xref: /freebsd/sys/amd64/vmm/amd/svm.c (revision a812392203d7c4c3f0db9d8a0f3391374c49c71f)
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->lastcpu = NOCPU;
558 		vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
559 		vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
560 		svm_msr_guest_init(svm_sc, i);
561 	}
562 	return (svm_sc);
563 }
564 
565 static int
566 svm_cpl(struct vmcb_state *state)
567 {
568 
569 	/*
570 	 * From APMv2:
571 	 *   "Retrieve the CPL from the CPL field in the VMCB, not
572 	 *    from any segment DPL"
573 	 */
574 	return (state->cpl);
575 }
576 
577 static enum vm_cpu_mode
578 svm_vcpu_mode(struct vmcb *vmcb)
579 {
580 	struct vmcb_segment seg;
581 	struct vmcb_state *state;
582 	int error;
583 
584 	state = &vmcb->state;
585 
586 	if (state->efer & EFER_LMA) {
587 		error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
588 		KASSERT(error == 0, ("%s: vmcb_seg(cs) error %d", __func__,
589 		    error));
590 
591 		/*
592 		 * Section 4.8.1 for APM2, check if Code Segment has
593 		 * Long attribute set in descriptor.
594 		 */
595 		if (seg.attrib & VMCB_CS_ATTRIB_L)
596 			return (CPU_MODE_64BIT);
597 		else
598 			return (CPU_MODE_COMPATIBILITY);
599 	} else  if (state->cr0 & CR0_PE) {
600 		return (CPU_MODE_PROTECTED);
601 	} else {
602 		return (CPU_MODE_REAL);
603 	}
604 }
605 
606 static enum vm_paging_mode
607 svm_paging_mode(uint64_t cr0, uint64_t cr4, uint64_t efer)
608 {
609 
610 	if ((cr0 & CR0_PG) == 0)
611 		return (PAGING_MODE_FLAT);
612 	if ((cr4 & CR4_PAE) == 0)
613 		return (PAGING_MODE_32);
614 	if (efer & EFER_LME)
615 		return (PAGING_MODE_64);
616 	else
617 		return (PAGING_MODE_PAE);
618 }
619 
620 /*
621  * ins/outs utility routines
622  */
623 static uint64_t
624 svm_inout_str_index(struct svm_regctx *regs, int in)
625 {
626 	uint64_t val;
627 
628 	val = in ? regs->sctx_rdi : regs->sctx_rsi;
629 
630 	return (val);
631 }
632 
633 static uint64_t
634 svm_inout_str_count(struct svm_regctx *regs, int rep)
635 {
636 	uint64_t val;
637 
638 	val = rep ? regs->sctx_rcx : 1;
639 
640 	return (val);
641 }
642 
643 static void
644 svm_inout_str_seginfo(struct svm_softc *svm_sc, int vcpu, int64_t info1,
645     int in, struct vm_inout_str *vis)
646 {
647 	int error, s;
648 
649 	if (in) {
650 		vis->seg_name = VM_REG_GUEST_ES;
651 	} else {
652 		/* The segment field has standard encoding */
653 		s = (info1 >> 10) & 0x7;
654 		vis->seg_name = vm_segment_name(s);
655 	}
656 
657 	error = vmcb_getdesc(svm_sc, vcpu, vis->seg_name, &vis->seg_desc);
658 	KASSERT(error == 0, ("%s: svm_getdesc error %d", __func__, error));
659 }
660 
661 static int
662 svm_inout_str_addrsize(uint64_t info1)
663 {
664         uint32_t size;
665 
666         size = (info1 >> 7) & 0x7;
667         switch (size) {
668         case 1:
669                 return (2);     /* 16 bit */
670         case 2:
671                 return (4);     /* 32 bit */
672         case 4:
673                 return (8);     /* 64 bit */
674         default:
675                 panic("%s: invalid size encoding %d", __func__, size);
676         }
677 }
678 
679 static void
680 svm_paging_info(struct vmcb *vmcb, struct vm_guest_paging *paging)
681 {
682 	struct vmcb_state *state;
683 
684 	state = &vmcb->state;
685 	paging->cr3 = state->cr3;
686 	paging->cpl = svm_cpl(state);
687 	paging->cpu_mode = svm_vcpu_mode(vmcb);
688 	paging->paging_mode = svm_paging_mode(state->cr0, state->cr4,
689 	    state->efer);
690 }
691 
692 #define	UNHANDLED 0
693 
694 /*
695  * Handle guest I/O intercept.
696  */
697 static int
698 svm_handle_io(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
699 {
700 	struct vmcb_ctrl *ctrl;
701 	struct vmcb_state *state;
702 	struct svm_regctx *regs;
703 	struct vm_inout_str *vis;
704 	uint64_t info1;
705 	int inout_string;
706 
707 	state = svm_get_vmcb_state(svm_sc, vcpu);
708 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
709 	regs  = svm_get_guest_regctx(svm_sc, vcpu);
710 
711 	info1 = ctrl->exitinfo1;
712 	inout_string = info1 & BIT(2) ? 1 : 0;
713 
714 	/*
715 	 * The effective segment number in EXITINFO1[12:10] is populated
716 	 * only if the processor has the DecodeAssist capability.
717 	 *
718 	 * XXX this is not specified explicitly in APMv2 but can be verified
719 	 * empirically.
720 	 */
721 	if (inout_string && !decode_assist())
722 		return (UNHANDLED);
723 
724 	vmexit->exitcode 	= VM_EXITCODE_INOUT;
725 	vmexit->u.inout.in 	= (info1 & BIT(0)) ? 1 : 0;
726 	vmexit->u.inout.string 	= inout_string;
727 	vmexit->u.inout.rep 	= (info1 & BIT(3)) ? 1 : 0;
728 	vmexit->u.inout.bytes 	= (info1 >> 4) & 0x7;
729 	vmexit->u.inout.port 	= (uint16_t)(info1 >> 16);
730 	vmexit->u.inout.eax 	= (uint32_t)(state->rax);
731 
732 	if (inout_string) {
733 		vmexit->exitcode = VM_EXITCODE_INOUT_STR;
734 		vis = &vmexit->u.inout_str;
735 		svm_paging_info(svm_get_vmcb(svm_sc, vcpu), &vis->paging);
736 		vis->rflags = state->rflags;
737 		vis->cr0 = state->cr0;
738 		vis->index = svm_inout_str_index(regs, vmexit->u.inout.in);
739 		vis->count = svm_inout_str_count(regs, vmexit->u.inout.rep);
740 		vis->addrsize = svm_inout_str_addrsize(info1);
741 		svm_inout_str_seginfo(svm_sc, vcpu, info1,
742 		    vmexit->u.inout.in, vis);
743 	}
744 
745 	return (UNHANDLED);
746 }
747 
748 static int
749 npf_fault_type(uint64_t exitinfo1)
750 {
751 
752 	if (exitinfo1 & VMCB_NPF_INFO1_W)
753 		return (VM_PROT_WRITE);
754 	else if (exitinfo1 & VMCB_NPF_INFO1_ID)
755 		return (VM_PROT_EXECUTE);
756 	else
757 		return (VM_PROT_READ);
758 }
759 
760 static bool
761 svm_npf_emul_fault(uint64_t exitinfo1)
762 {
763 
764 	if (exitinfo1 & VMCB_NPF_INFO1_ID) {
765 		return (false);
766 	}
767 
768 	if (exitinfo1 & VMCB_NPF_INFO1_GPT) {
769 		return (false);
770 	}
771 
772 	if ((exitinfo1 & VMCB_NPF_INFO1_GPA) == 0) {
773 		return (false);
774 	}
775 
776 	return (true);
777 }
778 
779 static void
780 svm_handle_inst_emul(struct vmcb *vmcb, uint64_t gpa, struct vm_exit *vmexit)
781 {
782 	struct vm_guest_paging *paging;
783 	struct vmcb_segment seg;
784 	struct vmcb_ctrl *ctrl;
785 	char *inst_bytes;
786 	int error, inst_len;
787 
788 	ctrl = &vmcb->ctrl;
789 	paging = &vmexit->u.inst_emul.paging;
790 
791 	vmexit->exitcode = VM_EXITCODE_INST_EMUL;
792 	vmexit->u.inst_emul.gpa = gpa;
793 	vmexit->u.inst_emul.gla = VIE_INVALID_GLA;
794 	svm_paging_info(vmcb, paging);
795 
796 	error = vmcb_seg(vmcb, VM_REG_GUEST_CS, &seg);
797 	KASSERT(error == 0, ("%s: vmcb_seg(CS) error %d", __func__, error));
798 
799 	switch(paging->cpu_mode) {
800 	case CPU_MODE_PROTECTED:
801 	case CPU_MODE_COMPATIBILITY:
802 		/*
803 		 * Section 4.8.1 of APM2, Default Operand Size or D bit.
