arch/x86_64/kernel/io_apic.c

   1 /*
   2  *      Intel IO-APIC support for multi-Pentium hosts.
   3  *
   4  *      Copyright (C) 1997, 1998, 1999, 2000 Ingo Molnar, Hajnalka Szabo
   5  *
   6  *      Many thanks to Stig Venaas for trying out countless experimental
   7  *      patches and reporting/debugging problems patiently!
   8  *
   9  *      (c) 1999, Multiple IO-APIC support, developed by
  10  *      Ken-ichi Yaku <yaku@css1.kbnes.nec.co.jp> and
  11  *      Hidemi Kishimoto <kisimoto@css1.kbnes.nec.co.jp>,
  12  *      further tested and cleaned up by Zach Brown <zab@redhat.com>
  13  *      and Ingo Molnar <mingo@redhat.com>
  14  *
  15  *      Fixes
  16  *      Maciej W. Rozycki       :       Bits for genuine 82489DX APICs;
  17  *                                      thanks to Eric Gilmore
  18  *                                      and Rolf G. Tews
  19  *                                      for testing these extensively
  20  *      Paul Diefenbaugh        :       Added full ACPI support
  21  */
  22
  23 #include <linux/mm.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/init.h>
  26 #include <linux/delay.h>
  27 #include <linux/sched.h>
  28 #include <linux/pci.h>
  29 #include <linux/mc146818rtc.h>
  30 #include <linux/acpi.h>
  31 #include <linux/sysdev.h>
  32 #include <linux/msi.h>
  33 #include <linux/htirq.h>
  34 #ifdef CONFIG_ACPI
  35 #include <acpi/acpi_bus.h>
  36 #endif
  37
  38 #include <asm/idle.h>
  39 #include <asm/io.h>
  40 #include <asm/smp.h>
  41 #include <asm/desc.h>
  42 #include <asm/proto.h>
  43 #include <asm/mach_apic.h>
  44 #include <asm/acpi.h>
  45 #include <asm/dma.h>
  46 #include <asm/nmi.h>
  47 #include <asm/msidef.h>
  48 #include <asm/hypertransport.h>
  49
  50 struct irq_cfg {
  51         cpumask_t domain;
  52         cpumask_t old_domain;
  53         unsigned move_cleanup_count;
  54         u8 vector;
  55         u8 move_in_progress : 1;
  56 };
  57
  58 /* irq_cfg is indexed by the sum of all RTEs in all I/O APICs. */
  59 struct irq_cfg irq_cfg[NR_IRQS] __read_mostly = {
  60         [0]  = { .domain = CPU_MASK_ALL, .vector = IRQ0_VECTOR,  },
  61         [1]  = { .domain = CPU_MASK_ALL, .vector = IRQ1_VECTOR,  },
  62         [2]  = { .domain = CPU_MASK_ALL, .vector = IRQ2_VECTOR,  },
  63         [3]  = { .domain = CPU_MASK_ALL, .vector = IRQ3_VECTOR,  },
  64         [4]  = { .domain = CPU_MASK_ALL, .vector = IRQ4_VECTOR,  },
  65         [5]  = { .domain = CPU_MASK_ALL, .vector = IRQ5_VECTOR,  },
  66         [6]  = { .domain = CPU_MASK_ALL, .vector = IRQ6_VECTOR,  },
  67         [7]  = { .domain = CPU_MASK_ALL, .vector = IRQ7_VECTOR,  },
  68         [8]  = { .domain = CPU_MASK_ALL, .vector = IRQ8_VECTOR,  },
  69         [9]  = { .domain = CPU_MASK_ALL, .vector = IRQ9_VECTOR,  },
  70         [10] = { .domain = CPU_MASK_ALL, .vector = IRQ10_VECTOR, },
  71         [11] = { .domain = CPU_MASK_ALL, .vector = IRQ11_VECTOR, },
  72         [12] = { .domain = CPU_MASK_ALL, .vector = IRQ12_VECTOR, },
  73         [13] = { .domain = CPU_MASK_ALL, .vector = IRQ13_VECTOR, },
  74         [14] = { .domain = CPU_MASK_ALL, .vector = IRQ14_VECTOR, },
  75         [15] = { .domain = CPU_MASK_ALL, .vector = IRQ15_VECTOR, },
  76 };
  77
  78 static int assign_irq_vector(int irq, cpumask_t mask);
  79
  80 #define __apicdebuginit  __init
  81
  82 int sis_apic_bug; /* not actually supported, dummy for compile */
  83
  84 static int no_timer_check;
  85
  86 static int disable_timer_pin_1 __initdata;
  87
  88 int timer_over_8254 __initdata = 1;
  89
  90 /* Where if anywhere is the i8259 connect in external int mode */
  91 static struct { int pin, apic; } ioapic_i8259 = { -1, -1 };
  92
  93 static DEFINE_SPINLOCK(ioapic_lock);
  94 DEFINE_SPINLOCK(vector_lock);
  95
  96 /*
  97  * # of IRQ routing registers
  98  */
  99 int nr_ioapic_registers[MAX_IO_APICS];
 100
 101 /*
 102  * Rough estimation of how many shared IRQs there are, can
 103  * be changed anytime.
