xref: /freebsd/sys/amd64/vmm/amd/svm.c (revision 28f42739a547ffe0b5dfaaf9f49fb4c4813aa232)
1 /*-
2  * Copyright (c) 2013, Anish Gupta (akgupt3@gmail.com)
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/smp.h>
33 #include <sys/kernel.h>
34 #include <sys/malloc.h>
35 #include <sys/pcpu.h>
36 #include <sys/proc.h>
37 #include <sys/sysctl.h>
38 
39 #include <vm/vm.h>
40 #include <vm/pmap.h>
41 
42 #include <machine/cpufunc.h>
43 #include <machine/psl.h>
44 #include <machine/pmap.h>
45 #include <machine/md_var.h>
46 #include <machine/specialreg.h>
47 #include <machine/smp.h>
48 #include <machine/vmm.h>
49 #include <machine/vmm_dev.h>
50 #include <machine/vmm_instruction_emul.h>
51 
52 #include "vmm_lapic.h"
53 #include "vmm_stat.h"
54 #include "vmm_ktr.h"
55 #include "vmm_ioport.h"
56 #include "vatpic.h"
57 #include "vlapic.h"
58 #include "vlapic_priv.h"
59 
60 #include "x86.h"
61 #include "vmcb.h"
62 #include "svm.h"
63 #include "svm_softc.h"
64 #include "svm_msr.h"
65 #include "npt.h"
66 
67 SYSCTL_DECL(_hw_vmm);
68 SYSCTL_NODE(_hw_vmm, OID_AUTO, svm, CTLFLAG_RW, NULL, NULL);
69 
70 /*
71  * SVM CPUID function 0x8000_000A, edx bit decoding.
72  */
73 #define AMD_CPUID_SVM_NP		BIT(0)  /* Nested paging or RVI */
74 #define AMD_CPUID_SVM_LBR		BIT(1)  /* Last branch virtualization */
75 #define AMD_CPUID_SVM_SVML		BIT(2)  /* SVM lock */
76 #define AMD_CPUID_SVM_NRIP_SAVE		BIT(3)  /* Next RIP is saved */
77 #define AMD_CPUID_SVM_TSC_RATE		BIT(4)  /* TSC rate control. */
78 #define AMD_CPUID_SVM_VMCB_CLEAN	BIT(5)  /* VMCB state caching */
79 #define AMD_CPUID_SVM_FLUSH_BY_ASID	BIT(6)  /* Flush by ASID */
80 #define AMD_CPUID_SVM_DECODE_ASSIST	BIT(7)  /* Decode assist */
81 #define AMD_CPUID_SVM_PAUSE_INC		BIT(10) /* Pause intercept filter. */
82 #define AMD_CPUID_SVM_PAUSE_FTH		BIT(12) /* Pause filter threshold */
83 
84 #define	VMCB_CACHE_DEFAULT	(VMCB_CACHE_ASID 	|	\
85 				VMCB_CACHE_IOPM		|	\
86 				VMCB_CACHE_I		|	\
87 				VMCB_CACHE_TPR		|	\
88 				VMCB_CACHE_CR2		|	\
89 				VMCB_CACHE_CR		|	\
90 				VMCB_CACHE_DT		|	\
91 				VMCB_CACHE_SEG		|	\
92 				VMCB_CACHE_NP)
93 
94 static uint32_t vmcb_clean = VMCB_CACHE_DEFAULT;
95 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, vmcb_clean, CTLFLAG_RDTUN, &vmcb_clean,
96     0, NULL);
97 
98 static MALLOC_DEFINE(M_SVM, "svm", "svm");
99 static MALLOC_DEFINE(M_SVM_VLAPIC, "svm-vlapic", "svm-vlapic");
100 
101 /* Per-CPU context area. */
102 extern struct pcpu __pcpu[];
103 
104 static uint32_t svm_feature;	/* AMD SVM features. */
105 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, features, CTLFLAG_RD, &svm_feature, 0,
106     "SVM features advertised by CPUID.8000000AH:EDX");
107 
108 static int disable_npf_assist;
109 SYSCTL_INT(_hw_vmm_svm, OID_AUTO, disable_npf_assist, CTLFLAG_RWTUN,
110     &disable_npf_assist, 0, NULL);
111 
112 /* Maximum ASIDs supported by the processor */
113 static uint32_t nasid;
114 SYSCTL_UINT(_hw_vmm_svm, OID_AUTO, num_asids, CTLFLAG_RD, &nasid, 0,
115     "Number of ASIDs supported by this processor");
116 
117 /* Current ASID generation for each host cpu */
118 static struct asid asid[MAXCPU];
119 
120 /*
121  * SVM host state saved area of size 4KB for each core.
122  */
123 static uint8_t hsave[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE);
124 
125 static VMM_STAT_AMD(VCPU_EXITINTINFO, "VM exits during event delivery");
126 static VMM_STAT_AMD(VCPU_INTINFO_INJECTED, "Events pending at VM entry");
127 static VMM_STAT_AMD(VMEXIT_VINTR, "VM exits due to interrupt window");
128 
129 static int svm_setreg(void *arg, int vcpu, int ident, uint64_t val);
130 
131 static __inline int
132 flush_by_asid(void)
133 {
134 
135 	return (svm_feature & AMD_CPUID_SVM_FLUSH_BY_ASID);
136 }
137 
138 static __inline int
139 decode_assist(void)
140 {
141 
142 	return (svm_feature & AMD_CPUID_SVM_DECODE_ASSIST);
143 }
144 
145 static void
146 svm_disable(void *arg __unused)
147 {
148 	uint64_t efer;
149 
150 	efer = rdmsr(MSR_EFER);
151 	efer &= ~EFER_SVM;
152 	wrmsr(MSR_EFER, efer);
153 }
154 
155 /*
156  * Disable SVM on all CPUs.
157  */
158 static int
159 svm_cleanup(void)
160 {
161 
162 	smp_rendezvous(NULL, svm_disable, NULL, NULL);
163 	return (0);
164 }
165 
166 /*
167  * Verify that all the features required by bhyve are available.
168  */
169 static int
170 check_svm_features(void)
171 {
172 	u_int regs[4];
173 
174 	/* CPUID Fn8000_000A is for SVM */
175 	do_cpuid(0x8000000A, regs);
176 	svm_feature = regs[3];
177 
178 	nasid = regs[1];
179 	KASSERT(nasid > 1, ("Insufficient ASIDs for guests: %#x", nasid));
180 
181 	/* bhyve requires the Nested Paging feature */
182 	if (!(svm_feature & AMD_CPUID_SVM_NP)) {
183 		printf("SVM: Nested Paging feature not available.\n");
184 		return (ENXIO);
185 	}
186 
187 	/* bhyve requires the NRIP Save feature */
188 	if (!(svm_feature & AMD_CPUID_SVM_NRIP_SAVE)) {
189 		printf("SVM: NRIP Save feature not available.\n");
190 		return (ENXIO);
191 	}
192 
193 	return (0);
194 }
195 
196 static void
197 svm_enable(void *arg __unused)
198 {
199 	uint64_t efer;
200 
201 	efer = rdmsr(MSR_EFER);
202 	efer |= EFER_SVM;
203 	wrmsr(MSR_EFER, efer);
204 
205 	wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave[curcpu]));
206 }
207 
208 /*
209  * Return 1 if SVM is enabled on this processor and 0 otherwise.
210  */
211 static int
212 svm_available(void)
213 {
214 	uint64_t msr;
215 
216 	/* Section 15.4 Enabling SVM from APM2. */
217 	if ((amd_feature2 & AMDID2_SVM) == 0) {
218 		printf("SVM: not available.\n");
219 		return (0);
220 	}
221 
222 	msr = rdmsr(MSR_VM_CR);
223 	if ((msr & VM_CR_SVMDIS) != 0) {
224 		printf("SVM: disabled by BIOS.\n");
225 		return (0);
226 	}
227 
228 	return (1);
229 }
230 
231 static int
232 svm_init(int ipinum)
233 {
234 	int error, cpu;
235 
236 	if (!svm_available())
237 		return (ENXIO);
238 
239 	error = check_svm_features();
240 	if (error)
241 		return (error);
242 
243 	vmcb_clean &= VMCB_CACHE_DEFAULT;
244 
245 	for (cpu = 0; cpu < MAXCPU; cpu++) {
246 		/*
247 		 * Initialize the host ASIDs to their "highest" valid values.
248 		 *
249 		 * The next ASID allocation will rollover both 'gen' and 'num'
250 		 * and start off the sequence at {1,1}.
251 		 */
252 		asid[cpu].gen = ~0UL;
253 		asid[cpu].num = nasid - 1;
254 	}
255 
256 	svm_msr_init();
257 	svm_npt_init(ipinum);
258 
259 	/* Enable SVM on all CPUs */
260 	smp_rendezvous(NULL, svm_enable, NULL, NULL);
261 
262 	return (0);
263 }
264 
265 static void
266 svm_restore(void)
267 {
268 
269 	svm_enable(NULL);
270 }
271 
272 /* Pentium compatible MSRs */
273 #define MSR_PENTIUM_START 	0
274 #define MSR_PENTIUM_END 	0x1FFF
275 /* AMD 6th generation and Intel compatible MSRs */
276 #define MSR_AMD6TH_START 	0xC0000000UL
277 #define MSR_AMD6TH_END 		0xC0001FFFUL
278 /* AMD 7th and 8th generation compatible MSRs */
279 #define MSR_AMD7TH_START 	0xC0010000UL
280 #define MSR_AMD7TH_END 		0xC0011FFFUL
281 
282 /*
283  * Get the index and bit position for a MSR in permission bitmap.