804 		 */
805 		vmexit->u.inst_emul.cs_d = (seg.attrib & VMCB_CS_ATTRIB_D) ?
806 		    1 : 0;
807 		break;
808 	default:
809 		vmexit->u.inst_emul.cs_d = 0;
810 		break;
811 	}
812 
813 	/*
814 	 * Copy the instruction bytes into 'vie' if available.
815 	 */
816 	if (decode_assist() && !disable_npf_assist) {
817 		inst_len = ctrl->inst_len;
818 		inst_bytes = ctrl->inst_bytes;
819 	} else {
820 		inst_len = 0;
821 		inst_bytes = NULL;
822 	}
823 	vie_init(&vmexit->u.inst_emul.vie, inst_bytes, inst_len);
824 }
825 
826 #ifdef KTR
827 static const char *
828 intrtype_to_str(int intr_type)
829 {
830 	switch (intr_type) {
831 	case VMCB_EVENTINJ_TYPE_INTR:
832 		return ("hwintr");
833 	case VMCB_EVENTINJ_TYPE_NMI:
834 		return ("nmi");
835 	case VMCB_EVENTINJ_TYPE_INTn:
836 		return ("swintr");
837 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
838 		return ("exception");
839 	default:
840 		panic("%s: unknown intr_type %d", __func__, intr_type);
841 	}
842 }
843 #endif
844 
845 /*
846  * Inject an event to vcpu as described in section 15.20, "Event injection".
847  */
848 static void
849 svm_eventinject(struct svm_softc *sc, int vcpu, int intr_type, int vector,
850 		 uint32_t error, bool ec_valid)
851 {
852 	struct vmcb_ctrl *ctrl;
853 
854 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
855 
856 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0,
857 	    ("%s: event already pending %#lx", __func__, ctrl->eventinj));
858 
859 	KASSERT(vector >=0 && vector <= 255, ("%s: invalid vector %d",
860 	    __func__, vector));
861 
862 	switch (intr_type) {
863 	case VMCB_EVENTINJ_TYPE_INTR:
864 	case VMCB_EVENTINJ_TYPE_NMI:
865 	case VMCB_EVENTINJ_TYPE_INTn:
866 		break;
867 	case VMCB_EVENTINJ_TYPE_EXCEPTION:
868 		if (vector >= 0 && vector <= 31 && vector != 2)
869 			break;
870 		/* FALLTHROUGH */
871 	default:
872 		panic("%s: invalid intr_type/vector: %d/%d", __func__,
873 		    intr_type, vector);
874 	}
875 	ctrl->eventinj = vector | (intr_type << 8) | VMCB_EVENTINJ_VALID;
876 	if (ec_valid) {
877 		ctrl->eventinj |= VMCB_EVENTINJ_EC_VALID;
878 		ctrl->eventinj |= (uint64_t)error << 32;
879 		VCPU_CTR3(sc->vm, vcpu, "Injecting %s at vector %d errcode %#x",
880 		    intrtype_to_str(intr_type), vector, error);
881 	} else {
882 		VCPU_CTR2(sc->vm, vcpu, "Injecting %s at vector %d",
883 		    intrtype_to_str(intr_type), vector);
884 	}
885 }
886 
887 static void
888 svm_update_virqinfo(struct svm_softc *sc, int vcpu)
889 {
890 	struct vm *vm;
891 	struct vlapic *vlapic;
892 	struct vmcb_ctrl *ctrl;
893 	int pending;
894 
895 	vm = sc->vm;
896 	vlapic = vm_lapic(vm, vcpu);
897 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
898 
899 	/* Update %cr8 in the emulated vlapic */
900 	vlapic_set_cr8(vlapic, ctrl->v_tpr);
901 
902 	/*
903 	 * If V_IRQ indicates that the interrupt injection attempted on then
904 	 * last VMRUN was successful then update the vlapic accordingly.
905 	 */
906 	if (ctrl->v_intr_vector != 0) {
907 		pending = ctrl->v_irq;
908 		KASSERT(ctrl->v_intr_vector >= 16, ("%s: invalid "
909 		    "v_intr_vector %d", __func__, ctrl->v_intr_vector));
910 		KASSERT(!ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
911 		VCPU_CTR2(vm, vcpu, "v_intr_vector %d %s", ctrl->v_intr_vector,
912 		    pending ? "pending" : "accepted");
913 		if (!pending)
914 			vlapic_intr_accepted(vlapic, ctrl->v_intr_vector);
915 	}
916 }
917 
918 static void
919 svm_save_intinfo(struct svm_softc *svm_sc, int vcpu)
920 {
921 	struct vmcb_ctrl *ctrl;
922 	uint64_t intinfo;
923 
924 	ctrl  = svm_get_vmcb_ctrl(svm_sc, vcpu);
925 	intinfo = ctrl->exitintinfo;
926 	if (!VMCB_EXITINTINFO_VALID(intinfo))
927 		return;
928 
929 	/*
930 	 * From APMv2, Section "Intercepts during IDT interrupt delivery"
931 	 *
932 	 * If a #VMEXIT happened during event delivery then record the event
933 	 * that was being delivered.
934 	 */
935 	VCPU_CTR2(svm_sc->vm, vcpu, "SVM:Pending INTINFO(0x%lx), vector=%d.\n",
936 		intinfo, VMCB_EXITINTINFO_VECTOR(intinfo));
937 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_EXITINTINFO, 1);
938 	vm_exit_intinfo(svm_sc->vm, vcpu, intinfo);
939 }
940 
941 static __inline int
942 vintr_intercept_enabled(struct svm_softc *sc, int vcpu)
943 {
944 
945 	return (svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
946 	    VMCB_INTCPT_VINTR));
947 }
948 
949 static __inline void
950 enable_intr_window_exiting(struct svm_softc *sc, int vcpu)
951 {
952 	struct vmcb_ctrl *ctrl;
953 
954 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
955 
956 	if (ctrl->v_irq && ctrl->v_intr_vector == 0) {
957 		KASSERT(ctrl->v_ign_tpr, ("%s: invalid v_ign_tpr", __func__));
958 		KASSERT(vintr_intercept_enabled(sc, vcpu),
959 		    ("%s: vintr intercept should be enabled", __func__));
960 		return;
961 	}
962 
963 	VCPU_CTR0(sc->vm, vcpu, "Enable intr window exiting");
964 	ctrl->v_irq = 1;
965 	ctrl->v_ign_tpr = 1;
966 	ctrl->v_intr_vector = 0;
967 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
968 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
969 }
970 
971 static __inline void
972 disable_intr_window_exiting(struct svm_softc *sc, int vcpu)
973 {
974 	struct vmcb_ctrl *ctrl;
975 
976 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
977 
978 	if (!ctrl->v_irq && ctrl->v_intr_vector == 0) {
979 		KASSERT(!vintr_intercept_enabled(sc, vcpu),
980 		    ("%s: vintr intercept should be disabled", __func__));
981 		return;
982 	}
983 
984 #ifdef KTR
985 	if (ctrl->v_intr_vector == 0)
986 		VCPU_CTR0(sc->vm, vcpu, "Disable intr window exiting");
987 	else
988 		VCPU_CTR0(sc->vm, vcpu, "Clearing V_IRQ interrupt injection");
989 #endif
990 	ctrl->v_irq = 0;
991 	ctrl->v_intr_vector = 0;
992 	svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
993 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_VINTR);
994 }
995 
996 static int
997 svm_modify_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t val)
998 {
999 	struct vmcb_ctrl *ctrl;
1000 	int oldval, newval;
1001 
1002 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1003 	oldval = ctrl->intr_shadow;
1004 	newval = val ? 1 : 0;
1005 	if (newval != oldval) {
1006 		ctrl->intr_shadow = newval;
1007 		VCPU_CTR1(sc->vm, vcpu, "Setting intr_shadow to %d", newval);
1008 	}
1009 	return (0);
1010 }
1011 
1012 static int
1013 svm_get_intr_shadow(struct svm_softc *sc, int vcpu, uint64_t *val)
1014 {
1015 	struct vmcb_ctrl *ctrl;
1016 
1017 	ctrl = svm_get_vmcb_ctrl(sc, vcpu);
1018 	*val = ctrl->intr_shadow;
1019 	return (0);
1020 }
1021 
1022 /*
1023  * Once an NMI is injected it blocks delivery of further NMIs until the handler
1024  * executes an IRET. The IRET intercept is enabled when an NMI is injected to
1025  * to track when the vcpu is done handling the NMI.
