1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2011 NetApp, Inc.
5 * All rights reserved.
6 * Copyright (c) 2018 Joyent, Inc.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 *
29 * $FreeBSD$
30 */
31
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/smp.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/pcpu.h>
41 #include <sys/proc.h>
42 #include <sys/sysctl.h>
43
44 #include <vm/vm.h>
45 #include <vm/pmap.h>
46
47 #include <machine/psl.h>
48 #include <machine/cpufunc.h>
49 #include <machine/md_var.h>
50 #include <machine/reg.h>
51 #include <machine/segments.h>
52 #include <machine/smp.h>
53 #include <machine/specialreg.h>
54 #include <machine/vmparam.h>
55
56 #include <machine/vmm.h>
57 #include <machine/vmm_dev.h>
58 #include <machine/vmm_instruction_emul.h>
59 #include "vmm_lapic.h"
60 #include "vmm_host.h"
61 #include "vmm_ioport.h"
62 #include "vmm_ktr.h"
63 #include "vmm_stat.h"
64 #include "vatpic.h"
65 #include "vlapic.h"
66 #include "vlapic_priv.h"
67
68 #include "ept.h"
69 #include "vmx_cpufunc.h"
70 #include "vmx.h"
71 #include "vmx_msr.h"
72 #include "x86.h"
73 #include "vmx_controls.h"
74
75 #define PINBASED_CTLS_ONE_SETTING \
76 (PINBASED_EXTINT_EXITING | \
77 PINBASED_NMI_EXITING | \
78 PINBASED_VIRTUAL_NMI)
79 #define PINBASED_CTLS_ZERO_SETTING 0
80
81 #define PROCBASED_CTLS_WINDOW_SETTING \
82 (PROCBASED_INT_WINDOW_EXITING | \
83 PROCBASED_NMI_WINDOW_EXITING)
84
85 #define PROCBASED_CTLS_ONE_SETTING \
86 (PROCBASED_SECONDARY_CONTROLS | \
87 PROCBASED_MWAIT_EXITING | \
88 PROCBASED_MONITOR_EXITING | \
89 PROCBASED_IO_EXITING | \
90 PROCBASED_MSR_BITMAPS | \
91 PROCBASED_CTLS_WINDOW_SETTING | \
92 PROCBASED_CR8_LOAD_EXITING | \
93 PROCBASED_CR8_STORE_EXITING)
94 #define PROCBASED_CTLS_ZERO_SETTING \
95 (PROCBASED_CR3_LOAD_EXITING | \
96 PROCBASED_CR3_STORE_EXITING | \
97 PROCBASED_IO_BITMAPS)
98
99 #define PROCBASED_CTLS2_ONE_SETTING PROCBASED2_ENABLE_EPT
100 #define PROCBASED_CTLS2_ZERO_SETTING 0
101
102 #define VM_EXIT_CTLS_ONE_SETTING \
103 (VM_EXIT_SAVE_DEBUG_CONTROLS | \
104 VM_EXIT_HOST_LMA | \
105 VM_EXIT_SAVE_EFER | \
106 VM_EXIT_LOAD_EFER | \
107 VM_EXIT_ACKNOWLEDGE_INTERRUPT)
108
109 #define VM_EXIT_CTLS_ZERO_SETTING 0
110
111 #define VM_ENTRY_CTLS_ONE_SETTING \
112 (VM_ENTRY_LOAD_DEBUG_CONTROLS | \
113 VM_ENTRY_LOAD_EFER)
114
115 #define VM_ENTRY_CTLS_ZERO_SETTING \
116 (VM_ENTRY_INTO_SMM | \
117 VM_ENTRY_DEACTIVATE_DUAL_MONITOR)
118
119 #define HANDLED 1
120 #define UNHANDLED 0
121
122 static MALLOC_DEFINE(M_VMX, "vmx", "vmx");
123 static MALLOC_DEFINE(M_VLAPIC, "vlapic", "vlapic");
124
125 SYSCTL_DECL(_hw_vmm);
126 SYSCTL_NODE(_hw_vmm, OID_AUTO, vmx, CTLFLAG_RW, NULL, NULL);
127
128 int vmxon_enabled[MAXCPU];
129 static char vmxon_region[MAXCPU][PAGE_SIZE] __aligned(PAGE_SIZE);
130
131 static uint32_t pinbased_ctls, procbased_ctls, procbased_ctls2;
132 static uint32_t exit_ctls, entry_ctls;
133
134 static uint64_t cr0_ones_mask, cr0_zeros_mask;
135 SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr0_ones_mask, CTLFLAG_RD,
136 &cr0_ones_mask, 0, NULL);
137 SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr0_zeros_mask, CTLFLAG_RD,
138 &cr0_zeros_mask, 0, NULL);
139
140 static uint64_t cr4_ones_mask, cr4_zeros_mask;
141 SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr4_ones_mask, CTLFLAG_RD,
142 &cr4_ones_mask, 0, NULL);
143 SYSCTL_ULONG(_hw_vmm_vmx, OID_AUTO, cr4_zeros_mask, CTLFLAG_RD,
144 &cr4_zeros_mask, 0, NULL);
145
146 static int vmx_initialized;
147 SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, initialized, CTLFLAG_RD,
148 &vmx_initialized, 0, "Intel VMX initialized");
149
150 /*
151 * Optional capabilities
152 */
153 static SYSCTL_NODE(_hw_vmm_vmx, OID_AUTO, cap, CTLFLAG_RW, NULL, NULL);
154
155 static int cap_halt_exit;
156 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, halt_exit, CTLFLAG_RD, &cap_halt_exit, 0,
157 "HLT triggers a VM-exit");
158
159 static int cap_pause_exit;
160 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, pause_exit, CTLFLAG_RD, &cap_pause_exit,
161 0, "PAUSE triggers a VM-exit");
162
163 static int cap_rdpid;
164 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, rdpid, CTLFLAG_RD, &cap_rdpid, 0,
165 "Guests are allowed to use RDPID");
166
167 static int cap_rdtscp;
168 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, rdtscp, CTLFLAG_RD, &cap_rdtscp, 0,
169 "Guests are allowed to use RDTSCP");
170
171 static int cap_unrestricted_guest;
172 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, unrestricted_guest, CTLFLAG_RD,
173 &cap_unrestricted_guest, 0, "Unrestricted guests");
174
175 static int cap_monitor_trap;
176 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, monitor_trap, CTLFLAG_RD,
177 &cap_monitor_trap, 0, "Monitor trap flag");
178
179 static int cap_invpcid;
180 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, invpcid, CTLFLAG_RD, &cap_invpcid,
181 0, "Guests are allowed to use INVPCID");
182
183 static int tpr_shadowing;
184 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, tpr_shadowing, CTLFLAG_RD,
185 &tpr_shadowing, 0, "TPR shadowing support");
186
187 static int virtual_interrupt_delivery;
188 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, virtual_interrupt_delivery, CTLFLAG_RD,
189 &virtual_interrupt_delivery, 0, "APICv virtual interrupt delivery support");
190
191 static int posted_interrupts;
192 SYSCTL_INT(_hw_vmm_vmx_cap, OID_AUTO, posted_interrupts, CTLFLAG_RD,
193 &posted_interrupts, 0, "APICv posted interrupt support");
194
195 static int pirvec = -1;
196 SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, posted_interrupt_vector, CTLFLAG_RD,
197 &pirvec, 0, "APICv posted interrupt vector");
198
199 static struct unrhdr *vpid_unr;
200 static u_int vpid_alloc_failed;
201 SYSCTL_UINT(_hw_vmm_vmx, OID_AUTO, vpid_alloc_failed, CTLFLAG_RD,
202 &vpid_alloc_failed, 0, NULL);
203
204 int guest_l1d_flush;
205 SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, l1d_flush, CTLFLAG_RD,
206 &guest_l1d_flush, 0, NULL);
207 int guest_l1d_flush_sw;
208 SYSCTL_INT(_hw_vmm_vmx, OID_AUTO, l1d_flush_sw, CTLFLAG_RD,
209 &guest_l1d_flush_sw, 0, NULL);
210
211 static struct msr_entry msr_load_list[1] __aligned(16);
212
213 /*
214 * The definitions of SDT probes for VMX.
215 */
216
217 SDT_PROBE_DEFINE3(vmm, vmx, exit, entry,
218 "struct vmx *", "int", "struct vm_exit *");
219
220 SDT_PROBE_DEFINE4(vmm, vmx, exit, taskswitch,
221 "struct vmx *", "int", "struct vm_exit *", "struct vm_task_switch *");
222
223 SDT_PROBE_DEFINE4(vmm, vmx, exit, craccess,
224 "struct vmx *", "int", "struct vm_exit *", "uint64_t");
225
226 SDT_PROBE_DEFINE4(vmm, vmx, exit, rdmsr,
227 "struct vmx *", "int", "struct vm_exit *", "uint32_t");
228
229 SDT_PROBE_DEFINE5(vmm, vmx, exit, wrmsr,
230 "struct vmx *", "int", "struct vm_exit *", "uint32_t", "uint64_t");
231
232 SDT_PROBE_DEFINE3(vmm, vmx, exit, halt,
233 "struct vmx *", "int", "struct vm_exit *");
234
235 SDT_PROBE_DEFINE3(vmm, vmx, exit, mtrap,
236 "struct vmx *", "int", "struct vm_exit *");
237
238 SDT_PROBE_DEFINE3(vmm, vmx, exit, pause,
239 "struct vmx *", "int", "struct vm_exit *");
240
241 SDT_PROBE_DEFINE3(vmm, vmx, exit, intrwindow,
242 "struct vmx *", "int", "struct vm_exit *");
243
244 SDT_PROBE_DEFINE4(vmm, vmx, exit, interrupt,
245 "struct vmx *", "int", "struct vm_exit *", "uint32_t");
246
247 SDT_PROBE_DEFINE3(vmm, vmx, exit, nmiwindow,
248 "struct vmx *", "int", "struct vm_exit *");
249
250 SDT_PROBE_DEFINE3(vmm, vmx, exit, inout,
251 "struct vmx *", "int", "struct vm_exit *");
252
253 SDT_PROBE_DEFINE3(vmm, vmx, exit, cpuid,
254 "struct vmx *", "int", "struct vm_exit *");
255
256 SDT_PROBE_DEFINE5(vmm, vmx, exit, exception,
257 "struct vmx *", "int", "struct vm_exit *", "uint32_t", "int");
258
259 SDT_PROBE_DEFINE5(vmm, vmx, exit, nestedfault,
260 "struct vmx *", "int", "struct vm_exit *", "uint64_t", "uint64_t");
261
262 SDT_PROBE_DEFINE4(vmm, vmx, exit, mmiofault,
263 "struct vmx *", "int", "struct vm_exit *", "uint64_t");
264
265 SDT_PROBE_DEFINE3(vmm, vmx, exit, eoi,
266 "struct vmx *", "int", "struct vm_exit *");
267
268 SDT_PROBE_DEFINE3(vmm, vmx, exit, apicaccess,
269 "struct vmx *", "int", "struct vm_exit *");
270
271 SDT_PROBE_DEFINE4(vmm, vmx, exit, apicwrite,
272 "struct vmx *", "int", "struct vm_exit *", "struct vlapic *");
273
274 SDT_PROBE_DEFINE3(vmm, vmx, exit, xsetbv,
275 "struct vmx *", "int", "struct vm_exit *");
276
277 SDT_PROBE_DEFINE3(vmm, vmx, exit, monitor,
278 "struct vmx *", "int", "struct vm_exit *");
279
280 SDT_PROBE_DEFINE3(vmm, vmx, exit, mwait,
281 "struct vmx *", "int", "struct vm_exit *");
282
283 SDT_PROBE_DEFINE3(vmm, vmx, exit, vminsn,
284 "struct vmx *", "int", "struct vm_exit *");
285
286 SDT_PROBE_DEFINE4(vmm, vmx, exit, unknown,
287 "struct vmx *", "int", "struct vm_exit *", "uint32_t");
288
289 SDT_PROBE_DEFINE4(vmm, vmx, exit, return,
290 "struct vmx *", "int", "struct vm_exit *", "int");
291
292 /*
293 * Use the last page below 4GB as the APIC access address. This address is
294 * occupied by the boot firmware so it is guaranteed that it will not conflict
295 * with a page in system memory.
296 */
297 #define APIC_ACCESS_ADDRESS 0xFFFFF000
298
299 static int vmx_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc);
300 static int vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval);
301 static int vmxctx_setreg(struct vmxctx *vmxctx, int reg, uint64_t val);
302 static void vmx_inject_pir(struct vlapic *vlapic);
303
304 static inline bool
305 host_has_rdpid(void)
306 {
307 return ((cpu_stdext_feature2 & CPUID_STDEXT2_RDPID) != 0);
308 }
309
310 static inline bool
311 host_has_rdtscp(void)
312 {
313 return ((amd_feature & AMDID_RDTSCP) != 0);
314 }
315
316 #ifdef KTR
317 static const char *
318 exit_reason_to_str(int reason)
319 {
320 static char reasonbuf[32];
321
322 switch (reason) {
323 case EXIT_REASON_EXCEPTION:
324 return "exception";
325 case EXIT_REASON_EXT_INTR:
326 return "extint";
327 case EXIT_REASON_TRIPLE_FAULT:
328 return "triplefault";
329 case EXIT_REASON_INIT:
330 return "init";
331 case EXIT_REASON_SIPI:
332 return "sipi";
333 case EXIT_REASON_IO_SMI:
334 return "iosmi";
335 case EXIT_REASON_SMI:
336 return "smi";
337 case EXIT_REASON_INTR_WINDOW:
338 return "intrwindow";
339 case EXIT_REASON_NMI_WINDOW:
340 return "nmiwindow";
341 case EXIT_REASON_TASK_SWITCH:
342 return "taskswitch";
343 case EXIT_REASON_CPUID:
344 return "cpuid";
345 case EXIT_REASON_GETSEC:
346 return "getsec";
347 case EXIT_REASON_HLT:
348 return "hlt";
349 case EXIT_REASON_INVD:
350 return "invd";
351 case EXIT_REASON_INVLPG:
352 return "invlpg";
353 case EXIT_REASON_RDPMC:
354 return "rdpmc";
355 case EXIT_REASON_RDTSC:
356 return "rdtsc";
357 case EXIT_REASON_RSM:
358 return "rsm";
359 case EXIT_REASON_VMCALL:
360 return "vmcall";
361 case EXIT_REASON_VMCLEAR:
362 return "vmclear";
363 case EXIT_REASON_VMLAUNCH:
364 return "vmlaunch";
365 case EXIT_REASON_VMPTRLD:
366 return "vmptrld";
367 case EXIT_REASON_VMPTRST:
368 return "vmptrst";
369 case EXIT_REASON_VMREAD:
370 return "vmread";
371 case EXIT_REASON_VMRESUME:
372 return "vmresume";
373 case EXIT_REASON_VMWRITE:
374 return "vmwrite";
375 case EXIT_REASON_VMXOFF:
376 return "vmxoff";
377 case EXIT_REASON_VMXON:
378 return "vmxon";
379 case EXIT_REASON_CR_ACCESS:
380 return "craccess";
381 case EXIT_REASON_DR_ACCESS:
382 return "draccess";
383 case EXIT_REASON_INOUT:
384 return "inout";
385 case EXIT_REASON_RDMSR:
386 return "rdmsr";
387 case EXIT_REASON_WRMSR:
388 return "wrmsr";
389 case EXIT_REASON_INVAL_VMCS:
390 return "invalvmcs";
391 case EXIT_REASON_INVAL_MSR:
392 return "invalmsr";
393 case EXIT_REASON_MWAIT:
394 return "mwait";
395 case EXIT_REASON_MTF:
396 return "mtf";
397 case EXIT_REASON_MONITOR:
398 return "monitor";
399 case EXIT_REASON_PAUSE:
400 return "pause";
401 case EXIT_REASON_MCE_DURING_ENTRY:
402 return "mce-during-entry";
403 case EXIT_REASON_TPR:
404 return "tpr";
405 case EXIT_REASON_APIC_ACCESS:
406 return "apic-access";
407 case EXIT_REASON_GDTR_IDTR:
408 return "gdtridtr";
409 case EXIT_REASON_LDTR_TR:
410 return "ldtrtr";
411 case EXIT_REASON_EPT_FAULT:
412 return "eptfault";
413 case EXIT_REASON_EPT_MISCONFIG:
414 return "eptmisconfig";
415 case EXIT_REASON_INVEPT:
416 return "invept";
417 case EXIT_REASON_RDTSCP:
418 return "rdtscp";
419 case EXIT_REASON_VMX_PREEMPT:
420 return "vmxpreempt";
421 case EXIT_REASON_INVVPID:
422 return "invvpid";
423 case EXIT_REASON_WBINVD:
424 return "wbinvd";
425 case EXIT_REASON_XSETBV:
426 return "xsetbv";
427 case EXIT_REASON_APIC_WRITE:
428 return "apic-write";
429 default:
430 snprintf(reasonbuf, sizeof(reasonbuf), "%d", reason);
431 return (reasonbuf);
432 }
433 }
434 #endif /* KTR */
435
436 static int
437 vmx_allow_x2apic_msrs(struct vmx *vmx)
438 {
439 int i, error;
440
441 error = 0;
442
443 /*
444 * Allow readonly access to the following x2APIC MSRs from the guest.
445 */
446 error += guest_msr_ro(vmx, MSR_APIC_ID);
447 error += guest_msr_ro(vmx, MSR_APIC_VERSION);
448 error += guest_msr_ro(vmx, MSR_APIC_LDR);
449 error += guest_msr_ro(vmx, MSR_APIC_SVR);
450
451 for (i = 0; i < 8; i++)
452 error += guest_msr_ro(vmx, MSR_APIC_ISR0 + i);
453
454 for (i = 0; i < 8; i++)
455 error += guest_msr_ro(vmx, MSR_APIC_TMR0 + i);
456
457 for (i = 0; i < 8; i++)
458 error += guest_msr_ro(vmx, MSR_APIC_IRR0 + i);
459
460 error += guest_msr_ro(vmx, MSR_APIC_ESR);
461 error += guest_msr_ro(vmx, MSR_APIC_LVT_TIMER);
462 error += guest_msr_ro(vmx, MSR_APIC_LVT_THERMAL);
463 error += guest_msr_ro(vmx, MSR_APIC_LVT_PCINT);
464 error += guest_msr_ro(vmx, MSR_APIC_LVT_LINT0);
465 error += guest_msr_ro(vmx, MSR_APIC_LVT_LINT1);
466 error += guest_msr_ro(vmx, MSR_APIC_LVT_ERROR);
467 error += guest_msr_ro(vmx, MSR_APIC_ICR_TIMER);
468 error += guest_msr_ro(vmx, MSR_APIC_DCR_TIMER);
469 error += guest_msr_ro(vmx, MSR_APIC_ICR);
470
471 /*
472 * Allow TPR, EOI and SELF_IPI MSRs to be read and written by the guest.
473 *
474 * These registers get special treatment described in the section
475 * "Virtualizing MSR-Based APIC Accesses".
476 */
477 error += guest_msr_rw(vmx, MSR_APIC_TPR);
478 error += guest_msr_rw(vmx, MSR_APIC_EOI);
479 error += guest_msr_rw(vmx, MSR_APIC_SELF_IPI);
480
481 return (error);
482 }
483
484 u_long
485 vmx_fix_cr0(u_long cr0)
486 {
487
488 return ((cr0 | cr0_ones_mask) & ~cr0_zeros_mask);
489 }
490
491 u_long
492 vmx_fix_cr4(u_long cr4)
493 {
494
495 return ((cr4 | cr4_ones_mask) & ~cr4_zeros_mask);
496 }
497
498 static void
499 vpid_free(int vpid)
500 {
501 if (vpid < 0 || vpid > 0xffff)
502 panic("vpid_free: invalid vpid %d", vpid);
503
504 /*
505 * VPIDs [0,VM_MAXCPU] are special and are not allocated from
506 * the unit number allocator.
507 */
508
509 if (vpid > VM_MAXCPU)
510 free_unr(vpid_unr, vpid);
511 }
512
513 static void
514 vpid_alloc(uint16_t *vpid, int num)
515 {
516 int i, x;
517
518 if (num <= 0 || num > VM_MAXCPU)
519 panic("invalid number of vpids requested: %d", num);
520
521 /*
522 * If the "enable vpid" execution control is not enabled then the
523 * VPID is required to be 0 for all vcpus.
