SAFE public projects git trees. - safe/jmp/linux-2.6/blob - arch/x86/kvm/lapic.c

   1
   2 /*
   3  * Local APIC virtualization
   4  *
   5  * Copyright (C) 2006 Qumranet, Inc.
   6  * Copyright (C) 2007 Novell
   7  * Copyright (C) 2007 Intel
   8  *
   9  * Authors:
  10  *   Dor Laor <dor.laor@qumranet.com>
  11  *   Gregory Haskins <ghaskins@novell.com>
  12  *   Yaozu (Eddie) Dong <eddie.dong@intel.com>
  13  *
  14  * Based on Xen 3.1 code, Copyright (c) 2004, Intel Corporation.
  15  *
  16  * This work is licensed under the terms of the GNU GPL, version 2.  See
  17  * the COPYING file in the top-level directory.
  18  */
  19
  20 #include <linux/kvm_host.h>
  21 #include <linux/kvm.h>
  22 #include <linux/mm.h>
  23 #include <linux/highmem.h>
  24 #include <linux/smp.h>
  25 #include <linux/hrtimer.h>
  26 #include <linux/io.h>
  27 #include <linux/module.h>
  28 #include <linux/math64.h>
  29 #include <asm/processor.h>
  30 #include <asm/msr.h>
  31 #include <asm/page.h>
  32 #include <asm/current.h>
  33 #include <asm/apicdef.h>
  34 #include <asm/atomic.h>
  35 #include "kvm_cache_regs.h"
  36 #include "irq.h"
  37
  38 #ifndef CONFIG_X86_64
  39 #define mod_64(x, y) ((x) - (y) * div64_u64(x, y))
  40 #else
  41 #define mod_64(x, y) ((x) % (y))
  42 #endif
  43
  44 #define PRId64 "d"
  45 #define PRIx64 "llx"
  46 #define PRIu64 "u"
  47 #define PRIo64 "o"
  48
  49 #define APIC_BUS_CYCLE_NS 1
  50
  51 /* #define apic_debug(fmt,arg...) printk(KERN_WARNING fmt,##arg) */
  52 #define apic_debug(fmt, arg...)
  53
  54 #define APIC_LVT_NUM                    6
  55 /* 14 is the version for Xeon and Pentium 8.4.8*/
  56 #define APIC_VERSION                    (0x14UL | ((APIC_LVT_NUM - 1) << 16))
  57 #define LAPIC_MMIO_LENGTH               (1 << 12)
  58 /* followed define is not in apicdef.h */
  59 #define APIC_SHORT_MASK                 0xc0000
  60 #define APIC_DEST_NOSHORT               0x0
  61 #define APIC_DEST_MASK                  0x800
  62 #define MAX_APIC_VECTOR                 256
  63
  64 #define VEC_POS(v) ((v) & (32 - 1))
  65 #define REG_POS(v) (((v) >> 5) << 4)
  66
  67 static inline u32 apic_get_reg(struct kvm_lapic *apic, int reg_off)
  68 {
  69         return *((u32 *) (apic->regs + reg_off));
  70 }
  71
  72 static inline void apic_set_reg(struct kvm_lapic *apic, int reg_off, u32 val)
  73 {
  74         *((u32 *) (apic->regs + reg_off)) = val;
  75 }
  76
  77 static inline int apic_test_and_set_vector(int vec, void *bitmap)
  78 {
  79         return test_and_set_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
  80 }
  81
  82 static inline int apic_test_and_clear_vector(int vec, void *bitmap)
  83 {
  84         return test_and_clear_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
  85 }
  86
  87 static inline void apic_set_vector(int vec, void *bitmap)
  88 {
  89         set_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
  90 }
  91
  92 static inline void apic_clear_vector(int vec, void *bitmap)
  93 {
  94         clear_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
  95 }
  96
  97 static inline int apic_hw_enabled(struct kvm_lapic *apic)
  98 {
  99         return (apic)->vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE;
 100 }
 101
 102 static inline int  apic_sw_enabled(struct kvm_lapic *apic)
 103 {
 104         return apic_get_reg(apic, APIC_SPIV) & APIC_SPIV_APIC_ENABLED;
 105 }
 106
 107 static inline int apic_enabled(struct kvm_lapic *apic)
 108 {
 109         return apic_sw_enabled(apic) && apic_hw_enabled(apic);
 110 }
 111
 112 #define LVT_MASK        \
 113         (APIC_LVT_MASKED | APIC_SEND_PENDING | APIC_VECTOR_MASK)
 114
 115 #define LINT_MASK       \
 116         (LVT_MASK | APIC_MODE_MASK | APIC_INPUT_POLARITY | \
 117          APIC_LVT_REMOTE_IRR | APIC_LVT_LEVEL_TRIGGER)
 118
 119 static inline int kvm_apic_id(struct kvm_lapic *apic)
 120 {
 121         return (apic_get_reg(apic, APIC_ID) >> 24) & 0xff;
 122 }
 123
 124 static inline int apic_lvt_enabled(struct kvm_lapic *apic, int lvt_type)
 125 {
 126         return !(apic_get_reg(apic, lvt_type) & APIC_LVT_MASKED);
 127 }
 128
 129 static inline int apic_lvt_vector(struct kvm_lapic *apic, int lvt_type)
 130 {
 131         return apic_get_reg(apic, lvt_type) & APIC_VECTOR_MASK;
 132 }
 133
 134 static inline int apic_lvtt_period(struct kvm_lapic *apic)
 135 {
 136         return apic_get_reg(apic, APIC_LVTT) & APIC_LVT_TIMER_PERIODIC;
 137 }
 138
 139 static inline int apic_lvt_nmi_mode(u32 lvt_val)
 140 {
 141         return (lvt_val & (APIC_MODE_MASK | APIC_LVT_MASKED)) == APIC_DM_NMI;
 142 }
 143
 144 static unsigned int apic_lvt_mask[APIC_LVT_NUM] = {
 145         LVT_MASK | APIC_LVT_TIMER_PERIODIC,     /* LVTT */
 146         LVT_MASK | APIC_MODE_MASK,      /* LVTTHMR */
 147         LVT_MASK | APIC_MODE_MASK,      /* LVTPC */
 148         LINT_MASK, LINT_MASK,   /* LVT0-1 */
 149         LVT_MASK                /* LVTERR */
 150 };
 151
 152 static int find_highest_vector(void *bitmap)
 153 {
 154         u32 *word = bitmap;
 155         int word_offset = MAX_APIC_VECTOR >> 5;
 156
 157         while ((word_offset != 0) && (word[(--word_offset) << 2] == 0))
 158                 continue;
 159
 160         if (likely(!word_offset && !word[0]))
 161                 return -1;
 162         else