1026  */
1027 static int
1028 nmi_blocked(struct svm_softc *sc, int vcpu)
1029 {
1030 	int blocked;
1031 
1032 	blocked = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
1033 	    VMCB_INTCPT_IRET);
1034 	return (blocked);
1035 }
1036 
1037 static void
1038 enable_nmi_blocking(struct svm_softc *sc, int vcpu)
1039 {
1040 
1041 	KASSERT(!nmi_blocked(sc, vcpu), ("vNMI already blocked"));
1042 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking enabled");
1043 	svm_enable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1044 }
1045 
1046 static void
1047 clear_nmi_blocking(struct svm_softc *sc, int vcpu)
1048 {
1049 	int error;
1050 
1051 	KASSERT(nmi_blocked(sc, vcpu), ("vNMI already unblocked"));
1052 	VCPU_CTR0(sc->vm, vcpu, "vNMI blocking cleared");
1053 	/*
1054 	 * When the IRET intercept is cleared the vcpu will attempt to execute
1055 	 * the "iret" when it runs next. However, it is possible to inject
1056 	 * another NMI into the vcpu before the "iret" has actually executed.
1057 	 *
1058 	 * For e.g. if the "iret" encounters a #NPF when accessing the stack
1059 	 * it will trap back into the hypervisor. If an NMI is pending for
1060 	 * the vcpu it will be injected into the guest.
1061 	 *
1062 	 * XXX this needs to be fixed
1063 	 */
1064 	svm_disable_intercept(sc, vcpu, VMCB_CTRL1_INTCPT, VMCB_INTCPT_IRET);
1065 
1066 	/*
1067 	 * Set 'intr_shadow' to prevent an NMI from being injected on the
1068 	 * immediate VMRUN.
1069 	 */
1070 	error = svm_modify_intr_shadow(sc, vcpu, 1);
1071 	KASSERT(!error, ("%s: error %d setting intr_shadow", __func__, error));
1072 }
1073 
1074 static int
1075 emulate_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val,
1076     bool *retu)
1077 {
1078 	int error;
1079 
1080 	if (lapic_msr(num))
1081 		error = lapic_wrmsr(sc->vm, vcpu, num, val, retu);
1082 	else if (num == MSR_EFER)
1083 		error = svm_setreg(sc, vcpu, VM_REG_GUEST_EFER, val);
1084 	else
1085 		error = svm_wrmsr(sc, vcpu, num, val, retu);
1086 
1087 	return (error);
1088 }
1089 
1090 static int
1091 emulate_rdmsr(struct svm_softc *sc, int vcpu, u_int num, bool *retu)
1092 {
1093 	struct vmcb_state *state;
1094 	struct svm_regctx *ctx;
1095 	uint64_t result;
1096 	int error;
1097 
1098 	if (lapic_msr(num))
1099 		error = lapic_rdmsr(sc->vm, vcpu, num, &result, retu);
1100 	else
1101 		error = svm_rdmsr(sc, vcpu, num, &result, retu);
1102 
1103 	if (error == 0) {
1104 		state = svm_get_vmcb_state(sc, vcpu);
1105 		ctx = svm_get_guest_regctx(sc, vcpu);
1106 		state->rax = result & 0xffffffff;
1107 		ctx->sctx_rdx = result >> 32;
1108 	}
1109 
1110 	return (error);
1111 }
1112 
1113 #ifdef KTR
1114 static const char *
1115 exit_reason_to_str(uint64_t reason)
1116 {
1117 	static char reasonbuf[32];
1118 
1119 	switch (reason) {
1120 	case VMCB_EXIT_INVALID:
1121 		return ("invalvmcb");
1122 	case VMCB_EXIT_SHUTDOWN:
1123 		return ("shutdown");
1124 	case VMCB_EXIT_NPF:
1125 		return ("nptfault");
1126 	case VMCB_EXIT_PAUSE:
1127 		return ("pause");
1128 	case VMCB_EXIT_HLT:
1129 		return ("hlt");
1130 	case VMCB_EXIT_CPUID:
1131 		return ("cpuid");
1132 	case VMCB_EXIT_IO:
1133 		return ("inout");
1134 	case VMCB_EXIT_MC:
1135 		return ("mchk");
1136 	case VMCB_EXIT_INTR:
1137 		return ("extintr");
1138 	case VMCB_EXIT_NMI:
1139 		return ("nmi");
1140 	case VMCB_EXIT_VINTR:
1141 		return ("vintr");
1142 	case VMCB_EXIT_MSR:
1143 		return ("msr");
1144 	case VMCB_EXIT_IRET:
1145 		return ("iret");
1146 	case VMCB_EXIT_MONITOR:
1147 		return ("monitor");
1148 	case VMCB_EXIT_MWAIT:
1149 		return ("mwait");
1150 	default:
1151 		snprintf(reasonbuf, sizeof(reasonbuf), "%#lx", reason);
1152 		return (reasonbuf);
1153 	}
1154 }
1155 #endif	/* KTR */
1156 
1157 /*
1158  * From section "State Saved on Exit" in APMv2: nRIP is saved for all #VMEXITs
1159  * that are due to instruction intercepts as well as MSR and IOIO intercepts
1160  * and exceptions caused by INT3, INTO and BOUND instructions.
1161  *
1162  * Return 1 if the nRIP is valid and 0 otherwise.
1163  */
1164 static int
1165 nrip_valid(uint64_t exitcode)
1166 {
1167 	switch (exitcode) {
1168 	case 0x00 ... 0x0F:	/* read of CR0 through CR15 */
1169 	case 0x10 ... 0x1F:	/* write of CR0 through CR15 */
1170 	case 0x20 ... 0x2F:	/* read of DR0 through DR15 */
1171 	case 0x30 ... 0x3F:	/* write of DR0 through DR15 */
1172 	case 0x43:		/* INT3 */
1173 	case 0x44:		/* INTO */
1174 	case 0x45:		/* BOUND */
1175 	case 0x65 ... 0x7C:	/* VMEXIT_CR0_SEL_WRITE ... VMEXIT_MSR */
1176 	case 0x80 ... 0x8D:	/* VMEXIT_VMRUN ... VMEXIT_XSETBV */
1177 		return (1);
1178 	default:
1179 		return (0);
1180 	}
1181 }
1182 
1183 /*
1184  * Collateral for a generic SVM VM-exit.