284  * Two bits are used for each MSR: lower bit for read and higher bit for write.
285  */
286 static int
287 svm_msr_index(uint64_t msr, int *index, int *bit)
288 {
289 	uint32_t base, off;
290 
291 	*index = -1;
292 	*bit = (msr % 4) * 2;
293 	base = 0;
294 
295 	if (msr >= MSR_PENTIUM_START && msr <= MSR_PENTIUM_END) {
296 		*index = msr / 4;
297 		return (0);
298 	}
299 
300 	base += (MSR_PENTIUM_END - MSR_PENTIUM_START + 1);
301 	if (msr >= MSR_AMD6TH_START && msr <= MSR_AMD6TH_END) {
302 		off = (msr - MSR_AMD6TH_START);
303 		*index = (off + base) / 4;
304 		return (0);
305 	}
306 
307 	base += (MSR_AMD6TH_END - MSR_AMD6TH_START + 1);
308 	if (msr >= MSR_AMD7TH_START && msr <= MSR_AMD7TH_END) {
309 		off = (msr - MSR_AMD7TH_START);
310 		*index = (off + base) / 4;
311 		return (0);
312 	}
313 
314 	return (EINVAL);
315 }
316 
317 /*
318  * Allow vcpu to read or write the 'msr' without trapping into the hypervisor.
319  */
320 static void
321 svm_msr_perm(uint8_t *perm_bitmap, uint64_t msr, bool read, bool write)
322 {
323 	int index, bit, error;
324 
325 	error = svm_msr_index(msr, &index, &bit);
326 	KASSERT(error == 0, ("%s: invalid msr %#lx", __func__, msr));
327 	KASSERT(index >= 0 && index < SVM_MSR_BITMAP_SIZE,
328 	    ("%s: invalid index %d for msr %#lx", __func__, index, msr));
329 	KASSERT(bit >= 0 && bit <= 6, ("%s: invalid bit position %d "
330 	    "msr %#lx", __func__, bit, msr));
331 
332 	if (read)
333 		perm_bitmap[index] &= ~(1UL << bit);
334 
335 	if (write)
336 		perm_bitmap[index] &= ~(2UL << bit);
337 }
338 
339 static void
340 svm_msr_rw_ok(uint8_t *perm_bitmap, uint64_t msr)
341 {
342 
343 	svm_msr_perm(perm_bitmap, msr, true, true);
344 }
345 
346 static void
347 svm_msr_rd_ok(uint8_t *perm_bitmap, uint64_t msr)
348 {
349 
350 	svm_msr_perm(perm_bitmap, msr, true, false);
351 }
352 
353 static __inline int
354 svm_get_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask)
355 {
356 	struct vmcb_ctrl *ctrl;
357 
358 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
359 
360 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
361 	return (ctrl->intercept[idx] & bitmask ? 1 : 0);
362 }
363 
364 static __inline void
365 svm_set_intercept(struct svm_softc *sc, int vcpu, int idx, uint32_t bitmask,
366     int enabled)
367 {
368 	struct vmcb_ctrl *ctrl;
369 	uint32_t oldval;
370 
371 	KASSERT(idx >=0 && idx < 5, ("invalid intercept index %d", idx));
372 
373 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
374 	oldval = ctrl->intercept[idx];
375 
376 	if (enabled)
377 		ctrl->intercept[idx] |= bitmask;
378 	else
379 		ctrl->intercept[idx] &= ~bitmask;
380 
381 	if (ctrl->intercept[idx] != oldval) {
382 		svm_set_dirty(sc, vcpu, VMCB_CACHE_I);
383 		VCPU_CTR3(sc->vm, vcpu, "intercept[%d] modified "
384 		    "from %#x to %#x", idx, oldval, ctrl->intercept[idx]);
385 	}
386 }
387 
388 static __inline void
389 svm_disable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
390 {
391 
392 	svm_set_intercept(sc, vcpu, off, bitmask, 0);
393 }
394 
395 static __inline void
396 svm_enable_intercept(struct svm_softc *sc, int vcpu, int off, uint32_t bitmask)
397 {
398 
399 	svm_set_intercept(sc, vcpu, off, bitmask, 1);
400 }
401 
402 static void
403 vmcb_init(struct svm_softc *sc, int vcpu, uint64_t iopm_base_pa,
404     uint64_t msrpm_base_pa, uint64_t np_pml4)
405 {
406 	struct vmcb_ctrl *ctrl;
407 	struct vmcb_state *state;
408 	uint32_t mask;
409 	int n;
410 
411 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
412 	state = svm_get_vmcb_state(sc, vcpu);
413 
414 	ctrl->iopm_base_pa = iopm_base_pa;
415 	ctrl->msrpm_base_pa = msrpm_base_pa;
416 
417 	/* Enable nested paging */
418 	ctrl->np_enable = 1;
419 	ctrl->n_cr3 = np_pml4;
420 
421 	/*
422 	 * Intercept accesses to the control registers that are not shadowed
423 	 * in the VMCB - i.e. all except cr0, cr2, cr3, cr4 and cr8.
424 	 */
425 	for (n = 0; n < 16; n++) {
426 		mask = (BIT(n) << 16) | BIT(n);
427 		if (n == 0 || n == 2 || n == 3 || n == 4 || n == 8)
428 			svm_disable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
429 		else
430 			svm_enable_intercept(sc, vcpu, VMCB_CR_INTCPT, mask);
431 	}
432 
433 
434 	/*
435 	 * Intercept everything when tracing guest exceptions otherwise
436 	 * just intercept machine check exception.
437 	 */
438 	if (vcpu_trace_exceptions(sc->vm, vcpu)) {
439 		for (n = 0; n < 32; n++) {
440 			/*
441 			 * Skip unimplemented vectors in the exception bitmap.
442 			 */
443 			if (n == 2 || n == 9) {
444 				continue;
445 			}
446 			svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(n));
447 		}
448 	} else {
449 		svm_enable_intercept(sc, vcpu, VMCB_EXC_INTCPT, BIT(IDT_MC));
450 	}
451 
452 	/* Intercept various events (for e.g. I/O, MSR and CPUID accesses) */
453 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IO);
454 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_MSR);
455 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_CPUID);
456 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INTR);
457 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_INIT);
458 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_NMI);
459 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SMI);
460 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_SHUTDOWN);
461 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
462 	    VMCB_INTCPT_FERR_FREEZE);
463 
464 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MONITOR);
465 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_MWAIT);
466 
467 	/*
468 	 * From section "Canonicalization and Consistency Checks" in APMv2
469 	 * the VMRUN intercept bit must be set to pass the consistency check.
470 	 */
471 	svm_enable_intercept(sc, vcpu, VMCB_CTRL2_INTCPT, VMCB_INTCPT_VMRUN);
472 
473 	/*
474 	 * The ASID will be set to a non-zero value just before VMRUN.
475 	 */
476 	ctrl->asid = 0;
477 
478 	/*
479 	 * Section 15.21.1, Interrupt Masking in EFLAGS
480 	 * Section 15.21.2, Virtualizing APIC.TPR
481 	 *
482 	 * This must be set for %rflag and %cr8 isolation of guest and host.
483 	 */
484 	ctrl->v_intr_masking = 1;
485 
486 	/* Enable Last Branch Record aka LBR for debugging */
487 	ctrl->lbr_virt_en = 1;
488 	state->dbgctl = BIT(0);
489 
490 	/* EFER_SVM must always be set when the guest is executing */
491 	state->efer = EFER_SVM;
492 
493 	/* Set up the PAT to power-on state */
494 	state->g_pat = PAT_VALUE(0, PAT_WRITE_BACK)	|
495 	    PAT_VALUE(1, PAT_WRITE_THROUGH)	|
496 	    PAT_VALUE(2, PAT_UNCACHED)		|
497 	    PAT_VALUE(3, PAT_UNCACHEABLE)	|
498 	    PAT_VALUE(4, PAT_WRITE_BACK)	|
499 	    PAT_VALUE(5, PAT_WRITE_THROUGH)	|
500 	    PAT_VALUE(6, PAT_UNCACHED)		|
501 	    PAT_VALUE(7, PAT_UNCACHEABLE);
502 }
503 
504 /*
505  * Initialize a virtual machine.
506  */
507 static void *
508 svm_vminit(struct vm *vm, pmap_t pmap)
509 {
510 	struct svm_softc *svm_sc;
511 	struct svm_vcpu *vcpu;
512 	vm_paddr_t msrpm_pa, iopm_pa, pml4_pa;
513 	int i;
514 
515 	svm_sc = malloc(sizeof (struct svm_softc), M_SVM, M_WAITOK | M_ZERO);
516 	svm_sc->vm = vm;
517 	svm_sc->nptp = (vm_offset_t)vtophys(pmap->pm_pml4);
518 
519 	/*
520 	 * Intercept read and write accesses to all MSRs.
