2407 	struct vmcb *vmcb;
2408 	uint64_t val;
2409 	int i;
2410 	int ret;
2411 
2412 	sc = arg;
2413 
2414 	KASSERT(sc != NULL, ("%s: arg was NULL", __func__));
2415 
2416 	SNAPSHOT_VAR_OR_LEAVE(sc->nptp, meta, ret, done);
2417 
2418 	for (i = 0; i < VM_MAXCPU; i++) {
2419 		vcpu = &sc->vcpu[i];
2420 		vmcb = &vcpu->vmcb;
2421 
2422 		/* VMCB fields for virtual cpu i */
2423 		SNAPSHOT_VAR_OR_LEAVE(vmcb->ctrl.v_tpr, meta, ret, done);
2424 		val = vmcb->ctrl.v_tpr;
2425 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2426 		vmcb->ctrl.v_tpr = val;
2427 
2428 		SNAPSHOT_VAR_OR_LEAVE(vmcb->ctrl.asid, meta, ret, done);
2429 		val = vmcb->ctrl.np_enable;
2430 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2431 		vmcb->ctrl.np_enable = val;
2432 
2433 		val = vmcb->ctrl.intr_shadow;
2434 		SNAPSHOT_VAR_OR_LEAVE(val, meta, ret, done);
2435 		vmcb->ctrl.intr_shadow = val;
2436 		SNAPSHOT_VAR_OR_LEAVE(vmcb->ctrl.tlb_ctrl, meta, ret, done);
2437 
2438 		SNAPSHOT_BUF_OR_LEAVE(vmcb->state.pad1,
2439 				      sizeof(vmcb->state.pad1),
2440 				      meta, ret, done);
2441 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.cpl, meta, ret, done);
2442 		SNAPSHOT_BUF_OR_LEAVE(vmcb->state.pad2,
2443 				      sizeof(vmcb->state.pad2),
2444 				      meta, ret, done);
2445 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.efer, meta, ret, done);
2446 		SNAPSHOT_BUF_OR_LEAVE(vmcb->state.pad3,
2447 				      sizeof(vmcb->state.pad3),
2448 				      meta, ret, done);
2449 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.cr4, meta, ret, done);
2450 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.cr3, meta, ret, done);
2451 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.cr0, meta, ret, done);
2452 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.dr7, meta, ret, done);
2453 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.dr6, meta, ret, done);
2454 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.rflags, meta, ret, done);
2455 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.rip, meta, ret, done);
2456 		SNAPSHOT_BUF_OR_LEAVE(vmcb->state.pad4,
2457 				      sizeof(vmcb->state.pad4),
2458 				      meta, ret, done);
2459 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.rsp, meta, ret, done);
2460 		SNAPSHOT_BUF_OR_LEAVE(vmcb->state.pad5,
2461 				      sizeof(vmcb->state.pad5),
2462 				      meta, ret, done);
2463 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.rax, meta, ret, done);
2464 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.star, meta, ret, done);
2465 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.lstar, meta, ret, done);
2466 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.cstar, meta, ret, done);
2467 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.sfmask, meta, ret, done);
2468 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.kernelgsbase,
2469 				      meta, ret, done);
2470 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.sysenter_cs, meta, ret, done);
2471 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.sysenter_esp,
2472 				      meta, ret, done);
2473 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.sysenter_eip,
2474 				      meta, ret, done);
2475 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.cr2, meta, ret, done);
2476 		SNAPSHOT_BUF_OR_LEAVE(vmcb->state.pad6,
2477 				      sizeof(vmcb->state.pad6),
2478 				      meta, ret, done);
2479 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.g_pat, meta, ret, done);
2480 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.dbgctl, meta, ret, done);
2481 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.br_from, meta, ret, done);
2482 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.br_to, meta, ret, done);
2483 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.int_from, meta, ret, done);
2484 		SNAPSHOT_VAR_OR_LEAVE(vmcb->state.int_to, meta, ret, done);
2485 		SNAPSHOT_BUF_OR_LEAVE(vmcb->state.pad7,
2486 				      sizeof(vmcb->state.pad7),
2487 				      meta, ret, done);
2488 
2489 		/* Snapshot swctx for virtual cpu i */
2490 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbp, meta, ret, done);
2491 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rbx, meta, ret, done);
2492 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rcx, meta, ret, done);
2493 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdx, meta, ret, done);
2494 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rdi, meta, ret, done);