1185  */
1186 static void
1187 vm_exit_svm(struct vm_exit *vme, uint64_t code, uint64_t info1, uint64_t info2)
1188 {
1189 
1190 	vme->exitcode = VM_EXITCODE_SVM;
1191 	vme->u.svm.exitcode = code;
1192 	vme->u.svm.exitinfo1 = info1;
1193 	vme->u.svm.exitinfo2 = info2;
1194 }
1195 
1196 static int
1197 svm_vmexit(struct svm_softc *svm_sc, int vcpu, struct vm_exit *vmexit)
1198 {
1199 	struct vmcb *vmcb;
1200 	struct vmcb_state *state;
1201 	struct vmcb_ctrl *ctrl;
1202 	struct svm_regctx *ctx;
1203 	struct vm_exception exception;
1204 	uint64_t code, info1, info2, val;
1205 	uint32_t eax, ecx, edx;
1206 	int error, errcode_valid, handled, idtvec, reflect;
1207 	bool retu;
1208 
1209 	ctx = svm_get_guest_regctx(svm_sc, vcpu);
1210 	vmcb = svm_get_vmcb(svm_sc, vcpu);
1211 	state = &vmcb->state;
1212 	ctrl = &vmcb->ctrl;
1213 
1214 	handled = 0;
1215 	code = ctrl->exitcode;
1216 	info1 = ctrl->exitinfo1;
1217 	info2 = ctrl->exitinfo2;
1218 
1219 	vmexit->exitcode = VM_EXITCODE_BOGUS;
1220 	vmexit->rip = state->rip;
1221 	vmexit->inst_length = nrip_valid(code) ? ctrl->nrip - state->rip : 0;
1222 
1223 	vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_COUNT, 1);
1224 
1225 	/*
1226 	 * #VMEXIT(INVALID) needs to be handled early because the VMCB is
1227 	 * in an inconsistent state and can trigger assertions that would
1228 	 * never happen otherwise.
1229 	 */
1230 	if (code == VMCB_EXIT_INVALID) {
1231 		vm_exit_svm(vmexit, code, info1, info2);
1232 		return (0);
1233 	}
1234 
1235 	KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) == 0, ("%s: event "
1236 	    "injection valid bit is set %#lx", __func__, ctrl->eventinj));
1237 
1238 	KASSERT(vmexit->inst_length >= 0 && vmexit->inst_length <= 15,
1239 	    ("invalid inst_length %d: code (%#lx), info1 (%#lx), info2 (%#lx)",
1240 	    vmexit->inst_length, code, info1, info2));
1241 
1242 	svm_update_virqinfo(svm_sc, vcpu);
1243 	svm_save_intinfo(svm_sc, vcpu);
1244 
1245 	switch (code) {
1246 	case VMCB_EXIT_IRET:
1247 		/*
1248 		 * Restart execution at "iret" but with the intercept cleared.
1249 		 */
1250 		vmexit->inst_length = 0;
1251 		clear_nmi_blocking(svm_sc, vcpu);
1252 		handled = 1;
1253 		break;
1254 	case VMCB_EXIT_VINTR:	/* interrupt window exiting */
1255 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_VINTR, 1);
1256 		handled = 1;
1257 		break;
1258 	case VMCB_EXIT_INTR:	/* external interrupt */
1259 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXTINT, 1);
1260 		handled = 1;
1261 		break;
1262 	case VMCB_EXIT_NMI:	/* external NMI */
1263 		handled = 1;
1264 		break;
1265 	case 0x40 ... 0x5F:
1266 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_EXCEPTION, 1);
1267 		reflect = 1;
1268 		idtvec = code - 0x40;
1269 		switch (idtvec) {
1270 		case IDT_MC:
1271 			/*
1272 			 * Call the machine check handler by hand. Also don't
1273 			 * reflect the machine check back into the guest.
1274 			 */
1275 			reflect = 0;
1276 			VCPU_CTR0(svm_sc->vm, vcpu, "Vectoring to MCE handler");
1277 			__asm __volatile("int $18");
1278 			break;
1279 		case IDT_PF:
1280 			error = svm_setreg(svm_sc, vcpu, VM_REG_GUEST_CR2,
1281 			    info2);
1282 			KASSERT(error == 0, ("%s: error %d updating cr2",
1283 			    __func__, error));
1284 			/* fallthru */
1285 		case IDT_NP:
1286 		case IDT_SS:
1287 		case IDT_GP:
1288 		case IDT_AC:
1289 		case IDT_TS:
1290 			errcode_valid = 1;
1291 			break;
1292 
1293 		case IDT_DF:
1294 			errcode_valid = 1;
1295 			info1 = 0;
1296 			break;
1297 
1298 		case IDT_BP:
1299 		case IDT_OF:
1300 		case IDT_BR:
1301 			/*
1302 			 * The 'nrip' field is populated for INT3, INTO and
1303 			 * BOUND exceptions and this also implies that
1304 			 * 'inst_length' is non-zero.
1305 			 *
1306 			 * Reset 'inst_length' to zero so the guest %rip at
1307 			 * event injection is identical to what it was when
1308 			 * the exception originally happened.
1309 			 */
1310 			VCPU_CTR2(svm_sc->vm, vcpu, "Reset inst_length from %d "
1311 			    "to zero before injecting exception %d",
1312 			    vmexit->inst_length, idtvec);
1313 			vmexit->inst_length = 0;
1314 			/* fallthru */
1315 		default:
1316 			errcode_valid = 0;
1317 			break;
1318 		}
1319 		KASSERT(vmexit->inst_length == 0, ("invalid inst_length (%d) "
1320 		    "when reflecting exception %d into guest",
1321 		    vmexit->inst_length, idtvec));
1322 
1323 		if (reflect) {
1324 			/* Reflect the exception back into the guest */
1325 			bzero(&exception, sizeof(struct vm_exception));
1326 			exception.vector = idtvec;
1327 			if (errcode_valid) {
1328 				exception.error_code = info1;
1329 				exception.error_code_valid = 1;
1330 			}
1331 			VCPU_CTR2(svm_sc->vm, vcpu, "Reflecting exception "
1332 			    "%d/%#x into the guest", exception.vector,
1333 			    exception.error_code);
1334 			error = vm_inject_exception(svm_sc->vm, vcpu,
1335 			    &exception);
1336 			KASSERT(error == 0, ("%s: vm_inject_exception error %d",
1337 			    __func__, error));
1338 		}
1339 		handled = 1;
1340 		break;
1341 	case VMCB_EXIT_MSR:	/* MSR access. */
1342 		eax = state->rax;
1343 		ecx = ctx->sctx_rcx;
1344 		edx = ctx->sctx_rdx;
1345 		retu = false;
1346 
1347 		if (info1) {
1348 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_WRMSR, 1);
1349 			val = (uint64_t)edx << 32 | eax;
1350 			VCPU_CTR2(svm_sc->vm, vcpu, "wrmsr %#x val %#lx",
1351 			    ecx, val);
1352 			if (emulate_wrmsr(svm_sc, vcpu, ecx, val, &retu)) {
1353 				vmexit->exitcode = VM_EXITCODE_WRMSR;
1354 				vmexit->u.msr.code = ecx;
1355 				vmexit->u.msr.wval = val;
1356 			} else if (!retu) {
1357 				handled = 1;
1358 			} else {
1359 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1360 				    ("emulate_wrmsr retu with bogus exitcode"));
1361 			}
1362 		} else {
1363 			VCPU_CTR1(svm_sc->vm, vcpu, "rdmsr %#x", ecx);
1364 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_RDMSR, 1);
1365 			if (emulate_rdmsr(svm_sc, vcpu, ecx, &retu)) {
1366 				vmexit->exitcode = VM_EXITCODE_RDMSR;
1367 				vmexit->u.msr.code = ecx;
1368 			} else if (!retu) {
1369 				handled = 1;
1370 			} else {
1371 				KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
1372 				    ("emulate_rdmsr retu with bogus exitcode"));
1373 			}
1374 		}
1375 		break;
1376 	case VMCB_EXIT_IO:
1377 		handled = svm_handle_io(svm_sc, vcpu, vmexit);
1378 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INOUT, 1);
1379 		break;
1380 	case VMCB_EXIT_CPUID:
1381 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_CPUID, 1);
1382 		handled = x86_emulate_cpuid(svm_sc->vm, vcpu,
1383 		    (uint32_t *)&state->rax,
1384 		    (uint32_t *)&ctx->sctx_rbx,
1385 		    (uint32_t *)&ctx->sctx_rcx,
1386 		    (uint32_t *)&ctx->sctx_rdx);
1387 		break;