524 */
525 if ((procbased_ctls2 & PROCBASED2_ENABLE_VPID) == 0) {
526 for (i = 0; i < num; i++)
527 vpid[i] = 0;
528 return;
529 }
530
531 /*
532 * Allocate a unique VPID for each vcpu from the unit number allocator.
533 */
534 for (i = 0; i < num; i++) {
535 x = alloc_unr(vpid_unr);
536 if (x == -1)
537 break;
538 else
539 vpid[i] = x;
540 }
541
542 if (i < num) {
543 atomic_add_int(&vpid_alloc_failed, 1);
544
545 /*
546 * If the unit number allocator does not have enough unique
547 * VPIDs then we need to allocate from the [1,VM_MAXCPU] range.
548 *
549 * These VPIDs are not be unique across VMs but this does not
550 * affect correctness because the combined mappings are also
551 * tagged with the EP4TA which is unique for each VM.
552 *
553 * It is still sub-optimal because the invvpid will invalidate
554 * combined mappings for a particular VPID across all EP4TAs.
555 */
556 while (i-- > 0)
557 vpid_free(vpid[i]);
558
559 for (i = 0; i < num; i++)
560 vpid[i] = i + 1;
561 }
562 }
563
564 static void
565 vpid_init(void)
566 {
567 /*
568 * VPID 0 is required when the "enable VPID" execution control is
569 * disabled.
570 *
571 * VPIDs [1,VM_MAXCPU] are used as the "overflow namespace" when the
572 * unit number allocator does not have sufficient unique VPIDs to
573 * satisfy the allocation.
574 *
575 * The remaining VPIDs are managed by the unit number allocator.
576 */
577 vpid_unr = new_unrhdr(VM_MAXCPU + 1, 0xffff, NULL);
578 }
579
580 static void
581 vmx_disable(void *arg __unused)
582 {
583 struct invvpid_desc invvpid_desc = { 0 };
584 struct invept_desc invept_desc = { 0 };
585
586 if (vmxon_enabled[curcpu]) {
587 /*
588 * See sections 25.3.3.3 and 25.3.3.4 in Intel Vol 3b.
589 *
590 * VMXON or VMXOFF are not required to invalidate any TLB
591 * caching structures. This prevents potential retention of
592 * cached information in the TLB between distinct VMX episodes.
593 */
594 invvpid(INVVPID_TYPE_ALL_CONTEXTS, invvpid_desc);
595 invept(INVEPT_TYPE_ALL_CONTEXTS, invept_desc);
596 vmxoff();
597 }
598 load_cr4(rcr4() & ~CR4_VMXE);
599 }
600
601 static int
602 vmx_cleanup(void)
603 {
604
605 if (pirvec >= 0)
606 lapic_ipi_free(pirvec);
607
608 if (vpid_unr != NULL) {
609 delete_unrhdr(vpid_unr);
610 vpid_unr = NULL;
611 }
612
613 if (nmi_flush_l1d_sw == 1)
614 nmi_flush_l1d_sw = 0;
615
616 smp_rendezvous(NULL, vmx_disable, NULL, NULL);
617
618 return (0);
619 }
620
621 static void
622 vmx_enable(void *arg __unused)
623 {
624 int error;
625 uint64_t feature_control;
626
627 feature_control = rdmsr(MSR_IA32_FEATURE_CONTROL);
628 if ((feature_control & IA32_FEATURE_CONTROL_LOCK) == 0 ||
629 (feature_control & IA32_FEATURE_CONTROL_VMX_EN) == 0) {
630 wrmsr(MSR_IA32_FEATURE_CONTROL,
631 feature_control | IA32_FEATURE_CONTROL_VMX_EN |
632 IA32_FEATURE_CONTROL_LOCK);
633 }
634
635 load_cr4(rcr4() | CR4_VMXE);
636
637 *(uint32_t *)vmxon_region[curcpu] = vmx_revision();
638 error = vmxon(vmxon_region[curcpu]);
639 if (error == 0)
640 vmxon_enabled[curcpu] = 1;
641 }
642
643 static void
644 vmx_restore(void)
645 {
646
647 if (vmxon_enabled[curcpu])
648 vmxon(vmxon_region[curcpu]);
649 }
650
651 static int
652 vmx_init(int ipinum)
653 {
654 int error;
655 uint64_t basic, fixed0, fixed1, feature_control;
656 uint32_t tmp, procbased2_vid_bits;
657
658 /* CPUID.1:ECX[bit 5] must be 1 for processor to support VMX */
659 if (!(cpu_feature2 & CPUID2_VMX)) {
660 printf("vmx_init: processor does not support VMX operation\n");
661 return (ENXIO);
662 }
663
664 /*
665 * Verify that MSR_IA32_FEATURE_CONTROL lock and VMXON enable bits
666 * are set (bits 0 and 2 respectively).
667 */
668 feature_control = rdmsr(MSR_IA32_FEATURE_CONTROL);
669 if ((feature_control & IA32_FEATURE_CONTROL_LOCK) == 1 &&
670 (feature_control & IA32_FEATURE_CONTROL_VMX_EN) == 0) {
671 printf("vmx_init: VMX operation disabled by BIOS\n");
672 return (ENXIO);
673 }
674
675 /*
676 * Verify capabilities MSR_VMX_BASIC:
677 * - bit 54 indicates support for INS/OUTS decoding
678 */
679 basic = rdmsr(MSR_VMX_BASIC);
680 if ((basic & (1UL << 54)) == 0) {
681 printf("vmx_init: processor does not support desired basic "
682 "capabilities\n");
683 return (EINVAL);
684 }
685
686 /* Check support for primary processor-based VM-execution controls */
687 error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS,
688 MSR_VMX_TRUE_PROCBASED_CTLS,
689 PROCBASED_CTLS_ONE_SETTING,
690 PROCBASED_CTLS_ZERO_SETTING, &procbased_ctls);
691 if (error) {
692 printf("vmx_init: processor does not support desired primary "
693 "processor-based controls\n");
694 return (error);
695 }
696
697 /* Clear the processor-based ctl bits that are set on demand */
698 procbased_ctls &= ~PROCBASED_CTLS_WINDOW_SETTING;
699
700 /* Check support for secondary processor-based VM-execution controls */
701 error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2,
702 MSR_VMX_PROCBASED_CTLS2,
703 PROCBASED_CTLS2_ONE_SETTING,
704 PROCBASED_CTLS2_ZERO_SETTING, &procbased_ctls2);
705 if (error) {
706 printf("vmx_init: processor does not support desired secondary "
707 "processor-based controls\n");
708 return (error);
709 }
710
711 /* Check support for VPID */
712 error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2,
713 PROCBASED2_ENABLE_VPID, 0, &tmp);
714 if (error == 0)
715 procbased_ctls2 |= PROCBASED2_ENABLE_VPID;
716
717 /* Check support for pin-based VM-execution controls */
718 error = vmx_set_ctlreg(MSR_VMX_PINBASED_CTLS,
719 MSR_VMX_TRUE_PINBASED_CTLS,
720 PINBASED_CTLS_ONE_SETTING,
721 PINBASED_CTLS_ZERO_SETTING, &pinbased_ctls);
722 if (error) {
723 printf("vmx_init: processor does not support desired "
724 "pin-based controls\n");
725 return (error);
726 }
727
728 /* Check support for VM-exit controls */
729 error = vmx_set_ctlreg(MSR_VMX_EXIT_CTLS, MSR_VMX_TRUE_EXIT_CTLS,
730 VM_EXIT_CTLS_ONE_SETTING,
731 VM_EXIT_CTLS_ZERO_SETTING,
732 &exit_ctls);
733 if (error) {
734 printf("vmx_init: processor does not support desired "
735 "exit controls\n");
736 return (error);
737 }
738
739 /* Check support for VM-entry controls */
740 error = vmx_set_ctlreg(MSR_VMX_ENTRY_CTLS, MSR_VMX_TRUE_ENTRY_CTLS,
741 VM_ENTRY_CTLS_ONE_SETTING, VM_ENTRY_CTLS_ZERO_SETTING,
742 &entry_ctls);
743 if (error) {
744 printf("vmx_init: processor does not support desired "
745 "entry controls\n");
746 return (error);
747 }
748
749 /*
750 * Check support for optional features by testing them
751 * as individual bits
752 */
753 cap_halt_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS,
754 MSR_VMX_TRUE_PROCBASED_CTLS,
755 PROCBASED_HLT_EXITING, 0,
756 &tmp) == 0);
757
758 cap_monitor_trap = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS,
759 MSR_VMX_PROCBASED_CTLS,
760 PROCBASED_MTF, 0,
761 &tmp) == 0);
762
763 cap_pause_exit = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS,
764 MSR_VMX_TRUE_PROCBASED_CTLS,
765 PROCBASED_PAUSE_EXITING, 0,
766 &tmp) == 0);
767
768 /*
769 * Check support for RDPID and/or RDTSCP.
770 *
771 * Support a pass-through-based implementation of these via the
772 * "enable RDTSCP" VM-execution control and the "RDTSC exiting"
773 * VM-execution control.
774 *
775 * The "enable RDTSCP" VM-execution control applies to both RDPID
776 * and RDTSCP (see SDM volume 3, section 25.3, "Changes to
777 * Instruction Behavior in VMX Non-root operation"); this is why
778 * only this VM-execution control needs to be enabled in order to
779 * enable passing through whichever of RDPID and/or RDTSCP are
780 * supported by the host.
781 *
782 * The "RDTSC exiting" VM-execution control applies to both RDTSC
783 * and RDTSCP (again, per SDM volume 3, section 25.3), and is
784 * already set up for RDTSC and RDTSCP pass-through by the current
785 * implementation of RDTSC.
786 *
787 * Although RDPID and RDTSCP are optional capabilities, since there
788 * does not currently seem to be a use case for enabling/disabling
789 * these via libvmmapi, choose not to support this and, instead,
790 * just statically always enable or always disable this support
791 * across all vCPUs on all VMs. (Note that there may be some
792 * complications to providing this functionality, e.g., the MSR
793 * bitmap is currently per-VM rather than per-vCPU while the
794 * capability API wants to be able to control capabilities on a
795 * per-vCPU basis).
796 */
797 error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2,
798 MSR_VMX_PROCBASED_CTLS2,
799 PROCBASED2_ENABLE_RDTSCP, 0, &tmp);
800 cap_rdpid = error == 0 && host_has_rdpid();
801 cap_rdtscp = error == 0 && host_has_rdtscp();
802 if (cap_rdpid || cap_rdtscp)
803 procbased_ctls2 |= PROCBASED2_ENABLE_RDTSCP;
804
805 cap_unrestricted_guest = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2,
806 MSR_VMX_PROCBASED_CTLS2,
807 PROCBASED2_UNRESTRICTED_GUEST, 0,
808 &tmp) == 0);
809
810 cap_invpcid = (vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2,
811 MSR_VMX_PROCBASED_CTLS2, PROCBASED2_ENABLE_INVPCID, 0,
812 &tmp) == 0);
813
814 /*
815 * Check support for TPR shadow.
816 */
817 error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS,
818 MSR_VMX_TRUE_PROCBASED_CTLS, PROCBASED_USE_TPR_SHADOW, 0,
819 &tmp);
820 if (error == 0) {
821 tpr_shadowing = 1;
822 TUNABLE_INT_FETCH("hw.vmm.vmx.use_tpr_shadowing",
823 &tpr_shadowing);
824 }
825
826 if (tpr_shadowing) {
827 procbased_ctls |= PROCBASED_USE_TPR_SHADOW;
828 procbased_ctls &= ~PROCBASED_CR8_LOAD_EXITING;
829 procbased_ctls &= ~PROCBASED_CR8_STORE_EXITING;
830 }
831
832 /*
833 * Check support for virtual interrupt delivery.
834 */
835 procbased2_vid_bits = (PROCBASED2_VIRTUALIZE_APIC_ACCESSES |
836 PROCBASED2_VIRTUALIZE_X2APIC_MODE |
837 PROCBASED2_APIC_REGISTER_VIRTUALIZATION |
838 PROCBASED2_VIRTUAL_INTERRUPT_DELIVERY);
839
840 error = vmx_set_ctlreg(MSR_VMX_PROCBASED_CTLS2, MSR_VMX_PROCBASED_CTLS2,
841 procbased2_vid_bits, 0, &tmp);
842 if (error == 0 && tpr_shadowing) {
843 virtual_interrupt_delivery = 1;
844 TUNABLE_INT_FETCH("hw.vmm.vmx.use_apic_vid",
845 &virtual_interrupt_delivery);
846 }
847
848 if (virtual_interrupt_delivery) {
849 procbased_ctls |= PROCBASED_USE_TPR_SHADOW;
850 procbased_ctls2 |= procbased2_vid_bits;
851 procbased_ctls2 &= ~PROCBASED2_VIRTUALIZE_X2APIC_MODE;
852
853 /*
854 * Check for Posted Interrupts only if Virtual Interrupt
855 * Delivery is enabled.
856 */
857 error = vmx_set_ctlreg(MSR_VMX_PINBASED_CTLS,
858 MSR_VMX_TRUE_PINBASED_CTLS, PINBASED_POSTED_INTERRUPT, 0,
859 &tmp);
860 if (error == 0) {
861 pirvec = lapic_ipi_alloc(pti ? &IDTVEC(justreturn1_pti) :
862 &IDTVEC(justreturn));
863 if (pirvec < 0) {
864 if (bootverbose) {
865 printf("vmx_init: unable to allocate "
866 "posted interrupt vector\n");
867 }
868 } else {
869 posted_interrupts = 1;
870 TUNABLE_INT_FETCH("hw.vmm.vmx.use_apic_pir",
871 &posted_interrupts);
872 }
873 }
874 }
875
876 if (posted_interrupts)
877 pinbased_ctls |= PINBASED_POSTED_INTERRUPT;
878
879 /* Initialize EPT */
880 error = ept_init(ipinum);
881 if (error) {
882 printf("vmx_init: ept initialization failed (%d)\n", error);
883 return (error);
884 }
885
886 guest_l1d_flush = (cpu_ia32_arch_caps &
887 IA32_ARCH_CAP_SKIP_L1DFL_VMENTRY) == 0;
888 TUNABLE_INT_FETCH("hw.vmm.l1d_flush", &guest_l1d_flush);
889
890 /*
891 * L1D cache flush is enabled. Use IA32_FLUSH_CMD MSR when
892 * available. Otherwise fall back to the software flush
893 * method which loads enough data from the kernel text to
894 * flush existing L1D content, both on VMX entry and on NMI
895 * return.
896 */
897 if (guest_l1d_flush) {
898 if ((cpu_stdext_feature3 & CPUID_STDEXT3_L1D_FLUSH) == 0) {
899 guest_l1d_flush_sw = 1;
900 TUNABLE_INT_FETCH("hw.vmm.l1d_flush_sw",
901 &guest_l1d_flush_sw);
902 }
903 if (guest_l1d_flush_sw) {
904 if (nmi_flush_l1d_sw <= 1)
905 nmi_flush_l1d_sw = 1;
906 } else {
907 msr_load_list[0].index = MSR_IA32_FLUSH_CMD;
908 msr_load_list[0].val = IA32_FLUSH_CMD_L1D;
909 }
910 }
911
912 /*
913 * Stash the cr0 and cr4 bits that must be fixed to 0 or 1
914 */
915 fixed0 = rdmsr(MSR_VMX_CR0_FIXED0);
916 fixed1 = rdmsr(MSR_VMX_CR0_FIXED1);
917 cr0_ones_mask = fixed0 & fixed1;
918 cr0_zeros_mask = ~fixed0 & ~fixed1;
919
920 /*
921 * CR0_PE and CR0_PG can be set to zero in VMX non-root operation
922 * if unrestricted guest execution is allowed.
923 */
924 if (cap_unrestricted_guest)
925 cr0_ones_mask &= ~(CR0_PG | CR0_PE);
926
927 /*
928 * Do not allow the guest to set CR0_NW or CR0_CD.
929 */
930 cr0_zeros_mask |= (CR0_NW | CR0_CD);
931
932 fixed0 = rdmsr(MSR_VMX_CR4_FIXED0);
933 fixed1 = rdmsr(MSR_VMX_CR4_FIXED1);
934 cr4_ones_mask = fixed0 & fixed1;
935 cr4_zeros_mask = ~fixed0 & ~fixed1;
936
937 vpid_init();
938
939 vmx_msr_init();
940
941 /* enable VMX operation */
942 smp_rendezvous(NULL, vmx_enable, NULL, NULL);
943
944 vmx_initialized = 1;
945
946 return (0);
947 }
948
949 static void
950 vmx_trigger_hostintr(int vector)
951 {
952 uintptr_t func;
953 struct gate_descriptor *gd;
954
955 gd = &idt[vector];
956
957 KASSERT(vector >= 32 && vector <= 255, ("vmx_trigger_hostintr: "
958 "invalid vector %d", vector));
959 KASSERT(gd->gd_p == 1, ("gate descriptor for vector %d not present",
960 vector));
961 KASSERT(gd->gd_type == SDT_SYSIGT, ("gate descriptor for vector %d "
962 "has invalid type %d", vector, gd->gd_type));
963 KASSERT(gd->gd_dpl == SEL_KPL, ("gate descriptor for vector %d "
964 "has invalid dpl %d", vector, gd->gd_dpl));
965 KASSERT(gd->gd_selector == GSEL(GCODE_SEL, SEL_KPL), ("gate descriptor "
966 "for vector %d has invalid selector %d", vector, gd->gd_selector));
967 KASSERT(gd->gd_ist == 0, ("gate descriptor for vector %d has invalid "
968 "IST %d", vector, gd->gd_ist));
969
970 func = ((long)gd->gd_hioffset << 16 | gd->gd_looffset);
971 vmx_call_isr(func);
972 }
973
974 static int
975 vmx_setup_cr_shadow(int which, struct vmcs *vmcs, uint32_t initial)
976 {
977 int error, mask_ident, shadow_ident;
978 uint64_t mask_value;
979
980 if (which != 0 && which != 4)
981 panic("vmx_setup_cr_shadow: unknown cr%d", which);
982
983 if (which == 0) {
984 mask_ident = VMCS_CR0_MASK;
985 mask_value = cr0_ones_mask | cr0_zeros_mask;
986 shadow_ident = VMCS_CR0_SHADOW;
987 } else {
988 mask_ident = VMCS_CR4_MASK;
989 mask_value = cr4_ones_mask | cr4_zeros_mask;
990 shadow_ident = VMCS_CR4_SHADOW;
991 }
992
993 error = vmcs_setreg(vmcs, 0, VMCS_IDENT(mask_ident), mask_value);
994 if (error)
995 return (error);
996
997 error = vmcs_setreg(vmcs, 0, VMCS_IDENT(shadow_ident), initial);
998 if (error)
999 return (error);
1000
1001 return (0);
1002 }
1003 #define vmx_setup_cr0_shadow(vmcs,init) vmx_setup_cr_shadow(0, (vmcs), (init))
1004 #define vmx_setup_cr4_shadow(vmcs,init) vmx_setup_cr_shadow(4, (vmcs), (init))
1005
1006 static void *
1007 vmx_vminit(struct vm *vm, pmap_t pmap)
1008 {
1009 uint16_t vpid[VM_MAXCPU];
1010 int i, error;
1011 struct vmx *vmx;
1012 struct vmcs *vmcs;
1013 uint32_t exc_bitmap;
1014 uint16_t maxcpus;
1015
1016 vmx = malloc(sizeof(struct vmx), M_VMX, M_WAITOK | M_ZERO);
1017 if ((uintptr_t)vmx & PAGE_MASK) {
1018 panic("malloc of struct vmx not aligned on %d byte boundary",
1019 PAGE_SIZE);
1020 }
1021 vmx->vm = vm;
1022
1023 vmx->eptp = eptp(vtophys((vm_offset_t)pmap->pm_pml4));
1024
1025 /*
1026 * Clean up EPTP-tagged guest physical and combined mappings
1027 *
1028 * VMX transitions are not required to invalidate any guest physical
1029 * mappings. So, it may be possible for stale guest physical mappings
1030 * to be present in the processor TLBs.
1031 *
1032 * Combined mappings for this EP4TA are also invalidated for all VPIDs.
1033 */
1034 ept_invalidate_mappings(vmx->eptp);
1035
1036 msr_bitmap_initialize(vmx->msr_bitmap);
1037
1038 /*
1039 * It is safe to allow direct access to MSR_GSBASE and MSR_FSBASE.
1040 * The guest FSBASE and GSBASE are saved and restored during
1041 * vm-exit and vm-entry respectively. The host FSBASE and GSBASE are
1042 * always restored from the vmcs host state area on vm-exit.
