]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/amd64/vmm/io/vlapic.c
Use V_IRQ, V_INTR_VECTOR and V_TPR to offload APIC interrupt delivery to the
[FreeBSD/FreeBSD.git] / sys / amd64 / vmm / io / vlapic.c
1 /*-
2  * Copyright (c) 2011 NetApp, Inc.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  */
28
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31
32 #include <sys/param.h>
33 #include <sys/lock.h>
34 #include <sys/kernel.h>
35 #include <sys/malloc.h>
36 #include <sys/mutex.h>
37 #include <sys/systm.h>
38 #include <sys/smp.h>
39
40 #include <x86/specialreg.h>
41 #include <x86/apicreg.h>
42
43 #include <machine/clock.h>
44 #include <machine/smp.h>
45
46 #include <machine/vmm.h>
47
48 #include "vmm_ipi.h"
49 #include "vmm_lapic.h"
50 #include "vmm_ktr.h"
51 #include "vmm_stat.h"
52
53 #include "vlapic.h"
54 #include "vlapic_priv.h"
55 #include "vioapic.h"
56
57 #define PRIO(x)                 ((x) >> 4)
58
59 #define VLAPIC_VERSION          (16)
60
61 #define x2apic(vlapic)  (((vlapic)->msr_apicbase & APICBASE_X2APIC) ? 1 : 0)
62
63 /*
64  * The 'vlapic->timer_mtx' is used to provide mutual exclusion between the
65  * vlapic_callout_handler() and vcpu accesses to:
66  * - timer_freq_bt, timer_period_bt, timer_fire_bt
67  * - timer LVT register
68  */
69 #define VLAPIC_TIMER_LOCK(vlapic)       mtx_lock_spin(&((vlapic)->timer_mtx))
70 #define VLAPIC_TIMER_UNLOCK(vlapic)     mtx_unlock_spin(&((vlapic)->timer_mtx))
71 #define VLAPIC_TIMER_LOCKED(vlapic)     mtx_owned(&((vlapic)->timer_mtx))
72
73 /*
74  * APIC timer frequency:
75  * - arbitrary but chosen to be in the ballpark of contemporary hardware.
76  * - power-of-two to avoid loss of precision when converted to a bintime.
77  */
78 #define VLAPIC_BUS_FREQ         (128 * 1024 * 1024)
79
80 static __inline uint32_t
81 vlapic_get_id(struct vlapic *vlapic)
82 {
83
84         if (x2apic(vlapic))
85                 return (vlapic->vcpuid);
86         else
87                 return (vlapic->vcpuid << 24);
88 }
89
90 static uint32_t
91 x2apic_ldr(struct vlapic *vlapic)
92 {
93         int apicid;
94         uint32_t ldr;
95
96         apicid = vlapic_get_id(vlapic);
97         ldr = 1 << (apicid & 0xf);
98         ldr |= (apicid & 0xffff0) << 12;
99         return (ldr);
100 }
101
102 void
103 vlapic_dfr_write_handler(struct vlapic *vlapic)
104 {
105         struct LAPIC *lapic;
106
107         lapic = vlapic->apic_page;
108         if (x2apic(vlapic)) {
109                 VM_CTR1(vlapic->vm, "ignoring write to DFR in x2apic mode: %#x",
110                     lapic->dfr);
111                 lapic->dfr = 0;
112                 return;
113         }
114
115         lapic->dfr &= APIC_DFR_MODEL_MASK;
116         lapic->dfr |= APIC_DFR_RESERVED;
117
118         if ((lapic->dfr & APIC_DFR_MODEL_MASK) == APIC_DFR_MODEL_FLAT)
119                 VLAPIC_CTR0(vlapic, "vlapic DFR in Flat Model");
120         else if ((lapic->dfr & APIC_DFR_MODEL_MASK) == APIC_DFR_MODEL_CLUSTER)
121                 VLAPIC_CTR0(vlapic, "vlapic DFR in Cluster Model");
122         else
123                 VLAPIC_CTR1(vlapic, "DFR in Unknown Model %#x", lapic->dfr);
124 }
125
126 void
127 vlapic_ldr_write_handler(struct vlapic *vlapic)
128 {
129         struct LAPIC *lapic;
130
131         lapic = vlapic->apic_page;
132
133         /* LDR is read-only in x2apic mode */
134         if (x2apic(vlapic)) {
135                 VLAPIC_CTR1(vlapic, "ignoring write to LDR in x2apic mode: %#x",
136                     lapic->ldr);
137                 lapic->ldr = x2apic_ldr(vlapic);
138         } else {
139                 lapic->ldr &= ~APIC_LDR_RESERVED;
140                 VLAPIC_CTR1(vlapic, "vlapic LDR set to %#x", lapic->ldr);
141         }
142 }
143
144 void
145 vlapic_id_write_handler(struct vlapic *vlapic)
146 {
147         struct LAPIC *lapic;
148         
149         /*
150          * We don't allow the ID register to be modified so reset it back to
151          * its default value.
152          */
153         lapic = vlapic->apic_page;
154         lapic->id = vlapic_get_id(vlapic);
155 }
156
157 static int
158 vlapic_timer_divisor(uint32_t dcr)
159 {
160         switch (dcr & 0xB) {
161         case APIC_TDCR_1:
162                 return (1);
163         case APIC_TDCR_2:
164                 return (2);
165         case APIC_TDCR_4:
166                 return (4);
167         case APIC_TDCR_8:
168                 return (8);
169         case APIC_TDCR_16:
170                 return (16);
171         case APIC_TDCR_32:
172                 return (32);
173         case APIC_TDCR_64:
174                 return (64);
175         case APIC_TDCR_128:
176                 return (128);
177         default:
178                 panic("vlapic_timer_divisor: invalid dcr 0x%08x", dcr);
179         }
180 }
181
182 #if 0
183 static inline void
184 vlapic_dump_lvt(uint32_t offset, uint32_t *lvt)
185 {
186         printf("Offset %x: lvt %08x (V:%02x DS:%x M:%x)\n", offset,
187             *lvt, *lvt & APIC_LVTT_VECTOR, *lvt & APIC_LVTT_DS,
188             *lvt & APIC_LVTT_M);
189 }
190 #endif
191
192 static uint32_t
193 vlapic_get_ccr(struct vlapic *vlapic)
194 {
195         struct bintime bt_now, bt_rem;
196         struct LAPIC *lapic;
197         uint32_t ccr;
198         
199         ccr = 0;
200         lapic = vlapic->apic_page;
201
202         VLAPIC_TIMER_LOCK(vlapic);
203         if (callout_active(&vlapic->callout)) {
204                 /*
205                  * If the timer is scheduled to expire in the future then
206                  * compute the value of 'ccr' based on the remaining time.
207                  */
208                 binuptime(&bt_now);
209                 if (bintime_cmp(&vlapic->timer_fire_bt, &bt_now, >)) {
210                         bt_rem = vlapic->timer_fire_bt;
211                         bintime_sub(&bt_rem, &bt_now);
212                         ccr += bt_rem.sec * BT2FREQ(&vlapic->timer_freq_bt);
213                         ccr += bt_rem.frac / vlapic->timer_freq_bt.frac;
214                 }
215         }
216         KASSERT(ccr <= lapic->icr_timer, ("vlapic_get_ccr: invalid ccr %#x, "
217             "icr_timer is %#x", ccr, lapic->icr_timer));
218         VLAPIC_CTR2(vlapic, "vlapic ccr_timer = %#x, icr_timer = %#x",
219             ccr, lapic->icr_timer);
220         VLAPIC_TIMER_UNLOCK(vlapic);
221         return (ccr);
222 }
223
224 void
225 vlapic_dcr_write_handler(struct vlapic *vlapic)
226 {
227         struct LAPIC *lapic;
228         int divisor;
229         
230         lapic = vlapic->apic_page;
231         VLAPIC_TIMER_LOCK(vlapic);
232
233         divisor = vlapic_timer_divisor(lapic->dcr_timer);
234         VLAPIC_CTR2(vlapic, "vlapic dcr_timer=%#x, divisor=%d",
235             lapic->dcr_timer, divisor);
236
237         /*
238          * Update the timer frequency and the timer period.