2495 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_rsi, meta, ret, done);
2496 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r8, meta, ret, done);
2497 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r9, meta, ret, done);
2498 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r10, meta, ret, done);
2499 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r11, meta, ret, done);
2500 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r12, meta, ret, done);
2501 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r13, meta, ret, done);
2502 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r14, meta, ret, done);
2503 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_r15, meta, ret, done);
2504 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr0, meta, ret, done);
2505 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr1, meta, ret, done);
2506 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr2, meta, ret, done);
2507 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.sctx_dr3, meta, ret, done);
2508 
2509 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.host_dr0, meta, ret, done);
2510 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.host_dr1, meta, ret, done);
2511 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.host_dr2, meta, ret, done);
2512 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.host_dr3, meta, ret, done);
2513 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.host_dr6, meta, ret, done);
2514 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.host_dr7, meta, ret, done);
2515 		SNAPSHOT_VAR_OR_LEAVE(vcpu->swctx.host_debugctl, meta, ret,
2516 				      done);
2517 
2518 		/* Restore other svm_vcpu struct fields */
2519 
2520 		/* Restore NEXTRIP field */
2521 		SNAPSHOT_VAR_OR_LEAVE(vcpu->nextrip, meta, ret, done);
2522 
2523 		/* Restore lastcpu field */
2524 		SNAPSHOT_VAR_OR_LEAVE(vcpu->lastcpu, meta, ret, done);
2525 		SNAPSHOT_VAR_OR_LEAVE(vcpu->dirty, meta, ret, done);
2526 
2527 		/* Restore EPTGEN field - EPT is Extended Page Tabel */
2528 		SNAPSHOT_VAR_OR_LEAVE(vcpu->eptgen, meta, ret, done);
2529 
2530 		SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.gen, meta, ret, done);
2531 		SNAPSHOT_VAR_OR_LEAVE(vcpu->asid.num, meta, ret, done);
2532 
2533 		/* Set all caches dirty */
2534 		if (meta->op == VM_SNAPSHOT_RESTORE) {
2535 			svm_set_dirty(sc, i, VMCB_CACHE_ASID);
2536 			svm_set_dirty(sc, i, VMCB_CACHE_IOPM);
2537 			svm_set_dirty(sc, i, VMCB_CACHE_I);
2538 			svm_set_dirty(sc, i, VMCB_CACHE_TPR);
2539 			svm_set_dirty(sc, i, VMCB_CACHE_CR2);
2540 			svm_set_dirty(sc, i, VMCB_CACHE_CR);
2541 			svm_set_dirty(sc, i, VMCB_CACHE_DT);
2542 			svm_set_dirty(sc, i, VMCB_CACHE_SEG);
2543 			svm_set_dirty(sc, i, VMCB_CACHE_NP);
2544 		}
2545 	}
2546 
2547 	if (meta->op == VM_SNAPSHOT_RESTORE)
2548 		flush_by_asid();
2549 
2550 done:
2551 	return (ret);
2552 }
2553 
2554 static int
2555 svm_vmcx_snapshot(void *arg, struct vm_snapshot_meta *meta, int vcpu)
2556 {
2557 	struct svm_softc *sc;
2558 	int err, running, hostcpu;
2559 
2560 	sc = (struct svm_softc *)arg;
2561 	err = 0;
2562 
2563 	KASSERT(arg != NULL, ("%s: arg was NULL", __func__));
2564 
2565 	running = vcpu_is_running(sc->vm, vcpu, &hostcpu);
2566 	if (running && hostcpu !=curcpu) {
2567 		printf("%s: %s%d is running", __func__, vm_name(sc->vm), vcpu);
2568 		return (EINVAL);
2569 	}
2570 
2571 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR0, meta);
2572 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR2, meta);
2573 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR3, meta);
2574 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CR4, meta);
2575 
2576 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_DR7, meta);
2577 
2578 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RAX, meta);
2579 
2580 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RSP, meta);
2581 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RIP, meta);
2582 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_RFLAGS, meta);
2583 
2584 	/* Guest segments */
2585 	/* ES */
2586 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_ES, meta);
2587 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_ES, meta);
2588 
2589 	/* CS */
2590 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_CS, meta);
2591 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_CS, meta);
2592 
2593 	/* SS */
2594 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_SS, meta);