1043 *
1044 * The SYSENTER_CS/ESP/EIP MSRs are identical to FS/GSBASE in
1045 * how they are saved/restored so can be directly accessed by the
1046 * guest.
1047 *
1048 * MSR_EFER is saved and restored in the guest VMCS area on a
1049 * VM exit and entry respectively. It is also restored from the
1050 * host VMCS area on a VM exit.
1051 *
1052 * The TSC MSR is exposed read-only. Writes are disallowed as
1053 * that will impact the host TSC. If the guest does a write
1054 * the "use TSC offsetting" execution control is enabled and the
1055 * difference between the host TSC and the guest TSC is written
1056 * into the TSC offset in the VMCS.
1057 *
1058 * Guest TSC_AUX support is enabled if any of guest RDPID and/or
1059 * guest RDTSCP support are enabled (since, as per Table 2-2 in SDM
1060 * volume 4, TSC_AUX is supported if any of RDPID and/or RDTSCP are
1061 * supported). If guest TSC_AUX support is enabled, TSC_AUX is
1062 * exposed read-only so that the VMM can do one fewer MSR read per
1063 * exit than if this register were exposed read-write; the guest
1064 * restore value can be updated during guest writes (expected to be
1065 * rare) instead of during all exits (common).
1066 */
1067 if (guest_msr_rw(vmx, MSR_GSBASE) ||
1068 guest_msr_rw(vmx, MSR_FSBASE) ||
1069 guest_msr_rw(vmx, MSR_SYSENTER_CS_MSR) ||
1070 guest_msr_rw(vmx, MSR_SYSENTER_ESP_MSR) ||
1071 guest_msr_rw(vmx, MSR_SYSENTER_EIP_MSR) ||
1072 guest_msr_rw(vmx, MSR_EFER) ||
1073 guest_msr_ro(vmx, MSR_TSC) ||
1074 ((cap_rdpid || cap_rdtscp) && guest_msr_ro(vmx, MSR_TSC_AUX)))
1075 panic("vmx_vminit: error setting guest msr access");
1076
1077 vpid_alloc(vpid, VM_MAXCPU);
1078
1079 if (virtual_interrupt_delivery) {
1080 error = vm_map_mmio(vm, DEFAULT_APIC_BASE, PAGE_SIZE,
1081 APIC_ACCESS_ADDRESS);
1082 /* XXX this should really return an error to the caller */
1083 KASSERT(error == 0, ("vm_map_mmio(apicbase) error %d", error));
1084 }
1085
1086 maxcpus = vm_get_maxcpus(vm);
1087 for (i = 0; i < maxcpus; i++) {
1088 vmcs = &vmx->vmcs[i];
1089 vmcs->identifier = vmx_revision();
1090 error = vmclear(vmcs);
1091 if (error != 0) {
1092 panic("vmx_vminit: vmclear error %d on vcpu %d\n",
1093 error, i);
1094 }
1095
1096 vmx_msr_guest_init(vmx, i);
1097
1098 error = vmcs_init(vmcs);
1099 KASSERT(error == 0, ("vmcs_init error %d", error));
1100
1101 VMPTRLD(vmcs);
1102 error = 0;
1103 error += vmwrite(VMCS_HOST_RSP, (u_long)&vmx->ctx[i]);
1104 error += vmwrite(VMCS_EPTP, vmx->eptp);
1105 error += vmwrite(VMCS_PIN_BASED_CTLS, pinbased_ctls);
1106 error += vmwrite(VMCS_PRI_PROC_BASED_CTLS, procbased_ctls);
1107 error += vmwrite(VMCS_SEC_PROC_BASED_CTLS, procbased_ctls2);
1108 error += vmwrite(VMCS_EXIT_CTLS, exit_ctls);
1109 error += vmwrite(VMCS_ENTRY_CTLS, entry_ctls);
1110 error += vmwrite(VMCS_MSR_BITMAP, vtophys(vmx->msr_bitmap));
1111 error += vmwrite(VMCS_VPID, vpid[i]);
1112
1113 if (guest_l1d_flush && !guest_l1d_flush_sw) {
1114 vmcs_write(VMCS_ENTRY_MSR_LOAD, pmap_kextract(
1115 (vm_offset_t)&msr_load_list[0]));
1116 vmcs_write(VMCS_ENTRY_MSR_LOAD_COUNT,
1117 nitems(msr_load_list));
1118 vmcs_write(VMCS_EXIT_MSR_STORE, 0);
1119 vmcs_write(VMCS_EXIT_MSR_STORE_COUNT, 0);
1120 }
1121
1122 /* exception bitmap */
1123 if (vcpu_trace_exceptions(vm, i))
1124 exc_bitmap = 0xffffffff;
1125 else
1126 exc_bitmap = 1 << IDT_MC;
1127 error += vmwrite(VMCS_EXCEPTION_BITMAP, exc_bitmap);
1128
1129 vmx->ctx[i].guest_dr6 = DBREG_DR6_RESERVED1;
1130 error += vmwrite(VMCS_GUEST_DR7, DBREG_DR7_RESERVED1);
1131
1132 if (tpr_shadowing) {
1133 error += vmwrite(VMCS_VIRTUAL_APIC,
1134 vtophys(&vmx->apic_page[i]));
1135 }
1136
1137 if (virtual_interrupt_delivery) {
1138 error += vmwrite(VMCS_APIC_ACCESS, APIC_ACCESS_ADDRESS);
1139 error += vmwrite(VMCS_EOI_EXIT0, 0);
1140 error += vmwrite(VMCS_EOI_EXIT1, 0);
1141 error += vmwrite(VMCS_EOI_EXIT2, 0);
1142 error += vmwrite(VMCS_EOI_EXIT3, 0);
1143 }
1144 if (posted_interrupts) {
1145 error += vmwrite(VMCS_PIR_VECTOR, pirvec);
1146 error += vmwrite(VMCS_PIR_DESC,
1147 vtophys(&vmx->pir_desc[i]));
1148 }
1149 VMCLEAR(vmcs);
1150 KASSERT(error == 0, ("vmx_vminit: error customizing the vmcs"));
1151
1152 vmx->cap[i].set = 0;
1153 vmx->cap[i].set |= cap_rdpid != 0 ? 1 << VM_CAP_RDPID : 0;
1154 vmx->cap[i].set |= cap_rdtscp != 0 ? 1 << VM_CAP_RDTSCP : 0;
1155 vmx->cap[i].proc_ctls = procbased_ctls;
1156 vmx->cap[i].proc_ctls2 = procbased_ctls2;
1157 vmx->cap[i].exc_bitmap = exc_bitmap;
1158
1159 vmx->state[i].nextrip = ~0;
1160 vmx->state[i].lastcpu = NOCPU;
1161 vmx->state[i].vpid = vpid[i];
1162
1163 /*
1164 * Set up the CR0/4 shadows, and init the read shadow
1165 * to the power-on register value from the Intel Sys Arch.
1166 * CR0 - 0x60000010
1167 * CR4 - 0
1168 */
1169 error = vmx_setup_cr0_shadow(vmcs, 0x60000010);
1170 if (error != 0)
1171 panic("vmx_setup_cr0_shadow %d", error);
1172
1173 error = vmx_setup_cr4_shadow(vmcs, 0);
1174 if (error != 0)
1175 panic("vmx_setup_cr4_shadow %d", error);
1176
1177 vmx->ctx[i].pmap = pmap;
1178 }
1179
1180 return (vmx);
1181 }
1182
1183 static int
1184 vmx_handle_cpuid(struct vm *vm, int vcpu, struct vmxctx *vmxctx)
1185 {
1186 int handled, func;
1187
1188 func = vmxctx->guest_rax;
1189
1190 handled = x86_emulate_cpuid(vm, vcpu,
1191 (uint32_t*)(&vmxctx->guest_rax),
1192 (uint32_t*)(&vmxctx->guest_rbx),
1193 (uint32_t*)(&vmxctx->guest_rcx),
1194 (uint32_t*)(&vmxctx->guest_rdx));
1195 return (handled);
1196 }
1197
1198 static __inline void
1199 vmx_run_trace(struct vmx *vmx, int vcpu)
1200 {
1201 #ifdef KTR
1202 VCPU_CTR1(vmx->vm, vcpu, "Resume execution at %#lx", vmcs_guest_rip());
1203 #endif
1204 }
1205
1206 static __inline void
1207 vmx_exit_trace(struct vmx *vmx, int vcpu, uint64_t rip, uint32_t exit_reason,
1208 int handled)
1209 {
1210 #ifdef KTR
1211 VCPU_CTR3(vmx->vm, vcpu, "%s %s vmexit at 0x%0lx",
1212 handled ? "handled" : "unhandled",
1213 exit_reason_to_str(exit_reason), rip);
1214 #endif
1215 }
1216
1217 static __inline void
1218 vmx_astpending_trace(struct vmx *vmx, int vcpu, uint64_t rip)
1219 {
1220 #ifdef KTR
1221 VCPU_CTR1(vmx->vm, vcpu, "astpending vmexit at 0x%0lx", rip);
1222 #endif
1223 }
1224
1225 static VMM_STAT_INTEL(VCPU_INVVPID_SAVED, "Number of vpid invalidations saved");
1226 static VMM_STAT_INTEL(VCPU_INVVPID_DONE, "Number of vpid invalidations done");
1227
1228 /*
1229 * Invalidate guest mappings identified by its vpid from the TLB.
1230 */
1231 static __inline void
1232 vmx_invvpid(struct vmx *vmx, int vcpu, pmap_t pmap, int running)
1233 {
1234 struct vmxstate *vmxstate;
1235 struct invvpid_desc invvpid_desc;
1236
1237 vmxstate = &vmx->state[vcpu];
1238 if (vmxstate->vpid == 0)
1239 return;
1240
1241 if (!running) {
1242 /*
1243 * Set the 'lastcpu' to an invalid host cpu.
1244 *
1245 * This will invalidate TLB entries tagged with the vcpu's
1246 * vpid the next time it runs via vmx_set_pcpu_defaults().
1247 */
1248 vmxstate->lastcpu = NOCPU;
1249 return;
1250 }
1251
1252 KASSERT(curthread->td_critnest > 0, ("%s: vcpu %d running outside "
1253 "critical section", __func__, vcpu));
1254
1255 /*
1256 * Invalidate all mappings tagged with 'vpid'
1257 *
1258 * We do this because this vcpu was executing on a different host
1259 * cpu when it last ran. We do not track whether it invalidated
1260 * mappings associated with its 'vpid' during that run. So we must
1261 * assume that the mappings associated with 'vpid' on 'curcpu' are
1262 * stale and invalidate them.
1263 *
1264 * Note that we incur this penalty only when the scheduler chooses to
1265 * move the thread associated with this vcpu between host cpus.
1266 *
1267 * Note also that this will invalidate mappings tagged with 'vpid'
1268 * for "all" EP4TAs.
1269 */
1270 if (pmap->pm_eptgen == vmx->eptgen[curcpu]) {
1271 invvpid_desc._res1 = 0;
1272 invvpid_desc._res2 = 0;
1273 invvpid_desc.vpid = vmxstate->vpid;
1274 invvpid_desc.linear_addr = 0;
1275 invvpid(INVVPID_TYPE_SINGLE_CONTEXT, invvpid_desc);
1276 vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_DONE, 1);
1277 } else {
1278 /*
1279 * The invvpid can be skipped if an invept is going to
1280 * be performed before entering the guest. The invept
1281 * will invalidate combined mappings tagged with
1282 * 'vmx->eptp' for all vpids.
1283 */
1284 vmm_stat_incr(vmx->vm, vcpu, VCPU_INVVPID_SAVED, 1);
1285 }
1286 }
1287
1288 static void
1289 vmx_set_pcpu_defaults(struct vmx *vmx, int vcpu, pmap_t pmap)
1290 {
1291 struct vmxstate *vmxstate;
1292
1293 vmxstate = &vmx->state[vcpu];
1294 if (vmxstate->lastcpu == curcpu)
1295 return;
1296
1297 vmxstate->lastcpu = curcpu;
1298
1299 vmm_stat_incr(vmx->vm, vcpu, VCPU_MIGRATIONS, 1);
1300
1301 vmcs_write(VMCS_HOST_TR_BASE, vmm_get_host_trbase());
1302 vmcs_write(VMCS_HOST_GDTR_BASE, vmm_get_host_gdtrbase());
1303 vmcs_write(VMCS_HOST_GS_BASE, vmm_get_host_gsbase());
1304 vmx_invvpid(vmx, vcpu, pmap, 1);
1305 }
1306
1307 /*
1308 * We depend on 'procbased_ctls' to have the Interrupt Window Exiting bit set.
1309 */
1310 CTASSERT((PROCBASED_CTLS_ONE_SETTING & PROCBASED_INT_WINDOW_EXITING) != 0);
1311
1312 static void __inline
1313 vmx_set_int_window_exiting(struct vmx *vmx, int vcpu)
1314 {
1315
1316 if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) == 0) {
1317 vmx->cap[vcpu].proc_ctls |= PROCBASED_INT_WINDOW_EXITING;
1318 vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls);
1319 VCPU_CTR0(vmx->vm, vcpu, "Enabling interrupt window exiting");
1320 }
1321 }
1322
1323 static void __inline
1324 vmx_clear_int_window_exiting(struct vmx *vmx, int vcpu)
1325 {
1326
1327 KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0,
1328 ("intr_window_exiting not set: %#x", vmx->cap[vcpu].proc_ctls));
1329 vmx->cap[vcpu].proc_ctls &= ~PROCBASED_INT_WINDOW_EXITING;
1330 vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls);
1331 VCPU_CTR0(vmx->vm, vcpu, "Disabling interrupt window exiting");
1332 }
1333
1334 static void __inline
1335 vmx_set_nmi_window_exiting(struct vmx *vmx, int vcpu)
1336 {
1337
1338 if ((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) == 0) {
1339 vmx->cap[vcpu].proc_ctls |= PROCBASED_NMI_WINDOW_EXITING;
1340 vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls);
1341 VCPU_CTR0(vmx->vm, vcpu, "Enabling NMI window exiting");
1342 }
1343 }
1344
1345 static void __inline
1346 vmx_clear_nmi_window_exiting(struct vmx *vmx, int vcpu)
1347 {
1348
1349 KASSERT((vmx->cap[vcpu].proc_ctls & PROCBASED_NMI_WINDOW_EXITING) != 0,
1350 ("nmi_window_exiting not set %#x", vmx->cap[vcpu].proc_ctls));
1351 vmx->cap[vcpu].proc_ctls &= ~PROCBASED_NMI_WINDOW_EXITING;
1352 vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls);
1353 VCPU_CTR0(vmx->vm, vcpu, "Disabling NMI window exiting");
1354 }
1355
1356 int
1357 vmx_set_tsc_offset(struct vmx *vmx, int vcpu, uint64_t offset)
1358 {
1359 int error;
1360
1361 if ((vmx->cap[vcpu].proc_ctls & PROCBASED_TSC_OFFSET) == 0) {
1362 vmx->cap[vcpu].proc_ctls |= PROCBASED_TSC_OFFSET;
1363 vmcs_write(VMCS_PRI_PROC_BASED_CTLS, vmx->cap[vcpu].proc_ctls);
1364 VCPU_CTR0(vmx->vm, vcpu, "Enabling TSC offsetting");
1365 }
1366
1367 error = vmwrite(VMCS_TSC_OFFSET, offset);
1368
1369 return (error);
1370 }
1371
1372 #define NMI_BLOCKING (VMCS_INTERRUPTIBILITY_NMI_BLOCKING | \
1373 VMCS_INTERRUPTIBILITY_MOVSS_BLOCKING)
1374 #define HWINTR_BLOCKING (VMCS_INTERRUPTIBILITY_STI_BLOCKING | \
1375 VMCS_INTERRUPTIBILITY_MOVSS_BLOCKING)
1376
1377 static void
1378 vmx_inject_nmi(struct vmx *vmx, int vcpu)
1379 {
1380 uint32_t gi, info;
1381
1382 gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY);
1383 KASSERT((gi & NMI_BLOCKING) == 0, ("vmx_inject_nmi: invalid guest "
1384 "interruptibility-state %#x", gi));
1385
1386 info = vmcs_read(VMCS_ENTRY_INTR_INFO);
1387 KASSERT((info & VMCS_INTR_VALID) == 0, ("vmx_inject_nmi: invalid "
1388 "VM-entry interruption information %#x", info));
1389
1390 /*
1391 * Inject the virtual NMI. The vector must be the NMI IDT entry
1392 * or the VMCS entry check will fail.
1393 */
1394 info = IDT_NMI | VMCS_INTR_T_NMI | VMCS_INTR_VALID;
1395 vmcs_write(VMCS_ENTRY_INTR_INFO, info);
1396
1397 VCPU_CTR0(vmx->vm, vcpu, "Injecting vNMI");
1398
1399 /* Clear the request */
1400 vm_nmi_clear(vmx->vm, vcpu);
1401 }
1402
1403 static void
1404 vmx_inject_interrupts(struct vmx *vmx, int vcpu, struct vlapic *vlapic,
1405 uint64_t guestrip)
1406 {
1407 int vector, need_nmi_exiting, extint_pending;
1408 uint64_t rflags, entryinfo;
1409 uint32_t gi, info;
1410
1411 if (vmx->state[vcpu].nextrip != guestrip) {
1412 gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY);
1413 if (gi & HWINTR_BLOCKING) {
1414 VCPU_CTR2(vmx->vm, vcpu, "Guest interrupt blocking "
1415 "cleared due to rip change: %#lx/%#lx",
1416 vmx->state[vcpu].nextrip, guestrip);
1417 gi &= ~HWINTR_BLOCKING;
1418 vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi);
1419 }
1420 }
1421
1422 if (vm_entry_intinfo(vmx->vm, vcpu, &entryinfo)) {
1423 KASSERT((entryinfo & VMCS_INTR_VALID) != 0, ("%s: entry "
1424 "intinfo is not valid: %#lx", __func__, entryinfo));
1425
1426 info = vmcs_read(VMCS_ENTRY_INTR_INFO);
1427 KASSERT((info & VMCS_INTR_VALID) == 0, ("%s: cannot inject "
1428 "pending exception: %#lx/%#x", __func__, entryinfo, info));
1429
1430 info = entryinfo;
1431 vector = info & 0xff;
1432 if (vector == IDT_BP || vector == IDT_OF) {
1433 /*
1434 * VT-x requires #BP and #OF to be injected as software
1435 * exceptions.
1436 */
1437 info &= ~VMCS_INTR_T_MASK;
1438 info |= VMCS_INTR_T_SWEXCEPTION;
1439 }
1440
1441 if (info & VMCS_INTR_DEL_ERRCODE)
1442 vmcs_write(VMCS_ENTRY_EXCEPTION_ERROR, entryinfo >> 32);
1443
1444 vmcs_write(VMCS_ENTRY_INTR_INFO, info);
1445 }
1446
1447 if (vm_nmi_pending(vmx->vm, vcpu)) {
1448 /*
1449 * If there are no conditions blocking NMI injection then
1450 * inject it directly here otherwise enable "NMI window
1451 * exiting" to inject it as soon as we can.
1452 *
1453 * We also check for STI_BLOCKING because some implementations
1454 * don't allow NMI injection in this case. If we are running
1455 * on a processor that doesn't have this restriction it will
1456 * immediately exit and the NMI will be injected in the
1457 * "NMI window exiting" handler.
1458 */
1459 need_nmi_exiting = 1;
1460 gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY);
1461 if ((gi & (HWINTR_BLOCKING | NMI_BLOCKING)) == 0) {
1462 info = vmcs_read(VMCS_ENTRY_INTR_INFO);
1463 if ((info & VMCS_INTR_VALID) == 0) {
1464 vmx_inject_nmi(vmx, vcpu);
1465 need_nmi_exiting = 0;
1466 } else {
1467 VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI "
1468 "due to VM-entry intr info %#x", info);
1469 }
1470 } else {
1471 VCPU_CTR1(vmx->vm, vcpu, "Cannot inject NMI due to "
1472 "Guest Interruptibility-state %#x", gi);
1473 }
1474
1475 if (need_nmi_exiting)
1476 vmx_set_nmi_window_exiting(vmx, vcpu);
1477 }
1478
1479 extint_pending = vm_extint_pending(vmx->vm, vcpu);
1480
1481 if (!extint_pending && virtual_interrupt_delivery) {
1482 vmx_inject_pir(vlapic);
1483 return;
1484 }
1485
1486 /*
1487 * If interrupt-window exiting is already in effect then don't bother
1488 * checking for pending interrupts. This is just an optimization and
1489 * not needed for correctness.