 163                 return fls(word[word_offset << 2]) - 1 + (word_offset << 5);
 164 }
 165
 166 static inline int apic_test_and_set_irr(int vec, struct kvm_lapic *apic)
 167 {
 168         return apic_test_and_set_vector(vec, apic->regs + APIC_IRR);
 169 }
 170
 171 static inline void apic_clear_irr(int vec, struct kvm_lapic *apic)
 172 {
 173         apic_clear_vector(vec, apic->regs + APIC_IRR);
 174 }
 175
 176 static inline int apic_find_highest_irr(struct kvm_lapic *apic)
 177 {
 178         int result;
 179
 180         result = find_highest_vector(apic->regs + APIC_IRR);
 181         ASSERT(result == -1 || result >= 16);
 182
 183         return result;
 184 }
 185
 186 int kvm_lapic_find_highest_irr(struct kvm_vcpu *vcpu)
 187 {
 188         struct kvm_lapic *apic = vcpu->arch.apic;
 189         int highest_irr;
 190
 191         if (!apic)
 192                 return 0;
 193         highest_irr = apic_find_highest_irr(apic);
 194
 195         return highest_irr;
 196 }
 197 EXPORT_SYMBOL_GPL(kvm_lapic_find_highest_irr);
 198
 199 int kvm_apic_set_irq(struct kvm_vcpu *vcpu, u8 vec, u8 trig)
 200 {
 201         struct kvm_lapic *apic = vcpu->arch.apic;
 202
 203         if (!apic_test_and_set_irr(vec, apic)) {
 204                 /* a new pending irq is set in IRR */
 205                 if (trig)
 206                         apic_set_vector(vec, apic->regs + APIC_TMR);
 207                 else
 208                         apic_clear_vector(vec, apic->regs + APIC_TMR);
 209                 kvm_vcpu_kick(apic->vcpu);
 210                 return 1;
 211         }
 212         return 0;
 213 }
 214
 215 static inline int apic_find_highest_isr(struct kvm_lapic *apic)
 216 {
 217         int result;
 218
 219         result = find_highest_vector(apic->regs + APIC_ISR);
 220         ASSERT(result == -1 || result >= 16);
 221
 222         return result;
 223 }
 224
 225 static void apic_update_ppr(struct kvm_lapic *apic)
 226 {
 227         u32 tpr, isrv, ppr;
 228         int isr;
 229
 230         tpr = apic_get_reg(apic, APIC_TASKPRI);
 231         isr = apic_find_highest_isr(apic);
 232         isrv = (isr != -1) ? isr : 0;
 233
 234         if ((tpr & 0xf0) >= (isrv & 0xf0))
 235                 ppr = tpr & 0xff;
 236         else
 237                 ppr = isrv & 0xf0;
 238
 239         apic_debug("vlapic %p, ppr 0x%x, isr 0x%x, isrv 0x%x",
 240                    apic, ppr, isr, isrv);
 241
 242         apic_set_reg(apic, APIC_PROCPRI, ppr);
 243 }
 244
 245 static void apic_set_tpr(struct kvm_lapic *apic, u32 tpr)
 246 {
 247         apic_set_reg(apic, APIC_TASKPRI, tpr);
 248         apic_update_ppr(apic);
 249 }
 250
 251 int kvm_apic_match_physical_addr(struct kvm_lapic *apic, u16 dest)
 252 {
 253         return kvm_apic_id(apic) == dest;
 254 }
 255
 256 int kvm_apic_match_logical_addr(struct kvm_lapic *apic, u8 mda)
 257 {
 258         int result = 0;
 259         u8 logical_id;
 260
 261         logical_id = GET_APIC_LOGICAL_ID(apic_get_reg(apic, APIC_LDR));
 262
 263         switch (apic_get_reg(apic, APIC_DFR)) {
 264         case APIC_DFR_FLAT:
 265                 if (logical_id & mda)
 266                         result = 1;
 267                 break;
 268         case APIC_DFR_CLUSTER:
 269                 if (((logical_id >> 4) == (mda >> 0x4))
 270                     && (logical_id & mda & 0xf))
 271                         result = 1;
 272                 break;
 273         default:
 274                 printk(KERN_WARNING "Bad DFR vcpu %d: %08x\n",
 275                        apic->vcpu->vcpu_id, apic_get_reg(apic, APIC_DFR));
 276                 break;
 277         }
 278
 279         return result;
 280 }
 281
 282 static int apic_match_dest(struct kvm_vcpu *vcpu, struct kvm_lapic *source,
 283                            int short_hand, int dest, int dest_mode)
 284 {
 285         int result = 0;
 286         struct kvm_lapic *target = vcpu->arch.apic;
 287
 288         apic_debug("target %p, source %p, dest 0x%x, "
 289                    "dest_mode 0x%x, short_hand 0x%x",
 290                    target, source, dest, dest_mode, short_hand);
 291
 292         ASSERT(!target);
 293         switch (short_hand) {
 294         case APIC_DEST_NOSHORT:
 295                 if (dest_mode == 0) {
 296                         /* Physical mode. */
 297                         if ((dest == 0xFF) || (dest == kvm_apic_id(target)))
 298                                 result = 1;
 299                 } else
 300                         /* Logical mode. */
 301                         result = kvm_apic_match_logical_addr(target, dest);
 302                 break;
 303         case APIC_DEST_SELF:
 304                 if (target == source)
 305                         result = 1;
 306                 break;
 307         case APIC_DEST_ALLINC:
 308                 result = 1;
 309                 break;
 310         case APIC_DEST_ALLBUT:
 311                 if (target != source)
 312                         result = 1;
 313                 break;
 314         default:
 315                 printk(KERN_WARNING "Bad dest shorthand value %x\n",
 316                        short_hand);
 317                 break;
 318         }
 319
 320         return result;
 321 }
 322
 323 /*
 324  * Add a pending IRQ into lapic.