 104  */
 105 #define MAX_PLUS_SHARED_IRQS NR_IRQS
 106 #define PIN_MAP_SIZE (MAX_PLUS_SHARED_IRQS + NR_IRQS)
 107
 108 /*
 109  * This is performance-critical, we want to do it O(1)
 110  *
 111  * the indexing order of this array favors 1:1 mappings
 112  * between pins and IRQs.
 113  */
 114
 115 static struct irq_pin_list {
 116         short apic, pin, next;
 117 } irq_2_pin[PIN_MAP_SIZE];
 118
 119 struct io_apic {
 120         unsigned int index;
 121         unsigned int unused[3];
 122         unsigned int data;
 123 };
 124
 125 static __attribute_const__ struct io_apic __iomem *io_apic_base(int idx)
 126 {
 127         return (void __iomem *) __fix_to_virt(FIX_IO_APIC_BASE_0 + idx)
 128                 + (mp_ioapics[idx].mpc_apicaddr & ~PAGE_MASK);
 129 }
 130
 131 static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)
 132 {
 133         struct io_apic __iomem *io_apic = io_apic_base(apic);
 134         writel(reg, &io_apic->index);
 135         return readl(&io_apic->data);
 136 }
 137
 138 static inline void io_apic_write(unsigned int apic, unsigned int reg, unsigned int value)
 139 {
 140         struct io_apic __iomem *io_apic = io_apic_base(apic);
 141         writel(reg, &io_apic->index);
 142         writel(value, &io_apic->data);
 143 }
 144
 145 /*
 146  * Re-write a value: to be used for read-modify-write
 147  * cycles where the read already set up the index register.
 148  */
 149 static inline void io_apic_modify(unsigned int apic, unsigned int value)
 150 {
 151         struct io_apic __iomem *io_apic = io_apic_base(apic);
 152         writel(value, &io_apic->data);
 153 }
 154
 155 /*
 156  * Synchronize the IO-APIC and the CPU by doing
 157  * a dummy read from the IO-APIC
 158  */
 159 static inline void io_apic_sync(unsigned int apic)
 160 {
 161         struct io_apic __iomem *io_apic = io_apic_base(apic);
 162         readl(&io_apic->data);
 163 }
 164
 165 #define __DO_ACTION(R, ACTION, FINAL)                                   \
 166                                                                         \
 167 {                                                                       \
 168         int pin;                                                        \
 169         struct irq_pin_list *entry = irq_2_pin + irq;                   \
 170                                                                         \
 171         BUG_ON(irq >= NR_IRQS);                                         \
 172         for (;;) {                                                      \
 173                 unsigned int reg;                                       \
 174                 pin = entry->pin;                                       \
 175                 if (pin == -1)                                          \
 176                         break;                                          \
 177                 reg = io_apic_read(entry->apic, 0x10 + R + pin*2);      \
 178                 reg ACTION;                                             \
 179                 io_apic_modify(entry->apic, reg);                       \
 180                 FINAL;                                                  \
 181                 if (!entry->next)                                       \
 182                         break;                                          \
 183                 entry = irq_2_pin + entry->next;                        \
 184         }                                                               \
 185 }
 186
 187 union entry_union {
 188         struct { u32 w1, w2; };
 189         struct IO_APIC_route_entry entry;
 190 };
 191
 192 static struct IO_APIC_route_entry ioapic_read_entry(int apic, int pin)
 193 {
 194         union entry_union eu;
 195         unsigned long flags;
 196         spin_lock_irqsave(&ioapic_lock, flags);
 197         eu.w1 = io_apic_read(apic, 0x10 + 2 * pin);
 198         eu.w2 = io_apic_read(apic, 0x11 + 2 * pin);
 199         spin_unlock_irqrestore(&ioapic_lock, flags);
 200         return eu.entry;
 201 }
 202
 203 /*
 204  * When we write a new IO APIC routing entry, we need to write the high
 205  * word first! If the mask bit in the low word is clear, we will enable
 206  * the interrupt, and we need to make sure the entry is fully populated
 207  * before that happens.