239          *
240          * XXX changes to the frequency divider will not take effect until
241          * the timer is reloaded.
242          */
243         FREQ2BT(VLAPIC_BUS_FREQ / divisor, &vlapic->timer_freq_bt);
244         vlapic->timer_period_bt = vlapic->timer_freq_bt;
245         bintime_mul(&vlapic->timer_period_bt, lapic->icr_timer);
246
247         VLAPIC_TIMER_UNLOCK(vlapic);
248 }
249
250 void
251 vlapic_esr_write_handler(struct vlapic *vlapic)
252 {
253         struct LAPIC *lapic;
254         
255         lapic = vlapic->apic_page;
256         lapic->esr = vlapic->esr_pending;
257         vlapic->esr_pending = 0;
258 }
259
260 int
261 vlapic_set_intr_ready(struct vlapic *vlapic, int vector, bool level)
262 {
263         struct LAPIC *lapic;
264         uint32_t *irrptr, *tmrptr, mask;
265         int idx;
266
267         KASSERT(vector >= 0 && vector < 256, ("invalid vector %d", vector));
268
269         lapic = vlapic->apic_page;
270         if (!(lapic->svr & APIC_SVR_ENABLE)) {
271                 VLAPIC_CTR1(vlapic, "vlapic is software disabled, ignoring "
272                     "interrupt %d", vector);
273                 return (0);
274         }
275
276         if (vector < 16) {
277                 vlapic_set_error(vlapic, APIC_ESR_RECEIVE_ILLEGAL_VECTOR);
278                 VLAPIC_CTR1(vlapic, "vlapic ignoring interrupt to vector %d",
279                     vector);
280                 return (1);
281         }
282
283         if (vlapic->ops.set_intr_ready)
284                 return ((*vlapic->ops.set_intr_ready)(vlapic, vector, level));
285
286         idx = (vector / 32) * 4;
287         mask = 1 << (vector % 32);
288
289         irrptr = &lapic->irr0;
290         atomic_set_int(&irrptr[idx], mask);
291
292         /*
293          * Verify that the trigger-mode of the interrupt matches with
294          * the vlapic TMR registers.
295          */
296         tmrptr = &lapic->tmr0;
297         if ((tmrptr[idx] & mask) != (level ? mask : 0)) {
298                 VLAPIC_CTR3(vlapic, "vlapic TMR[%d] is 0x%08x but "
299                     "interrupt is %s-triggered", idx / 4, tmrptr[idx],
300                     level ? "level" : "edge");
301         }
302
303         VLAPIC_CTR_IRR(vlapic, "vlapic_set_intr_ready");
304         return (1);
305 }
306
307 static __inline uint32_t *
308 vlapic_get_lvtptr(struct vlapic *vlapic, uint32_t offset)
309 {
310         struct LAPIC    *lapic = vlapic->apic_page;
311         int              i;
312
313         switch (offset) {
314         case APIC_OFFSET_CMCI_LVT:
315                 return (&lapic->lvt_cmci);
316         case APIC_OFFSET_TIMER_LVT ... APIC_OFFSET_ERROR_LVT:
317                 i = (offset - APIC_OFFSET_TIMER_LVT) >> 2;
318                 return ((&lapic->lvt_timer) + i);;
319         default:
320                 panic("vlapic_get_lvt: invalid LVT\n");
321         }
322 }
323
324 static __inline int
325 lvt_off_to_idx(uint32_t offset)
326 {
327         int index;
328
329         switch (offset) {
330         case APIC_OFFSET_CMCI_LVT:
331                 index = APIC_LVT_CMCI;
332                 break;
333         case APIC_OFFSET_TIMER_LVT:
334                 index = APIC_LVT_TIMER;
335                 break;
336         case APIC_OFFSET_THERM_LVT:
337                 index = APIC_LVT_THERMAL;
338                 break;
339         case APIC_OFFSET_PERF_LVT:
340                 index = APIC_LVT_PMC;
341                 break;
342         case APIC_OFFSET_LINT0_LVT:
343                 index = APIC_LVT_LINT0;
344                 break;
345         case APIC_OFFSET_LINT1_LVT:
346                 index = APIC_LVT_LINT1;
347                 break;
348         case APIC_OFFSET_ERROR_LVT:
349                 index = APIC_LVT_ERROR;
350                 break;
351         default:
352                 index = -1;
353                 break;
354         }
355         KASSERT(index >= 0 && index <= VLAPIC_MAXLVT_INDEX, ("lvt_off_to_idx: "
356             "invalid lvt index %d for offset %#x", index, offset));
357
358         return (index);
359 }
360
361 static __inline uint32_t
362 vlapic_get_lvt(struct vlapic *vlapic, uint32_t offset)
363 {
364         int idx;
365         uint32_t val;
366
367         idx = lvt_off_to_idx(offset);
368         val = atomic_load_acq_32(&vlapic->lvt_last[idx]);
369         return (val);
370 }
371
372 void
373 vlapic_lvt_write_handler(struct vlapic *vlapic, uint32_t offset)
374 {
375         uint32_t *lvtptr, mask, val;
376         struct LAPIC *lapic;
377         int idx;
378         
379         lapic = vlapic->apic_page;
380         lvtptr = vlapic_get_lvtptr(vlapic, offset);     
381         val = *lvtptr;
382         idx = lvt_off_to_idx(offset);
383
384         if (!(lapic->svr & APIC_SVR_ENABLE))
385                 val |= APIC_LVT_M;
386         mask = APIC_LVT_M | APIC_LVT_DS | APIC_LVT_VECTOR;
387         switch (offset) {
388         case APIC_OFFSET_TIMER_LVT:
389                 mask |= APIC_LVTT_TM;
390                 break;
391         case APIC_OFFSET_ERROR_LVT:
392                 break;
393         case APIC_OFFSET_LINT0_LVT:
394         case APIC_OFFSET_LINT1_LVT:
395                 mask |= APIC_LVT_TM | APIC_LVT_RIRR | APIC_LVT_IIPP;
396                 /* FALLTHROUGH */
397         default:
398                 mask |= APIC_LVT_DM;
399                 break;
400         }
401         val &= mask;
402         *lvtptr = val;
403         atomic_store_rel_32(&vlapic->lvt_last[idx], val);
404 }
405
406 static void
407 vlapic_mask_lvts(struct vlapic *vlapic)
408 {
409         struct LAPIC *lapic = vlapic->apic_page;
410
411         lapic->lvt_cmci |= APIC_LVT_M;
412         vlapic_lvt_write_handler(vlapic, APIC_OFFSET_CMCI_LVT);
413
414         lapic->lvt_timer |= APIC_LVT_M;
415         vlapic_lvt_write_handler(vlapic, APIC_OFFSET_TIMER_LVT);
416
417         lapic->lvt_thermal |= APIC_LVT_M;
418         vlapic_lvt_write_handler(vlapic, APIC_OFFSET_THERM_LVT);
419
420         lapic->lvt_pcint |= APIC_LVT_M;
421         vlapic_lvt_write_handler(vlapic, APIC_OFFSET_PERF_LVT);
422
423         lapic->lvt_lint0 |= APIC_LVT_M;
424         vlapic_lvt_write_handler(vlapic, APIC_OFFSET_LINT0_LVT);
425
426         lapic->lvt_lint1 |= APIC_LVT_M;
427         vlapic_lvt_write_handler(vlapic, APIC_OFFSET_LINT1_LVT);
428
429         lapic->lvt_error |= APIC_LVT_M;
430         vlapic_lvt_write_handler(vlapic, APIC_OFFSET_ERROR_LVT);
431 }
432
433 static int
434 vlapic_fire_lvt(struct vlapic *vlapic, uint32_t lvt)
435 {
436         uint32_t vec, mode;
437
438         if (lvt & APIC_LVT_M)
439                 return (0);
440
441         vec = lvt & APIC_LVT_VECTOR;
442         mode = lvt & APIC_LVT_DM;
443
444         switch (mode) {
445         case APIC_LVT_DM_FIXED:
446                 if (vec < 16) {
447                         vlapic_set_error(vlapic, APIC_ESR_SEND_ILLEGAL_VECTOR);
448                         return (0);
449                 }
450                 if (vlapic_set_intr_ready(vlapic, vec, false))
451                         vcpu_notify_event(vlapic->vm, vlapic->vcpuid, true);
452                 break;
453         case APIC_LVT_DM_NMI:
454                 vm_inject_nmi(vlapic->vm, vlapic->vcpuid);
455                 break;
456         case APIC_LVT_DM_EXTINT:
457                 vm_inject_extint(vlapic->vm, vlapic->vcpuid);
458                 break;
459         default:
460                 // Other modes ignored
461                 return (0);
462         }
463         return (1);
464 }
465
466 #if 1
467 static void
468 dump_isrvec_stk(struct vlapic *vlapic)
469 {
470         int i;
471         uint32_t *isrptr;
472
473         isrptr = &vlapic->apic_page->isr0;
474         for (i = 0; i < 8; i++)
475                 printf("ISR%d 0x%08x\n", i, isrptr[i * 4]);
476
477         for (i = 0; i <= vlapic->isrvec_stk_top; i++)
478                 printf("isrvec_stk[%d] = %d\n", i, vlapic->isrvec_stk[i]);
479 }
480 #endif
481
482 /*
483  * Algorithm adopted from section "Interrupt, Task and Processor Priority"
484  * in Intel Architecture Manual Vol 3a.