 325  * Return 1 if successfully added and 0 if discarded.
 326  */
 327 static int __apic_accept_irq(struct kvm_lapic *apic, int delivery_mode,
 328                              int vector, int level, int trig_mode)
 329 {
 330         int orig_irr, result = 0;
 331         struct kvm_vcpu *vcpu = apic->vcpu;
 332
 333         switch (delivery_mode) {
 334         case APIC_DM_FIXED:
 335         case APIC_DM_LOWEST:
 336                 /* FIXME add logic for vcpu on reset */
 337                 if (unlikely(!apic_enabled(apic)))
 338                         break;
 339
 340                 orig_irr = apic_test_and_set_irr(vector, apic);
 341                 if (orig_irr && trig_mode) {
 342                         apic_debug("level trig mode repeatedly for vector %d",
 343                                    vector);
 344                         break;
 345                 }
 346
 347                 if (trig_mode) {
 348                         apic_debug("level trig mode for vector %d", vector);
 349                         apic_set_vector(vector, apic->regs + APIC_TMR);
 350                 } else
 351                         apic_clear_vector(vector, apic->regs + APIC_TMR);
 352
 353                 kvm_vcpu_kick(vcpu);
 354
 355                 result = (orig_irr == 0);
 356                 break;
 357
 358         case APIC_DM_REMRD:
 359                 printk(KERN_DEBUG "Ignoring delivery mode 3\n");
 360                 break;
 361
 362         case APIC_DM_SMI:
 363                 printk(KERN_DEBUG "Ignoring guest SMI\n");
 364                 break;
 365
 366         case APIC_DM_NMI:
 367                 kvm_inject_nmi(vcpu);
 368                 kvm_vcpu_kick(vcpu);
 369                 break;
 370
 371         case APIC_DM_INIT:
 372                 if (level) {
 373                         if (vcpu->arch.mp_state == KVM_MP_STATE_RUNNABLE)
 374                                 printk(KERN_DEBUG
 375                                        "INIT on a runnable vcpu %d\n",
 376                                        vcpu->vcpu_id);
 377                         vcpu->arch.mp_state = KVM_MP_STATE_INIT_RECEIVED;
 378                         kvm_vcpu_kick(vcpu);
 379                 } else {
 380                         apic_debug("Ignoring de-assert INIT to vcpu %d\n",
 381                                    vcpu->vcpu_id);
 382                 }
 383                 break;
 384
 385         case APIC_DM_STARTUP:
 386                 apic_debug("SIPI to vcpu %d vector 0x%02x\n",
 387                            vcpu->vcpu_id, vector);
 388                 if (vcpu->arch.mp_state == KVM_MP_STATE_INIT_RECEIVED) {
 389                         vcpu->arch.sipi_vector = vector;
 390                         vcpu->arch.mp_state = KVM_MP_STATE_SIPI_RECEIVED;
 391                         kvm_vcpu_kick(vcpu);
 392                 }
 393                 break;
 394
 395         case APIC_DM_EXTINT:
 396                 /*
 397                  * Should only be called by kvm_apic_local_deliver() with LVT0,
 398                  * before NMI watchdog was enabled. Already handled by
 399                  * kvm_apic_accept_pic_intr().
 400                  */
 401                 break;
 402
 403         default:
 404                 printk(KERN_ERR "TODO: unsupported delivery mode %x\n",
 405                        delivery_mode);
 406                 break;
 407         }
 408         return result;
 409 }
 410
 411 static struct kvm_lapic *kvm_apic_round_robin(struct kvm *kvm, u8 vector,
 412                                        unsigned long *bitmap)
 413 {
 414         int last;
 415         int next;
 416         struct kvm_lapic *apic = NULL;
 417
 418         last = kvm->arch.round_robin_prev_vcpu;
 419         next = last;
 420
 421         do {
 422                 if (++next == KVM_MAX_VCPUS)
 423                         next = 0;
 424                 if (kvm->vcpus[next] == NULL || !test_bit(next, bitmap))
 425                         continue;
 426                 apic = kvm->vcpus[next]->arch.apic;
 427                 if (apic && apic_enabled(apic))
 428                         break;
 429                 apic = NULL;
 430         } while (next != last);
 431         kvm->arch.round_robin_prev_vcpu = next;
 432
 433         if (!apic)
 434                 printk(KERN_DEBUG "vcpu not ready for apic_round_robin\n");
 435
 436         return apic;
 437 }
 438
 439 struct kvm_vcpu *kvm_get_lowest_prio_vcpu(struct kvm *kvm, u8 vector,
 440                 unsigned long *bitmap)
 441 {
 442         struct kvm_lapic *apic;
 443
 444         apic = kvm_apic_round_robin(kvm, vector, bitmap);
 445         if (apic)
 446                 return apic->vcpu;
 447         return NULL;
 448 }
 449
 450 static void apic_set_eoi(struct kvm_lapic *apic)
 451 {
 452         int vector = apic_find_highest_isr(apic);
 453         int trigger_mode;
 454         /*
 455          * Not every write EOI will has corresponding ISR,
 456          * one example is when Kernel check timer on setup_IO_APIC