521 	 */
522 	memset(svm_sc->msr_bitmap, 0xFF, sizeof(svm_sc->msr_bitmap));
523 
524 	/*
525 	 * Access to the following MSRs is redirected to the VMCB when the
526 	 * guest is executing. Therefore it is safe to allow the guest to
527 	 * read/write these MSRs directly without hypervisor involvement.
528 	 */
529 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_GSBASE);
530 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_FSBASE);
531 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_KGSBASE);
532 
533 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_STAR);
534 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_LSTAR);
535 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_CSTAR);
536 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SF_MASK);
537 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_CS_MSR);
538 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_ESP_MSR);
539 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_SYSENTER_EIP_MSR);
540 	svm_msr_rw_ok(svm_sc->msr_bitmap, MSR_PAT);
541 
542 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_TSC);
543 
544 	/*
545 	 * Intercept writes to make sure that the EFER_SVM bit is not cleared.
546 	 */
547 	svm_msr_rd_ok(svm_sc->msr_bitmap, MSR_EFER);
548 
549 	/* Intercept access to all I/O ports. */
550 	memset(svm_sc->iopm_bitmap, 0xFF, sizeof(svm_sc->iopm_bitmap));
551 
552 	iopm_pa = vtophys(svm_sc->iopm_bitmap);
553 	msrpm_pa = vtophys(svm_sc->msr_bitmap);
554 	pml4_pa = svm_sc->nptp;
555 	for (i = 0; i < VM_MAXCPU; i++) {
556 		vcpu = svm_get_vcpu(svm_sc, i);
557 		vcpu->nextrip = ~0;
558 		vcpu->lastcpu = NOCPU;
559 		vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
560 		vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
561 		svm_msr_guest_init(svm_sc, i);
562 	}
563 	return (svm_sc);
564 }
565 
566 static int
567 svm_cpl(struct vmcb_state *state)
568 {
569 
570 	/*
571 	 * From APMv2:
572 	 *   "Retrieve the CPL from the CPL field in the VMCB, not
573 	 *    from any segment DPL"
574 	 */
575 	return (state->cpl);
576 }
577 
578 static enum vm_cpu_mode
579 svm_vcpu_mode(struct vmcb *vmcb)
580 {
581 	struct vmcb_segment seg;
582 	struct vmcb_state *state;
583 	int error;
584 
585 	state = &vmcb->state;
586 
587 	if (state->efer & EFER_LMA) {
588 		error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
589 		KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
590 		    error));
591 
592 		/*
593 		 * Section 4.8.1 for APM2, check if Code Segment has
594 		 * Long attribute set in descriptor.
595 		 */
596 		if (seg.attrib & VMCB_CS_ATTRIB_L)
597 			return (CPU_MODE_64BIT);
598 		else
599 			return (CPU_MODE_COMPATIBILITY);
600 	} else  if (state->cr0 & CR0_PE) {
601 		return (CPU_MODE_PROTECTED);
602 	} else {
603 		return (CPU_MODE_REAL);
604 	}
605 }
606 
607 static enum vm_paging_mode
608 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
609 {
610 
611 	if ((cr0 & CR0_PG) == 0)
612 		return (PAGING_MODE_FLAT);
613 	if ((cr4 & CR4_PAE) == 0)
614 		return (PAGING_MODE_32);
615 	if (efer & EFER_LME)
616 		return (PAGING_MODE_64);
617 	else
618 		return (PAGING_MODE_PAE);
619 }
620 
621 /*
622  * ins/outs utility routines
623  */
624 static uint64_t
625 svm_inout_str_index(struct svm_regctx *regs, int in)
626 {
627 	uint64_t val;
628 
629 	val = in ? regs->sctx_rdi : regs->sctx_rsi;
630 
631 	return (val);
632 }
633 
634 static uint64_t
635 svm_inout_str_count(struct svm_regctx *regs, int rep)
636 {
637 	uint64_t val;
638 
639 	val = rep ? regs->sctx_rcx : 1;
640 
641 	return (val);
642 }
643 
644 static void
645 svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
646     int in, struct vm_inout_str *vis)
647 {
648 	int error, s;
649 
650 	if (in) {
651 		vis->seg_name = VM_REG_GUEST_ES;
652 	} else {
653 		/* The segment field has standard encoding */
654 		s = (info1 >> 10) & 0x7;
655 		vis->seg_name = vm_segment_name(s);
656 	}
657 
658 	error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
659 	KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
660 }
661 
662 static int
663 svm_inout_str_addrsize(uint64_t info1)
664 {
665         uint32_t size;
666 
667         size = (info1 >> 7) & 0x7;
668         switch (size) {
669         case 1:
670                 return (2);     /* 16 bit */
671         case 2:
672                 return (4);     /* 32 bit */
673         case 4:
674                 return (8);     /* 64 bit */
675         default:
676                 panic("%s: invalid size encoding %d", __func__, size);
677         }
678 }
679 
680 static void
681 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
682 {
683 	struct vmcb_state *state;
684 
685 	state = &vmcb->state;
686 	paging->cr3 = state->cr3;
687 	paging->cpl = svm_cpl(state);
688 	paging->cpu_mode = svm_vcpu_mode(vmcb);
689 	paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
690 	    state->efer);
691 }
692 
693 #define	UNHANDLED 0
694 
695 /*
696  * Handle guest I/O intercept.
697  */
698 static int
699 svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
700 {
701 	struct vmcb_ctrl *ctrl;
702 	struct vmcb_state *state;
703 	struct svm_regctx *regs;
704 	struct vm_inout_str *vis;
705 	uint64_t info1;
706 	int inout_string;
707 
708 	state = svm_get_vmcb_state(svm_sc, vcpu);
709 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
710 	regs  = svm_get_guest_regctx(svm_sc, vcpu);
711 
712 	info1 = ctrl->exitinfo1;
713 	inout_string = info1 & BIT(2) ? 1 : 0;
714 
715 	/*
716 	 * The effective segment number in EXITINFO1[12:10] is populated
717 	 * only if the processor has the DecodeAssist capability.
718 	 *
719 	 * XXX this is not specified explicitly in APMv2 but can be verified
720 	 * empirically.
721 	 */
722 	if (inout_string && !decode_assist())
723 		return (UNHANDLED);
724 
725 	vmexit->exitcode 	= VM_EXITCODE_INOUT;
726 	vmexit->u.inout.in 	= (info1 & BIT(0)) ? 1 : 0;
727 	vmexit->u.inout.string 	= inout_string;
728 	vmexit->u.inout.rep 	= (info1 & BIT(3)) ? 1 : 0;
729 	vmexit->u.inout.bytes 	= (info1 >> 4) & 0x7;
730 	vmexit->u.inout.port 	= (uint16_t)(info1 >> 16);
731 	vmexit->u.inout.eax 	= (uint32_t)(state->rax);
732 
733 	if (inout_string) {
734 		vmexit->exitcode = VM_EXITCODE_INOUT_STR;
735 		vis = &vmexit->u.inout_str;
736 		svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
737 		vis->rflags = state->rflags;
738 		vis->cr0 = state->cr0;
739 		vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
740 		vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
741 		vis->addrsize = svm_inout_str_addrsize(info1);
742 		svm_inout_str_seginfo(svm_sc, vcpu, info1,
743 		    vmexit->u.inout.in, vis);
744 	}
745 
746 	return (UNHANDLED);
747 }
748 
749 static int
750 npf_fault_type(uint64_t exitinfo1)
751 {
752 
753 	if (exitinfo1 & VMCB_NPF_INFO1_W)
754 		return (VM_PROT_WRITE);
755 	else if (exitinfo1 & VMCB_NPF_INFO1_ID)
756 		return (VM_PROT_EXECUTE);
757 	else
758 		return (VM_PROT_READ);
759 }
760 
761 static bool
762 svm_npf_emul_fault(uint64_t exitinfo1)
763 {
764 
765 	if (exitinfo1 & VMCB_NPF_INFO1_ID) {
766 		return (false);
767 	}
768 
769 	if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
770 		return (false);
771 	}
772 
773 	if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
774 		return (false);
775 	}
776 
777 	return (true);
778 }
779 
780 static void
781 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
782 {
783 	struct vm_guest_paging *paging;
784 	struct vmcb_segment seg;
785 	struct vmcb_ctrl *ctrl;
786 	char *inst_bytes;
787 	int error, inst_len;
788 
789 	ctrl = &vmcb->ctrl;
790 	paging = &vmexit->u.inst_emul.paging;
791 
792 	vmexit->exitcode = VM_EXITCODE_INST_EMUL;
793 	vmexit->u.inst_emul.gpa = gpa;
794 	vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
795 	svm_paging_info(vmcb, paging);
796 
797 	error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
798 	KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
799 
800 	switch(paging->cpu_mode) {
801 	case CPU_MODE_PROTECTED:
802 	case CPU_MODE_COMPATIBILITY:
803 		/*
804 		 * Section 4.8.1 of APM2, Default Operand Size or D bit.
805 		 */
806 		vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
807 		    1 : 0;
808 		break;
809 	default:
810 		vmexit->u.inst_emul.cs_d = 0;
811 		break;
812 	}
813 
814 	/*
815 	 * Copy the instruction bytes into 'vie' if available.