485  */
486 static void
487 vlapic_update_ppr(struct vlapic *vlapic)
488 {
489         int isrvec, tpr, ppr;
490
491         /*
492          * Note that the value on the stack at index 0 is always 0.
493          *
494          * This is a placeholder for the value of ISRV when none of the
495          * bits is set in the ISRx registers.
496          */
497         isrvec = vlapic->isrvec_stk[vlapic->isrvec_stk_top];
498         tpr = vlapic->apic_page->tpr;
499
500 #if 1
501         {
502                 int i, lastprio, curprio, vector, idx;
503                 uint32_t *isrptr;
504
505                 if (vlapic->isrvec_stk_top == 0 && isrvec != 0)
506                         panic("isrvec_stk is corrupted: %d", isrvec);
507
508                 /*
509                  * Make sure that the priority of the nested interrupts is
510                  * always increasing.
511                  */
512                 lastprio = -1;
513                 for (i = 1; i <= vlapic->isrvec_stk_top; i++) {
514                         curprio = PRIO(vlapic->isrvec_stk[i]);
515                         if (curprio <= lastprio) {
516                                 dump_isrvec_stk(vlapic);
517                                 panic("isrvec_stk does not satisfy invariant");
518                         }
519                         lastprio = curprio;
520                 }
521
522                 /*
523                  * Make sure that each bit set in the ISRx registers has a
524                  * corresponding entry on the isrvec stack.
525                  */
526                 i = 1;
527                 isrptr = &vlapic->apic_page->isr0;
528                 for (vector = 0; vector < 256; vector++) {
529                         idx = (vector / 32) * 4;
530                         if (isrptr[idx] & (1 << (vector % 32))) {
531                                 if (i > vlapic->isrvec_stk_top ||
532                                     vlapic->isrvec_stk[i] != vector) {
533                                         dump_isrvec_stk(vlapic);
534                                         panic("ISR and isrvec_stk out of sync");
535                                 }
536                                 i++;
537                         }
538                 }
539         }
540 #endif
541
542         if (PRIO(tpr) >= PRIO(isrvec))
543                 ppr = tpr;
544         else
545                 ppr = isrvec & 0xf0;
546
547         vlapic->apic_page->ppr = ppr;
548         VLAPIC_CTR1(vlapic, "vlapic_update_ppr 0x%02x", ppr);
549 }
550
551 static void
552 vlapic_process_eoi(struct vlapic *vlapic)
553 {
554         struct LAPIC    *lapic = vlapic->apic_page;
555         uint32_t        *isrptr, *tmrptr;
556         int             i, idx, bitpos, vector;
557
558         isrptr = &lapic->isr0;
559         tmrptr = &lapic->tmr0;
560
561         /*
562          * The x86 architecture reserves the the first 32 vectors for use
563          * by the processor.
564          */
565         for (i = 7; i > 0; i--) {
566                 idx = i * 4;
567                 bitpos = fls(isrptr[idx]);
568                 if (bitpos-- != 0) {
569                         if (vlapic->isrvec_stk_top <= 0) {
570                                 panic("invalid vlapic isrvec_stk_top %d",
571                                       vlapic->isrvec_stk_top);
572                         }
573                         isrptr[idx] &= ~(1 << bitpos);
574                         VLAPIC_CTR_ISR(vlapic, "vlapic_process_eoi");
575                         vlapic->isrvec_stk_top--;
576                         vlapic_update_ppr(vlapic);
577                         if ((tmrptr[idx] & (1 << bitpos)) != 0) {
578                                 vector = i * 32 + bitpos;
579                                 vioapic_process_eoi(vlapic->vm, vlapic->vcpuid,
580                                     vector);
581                         }
582                         return;
583                 }
584         }
585 }
586
587 static __inline int
588 vlapic_get_lvt_field(uint32_t lvt, uint32_t mask)
589 {
590
591         return (lvt & mask);
592 }
593
594 static __inline int
595 vlapic_periodic_timer(struct vlapic *vlapic)
596 {
597         uint32_t lvt;
598         
599         lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_TIMER_LVT);
600
601         return (vlapic_get_lvt_field(lvt, APIC_LVTT_TM_PERIODIC));
602 }
603
604 static VMM_STAT(VLAPIC_INTR_ERROR, "error interrupts generated by vlapic");
605
606 void
607 vlapic_set_error(struct vlapic *vlapic, uint32_t mask)
608 {
609         uint32_t lvt;
610
611         vlapic->esr_pending |= mask;
612         if (vlapic->esr_firing)
613                 return;
614         vlapic->esr_firing = 1;
615
616         // The error LVT always uses the fixed delivery mode.
617         lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_ERROR_LVT);
618         if (vlapic_fire_lvt(vlapic, lvt | APIC_LVT_DM_FIXED)) {
619                 vmm_stat_incr(vlapic->vm, vlapic->vcpuid, VLAPIC_INTR_ERROR, 1);
620         }
621         vlapic->esr_firing = 0;
622 }
623
624 static VMM_STAT(VLAPIC_INTR_TIMER, "timer interrupts generated by vlapic");
625
626 static void
627 vlapic_fire_timer(struct vlapic *vlapic)
628 {
629         uint32_t lvt;
630
631         KASSERT(VLAPIC_TIMER_LOCKED(vlapic), ("vlapic_fire_timer not locked"));
632         
633         // The timer LVT always uses the fixed delivery mode.