1490 */
1491 if ((vmx->cap[vcpu].proc_ctls & PROCBASED_INT_WINDOW_EXITING) != 0) {
1492 VCPU_CTR0(vmx->vm, vcpu, "Skip interrupt injection due to "
1493 "pending int_window_exiting");
1494 return;
1495 }
1496
1497 if (!extint_pending) {
1498 /* Ask the local apic for a vector to inject */
1499 if (!vlapic_pending_intr(vlapic, &vector))
1500 return;
1501
1502 /*
1503 * From the Intel SDM, Volume 3, Section "Maskable
1504 * Hardware Interrupts":
1505 * - maskable interrupt vectors [16,255] can be delivered
1506 * through the local APIC.
1507 */
1508 KASSERT(vector >= 16 && vector <= 255,
1509 ("invalid vector %d from local APIC", vector));
1510 } else {
1511 /* Ask the legacy pic for a vector to inject */
1512 vatpic_pending_intr(vmx->vm, &vector);
1513
1514 /*
1515 * From the Intel SDM, Volume 3, Section "Maskable
1516 * Hardware Interrupts":
1517 * - maskable interrupt vectors [0,255] can be delivered
1518 * through the INTR pin.
1519 */
1520 KASSERT(vector >= 0 && vector <= 255,
1521 ("invalid vector %d from INTR", vector));
1522 }
1523
1524 /* Check RFLAGS.IF and the interruptibility state of the guest */
1525 rflags = vmcs_read(VMCS_GUEST_RFLAGS);
1526 if ((rflags & PSL_I) == 0) {
1527 VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to "
1528 "rflags %#lx", vector, rflags);
1529 goto cantinject;
1530 }
1531
1532 gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY);
1533 if (gi & HWINTR_BLOCKING) {
1534 VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to "
1535 "Guest Interruptibility-state %#x", vector, gi);
1536 goto cantinject;
1537 }
1538
1539 info = vmcs_read(VMCS_ENTRY_INTR_INFO);
1540 if (info & VMCS_INTR_VALID) {
1541 /*
1542 * This is expected and could happen for multiple reasons:
1543 * - A vectoring VM-entry was aborted due to astpending
1544 * - A VM-exit happened during event injection.
1545 * - An exception was injected above.
1546 * - An NMI was injected above or after "NMI window exiting"
1547 */
1548 VCPU_CTR2(vmx->vm, vcpu, "Cannot inject vector %d due to "
1549 "VM-entry intr info %#x", vector, info);
1550 goto cantinject;
1551 }
1552
1553 /* Inject the interrupt */
1554 info = VMCS_INTR_T_HWINTR | VMCS_INTR_VALID;
1555 info |= vector;
1556 vmcs_write(VMCS_ENTRY_INTR_INFO, info);
1557
1558 if (!extint_pending) {
1559 /* Update the Local APIC ISR */
1560 vlapic_intr_accepted(vlapic, vector);
1561 } else {
1562 vm_extint_clear(vmx->vm, vcpu);
1563 vatpic_intr_accepted(vmx->vm, vector);
1564
1565 /*
1566 * After we accepted the current ExtINT the PIC may
1567 * have posted another one. If that is the case, set
1568 * the Interrupt Window Exiting execution control so
1569 * we can inject that one too.
1570 *
1571 * Also, interrupt window exiting allows us to inject any
1572 * pending APIC vector that was preempted by the ExtINT
1573 * as soon as possible. This applies both for the software
1574 * emulated vlapic and the hardware assisted virtual APIC.
1575 */
1576 vmx_set_int_window_exiting(vmx, vcpu);
1577 }
1578
1579 VCPU_CTR1(vmx->vm, vcpu, "Injecting hwintr at vector %d", vector);
1580
1581 return;
1582
1583 cantinject:
1584 /*
1585 * Set the Interrupt Window Exiting execution control so we can inject
1586 * the interrupt as soon as blocking condition goes away.
1587 */
1588 vmx_set_int_window_exiting(vmx, vcpu);
1589 }
1590
1591 /*
1592 * If the Virtual NMIs execution control is '1' then the logical processor
1593 * tracks virtual-NMI blocking in the Guest Interruptibility-state field of
1594 * the VMCS. An IRET instruction in VMX non-root operation will remove any
1595 * virtual-NMI blocking.
1596 *
1597 * This unblocking occurs even if the IRET causes a fault. In this case the
1598 * hypervisor needs to restore virtual-NMI blocking before resuming the guest.
1599 */
1600 static void
1601 vmx_restore_nmi_blocking(struct vmx *vmx, int vcpuid)
1602 {
1603 uint32_t gi;
1604
1605 VCPU_CTR0(vmx->vm, vcpuid, "Restore Virtual-NMI blocking");
1606 gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY);
1607 gi |= VMCS_INTERRUPTIBILITY_NMI_BLOCKING;
1608 vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi);
1609 }
1610
1611 static void
1612 vmx_clear_nmi_blocking(struct vmx *vmx, int vcpuid)
1613 {
1614 uint32_t gi;
1615
1616 VCPU_CTR0(vmx->vm, vcpuid, "Clear Virtual-NMI blocking");
1617 gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY);
1618 gi &= ~VMCS_INTERRUPTIBILITY_NMI_BLOCKING;
1619 vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi);
1620 }
1621
1622 static void
1623 vmx_assert_nmi_blocking(struct vmx *vmx, int vcpuid)
1624 {
1625 uint32_t gi;
1626
1627 gi = vmcs_read(VMCS_GUEST_INTERRUPTIBILITY);
1628 KASSERT(gi & VMCS_INTERRUPTIBILITY_NMI_BLOCKING,
1629 ("NMI blocking is not in effect %#x", gi));
1630 }
1631
1632 static int
1633 vmx_emulate_xsetbv(struct vmx *vmx, int vcpu, struct vm_exit *vmexit)
1634 {
1635 struct vmxctx *vmxctx;
1636 uint64_t xcrval;
1637 const struct xsave_limits *limits;
1638
1639 vmxctx = &vmx->ctx[vcpu];
1640 limits = vmm_get_xsave_limits();
1641
1642 /*
1643 * Note that the processor raises a GP# fault on its own if
1644 * xsetbv is executed for CPL != 0, so we do not have to
1645 * emulate that fault here.
1646 */
1647
1648 /* Only xcr0 is supported. */
1649 if (vmxctx->guest_rcx != 0) {
1650 vm_inject_gp(vmx->vm, vcpu);
1651 return (HANDLED);
1652 }
1653
1654 /* We only handle xcr0 if both the host and guest have XSAVE enabled. */
1655 if (!limits->xsave_enabled || !(vmcs_read(VMCS_GUEST_CR4) & CR4_XSAVE)) {
1656 vm_inject_ud(vmx->vm, vcpu);
1657 return (HANDLED);
1658 }
1659
1660 xcrval = vmxctx->guest_rdx << 32 | (vmxctx->guest_rax & 0xffffffff);
1661 if ((xcrval & ~limits->xcr0_allowed) != 0) {
1662 vm_inject_gp(vmx->vm, vcpu);
1663 return (HANDLED);
1664 }
1665
1666 if (!(xcrval & XFEATURE_ENABLED_X87)) {
1667 vm_inject_gp(vmx->vm, vcpu);
1668 return (HANDLED);
1669 }
1670
1671 /* AVX (YMM_Hi128) requires SSE. */
1672 if (xcrval & XFEATURE_ENABLED_AVX &&
1673 (xcrval & XFEATURE_AVX) != XFEATURE_AVX) {
1674 vm_inject_gp(vmx->vm, vcpu);
1675 return (HANDLED);
1676 }
1677
1678 /*
1679 * AVX512 requires base AVX (YMM_Hi128) as well as OpMask,
1680 * ZMM_Hi256, and Hi16_ZMM.
1681 */
1682 if (xcrval & XFEATURE_AVX512 &&
1683 (xcrval & (XFEATURE_AVX512 | XFEATURE_AVX)) !=
1684 (XFEATURE_AVX512 | XFEATURE_AVX)) {
1685 vm_inject_gp(vmx->vm, vcpu);
1686 return (HANDLED);
1687 }
1688
1689 /*
1690 * Intel MPX requires both bound register state flags to be
1691 * set.
1692 */
1693 if (((xcrval & XFEATURE_ENABLED_BNDREGS) != 0) !=
1694 ((xcrval & XFEATURE_ENABLED_BNDCSR) != 0)) {
1695 vm_inject_gp(vmx->vm, vcpu);
1696 return (HANDLED);
1697 }
1698
1699 /*
1700 * This runs "inside" vmrun() with the guest's FPU state, so
1701 * modifying xcr0 directly modifies the guest's xcr0, not the
1702 * host's.
1703 */
1704 load_xcr(0, xcrval);
1705 return (HANDLED);
1706 }
1707
1708 static uint64_t
1709 vmx_get_guest_reg(struct vmx *vmx, int vcpu, int ident)
1710 {
1711 const struct vmxctx *vmxctx;
1712
1713 vmxctx = &vmx->ctx[vcpu];
1714
1715 switch (ident) {
1716 case 0:
1717 return (vmxctx->guest_rax);
1718 case 1:
1719 return (vmxctx->guest_rcx);
1720 case 2:
1721 return (vmxctx->guest_rdx);
1722 case 3:
1723 return (vmxctx->guest_rbx);
1724 case 4:
1725 return (vmcs_read(VMCS_GUEST_RSP));
1726 case 5:
1727 return (vmxctx->guest_rbp);
1728 case 6:
1729 return (vmxctx->guest_rsi);
1730 case 7:
1731 return (vmxctx->guest_rdi);
1732 case 8:
1733 return (vmxctx->guest_r8);
1734 case 9:
1735 return (vmxctx->guest_r9);
1736 case 10:
1737 return (vmxctx->guest_r10);
1738 case 11:
1739 return (vmxctx->guest_r11);
1740 case 12:
1741 return (vmxctx->guest_r12);
1742 case 13:
1743 return (vmxctx->guest_r13);
1744 case 14:
1745 return (vmxctx->guest_r14);
1746 case 15:
1747 return (vmxctx->guest_r15);
1748 default:
1749 panic("invalid vmx register %d", ident);
1750 }
1751 }
1752
1753 static void
1754 vmx_set_guest_reg(struct vmx *vmx, int vcpu, int ident, uint64_t regval)
1755 {
1756 struct vmxctx *vmxctx;
1757
1758 vmxctx = &vmx->ctx[vcpu];
1759
1760 switch (ident) {
1761 case 0:
1762 vmxctx->guest_rax = regval;
1763 break;
1764 case 1:
1765 vmxctx->guest_rcx = regval;
1766 break;
1767 case 2:
1768 vmxctx->guest_rdx = regval;
1769 break;
1770 case 3:
1771 vmxctx->guest_rbx = regval;
1772 break;
1773 case 4:
1774 vmcs_write(VMCS_GUEST_RSP, regval);
1775 break;
1776 case 5:
1777 vmxctx->guest_rbp = regval;
1778 break;
1779 case 6:
1780 vmxctx->guest_rsi = regval;
1781 break;
1782 case 7:
1783 vmxctx->guest_rdi = regval;
1784 break;
1785 case 8:
1786 vmxctx->guest_r8 = regval;
1787 break;
1788 case 9:
1789 vmxctx->guest_r9 = regval;
1790 break;
1791 case 10:
1792 vmxctx->guest_r10 = regval;
1793 break;
1794 case 11:
1795 vmxctx->guest_r11 = regval;
1796 break;
1797 case 12:
1798 vmxctx->guest_r12 = regval;
1799 break;
1800 case 13:
1801 vmxctx->guest_r13 = regval;
1802 break;
1803 case 14:
1804 vmxctx->guest_r14 = regval;
1805 break;
1806 case 15:
1807 vmxctx->guest_r15 = regval;
1808 break;
1809 default:
1810 panic("invalid vmx register %d", ident);
1811 }
1812 }
1813
1814 static int
1815 vmx_emulate_cr0_access(struct vmx *vmx, int vcpu, uint64_t exitqual)
1816 {
1817 uint64_t crval, regval;
1818
1819 /* We only handle mov to %cr0 at this time */
1820 if ((exitqual & 0xf0) != 0x00)
1821 return (UNHANDLED);
1822
1823 regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf);
1824
1825 vmcs_write(VMCS_CR0_SHADOW, regval);
1826
1827 crval = regval | cr0_ones_mask;
1828 crval &= ~cr0_zeros_mask;
1829 vmcs_write(VMCS_GUEST_CR0, crval);
1830
1831 if (regval & CR0_PG) {
1832 uint64_t efer, entry_ctls;
1833
1834 /*
1835 * If CR0.PG is 1 and EFER.LME is 1 then EFER.LMA and
1836 * the "IA-32e mode guest" bit in VM-entry control must be
1837 * equal.
1838 */
1839 efer = vmcs_read(VMCS_GUEST_IA32_EFER);
1840 if (efer & EFER_LME) {
1841 efer |= EFER_LMA;
1842 vmcs_write(VMCS_GUEST_IA32_EFER, efer);
1843 entry_ctls = vmcs_read(VMCS_ENTRY_CTLS);
1844 entry_ctls |= VM_ENTRY_GUEST_LMA;
1845 vmcs_write(VMCS_ENTRY_CTLS, entry_ctls);
1846 }
1847 }
1848
1849 return (HANDLED);
1850 }
1851
1852 static int
1853 vmx_emulate_cr4_access(struct vmx *vmx, int vcpu, uint64_t exitqual)
1854 {
1855 uint64_t crval, regval;
1856
1857 /* We only handle mov to %cr4 at this time */
1858 if ((exitqual & 0xf0) != 0x00)
1859 return (UNHANDLED);
1860
1861 regval = vmx_get_guest_reg(vmx, vcpu, (exitqual >> 8) & 0xf);
1862
1863 vmcs_write(VMCS_CR4_SHADOW, regval);
1864
1865 crval = regval | cr4_ones_mask;
1866 crval &= ~cr4_zeros_mask;
1867 vmcs_write(VMCS_GUEST_CR4, crval);
1868
1869 return (HANDLED);
1870 }
1871
1872 static int
1873 vmx_emulate_cr8_access(struct vmx *vmx, int vcpu, uint64_t exitqual)
1874 {
1875 struct vlapic *vlapic;
1876 uint64_t cr8;
1877 int regnum;
1878
1879 /* We only handle mov %cr8 to/from a register at this time. */
1880 if ((exitqual & 0xe0) != 0x00) {
1881 return (UNHANDLED);
1882 }
1883
1884 vlapic = vm_lapic(vmx->vm, vcpu);
1885 regnum = (exitqual >> 8) & 0xf;
1886 if (exitqual & 0x10) {
1887 cr8 = vlapic_get_cr8(vlapic);
1888 vmx_set_guest_reg(vmx, vcpu, regnum, cr8);
1889 } else {
1890 cr8 = vmx_get_guest_reg(vmx, vcpu, regnum);
1891 vlapic_set_cr8(vlapic, cr8);
1892 }
1893
1894 return (HANDLED);
1895 }
1896
1897 /*
1898 * From section "Guest Register State" in the Intel SDM: CPL = SS.DPL
1899 */
1900 static int
1901 vmx_cpl(void)
1902 {
1903 uint32_t ssar;
1904
1905 ssar = vmcs_read(VMCS_GUEST_SS_ACCESS_RIGHTS);
1906 return ((ssar >> 5) & 0x3);
1907 }
1908
1909 static enum vm_cpu_mode
1910 vmx_cpu_mode(void)
1911 {
1912 uint32_t csar;
1913
1914 if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LMA) {
1915 csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS);
1916 if (csar & 0x2000)
1917 return (CPU_MODE_64BIT); /* CS.L = 1 */
1918 else
1919 return (CPU_MODE_COMPATIBILITY);
1920 } else if (vmcs_read(VMCS_GUEST_CR0) & CR0_PE) {
1921 return (CPU_MODE_PROTECTED);
1922 } else {
1923 return (CPU_MODE_REAL);
1924 }
1925 }
1926
1927 static enum vm_paging_mode
1928 vmx_paging_mode(void)
1929 {
1930
1931 if (!(vmcs_read(VMCS_GUEST_CR0) & CR0_PG))
1932 return (PAGING_MODE_FLAT);
1933 if (!(vmcs_read(VMCS_GUEST_CR4) & CR4_PAE))
1934 return (PAGING_MODE_32);
1935 if (vmcs_read(VMCS_GUEST_IA32_EFER) & EFER_LME)
1936 return (PAGING_MODE_64);
1937 else
1938 return (PAGING_MODE_PAE);
1939 }
1940
1941 static uint64_t
1942 inout_str_index(struct vmx *vmx, int vcpuid, int in)
1943 {
1944 uint64_t val;
1945 int error;
1946 enum vm_reg_name reg;
1947
1948 reg = in ? VM_REG_GUEST_RDI : VM_REG_GUEST_RSI;
1949 error = vmx_getreg(vmx, vcpuid, reg, &val);
1950 KASSERT(error == 0, ("%s: vmx_getreg error %d", __func__, error));
1951 return (val);
1952 }
1953
1954 static uint64_t
1955 inout_str_count(struct vmx *vmx, int vcpuid, int rep)
1956 {
1957 uint64_t val;
1958 int error;
1959
1960 if (rep) {
1961 error = vmx_getreg(vmx, vcpuid, VM_REG_GUEST_RCX, &val);
1962 KASSERT(!error, ("%s: vmx_getreg error %d", __func__, error));
1963 } else {
1964 val = 1;
1965 }
1966 return (val);
1967 }
1968
1969 static int
1970 inout_str_addrsize(uint32_t inst_info)
1971 {
1972 uint32_t size;
1973
1974 size = (inst_info >> 7) & 0x7;
1975 switch (size) {
1976 case 0:
1977 return (2); /* 16 bit */
1978 case 1:
1979 return (4); /* 32 bit */
1980 case 2:
1981 return (8); /* 64 bit */
1982 default:
1983 panic("%s: invalid size encoding %d", __func__, size);
1984 }
1985 }
1986
1987 static void
1988 inout_str_seginfo(struct vmx *vmx, int vcpuid, uint32_t inst_info, int in,
1989 struct vm_inout_str *vis)
1990 {
1991 int error, s;
1992
1993 if (in) {
1994 vis->seg_name = VM_REG_GUEST_ES;
1995 } else {
1996 s = (inst_info >> 15) & 0x7;
1997 vis->seg_name = vm_segment_name(s);
1998 }
1999
2000 error = vmx_getdesc(vmx, vcpuid, vis->seg_name, &vis->seg_desc);
2001 KASSERT(error == 0, ("%s: vmx_getdesc error %d", __func__, error));
2002 }
2003
2004 static void
2005 vmx_paging_info(struct vm_guest_paging *paging)
2006 {
2007 paging->cr3 = vmcs_guest_cr3();
2008 paging->cpl = vmx_cpl();
2009 paging->cpu_mode = vmx_cpu_mode();
2010 paging->paging_mode = vmx_paging_mode();
2011 }
2012
2013 static void
2014 vmexit_inst_emul(struct vm_exit *vmexit, uint64_t gpa, uint64_t gla)
2015 {
2016 struct vm_guest_paging *paging;
2017 uint32_t csar;
2018
2019 paging = &vmexit->u.inst_emul.paging;
2020
2021 vmexit->exitcode = VM_EXITCODE_INST_EMUL;
2022 vmexit->inst_length = 0;
2023 vmexit->u.inst_emul.gpa = gpa;
2024 vmexit->u.inst_emul.gla = gla;
2025 vmx_paging_info(paging);
2026 switch (paging->cpu_mode) {
2027 case CPU_MODE_REAL:
2028 vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE);
2029 vmexit->u.inst_emul.cs_d = 0;
2030 break;
2031 case CPU_MODE_PROTECTED:
2032 case CPU_MODE_COMPATIBILITY:
2033 vmexit->u.inst_emul.cs_base = vmcs_read(VMCS_GUEST_CS_BASE);
2034 csar = vmcs_read(VMCS_GUEST_CS_ACCESS_RIGHTS);
2035 vmexit->u.inst_emul.cs_d = SEG_DESC_DEF32(csar);
2036 break;
2037 default:
2038 vmexit->u.inst_emul.cs_base = 0;
2039 vmexit->u.inst_emul.cs_d = 0;
2040 break;
2041 }
2042 vie_init(&vmexit->u.inst_emul.vie, NULL, 0);
2043 }
2044
2045 static int
2046 ept_fault_type(uint64_t ept_qual)
2047 {
2048 int fault_type;
2049
2050 if (ept_qual & EPT_VIOLATION_DATA_WRITE)
2051 fault_type = VM_PROT_WRITE;
2052 else if (ept_qual & EPT_VIOLATION_INST_FETCH)
2053 fault_type = VM_PROT_EXECUTE;
2054 else
2055 fault_type= VM_PROT_READ;
2056
2057 return (fault_type);
2058 }
2059
2060 static bool
2061 ept_emulation_fault(uint64_t ept_qual)
2062 {
2063 int read, write;
2064
2065 /* EPT fault on an instruction fetch doesn't make sense here */
2066 if (ept_qual & EPT_VIOLATION_INST_FETCH)
2067 return (false);
2068
2069 /* EPT fault must be a read fault or a write fault */
2070 read = ept_qual & EPT_VIOLATION_DATA_READ ? 1 : 0;
2071 write = ept_qual & EPT_VIOLATION_DATA_WRITE ? 1 : 0;
2072 if ((read | write) == 0)
2073 return (false);
2074
2075 /*
2076 * The EPT violation must have been caused by accessing a
2077 * guest-physical address that is a translation of a guest-linear
2078 * address.