1388 	case VMCB_EXIT_HLT:
1389 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_HLT, 1);
1390 		vmexit->exitcode = VM_EXITCODE_HLT;
1391 		vmexit->u.hlt.rflags = state->rflags;
1392 		break;
1393 	case VMCB_EXIT_PAUSE:
1394 		vmexit->exitcode = VM_EXITCODE_PAUSE;
1395 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_PAUSE, 1);
1396 		break;
1397 	case VMCB_EXIT_NPF:
1398 		/* EXITINFO2 contains the faulting guest physical address */
1399 		if (info1 & VMCB_NPF_INFO1_RSV) {
1400 			VCPU_CTR2(svm_sc->vm, vcpu, "nested page fault with "
1401 			    "reserved bits set: info1(%#lx) info2(%#lx)",
1402 			    info1, info2);
1403 		} else if (vm_mem_allocated(svm_sc->vm, info2)) {
1404 			vmexit->exitcode = VM_EXITCODE_PAGING;
1405 			vmexit->u.paging.gpa = info2;
1406 			vmexit->u.paging.fault_type = npf_fault_type(info1);
1407 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
1408 			VCPU_CTR3(svm_sc->vm, vcpu, "nested page fault "
1409 			    "on gpa %#lx/%#lx at rip %#lx",
1410 			    info2, info1, state->rip);
1411 		} else if (svm_npf_emul_fault(info1)) {
1412 			svm_handle_inst_emul(vmcb, info2, vmexit);
1413 			vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_INST_EMUL, 1);
1414 			VCPU_CTR3(svm_sc->vm, vcpu, "inst_emul fault "
1415 			    "for gpa %#lx/%#lx at rip %#lx",
1416 			    info2, info1, state->rip);
1417 		}
1418 		break;
1419 	case VMCB_EXIT_MONITOR:
1420 		vmexit->exitcode = VM_EXITCODE_MONITOR;
1421 		break;
1422 	case VMCB_EXIT_MWAIT:
1423 		vmexit->exitcode = VM_EXITCODE_MWAIT;
1424 		break;
1425 	default:
1426 		vmm_stat_incr(svm_sc->vm, vcpu, VMEXIT_UNKNOWN, 1);
1427 		break;
1428 	}
1429 
1430 	VCPU_CTR4(svm_sc->vm, vcpu, "%s %s vmexit at %#lx/%d",
1431 	    handled ? "handled" : "unhandled", exit_reason_to_str(code),
1432 	    vmexit->rip, vmexit->inst_length);
1433 
1434 	if (handled) {
1435 		vmexit->rip += vmexit->inst_length;
1436 		vmexit->inst_length = 0;
1437 		state->rip = vmexit->rip;
1438 	} else {
1439 		if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
1440 			/*
1441 			 * If this VM exit was not claimed by anybody then
1442 			 * treat it as a generic SVM exit.
1443 			 */
1444 			vm_exit_svm(vmexit, code, info1, info2);
1445 		} else {
1446 			/*
1447 			 * The exitcode and collateral have been populated.
1448 			 * The VM exit will be processed further in userland.
1449 			 */
1450 		}
1451 	}
1452 	return (handled);
1453 }
1454 
1455 static void
1456 svm_inj_intinfo(struct svm_softc *svm_sc, int vcpu)
1457 {
1458 	uint64_t intinfo;
1459 
1460 	if (!vm_entry_intinfo(svm_sc->vm, vcpu, &intinfo))
1461 		return;
1462 
1463 	KASSERT(VMCB_EXITINTINFO_VALID(intinfo), ("%s: entry intinfo is not "
1464 	    "valid: %#lx", __func__, intinfo));
1465 
1466 	svm_eventinject(svm_sc, vcpu, VMCB_EXITINTINFO_TYPE(intinfo),
1467 		VMCB_EXITINTINFO_VECTOR(intinfo),
1468 		VMCB_EXITINTINFO_EC(intinfo),
1469 		VMCB_EXITINTINFO_EC_VALID(intinfo));
1470 	vmm_stat_incr(svm_sc->vm, vcpu, VCPU_INTINFO_INJECTED, 1);
1471 	VCPU_CTR1(svm_sc->vm, vcpu, "Injected entry intinfo: %#lx", intinfo);
1472 }
1473 
1474 /*
1475  * Inject event to virtual cpu.
1476  */
1477 static void
1478 svm_inj_interrupts(struct svm_softc *sc, int vcpu, struct vlapic *vlapic)
1479 {
1480 	struct vmcb_ctrl *ctrl;
1481 	struct vmcb_state *state;
1482 	uint8_t v_tpr;
1483 	int vector, need_intr_window, pending_apic_vector;
1484 
1485 	state = svm_get_vmcb_state(sc, vcpu);
1486 	ctrl  = svm_get_vmcb_ctrl(sc, vcpu);
1487 
1488 	need_intr_window = 0;
1489 	pending_apic_vector = 0;
1490 
1491 	/*
1492 	 * Inject pending events or exceptions for this vcpu.
1493 	 *
1494 	 * An event might be pending because the previous #VMEXIT happened
1495 	 * during event delivery (i.e. ctrl->exitintinfo).
1496 	 *
1497 	 * An event might also be pending because an exception was injected
1498 	 * by the hypervisor (e.g. #PF during instruction emulation).
1499 	 */
1500 	svm_inj_intinfo(sc, vcpu);
1501 
1502 	/* NMI event has priority over interrupts. */
1503 	if (vm_nmi_pending(sc->vm, vcpu)) {
1504 		if (nmi_blocked(sc, vcpu)) {
1505 			/*
1506 			 * Can't inject another NMI if the guest has not
1507 			 * yet executed an "iret" after the last NMI.
1508 			 */
1509 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due "
1510 			    "to NMI-blocking");
1511 		} else if (ctrl->intr_shadow) {
1512 			/*
1513 			 * Can't inject an NMI if the vcpu is in an intr_shadow.
1514 			 */
1515 			VCPU_CTR0(sc->vm, vcpu, "Cannot inject NMI due to "
1516 			    "interrupt shadow");
1517 			need_intr_window = 1;
1518 			goto done;
1519 		} else if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1520 			/*
1521 			 * If there is already an exception/interrupt pending
1522 			 * then defer the NMI until after that.
1523 			 */
1524 			VCPU_CTR1(sc->vm, vcpu, "Cannot inject NMI due to "
1525 			    "eventinj %#lx", ctrl->eventinj);
1526 
1527 			/*
1528 			 * Use self-IPI to trigger a VM-exit as soon as
1529 			 * possible after the event injection is completed.
1530 			 *
1531 			 * This works only if the external interrupt exiting
1532 			 * is at a lower priority than the event injection.
1533 			 *
1534 			 * Although not explicitly specified in APMv2 the
1535 			 * relative priorities were verified empirically.
1536 			 */
1537 			ipi_cpu(curcpu, IPI_AST);	/* XXX vmm_ipinum? */
1538 		} else {
1539 			vm_nmi_clear(sc->vm, vcpu);
1540 
1541 			/* Inject NMI, vector number is not used */
1542 			svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_NMI,
1543 			    IDT_NMI, 0, false);
1544 
1545 			/* virtual NMI blocking is now in effect */
1546 			enable_nmi_blocking(sc, vcpu);
1547 
1548 			VCPU_CTR0(sc->vm, vcpu, "Injecting vNMI");
1549 		}
1550 	}
1551 
1552 	if (!vm_extint_pending(sc->vm, vcpu)) {
1553 		/*
1554 		 * APIC interrupts are delivered using the V_IRQ offload.
1555 		 *
1556 		 * The primary benefit is that the hypervisor doesn't need to
1557 		 * deal with the various conditions that inhibit interrupts.
1558 		 * It also means that TPR changes via CR8 will be handled
1559 		 * without any hypervisor involvement.
1560 		 *
1561 		 * Note that the APIC vector must remain pending in the vIRR
1562 		 * until it is confirmed that it was delivered to the guest.
1563 		 * This can be confirmed based on the value of V_IRQ at the
1564 		 * next #VMEXIT (1 = pending, 0 = delivered).
1565 		 *
1566 		 * Also note that it is possible that another higher priority
1567 		 * vector can become pending before this vector is delivered
1568 		 * to the guest. This is alright because vcpu_notify_event()
1569 		 * will send an IPI and force the vcpu to trap back into the
1570 		 * hypervisor. The higher priority vector will be injected on
1571 		 * the next VMRUN.