634         lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_TIMER_LVT);
635         if (vlapic_fire_lvt(vlapic, lvt | APIC_LVT_DM_FIXED)) {
636                 vmm_stat_incr(vlapic->vm, vlapic->vcpuid, VLAPIC_INTR_TIMER, 1);
637         }
638 }
639
640 static VMM_STAT(VLAPIC_INTR_CMC,
641     "corrected machine check interrupts generated by vlapic");
642
643 void
644 vlapic_fire_cmci(struct vlapic *vlapic)
645 {
646         uint32_t lvt;
647
648         lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_CMCI_LVT);
649         if (vlapic_fire_lvt(vlapic, lvt)) {
650                 vmm_stat_incr(vlapic->vm, vlapic->vcpuid, VLAPIC_INTR_CMC, 1);
651         }
652 }
653
654 static VMM_STAT_ARRAY(LVTS_TRIGGERRED, VLAPIC_MAXLVT_INDEX + 1,
655     "lvts triggered");
656
657 int
658 vlapic_trigger_lvt(struct vlapic *vlapic, int vector)
659 {
660         uint32_t lvt;
661
662         if (vlapic_enabled(vlapic) == false) {
663                 /*
664                  * When the local APIC is global/hardware disabled,
665                  * LINT[1:0] pins are configured as INTR and NMI pins,
666                  * respectively.
667                 */
668                 switch (vector) {
669                         case APIC_LVT_LINT0:
670                                 vm_inject_extint(vlapic->vm, vlapic->vcpuid);
671                                 break;
672                         case APIC_LVT_LINT1:
673                                 vm_inject_nmi(vlapic->vm, vlapic->vcpuid);
674                                 break;
675                         default:
676                                 break;
677                 }
678                 return (0);
679         }
680
681         switch (vector) {
682         case APIC_LVT_LINT0:
683                 lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_LINT0_LVT);
684                 break;
685         case APIC_LVT_LINT1:
686                 lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_LINT1_LVT);
687                 break;
688         case APIC_LVT_TIMER:
689                 lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_TIMER_LVT);
690                 lvt |= APIC_LVT_DM_FIXED;
691                 break;
692         case APIC_LVT_ERROR:
693                 lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_ERROR_LVT);
694                 lvt |= APIC_LVT_DM_FIXED;
695                 break;
696         case APIC_LVT_PMC:
697                 lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_PERF_LVT);
698                 break;
699         case APIC_LVT_THERMAL:
700                 lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_THERM_LVT);
701                 break;
702         case APIC_LVT_CMCI:
703                 lvt = vlapic_get_lvt(vlapic, APIC_OFFSET_CMCI_LVT);
704                 break;
705         default:
706                 return (EINVAL);
707         }
708         if (vlapic_fire_lvt(vlapic, lvt)) {
709                 vmm_stat_array_incr(vlapic->vm, vlapic->vcpuid,
710                     LVTS_TRIGGERRED, vector, 1);
711         }
712         return (0);
713 }
714
715 static void
716 vlapic_callout_handler(void *arg)
717 {
718         struct vlapic *vlapic;
719         struct bintime bt, btnow;
720         sbintime_t rem_sbt;
721
722         vlapic = arg;
723
724         VLAPIC_TIMER_LOCK(vlapic);
725         if (callout_pending(&vlapic->callout))  /* callout was reset */
726                 goto done;
727
728         if (!callout_active(&vlapic->callout))  /* callout was stopped */
729                 goto done;
730
731         callout_deactivate(&vlapic->callout);
732
733         vlapic_fire_timer(vlapic);
734
735         if (vlapic_periodic_timer(vlapic)) {
736                 binuptime(&btnow);
737                 KASSERT(bintime_cmp(&btnow, &vlapic->timer_fire_bt, >=),
738                     ("vlapic callout at %#lx.%#lx, expected at %#lx.#%lx",
739                     btnow.sec, btnow.frac, vlapic->timer_fire_bt.sec,
740                     vlapic->timer_fire_bt.frac));
741
742                 /*
743                  * Compute the delta between when the timer was supposed to
744                  * fire and the present time.
745                  */
746                 bt = btnow;
747                 bintime_sub(&bt, &vlapic->timer_fire_bt);
748
749                 rem_sbt = bttosbt(vlapic->timer_period_bt);
750                 if (bintime_cmp(&bt, &vlapic->timer_period_bt, <)) {
751                         /*
752                          * Adjust the time until the next countdown downward
753                          * to account for the lost time.
754                          */
755                         rem_sbt -= bttosbt(bt);
756                 } else {
757                         /*
758                          * If the delta is greater than the timer period then
759                          * just reset our time base instead of trying to catch
760                          * up.
761                          */
762                         vlapic->timer_fire_bt = btnow;
763                         VLAPIC_CTR2(vlapic, "vlapic timer lagging by %lu "
764                             "usecs, period is %lu usecs - resetting time base",
765                             bttosbt(bt) / SBT_1US,
766                             bttosbt(vlapic->timer_period_bt) / SBT_1US);
767                 }
768
769                 bintime_add(&vlapic->timer_fire_bt, &vlapic->timer_period_bt);
770                 callout_reset_sbt(&vlapic->callout, rem_sbt, 0,
771                     vlapic_callout_handler, vlapic, 0);
772         }
773 done:
774         VLAPIC_TIMER_UNLOCK(vlapic);
775 }
776
777 void
778 vlapic_icrtmr_write_handler(struct vlapic *vlapic)
779 {
780         struct LAPIC *lapic;
781         sbintime_t sbt;
782         uint32_t icr_timer;
783
784         VLAPIC_TIMER_LOCK(vlapic);
785
786         lapic = vlapic->apic_page;
787         icr_timer = lapic->icr_timer;
788
789         vlapic->timer_period_bt = vlapic->timer_freq_bt;
790         bintime_mul(&vlapic->timer_period_bt, icr_timer);
791
792         if (icr_timer != 0) {
793                 binuptime(&vlapic->timer_fire_bt);
794                 bintime_add(&vlapic->timer_fire_bt, &vlapic->timer_period_bt);
795
796                 sbt = bttosbt(vlapic->timer_period_bt);
797                 callout_reset_sbt(&vlapic->callout, sbt, 0,
798                     vlapic_callout_handler, vlapic, 0);
799         } else
800                 callout_stop(&vlapic->callout);
801
802         VLAPIC_TIMER_UNLOCK(vlapic);
803 }
804
805 /*
806  * This function populates 'dmask' with the set of vcpus that match the
807  * addressing specified by the (dest, phys, lowprio) tuple.
808  * 
809  * 'x2apic_dest' specifies whether 'dest' is interpreted as x2APIC (32-bit)
810  * or xAPIC (8-bit) destination field.
811  */
812 static void
813 vlapic_calcdest(struct vm *vm, cpuset_t *dmask, uint32_t dest, bool phys,
814     bool lowprio, bool x2apic_dest)
815 {
816         struct vlapic *vlapic;
817         uint32_t dfr, ldr, ldest, cluster;
818         uint32_t mda_flat_ldest, mda_cluster_ldest, mda_ldest, mda_cluster_id;
819         cpuset_t amask;
820         int vcpuid;
821
822         if ((x2apic_dest && dest == 0xffffffff) ||
823             (!x2apic_dest && dest == 0xff)) {
824                 /*
825                  * Broadcast in both logical and physical modes.
826                  */
827                 *dmask = vm_active_cpus(vm);
828                 return;
829         }
830
831         if (phys) {
832                 /*
833                  * Physical mode: destination is APIC ID.
834                  */
835                 CPU_ZERO(dmask);
836                 vcpuid = vm_apicid2vcpuid(vm, dest);
837                 if (vcpuid < VM_MAXCPU)
838                         CPU_SET(vcpuid, dmask);
839         } else {
840                 /*
841                  * In the "Flat Model" the MDA is interpreted as an 8-bit wide
842                  * bitmask. This model is only avilable in the xAPIC mode.
843                  */
844                 mda_flat_ldest = dest & 0xff;
845
846                 /*
847                  * In the "Cluster Model" the MDA is used to identify a
848                  * specific cluster and a set of APICs in that cluster.
849                  */
850                 if (x2apic_dest) {
851                         mda_cluster_id = dest >> 16;
852                         mda_cluster_ldest = dest & 0xffff;
853                 } else {
854                         mda_cluster_id = (dest >> 4) & 0xf;
855                         mda_cluster_ldest = dest & 0xf;
856                 }
857
858                 /*
859                  * Logical mode: match each APIC that has a bit set
860                  * in it's LDR that matches a bit in the ldest.