2079 */
2080 if ((ept_qual & EPT_VIOLATION_GLA_VALID) == 0 ||
2081 (ept_qual & EPT_VIOLATION_XLAT_VALID) == 0) {
2082 return (false);
2083 }
2084
2085 return (true);
2086 }
2087
2088 static __inline int
2089 apic_access_virtualization(struct vmx *vmx, int vcpuid)
2090 {
2091 uint32_t proc_ctls2;
2092
2093 proc_ctls2 = vmx->cap[vcpuid].proc_ctls2;
2094 return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) ? 1 : 0);
2095 }
2096
2097 static __inline int
2098 x2apic_virtualization(struct vmx *vmx, int vcpuid)
2099 {
2100 uint32_t proc_ctls2;
2101
2102 proc_ctls2 = vmx->cap[vcpuid].proc_ctls2;
2103 return ((proc_ctls2 & PROCBASED2_VIRTUALIZE_X2APIC_MODE) ? 1 : 0);
2104 }
2105
2106 static int
2107 vmx_handle_apic_write(struct vmx *vmx, int vcpuid, struct vlapic *vlapic,
2108 uint64_t qual)
2109 {
2110 int error, handled, offset;
2111 uint32_t *apic_regs, vector;
2112 bool retu;
2113
2114 handled = HANDLED;
2115 offset = APIC_WRITE_OFFSET(qual);
2116
2117 if (!apic_access_virtualization(vmx, vcpuid)) {
2118 /*
2119 * In general there should not be any APIC write VM-exits
2120 * unless APIC-access virtualization is enabled.
2121 *
2122 * However self-IPI virtualization can legitimately trigger
2123 * an APIC-write VM-exit so treat it specially.
2124 */
2125 if (x2apic_virtualization(vmx, vcpuid) &&
2126 offset == APIC_OFFSET_SELF_IPI) {
2127 apic_regs = (uint32_t *)(vlapic->apic_page);
2128 vector = apic_regs[APIC_OFFSET_SELF_IPI / 4];
2129 vlapic_self_ipi_handler(vlapic, vector);
2130 return (HANDLED);
2131 } else
2132 return (UNHANDLED);
2133 }
2134
2135 switch (offset) {
2136 case APIC_OFFSET_ID:
2137 vlapic_id_write_handler(vlapic);
2138 break;
2139 case APIC_OFFSET_LDR:
2140 vlapic_ldr_write_handler(vlapic);
2141 break;
2142 case APIC_OFFSET_DFR:
2143 vlapic_dfr_write_handler(vlapic);
2144 break;
2145 case APIC_OFFSET_SVR:
2146 vlapic_svr_write_handler(vlapic);
2147 break;
2148 case APIC_OFFSET_ESR:
2149 vlapic_esr_write_handler(vlapic);
2150 break;
2151 case APIC_OFFSET_ICR_LOW:
2152 retu = false;
2153 error = vlapic_icrlo_write_handler(vlapic, &retu);
2154 if (error != 0 || retu)
2155 handled = UNHANDLED;
2156 break;
2157 case APIC_OFFSET_CMCI_LVT:
2158 case APIC_OFFSET_TIMER_LVT ... APIC_OFFSET_ERROR_LVT:
2159 vlapic_lvt_write_handler(vlapic, offset);
2160 break;
2161 case APIC_OFFSET_TIMER_ICR:
2162 vlapic_icrtmr_write_handler(vlapic);
2163 break;
2164 case APIC_OFFSET_TIMER_DCR:
2165 vlapic_dcr_write_handler(vlapic);
2166 break;
2167 default:
2168 handled = UNHANDLED;
2169 break;
2170 }
2171 return (handled);
2172 }
2173
2174 static bool
2175 apic_access_fault(struct vmx *vmx, int vcpuid, uint64_t gpa)
2176 {
2177
2178 if (apic_access_virtualization(vmx, vcpuid) &&
2179 (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE))
2180 return (true);
2181 else
2182 return (false);
2183 }
2184
2185 static int
2186 vmx_handle_apic_access(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit)
2187 {
2188 uint64_t qual;
2189 int access_type, offset, allowed;
2190
2191 if (!apic_access_virtualization(vmx, vcpuid))
2192 return (UNHANDLED);
2193
2194 qual = vmexit->u.vmx.exit_qualification;
2195 access_type = APIC_ACCESS_TYPE(qual);
2196 offset = APIC_ACCESS_OFFSET(qual);
2197
2198 allowed = 0;
2199 if (access_type == 0) {
2200 /*
2201 * Read data access to the following registers is expected.
2202 */
2203 switch (offset) {
2204 case APIC_OFFSET_APR:
2205 case APIC_OFFSET_PPR:
2206 case APIC_OFFSET_RRR:
2207 case APIC_OFFSET_CMCI_LVT:
2208 case APIC_OFFSET_TIMER_CCR:
2209 allowed = 1;
2210 break;
2211 default:
2212 break;
2213 }
2214 } else if (access_type == 1) {
2215 /*
2216 * Write data access to the following registers is expected.
2217 */
2218 switch (offset) {
2219 case APIC_OFFSET_VER:
2220 case APIC_OFFSET_APR:
2221 case APIC_OFFSET_PPR:
2222 case APIC_OFFSET_RRR:
2223 case APIC_OFFSET_ISR0 ... APIC_OFFSET_ISR7:
2224 case APIC_OFFSET_TMR0 ... APIC_OFFSET_TMR7:
2225 case APIC_OFFSET_IRR0 ... APIC_OFFSET_IRR7:
2226 case APIC_OFFSET_CMCI_LVT:
2227 case APIC_OFFSET_TIMER_CCR:
2228 allowed = 1;
2229 break;
2230 default:
2231 break;
2232 }
2233 }
2234
2235 if (allowed) {
2236 vmexit_inst_emul(vmexit, DEFAULT_APIC_BASE + offset,
2237 VIE_INVALID_GLA);
2238 }
2239
2240 /*
2241 * Regardless of whether the APIC-access is allowed this handler
2242 * always returns UNHANDLED:
2243 * - if the access is allowed then it is handled by emulating the
2244 * instruction that caused the VM-exit (outside the critical section)
2245 * - if the access is not allowed then it will be converted to an
2246 * exitcode of VM_EXITCODE_VMX and will be dealt with in userland.
2247 */
2248 return (UNHANDLED);
2249 }
2250
2251 static enum task_switch_reason
2252 vmx_task_switch_reason(uint64_t qual)
2253 {
2254 int reason;
2255
2256 reason = (qual >> 30) & 0x3;
2257 switch (reason) {
2258 case 0:
2259 return (TSR_CALL);
2260 case 1:
2261 return (TSR_IRET);
2262 case 2:
2263 return (TSR_JMP);
2264 case 3:
2265 return (TSR_IDT_GATE);
2266 default:
2267 panic("%s: invalid reason %d", __func__, reason);
2268 }
2269 }
2270
2271 static int
2272 emulate_wrmsr(struct vmx *vmx, int vcpuid, u_int num, uint64_t val, bool *retu)
2273 {
2274 int error;
2275
2276 if (lapic_msr(num))
2277 error = lapic_wrmsr(vmx->vm, vcpuid, num, val, retu);
2278 else
2279 error = vmx_wrmsr(vmx, vcpuid, num, val, retu);
2280
2281 return (error);
2282 }
2283
2284 static int
2285 emulate_rdmsr(struct vmx *vmx, int vcpuid, u_int num, bool *retu)
2286 {
2287 struct vmxctx *vmxctx;
2288 uint64_t result;
2289 uint32_t eax, edx;
2290 int error;
2291
2292 if (lapic_msr(num))
2293 error = lapic_rdmsr(vmx->vm, vcpuid, num, &result, retu);
2294 else
2295 error = vmx_rdmsr(vmx, vcpuid, num, &result, retu);
2296
2297 if (error == 0) {
2298 eax = result;
2299 vmxctx = &vmx->ctx[vcpuid];
2300 error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RAX, eax);
2301 KASSERT(error == 0, ("vmxctx_setreg(rax) error %d", error));
2302
2303 edx = result >> 32;
2304 error = vmxctx_setreg(vmxctx, VM_REG_GUEST_RDX, edx);
2305 KASSERT(error == 0, ("vmxctx_setreg(rdx) error %d", error));
2306 }
2307
2308 return (error);
2309 }
2310
2311 static int
2312 vmx_exit_process(struct vmx *vmx, int vcpu, struct vm_exit *vmexit)
2313 {
2314 int error, errcode, errcode_valid, handled, in;
2315 struct vmxctx *vmxctx;
2316 struct vlapic *vlapic;
2317 struct vm_inout_str *vis;
2318 struct vm_task_switch *ts;
2319 uint32_t eax, ecx, edx, idtvec_info, idtvec_err, intr_info, inst_info;
2320 uint32_t intr_type, intr_vec, reason;
2321 uint64_t exitintinfo, qual, gpa;
2322 bool retu;
2323
2324 CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_VIRTUAL_NMI) != 0);
2325 CTASSERT((PINBASED_CTLS_ONE_SETTING & PINBASED_NMI_EXITING) != 0);
2326
2327 handled = UNHANDLED;
2328 vmxctx = &vmx->ctx[vcpu];
2329
2330 qual = vmexit->u.vmx.exit_qualification;
2331 reason = vmexit->u.vmx.exit_reason;
2332 vmexit->exitcode = VM_EXITCODE_BOGUS;
2333
2334 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_COUNT, 1);
2335 SDT_PROBE3(vmm, vmx, exit, entry, vmx, vcpu, vmexit);
2336
2337 /*
2338 * VM-entry failures during or after loading guest state.
2339 *
2340 * These VM-exits are uncommon but must be handled specially
2341 * as most VM-exit fields are not populated as usual.
2342 */
2343 if (__predict_false(reason == EXIT_REASON_MCE_DURING_ENTRY)) {
2344 VCPU_CTR0(vmx->vm, vcpu, "Handling MCE during VM-entry");
2345 __asm __volatile("int $18");
2346 return (1);
2347 }
2348
2349 /*
2350 * VM exits that can be triggered during event delivery need to
2351 * be handled specially by re-injecting the event if the IDT
2352 * vectoring information field's valid bit is set.
2353 *
2354 * See "Information for VM Exits During Event Delivery" in Intel SDM
2355 * for details.
2356 */
2357 idtvec_info = vmcs_idt_vectoring_info();
2358 if (idtvec_info & VMCS_IDT_VEC_VALID) {
2359 idtvec_info &= ~(1 << 12); /* clear undefined bit */
2360 exitintinfo = idtvec_info;
2361 if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) {
2362 idtvec_err = vmcs_idt_vectoring_err();
2363 exitintinfo |= (uint64_t)idtvec_err << 32;
2364 }
2365 error = vm_exit_intinfo(vmx->vm, vcpu, exitintinfo);
2366 KASSERT(error == 0, ("%s: vm_set_intinfo error %d",
2367 __func__, error));
2368
2369 /*
2370 * If 'virtual NMIs' are being used and the VM-exit
2371 * happened while injecting an NMI during the previous
2372 * VM-entry, then clear "blocking by NMI" in the
2373 * Guest Interruptibility-State so the NMI can be
2374 * reinjected on the subsequent VM-entry.
2375 *
2376 * However, if the NMI was being delivered through a task
2377 * gate, then the new task must start execution with NMIs
2378 * blocked so don't clear NMI blocking in this case.
2379 */
2380 intr_type = idtvec_info & VMCS_INTR_T_MASK;
2381 if (intr_type == VMCS_INTR_T_NMI) {
2382 if (reason != EXIT_REASON_TASK_SWITCH)
2383 vmx_clear_nmi_blocking(vmx, vcpu);
2384 else
2385 vmx_assert_nmi_blocking(vmx, vcpu);
2386 }
2387
2388 /*
2389 * Update VM-entry instruction length if the event being
2390 * delivered was a software interrupt or software exception.
2391 */
2392 if (intr_type == VMCS_INTR_T_SWINTR ||
2393 intr_type == VMCS_INTR_T_PRIV_SWEXCEPTION ||
2394 intr_type == VMCS_INTR_T_SWEXCEPTION) {
2395 vmcs_write(VMCS_ENTRY_INST_LENGTH, vmexit->inst_length);
2396 }
2397 }
2398
2399 switch (reason) {
2400 case EXIT_REASON_TASK_SWITCH:
2401 ts = &vmexit->u.task_switch;
2402 ts->tsssel = qual & 0xffff;
2403 ts->reason = vmx_task_switch_reason(qual);
2404 ts->ext = 0;
2405 ts->errcode_valid = 0;
2406 vmx_paging_info(&ts->paging);
2407 /*
2408 * If the task switch was due to a CALL, JMP, IRET, software
2409 * interrupt (INT n) or software exception (INT3, INTO),
2410 * then the saved %rip references the instruction that caused
2411 * the task switch. The instruction length field in the VMCS
2412 * is valid in this case.
2413 *
2414 * In all other cases (e.g., NMI, hardware exception) the
2415 * saved %rip is one that would have been saved in the old TSS
2416 * had the task switch completed normally so the instruction
2417 * length field is not needed in this case and is explicitly
2418 * set to 0.
2419 */
2420 if (ts->reason == TSR_IDT_GATE) {
2421 KASSERT(idtvec_info & VMCS_IDT_VEC_VALID,
2422 ("invalid idtvec_info %#x for IDT task switch",
2423 idtvec_info));
2424 intr_type = idtvec_info & VMCS_INTR_T_MASK;
2425 if (intr_type != VMCS_INTR_T_SWINTR &&
2426 intr_type != VMCS_INTR_T_SWEXCEPTION &&
2427 intr_type != VMCS_INTR_T_PRIV_SWEXCEPTION) {
2428 /* Task switch triggered by external event */
2429 ts->ext = 1;
2430 vmexit->inst_length = 0;
2431 if (idtvec_info & VMCS_IDT_VEC_ERRCODE_VALID) {
2432 ts->errcode_valid = 1;
2433 ts->errcode = vmcs_idt_vectoring_err();
2434 }
2435 }
2436 }
2437 vmexit->exitcode = VM_EXITCODE_TASK_SWITCH;
2438 SDT_PROBE4(vmm, vmx, exit, taskswitch, vmx, vcpu, vmexit, ts);
2439 VCPU_CTR4(vmx->vm, vcpu, "task switch reason %d, tss 0x%04x, "
2440 "%s errcode 0x%016lx", ts->reason, ts->tsssel,
2441 ts->ext ? "external" : "internal",
2442 ((uint64_t)ts->errcode << 32) | ts->errcode_valid);
2443 break;
2444 case EXIT_REASON_CR_ACCESS:
2445 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CR_ACCESS, 1);
2446 SDT_PROBE4(vmm, vmx, exit, craccess, vmx, vcpu, vmexit, qual);
2447 switch (qual & 0xf) {
2448 case 0:
2449 handled = vmx_emulate_cr0_access(vmx, vcpu, qual);
2450 break;
2451 case 4:
2452 handled = vmx_emulate_cr4_access(vmx, vcpu, qual);
2453 break;
2454 case 8:
2455 handled = vmx_emulate_cr8_access(vmx, vcpu, qual);
2456 break;
2457 }
2458 break;
2459 case EXIT_REASON_RDMSR:
2460 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_RDMSR, 1);
2461 retu = false;
2462 ecx = vmxctx->guest_rcx;
2463 VCPU_CTR1(vmx->vm, vcpu, "rdmsr 0x%08x", ecx);
2464 SDT_PROBE4(vmm, vmx, exit, rdmsr, vmx, vcpu, vmexit, ecx);
2465 error = emulate_rdmsr(vmx, vcpu, ecx, &retu);
2466 if (error) {
2467 vmexit->exitcode = VM_EXITCODE_RDMSR;
2468 vmexit->u.msr.code = ecx;
2469 } else if (!retu) {
2470 handled = HANDLED;
2471 } else {
2472 /* Return to userspace with a valid exitcode */
2473 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
2474 ("emulate_rdmsr retu with bogus exitcode"));
2475 }
2476 break;
2477 case EXIT_REASON_WRMSR:
2478 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_WRMSR, 1);
2479 retu = false;
2480 eax = vmxctx->guest_rax;
2481 ecx = vmxctx->guest_rcx;
2482 edx = vmxctx->guest_rdx;
2483 VCPU_CTR2(vmx->vm, vcpu, "wrmsr 0x%08x value 0x%016lx",
2484 ecx, (uint64_t)edx << 32 | eax);
2485 SDT_PROBE5(vmm, vmx, exit, wrmsr, vmx, vmexit, vcpu, ecx,
2486 (uint64_t)edx << 32 | eax);
2487 error = emulate_wrmsr(vmx, vcpu, ecx,
2488 (uint64_t)edx << 32 | eax, &retu);
2489 if (error) {
2490 vmexit->exitcode = VM_EXITCODE_WRMSR;
2491 vmexit->u.msr.code = ecx;
2492 vmexit->u.msr.wval = (uint64_t)edx << 32 | eax;
2493 } else if (!retu) {
2494 handled = HANDLED;
2495 } else {
2496 /* Return to userspace with a valid exitcode */
2497 KASSERT(vmexit->exitcode != VM_EXITCODE_BOGUS,
2498 ("emulate_wrmsr retu with bogus exitcode"));
2499 }
2500 break;
2501 case EXIT_REASON_HLT:
2502 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_HLT, 1);
2503 SDT_PROBE3(vmm, vmx, exit, halt, vmx, vcpu, vmexit);
2504 vmexit->exitcode = VM_EXITCODE_HLT;
2505 vmexit->u.hlt.rflags = vmcs_read(VMCS_GUEST_RFLAGS);
2506 if (virtual_interrupt_delivery)
2507 vmexit->u.hlt.intr_status =
2508 vmcs_read(VMCS_GUEST_INTR_STATUS);
2509 else
2510 vmexit->u.hlt.intr_status = 0;
2511 break;
2512 case EXIT_REASON_MTF:
2513 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_MTRAP, 1);
2514 SDT_PROBE3(vmm, vmx, exit, mtrap, vmx, vcpu, vmexit);
2515 vmexit->exitcode = VM_EXITCODE_MTRAP;
2516 vmexit->inst_length = 0;
2517 break;
2518 case EXIT_REASON_PAUSE:
2519 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_PAUSE, 1);
2520 SDT_PROBE3(vmm, vmx, exit, pause, vmx, vcpu, vmexit);
2521 vmexit->exitcode = VM_EXITCODE_PAUSE;
2522 break;
2523 case EXIT_REASON_INTR_WINDOW:
2524 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INTR_WINDOW, 1);
2525 SDT_PROBE3(vmm, vmx, exit, intrwindow, vmx, vcpu, vmexit);
2526 vmx_clear_int_window_exiting(vmx, vcpu);
2527 return (1);
2528 case EXIT_REASON_EXT_INTR:
2529 /*
2530 * External interrupts serve only to cause VM exits and allow
2531 * the host interrupt handler to run.
2532 *
2533 * If this external interrupt triggers a virtual interrupt
2534 * to a VM, then that state will be recorded by the
2535 * host interrupt handler in the VM's softc. We will inject
2536 * this virtual interrupt during the subsequent VM enter.
2537 */
2538 intr_info = vmcs_read(VMCS_EXIT_INTR_INFO);
2539 SDT_PROBE4(vmm, vmx, exit, interrupt,
2540 vmx, vcpu, vmexit, intr_info);
2541
2542 /*
2543 * XXX: Ignore this exit if VMCS_INTR_VALID is not set.