 457          */
 458         if (vector == -1)
 459                 return;
 460
 461         apic_clear_vector(vector, apic->regs + APIC_ISR);
 462         apic_update_ppr(apic);
 463
 464         if (apic_test_and_clear_vector(vector, apic->regs + APIC_TMR))
 465                 trigger_mode = IOAPIC_LEVEL_TRIG;
 466         else
 467                 trigger_mode = IOAPIC_EDGE_TRIG;
 468         kvm_ioapic_update_eoi(apic->vcpu->kvm, vector, trigger_mode);
 469 }
 470
 471 static void apic_send_ipi(struct kvm_lapic *apic)
 472 {
 473         u32 icr_low = apic_get_reg(apic, APIC_ICR);
 474         u32 icr_high = apic_get_reg(apic, APIC_ICR2);
 475
 476         unsigned int dest = GET_APIC_DEST_FIELD(icr_high);
 477         unsigned int short_hand = icr_low & APIC_SHORT_MASK;
 478         unsigned int trig_mode = icr_low & APIC_INT_LEVELTRIG;
 479         unsigned int level = icr_low & APIC_INT_ASSERT;
 480         unsigned int dest_mode = icr_low & APIC_DEST_MASK;
 481         unsigned int delivery_mode = icr_low & APIC_MODE_MASK;
 482         unsigned int vector = icr_low & APIC_VECTOR_MASK;
 483
 484         struct kvm_vcpu *target;
 485         struct kvm_vcpu *vcpu;
 486         DECLARE_BITMAP(lpr_map, KVM_MAX_VCPUS);
 487         int i;
 488
 489         bitmap_zero(lpr_map, KVM_MAX_VCPUS);
 490         apic_debug("icr_high 0x%x, icr_low 0x%x, "
 491                    "short_hand 0x%x, dest 0x%x, trig_mode 0x%x, level 0x%x, "
 492                    "dest_mode 0x%x, delivery_mode 0x%x, vector 0x%x\n",
 493                    icr_high, icr_low, short_hand, dest,
 494                    trig_mode, level, dest_mode, delivery_mode, vector);
 495
 496         for (i = 0; i < KVM_MAX_VCPUS; i++) {
 497                 vcpu = apic->vcpu->kvm->vcpus[i];
 498                 if (!vcpu)
 499                         continue;
 500
 501                 if (vcpu->arch.apic &&
 502                     apic_match_dest(vcpu, apic, short_hand, dest, dest_mode)) {
 503                         if (delivery_mode == APIC_DM_LOWEST)
 504                                 __set_bit(vcpu->vcpu_id, lpr_map);
 505                         else
 506                                 __apic_accept_irq(vcpu->arch.apic, delivery_mode,
 507                                                   vector, level, trig_mode);
 508                 }
 509         }
 510
 511         if (delivery_mode == APIC_DM_LOWEST) {
 512                 target = kvm_get_lowest_prio_vcpu(vcpu->kvm, vector, lpr_map);
 513                 if (target != NULL)
 514                         __apic_accept_irq(target->arch.apic, delivery_mode,
 515                                           vector, level, trig_mode);
 516         }
 517 }
 518
 519 static u32 apic_get_tmcct(struct kvm_lapic *apic)
 520 {
 521         ktime_t remaining;
 522         s64 ns;
 523         u32 tmcct;
 524
 525         ASSERT(apic != NULL);
 526
 527         /* if initial count is 0, current count should also be 0 */
 528         if (apic_get_reg(apic, APIC_TMICT) == 0)
 529                 return 0;
 530
 531         remaining = hrtimer_expires_remaining(&apic->lapic_timer.timer);
 532         if (ktime_to_ns(remaining) < 0)
 533                 remaining = ktime_set(0, 0);
 534
 535         ns = mod_64(ktime_to_ns(remaining), apic->lapic_timer.period);
 536         tmcct = div64_u64(ns,
 537                          (APIC_BUS_CYCLE_NS * apic->divide_count));
 538
 539         return tmcct;
 540 }
 541
 542 static void __report_tpr_access(struct kvm_lapic *apic, bool write)
 543 {
 544         struct kvm_vcpu *vcpu = apic->vcpu;
 545         struct kvm_run *run = vcpu->run;
 546
 547         set_bit(KVM_REQ_REPORT_TPR_ACCESS, &vcpu->requests);
 548         run->tpr_access.rip = kvm_rip_read(vcpu);
 549         run->tpr_access.is_write = write;
 550 }
 551
 552 static inline void report_tpr_access(struct kvm_lapic *apic, bool write)
 553 {
 554         if (apic->vcpu->arch.tpr_access_reporting)
 555                 __report_tpr_access(apic, write);
 556 }
 557
 558 static u32 __apic_read(struct kvm_lapic *apic, unsigned int offset)
 559 {
 560         u32 val = 0;
 561
 562         KVMTRACE_1D(APIC_ACCESS, apic->vcpu, (u32)offset, handler);
 563
 564         if (offset >= LAPIC_MMIO_LENGTH)
 565                 return 0;
 566
 567         switch (offset) {
 568         case APIC_ARBPRI:
 569                 printk(KERN_WARNING "Access APIC ARBPRI register "
 570                        "which is for P6\n");
 571                 break;
 572
 573         case APIC_TMCCT:        /* Timer CCR */
 574                 val = apic_get_tmcct(apic);
 575                 break;
 576
 577         case APIC_TASKPRI:
 578                 report_tpr_access(apic, false);
 579                 /* fall thru */
 580         default:
 581                 apic_update_ppr(apic);
 582                 val = apic_get_reg(apic, offset);
 583                 break;
 584         }
 585
 586         return val;
 587 }
 588
 589 static void apic_mmio_read(struct kvm_io_device *this,