816 	 */
817 	if (decode_assist() && !disable_npf_assist) {
818 		inst_len = ctrl->inst_len;
819 		inst_bytes = ctrl->inst_bytes;
820 	} else {
821 		inst_len = 0;
822 		inst_bytes = NULL;
823 	}
824 	vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
825 }
826 
827 #ifdef KTR
828 static const char *
829 intrtype_to_str(int intr_type)
830 {
831 	switch (intr_type) {
832 	case VMCB_EVENTINJ_TYPE_INTR:
833 		return ("hwintr");
834 	case VMCB_EVENTINJ_TYPE_NMI:
835 		return ("nmi");
836 	case VMCB_EVENTINJ_TYPE_INTn:
837 		return ("swintr");
838 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
839 		return ("exception");
840 	default:
841 		panic("%s: unknown intr_type %d", __func__, intr_type);
842 	}
843 }
844 #endif
845 
846 /*
847  * Inject an event to vcpu as described in section 15.20, "Event injection".
848  */
849 static void
850 svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
851 		 uint32_t error, bool ec_valid)
852 {
853 	struct vmcb_ctrl *ctrl;
854 
855 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
856 
857 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
858 	    ("%s: event already pending %#lx", __func__, ctrl->eventinj));
859 
860 	KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
861 	    __func__, vector));
862 
863 	switch (intr_type) {
864 	case VMCB_EVENTINJ_TYPE_INTR:
865 	case VMCB_EVENTINJ_TYPE_NMI:
866 	case VMCB_EVENTINJ_TYPE_INTn:
867 		break;
868 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
869 		if (vector >= 0 && vector <= 31 && vector != 2)
870 			break;
871 		/* FALLTHROUGH */
872 	default:
873 		panic("%s: invalid intr_type/vector: %d/%d", __func__,
874 		    intr_type, vector);
875 	}
876 	ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
877 	if (ec_valid) {
878 		ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
879 		ctrl->eventinj |= (uint64_t)error << 32;
880 		VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
881 		    intrtype_to_str(intr_type), vector, error);
882 	} else {
883 		VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
884 		    intrtype_to_str(intr_type), vector);
885 	}
886 }
887 
888 static void
889 svm_update_virqinfo(struct svm_softc *sc, int vcpu)
890 {
891 	struct vm *vm;
892 	struct vlapic *vlapic;
893 	struct vmcb_ctrl *ctrl;
894 	int pending;
895 
896 	vm = sc->vm;
897 	vlapic = vm_lapic(vm, vcpu);
898 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
899 
900 	/* Update %cr8 in the emulated vlapic */
901 	vlapic_set_cr8(vlapic, ctrl->v_tpr);
902 
903 	/*
904 	 * If V_IRQ indicates that the interrupt injection attempted on then
905 	 * last VMRUN was successful then update the vlapic accordingly.
906 	 */
907 	if (ctrl->v_intr_vector != 0) {
908 		pending = ctrl->v_irq;
909 		KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid "
910 		    "v_intr_vector %d", __func__, ctrl->v_intr_vector));
911 		KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
912 		VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector,
913 		    pending ? "pending" : "accepted");
914 		if (!pending)
915 			vlapic_intr_accepted(vlapic, ctrl->v_intr_vector);
916 	}
917 }
918 
919 static void
920 svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
921 {
922 	struct vmcb_ctrl *ctrl;
923 	uint64_t intinfo;
924 
925 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
926 	intinfo = ctrl->exitintinfo;
927 	if (!VMCB_EXITINTINFO_VALID(intinfo))
928 		return;
929 
930 	/*
931 	 * From APMv2, Section "Intercepts during IDT interrupt delivery"
932 	 *
933 	 * If a #VMEXIT happened during event delivery then record the event
934 	 * that was being delivered.
935 	 */
936 	VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
937 		intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
938 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
939 	vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
940 }
941 
942 static __inline int
943 vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
944 {
945 
946 	return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
947 	    VMCB_INTCPT_VINTR));
948 }
949 
950 static __inline void
951 enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
952 {
953 	struct vmcb_ctrl *ctrl;
954 
955 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
956 
957 	if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
958 		KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
959 		KASSERT(vintr_intercept_enabled(sc, vcpu),
960 		    ("%s: vintr intercept should be enabled", __func__));
961 		return;
962 	}
963 
964 	VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
965 	ctrl->v_irq = 1;
966 	ctrl->v_ign_tpr = 1;
967 	ctrl->v_intr_vector = 0;
968 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
969 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
970 }
971 
972 static __inline void
973 disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
974 {
975 	struct vmcb_ctrl *ctrl;
976 
977 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
978 
979 	if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
980 		KASSERT(!vintr_intercept_enabled(sc, vcpu),
981 		    ("%s: vintr intercept should be disabled", __func__));
982 		return;
983 	}
984 
985 #ifdef KTR
986 	if (ctrl->v_intr_vector == 0)
987 		VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
988 	else
989 		VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection");
990 #endif
991 	ctrl->v_irq = 0;
992 	ctrl->v_intr_vector = 0;
993 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
994 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
995 }
996 
997 static int
998 svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
999 {
1000 	struct vmcb_ctrl *ctrl;
1001 	int oldval, newval;
1002 
1003 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1004 	oldval = ctrl->intr_shadow;
1005 	newval = val ? 1 : 0;
1006 	if (newval != oldval) {
1007 		ctrl->intr_shadow = newval;
1008 		VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1009 	}
1010 	return (0);
1011 }
1012 
1013 static int
1014 svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1015 {
1016 	struct vmcb_ctrl *ctrl;
1017 
1018 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1019 	*val = ctrl->intr_shadow;
1020 	return (0);
1021 }
1022 
1023 /*
1024  * Once an NMI is injected it blocks delivery of further NMIs until the handler
1025  * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1026  * to track when the vcpu is done handling the NMI.
1027  */
1028 static int
1029 nmi_blocked(struct svm_softc *sc, int vcpu)
1030 {
1031 	int blocked;
1032 
1033 	blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1034 	    VMCB_INTCPT_IRET);
1035 	return (blocked);
1036 }
1037 
1038 static void
1039 enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1040 {
1041 
1042 	KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1043 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1044 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1045 }
1046 
1047 static void
1048 clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1049 {
1050 	int error;
1051 
1052 	KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1053 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1054 	/*
1055 	 * When the IRET intercept is cleared the vcpu will attempt to execute
1056 	 * the "iret" when it runs next. However, it is possible to inject
1057 	 * another NMI into the vcpu before the "iret" has actually executed.
1058 	 *
1059 	 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1060 	 * it will trap back into the hypervisor. If an NMI is pending for
1061 	 * the vcpu it will be injected into the guest.
1062 	 *
1063 	 * XXX this needs to be fixed
1064 	 */
1065 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1066 
1067 	/*
1068 	 * Set 'intr_shadow' to prevent an NMI from being injected on the
1069 	 * immediate VMRUN.
1070 	 */
1071 	error = svm_modify_intr_shadow(sc, vcpu, 1);
1072 	KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1073 }
1074 
1075 static int
1076 emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1077     bool *retu)
1078 {
1079 	int error;
1080 
1081 	if (lapic_msr(num))
1082 		error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1083 	else if (num == MSR_EFER)
1084 		error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, val);
1085 	else
1086 		error = svm_wrmsr(sc, vcpu, num, val, retu);
1087 
1088 	return (error);
1089 }
1090 
1091 static int
1092 emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1093 {
1094 	struct vmcb_state *state;
1095 	struct svm_regctx *ctx;
1096 	uint64_t result;
1097 	int error;
1098 
1099 	if (lapic_msr(num))
1100 		error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1101 	else
1102 		error = svm_rdmsr(sc, vcpu, num, &result, retu);
1103 
1104 	if (error == 0) {
1105 		state = svm_get_vmcb_state(sc, vcpu);
1106 		ctx = svm_get_guest_regctx(sc, vcpu);
1107 		state->rax = result & 0xffffffff;
1108 		ctx->sctx_rdx = result >> 32;
1109 	}
1110 
1111 	return (error);
1112 }
1113 
1114 #ifdef KTR
1115 static const char *
1116 exit_reason_to_str(uint64_t reason)
1117 {
1118 	static char reasonbuf[32];
1119 
1120 	switch (reason) {
1121 	case VMCB_EXIT_INVALID:
1122 		return ("invalvmcb");
1123 	case VMCB_EXIT_SHUTDOWN:
1124 		return ("shutdown");
1125 	case VMCB_EXIT_NPF:
1126 		return ("nptfault");
1127 	case VMCB_EXIT_PAUSE:
1128 		return ("pause");
1129 	case VMCB_EXIT_HLT:
1130 		return ("hlt");
1131 	case VMCB_EXIT_CPUID:
1132 		return ("cpuid");
1133 	case VMCB_EXIT_IO:
1134 		return ("inout");
1135 	case VMCB_EXIT_MC:
1136 		return ("mchk");
1137 	case VMCB_EXIT_INTR:
1138 		return ("extintr");
1139 	case VMCB_EXIT_NMI:
1140 		return ("nmi");
1141 	case VMCB_EXIT_VINTR:
1142 		return ("vintr");
1143 	case VMCB_EXIT_MSR:
1144 		return ("msr");
1145 	case VMCB_EXIT_IRET:
1146 		return ("iret");
1147 	case VMCB_EXIT_MONITOR:
1148 		return ("monitor");
1149 	case VMCB_EXIT_MWAIT:
1150 		return ("mwait");
1151 	default:
1152 		snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1153 		return (reasonbuf);
1154 	}
1155 }
1156 #endif	/* KTR */
1157 
1158 /*
1159  * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1160  * that are due to instruction intercepts as well as MSR and IOIO intercepts
1161  * and exceptions caused by INT3, INTO and BOUND instructions.