1572 		 */
1573 		if (vlapic_pending_intr(vlapic, &vector)) {
1574 			KASSERT(vector >= 16 && vector <= 255,
1575 			    ("invalid vector %d from local APIC", vector));
1576 			pending_apic_vector = vector;
1577 		}
1578 		goto done;
1579 	}
1580 
1581 	/* Ask the legacy pic for a vector to inject */
1582 	vatpic_pending_intr(sc->vm, &vector);
1583 	KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d from INTR",
1584 	    vector));
1585 
1586 	/*
1587 	 * If the guest has disabled interrupts or is in an interrupt shadow
1588 	 * then we cannot inject the pending interrupt.
1589 	 */
1590 	if ((state->rflags & PSL_I) == 0) {
1591 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1592 		    "rflags %#lx", vector, state->rflags);
1593 		need_intr_window = 1;
1594 		goto done;
1595 	}
1596 
1597 	if (ctrl->intr_shadow) {
1598 		VCPU_CTR1(sc->vm, vcpu, "Cannot inject vector %d due to "
1599 		    "interrupt shadow", vector);
1600 		need_intr_window = 1;
1601 		goto done;
1602 	}
1603 
1604 	if (ctrl->eventinj & VMCB_EVENTINJ_VALID) {
1605 		VCPU_CTR2(sc->vm, vcpu, "Cannot inject vector %d due to "
1606 		    "eventinj %#lx", vector, ctrl->eventinj);
1607 		need_intr_window = 1;
1608 		goto done;
1609 	}
1610 
1611 	/*
1612 	 * Legacy PIC interrupts are delivered via the event injection
1613 	 * mechanism.
1614 	 */
1615 	svm_eventinject(sc, vcpu, VMCB_EVENTINJ_TYPE_INTR, vector, 0, false);
1616 
1617 	vm_extint_clear(sc->vm, vcpu);
1618 	vatpic_intr_accepted(sc->vm, vector);
1619 
1620 	/*
1621 	 * Force a VM-exit as soon as the vcpu is ready to accept another
1622 	 * interrupt. This is done because the PIC might have another vector
1623 	 * that it wants to inject. Also, if the APIC has a pending interrupt
1624 	 * that was preempted by the ExtInt then it allows us to inject the
1625 	 * APIC vector as soon as possible.
1626 	 */
1627 	need_intr_window = 1;
1628 done:
1629 	/*
1630 	 * The guest can modify the TPR by writing to %CR8. In guest mode
1631 	 * the processor reflects this write to V_TPR without hypervisor
1632 	 * intervention.
1633 	 *
1634 	 * The guest can also modify the TPR by writing to it via the memory
1635 	 * mapped APIC page. In this case, the write will be emulated by the
1636 	 * hypervisor. For this reason V_TPR must be updated before every
1637 	 * VMRUN.
1638 	 */
1639 	v_tpr = vlapic_get_cr8(vlapic);
1640 	KASSERT(v_tpr >= 0 && v_tpr <= 15, ("invalid v_tpr %#x", v_tpr));
1641 	if (ctrl->v_tpr != v_tpr) {
1642 		VCPU_CTR2(sc->vm, vcpu, "VMCB V_TPR changed from %#x to %#x",
1643 		    ctrl->v_tpr, v_tpr);
1644 		ctrl->v_tpr = v_tpr;
1645 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1646 	}
1647 
1648 	if (pending_apic_vector) {
1649 		/*
1650 		 * If an APIC vector is being injected then interrupt window
1651 		 * exiting is not possible on this VMRUN.
1652 		 */
1653 		KASSERT(!need_intr_window, ("intr_window exiting impossible"));
1654 		VCPU_CTR1(sc->vm, vcpu, "Injecting vector %d using V_IRQ",
1655 		    pending_apic_vector);
1656 
1657 		ctrl->v_irq = 1;
1658 		ctrl->v_ign_tpr = 0;
1659 		ctrl->v_intr_vector = pending_apic_vector;
1660 		ctrl->v_intr_prio = pending_apic_vector >> 4;
1661 		svm_set_dirty(sc, vcpu, VMCB_CACHE_TPR);
1662 	} else if (need_intr_window) {
1663 		/*
1664 		 * We use V_IRQ in conjunction with the VINTR intercept to
1665 		 * trap into the hypervisor as soon as a virtual interrupt
1666 		 * can be delivered.
1667 		 *
1668 		 * Since injected events are not subject to intercept checks
1669 		 * we need to ensure that the V_IRQ is not actually going to
1670 		 * be delivered on VM entry. The KASSERT below enforces this.
1671 		 */
1672 		KASSERT((ctrl->eventinj & VMCB_EVENTINJ_VALID) != 0 ||
1673 		    (state->rflags & PSL_I) == 0 || ctrl->intr_shadow,
1674 		    ("Bogus intr_window_exiting: eventinj (%#lx), "
1675 		    "intr_shadow (%u), rflags (%#lx)",
1676 		    ctrl->eventinj, ctrl->intr_shadow, state->rflags));
1677 		enable_intr_window_exiting(sc, vcpu);
1678 	} else {
1679 		disable_intr_window_exiting(sc, vcpu);
1680 	}
1681 }
1682 
1683 static __inline void
1684 restore_host_tss(void)
1685 {
1686 	struct system_segment_descriptor *tss_sd;
1687 
1688 	/*
1689 	 * The TSS descriptor was in use prior to launching the guest so it
1690 	 * has been marked busy.
1691 	 *
1692 	 * 'ltr' requires the descriptor to be marked available so change the
1693 	 * type to "64-bit available TSS".
1694 	 */
1695 	tss_sd = PCPU_GET(tss);
1696 	tss_sd->sd_type = SDT_SYSTSS;
1697 	ltr(GSEL(GPROC0_SEL, SEL_KPL));
1698 }
1699 
1700 static void
1701 check_asid(struct svm_softc *sc, int vcpuid, pmap_t pmap, u_int thiscpu)
1702 {
1703 	struct svm_vcpu *vcpustate;
1704 	struct vmcb_ctrl *ctrl;
1705 	long eptgen;
1706 	bool alloc_asid;
1707 
1708 	KASSERT(CPU_ISSET(thiscpu, &pmap->pm_active), ("%s: nested pmap not "
1709 	    "active on cpu %u", __func__, thiscpu));
1710 
1711 	vcpustate = svm_get_vcpu(sc, vcpuid);
1712 	ctrl = svm_get_vmcb_ctrl(sc, vcpuid);
1713 
1714 	/*
1715 	 * The TLB entries associated with the vcpu's ASID are not valid
1716 	 * if either of the following conditions is true:
1717 	 *
1718 	 * 1. The vcpu's ASID generation is different than the host cpu's
1719 	 *    ASID generation. This happens when the vcpu migrates to a new
1720 	 *    host cpu. It can also happen when the number of vcpus executing
1721 	 *    on a host cpu is greater than the number of ASIDs available.
1722 	 *
1723 	 * 2. The pmap generation number is different than the value cached in
1724 	 *    the 'vcpustate'. This happens when the host invalidates pages
1725 	 *    belonging to the guest.
1726 	 *
1727 	 *	asidgen		eptgen	      Action
1728 	 *	mismatch	mismatch
1729 	 *	   0		   0		(a)
1730 	 *	   0		   1		(b1) or (b2)
1731 	 *	   1		   0		(c)
1732 	 *	   1		   1		(d)
1733 	 *
1734 	 * (a) There is no mismatch in eptgen or ASID generation and therefore
1735 	 *     no further action is needed.
1736 	 *
1737 	 * (b1) If the cpu supports FlushByAsid then the vcpu's ASID is
1738 	 *      retained and the TLB entries associated with this ASID
1739 	 *      are flushed by VMRUN.
1740 	 *
1741 	 * (b2) If the cpu does not support FlushByAsid then a new ASID is
1742 	 *      allocated.
1743 	 *
1744 	 * (c) A new ASID is allocated.
1745 	 *
1746 	 * (d) A new ASID is allocated.