 590                            gpa_t address, int len, void *data)
 591 {
 592         struct kvm_lapic *apic = (struct kvm_lapic *)this->private;
 593         unsigned int offset = address - apic->base_address;
 594         unsigned char alignment = offset & 0xf;
 595         u32 result;
 596
 597         if ((alignment + len) > 4) {
 598                 printk(KERN_ERR "KVM_APIC_READ: alignment error %lx %d",
 599                        (unsigned long)address, len);
 600                 return;
 601         }
 602         result = __apic_read(apic, offset & ~0xf);
 603
 604         switch (len) {
 605         case 1:
 606         case 2:
 607         case 4:
 608                 memcpy(data, (char *)&result + alignment, len);
 609                 break;
 610         default:
 611                 printk(KERN_ERR "Local APIC read with len = %x, "
 612                        "should be 1,2, or 4 instead\n", len);
 613                 break;
 614         }
 615 }
 616
 617 static void update_divide_count(struct kvm_lapic *apic)
 618 {
 619         u32 tmp1, tmp2, tdcr;
 620
 621         tdcr = apic_get_reg(apic, APIC_TDCR);
 622         tmp1 = tdcr & 0xf;
 623         tmp2 = ((tmp1 & 0x3) | ((tmp1 & 0x8) >> 1)) + 1;
 624         apic->divide_count = 0x1 << (tmp2 & 0x7);
 625
 626         apic_debug("timer divide count is 0x%x\n",
 627                                    apic->lapic_timer.divide_count);
 628 }
 629
 630 static void start_apic_timer(struct kvm_lapic *apic)
 631 {
 632         ktime_t now = apic->lapic_timer.timer.base->get_time();
 633
 634         apic->lapic_timer.period = apic_get_reg(apic, APIC_TMICT) *
 635                     APIC_BUS_CYCLE_NS * apic->divide_count;
 636         atomic_set(&apic->lapic_timer.pending, 0);
 637
 638         if (!apic->lapic_timer.period)
 639                 return;
 640
 641         hrtimer_start(&apic->lapic_timer.timer,
 642                       ktime_add_ns(now, apic->lapic_timer.period),
 643                       HRTIMER_MODE_ABS);
 644
 645         apic_debug("%s: bus cycle is %" PRId64 "ns, now 0x%016"
 646                            PRIx64 ", "
 647                            "timer initial count 0x%x, period %lldns, "
 648                            "expire @ 0x%016" PRIx64 ".\n", __func__,
 649                            APIC_BUS_CYCLE_NS, ktime_to_ns(now),
 650                            apic_get_reg(apic, APIC_TMICT),
 651                            apic->lapic_timer.period,
 652                            ktime_to_ns(ktime_add_ns(now,
 653                                         apic->lapic_timer.period)));
 654 }
 655
 656 static void apic_manage_nmi_watchdog(struct kvm_lapic *apic, u32 lvt0_val)
 657 {
 658         int nmi_wd_enabled = apic_lvt_nmi_mode(apic_get_reg(apic, APIC_LVT0));
 659
 660         if (apic_lvt_nmi_mode(lvt0_val)) {
 661                 if (!nmi_wd_enabled) {
 662                         apic_debug("Receive NMI setting on APIC_LVT0 "
 663                                    "for cpu %d\n", apic->vcpu->vcpu_id);
 664                         apic->vcpu->kvm->arch.vapics_in_nmi_mode++;
 665                 }
 666         } else if (nmi_wd_enabled)
 667                 apic->vcpu->kvm->arch.vapics_in_nmi_mode--;
 668 }
 669
 670 static void apic_mmio_write(struct kvm_io_device *this,
 671                             gpa_t address, int len, const void *data)
 672 {
 673         struct kvm_lapic *apic = (struct kvm_lapic *)this->private;
 674         unsigned int offset = address - apic->base_address;
 675         unsigned char alignment = offset & 0xf;
 676         u32 val;
 677
 678         /*
 679          * APIC register must be aligned on 128-bits boundary.
 680          * 32/64/128 bits registers must be accessed thru 32 bits.
 681          * Refer SDM 8.4.1
 682          */
 683         if (len != 4 || alignment) {
 684                 /* Don't shout loud, $infamous_os would cause only noise. */
 685                 apic_debug("apic write: bad size=%d %lx\n",
 686                            len, (long)address);
 687                 return;
 688         }
 689
 690         val = *(u32 *) data;
 691
 692         /* too common printing */
 693         if (offset != APIC_EOI)
 694                 apic_debug("%s: offset 0x%x with length 0x%x, and value is "
 695                            "0x%x\n", __func__, offset, len, val);
 696
 697         offset &= 0xff0;
 698
 699         KVMTRACE_1D(APIC_ACCESS, apic->vcpu, (u32)offset, handler);
 700
 701         switch (offset) {
 702         case APIC_ID:           /* Local APIC ID */
 703                 apic_set_reg(apic, APIC_ID, val);
 704                 break;
 705
 706         case APIC_TASKPRI:
 707                 report_tpr_access(apic, true);
 708                 apic_set_tpr(apic, val & 0xff);
 709                 break;
 710
 711         case APIC_EOI:
 712                 apic_set_eoi(apic);
 713                 break;
 714
 715         case APIC_LDR:
 716                 apic_set_reg(apic, APIC_LDR, val & APIC_LDR_MASK);
 717                 break;
 718
 719         case APIC_DFR:
 720                 apic_set_reg(apic, APIC_DFR, val | 0x0FFFFFFF);
 721                 break;
 722
 723         case APIC_SPIV:
 724                 apic_set_reg(apic, APIC_SPIV, val & 0x3ff);
 725                 if (!(val & APIC_SPIV_APIC_ENABLED)) {
 726                         int i;
 727                         u32 lvt_val;
 728
 729                         for (i = 0; i < APIC_LVT_NUM; i++) {
 730                                 lvt_val = apic_get_reg(apic,
 731                                                        APIC_LVTT + 0x10 * i);
 732                                 apic_set_reg(apic, APIC_LVTT + 0x10 * i,
 733                                              lvt_val | APIC_LVT_MASKED);
 734                         }
 735                         atomic_set(&apic->lapic_timer.pending, 0);
 736
 737                 }
 738                 break;
 739
 740         case APIC_ICR:
 741                 /* No delay here, so we always clear the pending bit */
 742                 apic_set_reg(apic, APIC_ICR, val & ~(1 << 12));
 743                 apic_send_ipi(apic);
 744                 break;
 745
 746         case APIC_ICR2:
 747                 apic_set_reg(apic, APIC_ICR2, val & 0xff000000);
 748                 break;
 749
 750         case APIC_LVT0:
 751                 apic_manage_nmi_watchdog(apic, val);
 752         case APIC_LVTT:
 753         case APIC_LVTTHMR:
 754         case APIC_LVTPC:
 755         case APIC_LVT1:
 756         case APIC_LVTERR:
 757                 /* TODO: Check vector */
 758                 if (!apic_sw_enabled(apic))
 759                         val |= APIC_LVT_MASKED;
 760
 761                 val &= apic_lvt_mask[(offset - APIC_LVTT) >> 4];
 762                 apic_set_reg(apic, offset, val);
 763
 764                 break;
 765
 766         case APIC_TMICT:
 767                 hrtimer_cancel(&apic->lapic_timer.timer);
 768                 apic_set_reg(apic, APIC_TMICT, val);
 769                 start_apic_timer(apic);
 770                 return;
 771
 772         case APIC_TDCR:
 773                 if (val & 4)
 774                         printk(KERN_ERR "KVM_WRITE:TDCR %x\n", val);
 775                 apic_set_reg(apic, APIC_TDCR, val);
 776                 update_divide_count(apic);
 777                 break;
 778
 779         default:
 780                 apic_debug("Local APIC Write to read-only register %x\n",
 781                            offset);
 782                 break;
 783         }
 784
 785 }
 786
 787 static int apic_mmio_range(struct kvm_io_device *this, gpa_t addr,
 788                            int len, int size)
 789 {
 790         struct kvm_lapic *apic = (struct kvm_lapic *)this->private;
 791         int ret = 0;
 792
 793
 794         if (apic_hw_enabled(apic) &&
 795             (addr >= apic->base_address) &&
 796             (addr < (apic->base_address + LAPIC_MMIO_LENGTH)))
 797                 ret = 1;
 798
 799         return ret;
 800 }
 801
 802 void kvm_free_lapic(struct kvm_vcpu *vcpu)
 803 {
 804         if (!vcpu->arch.apic)
 805                 return;
 806
 807         hrtimer_cancel(&vcpu->arch.apic->lapic_timer.timer);
 808
 809         if (vcpu->arch.apic->regs_page)
 810                 __free_page(vcpu->arch.apic->regs_page);
 811
 812         kfree(vcpu->arch.apic);
 813 }
 814
 815 /*
 816  *----------------------------------------------------------------------
 817  * LAPIC interface
 818  *----------------------------------------------------------------------
 819  */
 820
 821 void kvm_lapic_set_tpr(struct kvm_vcpu *vcpu, unsigned long cr8)
 822 {
 823         struct kvm_lapic *apic = vcpu->arch.apic;
 824
 825         if (!apic)
 826                 return;
 827         apic_set_tpr(apic, ((cr8 & 0x0f) << 4)
 828                      | (apic_get_reg(apic, APIC_TASKPRI) & 4));
 829 }
 830 EXPORT_SYMBOL_GPL(kvm_lapic_set_tpr);
 831
 832 u64 kvm_lapic_get_cr8(struct kvm_vcpu *vcpu)
 833 {
 834         struct kvm_lapic *apic = vcpu->arch.apic;
 835         u64 tpr;
 836
 837         if (!apic)
 838                 return 0;
 839         tpr = (u64) apic_get_reg(apic, APIC_TASKPRI);
 840
 841         return (tpr & 0xf0) >> 4;
 842 }
 843 EXPORT_SYMBOL_GPL(kvm_lapic_get_cr8);
 844
 845 void kvm_lapic_set_base(struct kvm_vcpu *vcpu, u64 value)
 846 {
 847         struct kvm_lapic *apic = vcpu->arch.apic;
 848
 849         if (!apic) {
 850                 value |= MSR_IA32_APICBASE_BSP;
 851                 vcpu->arch.apic_base = value;
 852                 return;
 853         }
 854         if (apic->vcpu->vcpu_id)
 855                 value &= ~MSR_IA32_APICBASE_BSP;
 856
 857         vcpu->arch.apic_base = value;
 858         apic->base_address = apic->vcpu->arch.apic_base &
 859                              MSR_IA32_APICBASE_BASE;
 860
 861         /* with FSB delivery interrupt, we can restart APIC functionality */
 862         apic_debug("apic base msr is 0x%016" PRIx64 ", and base address is "
 863                    "0x%lx.\n", apic->vcpu->arch.apic_base, apic->base_address);