1162  *
1163  * Return 1 if the nRIP is valid and 0 otherwise.
1164  */
1165 static int
1166 nrip_valid(uint64_t exitcode)
1167 {
1168 	switch (exitcode) {
1169 	case 0x00 ... 0x0F:	/* read of CR0 through CR15 */
1170 	case 0x10 ... 0x1F:	/* write of CR0 through CR15 */
1171 	case 0x20 ... 0x2F:	/* read of DR0 through DR15 */
1172 	case 0x30 ... 0x3F:	/* write of DR0 through DR15 */
1173 	case 0x43:		/* INT3 */
1174 	case 0x44:		/* INTO */
1175 	case 0x45:		/* BOUND */
1176 	case 0x65 ... 0x7C:	/* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1177 	case 0x80 ... 0x8D:	/* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1178 		return (1);
1179 	default:
1180 		return (0);
1181 	}
1182 }
1183 
1184 /*
1185  * Collateral for a generic SVM VM-exit.
1186  */
1187 static void
1188 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
1189 {
1190 
1191 	vme->exitcode = VM_EXITCODE_SVM;
1192 	vme->u.svm.exitcode = code;
1193 	vme->u.svm.exitinfo1 = info1;
1194 	vme->u.svm.exitinfo2 = info2;
1195 }
1196 
1197 static int
1198 svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1199 {
1200 	struct vmcb *vmcb;
1201 	struct vmcb_state *state;
1202 	struct vmcb_ctrl *ctrl;
1203 	struct svm_regctx *ctx;
1204 	struct vm_exception exception;
1205 	uint64_t code, info1, info2, val;
1206 	uint32_t eax, ecx, edx;
1207 	int error, errcode_valid, handled, idtvec, reflect;
1208 	bool retu;
1209 
1210 	ctx = svm_get_guest_regctx(svm_sc, vcpu);
1211 	vmcb = svm_get_vmcb(svm_sc, vcpu);
1212 	state = &vmcb->state;
1213 	ctrl = &vmcb->ctrl;
1214 
1215 	handled = 0;
1216 	code = ctrl->exitcode;
1217 	info1 = ctrl->exitinfo1;
1218 	info2 = ctrl->exitinfo2;
1219 
1220 	vmexit->exitcode = VM_EXITCODE_BOGUS;
1221 	vmexit->rip = state->rip;
1222 	vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1223 
1224 	vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1225 
1226 	/*
1227 	 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1228 	 * in an inconsistent state and can trigger assertions that would
1229 	 * never happen otherwise.
1230 	 */
1231 	if (code == VMCB_EXIT_INVALID) {
1232 		vm_exit_svm(vmexit, code, info1, info2);
1233 		return (0);
1234 	}
1235 
1236 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1237 	    "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1238 
1239 	KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1240 	    ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1241 	    vmexit->inst_length, code, info1, info2));
1242 
1243 	svm_update_virqinfo(svm_sc, vcpu);
1244 	svm_save_intinfo(svm_sc, vcpu);
1245 
1246 	switch (code) {
1247 	case VMCB_EXIT_IRET:
1248 		/*
1249 		 * Restart execution at "iret" but with the intercept cleared.
1250 		 */
1251 		vmexit->inst_length = 0;
1252 		clear_nmi_blocking(svm_sc, vcpu);
1253 		handled = 1;
1254 		break;
1255 	case VMCB_EXIT_VINTR:	/* interrupt window exiting */
1256 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1257 		handled = 1;
1258 		break;
1259 	case VMCB_EXIT_INTR:	/* external interrupt */
1260 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1261 		handled = 1;
1262 		break;
1263 	case VMCB_EXIT_NMI:	/* external NMI */
1264 		handled = 1;
1265 		break;
1266 	case 0x40 ... 0x5F:
1267 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1268 		reflect = 1;
1269 		idtvec = code - 0x40;
1270 		switch (idtvec) {
1271 		case IDT_MC:
1272 			/*
1273 			 * Call the machine check handler by hand. Also don't
1274 			 * reflect the machine check back into the guest.
1275 			 */
1276 			reflect = 0;
1277 			VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler");
1278 			__asm __volatile("int $18");
1279 			break;
1280 		case IDT_PF:
1281 			error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2,
1282 			    info2);
1283 			KASSERT(error == 0, ("%s: error %d updating cr2",
1284 			    __func__, error));
1285 			/* fallthru */
1286 		case IDT_NP:
1287 		case IDT_SS:
1288 		case IDT_GP:
1289 		case IDT_AC:
1290 		case IDT_TS:
1291 			errcode_valid = 1;
1292 			break;
1293 
1294 		case IDT_DF:
1295 			errcode_valid = 1;
1296 			info1 = 0;
1297 			break;
1298 
1299 		case IDT_BP:
1300 		case IDT_OF:
1301 		case IDT_BR:
1302 			/*
1303 			 * The 'nrip' field is populated for INT3, INTO and
1304 			 * BOUND exceptions and this also implies that
1305 			 * 'inst_length' is non-zero.
1306 			 *
1307 			 * Reset 'inst_length' to zero so the guest %rip at
1308 			 * event injection is identical to what it was when
1309 			 * the exception originally happened.
1310 			 */
1311 			VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d "
1312 			    "to zero before injecting exception %d",
1313 			    vmexit->inst_length, idtvec);
1314 			vmexit->inst_length = 0;
1315 			/* fallthru */
1316 		default:
1317 			errcode_valid = 0;
1318 			break;
1319 		}
1320 		KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1321 		    "when reflecting exception %d into guest",
1322 		    vmexit->inst_length, idtvec));
1323 
1324 		if (reflect) {
1325 			/* Reflect the exception back into the guest */
1326 			bzero(&exception, sizeof(struct vm_exception));
1327 			exception.vector = idtvec;
1328 			if (errcode_valid) {
1329 				exception.error_code = info1;
1330 				exception.error_code_valid = 1;
1331 			}
1332 			VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception "
1333 			    "%d/%#x into the guest", exception.vector,
1334 			    exception.error_code);
1335 			error = vm_inject_exception(svm_sc->vm, vcpu,
1336 			    &exception);
1337 			KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1338 			    __func__, error));
1339 		}
1340 		handled = 1;
1341 		break;
1342 	case VMCB_EXIT_MSR:	/* MSR access. */
1343 		eax = state->rax;
1344 		ecx = ctx->sctx_rcx;
1345 		edx = ctx->sctx_rdx;
1346 		retu = false;
1347 
1348 		if (info1) {
1349 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1350 			val = (uint64_t)edx << 32 | eax;
1351 			VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1352 			    ecx, val);
1353 			if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1354 				vmexit->exitcode = VM_EXITCODE_WRMSR;
1355 				vmexit->u.msr.code = ecx;
1356 				vmexit->u.msr.wval = val;
1357 			} else if (!retu) {
1358 				handled = 1;
1359 			} else {
1360 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1361 				    ("emulate_wrmsr retu with bogus exitcode"));
1362 			}
1363 		} else {
1364 			VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1365 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1366 			if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1367 				vmexit->exitcode = VM_EXITCODE_RDMSR;
1368 				vmexit->u.msr.code = ecx;
1369 			} else if (!retu) {
1370 				handled = 1;
1371 			} else {
1372 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1373 				    ("emulate_rdmsr retu with bogus exitcode"));
1374 			}
1375 		}
1376 		break;
1377 	case VMCB_EXIT_IO:
1378 		handled = svm_handle_io(svm_sc, vcpu, vmexit);
1379 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1380 		break;
1381 	case VMCB_EXIT_CPUID:
1382 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1383 		handled = x86_emulate_cpuid(svm_sc->vm, vcpu,
1384 		    (uint32_t *)&state->rax,
1385 		    (uint32_t *)&ctx->sctx_rbx,
1386 		    (uint32_t *)&ctx->sctx_rcx,
1387 		    (uint32_t *)&ctx->sctx_rdx);
1388 		break;
1389 	case VMCB_EXIT_HLT:
1390 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1391 		vmexit->exitcode = VM_EXITCODE_HLT;
1392 		vmexit->u.hlt.rflags = state->rflags;
1393 		break;
1394 	case VMCB_EXIT_PAUSE:
1395 		vmexit->exitcode = VM_EXITCODE_PAUSE;
1396 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1397 		break;
1398 	case VMCB_EXIT_NPF:
1399 		/* EXITINFO2 contains the faulting guest physical address */
1400 		if (info1 & VMCB_NPF_INFO1_RSV) {
1401 			VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1402 			    "reserved bits set: info1(%#lx) info2(%#lx)",
1403 			    info1, info2);
1404 		} else if (vm_mem_allocated(svm_sc->vm, info2)) {
1405 			vmexit->exitcode = VM_EXITCODE_PAGING;
1406 			vmexit->u.paging.gpa = info2;
1407 			vmexit->u.paging.fault_type = npf_fault_type(info1);
1408 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1409 			VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1410 			    "on gpa %#lx/%#lx at rip %#lx",
1411 			    info2, info1, state->rip);
1412 		} else if (svm_npf_emul_fault(info1)) {
1413 			svm_handle_inst_emul(vmcb, info2, vmexit);
1414 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1415 			VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1416 			    "for gpa %#lx/%#lx at rip %#lx",
1417 			    info2, info1, state->rip);
1418 		}
1419 		break;
1420 	case VMCB_EXIT_MONITOR:
1421 		vmexit->exitcode = VM_EXITCODE_MONITOR;
1422 		break;
1423 	case VMCB_EXIT_MWAIT:
1424 		vmexit->exitcode = VM_EXITCODE_MWAIT;
1425 		break;
1426 	default:
1427 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1428 		break;
1429 	}
1430 
1431 	VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1432 	    handled ? "handled" : "unhandled", exit_reason_to_str(code),
1433 	    vmexit->rip, vmexit->inst_length);
1434 
1435 	if (handled) {
1436 		vmexit->rip += vmexit->inst_length;
1437 		vmexit->inst_length = 0;
1438 		state->rip = vmexit->rip;
1439 	} else {
1440 		if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1441 			/*
1442 			 * If this VM exit was not claimed by anybody then
1443 			 * treat it as a generic SVM exit.