2544 * This appears to be a bug in VMware Fusion?
2545 */
2546 if (!(intr_info & VMCS_INTR_VALID))
2547 return (1);
2548 KASSERT((intr_info & VMCS_INTR_VALID) != 0 &&
2549 (intr_info & VMCS_INTR_T_MASK) == VMCS_INTR_T_HWINTR,
2550 ("VM exit interruption info invalid: %#x", intr_info));
2551 vmx_trigger_hostintr(intr_info & 0xff);
2552
2553 /*
2554 * This is special. We want to treat this as an 'handled'
2555 * VM-exit but not increment the instruction pointer.
2556 */
2557 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_EXTINT, 1);
2558 return (1);
2559 case EXIT_REASON_NMI_WINDOW:
2560 SDT_PROBE3(vmm, vmx, exit, nmiwindow, vmx, vcpu, vmexit);
2561 /* Exit to allow the pending virtual NMI to be injected */
2562 if (vm_nmi_pending(vmx->vm, vcpu))
2563 vmx_inject_nmi(vmx, vcpu);
2564 vmx_clear_nmi_window_exiting(vmx, vcpu);
2565 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NMI_WINDOW, 1);
2566 return (1);
2567 case EXIT_REASON_INOUT:
2568 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INOUT, 1);
2569 vmexit->exitcode = VM_EXITCODE_INOUT;
2570 vmexit->u.inout.bytes = (qual & 0x7) + 1;
2571 vmexit->u.inout.in = in = (qual & 0x8) ? 1 : 0;
2572 vmexit->u.inout.string = (qual & 0x10) ? 1 : 0;
2573 vmexit->u.inout.rep = (qual & 0x20) ? 1 : 0;
2574 vmexit->u.inout.port = (uint16_t)(qual >> 16);
2575 vmexit->u.inout.eax = (uint32_t)(vmxctx->guest_rax);
2576 if (vmexit->u.inout.string) {
2577 inst_info = vmcs_read(VMCS_EXIT_INSTRUCTION_INFO);
2578 vmexit->exitcode = VM_EXITCODE_INOUT_STR;
2579 vis = &vmexit->u.inout_str;
2580 vmx_paging_info(&vis->paging);
2581 vis->rflags = vmcs_read(VMCS_GUEST_RFLAGS);
2582 vis->cr0 = vmcs_read(VMCS_GUEST_CR0);
2583 vis->index = inout_str_index(vmx, vcpu, in);
2584 vis->count = inout_str_count(vmx, vcpu, vis->inout.rep);
2585 vis->addrsize = inout_str_addrsize(inst_info);
2586 inout_str_seginfo(vmx, vcpu, inst_info, in, vis);
2587 }
2588 SDT_PROBE3(vmm, vmx, exit, inout, vmx, vcpu, vmexit);
2589 break;
2590 case EXIT_REASON_CPUID:
2591 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_CPUID, 1);
2592 SDT_PROBE3(vmm, vmx, exit, cpuid, vmx, vcpu, vmexit);
2593 handled = vmx_handle_cpuid(vmx->vm, vcpu, vmxctx);
2594 break;
2595 case EXIT_REASON_EXCEPTION:
2596 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_EXCEPTION, 1);
2597 intr_info = vmcs_read(VMCS_EXIT_INTR_INFO);
2598 KASSERT((intr_info & VMCS_INTR_VALID) != 0,
2599 ("VM exit interruption info invalid: %#x", intr_info));
2600
2601 intr_vec = intr_info & 0xff;
2602 intr_type = intr_info & VMCS_INTR_T_MASK;
2603
2604 /*
2605 * If Virtual NMIs control is 1 and the VM-exit is due to a
2606 * fault encountered during the execution of IRET then we must
2607 * restore the state of "virtual-NMI blocking" before resuming
2608 * the guest.
2609 *
2610 * See "Resuming Guest Software after Handling an Exception".
2611 * See "Information for VM Exits Due to Vectored Events".
2612 */
2613 if ((idtvec_info & VMCS_IDT_VEC_VALID) == 0 &&
2614 (intr_vec != IDT_DF) &&
2615 (intr_info & EXIT_QUAL_NMIUDTI) != 0)
2616 vmx_restore_nmi_blocking(vmx, vcpu);
2617
2618 /*
2619 * The NMI has already been handled in vmx_exit_handle_nmi().
2620 */
2621 if (intr_type == VMCS_INTR_T_NMI)
2622 return (1);
2623
2624 /*
2625 * Call the machine check handler by hand. Also don't reflect
2626 * the machine check back into the guest.
2627 */
2628 if (intr_vec == IDT_MC) {
2629 VCPU_CTR0(vmx->vm, vcpu, "Vectoring to MCE handler");
2630 __asm __volatile("int $18");
2631 return (1);
2632 }
2633
2634 /*
2635 * If the hypervisor has requested user exits for
2636 * debug exceptions, bounce them out to userland.
2637 */
2638 if (intr_type == VMCS_INTR_T_SWEXCEPTION && intr_vec == IDT_BP &&
2639 (vmx->cap[vcpu].set & (1 << VM_CAP_BPT_EXIT))) {
2640 vmexit->exitcode = VM_EXITCODE_BPT;
2641 vmexit->u.bpt.inst_length = vmexit->inst_length;
2642 vmexit->inst_length = 0;
2643 break;
2644 }
2645
2646 if (intr_vec == IDT_PF) {
2647 error = vmxctx_setreg(vmxctx, VM_REG_GUEST_CR2, qual);
2648 KASSERT(error == 0, ("%s: vmxctx_setreg(cr2) error %d",
2649 __func__, error));
2650 }
2651
2652 /*
2653 * Software exceptions exhibit trap-like behavior. This in
2654 * turn requires populating the VM-entry instruction length
2655 * so that the %rip in the trap frame is past the INT3/INTO
2656 * instruction.
2657 */
2658 if (intr_type == VMCS_INTR_T_SWEXCEPTION)
2659 vmcs_write(VMCS_ENTRY_INST_LENGTH, vmexit->inst_length);
2660
2661 /* Reflect all other exceptions back into the guest */
2662 errcode_valid = errcode = 0;
2663 if (intr_info & VMCS_INTR_DEL_ERRCODE) {
2664 errcode_valid = 1;
2665 errcode = vmcs_read(VMCS_EXIT_INTR_ERRCODE);
2666 }
2667 VCPU_CTR2(vmx->vm, vcpu, "Reflecting exception %d/%#x into "
2668 "the guest", intr_vec, errcode);
2669 SDT_PROBE5(vmm, vmx, exit, exception,
2670 vmx, vcpu, vmexit, intr_vec, errcode);
2671 error = vm_inject_exception(vmx->vm, vcpu, intr_vec,
2672 errcode_valid, errcode, 0);
2673 KASSERT(error == 0, ("%s: vm_inject_exception error %d",
2674 __func__, error));
2675 return (1);
2676
2677 case EXIT_REASON_EPT_FAULT:
2678 /*
2679 * If 'gpa' lies within the address space allocated to
2680 * memory then this must be a nested page fault otherwise
2681 * this must be an instruction that accesses MMIO space.
2682 */
2683 gpa = vmcs_gpa();
2684 if (vm_mem_allocated(vmx->vm, vcpu, gpa) ||
2685 apic_access_fault(vmx, vcpu, gpa)) {
2686 vmexit->exitcode = VM_EXITCODE_PAGING;
2687 vmexit->inst_length = 0;
2688 vmexit->u.paging.gpa = gpa;
2689 vmexit->u.paging.fault_type = ept_fault_type(qual);
2690 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_NESTED_FAULT, 1);
2691 SDT_PROBE5(vmm, vmx, exit, nestedfault,
2692 vmx, vcpu, vmexit, gpa, qual);
2693 } else if (ept_emulation_fault(qual)) {
2694 vmexit_inst_emul(vmexit, gpa, vmcs_gla());
2695 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_INST_EMUL, 1);
2696 SDT_PROBE4(vmm, vmx, exit, mmiofault,
2697 vmx, vcpu, vmexit, gpa);
2698 }
2699 /*
2700 * If Virtual NMIs control is 1 and the VM-exit is due to an
2701 * EPT fault during the execution of IRET then we must restore
2702 * the state of "virtual-NMI blocking" before resuming.
2703 *
2704 * See description of "NMI unblocking due to IRET" in
2705 * "Exit Qualification for EPT Violations".
2706 */
2707 if ((idtvec_info & VMCS_IDT_VEC_VALID) == 0 &&
2708 (qual & EXIT_QUAL_NMIUDTI) != 0)
2709 vmx_restore_nmi_blocking(vmx, vcpu);
2710 break;
2711 case EXIT_REASON_VIRTUALIZED_EOI:
2712 vmexit->exitcode = VM_EXITCODE_IOAPIC_EOI;
2713 vmexit->u.ioapic_eoi.vector = qual & 0xFF;
2714 SDT_PROBE3(vmm, vmx, exit, eoi, vmx, vcpu, vmexit);
2715 vmexit->inst_length = 0; /* trap-like */
2716 break;
2717 case EXIT_REASON_APIC_ACCESS:
2718 SDT_PROBE3(vmm, vmx, exit, apicaccess, vmx, vcpu, vmexit);
2719 handled = vmx_handle_apic_access(vmx, vcpu, vmexit);
2720 break;
2721 case EXIT_REASON_APIC_WRITE:
2722 /*
2723 * APIC-write VM exit is trap-like so the %rip is already
2724 * pointing to the next instruction.
2725 */
2726 vmexit->inst_length = 0;
2727 vlapic = vm_lapic(vmx->vm, vcpu);
2728 SDT_PROBE4(vmm, vmx, exit, apicwrite,
2729 vmx, vcpu, vmexit, vlapic);
2730 handled = vmx_handle_apic_write(vmx, vcpu, vlapic, qual);
2731 break;
2732 case EXIT_REASON_XSETBV:
2733 SDT_PROBE3(vmm, vmx, exit, xsetbv, vmx, vcpu, vmexit);
2734 handled = vmx_emulate_xsetbv(vmx, vcpu, vmexit);
2735 break;
2736 case EXIT_REASON_MONITOR:
2737 SDT_PROBE3(vmm, vmx, exit, monitor, vmx, vcpu, vmexit);
2738 vmexit->exitcode = VM_EXITCODE_MONITOR;
2739 break;
2740 case EXIT_REASON_MWAIT:
2741 SDT_PROBE3(vmm, vmx, exit, mwait, vmx, vcpu, vmexit);
2742 vmexit->exitcode = VM_EXITCODE_MWAIT;
2743 break;
2744 case EXIT_REASON_TPR:
2745 vlapic = vm_lapic(vmx->vm, vcpu);
2746 vlapic_sync_tpr(vlapic);
2747 vmexit->inst_length = 0;
2748 handled = HANDLED;
2749 break;
2750 case EXIT_REASON_VMCALL:
2751 case EXIT_REASON_VMCLEAR:
2752 case EXIT_REASON_VMLAUNCH:
2753 case EXIT_REASON_VMPTRLD:
2754 case EXIT_REASON_VMPTRST:
2755 case EXIT_REASON_VMREAD:
2756 case EXIT_REASON_VMRESUME:
2757 case EXIT_REASON_VMWRITE:
2758 case EXIT_REASON_VMXOFF:
2759 case EXIT_REASON_VMXON:
2760 SDT_PROBE3(vmm, vmx, exit, vminsn, vmx, vcpu, vmexit);
2761 vmexit->exitcode = VM_EXITCODE_VMINSN;
2762 break;
2763 default:
2764 SDT_PROBE4(vmm, vmx, exit, unknown,
2765 vmx, vcpu, vmexit, reason);
2766 vmm_stat_incr(vmx->vm, vcpu, VMEXIT_UNKNOWN, 1);
2767 break;
2768 }
2769
2770 if (handled) {
2771 /*
2772 * It is possible that control is returned to userland
2773 * even though we were able to handle the VM exit in the
2774 * kernel.
2775 *
2776 * In such a case we want to make sure that the userland
2777 * restarts guest execution at the instruction *after*
2778 * the one we just processed. Therefore we update the
2779 * guest rip in the VMCS and in 'vmexit'.
2780 */
2781 vmexit->rip += vmexit->inst_length;
2782 vmexit->inst_length = 0;
2783 vmcs_write(VMCS_GUEST_RIP, vmexit->rip);
2784 } else {
2785 if (vmexit->exitcode == VM_EXITCODE_BOGUS) {
2786 /*
2787 * If this VM exit was not claimed by anybody then
2788 * treat it as a generic VMX exit.
2789 */
2790 vmexit->exitcode = VM_EXITCODE_VMX;
2791 vmexit->u.vmx.status = VM_SUCCESS;
2792 vmexit->u.vmx.inst_type = 0;
2793 vmexit->u.vmx.inst_error = 0;
2794 } else {
2795 /*
2796 * The exitcode and collateral have been populated.
2797 * The VM exit will be processed further in userland.
2798 */
2799 }
2800 }
2801
2802 SDT_PROBE4(vmm, vmx, exit, return,
2803 vmx, vcpu, vmexit, handled);
2804 return (handled);
2805 }
2806
2807 static __inline void
2808 vmx_exit_inst_error(struct vmxctx *vmxctx, int rc, struct vm_exit *vmexit)
2809 {
2810
2811 KASSERT(vmxctx->inst_fail_status != VM_SUCCESS,
2812 ("vmx_exit_inst_error: invalid inst_fail_status %d",
2813 vmxctx->inst_fail_status));
2814
2815 vmexit->inst_length = 0;
2816 vmexit->exitcode = VM_EXITCODE_VMX;
2817 vmexit->u.vmx.status = vmxctx->inst_fail_status;
2818 vmexit->u.vmx.inst_error = vmcs_instruction_error();
2819 vmexit->u.vmx.exit_reason = ~0;
2820 vmexit->u.vmx.exit_qualification = ~0;
2821
2822 switch (rc) {
2823 case VMX_VMRESUME_ERROR:
2824 case VMX_VMLAUNCH_ERROR:
2825 vmexit->u.vmx.inst_type = rc;
2826 break;
2827 default:
2828 panic("vm_exit_inst_error: vmx_enter_guest returned %d", rc);
2829 }
2830 }
2831
2832 /*
2833 * If the NMI-exiting VM execution control is set to '1' then an NMI in
2834 * non-root operation causes a VM-exit. NMI blocking is in effect so it is
2835 * sufficient to simply vector to the NMI handler via a software interrupt.
2836 * However, this must be done before maskable interrupts are enabled
2837 * otherwise the "iret" issued by an interrupt handler will incorrectly
2838 * clear NMI blocking.
2839 */
2840 static __inline void
2841 vmx_exit_handle_nmi(struct vmx *vmx, int vcpuid, struct vm_exit *vmexit)
2842 {
2843 uint32_t intr_info;
2844
2845 KASSERT((read_rflags() & PSL_I) == 0, ("interrupts enabled"));
2846
2847 if (vmexit->u.vmx.exit_reason != EXIT_REASON_EXCEPTION)
2848 return;
2849
2850 intr_info = vmcs_read(VMCS_EXIT_INTR_INFO);
2851 KASSERT((intr_info & VMCS_INTR_VALID) != 0,
2852 ("VM exit interruption info invalid: %#x", intr_info));
2853
2854 if ((intr_info & VMCS_INTR_T_MASK) == VMCS_INTR_T_NMI) {
2855 KASSERT((intr_info & 0xff) == IDT_NMI, ("VM exit due "
2856 "to NMI has invalid vector: %#x", intr_info));
2857 VCPU_CTR0(vmx->vm, vcpuid, "Vectoring to NMI handler");
2858 __asm __volatile("int $2");
2859 }
2860 }
2861
2862 static __inline void
2863 vmx_dr_enter_guest(struct vmxctx *vmxctx)
2864 {
2865 register_t rflags;
2866
2867 /* Save host control debug registers. */
2868 vmxctx->host_dr7 = rdr7();
2869 vmxctx->host_debugctl = rdmsr(MSR_DEBUGCTLMSR);
2870
2871 /*
2872 * Disable debugging in DR7 and DEBUGCTL to avoid triggering
2873 * exceptions in the host based on the guest DRx values. The
2874 * guest DR7 and DEBUGCTL are saved/restored in the VMCS.
2875 */
2876 load_dr7(0);
2877 wrmsr(MSR_DEBUGCTLMSR, 0);
2878
2879 /*
2880 * Disable single stepping the kernel to avoid corrupting the
2881 * guest DR6. A debugger might still be able to corrupt the
2882 * guest DR6 by setting a breakpoint after this point and then
2883 * single stepping.
2884 */
2885 rflags = read_rflags();
2886 vmxctx->host_tf = rflags & PSL_T;
2887 write_rflags(rflags & ~PSL_T);
2888
2889 /* Save host debug registers. */
2890 vmxctx->host_dr0 = rdr0();
2891 vmxctx->host_dr1 = rdr1();
2892 vmxctx->host_dr2 = rdr2();
2893 vmxctx->host_dr3 = rdr3();
2894 vmxctx->host_dr6 = rdr6();
2895
2896 /* Restore guest debug registers. */
2897 load_dr0(vmxctx->guest_dr0);
2898 load_dr1(vmxctx->guest_dr1);
2899 load_dr2(vmxctx->guest_dr2);
2900 load_dr3(vmxctx->guest_dr3);
2901 load_dr6(vmxctx->guest_dr6);
2902 }
2903
2904 static __inline void
2905 vmx_dr_leave_guest(struct vmxctx *vmxctx)
2906 {
2907
2908 /* Save guest debug registers. */
2909 vmxctx->guest_dr0 = rdr0();
2910 vmxctx->guest_dr1 = rdr1();
2911 vmxctx->guest_dr2 = rdr2();
2912 vmxctx->guest_dr3 = rdr3();
2913 vmxctx->guest_dr6 = rdr6();
2914
2915 /*
2916 * Restore host debug registers. Restore DR7, DEBUGCTL, and
2917 * PSL_T last.
2918 */
2919 load_dr0(vmxctx->host_dr0);
2920 load_dr1(vmxctx->host_dr1);
2921 load_dr2(vmxctx->host_dr2);
2922 load_dr3(vmxctx->host_dr3);
2923 load_dr6(vmxctx->host_dr6);
2924 wrmsr(MSR_DEBUGCTLMSR, vmxctx->host_debugctl);
2925 load_dr7(vmxctx->host_dr7);
2926 write_rflags(read_rflags() | vmxctx->host_tf);
2927 }
2928
2929 static __inline void
2930 vmx_pmap_activate(struct vmx *vmx, pmap_t pmap)
2931 {
2932 long eptgen;
2933 int cpu;
2934
2935 cpu = curcpu;
2936
2937 CPU_SET_ATOMIC(cpu, &pmap->pm_active);
2938 eptgen = atomic_load_long(&pmap->pm_eptgen);
2939 if (eptgen != vmx->eptgen[cpu]) {
2940 vmx->eptgen[cpu] = eptgen;
2941 invept(INVEPT_TYPE_SINGLE_CONTEXT,
2942 (struct invept_desc){ .eptp = vmx->eptp, ._res = 0 });
2943 }
2944 }
2945
2946 static __inline void
2947 vmx_pmap_deactivate(struct vmx *vmx, pmap_t pmap)
2948 {
2949 CPU_CLR_ATOMIC(curcpu, &pmap->pm_active);
2950 }
2951
2952 static int
2953 vmx_run(void *arg, int vcpu, register_t rip, pmap_t pmap,
2954 struct vm_eventinfo *evinfo)
2955 {
2956 int rc, handled, launched;
2957 struct vmx *vmx;
2958 struct vm *vm;
2959 struct vmxctx *vmxctx;
2960 struct vmcs *vmcs;
2961 struct vm_exit *vmexit;
2962 struct vlapic *vlapic;
2963 uint32_t exit_reason;
2964 struct region_descriptor gdtr, idtr;
2965 uint16_t ldt_sel;
2966
2967 vmx = arg;
2968 vm = vmx->vm;
2969 vmcs = &vmx->vmcs[vcpu];
2970 vmxctx = &vmx->ctx[vcpu];
2971 vlapic = vm_lapic(vm, vcpu);
2972 vmexit = vm_exitinfo(vm, vcpu);
2973 launched = 0;
2974
2975 KASSERT(vmxctx->pmap == pmap,
2976 ("pmap %p different than ctx pmap %p", pmap, vmxctx->pmap));
2977
2978 vmx_msr_guest_enter(vmx, vcpu);
2979
2980 VMPTRLD(vmcs);
2981
2982 /*
2983 * XXX
2984 * We do this every time because we may setup the virtual machine
2985 * from a different process than the one that actually runs it.