 864
 865 }
 866
 867 u64 kvm_lapic_get_base(struct kvm_vcpu *vcpu)
 868 {
 869         return vcpu->arch.apic_base;
 870 }
 871 EXPORT_SYMBOL_GPL(kvm_lapic_get_base);
 872
 873 void kvm_lapic_reset(struct kvm_vcpu *vcpu)
 874 {
 875         struct kvm_lapic *apic;
 876         int i;
 877
 878         apic_debug("%s\n", __func__);
 879
 880         ASSERT(vcpu);
 881         apic = vcpu->arch.apic;
 882         ASSERT(apic != NULL);
 883
 884         /* Stop the timer in case it's a reset to an active apic */
 885         hrtimer_cancel(&apic->lapic_timer.timer);
 886
 887         apic_set_reg(apic, APIC_ID, vcpu->vcpu_id << 24);
 888         apic_set_reg(apic, APIC_LVR, APIC_VERSION);
 889
 890         for (i = 0; i < APIC_LVT_NUM; i++)
 891                 apic_set_reg(apic, APIC_LVTT + 0x10 * i, APIC_LVT_MASKED);
 892         apic_set_reg(apic, APIC_LVT0,
 893                      SET_APIC_DELIVERY_MODE(0, APIC_MODE_EXTINT));
 894
 895         apic_set_reg(apic, APIC_DFR, 0xffffffffU);
 896         apic_set_reg(apic, APIC_SPIV, 0xff);
 897         apic_set_reg(apic, APIC_TASKPRI, 0);
 898         apic_set_reg(apic, APIC_LDR, 0);
 899         apic_set_reg(apic, APIC_ESR, 0);
 900         apic_set_reg(apic, APIC_ICR, 0);
 901         apic_set_reg(apic, APIC_ICR2, 0);
 902         apic_set_reg(apic, APIC_TDCR, 0);
 903         apic_set_reg(apic, APIC_TMICT, 0);
 904         for (i = 0; i < 8; i++) {
 905                 apic_set_reg(apic, APIC_IRR + 0x10 * i, 0);
 906                 apic_set_reg(apic, APIC_ISR + 0x10 * i, 0);
 907                 apic_set_reg(apic, APIC_TMR + 0x10 * i, 0);
 908         }
 909         update_divide_count(apic);
 910         atomic_set(&apic->lapic_timer.pending, 0);
 911         if (vcpu->vcpu_id == 0)
 912                 vcpu->arch.apic_base |= MSR_IA32_APICBASE_BSP;
 913         apic_update_ppr(apic);
 914
 915         apic_debug(KERN_INFO "%s: vcpu=%p, id=%d, base_msr="
 916                    "0x%016" PRIx64 ", base_address=0x%0lx.\n", __func__,
 917                    vcpu, kvm_apic_id(apic),
 918                    vcpu->arch.apic_base, apic->base_address);
 919 }
 920 EXPORT_SYMBOL_GPL(kvm_lapic_reset);
 921
 922 int kvm_lapic_enabled(struct kvm_vcpu *vcpu)
 923 {
 924         struct kvm_lapic *apic = vcpu->arch.apic;
 925         int ret = 0;
 926
 927         if (!apic)
 928                 return 0;
 929         ret = apic_enabled(apic);
 930
 931         return ret;
 932 }
 933 EXPORT_SYMBOL_GPL(kvm_lapic_enabled);
 934
 935 /*
 936  *----------------------------------------------------------------------
 937  * timer interface
 938  *----------------------------------------------------------------------
 939  */
 940
 941 static bool lapic_is_periodic(struct kvm_timer *ktimer)
 942 {
 943         struct kvm_lapic *apic = container_of(ktimer, struct kvm_lapic,
 944                                               lapic_timer);
 945         return apic_lvtt_period(apic);
 946 }
 947
 948 int apic_has_pending_timer(struct kvm_vcpu *vcpu)
 949 {
 950         struct kvm_lapic *lapic = vcpu->arch.apic;
 951
 952         if (lapic && apic_enabled(lapic) && apic_lvt_enabled(lapic, APIC_LVTT))
 953                 return atomic_read(&lapic->lapic_timer.pending);
 954
 955         return 0;
 956 }
 957
 958 static int kvm_apic_local_deliver(struct kvm_lapic *apic, int lvt_type)
 959 {
 960         u32 reg = apic_get_reg(apic, lvt_type);
 961         int vector, mode, trig_mode;
 962
 963         if (apic_hw_enabled(apic) && !(reg & APIC_LVT_MASKED)) {
 964                 vector = reg & APIC_VECTOR_MASK;
 965                 mode = reg & APIC_MODE_MASK;
 966                 trig_mode = reg & APIC_LVT_LEVEL_TRIGGER;
 967                 return __apic_accept_irq(apic, mode, vector, 1, trig_mode);
 968         }
 969         return 0;
 970 }
 971
 972 void kvm_apic_nmi_wd_deliver(struct kvm_vcpu *vcpu)
 973 {
 974         struct kvm_lapic *apic = vcpu->arch.apic;
 975
 976         if (apic)
 977                 kvm_apic_local_deliver(apic, APIC_LVT0);
 978 }
 979
 980 struct kvm_timer_ops lapic_timer_ops = {
 981         .is_periodic = lapic_is_periodic,
 982 };
 983
 984 int kvm_create_lapic(struct kvm_vcpu *vcpu)
 985 {
 986         struct kvm_lapic *apic;
 987
 988         ASSERT(vcpu != NULL);
 989         apic_debug("apic_init %d\n", vcpu->vcpu_id);
 990
 991         apic = kzalloc(sizeof(*apic), GFP_KERNEL);
 992         if (!apic)
 993                 goto nomem;
 994
 995         vcpu->arch.apic = apic;
 996
 997         apic->regs_page = alloc_page(GFP_KERNEL);
 998         if (apic->regs_page == NULL) {
 999                 printk(KERN_ERR "malloc apic regs error for vcpu %x\n",
1000                        vcpu->vcpu_id);
1001                 goto nomem_free_apic;
1002         }
1003         apic->regs = page_address(apic->regs_page);
1004         memset(apic->regs, 0, PAGE_SIZE);
1005         apic->vcpu = vcpu;
1006
1007         hrtimer_init(&apic->lapic_timer.timer, CLOCK_MONOTONIC,
1008                      HRTIMER_MODE_ABS);