 208  */
 209 static void
 210 __ioapic_write_entry(int apic, int pin, struct IO_APIC_route_entry e)
 211 {
 212         union entry_union eu;
 213         eu.entry = e;
 214         io_apic_write(apic, 0x11 + 2*pin, eu.w2);
 215         io_apic_write(apic, 0x10 + 2*pin, eu.w1);
 216 }
 217
 218 static void ioapic_write_entry(int apic, int pin, struct IO_APIC_route_entry e)
 219 {
 220         unsigned long flags;
 221         spin_lock_irqsave(&ioapic_lock, flags);
 222         __ioapic_write_entry(apic, pin, e);
 223         spin_unlock_irqrestore(&ioapic_lock, flags);
 224 }
 225
 226 /*
 227  * When we mask an IO APIC routing entry, we need to write the low
 228  * word first, in order to set the mask bit before we change the
 229  * high bits!
 230  */
 231 static void ioapic_mask_entry(int apic, int pin)
 232 {
 233         unsigned long flags;
 234         union entry_union eu = { .entry.mask = 1 };
 235
 236         spin_lock_irqsave(&ioapic_lock, flags);
 237         io_apic_write(apic, 0x10 + 2*pin, eu.w1);
 238         io_apic_write(apic, 0x11 + 2*pin, eu.w2);
 239         spin_unlock_irqrestore(&ioapic_lock, flags);
 240 }
 241
 242 #ifdef CONFIG_SMP
 243 static void __target_IO_APIC_irq(unsigned int irq, unsigned int dest, u8 vector)
 244 {
 245         int apic, pin;
 246         struct irq_pin_list *entry = irq_2_pin + irq;
 247
 248         BUG_ON(irq >= NR_IRQS);
 249         for (;;) {
 250                 unsigned int reg;
 251                 apic = entry->apic;
 252                 pin = entry->pin;
 253                 if (pin == -1)
 254                         break;
 255                 io_apic_write(apic, 0x11 + pin*2, dest);
 256                 reg = io_apic_read(apic, 0x10 + pin*2);
 257                 reg &= ~0x000000ff;
 258                 reg |= vector;
 259                 io_apic_modify(apic, reg);
 260                 if (!entry->next)
 261                         break;
 262                 entry = irq_2_pin + entry->next;
 263         }
 264 }
 265
 266 static void set_ioapic_affinity_irq(unsigned int irq, cpumask_t mask)
 267 {
 268         struct irq_cfg *cfg = irq_cfg + irq;
 269         unsigned long flags;
 270         unsigned int dest;
 271         cpumask_t tmp;
 272
 273         cpus_and(tmp, mask, cpu_online_map);
 274         if (cpus_empty(tmp))
 275                 return;
 276
 277         if (assign_irq_vector(irq, mask))
 278                 return;
 279
 280         cpus_and(tmp, cfg->domain, mask);
 281         dest = cpu_mask_to_apicid(tmp);
 282
 283         /*
 284          * Only the high 8 bits are valid.