2986 *
2987 * If the life of a virtual machine was spent entirely in the context
2988 * of a single process we could do this once in vmx_vminit().
2989 */
2990 vmcs_write(VMCS_HOST_CR3, rcr3());
2991
2992 vmcs_write(VMCS_GUEST_RIP, rip);
2993 vmx_set_pcpu_defaults(vmx, vcpu, pmap);
2994 do {
2995 KASSERT(vmcs_guest_rip() == rip, ("%s: vmcs guest rip mismatch "
2996 "%#lx/%#lx", __func__, vmcs_guest_rip(), rip));
2997
2998 handled = UNHANDLED;
2999 /*
3000 * Interrupts are disabled from this point on until the
3001 * guest starts executing. This is done for the following
3002 * reasons:
3003 *
3004 * If an AST is asserted on this thread after the check below,
3005 * then the IPI_AST notification will not be lost, because it
3006 * will cause a VM exit due to external interrupt as soon as
3007 * the guest state is loaded.
3008 *
3009 * A posted interrupt after 'vmx_inject_interrupts()' will
3010 * not be "lost" because it will be held pending in the host
3011 * APIC because interrupts are disabled. The pending interrupt
3012 * will be recognized as soon as the guest state is loaded.
3013 *
3014 * The same reasoning applies to the IPI generated by
3015 * pmap_invalidate_ept().
3016 */
3017 disable_intr();
3018 vmx_inject_interrupts(vmx, vcpu, vlapic, rip);
3019
3020 /*
3021 * Check for vcpu suspension after injecting events because
3022 * vmx_inject_interrupts() can suspend the vcpu due to a
3023 * triple fault.
3024 */
3025 if (vcpu_suspended(evinfo)) {
3026 enable_intr();
3027 vm_exit_suspended(vmx->vm, vcpu, rip);
3028 break;
3029 }
3030
3031 if (vcpu_rendezvous_pending(evinfo)) {
3032 enable_intr();
3033 vm_exit_rendezvous(vmx->vm, vcpu, rip);
3034 break;
3035 }
3036
3037 if (vcpu_reqidle(evinfo)) {
3038 enable_intr();
3039 vm_exit_reqidle(vmx->vm, vcpu, rip);
3040 break;
3041 }
3042
3043 if (vcpu_should_yield(vm, vcpu)) {
3044 enable_intr();
3045 vm_exit_astpending(vmx->vm, vcpu, rip);
3046 vmx_astpending_trace(vmx, vcpu, rip);
3047 handled = HANDLED;
3048 break;
3049 }
3050
3051 if (vcpu_debugged(vm, vcpu)) {
3052 enable_intr();
3053 vm_exit_debug(vmx->vm, vcpu, rip);
3054 break;
3055 }
3056
3057 /*
3058 * If TPR Shadowing is enabled, the TPR Threshold
3059 * must be updated right before entering the guest.
3060 */
3061 if (tpr_shadowing && !virtual_interrupt_delivery) {
3062 if ((vmx->cap[vcpu].proc_ctls & PROCBASED_USE_TPR_SHADOW) != 0) {
3063 vmcs_write(VMCS_TPR_THRESHOLD, vlapic_get_cr8(vlapic));
3064 }
3065 }
3066
3067 /*
3068 * VM exits restore the base address but not the
3069 * limits of GDTR and IDTR. The VMCS only stores the
3070 * base address, so VM exits set the limits to 0xffff.
3071 * Save and restore the full GDTR and IDTR to restore
3072 * the limits.
3073 *
3074 * The VMCS does not save the LDTR at all, and VM
3075 * exits clear LDTR as if a NULL selector were loaded.
3076 * The userspace hypervisor probably doesn't use a
3077 * LDT, but save and restore it to be safe.
3078 */
3079 sgdt(&gdtr);
3080 sidt(&idtr);
3081 ldt_sel = sldt();
3082
3083 /*
3084 * The TSC_AUX MSR must be saved/restored while interrupts
3085 * are disabled so that it is not possible for the guest
3086 * TSC_AUX MSR value to be overwritten by the resume
3087 * portion of the IPI_SUSPEND codepath. This is why the
3088 * transition of this MSR is handled separately from those
3089 * handled by vmx_msr_guest_{enter,exit}(), which are ok to
3090 * be transitioned with preemption disabled but interrupts
3091 * enabled.
3092 *
3093 * These vmx_msr_guest_{enter,exit}_tsc_aux() calls can be
3094 * anywhere in this loop so long as they happen with
3095 * interrupts disabled. This location is chosen for
3096 * simplicity.
3097 */
3098 vmx_msr_guest_enter_tsc_aux(vmx, vcpu);
3099
3100 vmx_dr_enter_guest(vmxctx);
3101
3102 /*
3103 * Mark the EPT as active on this host CPU and invalidate
3104 * EPTP-tagged TLB entries if required.
3105 */
3106 vmx_pmap_activate(vmx, pmap);
3107
3108 vmx_run_trace(vmx, vcpu);
3109 rc = vmx_enter_guest(vmxctx, vmx, launched);
3110
3111 vmx_pmap_deactivate(vmx, pmap);
3112 vmx_dr_leave_guest(vmxctx);
3113 vmx_msr_guest_exit_tsc_aux(vmx, vcpu);
3114
3115 bare_lgdt(&gdtr);
3116 lidt(&idtr);
3117 lldt(ldt_sel);
3118
3119 /* Collect some information for VM exit processing */
3120 vmexit->rip = rip = vmcs_guest_rip();
3121 vmexit->inst_length = vmexit_instruction_length();
3122 vmexit->u.vmx.exit_reason = exit_reason = vmcs_exit_reason();
3123 vmexit->u.vmx.exit_qualification = vmcs_exit_qualification();
3124
3125 /* Update 'nextrip' */
3126 vmx->state[vcpu].nextrip = rip;
3127
3128 if (rc == VMX_GUEST_VMEXIT) {
3129 vmx_exit_handle_nmi(vmx, vcpu, vmexit);
3130 enable_intr();
3131 handled = vmx_exit_process(vmx, vcpu, vmexit);
3132 } else {
3133 enable_intr();
3134 vmx_exit_inst_error(vmxctx, rc, vmexit);
3135 }
3136 launched = 1;
3137 vmx_exit_trace(vmx, vcpu, rip, exit_reason, handled);
3138 rip = vmexit->rip;
3139 } while (handled);
3140
3141 /*
3142 * If a VM exit has been handled then the exitcode must be BOGUS
3143 * If a VM exit is not handled then the exitcode must not be BOGUS
3144 */
3145 if ((handled && vmexit->exitcode != VM_EXITCODE_BOGUS) ||
3146 (!handled && vmexit->exitcode == VM_EXITCODE_BOGUS)) {
3147 panic("Mismatch between handled (%d) and exitcode (%d)",
3148 handled, vmexit->exitcode);
3149 }
3150
3151 if (!handled)
3152 vmm_stat_incr(vm, vcpu, VMEXIT_USERSPACE, 1);
3153
3154 VCPU_CTR1(vm, vcpu, "returning from vmx_run: exitcode %d",
3155 vmexit->exitcode);
3156
3157 VMCLEAR(vmcs);
3158 vmx_msr_guest_exit(vmx, vcpu);
3159
3160 return (0);
3161 }
3162
3163 static void
3164 vmx_vmcleanup(void *arg)
3165 {
3166 int i;
3167 struct vmx *vmx = arg;
3168 uint16_t maxcpus;
3169
3170 if (apic_access_virtualization(vmx, 0))
3171 vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE);
3172
3173 maxcpus = vm_get_maxcpus(vmx->vm);
3174 for (i = 0; i < maxcpus; i++)
3175 vpid_free(vmx->state[i].vpid);
3176
3177 free(vmx, M_VMX);
3178
3179 return;
3180 }
3181
3182 static register_t *
3183 vmxctx_regptr(struct vmxctx *vmxctx, int reg)
3184 {
3185
3186 switch (reg) {
3187 case VM_REG_GUEST_RAX:
3188 return (&vmxctx->guest_rax);
3189 case VM_REG_GUEST_RBX:
3190 return (&vmxctx->guest_rbx);
3191 case VM_REG_GUEST_RCX:
3192 return (&vmxctx->guest_rcx);
3193 case VM_REG_GUEST_RDX:
3194 return (&vmxctx->guest_rdx);
3195 case VM_REG_GUEST_RSI:
3196 return (&vmxctx->guest_rsi);
3197 case VM_REG_GUEST_RDI:
3198 return (&vmxctx->guest_rdi);
3199 case VM_REG_GUEST_RBP:
3200 return (&vmxctx->guest_rbp);
3201 case VM_REG_GUEST_R8:
3202 return (&vmxctx->guest_r8);
3203 case VM_REG_GUEST_R9:
3204 return (&vmxctx->guest_r9);
3205 case VM_REG_GUEST_R10:
3206 return (&vmxctx->guest_r10);
3207 case VM_REG_GUEST_R11:
3208 return (&vmxctx->guest_r11);
3209 case VM_REG_GUEST_R12:
3210 return (&vmxctx->guest_r12);
3211 case VM_REG_GUEST_R13:
3212 return (&vmxctx->guest_r13);
3213 case VM_REG_GUEST_R14:
3214 return (&vmxctx->guest_r14);
3215 case VM_REG_GUEST_R15:
3216 return (&vmxctx->guest_r15);
3217 case VM_REG_GUEST_CR2:
3218 return (&vmxctx->guest_cr2);
3219 case VM_REG_GUEST_DR0:
3220 return (&vmxctx->guest_dr0);
3221 case VM_REG_GUEST_DR1:
3222 return (&vmxctx->guest_dr1);
3223 case VM_REG_GUEST_DR2:
3224 return (&vmxctx->guest_dr2);
3225 case VM_REG_GUEST_DR3:
3226 return (&vmxctx->guest_dr3);
3227 case VM_REG_GUEST_DR6:
3228 return (&vmxctx->guest_dr6);
3229 default:
3230 break;
3231 }
3232 return (NULL);
3233 }
3234
3235 static int
3236 vmxctx_getreg(struct vmxctx *vmxctx, int reg, uint64_t *retval)
3237 {
3238 register_t *regp;
3239
3240 if ((regp = vmxctx_regptr(vmxctx, reg)) != NULL) {
3241 *retval = *regp;
3242 return (0);
3243 } else
3244 return (EINVAL);
3245 }
3246
3247 static int
3248 vmxctx_setreg(struct vmxctx *vmxctx, int reg, uint64_t val)
3249 {
3250 register_t *regp;
3251
3252 if ((regp = vmxctx_regptr(vmxctx, reg)) != NULL) {
3253 *regp = val;
3254 return (0);
3255 } else
3256 return (EINVAL);
3257 }
3258
3259 static int
3260 vmx_get_intr_shadow(struct vmx *vmx, int vcpu, int running, uint64_t *retval)
3261 {
3262 uint64_t gi;
3263 int error;
3264
3265 error = vmcs_getreg(&vmx->vmcs[vcpu], running,
3266 VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY), &gi);
3267 *retval = (gi & HWINTR_BLOCKING) ? 1 : 0;
3268 return (error);
3269 }
3270
3271 static int
3272 vmx_modify_intr_shadow(struct vmx *vmx, int vcpu, int running, uint64_t val)
3273 {
3274 struct vmcs *vmcs;
3275 uint64_t gi;
3276 int error, ident;
3277
3278 /*
3279 * Forcing the vcpu into an interrupt shadow is not supported.
3280 */
3281 if (val) {
3282 error = EINVAL;
3283 goto done;
3284 }
3285
3286 vmcs = &vmx->vmcs[vcpu];
3287 ident = VMCS_IDENT(VMCS_GUEST_INTERRUPTIBILITY);
3288 error = vmcs_getreg(vmcs, running, ident, &gi);
3289 if (error == 0) {
3290 gi &= ~HWINTR_BLOCKING;
3291 error = vmcs_setreg(vmcs, running, ident, gi);
3292 }
3293 done:
3294 VCPU_CTR2(vmx->vm, vcpu, "Setting intr_shadow to %#lx %s", val,
3295 error ? "failed" : "succeeded");
3296 return (error);
3297 }
3298
3299 static int
3300 vmx_shadow_reg(int reg)
3301 {
3302 int shreg;
3303
3304 shreg = -1;
3305
3306 switch (reg) {
3307 case VM_REG_GUEST_CR0:
3308 shreg = VMCS_CR0_SHADOW;
3309 break;
3310 case VM_REG_GUEST_CR4:
3311 shreg = VMCS_CR4_SHADOW;
3312 break;
3313 default:
3314 break;
3315 }
3316
3317 return (shreg);
3318 }
3319
3320 static int
3321 vmx_getreg(void *arg, int vcpu, int reg, uint64_t *retval)
3322 {
3323 int running, hostcpu;
3324 struct vmx *vmx = arg;
3325
3326 running = vcpu_is_running(vmx->vm, vcpu, &hostcpu);
3327 if (running && hostcpu != curcpu)
3328 panic("vmx_getreg: %s%d is running", vm_name(vmx->vm), vcpu);
3329
3330 if (reg == VM_REG_GUEST_INTR_SHADOW)
3331 return (vmx_get_intr_shadow(vmx, vcpu, running, retval));
3332
3333 if (vmxctx_getreg(&vmx->ctx[vcpu], reg, retval) == 0)
3334 return (0);
3335
3336 return (vmcs_getreg(&vmx->vmcs[vcpu], running, reg, retval));
3337 }
3338
3339 static int
3340 vmx_setreg(void *arg, int vcpu, int reg, uint64_t val)
3341 {
3342 int error, hostcpu, running, shadow;
3343 uint64_t ctls;
3344 pmap_t pmap;
3345 struct vmx *vmx = arg;
3346
3347 running = vcpu_is_running(vmx->vm, vcpu, &hostcpu);
3348 if (running && hostcpu != curcpu)
3349 panic("vmx_setreg: %s%d is running", vm_name(vmx->vm), vcpu);
3350
3351 if (reg == VM_REG_GUEST_INTR_SHADOW)
3352 return (vmx_modify_intr_shadow(vmx, vcpu, running, val));
3353
3354 if (vmxctx_setreg(&vmx->ctx[vcpu], reg, val) == 0)
3355 return (0);
3356
3357 /* Do not permit user write access to VMCS fields by offset. */
3358 if (reg < 0)
3359 return (EINVAL);
3360
3361 error = vmcs_setreg(&vmx->vmcs[vcpu], running, reg, val);
3362
3363 if (error == 0) {
3364 /*
3365 * If the "load EFER" VM-entry control is 1 then the
3366 * value of EFER.LMA must be identical to "IA-32e mode guest"
3367 * bit in the VM-entry control.
3368 */
3369 if ((entry_ctls & VM_ENTRY_LOAD_EFER) != 0 &&
3370 (reg == VM_REG_GUEST_EFER)) {
3371 vmcs_getreg(&vmx->vmcs[vcpu], running,
3372 VMCS_IDENT(VMCS_ENTRY_CTLS), &ctls);
3373 if (val & EFER_LMA)
3374 ctls |= VM_ENTRY_GUEST_LMA;
3375 else
3376 ctls &= ~VM_ENTRY_GUEST_LMA;
3377 vmcs_setreg(&vmx->vmcs[vcpu], running,
3378 VMCS_IDENT(VMCS_ENTRY_CTLS), ctls);
3379 }
3380
3381 shadow = vmx_shadow_reg(reg);
3382 if (shadow > 0) {
3383 /*
3384 * Store the unmodified value in the shadow
3385 */
3386 error = vmcs_setreg(&vmx->vmcs[vcpu], running,
3387 VMCS_IDENT(shadow), val);
3388 }
3389
3390 if (reg == VM_REG_GUEST_CR3) {
3391 /*
3392 * Invalidate the guest vcpu's TLB mappings to emulate
3393 * the behavior of updating %cr3.
3394 *
3395 * XXX the processor retains global mappings when %cr3
3396 * is updated but vmx_invvpid() does not.
3397 */
3398 pmap = vmx->ctx[vcpu].pmap;
3399 vmx_invvpid(vmx, vcpu, pmap, running);
3400 }
3401 }
3402
3403 return (error);
3404 }
3405
3406 static int
3407 vmx_getdesc(void *arg, int vcpu, int reg, struct seg_desc *desc)
3408 {
3409 int hostcpu, running;
3410 struct vmx *vmx = arg;
3411
3412 running = vcpu_is_running(vmx->vm, vcpu, &hostcpu);
3413 if (running && hostcpu != curcpu)
3414 panic("vmx_getdesc: %s%d is running", vm_name(vmx->vm), vcpu);
3415
3416 return (vmcs_getdesc(&vmx->vmcs[vcpu], running, reg, desc));
3417 }
3418
3419 static int
3420 vmx_setdesc(void *arg, int vcpu, int reg, struct seg_desc *desc)
3421 {
3422 int hostcpu, running;
3423 struct vmx *vmx = arg;
3424
3425 running = vcpu_is_running(vmx->vm, vcpu, &hostcpu);
3426 if (running && hostcpu != curcpu)
3427 panic("vmx_setdesc: %s%d is running", vm_name(vmx->vm), vcpu);
3428
3429 return (vmcs_setdesc(&vmx->vmcs[vcpu], running, reg, desc));
3430 }
3431
3432 static int
3433 vmx_getcap(void *arg, int vcpu, int type, int *retval)
3434 {
3435 struct vmx *vmx = arg;
3436 int vcap;
3437 int ret;
3438
3439 ret = ENOENT;
3440
3441 vcap = vmx->cap[vcpu].set;
3442
3443 switch (type) {
3444 case VM_CAP_HALT_EXIT:
3445 if (cap_halt_exit)
3446 ret = 0;
3447 break;
3448 case VM_CAP_PAUSE_EXIT:
3449 if (cap_pause_exit)
3450 ret = 0;
3451 break;
3452 case VM_CAP_MTRAP_EXIT:
3453 if (cap_monitor_trap)
3454 ret = 0;
3455 break;
3456 case VM_CAP_RDPID:
3457 if (cap_rdpid)
3458 ret = 0;
3459 break;
3460 case VM_CAP_RDTSCP:
3461 if (cap_rdtscp)
3462 ret = 0;
3463 break;
3464 case VM_CAP_UNRESTRICTED_GUEST:
3465 if (cap_unrestricted_guest)
3466 ret = 0;
3467 break;
3468 case VM_CAP_ENABLE_INVPCID:
3469 if (cap_invpcid)
3470 ret = 0;
3471 break;
3472 case VM_CAP_BPT_EXIT:
3473 ret = 0;
3474 break;
3475 default:
3476 break;
3477 }
3478
3479 if (ret == 0)
3480 *retval = (vcap & (1 << type)) ? 1 : 0;
3481
3482 return (ret);
3483 }
3484
3485 static int
3486 vmx_setcap(void *arg, int vcpu, int type, int val)
3487 {
3488 struct vmx *vmx = arg;
3489 struct vmcs *vmcs = &vmx->vmcs[vcpu];
3490 uint32_t baseval;
3491 uint32_t *pptr;
3492 int error;
3493 int flag;
3494 int reg;
3495 int retval;
3496
3497 retval = ENOENT;
3498 pptr = NULL;
3499
3500 switch (type) {
3501 case VM_CAP_HALT_EXIT:
3502 if (cap_halt_exit) {
3503 retval = 0;
3504 pptr = &vmx->cap[vcpu].proc_ctls;
3505 baseval = *pptr;
3506 flag = PROCBASED_HLT_EXITING;
3507 reg = VMCS_PRI_PROC_BASED_CTLS;
3508 }
3509 break;
3510 case VM_CAP_MTRAP_EXIT:
3511 if (cap_monitor_trap) {
3512 retval = 0;
3513 pptr = &vmx->cap[vcpu].proc_ctls;
3514 baseval = *pptr;
3515 flag = PROCBASED_MTF;
3516 reg = VMCS_PRI_PROC_BASED_CTLS;
3517 }
3518 break;
3519 case VM_CAP_PAUSE_EXIT:
3520 if (cap_pause_exit) {
3521 retval = 0;
3522 pptr = &vmx->cap[vcpu].proc_ctls;
3523 baseval = *pptr;
3524 flag = PROCBASED_PAUSE_EXITING;
3525 reg = VMCS_PRI_PROC_BASED_CTLS;
3526 }
3527 break;
3528 case VM_CAP_RDPID:
3529 case VM_CAP_RDTSCP:
3530 if (cap_rdpid || cap_rdtscp)
3531 /*
3532 * Choose not to support enabling/disabling
3533 * RDPID/RDTSCP via libvmmapi since, as per the
3534 * discussion in vmx_init(), RDPID/RDTSCP are
3535 * either always enabled or always disabled.