861                  */
862                 CPU_ZERO(dmask);
863                 amask = vm_active_cpus(vm);
864                 while ((vcpuid = CPU_FFS(&amask)) != 0) {
865                         vcpuid--;
866                         CPU_CLR(vcpuid, &amask);
867
868                         vlapic = vm_lapic(vm, vcpuid);
869                         dfr = vlapic->apic_page->dfr;
870                         ldr = vlapic->apic_page->ldr;
871
872                         if ((dfr & APIC_DFR_MODEL_MASK) ==
873                             APIC_DFR_MODEL_FLAT) {
874                                 ldest = ldr >> 24;
875                                 mda_ldest = mda_flat_ldest;
876                         } else if ((dfr & APIC_DFR_MODEL_MASK) ==
877                             APIC_DFR_MODEL_CLUSTER) {
878                                 if (x2apic(vlapic)) {
879                                         cluster = ldr >> 16;
880                                         ldest = ldr & 0xffff;
881                                 } else {
882                                         cluster = ldr >> 28;
883                                         ldest = (ldr >> 24) & 0xf;
884                                 }
885                                 if (cluster != mda_cluster_id)
886                                         continue;
887                                 mda_ldest = mda_cluster_ldest;
888                         } else {
889                                 /*
890                                  * Guest has configured a bad logical
891                                  * model for this vcpu - skip it.
892                                  */
893                                 VLAPIC_CTR1(vlapic, "vlapic has bad logical "
894                                     "model %x - cannot deliver interrupt", dfr);
895                                 continue;
896                         }
897
898                         if ((mda_ldest & ldest) != 0) {
899                                 CPU_SET(vcpuid, dmask);
900                                 if (lowprio)
901                                         break;
902                         }
903                 }
904         }
905 }
906
907 static VMM_STAT_ARRAY(IPIS_SENT, VM_MAXCPU, "ipis sent to vcpu");
908
909 static void
910 vlapic_set_tpr(struct vlapic *vlapic, uint8_t val)
911 {
912         struct LAPIC *lapic = vlapic->apic_page;
913
914         if (lapic->tpr != val) {
915                 VCPU_CTR2(vlapic->vm, vlapic->vcpuid, "vlapic TPR changed "
916                     "from %#x to %#x", lapic->tpr, val);
917                 lapic->tpr = val;
918                 vlapic_update_ppr(vlapic);
919         }
920 }
921
922 static uint8_t
923 vlapic_get_tpr(struct vlapic *vlapic)
924 {
925         struct LAPIC *lapic = vlapic->apic_page;
926
927         return (lapic->tpr);
928 }
929
930 void
931 vlapic_set_cr8(struct vlapic *vlapic, uint64_t val)
932 {
933         uint8_t tpr;
934
935         if (val & ~0xf) {
936                 vm_inject_gp(vlapic->vm, vlapic->vcpuid);
937                 return;
938         }
939
940         tpr = val << 4;
941         vlapic_set_tpr(vlapic, tpr);
942 }
943
944 uint64_t
945 vlapic_get_cr8(struct vlapic *vlapic)
946 {
947         uint8_t tpr;
948
949         tpr = vlapic_get_tpr(vlapic);
950         return (tpr >> 4);
951 }
952
953 int
954 vlapic_icrlo_write_handler(struct vlapic *vlapic, bool *retu)
955 {
956         int i;
957         bool phys;
958         cpuset_t dmask;
959         uint64_t icrval;
960         uint32_t dest, vec, mode;
961         struct vlapic *vlapic2;
962         struct vm_exit *vmexit;
963         struct LAPIC *lapic;
964
965         lapic = vlapic->apic_page;
966         lapic->icr_lo &= ~APIC_DELSTAT_PEND;
967         icrval = ((uint64_t)lapic->icr_hi << 32) | lapic->icr_lo;
968
969         if (x2apic(vlapic))
970                 dest = icrval >> 32;
971         else
972                 dest = icrval >> (32 + 24);
973         vec = icrval & APIC_VECTOR_MASK;
974         mode = icrval & APIC_DELMODE_MASK;
975
976         if (mode == APIC_DELMODE_FIXED && vec < 16) {
977                 vlapic_set_error(vlapic, APIC_ESR_SEND_ILLEGAL_VECTOR);
978                 VLAPIC_CTR1(vlapic, "Ignoring invalid IPI %d", vec);
979                 return (0);
980         }
981
982         VLAPIC_CTR2(vlapic, "icrlo 0x%016lx triggered ipi %d", icrval, vec);
983
984         if (mode == APIC_DELMODE_FIXED || mode == APIC_DELMODE_NMI) {
985                 switch (icrval & APIC_DEST_MASK) {
986                 case APIC_DEST_DESTFLD:
987                         phys = ((icrval & APIC_DESTMODE_LOG) == 0);
988                         vlapic_calcdest(vlapic->vm, &dmask, dest, phys, false,
989                             x2apic(vlapic));
990                         break;
991                 case APIC_DEST_SELF:
992                         CPU_SETOF(vlapic->vcpuid, &dmask);
993                         break;
994                 case APIC_DEST_ALLISELF:
995                         dmask = vm_active_cpus(vlapic->vm);
996                         break;
997                 case APIC_DEST_ALLESELF:
998                         dmask = vm_active_cpus(vlapic->vm);
999                         CPU_CLR(vlapic->vcpuid, &dmask);
1000                         break;
1001                 default:
1002                         CPU_ZERO(&dmask);       /* satisfy gcc */
1003                         break;
1004                 }
1005
1006                 while ((i = CPU_FFS(&dmask)) != 0) {
1007                         i--;
1008                         CPU_CLR(i, &dmask);
1009                         if (mode == APIC_DELMODE_FIXED) {
1010                                 lapic_intr_edge(vlapic->vm, i, vec);
1011                                 vmm_stat_array_incr(vlapic->vm, vlapic->vcpuid,
1012                                                     IPIS_SENT, i, 1);
1013                                 VLAPIC_CTR2(vlapic, "vlapic sending ipi %d "
1014                                     "to vcpuid %d", vec, i);
1015                         } else {
1016                                 vm_inject_nmi(vlapic->vm, i);
1017                                 VLAPIC_CTR1(vlapic, "vlapic sending ipi nmi "
1018                                     "to vcpuid %d", i);
1019                         }
1020                 }
1021
1022                 return (0);     /* handled completely in the kernel */
1023         }
1024
1025         if (mode == APIC_DELMODE_INIT) {
1026                 if ((icrval & APIC_LEVEL_MASK) == APIC_LEVEL_DEASSERT)
1027                         return (0);
1028
1029                 if (vlapic->vcpuid == 0 && dest != 0 && dest < VM_MAXCPU) {
1030                         vlapic2 = vm_lapic(vlapic->vm, dest);
1031
1032                         /* move from INIT to waiting-for-SIPI state */
1033                         if (vlapic2->boot_state == BS_INIT) {
1034                                 vlapic2->boot_state = BS_SIPI;
1035                         }
1036
1037                         return (0);
1038                 }
1039         }
1040
1041         if (mode == APIC_DELMODE_STARTUP) {
1042                 if (vlapic->vcpuid == 0 && dest != 0 && dest < VM_MAXCPU) {
1043                         vlapic2 = vm_lapic(vlapic->vm, dest);
1044
1045                         /*
1046                          * Ignore SIPIs in any state other than wait-for-SIPI
1047                          */
1048                         if (vlapic2->boot_state != BS_SIPI)
1049                                 return (0);
1050
1051                         vlapic2->boot_state = BS_RUNNING;
1052
1053                         *retu = true;
1054                         vmexit = vm_exitinfo(vlapic->vm, vlapic->vcpuid);
1055                         vmexit->exitcode = VM_EXITCODE_SPINUP_AP;
1056                         vmexit->u.spinup_ap.vcpu = dest;
1057                         vmexit->u.spinup_ap.rip = vec << PAGE_SHIFT;
1058
1059                         return (0);
1060                 }
1061         }
1062
1063         /*
1064          * This will cause a return to userland.