 285          */
 286         dest = SET_APIC_LOGICAL_ID(dest);
 287
 288         spin_lock_irqsave(&ioapic_lock, flags);
 289         __target_IO_APIC_irq(irq, dest, cfg->vector);
 290         irq_desc[irq].affinity = mask;
 291         spin_unlock_irqrestore(&ioapic_lock, flags);
 292 }
 293 #endif
 294
 295 /*
 296  * The common case is 1:1 IRQ<->pin mappings. Sometimes there are
 297  * shared ISA-space IRQs, so we have to support them. We are super
 298  * fast in the common case, and fast for shared ISA-space IRQs.
 299  */
 300 static void add_pin_to_irq(unsigned int irq, int apic, int pin)
 301 {
 302         static int first_free_entry = NR_IRQS;
 303         struct irq_pin_list *entry = irq_2_pin + irq;
 304
 305         BUG_ON(irq >= NR_IRQS);
 306         while (entry->next)
 307                 entry = irq_2_pin + entry->next;
 308
 309         if (entry->pin != -1) {
 310                 entry->next = first_free_entry;
 311                 entry = irq_2_pin + entry->next;
 312                 if (++first_free_entry >= PIN_MAP_SIZE)
 313                         panic("io_apic.c: ran out of irq_2_pin entries!");
 314         }
 315         entry->apic = apic;
 316         entry->pin = pin;
 317 }
 318
 319
 320 #define DO_ACTION(name,R,ACTION, FINAL)                                 \
 321                                                                         \
 322         static void name##_IO_APIC_irq (unsigned int irq)               \
 323         __DO_ACTION(R, ACTION, FINAL)
 324
 325 DO_ACTION( __mask,             0, |= 0x00010000, io_apic_sync(entry->apic) )
 326                                                 /* mask = 1 */
 327 DO_ACTION( __unmask,           0, &= 0xfffeffff, )
 328                                                 /* mask = 0 */
 329
 330 static void mask_IO_APIC_irq (unsigned int irq)
 331 {
 332         unsigned long flags;
 333
 334         spin_lock_irqsave(&ioapic_lock, flags);
 335         __mask_IO_APIC_irq(irq);
 336         spin_unlock_irqrestore(&ioapic_lock, flags);
 337 }
 338
 339 static void unmask_IO_APIC_irq (unsigned int irq)
 340 {
 341         unsigned long flags;
 342
 343         spin_lock_irqsave(&ioapic_lock, flags);
 344         __unmask_IO_APIC_irq(irq);
 345         spin_unlock_irqrestore(&ioapic_lock, flags);
 346 }
 347
 348 static void clear_IO_APIC_pin(unsigned int apic, unsigned int pin)
 349 {
 350         struct IO_APIC_route_entry entry;
 351
 352         /* Check delivery_mode to be sure we're not clearing an SMI pin */
 353         entry = ioapic_read_entry(apic, pin);
 354         if (entry.delivery_mode == dest_SMI)
 355                 return;
 356         /*
 357          * Disable it in the IO-APIC irq-routing table:
 358          */
 359         ioapic_mask_entry(apic, pin);
 360 }
 361
 362 static void clear_IO_APIC (void)
 363 {
 364         int apic, pin;
 365
 366         for (apic = 0; apic < nr_ioapics; apic++)
 367                 for (pin = 0; pin < nr_ioapic_registers[apic]; pin++)
 368                         clear_IO_APIC_pin(apic, pin);
 369 }
 370
 371 int skip_ioapic_setup;
 372 int ioapic_force;
 373
 374 /* dummy parsing: see setup.c */
 375
 376 static int __init disable_ioapic_setup(char *str)
 377 {
 378         skip_ioapic_setup = 1;
 379         return 0;
 380 }
 381 early_param("noapic", disable_ioapic_setup);
 382
 383 /* Actually the next is obsolete, but keep it for paranoid reasons -AK */
 384 static int __init disable_timer_pin_setup(char *arg)
 385 {
 386         disable_timer_pin_1 = 1;
 387         return 1;
 388 }
 389 __setup("disable_timer_pin_1", disable_timer_pin_setup);
 390
 391 static int __init setup_disable_8254_timer(char *s)
 392 {
 393         timer_over_8254 = -1;
 394         return 1;
 395 }
 396 static int __init setup_enable_8254_timer(char *s)
 397 {
 398         timer_over_8254 = 2;
 399         return 1;
 400 }
 401
 402 __setup("disable_8254_timer", setup_disable_8254_timer);
 403 __setup("enable_8254_timer", setup_enable_8254_timer);
 404
 405
 406 /*
 407  * Find the IRQ entry number of a certain pin.