1009         apic->lapic_timer.timer.function = kvm_timer_fn;
1010         apic->lapic_timer.t_ops = &lapic_timer_ops;
1011         apic->lapic_timer.kvm = vcpu->kvm;
1012         apic->lapic_timer.vcpu_id = vcpu->vcpu_id;
1013
1014         apic->base_address = APIC_DEFAULT_PHYS_BASE;
1015         vcpu->arch.apic_base = APIC_DEFAULT_PHYS_BASE;
1016
1017         kvm_lapic_reset(vcpu);
1018         apic->dev.read = apic_mmio_read;
1019         apic->dev.write = apic_mmio_write;
1020         apic->dev.in_range = apic_mmio_range;
1021         apic->dev.private = apic;
1022
1023         return 0;
1024 nomem_free_apic:
1025         kfree(apic);
1026 nomem:
1027         return -ENOMEM;
1028 }
1029 EXPORT_SYMBOL_GPL(kvm_create_lapic);
1030
1031 int kvm_apic_has_interrupt(struct kvm_vcpu *vcpu)
1032 {
1033         struct kvm_lapic *apic = vcpu->arch.apic;
1034         int highest_irr;
1035
1036         if (!apic || !apic_enabled(apic))
1037                 return -1;
1038
1039         apic_update_ppr(apic);
1040         highest_irr = apic_find_highest_irr(apic);
1041         if ((highest_irr == -1) ||
1042             ((highest_irr & 0xF0) <= apic_get_reg(apic, APIC_PROCPRI)))
1043                 return -1;
1044         return highest_irr;
1045 }
1046
1047 int kvm_apic_accept_pic_intr(struct kvm_vcpu *vcpu)
1048 {
1049         u32 lvt0 = apic_get_reg(vcpu->arch.apic, APIC_LVT0);
1050         int r = 0;
1051
1052         if (vcpu->vcpu_id == 0) {
1053                 if (!apic_hw_enabled(vcpu->arch.apic))
1054                         r = 1;
1055                 if ((lvt0 & APIC_LVT_MASKED) == 0 &&
1056                     GET_APIC_DELIVERY_MODE(lvt0) == APIC_MODE_EXTINT)
1057                         r = 1;
1058         }
1059         return r;
1060 }
1061
1062 void kvm_inject_apic_timer_irqs(struct kvm_vcpu *vcpu)
1063 {
1064         struct kvm_lapic *apic = vcpu->arch.apic;
1065
1066         if (apic && atomic_read(&apic->lapic_timer.pending) > 0) {
1067                 if (kvm_apic_local_deliver(apic, APIC_LVTT))
1068                         atomic_dec(&apic->lapic_timer.pending);
1069         }
1070 }
1071
1072 int kvm_get_apic_interrupt(struct kvm_vcpu *vcpu)
1073 {
1074         int vector = kvm_apic_has_interrupt(vcpu);
1075         struct kvm_lapic *apic = vcpu->arch.apic;
1076
1077         if (vector == -1)
1078                 return -1;
1079
1080         apic_set_vector(vector, apic->regs + APIC_ISR);
1081         apic_update_ppr(apic);
1082         apic_clear_irr(vector, apic);
1083         return vector;
1084 }
1085
1086 void kvm_apic_post_state_restore(struct kvm_vcpu *vcpu)
1087 {
1088         struct kvm_lapic *apic = vcpu->arch.apic;
1089
1090         apic->base_address = vcpu->arch.apic_base &
1091                              MSR_IA32_APICBASE_BASE;
1092         apic_set_reg(apic, APIC_LVR, APIC_VERSION);
1093         apic_update_ppr(apic);
1094         hrtimer_cancel(&apic->lapic_timer.timer);
1095         update_divide_count(apic);
1096         start_apic_timer(apic);
1097 }
1098
1099 void __kvm_migrate_apic_timer(struct kvm_vcpu *vcpu)
1100 {
1101         struct kvm_lapic *apic = vcpu->arch.apic;
1102         struct hrtimer *timer;
1103
1104         if (!apic)
1105                 return;
1106
1107         timer = &apic->lapic_timer.timer;
1108         if (hrtimer_cancel(timer))
1109                 hrtimer_start_expires(timer, HRTIMER_MODE_ABS);
1110 }
1111
1112 void kvm_lapic_sync_from_vapic(struct kvm_vcpu *vcpu)
1113 {
1114         u32 data;
1115         void *vapic;
1116
1117         if (!irqchip_in_kernel(vcpu->kvm) || !vcpu->arch.apic->vapic_addr)
1118                 return;
1119
1120         vapic = kmap_atomic(vcpu->arch.apic->vapic_page, KM_USER0);
1121         data = *(u32 *)(vapic + offset_in_page(vcpu->arch.apic->vapic_addr));
1122         kunmap_atomic(vapic, KM_USER0);
1123
1124         apic_set_tpr(vcpu->arch.apic, data & 0xff);
1125 }
1126
1127 void kvm_lapic_sync_to_vapic(struct kvm_vcpu *vcpu)
1128 {
1129         u32 data, tpr;
1130         int max_irr, max_isr;
1131         struct kvm_lapic *apic;
1132         void *vapic;
1133
1134         if (!irqchip_in_kernel(vcpu->kvm) || !vcpu->arch.apic->vapic_addr)
1135                 return;
1136
1137         apic = vcpu->arch.apic;
1138         tpr = apic_get_reg(apic, APIC_TASKPRI) & 0xff;
1139         max_irr = apic_find_highest_irr(apic);
1140         if (max_irr < 0)
1141                 max_irr = 0;
1142         max_isr = apic_find_highest_isr(apic);
1143         if (max_isr < 0)
1144                 max_isr = 0;
1145         data = (tpr & 0xff) | ((max_isr & 0xf0) << 8) | (max_irr << 24);
1146
1147         vapic = kmap_atomic(vcpu->arch.apic->vapic_page, KM_USER0);
1148         *(u32 *)(vapic + offset_in_page(vcpu->arch.apic->vapic_addr)) = data;
1149         kunmap_atomic(vapic, KM_USER0);
1150 }
1151
1152 void kvm_lapic_set_vapic_addr(struct kvm_vcpu *vcpu, gpa_t vapic_addr)
1153 {
1154         if (!irqchip_in_kernel(vcpu->kvm))
1155                 return;
1156
1157         vcpu->arch.apic->vapic_addr = vapic_addr;
1158 }