1065          */
1066         return (1);
1067 }
1068
1069 void
1070 vlapic_self_ipi_handler(struct vlapic *vlapic, uint64_t val)
1071 {
1072         int vec;
1073
1074         KASSERT(x2apic(vlapic), ("SELF_IPI does not exist in xAPIC mode"));
1075
1076         vec = val & 0xff;
1077         lapic_intr_edge(vlapic->vm, vlapic->vcpuid, vec);
1078         vmm_stat_array_incr(vlapic->vm, vlapic->vcpuid, IPIS_SENT,
1079             vlapic->vcpuid, 1);
1080         VLAPIC_CTR1(vlapic, "vlapic self-ipi %d", vec);
1081 }
1082
1083 int
1084 vlapic_pending_intr(struct vlapic *vlapic, int *vecptr)
1085 {
1086         struct LAPIC    *lapic = vlapic->apic_page;
1087         int              idx, i, bitpos, vector;
1088         uint32_t        *irrptr, val;
1089
1090         if (vlapic->ops.pending_intr)
1091                 return ((*vlapic->ops.pending_intr)(vlapic, vecptr));
1092
1093         irrptr = &lapic->irr0;
1094
1095         /*
1096          * The x86 architecture reserves the the first 32 vectors for use
1097          * by the processor.
1098          */
1099         for (i = 7; i > 0; i--) {
1100                 idx = i * 4;
1101                 val = atomic_load_acq_int(&irrptr[idx]);
1102                 bitpos = fls(val);
1103                 if (bitpos != 0) {
1104                         vector = i * 32 + (bitpos - 1);
1105                         if (PRIO(vector) > PRIO(lapic->ppr)) {
1106                                 VLAPIC_CTR1(vlapic, "pending intr %d", vector);
1107                                 if (vecptr != NULL)
1108                                         *vecptr = vector;
1109                                 return (1);
1110                         } else 
1111                                 break;
1112                 }
1113         }
1114         return (0);
1115 }
1116
1117 void
1118 vlapic_intr_accepted(struct vlapic *vlapic, int vector)
1119 {
1120         struct LAPIC    *lapic = vlapic->apic_page;
1121         uint32_t        *irrptr, *isrptr;
1122         int             idx, stk_top;
1123
1124         if (vlapic->ops.intr_accepted)
1125                 return ((*vlapic->ops.intr_accepted)(vlapic, vector));
1126
1127         /*
1128          * clear the ready bit for vector being accepted in irr 
1129          * and set the vector as in service in isr.
1130          */
1131         idx = (vector / 32) * 4;
1132
1133         irrptr = &lapic->irr0;
1134         atomic_clear_int(&irrptr[idx], 1 << (vector % 32));
1135         VLAPIC_CTR_IRR(vlapic, "vlapic_intr_accepted");
1136
1137         isrptr = &lapic->isr0;
1138         isrptr[idx] |= 1 << (vector % 32);
1139         VLAPIC_CTR_ISR(vlapic, "vlapic_intr_accepted");
1140
1141         /*
1142          * Update the PPR
1143          */
1144         vlapic->isrvec_stk_top++;
1145
1146         stk_top = vlapic->isrvec_stk_top;
1147         if (stk_top >= ISRVEC_STK_SIZE)
1148                 panic("isrvec_stk_top overflow %d", stk_top);
1149
1150         vlapic->isrvec_stk[stk_top] = vector;
1151         vlapic_update_ppr(vlapic);
1152 }
1153
1154 void
1155 vlapic_svr_write_handler(struct vlapic *vlapic)
1156 {
1157         struct LAPIC *lapic;
1158         uint32_t old, new, changed;
1159
1160         lapic = vlapic->apic_page;
1161
1162         new = lapic->svr;
1163         old = vlapic->svr_last;
1164         vlapic->svr_last = new;
1165
1166         changed = old ^ new;
1167         if ((changed & APIC_SVR_ENABLE) != 0) {
1168                 if ((new & APIC_SVR_ENABLE) == 0) {
1169                         /*
1170                          * The apic is now disabled so stop the apic timer
1171                          * and mask all the LVT entries.
1172                          */
1173                         VLAPIC_CTR0(vlapic, "vlapic is software-disabled");
1174                         VLAPIC_TIMER_LOCK(vlapic);
1175                         callout_stop(&vlapic->callout);
1176                         VLAPIC_TIMER_UNLOCK(vlapic);
1177                         vlapic_mask_lvts(vlapic);
1178                 } else {
1179                         /*
1180                          * The apic is now enabled so restart the apic timer
1181                          * if it is configured in periodic mode.
1182                          */
1183                         VLAPIC_CTR0(vlapic, "vlapic is software-enabled");
1184                         if (vlapic_periodic_timer(vlapic))
1185                                 vlapic_icrtmr_write_handler(vlapic);
1186                 }
1187         }
1188 }
1189
1190 int
1191 vlapic_read(struct vlapic *vlapic, int mmio_access, uint64_t offset,
1192     uint64_t *data, bool *retu)
1193 {
1194         struct LAPIC    *lapic = vlapic->apic_page;
1195         uint32_t        *reg;
1196         int              i;
1197
1198         /* Ignore MMIO accesses in x2APIC mode */
1199         if (x2apic(vlapic) && mmio_access) {
1200                 VLAPIC_CTR1(vlapic, "MMIO read from offset %#lx in x2APIC mode",
1201                     offset);
1202                 *data = 0;
1203                 goto done;
1204         }
1205
1206         if (!x2apic(vlapic) && !mmio_access) {
1207                 /*
1208                  * XXX Generate GP fault for MSR accesses in xAPIC mode
1209                  */
1210                 VLAPIC_CTR1(vlapic, "x2APIC MSR read from offset %#lx in "
1211                     "xAPIC mode", offset);
1212                 *data = 0;
1213                 goto done;
1214         }
1215
1216         if (offset > sizeof(*lapic)) {
1217                 *data = 0;
1218                 goto done;
1219         }
1220         
1221         offset &= ~3;
1222         switch(offset)
1223         {
1224                 case APIC_OFFSET_ID:
1225                         *data = lapic->id;
1226                         break;
1227                 case APIC_OFFSET_VER:
1228                         *data = lapic->version;
1229                         break;
1230                 case APIC_OFFSET_TPR:
1231                         *data = vlapic_get_tpr(vlapic);
1232                         break;
1233                 case APIC_OFFSET_APR:
1234                         *data = lapic->apr;
1235                         break;
1236                 case APIC_OFFSET_PPR:
1237                         *data = lapic->ppr;
1238                         break;
1239                 case APIC_OFFSET_EOI:
1240                         *data = lapic->eoi;
1241                         break;
1242                 case APIC_OFFSET_LDR:
1243                         *data = lapic->ldr;
1244                         break;
1245                 case APIC_OFFSET_DFR:
1246                         *data = lapic->dfr;
1247                         break;
1248                 case APIC_OFFSET_SVR:
1249                         *data = lapic->svr;
1250                         break;
1251                 case APIC_OFFSET_ISR0 ... APIC_OFFSET_ISR7:
1252                         i = (offset - APIC_OFFSET_ISR0) >> 2;
1253                         reg = &lapic->isr0;
1254                         *data = *(reg + i);
1255                         break;
1256                 case APIC_OFFSET_TMR0 ... APIC_OFFSET_TMR7:
1257                         i = (offset - APIC_OFFSET_TMR0) >> 2;
1258                         reg = &lapic->tmr0;
1259                         *data = *(reg + i);
1260                         break;
1261                 case APIC_OFFSET_IRR0 ... APIC_OFFSET_IRR7:
1262                         i = (offset - APIC_OFFSET_IRR0) >> 2;
1263                         reg = &lapic->irr0;
1264                         *data = atomic_load_acq_int(reg + i);
1265                         break;
1266                 case APIC_OFFSET_ESR:
1267                         *data = lapic->esr;
1268                         break;