1444 			 */
1445 			vm_exit_svm(vmexit, code, info1, info2);
1446 		} else {
1447 			/*
1448 			 * The exitcode and collateral have been populated.
1449 			 * The VM exit will be processed further in userland.
1450 			 */
1451 		}
1452 	}
1453 	return (handled);
1454 }
1455 
1456 static void
1457 svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1458 {
1459 	uint64_t intinfo;
1460 
1461 	if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1462 		return;
1463 
1464 	KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1465 	    "valid: %#lx", __func__, intinfo));
1466 
1467 	svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1468 		VMCB_EXITINTINFO_VECTOR(intinfo),
1469 		VMCB_EXITINTINFO_EC(intinfo),
1470 		VMCB_EXITINTINFO_EC_VALID(intinfo));
1471 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1472 	VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1473 }
1474 
1475 /*
1476  * Inject event to virtual cpu.
1477  */
1478 static void
1479 svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1480 {
1481 	struct vmcb_ctrl *ctrl;
1482 	struct vmcb_state *state;
1483 	struct svm_vcpu *vcpustate;
1484 	uint8_t v_tpr;
1485 	int vector, need_intr_window, pending_apic_vector;
1486 
1487 	state = svm_get_vmcb_state(sc, vcpu);
1488 	ctrl  = svm_get_vmcb_ctrl(sc, vcpu);
1489 	vcpustate = svm_get_vcpu(sc, vcpu);
1490 
1491 	need_intr_window = 0;
1492 	pending_apic_vector = 0;
1493 
1494 	if (vcpustate->nextrip != state->rip) {
1495 		ctrl->intr_shadow = 0;
1496 		VCPU_CTR2(sc->vm, vcpu, "Guest interrupt blocking "
1497 		    "cleared due to rip change: %#lx/%#lx",
1498 		    vcpustate->nextrip, state->rip);
1499 	}
1500 
1501 	/*
1502 	 * Inject pending events or exceptions for this vcpu.
1503 	 *
1504 	 * An event might be pending because the previous #VMEXIT happened
1505 	 * during event delivery (i.e. ctrl->exitintinfo).
1506 	 *
1507 	 * An event might also be pending because an exception was injected
1508 	 * by the hypervisor (e.g. #PF during instruction emulation).
1509 	 */
1510 	svm_inj_intinfo(sc, vcpu);
1511 
1512 	/* NMI event has priority over interrupts. */
1513 	if (vm_nmi_pending(sc->vm, vcpu)) {
1514 		if (nmi_blocked(sc, vcpu)) {
1515 			/*
1516 			 * Can't inject another NMI if the guest has not
1517 			 * yet executed an "iret" after the last NMI.
1518 			 */
1519 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1520 			    "to NMI-blocking");
1521 		} else if (ctrl->intr_shadow) {
1522 			/*
1523 			 * Can't inject an NMI if the vcpu is in an intr_shadow.
1524 			 */
1525 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1526 			    "interrupt shadow");
1527 			need_intr_window = 1;
1528 			goto done;
1529 		} else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1530 			/*
1531 			 * If there is already an exception/interrupt pending
1532 			 * then defer the NMI until after that.
1533 			 */
1534 			VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1535 			    "eventinj %#lx", ctrl->eventinj);
1536 
1537 			/*
1538 			 * Use self-IPI to trigger a VM-exit as soon as
1539 			 * possible after the event injection is completed.
1540 			 *
1541 			 * This works only if the external interrupt exiting
1542 			 * is at a lower priority than the event injection.
1543 			 *
1544 			 * Although not explicitly specified in APMv2 the
1545 			 * relative priorities were verified empirically.
1546 			 */
1547 			ipi_cpu(curcpu, IPI_AST);	/* XXX vmm_ipinum? */
1548 		} else {
1549 			vm_nmi_clear(sc->vm, vcpu);
1550 
1551 			/* Inject NMI, vector number is not used */
1552 			svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1553 			    IDT_NMI, 0, false);
1554 
1555 			/* virtual NMI blocking is now in effect */
1556 			enable_nmi_blocking(sc, vcpu);
1557 
1558 			VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1559 		}
1560 	}
1561 
1562 	if (!vm_extint_pending(sc->vm, vcpu)) {
1563 		/*
1564 		 * APIC interrupts are delivered using the V_IRQ offload.
1565 		 *
1566 		 * The primary benefit is that the hypervisor doesn't need to
1567 		 * deal with the various conditions that inhibit interrupts.
1568 		 * It also means that TPR changes via CR8 will be handled
1569 		 * without any hypervisor involvement.
1570 		 *
1571 		 * Note that the APIC vector must remain pending in the vIRR
1572 		 * until it is confirmed that it was delivered to the guest.
1573 		 * This can be confirmed based on the value of V_IRQ at the
1574 		 * next #VMEXIT (1 = pending, 0 = delivered).
1575 		 *
1576 		 * Also note that it is possible that another higher priority
1577 		 * vector can become pending before this vector is delivered
1578 		 * to the guest. This is alright because vcpu_notify_event()
1579 		 * will send an IPI and force the vcpu to trap back into the
1580 		 * hypervisor. The higher priority vector will be injected on
1581 		 * the next VMRUN.
1582 		 */
1583 		if (vlapic_pending_intr(vlapic, &vector)) {
1584 			KASSERT(vector >= 16 && vector <= 255,
1585 			    ("invalid vector %d from local APIC", vector));
1586 			pending_apic_vector = vector;
1587 		}
1588 		goto done;
1589 	}
1590 
1591 	/* Ask the legacy pic for a vector to inject */
1592 	vatpic_pending_intr(sc->vm, &vector);
1593 	KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR",
1594 	    vector));
1595 
1596 	/*
1597 	 * If the guest has disabled interrupts or is in an interrupt shadow
1598 	 * then we cannot inject the pending interrupt.
1599 	 */
1600 	if ((state->rflags & PSL_I) == 0) {
1601 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1602 		    "rflags %#lx", vector, state->rflags);
1603 		need_intr_window = 1;
1604 		goto done;
1605 	}
1606 
1607 	if (ctrl->intr_shadow) {
1608 		VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1609 		    "interrupt shadow", vector);
1610 		need_intr_window = 1;
1611 		goto done;
1612 	}
1613 
1614 	if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1615 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1616 		    "eventinj %#lx", vector, ctrl->eventinj);
1617 		need_intr_window = 1;
1618 		goto done;
1619 	}
1620 
1621 	/*
1622 	 * Legacy PIC interrupts are delivered via the event injection
1623 	 * mechanism.
1624 	 */
1625 	svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1626 
1627 	vm_extint_clear(sc->vm, vcpu);
1628 	vatpic_intr_accepted(sc->vm, vector);
1629 
1630 	/*
1631 	 * Force a VM-exit as soon as the vcpu is ready to accept another
1632 	 * interrupt. This is done because the PIC might have another vector
1633 	 * that it wants to inject. Also, if the APIC has a pending interrupt
1634 	 * that was preempted by the ExtInt then it allows us to inject the
1635 	 * APIC vector as soon as possible.
1636 	 */
1637 	need_intr_window = 1;
1638 done:
1639 	/*
1640 	 * The guest can modify the TPR by writing to %CR8. In guest mode
1641 	 * the processor reflects this write to V_TPR without hypervisor
1642 	 * intervention.
1643 	 *
1644 	 * The guest can also modify the TPR by writing to it via the memory
1645 	 * mapped APIC page. In this case, the write will be emulated by the
1646 	 * hypervisor. For this reason V_TPR must be updated before every
1647 	 * VMRUN.