1747 	 */
1748 
1749 	alloc_asid = false;
1750 	eptgen = pmap->pm_eptgen;
1751 	ctrl->tlb_ctrl = VMCB_TLB_FLUSH_NOTHING;
1752 
1753 	if (vcpustate->asid.gen != asid[thiscpu].gen) {
1754 		alloc_asid = true;	/* (c) and (d) */
1755 	} else if (vcpustate->eptgen != eptgen) {
1756 		if (flush_by_asid())
1757 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;	/* (b1) */
1758 		else
1759 			alloc_asid = true;			/* (b2) */
1760 	} else {
1761 		/*
1762 		 * This is the common case (a).
1763 		 */
1764 		KASSERT(!alloc_asid, ("ASID allocation not necessary"));
1765 		KASSERT(ctrl->tlb_ctrl == VMCB_TLB_FLUSH_NOTHING,
1766 		    ("Invalid VMCB tlb_ctrl: %#x", ctrl->tlb_ctrl));
1767 	}
1768 
1769 	if (alloc_asid) {
1770 		if (++asid[thiscpu].num >= nasid) {
1771 			asid[thiscpu].num = 1;
1772 			if (++asid[thiscpu].gen == 0)
1773 				asid[thiscpu].gen = 1;
1774 			/*
1775 			 * If this cpu does not support "flush-by-asid"
1776 			 * then flush the entire TLB on a generation
1777 			 * bump. Subsequent ASID allocation in this
1778 			 * generation can be done without a TLB flush.
1779 			 */
1780 			if (!flush_by_asid())
1781 				ctrl->tlb_ctrl = VMCB_TLB_FLUSH_ALL;
1782 		}
1783 		vcpustate->asid.gen = asid[thiscpu].gen;
1784 		vcpustate->asid.num = asid[thiscpu].num;
1785 
1786 		ctrl->asid = vcpustate->asid.num;
1787 		svm_set_dirty(sc, vcpuid, VMCB_CACHE_ASID);
1788 		/*
1789 		 * If this cpu supports "flush-by-asid" then the TLB
1790 		 * was not flushed after the generation bump. The TLB
1791 		 * is flushed selectively after every new ASID allocation.
1792 		 */
1793 		if (flush_by_asid())
1794 			ctrl->tlb_ctrl = VMCB_TLB_FLUSH_GUEST;
1795 	}
1796 	vcpustate->eptgen = eptgen;
1797 
1798 	KASSERT(ctrl->asid != 0, ("Guest ASID must be non-zero"));
1799 	KASSERT(ctrl->asid == vcpustate->asid.num,
1800 	    ("ASID mismatch: %u/%u", ctrl->asid, vcpustate->asid.num));
1801 }
1802 
1803 static __inline void
1804 disable_gintr(void)
1805 {
1806 
1807         __asm __volatile("clgi" : : :);
1808 }
1809 
1810 static __inline void
1811 enable_gintr(void)
1812 {
1813 
1814         __asm __volatile("stgi" : : :);
1815 }
1816 
1817 /*
1818  * Start vcpu with specified RIP.
1819  */
1820 static int
1821 svm_vmrun(void *arg, int vcpu, register_t rip, pmap_t pmap,
1822 	void *rend_cookie, void *suspended_cookie)
1823 {
1824 	struct svm_regctx *gctx;
1825 	struct svm_softc *svm_sc;
1826 	struct svm_vcpu *vcpustate;
1827 	struct vmcb_state *state;
1828 	struct vmcb_ctrl *ctrl;
1829 	struct vm_exit *vmexit;
1830 	struct vlapic *vlapic;
1831 	struct vm *vm;
1832 	uint64_t vmcb_pa;
1833 	u_int thiscpu;
1834 	int handled;
1835 
1836 	svm_sc = arg;
1837 	vm = svm_sc->vm;
1838 
1839 	vcpustate = svm_get_vcpu(svm_sc, vcpu);
1840 	state = svm_get_vmcb_state(svm_sc, vcpu);
1841 	ctrl = svm_get_vmcb_ctrl(svm_sc, vcpu);
1842 	vmexit = vm_exitinfo(vm, vcpu);
1843 	vlapic = vm_lapic(vm, vcpu);
1844 
1845 	/*
1846 	 * Stash 'curcpu' on the stack as 'thiscpu'.
1847 	 *
1848 	 * The per-cpu data area is not accessible until MSR_GSBASE is restored
1849 	 * after the #VMEXIT. Since VMRUN is executed inside a critical section
1850 	 * 'curcpu' and 'thiscpu' are guaranteed to identical.
1851 	 */
1852 	thiscpu = curcpu;
1853 
1854 	gctx = svm_get_guest_regctx(svm_sc, vcpu);
1855 	vmcb_pa = svm_sc->vcpu[vcpu].vmcb_pa;
1856 
1857 	if (vcpustate->lastcpu != thiscpu) {
1858 		/*
1859 		 * Force new ASID allocation by invalidating the generation.
1860 		 */
1861 		vcpustate->asid.gen = 0;
1862 
1863 		/*
1864 		 * Invalidate the VMCB state cache by marking all fields dirty.
1865 		 */
1866 		svm_set_dirty(svm_sc, vcpu, 0xffffffff);
1867 
1868 		/*
1869 		 * XXX
1870 		 * Setting 'vcpustate->lastcpu' here is bit premature because
1871 		 * we may return from this function without actually executing
1872 		 * the VMRUN  instruction. This could happen if a rendezvous
1873 		 * or an AST is pending on the first time through the loop.
1874 		 *
1875 		 * This works for now but any new side-effects of vcpu
1876 		 * migration should take this case into account.
1877 		 */
1878 		vcpustate->lastcpu = thiscpu;
1879 		vmm_stat_incr(vm, vcpu, VCPU_MIGRATIONS, 1);
1880 	}
1881 
1882 	svm_msr_guest_enter(svm_sc, vcpu);
1883 
1884 	/* Update Guest RIP */
1885 	state->rip = rip;
1886 
1887 	do {
1888 		/*
1889 		 * Disable global interrupts to guarantee atomicity during
1890 		 * loading of guest state. This includes not only the state
1891 		 * loaded by the "vmrun" instruction but also software state
1892 		 * maintained by the hypervisor: suspended and rendezvous
1893 		 * state, NPT generation number, vlapic interrupts etc.
1894 		 */
1895 		disable_gintr();
1896 
1897 		if (vcpu_suspended(suspended_cookie)) {
1898 			enable_gintr();
1899 			vm_exit_suspended(vm, vcpu, state->rip);
1900 			break;
1901 		}
1902 
1903 		if (vcpu_rendezvous_pending(rend_cookie)) {
1904 			enable_gintr();
1905 			vm_exit_rendezvous(vm, vcpu, state->rip);
1906 			break;
1907 		}
1908 
1909 		/* We are asked to give the cpu by scheduler. */
1910 		if (curthread->td_flags & (TDF_ASTPENDING | TDF_NEEDRESCHED)) {
1911 			enable_gintr();
1912 			vm_exit_astpending(vm, vcpu, state->rip);
1913 			break;
1914 		}
1915 
1916 		svm_inj_interrupts(svm_sc, vcpu, vlapic);
1917 
1918 		/* Activate the nested pmap on 'thiscpu' */
1919 		CPU_SET_ATOMIC_ACQ(thiscpu, &pmap->pm_active);
1920 
1921 		/*
1922 		 * Check the pmap generation and the ASID generation to
1923 		 * ensure that the vcpu does not use stale TLB mappings.
1924 		 */
1925 		check_asid(svm_sc, vcpu, pmap, thiscpu);
1926 
1927 		ctrl->vmcb_clean = vmcb_clean & ~vcpustate->dirty;
1928 		vcpustate->dirty = 0;
1929 		VCPU_CTR1(vm, vcpu, "vmcb clean %#x", ctrl->vmcb_clean);
1930 
1931 		/* Launch Virtual Machine. */
1932 		VCPU_CTR1(vm, vcpu, "Resume execution at %#lx", state->rip);
1933 		svm_launch(vmcb_pa, gctx);
1934 
1935 		CPU_CLR_ATOMIC(thiscpu, &pmap->pm_active);
1936 
1937 		/*
1938 		 * Restore MSR_GSBASE to point to the pcpu data area.