1269                 case APIC_OFFSET_ICR_LOW: 
1270                         *data = lapic->icr_lo;
1271                         if (x2apic(vlapic))
1272                                 *data |= (uint64_t)lapic->icr_hi << 32;
1273                         break;
1274                 case APIC_OFFSET_ICR_HI: 
1275                         *data = lapic->icr_hi;
1276                         break;
1277                 case APIC_OFFSET_CMCI_LVT:
1278                 case APIC_OFFSET_TIMER_LVT ... APIC_OFFSET_ERROR_LVT:
1279                         *data = vlapic_get_lvt(vlapic, offset); 
1280 #ifdef INVARIANTS
1281                         reg = vlapic_get_lvtptr(vlapic, offset);
1282                         KASSERT(*data == *reg, ("inconsistent lvt value at "
1283                             "offset %#lx: %#lx/%#x", offset, *data, *reg));
1284 #endif
1285                         break;
1286                 case APIC_OFFSET_TIMER_ICR:
1287                         *data = lapic->icr_timer;
1288                         break;
1289                 case APIC_OFFSET_TIMER_CCR:
1290                         *data = vlapic_get_ccr(vlapic);
1291                         break;
1292                 case APIC_OFFSET_TIMER_DCR:
1293                         *data = lapic->dcr_timer;
1294                         break;
1295                 case APIC_OFFSET_SELF_IPI:
1296                         /*
1297                          * XXX generate a GP fault if vlapic is in x2apic mode
1298                          */
1299                         *data = 0;
1300                         break;
1301                 case APIC_OFFSET_RRR:
1302                 default:
1303                         *data = 0;
1304                         break;
1305         }
1306 done:
1307         VLAPIC_CTR2(vlapic, "vlapic read offset %#x, data %#lx", offset, *data);
1308         return 0;
1309 }
1310
1311 int
1312 vlapic_write(struct vlapic *vlapic, int mmio_access, uint64_t offset,
1313     uint64_t data, bool *retu)
1314 {
1315         struct LAPIC    *lapic = vlapic->apic_page;
1316         uint32_t        *regptr;
1317         int             retval;
1318
1319         KASSERT((offset & 0xf) == 0 && offset < PAGE_SIZE,
1320             ("vlapic_write: invalid offset %#lx", offset));
1321
1322         VLAPIC_CTR2(vlapic, "vlapic write offset %#lx, data %#lx",
1323             offset, data);
1324
1325         if (offset > sizeof(*lapic))
1326                 return (0);
1327
1328         /* Ignore MMIO accesses in x2APIC mode */
1329         if (x2apic(vlapic) && mmio_access) {
1330                 VLAPIC_CTR2(vlapic, "MMIO write of %#lx to offset %#lx "
1331                     "in x2APIC mode", data, offset);
1332                 return (0);
1333         }
1334
1335         /*
1336          * XXX Generate GP fault for MSR accesses in xAPIC mode
1337          */
1338         if (!x2apic(vlapic) && !mmio_access) {
1339                 VLAPIC_CTR2(vlapic, "x2APIC MSR write of %#lx to offset %#lx "
1340                     "in xAPIC mode", data, offset);
1341                 return (0);
1342         }
1343
1344         retval = 0;
1345         switch(offset)
1346         {
1347                 case APIC_OFFSET_ID:
1348                         lapic->id = data;
1349                         vlapic_id_write_handler(vlapic);
1350                         break;
1351                 case APIC_OFFSET_TPR:
1352                         vlapic_set_tpr(vlapic, data & 0xff);
1353                         break;
1354                 case APIC_OFFSET_EOI:
1355                         vlapic_process_eoi(vlapic);
1356                         break;
1357                 case APIC_OFFSET_LDR:
1358                         lapic->ldr = data;
1359                         vlapic_ldr_write_handler(vlapic);
1360                         break;
1361                 case APIC_OFFSET_DFR:
1362                         lapic->dfr = data;
1363                         vlapic_dfr_write_handler(vlapic);
1364                         break;
1365                 case APIC_OFFSET_SVR:
1366                         lapic->svr = data;
1367                         vlapic_svr_write_handler(vlapic);
1368                         break;
1369                 case APIC_OFFSET_ICR_LOW: 
1370                         lapic->icr_lo = data;
1371                         if (x2apic(vlapic))
1372                                 lapic->icr_hi = data >> 32;
1373                         retval = vlapic_icrlo_write_handler(vlapic, retu);
1374                         break;
1375                 case APIC_OFFSET_ICR_HI:
1376                         lapic->icr_hi = data;
1377                         break;
1378                 case APIC_OFFSET_CMCI_LVT:
1379                 case APIC_OFFSET_TIMER_LVT ... APIC_OFFSET_ERROR_LVT:
1380                         regptr = vlapic_get_lvtptr(vlapic, offset);
1381                         *regptr = data;
1382                         vlapic_lvt_write_handler(vlapic, offset);
1383                         break;
1384                 case APIC_OFFSET_TIMER_ICR:
1385                         lapic->icr_timer = data;
1386                         vlapic_icrtmr_write_handler(vlapic);
1387                         break;
1388
1389                 case APIC_OFFSET_TIMER_DCR:
1390                         lapic->dcr_timer = data;
1391                         vlapic_dcr_write_handler(vlapic);
1392                         break;
1393
1394                 case APIC_OFFSET_ESR:
1395                         vlapic_esr_write_handler(vlapic);
1396                         break;
1397
1398                 case APIC_OFFSET_SELF_IPI:
1399                         if (x2apic(vlapic))
1400                                 vlapic_self_ipi_handler(vlapic, data);
1401                         break;
1402
1403                 case APIC_OFFSET_VER:
1404                 case APIC_OFFSET_APR:
1405                 case APIC_OFFSET_PPR:
1406                 case APIC_OFFSET_RRR:
1407                 case APIC_OFFSET_ISR0 ... APIC_OFFSET_ISR7:
1408                 case APIC_OFFSET_TMR0 ... APIC_OFFSET_TMR7:
1409                 case APIC_OFFSET_IRR0 ... APIC_OFFSET_IRR7:
1410                 case APIC_OFFSET_TIMER_CCR:
1411                 default:
1412                         // Read only.
1413                         break;
1414         }
1415
1416         return (retval);
1417 }
1418
1419 static void
1420 vlapic_reset(struct vlapic *vlapic)
1421 {
1422         struct LAPIC *lapic;
1423         
1424         lapic = vlapic->apic_page;
1425         bzero(lapic, sizeof(struct LAPIC));
1426
1427         lapic->id = vlapic_get_id(vlapic);
1428         lapic->version = VLAPIC_VERSION;
1429         lapic->version |= (VLAPIC_MAXLVT_INDEX << MAXLVTSHIFT);
1430         lapic->dfr = 0xffffffff;
1431         lapic->svr = APIC_SVR_VECTOR;
1432         vlapic_mask_lvts(vlapic);
1433         vlapic_reset_tmr(vlapic);
1434
1435         lapic->dcr_timer = 0;
1436         vlapic_dcr_write_handler(vlapic);
1437
1438         if (vlapic->vcpuid == 0)
1439                 vlapic->boot_state = BS_RUNNING;        /* BSP */
1440         else
1441                 vlapic->boot_state = BS_INIT;           /* AP */
1442
1443         vlapic->svr_last = lapic->svr;
1444 }
1445
1446 void
1447 vlapic_init(struct vlapic *vlapic)
1448 {
1449         KASSERT(vlapic->vm != NULL, ("vlapic_init: vm is not initialized"));
1450         KASSERT(vlapic->vcpuid >= 0 && vlapic->vcpuid < VM_MAXCPU,
1451             ("vlapic_init: vcpuid is not initialized"));
1452         KASSERT(vlapic->apic_page != NULL, ("vlapic_init: apic_page is not "
1453             "initialized"));
1454
1455         /*
1456          * If the vlapic is configured in x2apic mode then it will be
1457          * accessed in the critical section via the MSR emulation code.