2595 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_SS, meta);
2596 
2597 	/* DS */
2598 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_DS, meta);
2599 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_DS, meta);
2600 
2601 	/* FS */
2602 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_FS, meta);
2603 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_FS, meta);
2604 
2605 	/* GS */
2606 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_GS, meta);
2607 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_GS, meta);
2608 
2609 	/* TR */
2610 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_TR, meta);
2611 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_TR, meta);
2612 
2613 	/* LDTR */
2614 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_LDTR, meta);
2615 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_LDTR, meta);
2616 
2617 	/* EFER */
2618 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_EFER, meta);
2619 
2620 	/* IDTR and GDTR */
2621 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_IDTR, meta);
2622 	err += vmcb_snapshot_desc(sc, vcpu, VM_REG_GUEST_GDTR, meta);
2623 
2624 	/* Specific AMD registers */
2625 	err += vmcb_snapshot_any(sc, vcpu,
2626 				VMCB_ACCESS(VMCB_OFF_SYSENTER_CS, 8), meta);
2627 	err += vmcb_snapshot_any(sc, vcpu,
2628 				VMCB_ACCESS(VMCB_OFF_SYSENTER_ESP, 8), meta);
2629 	err += vmcb_snapshot_any(sc, vcpu,
2630 				VMCB_ACCESS(VMCB_OFF_SYSENTER_EIP, 8), meta);
2631 
2632 	err += vmcb_snapshot_any(sc, vcpu,
2633 				VMCB_ACCESS(VMCB_OFF_NPT_BASE, 8), meta);
2634 
2635 	err += vmcb_snapshot_any(sc, vcpu,
2636 				VMCB_ACCESS(VMCB_OFF_CR_INTERCEPT, 4), meta);
2637 	err += vmcb_snapshot_any(sc, vcpu,
2638 				VMCB_ACCESS(VMCB_OFF_DR_INTERCEPT, 4), meta);
2639 	err += vmcb_snapshot_any(sc, vcpu,
2640 				VMCB_ACCESS(VMCB_OFF_EXC_INTERCEPT, 4), meta);
2641 	err += vmcb_snapshot_any(sc, vcpu,
2642 				VMCB_ACCESS(VMCB_OFF_INST1_INTERCEPT, 4), meta);
2643 	err += vmcb_snapshot_any(sc, vcpu,
2644 				VMCB_ACCESS(VMCB_OFF_INST2_INTERCEPT, 4), meta);
2645 
2646 	err += vmcb_snapshot_any(sc, vcpu,
2647 				VMCB_ACCESS(VMCB_OFF_TLB_CTRL, 4), meta);
2648 
2649 	err += vmcb_snapshot_any(sc, vcpu,
2650 				VMCB_ACCESS(VMCB_OFF_EXITINFO1, 8), meta);
2651 	err += vmcb_snapshot_any(sc, vcpu,
2652 				VMCB_ACCESS(VMCB_OFF_EXITINFO2, 8), meta);
2653 	err += vmcb_snapshot_any(sc, vcpu,
2654 				VMCB_ACCESS(VMCB_OFF_EXITINTINFO, 8), meta);
2655 
2656 	err += vmcb_snapshot_any(sc, vcpu,
2657 				VMCB_ACCESS(VMCB_OFF_VIRQ, 8), meta);
2658 
2659 	err += vmcb_snapshot_any(sc, vcpu,
2660 				VMCB_ACCESS(VMCB_OFF_GUEST_PAT, 8), meta);
2661 
2662 	err += vmcb_snapshot_any(sc, vcpu,
2663 				VMCB_ACCESS(VMCB_OFF_AVIC_BAR, 8), meta);
2664 	err += vmcb_snapshot_any(sc, vcpu,
2665 				VMCB_ACCESS(VMCB_OFF_AVIC_PAGE, 8), meta);
2666 	err += vmcb_snapshot_any(sc, vcpu,
2667 				VMCB_ACCESS(VMCB_OFF_AVIC_LT, 8), meta);
2668 	err += vmcb_snapshot_any(sc, vcpu,
2669 				VMCB_ACCESS(VMCB_OFF_AVIC_PT, 8), meta);
2670 
2671 	err += vmcb_snapshot_any(sc, vcpu,
2672 				VMCB_ACCESS(VMCB_OFF_IO_PERM, 8), meta);
2673 	err += vmcb_snapshot_any(sc, vcpu,
2674 				VMCB_ACCESS(VMCB_OFF_MSR_PERM, 8), meta);
2675 
2676 	err += vmcb_snapshot_any(sc, vcpu,
2677 				VMCB_ACCESS(VMCB_OFF_ASID, 4), meta);
2678 
2679 	err += vmcb_snapshot_any(sc, vcpu,
2680 				VMCB_ACCESS(VMCB_OFF_EXIT_REASON, 8), meta);
2681 
2682 	err += svm_snapshot_reg(sc, vcpu, VM_REG_GUEST_INTR_SHADOW, meta);
2683 
2684 	return (err);
2685 }
2686 
2687 static int
2688 svm_restore_tsc(void *arg, int vcpu, uint64_t offset)
2689 {
2690 	int err;
2691 
2692 	err = svm_set_tsc_offset(arg, vcpu, offset);
2693 
2694 	return (err);
2695 }
2696 #endif
2697 
2698 const struct vmm_ops vmm_ops_amd = {
2699 	.modinit	= svm_modinit,
2700 	.modcleanup	= svm_modcleanup,
2701 	.modresume	= svm_modresume,
2702 	.init		= svm_init,
2703 	.run		= svm_run,
2704 	.cleanup	= svm_cleanup,
2705 	.getreg		= svm_getreg,
2706 	.setreg		= svm_setreg,
2707 	.getdesc	= svm_getdesc,
2708 	.setdesc	= svm_setdesc,
2709 	.getcap		= svm_getcap,
2710 	.setcap		= svm_setcap,
2711 	.vmspace_alloc	= svm_vmspace_alloc,
2712 	.vmspace_free	= svm_vmspace_free,
2713 	.vlapic_init	= svm_vlapic_init,
2714 	.vlapic_cleanup	= svm_vlapic_cleanup,
2715 #ifdef BHYVE_SNAPSHOT
2716 	.snapshot	= svm_snapshot,
2717 	.vmcx_snapshot	= svm_vmcx_snapshot,
2718 	.restore_tsc	= svm_restore_tsc,
2719 #endif
2720 };
2721