1648 	 */
1649 	v_tpr = vlapic_get_cr8(vlapic);
1650 	KASSERT(v_tpr >= 0 && v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1651 	if (ctrl->v_tpr != v_tpr) {
1652 		VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1653 		    ctrl->v_tpr, v_tpr);
1654 		ctrl->v_tpr = v_tpr;
1655 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1656 	}
1657 
1658 	if (pending_apic_vector) {
1659 		/*
1660 		 * If an APIC vector is being injected then interrupt window
1661 		 * exiting is not possible on this VMRUN.
1662 		 */
1663 		KASSERT(!need_intr_window, ("intr_window exiting impossible"));
1664 		VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ",
1665 		    pending_apic_vector);
1666 
1667 		ctrl->v_irq = 1;
1668 		ctrl->v_ign_tpr = 0;
1669 		ctrl->v_intr_vector = pending_apic_vector;
1670 		ctrl->v_intr_prio = pending_apic_vector >> 4;
1671 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1672 	} else if (need_intr_window) {
1673 		/*
1674 		 * We use V_IRQ in conjunction with the VINTR intercept to
1675 		 * trap into the hypervisor as soon as a virtual interrupt
1676 		 * can be delivered.
1677 		 *
1678 		 * Since injected events are not subject to intercept checks
1679 		 * we need to ensure that the V_IRQ is not actually going to
1680 		 * be delivered on VM entry. The KASSERT below enforces this.
1681 		 */
1682 		KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1683 		    (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1684 		    ("Bogus intr_window_exiting: eventinj (%#lx), "
1685 		    "intr_shadow (%u), rflags (%#lx)",
1686 		    ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1687 		enable_intr_window_exiting(sc, vcpu);
1688 	} else {
1689 		disable_intr_window_exiting(sc, vcpu);
1690 	}
1691 }
1692 
1693 static __inline void
1694 restore_host_tss(void)
1695 {
1696 	struct system_segment_descriptor *tss_sd;
1697 
1698 	/*
1699 	 * The TSS descriptor was in use prior to launching the guest so it
1700 	 * has been marked busy.
1701 	 *
1702 	 * 'ltr' requires the descriptor to be marked available so change the
1703 	 * type to "64-bit available TSS".
1704 	 */
1705 	tss_sd = PCPU_GET(tss);
1706 	tss_sd->sd_type = SDT_SYSTSS;
1707 	ltr(GSEL(GPROC0_SEL, SEL_KPL));
1708 }
1709 
1710 static void
1711 check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu)
1712 {
1713 	struct svm_vcpu *vcpustate;
1714 	struct vmcb_ctrl *ctrl;
1715 	long eptgen;
1716 	bool alloc_asid;
1717 
1718 	KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not "
1719 	    "active on cpu %u", __func__, thiscpu));
1720 
1721 	vcpustate = svm_get_vcpu(sc, vcpuid);
1722 	ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1723 
1724 	/*
1725 	 * The TLB entries associated with the vcpu's ASID are not valid
1726 	 * if either of the following conditions is true:
1727 	 *
1728 	 * 1. The vcpu's ASID generation is different than the host cpu's
1729 	 *    ASID generation. This happens when the vcpu migrates to a new
1730 	 *    host cpu. It can also happen when the number of vcpus executing
1731 	 *    on a host cpu is greater than the number of ASIDs available.
1732 	 *
1733 	 * 2. The pmap generation number is different than the value cached in
1734 	 *    the 'vcpustate'. This happens when the host invalidates pages
1735 	 *    belonging to the guest.
1736 	 *
1737 	 *	asidgen		eptgen	      Action
1738 	 *	mismatch	mismatch
1739 	 *	   0		   0		(a)
1740 	 *	   0		   1		(b1) or (b2)
1741 	 *	   1		   0		(c)
1742 	 *	   1		   1		(d)
1743 	 *
1744 	 * (a) There is no mismatch in eptgen or ASID generation and therefore
1745 	 *     no further action is needed.
1746 	 *
1747 	 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1748 	 *      retained and the TLB entries associated with this ASID
1749 	 *      are flushed by VMRUN.
1750 	 *
1751 	 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1752 	 *      allocated.
1753 	 *
1754 	 * (c) A new ASID is allocated.
1755 	 *
1756 	 * (d) A new ASID is allocated.
1757 	 */
1758 
1759 	alloc_asid = false;
1760 	eptgen = pmap->pm_eptgen;
1761 	ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1762 
1763 	if (vcpustate->asid.gen != asid[thiscpu].gen) {
1764 		alloc_asid = true;	/* (c) and (d) */
1765 	} else if (vcpustate->eptgen != eptgen) {
1766 		if (flush_by_asid())
1767 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;	/* (b1) */
1768 		else
1769 			alloc_asid = true;			/* (b2) */
1770 	} else {
1771 		/*
1772 		 * This is the common case (a).
1773 		 */
1774 		KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1775 		KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1776 		    ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1777 	}
1778 
1779 	if (alloc_asid) {
1780 		if (++asid[thiscpu].num >= nasid) {
1781 			asid[thiscpu].num = 1;
1782 			if (++asid[thiscpu].gen == 0)
1783 				asid[thiscpu].gen = 1;
1784 			/*
1785 			 * If this cpu does not support "flush-by-asid"
1786 			 * then flush the entire TLB on a generation
1787 			 * bump. Subsequent ASID allocation in this
1788 			 * generation can be done without a TLB flush.
1789 			 */
1790 			if (!flush_by_asid())
1791 				ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1792 		}
1793 		vcpustate->asid.gen = asid[thiscpu].gen;
1794 		vcpustate->asid.num = asid[thiscpu].num;
1795 
1796 		ctrl->asid = vcpustate->asid.num;
1797 		svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1798 		/*
1799 		 * If this cpu supports "flush-by-asid" then the TLB
1800 		 * was not flushed after the generation bump. The TLB
1801 		 * is flushed selectively after every new ASID allocation.
1802 		 */
1803 		if (flush_by_asid())
1804 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1805 	}
1806 	vcpustate->eptgen = eptgen;
1807 
1808 	KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1809 	KASSERT(ctrl->asid == vcpustate->asid.num,
1810 	    ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1811 }
1812 
1813 static __inline void
1814 disable_gintr(void)
1815 {
1816 
1817         __asm __volatile("clgi" : : :);
1818 }
1819 
1820 static __inline void
1821 enable_gintr(void)
1822 {
1823 
1824         __asm __volatile("stgi" : : :);
1825 }
1826 
1827 /*
1828  * Start vcpu with specified RIP.
1829  */
1830 static int
1831 svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap,
1832 	void *rend_cookie, void *suspended_cookie)
1833 {
1834 	struct svm_regctx *gctx;
1835 	struct svm_softc *svm_sc;
1836 	struct svm_vcpu *vcpustate;
1837 	struct vmcb_state *state;
1838 	struct vmcb_ctrl *ctrl;
1839 	struct vm_exit *vmexit;
1840 	struct vlapic *vlapic;
1841 	struct vm *vm;
1842 	uint64_t vmcb_pa;
1843 	u_int thiscpu;
1844 	int handled;
1845 
1846 	svm_sc = arg;
1847 	vm = svm_sc->vm;
1848 
1849 	vcpustate = svm_get_vcpu(svm_sc, vcpu);
1850 	state = svm_get_vmcb_state(svm_sc, vcpu);
1851 	ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
1852 	vmexit = vm_exitinfo(vm, vcpu);
1853 	vlapic = vm_lapic(vm, vcpu);
1854 
1855 	/*
1856 	 * Stash 'curcpu' on the stack as 'thiscpu'.
1857 	 *
1858 	 * The per-cpu data area is not accessible until MSR_GSBASE is restored
1859 	 * after the #VMEXIT. Since VMRUN is executed inside a critical section
1860 	 * 'curcpu' and 'thiscpu' are guaranteed to identical.
1861 	 */
1862 	thiscpu = curcpu;
1863 
1864 	gctx = svm_get_guest_regctx(svm_sc, vcpu);
1865 	vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
1866 
1867 	if (vcpustate->lastcpu != thiscpu) {
1868 		/*
1869 		 * Force new ASID allocation by invalidating the generation.
1870 		 */
1871 		vcpustate->asid.gen = 0;
1872 
1873 		/*
1874 		 * Invalidate the VMCB state cache by marking all fields dirty.
1875 		 */
1876 		svm_set_dirty(svm_sc, vcpu, 0xffffffff);
1877 
1878 		/*
1879 		 * XXX
1880 		 * Setting 'vcpustate->lastcpu' here is bit premature because
1881 		 * we may return from this function without actually executing
1882 		 * the VMRUN  instruction. This could happen if a rendezvous
1883 		 * or an AST is pending on the first time through the loop.
1884 		 *
1885 		 * This works for now but any new side-effects of vcpu
1886 		 * migration should take this case into account.
1887 		 */
1888 		vcpustate->lastcpu = thiscpu;
1889 		vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
1890 	}
1891 
1892 	svm_msr_guest_enter(svm_sc, vcpu);
1893 
1894 	/* Update Guest RIP */
1895 	state->rip = rip;
1896 
1897 	do {
1898 		/*
1899 		 * Disable global interrupts to guarantee atomicity during
1900 		 * loading of guest state. This includes not only the state
1901 		 * loaded by the "vmrun" instruction but also software state
1902 		 * maintained by the hypervisor: suspended and rendezvous
1903 		 * state, NPT generation number, vlapic interrupts etc.