1458          *
1459          * Therefore the timer mutex must be a spinlock because blockable
1460          * mutexes cannot be acquired in a critical section.
1461          */
1462         mtx_init(&vlapic->timer_mtx, "vlapic timer mtx", NULL, MTX_SPIN);
1463         callout_init(&vlapic->callout, 1);
1464
1465         vlapic->msr_apicbase = DEFAULT_APIC_BASE | APICBASE_ENABLED;
1466
1467         if (vlapic->vcpuid == 0)
1468                 vlapic->msr_apicbase |= APICBASE_BSP;
1469
1470         vlapic_reset(vlapic);
1471 }
1472
1473 void
1474 vlapic_cleanup(struct vlapic *vlapic)
1475 {
1476
1477         callout_drain(&vlapic->callout);
1478 }
1479
1480 uint64_t
1481 vlapic_get_apicbase(struct vlapic *vlapic)
1482 {
1483
1484         return (vlapic->msr_apicbase);
1485 }
1486
1487 int
1488 vlapic_set_apicbase(struct vlapic *vlapic, uint64_t new)
1489 {
1490
1491         if (vlapic->msr_apicbase != new) {
1492                 VLAPIC_CTR2(vlapic, "Changing APIC_BASE MSR from %#lx to %#lx "
1493                     "not supported", vlapic->msr_apicbase, new);
1494                 return (-1);
1495         }
1496
1497         return (0);
1498 }
1499
1500 void
1501 vlapic_set_x2apic_state(struct vm *vm, int vcpuid, enum x2apic_state state)
1502 {
1503         struct vlapic *vlapic;
1504         struct LAPIC *lapic;
1505
1506         vlapic = vm_lapic(vm, vcpuid);
1507
1508         if (state == X2APIC_DISABLED)
1509                 vlapic->msr_apicbase &= ~APICBASE_X2APIC;
1510         else
1511                 vlapic->msr_apicbase |= APICBASE_X2APIC;
1512
1513         /*
1514          * Reset the local APIC registers whose values are mode-dependent.
1515          *
1516          * XXX this works because the APIC mode can be changed only at vcpu
1517          * initialization time.
1518          */
1519         lapic = vlapic->apic_page;
1520         lapic->id = vlapic_get_id(vlapic);
1521         if (x2apic(vlapic)) {
1522                 lapic->ldr = x2apic_ldr(vlapic);
1523                 lapic->dfr = 0;
1524         } else {
1525                 lapic->ldr = 0;
1526                 lapic->dfr = 0xffffffff;
1527         }
1528
1529         if (state == X2APIC_ENABLED) {
1530                 if (vlapic->ops.enable_x2apic_mode)
1531                         (*vlapic->ops.enable_x2apic_mode)(vlapic);
1532         }
1533 }
1534
1535 void
1536 vlapic_deliver_intr(struct vm *vm, bool level, uint32_t dest, bool phys,
1537     int delmode, int vec)
1538 {
1539         bool lowprio;
1540         int vcpuid;
1541         cpuset_t dmask;
1542
1543         if (delmode != IOART_DELFIXED &&
1544             delmode != IOART_DELLOPRI &&
1545             delmode != IOART_DELEXINT) {
1546                 VM_CTR1(vm, "vlapic intr invalid delmode %#x", delmode);
1547                 return;
1548         }
1549         lowprio = (delmode == IOART_DELLOPRI);
1550
1551         /*
1552          * We don't provide any virtual interrupt redirection hardware so
1553          * all interrupts originating from the ioapic or MSI specify the
1554          * 'dest' in the legacy xAPIC format.
1555          */
1556         vlapic_calcdest(vm, &dmask, dest, phys, lowprio, false);
1557
1558         while ((vcpuid = CPU_FFS(&dmask)) != 0) {
1559                 vcpuid--;
1560                 CPU_CLR(vcpuid, &dmask);
1561                 if (delmode == IOART_DELEXINT) {
1562                         vm_inject_extint(vm, vcpuid);
1563                 } else {
1564                         lapic_set_intr(vm, vcpuid, vec, level);
1565                 }
1566         }
1567 }
1568
1569 void
1570 vlapic_post_intr(struct vlapic *vlapic, int hostcpu, int ipinum)
1571 {
1572         /*
1573          * Post an interrupt to the vcpu currently running on 'hostcpu'.
1574          *
1575          * This is done by leveraging features like Posted Interrupts (Intel)
1576          * Doorbell MSR (AMD AVIC) that avoid a VM exit.
1577          *
1578          * If neither of these features are available then fallback to
1579          * sending an IPI to 'hostcpu'.
1580          */
1581         if (vlapic->ops.post_intr)
1582                 (*vlapic->ops.post_intr)(vlapic, hostcpu);
1583         else
1584                 ipi_cpu(hostcpu, ipinum);
1585 }
1586
1587 bool
1588 vlapic_enabled(struct vlapic *vlapic)
1589 {
1590         struct LAPIC *lapic = vlapic->apic_page;
1591
1592         if ((vlapic->msr_apicbase & APICBASE_ENABLED) != 0 &&
1593             (lapic->svr & APIC_SVR_ENABLE) != 0)
1594                 return (true);
1595         else
1596                 return (false);
1597 }
1598
1599 static void
1600 vlapic_set_tmr(struct vlapic *vlapic, int vector, bool level)
1601 {
1602         struct LAPIC *lapic;
1603         uint32_t *tmrptr, mask;
1604         int idx;
1605
1606         lapic = vlapic->apic_page;
1607         tmrptr = &lapic->tmr0;
1608         idx = (vector / 32) * 4;
1609         mask = 1 << (vector % 32);
1610         if (level)
1611                 tmrptr[idx] |= mask;
1612         else
1613                 tmrptr[idx] &= ~mask;
1614
1615         if (vlapic->ops.set_tmr != NULL)
1616                 (*vlapic->ops.set_tmr)(vlapic, vector, level);
1617 }
1618
1619 void
1620 vlapic_reset_tmr(struct vlapic *vlapic)
1621 {
1622         int vector;
1623
1624         VLAPIC_CTR0(vlapic, "vlapic resetting all vectors to edge-triggered");
1625
1626         for (vector = 0; vector <= 255; vector++)
1627                 vlapic_set_tmr(vlapic, vector, false);
1628 }
1629
1630 void
1631 vlapic_set_tmr_level(struct vlapic *vlapic, uint32_t dest, bool phys,
1632     int delmode, int vector)
1633 {
1634         cpuset_t dmask;
1635         bool lowprio;
1636
1637         KASSERT(vector >= 0 && vector <= 255, ("invalid vector %d", vector));
1638
1639         /*
1640          * A level trigger is valid only for fixed and lowprio delivery modes.
1641          */
1642         if (delmode != APIC_DELMODE_FIXED && delmode != APIC_DELMODE_LOWPRIO) {
1643                 VLAPIC_CTR1(vlapic, "Ignoring level trigger-mode for "
1644                     "delivery-mode %d", delmode);
1645                 return;
1646         }
1647
1648         lowprio = (delmode == APIC_DELMODE_LOWPRIO);
1649         vlapic_calcdest(vlapic->vm, &dmask, dest, phys, lowprio, false);
1650
1651         if (!CPU_ISSET(vlapic->vcpuid, &dmask))
1652                 return;
1653
1654         VLAPIC_CTR1(vlapic, "vector %d set to level-triggered", vector);
1655         vlapic_set_tmr(vlapic, vector, true);
1656 }