1939 		 *
1940 		 * Note that accesses done via PCPU_GET/PCPU_SET will work
1941 		 * only after MSR_GSBASE is restored.
1942 		 *
1943 		 * Also note that we don't bother restoring MSR_KGSBASE
1944 		 * since it is not used in the kernel and will be restored
1945 		 * when the VMRUN ioctl returns to userspace.
1946 		 */
1947 		wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[thiscpu]);
1948 		KASSERT(curcpu == thiscpu, ("thiscpu/curcpu (%u/%u) mismatch",
1949 		    thiscpu, curcpu));
1950 
1951 		/*
1952 		 * The host GDTR and IDTR is saved by VMRUN and restored
1953 		 * automatically on #VMEXIT. However, the host TSS needs
1954 		 * to be restored explicitly.
1955 		 */
1956 		restore_host_tss();
1957 
1958 		/* #VMEXIT disables interrupts so re-enable them here. */
1959 		enable_gintr();
1960 
1961 		/* Handle #VMEXIT and if required return to user space. */
1962 		handled = svm_vmexit(svm_sc, vcpu, vmexit);
1963 	} while (handled);
1964 
1965 	svm_msr_guest_exit(svm_sc, vcpu);
1966 
1967 	return (0);
1968 }
1969 
1970 static void
1971 svm_vmcleanup(void *arg)
1972 {
1973 	struct svm_softc *sc = arg;
1974 
1975 	free(sc, M_SVM);
1976 }
1977 
1978 static register_t *
1979 swctx_regptr(struct svm_regctx *regctx, int reg)
1980 {
1981 
1982 	switch (reg) {
1983 	case VM_REG_GUEST_RBX:
1984 		return (&regctx->sctx_rbx);
1985 	case VM_REG_GUEST_RCX:
1986 		return (&regctx->sctx_rcx);
1987 	case VM_REG_GUEST_RDX:
1988 		return (&regctx->sctx_rdx);
1989 	case VM_REG_GUEST_RDI:
1990 		return (&regctx->sctx_rdi);
1991 	case VM_REG_GUEST_RSI:
1992 		return (&regctx->sctx_rsi);
1993 	case VM_REG_GUEST_RBP:
1994 		return (&regctx->sctx_rbp);
1995 	case VM_REG_GUEST_R8:
1996 		return (&regctx->sctx_r8);
1997 	case VM_REG_GUEST_R9:
1998 		return (&regctx->sctx_r9);
1999 	case VM_REG_GUEST_R10:
2000 		return (&regctx->sctx_r10);
2001 	case VM_REG_GUEST_R11:
2002 		return (&regctx->sctx_r11);
2003 	case VM_REG_GUEST_R12:
2004 		return (&regctx->sctx_r12);
2005 	case VM_REG_GUEST_R13:
2006 		return (&regctx->sctx_r13);
2007 	case VM_REG_GUEST_R14:
2008 		return (&regctx->sctx_r14);
2009 	case VM_REG_GUEST_R15:
2010 		return (&regctx->sctx_r15);
2011 	default:
2012 		return (NULL);
2013 	}
2014 }
2015 
2016 static int
2017 svm_getreg(void *arg, int vcpu, int ident, uint64_t *val)
2018 {
2019 	struct svm_softc *svm_sc;
2020 	register_t *reg;
2021 
2022 	svm_sc = arg;
2023 
2024 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2025 		return (svm_get_intr_shadow(svm_sc, vcpu, val));
2026 	}
2027 
2028 	if (vmcb_read(svm_sc, vcpu, ident, val) == 0) {
2029 		return (0);
2030 	}
2031 
2032 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2033 
2034 	if (reg != NULL) {
2035 		*val = *reg;
2036 		return (0);
2037 	}
2038 
2039 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_getreg: unknown register %#x", ident);
2040 	return (EINVAL);
2041 }
2042 
2043 static int
2044 svm_setreg(void *arg, int vcpu, int ident, uint64_t val)
2045 {
2046 	struct svm_softc *svm_sc;
2047 	register_t *reg;
2048 
2049 	svm_sc = arg;
2050 
2051 	if (ident == VM_REG_GUEST_INTR_SHADOW) {
2052 		return (svm_modify_intr_shadow(svm_sc, vcpu, val));
2053 	}
2054 
2055 	if (vmcb_write(svm_sc, vcpu, ident, val) == 0) {
2056 		return (0);
2057 	}
2058 
2059 	reg = swctx_regptr(svm_get_guest_regctx(svm_sc, vcpu), ident);
2060 
2061 	if (reg != NULL) {
2062 		*reg = val;
2063 		return (0);
2064 	}
2065 
2066 	/*
2067 	 * XXX deal with CR3 and invalidate TLB entries tagged with the
2068 	 * vcpu's ASID. This needs to be treated differently depending on
2069 	 * whether 'running' is true/false.
2070 	 */
2071 
2072 	VCPU_CTR1(svm_sc->vm, vcpu, "svm_setreg: unknown register %#x", ident);
2073 	return (EINVAL);
2074 }
2075 
2076 static int
2077 svm_setcap(void *arg, int vcpu, int type, int val)
2078 {
2079 	struct svm_softc *sc;
2080 	int error;
2081 
2082 	sc = arg;
2083 	error = 0;
2084 	switch (type) {
2085 	case VM_CAP_HALT_EXIT:
2086 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2087 		    VMCB_INTCPT_HLT, val);
2088 		break;
2089 	case VM_CAP_PAUSE_EXIT:
2090 		svm_set_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2091 		    VMCB_INTCPT_PAUSE, val);
2092 		break;
2093 	case VM_CAP_UNRESTRICTED_GUEST:
2094 		/* Unrestricted guest execution cannot be disabled in SVM */
2095 		if (val == 0)
2096 			error = EINVAL;
2097 		break;
2098 	default:
2099 		error = ENOENT;
2100 		break;
2101 	}
2102 	return (error);
2103 }
2104 
2105 static int
2106 svm_getcap(void *arg, int vcpu, int type, int *retval)
2107 {
2108 	struct svm_softc *sc;
2109 	int error;
2110 
2111 	sc = arg;
2112 	error = 0;
2113 
2114 	switch (type) {
2115 	case VM_CAP_HALT_EXIT:
2116 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2117 		    VMCB_INTCPT_HLT);
2118 		break;
2119 	case VM_CAP_PAUSE_EXIT:
2120 		*retval = svm_get_intercept(sc, vcpu, VMCB_CTRL1_INTCPT,
2121 		    VMCB_INTCPT_PAUSE);
2122 		break;
2123 	case VM_CAP_UNRESTRICTED_GUEST:
2124 		*retval = 1;	/* unrestricted guest is always enabled */
2125 		break;
2126 	default:
2127 		error = ENOENT;
2128 		break;
2129 	}
2130 	return (error);
2131 }
2132 
2133 static struct vlapic *
2134 svm_vlapic_init(void *arg, int vcpuid)
2135 {
2136 	struct svm_softc *svm_sc;
2137 	struct vlapic *vlapic;
2138 
2139 	svm_sc = arg;
2140 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
2141 	vlapic->vm = svm_sc->vm;
2142 	vlapic->vcpuid = vcpuid;
2143 	vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
2144 
2145 	vlapic_init(vlapic);
2146 
2147 	return (vlapic);
2148 }
2149 
2150 static void
2151 svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
2152 {
2153 
2154         vlapic_cleanup(vlapic);
2155         free(vlapic, M_SVM_VLAPIC);
2156 }
2157 
2158 struct vmm_ops vmm_ops_amd = {
2159 	svm_init,
2160 	svm_cleanup,
2161 	svm_restore,
2162 	svm_vminit,
2163 	svm_vmrun,
2164 	svm_vmcleanup,
2165 	svm_getreg,
2166 	svm_setreg,
2167 	vmcb_getdesc,
2168 	vmcb_setdesc,
2169 	svm_getcap,
2170 	svm_setcap,
2171 	svm_npt_alloc,
2172 	svm_npt_free,
2173 	svm_vlapic_init,
2174 	svm_vlapic_cleanup
2175 };
2176