1904 		 */
1905 		disable_gintr();
1906 
1907 		if (vcpu_suspended(suspended_cookie)) {
1908 			enable_gintr();
1909 			vm_exit_suspended(vm, vcpu, state->rip);
1910 			break;
1911 		}
1912 
1913 		if (vcpu_rendezvous_pending(rend_cookie)) {
1914 			enable_gintr();
1915 			vm_exit_rendezvous(vm, vcpu, state->rip);
1916 			break;
1917 		}
1918 
1919 		/* We are asked to give the cpu by scheduler. */
1920 		if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) {
1921 			enable_gintr();
1922 			vm_exit_astpending(vm, vcpu, state->rip);
1923 			break;
1924 		}
1925 
1926 		svm_inj_interrupts(svm_sc, vcpu, vlapic);
1927 
1928 		/* Activate the nested pmap on 'thiscpu' */
1929 		CPU_SET_ATOMIC_ACQ(thiscpu, &pmap->pm_active);
1930 
1931 		/*
1932 		 * Check the pmap generation and the ASID generation to
1933 		 * ensure that the vcpu does not use stale TLB mappings.
1934 		 */
1935 		check_asid(svm_sc, vcpu, pmap, thiscpu);
1936 
1937 		ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
1938 		vcpustate->dirty = 0;
1939 		VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
1940 
1941 		/* Launch Virtual Machine. */
1942 		VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
1943 		svm_launch(vmcb_pa, gctx);
1944 
1945 		CPU_CLR_ATOMIC(thiscpu, &pmap->pm_active);
1946 
1947 		/*
1948 		 * Restore MSR_GSBASE to point to the pcpu data area.
1949 		 *
1950 		 * Note that accesses done via PCPU_GET/PCPU_SET will work
1951 		 * only after MSR_GSBASE is restored.
1952 		 *
1953 		 * Also note that we don't bother restoring MSR_KGSBASE
1954 		 * since it is not used in the kernel and will be restored
1955 		 * when the VMRUN ioctl returns to userspace.
1956 		 */
1957 		wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[thiscpu]);
1958 		KASSERT(curcpu == thiscpu, ("thiscpu/curcpu (%u/%u) mismatch",
1959 		    thiscpu, curcpu));
1960 
1961 		/*
1962 		 * The host GDTR and IDTR is saved by VMRUN and restored
1963 		 * automatically on #VMEXIT. However, the host TSS needs
1964 		 * to be restored explicitly.
1965 		 */
1966 		restore_host_tss();
1967 
1968 		/* #VMEXIT disables interrupts so re-enable them here. */
1969 		enable_gintr();
1970 
1971 		/* Update 'nextrip' */
1972 		vcpustate->nextrip = state->rip;
1973 
1974 		/* Handle #VMEXIT and if required return to user space. */
1975 		handled = svm_vmexit(svm_sc, vcpu, vmexit);
1976 	} while (handled);
1977 
1978 	svm_msr_guest_exit(svm_sc, vcpu);
1979 
1980 	return (0);
1981 }
1982 
1983 static void
1984 svm_vmcleanup(void *arg)
1985 {
1986 	struct svm_softc *sc = arg;
1987 
1988 	free(sc, M_SVM);
1989 }
1990 
1991 static register_t *
1992 swctx_regptr(struct svm_regctx *regctx, int reg)
1993 {
1994 
1995 	switch (reg) {
1996 	case VM_REG_GUEST_RBX:
1997 		return (&regctx->sctx_rbx);
1998 	case VM_REG_GUEST_RCX:
1999 		return (&regctx->sctx_rcx);
2000 	case VM_REG_GUEST_RDX:
2001 		return (&regctx->sctx_rdx);
2002 	case VM_REG_GUEST_RDI:
2003 		return (&regctx->sctx_rdi);
2004 	case VM_REG_GUEST_RSI:
2005 		return (&regctx->sctx_rsi);
2006 	case VM_REG_GUEST_RBP:
2007 		return (&regctx->sctx_rbp);
2008 	case VM_REG_GUEST_R8:
2009 		return (&regctx->sctx_r8);
2010 	case VM_REG_GUEST_R9:
2011 		return (&regctx->sctx_r9);
2012 	case VM_REG_GUEST_R10:
2013 		return (&regctx->sctx_r10);
2014 	case VM_REG_GUEST_R11:
2015 		return (&regctx->sctx_r11);
2016 	case VM_REG_GUEST_R12:
2017 		return (&regctx->sctx_r12);
2018 	case VM_REG_GUEST_R13:
2019 		return (&regctx->sctx_r13);
2020 	case VM_REG_GUEST_R14:
2021 		return (&regctx->sctx_r14);
2022 	case VM_REG_GUEST_R15:
2023 		return (&regctx->sctx_r15);
2024 	default:
2025 		return (NULL);
2026 	}
2027 }
2028 
2029 static int
2030 svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
2031 {
2032 	struct svm_softc *svm_sc;
2033 	register_t *reg;
2034 
2035 	svm_sc = arg;
2036 
2037 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2038 		return (svm_get_intr_shadow(svm_sc, vcpu, val));
2039 	}
2040 
2041 	if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2042 		return (0);
2043 	}
2044 
2045 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2046 
2047 	if (reg != NULL) {
2048 		*val = *reg;
2049 		return (0);
2050 	}
2051 
2052 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident);
2053 	return (EINVAL);
2054 }
2055 
2056 static int
2057 svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2058 {
2059 	struct svm_softc *svm_sc;
2060 	register_t *reg;
2061 
2062 	svm_sc = arg;
2063 
2064 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2065 		return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2066 	}
2067 
2068 	if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2069 		return (0);
2070 	}
2071 
2072 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2073 
2074 	if (reg != NULL) {
2075 		*reg = val;
2076 		return (0);
2077 	}
2078 
2079 	/*
2080 	 * XXX deal with CR3 and invalidate TLB entries tagged with the
2081 	 * vcpu's ASID. This needs to be treated differently depending on
2082 	 * whether 'running' is true/false.
2083 	 */
2084 
2085 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident);
2086 	return (EINVAL);
2087 }
2088 
2089 static int
2090 svm_setcap(void *arg, int vcpu, int type, int val)
2091 {
2092 	struct svm_softc *sc;
2093 	int error;
2094 
2095 	sc = arg;
2096 	error = 0;
2097 	switch (type) {
2098 	case VM_CAP_HALT_EXIT:
2099 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2100 		    VMCB_INTCPT_HLT, val);
2101 		break;
2102 	case VM_CAP_PAUSE_EXIT:
2103 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2104 		    VMCB_INTCPT_PAUSE, val);
2105 		break;
2106 	case VM_CAP_UNRESTRICTED_GUEST:
2107 		/* Unrestricted guest execution cannot be disabled in SVM */
2108 		if (val == 0)
2109 			error = EINVAL;
2110 		break;
2111 	default:
2112 		error = ENOENT;
2113 		break;
2114 	}
2115 	return (error);
2116 }
2117 
2118 static int
2119 svm_getcap(void *arg, int vcpu, int type, int *retval)
2120 {
2121 	struct svm_softc *sc;
2122 	int error;
2123 
2124 	sc = arg;
2125 	error = 0;
2126 
2127 	switch (type) {
2128 	case VM_CAP_HALT_EXIT:
2129 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2130 		    VMCB_INTCPT_HLT);
2131 		break;
2132 	case VM_CAP_PAUSE_EXIT:
2133 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2134 		    VMCB_INTCPT_PAUSE);
2135 		break;
2136 	case VM_CAP_UNRESTRICTED_GUEST:
2137 		*retval = 1;	/* unrestricted guest is always enabled */
2138 		break;
2139 	default:
2140 		error = ENOENT;
2141 		break;
2142 	}
2143 	return (error);
2144 }
2145 
2146 static struct vlapic *
2147 svm_vlapic_init(void *arg, int vcpuid)
2148 {
2149 	struct svm_softc *svm_sc;
2150 	struct vlapic *vlapic;
2151 
2152 	svm_sc = arg;
2153 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2154 	vlapic->vm = svm_sc->vm;
2155 	vlapic->vcpuid = vcpuid;
2156 	vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2157 
2158 	vlapic_init(vlapic);
2159 
2160 	return (vlapic);
2161 }
2162 
2163 static void
2164 svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2165 {
2166 
2167         vlapic_cleanup(vlapic);
2168         free(vlapic, M_SVM_VLAPIC);
2169 }
2170 
2171 struct vmm_ops vmm_ops_amd = {
2172 	svm_init,
2173 	svm_cleanup,
2174 	svm_restore,
2175 	svm_vminit,
2176 	svm_vmrun,
2177 	svm_vmcleanup,
2178 	svm_getreg,
2179 	svm_setreg,
2180 	vmcb_getdesc,
2181 	vmcb_setdesc,
2182 	svm_getcap,
2183 	svm_setcap,
2184 	svm_npt_alloc,
2185 	svm_npt_free,
2186 	svm_vlapic_init,
2187 	svm_vlapic_cleanup
2188 };
2189