3536 */
3537 error = EOPNOTSUPP;
3538 break;
3539 case VM_CAP_UNRESTRICTED_GUEST:
3540 if (cap_unrestricted_guest) {
3541 retval = 0;
3542 pptr = &vmx->cap[vcpu].proc_ctls2;
3543 baseval = *pptr;
3544 flag = PROCBASED2_UNRESTRICTED_GUEST;
3545 reg = VMCS_SEC_PROC_BASED_CTLS;
3546 }
3547 break;
3548 case VM_CAP_ENABLE_INVPCID:
3549 if (cap_invpcid) {
3550 retval = 0;
3551 pptr = &vmx->cap[vcpu].proc_ctls2;
3552 baseval = *pptr;
3553 flag = PROCBASED2_ENABLE_INVPCID;
3554 reg = VMCS_SEC_PROC_BASED_CTLS;
3555 }
3556 break;
3557 case VM_CAP_BPT_EXIT:
3558 retval = 0;
3559
3560 /* Don't change the bitmap if we are tracing all exceptions. */
3561 if (vmx->cap[vcpu].exc_bitmap != 0xffffffff) {
3562 pptr = &vmx->cap[vcpu].exc_bitmap;
3563 baseval = *pptr;
3564 flag = (1 << IDT_BP);
3565 reg = VMCS_EXCEPTION_BITMAP;
3566 }
3567 break;
3568 default:
3569 break;
3570 }
3571
3572 if (retval)
3573 return (retval);
3574
3575 if (pptr != NULL) {
3576 if (val) {
3577 baseval |= flag;
3578 } else {
3579 baseval &= ~flag;
3580 }
3581 VMPTRLD(vmcs);
3582 error = vmwrite(reg, baseval);
3583 VMCLEAR(vmcs);
3584
3585 if (error)
3586 return (error);
3587
3588 /*
3589 * Update optional stored flags, and record
3590 * setting
3591 */
3592 *pptr = baseval;
3593 }
3594
3595 if (val) {
3596 vmx->cap[vcpu].set |= (1 << type);
3597 } else {
3598 vmx->cap[vcpu].set &= ~(1 << type);
3599 }
3600
3601 return (0);
3602 }
3603
3604 struct vlapic_vtx {
3605 struct vlapic vlapic;
3606 struct pir_desc *pir_desc;
3607 struct vmx *vmx;
3608 u_int pending_prio;
3609 };
3610
3611 #define VPR_PRIO_BIT(vpr) (1 << ((vpr) >> 4))
3612
3613 #define VMX_CTR_PIR(vm, vcpuid, pir_desc, notify, vector, level, msg) \
3614 do { \
3615 VCPU_CTR2(vm, vcpuid, msg " assert %s-triggered vector %d", \
3616 level ? "level" : "edge", vector); \
3617 VCPU_CTR1(vm, vcpuid, msg " pir0 0x%016lx", pir_desc->pir[0]); \
3618 VCPU_CTR1(vm, vcpuid, msg " pir1 0x%016lx", pir_desc->pir[1]); \
3619 VCPU_CTR1(vm, vcpuid, msg " pir2 0x%016lx", pir_desc->pir[2]); \
3620 VCPU_CTR1(vm, vcpuid, msg " pir3 0x%016lx", pir_desc->pir[3]); \
3621 VCPU_CTR1(vm, vcpuid, msg " notify: %s", notify ? "yes" : "no");\
3622 } while (0)
3623
3624 /*
3625 * vlapic->ops handlers that utilize the APICv hardware assist described in
3626 * Chapter 29 of the Intel SDM.
3627 */
3628 static int
3629 vmx_set_intr_ready(struct vlapic *vlapic, int vector, bool level)
3630 {
3631 struct vlapic_vtx *vlapic_vtx;
3632 struct pir_desc *pir_desc;
3633 uint64_t mask;
3634 int idx, notify = 0;
3635
3636 vlapic_vtx = (struct vlapic_vtx *)vlapic;
3637 pir_desc = vlapic_vtx->pir_desc;
3638
3639 /*
3640 * Keep track of interrupt requests in the PIR descriptor. This is
3641 * because the virtual APIC page pointed to by the VMCS cannot be
3642 * modified if the vcpu is running.
3643 */
3644 idx = vector / 64;
3645 mask = 1UL << (vector % 64);
3646 atomic_set_long(&pir_desc->pir[idx], mask);
3647
3648 /*
3649 * A notification is required whenever the 'pending' bit makes a
3650 * transition from 0->1.
3651 *
3652 * Even if the 'pending' bit is already asserted, notification about
3653 * the incoming interrupt may still be necessary. For example, if a
3654 * vCPU is HLTed with a high PPR, a low priority interrupt would cause
3655 * the 0->1 'pending' transition with a notification, but the vCPU
3656 * would ignore the interrupt for the time being. The same vCPU would
3657 * need to then be notified if a high-priority interrupt arrived which
3658 * satisfied the PPR.
3659 *
3660 * The priorities of interrupts injected while 'pending' is asserted
3661 * are tracked in a custom bitfield 'pending_prio'. Should the
3662 * to-be-injected interrupt exceed the priorities already present, the
3663 * notification is sent. The priorities recorded in 'pending_prio' are
3664 * cleared whenever the 'pending' bit makes another 0->1 transition.
3665 */
3666 if (atomic_cmpset_long(&pir_desc->pending, 0, 1) != 0) {
3667 notify = 1;
3668 vlapic_vtx->pending_prio = 0;
3669 } else {
3670 const u_int old_prio = vlapic_vtx->pending_prio;
3671 const u_int prio_bit = VPR_PRIO_BIT(vector & APIC_TPR_INT);
3672
3673 if ((old_prio & prio_bit) == 0 && prio_bit > old_prio) {
3674 atomic_set_int(&vlapic_vtx->pending_prio, prio_bit);
3675 notify = 1;
3676 }
3677 }
3678
3679 VMX_CTR_PIR(vlapic->vm, vlapic->vcpuid, pir_desc, notify, vector,
3680 level, "vmx_set_intr_ready");
3681 return (notify);
3682 }
3683
3684 static int
3685 vmx_pending_intr(struct vlapic *vlapic, int *vecptr)
3686 {
3687 struct vlapic_vtx *vlapic_vtx;
3688 struct pir_desc *pir_desc;
3689 struct LAPIC *lapic;
3690 uint64_t pending, pirval;
3691 uint32_t ppr, vpr;
3692 int i;
3693
3694 /*
3695 * This function is only expected to be called from the 'HLT' exit
3696 * handler which does not care about the vector that is pending.
3697 */
3698 KASSERT(vecptr == NULL, ("vmx_pending_intr: vecptr must be NULL"));
3699
3700 vlapic_vtx = (struct vlapic_vtx *)vlapic;
3701 pir_desc = vlapic_vtx->pir_desc;
3702
3703 pending = atomic_load_acq_long(&pir_desc->pending);
3704 if (!pending) {
3705 /*
3706 * While a virtual interrupt may have already been
3707 * processed the actual delivery maybe pending the
3708 * interruptibility of the guest. Recognize a pending
3709 * interrupt by reevaluating virtual interrupts
3710 * following Section 29.2.1 in the Intel SDM Volume 3.
3711 */
3712 struct vm_exit *vmexit;
3713 uint8_t rvi, ppr;
3714
3715 vmexit = vm_exitinfo(vlapic->vm, vlapic->vcpuid);
3716 KASSERT(vmexit->exitcode == VM_EXITCODE_HLT,
3717 ("vmx_pending_intr: exitcode not 'HLT'"));
3718 rvi = vmexit->u.hlt.intr_status & APIC_TPR_INT;
3719 lapic = vlapic->apic_page;
3720 ppr = lapic->ppr & APIC_TPR_INT;
3721 if (rvi > ppr) {
3722 return (1);
3723 }
3724
3725 return (0);
3726 }
3727
3728 /*
3729 * If there is an interrupt pending then it will be recognized only
3730 * if its priority is greater than the processor priority.
3731 *
3732 * Special case: if the processor priority is zero then any pending
3733 * interrupt will be recognized.
3734 */
3735 lapic = vlapic->apic_page;
3736 ppr = lapic->ppr & APIC_TPR_INT;
3737 if (ppr == 0)
3738 return (1);
3739
3740 VCPU_CTR1(vlapic->vm, vlapic->vcpuid, "HLT with non-zero PPR %d",
3741 lapic->ppr);
3742
3743 vpr = 0;
3744 for (i = 3; i >= 0; i--) {
3745 pirval = pir_desc->pir[i];
3746 if (pirval != 0) {
3747 vpr = (i * 64 + flsl(pirval) - 1) & APIC_TPR_INT;
3748 break;
3749 }
3750 }
3751
3752 /*
3753 * If the highest-priority pending interrupt falls short of the
3754 * processor priority of this vCPU, ensure that 'pending_prio' does not
3755 * have any stale bits which would preclude a higher-priority interrupt
3756 * from incurring a notification later.
3757 */
3758 if (vpr <= ppr) {
3759 const u_int prio_bit = VPR_PRIO_BIT(vpr);
3760 const u_int old = vlapic_vtx->pending_prio;
3761
3762 if (old > prio_bit && (old & prio_bit) == 0) {
3763 vlapic_vtx->pending_prio = prio_bit;
3764 }
3765 return (0);
3766 }
3767 return (1);
3768 }
3769
3770 static void
3771 vmx_intr_accepted(struct vlapic *vlapic, int vector)
3772 {
3773
3774 panic("vmx_intr_accepted: not expected to be called");
3775 }
3776
3777 static void
3778 vmx_set_tmr(struct vlapic *vlapic, int vector, bool level)
3779 {
3780 struct vlapic_vtx *vlapic_vtx;
3781 struct vmx *vmx;
3782 struct vmcs *vmcs;
3783 uint64_t mask, val;
3784
3785 KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d", vector));
3786 KASSERT(!vcpu_is_running(vlapic->vm, vlapic->vcpuid, NULL),
3787 ("vmx_set_tmr: vcpu cannot be running"));
3788
3789 vlapic_vtx = (struct vlapic_vtx *)vlapic;
3790 vmx = vlapic_vtx->vmx;
3791 vmcs = &vmx->vmcs[vlapic->vcpuid];
3792 mask = 1UL << (vector % 64);
3793
3794 VMPTRLD(vmcs);
3795 val = vmcs_read(VMCS_EOI_EXIT(vector));
3796 if (level)
3797 val |= mask;
3798 else
3799 val &= ~mask;
3800 vmcs_write(VMCS_EOI_EXIT(vector), val);
3801 VMCLEAR(vmcs);
3802 }
3803
3804 static void
3805 vmx_enable_x2apic_mode_ts(struct vlapic *vlapic)
3806 {
3807 struct vmx *vmx;
3808 struct vmcs *vmcs;
3809 uint32_t proc_ctls;
3810 int vcpuid;
3811
3812 vcpuid = vlapic->vcpuid;
3813 vmx = ((struct vlapic_vtx *)vlapic)->vmx;
3814 vmcs = &vmx->vmcs[vcpuid];
3815
3816 proc_ctls = vmx->cap[vcpuid].proc_ctls;
3817 proc_ctls &= ~PROCBASED_USE_TPR_SHADOW;
3818 proc_ctls |= PROCBASED_CR8_LOAD_EXITING;
3819 proc_ctls |= PROCBASED_CR8_STORE_EXITING;
3820 vmx->cap[vcpuid].proc_ctls = proc_ctls;
3821
3822 VMPTRLD(vmcs);
3823 vmcs_write(VMCS_PRI_PROC_BASED_CTLS, proc_ctls);
3824 VMCLEAR(vmcs);
3825 }
3826
3827 static void
3828 vmx_enable_x2apic_mode_vid(struct vlapic *vlapic)
3829 {
3830 struct vmx *vmx;
3831 struct vmcs *vmcs;
3832 uint32_t proc_ctls2;
3833 int vcpuid, error;
3834
3835 vcpuid = vlapic->vcpuid;
3836 vmx = ((struct vlapic_vtx *)vlapic)->vmx;
3837 vmcs = &vmx->vmcs[vcpuid];
3838
3839 proc_ctls2 = vmx->cap[vcpuid].proc_ctls2;
3840 KASSERT((proc_ctls2 & PROCBASED2_VIRTUALIZE_APIC_ACCESSES) != 0,
3841 ("%s: invalid proc_ctls2 %#x", __func__, proc_ctls2));
3842
3843 proc_ctls2 &= ~PROCBASED2_VIRTUALIZE_APIC_ACCESSES;
3844 proc_ctls2 |= PROCBASED2_VIRTUALIZE_X2APIC_MODE;
3845 vmx->cap[vcpuid].proc_ctls2 = proc_ctls2;
3846
3847 VMPTRLD(vmcs);
3848 vmcs_write(VMCS_SEC_PROC_BASED_CTLS, proc_ctls2);
3849 VMCLEAR(vmcs);
3850
3851 if (vlapic->vcpuid == 0) {
3852 /*
3853 * The nested page table mappings are shared by all vcpus
3854 * so unmap the APIC access page just once.
3855 */
3856 error = vm_unmap_mmio(vmx->vm, DEFAULT_APIC_BASE, PAGE_SIZE);
3857 KASSERT(error == 0, ("%s: vm_unmap_mmio error %d",
3858 __func__, error));
3859
3860 /*
3861 * The MSR bitmap is shared by all vcpus so modify it only
3862 * once in the context of vcpu 0.
3863 */
3864 error = vmx_allow_x2apic_msrs(vmx);
3865 KASSERT(error == 0, ("%s: vmx_allow_x2apic_msrs error %d",
3866 __func__, error));
3867 }
3868 }
3869
3870 static void
3871 vmx_post_intr(struct vlapic *vlapic, int hostcpu)
3872 {
3873
3874 ipi_cpu(hostcpu, pirvec);
3875 }
3876
3877 /*
3878 * Transfer the pending interrupts in the PIR descriptor to the IRR
3879 * in the virtual APIC page.
3880 */
3881 static void
3882 vmx_inject_pir(struct vlapic *vlapic)
3883 {
3884 struct vlapic_vtx *vlapic_vtx;
3885 struct pir_desc *pir_desc;
3886 struct LAPIC *lapic;
3887 uint64_t val, pirval;
3888 int rvi, pirbase = -1;
3889 uint16_t intr_status_old, intr_status_new;
3890
3891 vlapic_vtx = (struct vlapic_vtx *)vlapic;
3892 pir_desc = vlapic_vtx->pir_desc;
3893 if (atomic_cmpset_long(&pir_desc->pending, 1, 0) == 0) {
3894 VCPU_CTR0(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: "
3895 "no posted interrupt pending");
3896 return;
3897 }
3898
3899 pirval = 0;
3900 pirbase = -1;
3901 lapic = vlapic->apic_page;
3902
3903 val = atomic_readandclear_long(&pir_desc->pir[0]);
3904 if (val != 0) {
3905 lapic->irr0 |= val;
3906 lapic->irr1 |= val >> 32;
3907 pirbase = 0;
3908 pirval = val;
3909 }
3910
3911 val = atomic_readandclear_long(&pir_desc->pir[1]);
3912 if (val != 0) {
3913 lapic->irr2 |= val;
3914 lapic->irr3 |= val >> 32;
3915 pirbase = 64;
3916 pirval = val;
3917 }
3918
3919 val = atomic_readandclear_long(&pir_desc->pir[2]);
3920 if (val != 0) {
3921 lapic->irr4 |= val;
3922 lapic->irr5 |= val >> 32;
3923 pirbase = 128;
3924 pirval = val;
3925 }
3926
3927 val = atomic_readandclear_long(&pir_desc->pir[3]);
3928 if (val != 0) {
3929 lapic->irr6 |= val;
3930 lapic->irr7 |= val >> 32;
3931 pirbase = 192;
3932 pirval = val;
3933 }
3934
3935 VLAPIC_CTR_IRR(vlapic, "vmx_inject_pir");
3936
3937 /*
3938 * Update RVI so the processor can evaluate pending virtual
3939 * interrupts on VM-entry.
3940 *
3941 * It is possible for pirval to be 0 here, even though the
3942 * pending bit has been set. The scenario is:
3943 * CPU-Y is sending a posted interrupt to CPU-X, which
3944 * is running a guest and processing posted interrupts in h/w.
3945 * CPU-X will eventually exit and the state seen in s/w is
3946 * the pending bit set, but no PIR bits set.
3947 *
3948 * CPU-X CPU-Y
3949 * (vm running) (host running)
3950 * rx posted interrupt
3951 * CLEAR pending bit
3952 * SET PIR bit
3953 * READ/CLEAR PIR bits
3954 * SET pending bit
3955 * (vm exit)
3956 * pending bit set, PIR 0
3957 */
3958 if (pirval != 0) {
3959 rvi = pirbase + flsl(pirval) - 1;
3960 intr_status_old = vmcs_read(VMCS_GUEST_INTR_STATUS);
3961 intr_status_new = (intr_status_old & 0xFF00) | rvi;
3962 if (intr_status_new > intr_status_old) {
3963 vmcs_write(VMCS_GUEST_INTR_STATUS, intr_status_new);
3964 VCPU_CTR2(vlapic->vm, vlapic->vcpuid, "vmx_inject_pir: "
3965 "guest_intr_status changed from 0x%04x to 0x%04x",
3966 intr_status_old, intr_status_new);
3967 }
3968 }
3969 }
3970
3971 static struct vlapic *
3972 vmx_vlapic_init(void *arg, int vcpuid)
3973 {
3974 struct vmx *vmx;
3975 struct vlapic *vlapic;
3976 struct vlapic_vtx *vlapic_vtx;
3977
3978 vmx = arg;
3979
3980 vlapic = malloc(sizeof(struct vlapic_vtx), M_VLAPIC, M_WAITOK | M_ZERO);
3981 vlapic->vm = vmx->vm;
3982 vlapic->vcpuid = vcpuid;
3983 vlapic->apic_page = (struct LAPIC *)&vmx->apic_page[vcpuid];
3984
3985 vlapic_vtx = (struct vlapic_vtx *)vlapic;
3986 vlapic_vtx->pir_desc = &vmx->pir_desc[vcpuid];
3987 vlapic_vtx->vmx = vmx;
3988
3989 if (tpr_shadowing) {
3990 vlapic->ops.enable_x2apic_mode = vmx_enable_x2apic_mode_ts;
3991 }
3992
3993 if (virtual_interrupt_delivery) {
3994 vlapic->ops.set_intr_ready = vmx_set_intr_ready;
3995 vlapic->ops.pending_intr = vmx_pending_intr;
3996 vlapic->ops.intr_accepted = vmx_intr_accepted;
3997 vlapic->ops.set_tmr = vmx_set_tmr;
3998 vlapic->ops.enable_x2apic_mode = vmx_enable_x2apic_mode_vid;
3999 }
4000
4001 if (posted_interrupts)
4002 vlapic->ops.post_intr = vmx_post_intr;
4003
4004 vlapic_init(vlapic);
4005
4006 return (vlapic);
4007 }
4008
4009 static void
4010 vmx_vlapic_cleanup(void *arg, struct vlapic *vlapic)
4011 {
4012
4013 vlapic_cleanup(vlapic);
4014 free(vlapic, M_VLAPIC);
4015 }
4016
4017 struct vmm_ops vmm_ops_intel = {
4018 .init = vmx_init,
4019 .cleanup = vmx_cleanup,
4020 .resume = vmx_restore,
4021 .vminit = vmx_vminit,
4022 .vmrun = vmx_run,
4023 .vmcleanup = vmx_vmcleanup,
4024 .vmgetreg = vmx_getreg,
4025 .vmsetreg = vmx_setreg,
4026 .vmgetdesc = vmx_getdesc,
4027 .vmsetdesc = vmx_setdesc,
4028 .vmgetcap = vmx_getcap,
4029 .vmsetcap = vmx_setcap,
4030 .vmspace_alloc = ept_vmspace_alloc,
4031 .vmspace_free = ept_vmspace_free,
4032 .vlapic_init = vmx_vlapic_init,
4033 .vlapic_cleanup = vmx_vlapic_cleanup,
4034 };
Cache object: 2e437d1aa1ad06b7432cc933a5a34d0e
|