 408  */
 409 static int find_irq_entry(int apic, int pin, int type)
 410 {
 411         int i;
 412
 413         for (i = 0; i < mp_irq_entries; i++)
 414                 if (mp_irqs[i].mpc_irqtype == type &&
 415                     (mp_irqs[i].mpc_dstapic == mp_ioapics[apic].mpc_apicid ||
 416                      mp_irqs[i].mpc_dstapic == MP_APIC_ALL) &&
 417                     mp_irqs[i].mpc_dstirq == pin)
 418                         return i;
 419
 420         return -1;
 421 }
 422
 423 /*
 424  * Find the pin to which IRQ[irq] (ISA) is connected
 425  */
 426 static int __init find_isa_irq_pin(int irq, int type)
 427 {
 428         int i;
 429
 430         for (i = 0; i < mp_irq_entries; i++) {
 431                 int lbus = mp_irqs[i].mpc_srcbus;
 432
 433                 if (test_bit(lbus, mp_bus_not_pci) &&
 434                     (mp_irqs[i].mpc_irqtype == type) &&
 435                     (mp_irqs[i].mpc_srcbusirq == irq))
 436
 437                         return mp_irqs[i].mpc_dstirq;
 438         }
 439         return -1;
 440 }
 441
 442 static int __init find_isa_irq_apic(int irq, int type)
 443 {
 444         int i;
 445
 446         for (i = 0; i < mp_irq_entries; i++) {
 447                 int lbus = mp_irqs[i].mpc_srcbus;
 448
 449                 if (test_bit(lbus, mp_bus_not_pci) &&
 450                     (mp_irqs[i].mpc_irqtype == type) &&
 451                     (mp_irqs[i].mpc_srcbusirq == irq))
 452                         break;
 453         }
 454         if (i < mp_irq_entries) {
 455                 int apic;
 456                 for(apic = 0; apic < nr_ioapics; apic++) {
 457                         if (mp_ioapics[apic].mpc_apicid == mp_irqs[i].mpc_dstapic)
 458                                 return apic;
 459                 }
 460         }
 461
 462         return -1;
 463 }
 464
 465 /*
 466  * Find a specific PCI IRQ entry.
 467  * Not an __init, possibly needed by modules
 468  */
 469 static int pin_2_irq(int idx, int apic, int pin);
 470
 471 int IO_APIC_get_PCI_irq_vector(int bus, int slot, int pin)
 472 {
 473         int apic, i, best_guess = -1;
 474
 475         apic_printk(APIC_DEBUG, "querying PCI -> IRQ mapping bus:%d, slot:%d, pin:%d.\n",
 476                 bus, slot, pin);
 477         if (mp_bus_id_to_pci_bus[bus] == -1) {
 478                 apic_printk(APIC_VERBOSE, "PCI BIOS passed nonexistent PCI bus %d!\n", bus);
 479                 return -1;
 480         }
 481         for (i = 0; i < mp_irq_entries; i++) {
 482                 int lbus = mp_irqs[i].mpc_srcbus;
 483
 484                 for (apic = 0; apic < nr_ioapics; apic++)
 485                         if (mp_ioapics[apic].mpc_apicid == mp_irqs[i].mpc_dstapic ||
 486                             mp_irqs[i].mpc_dstapic == MP_APIC_ALL)
 487                                 break;
 488
 489                 if (!test_bit(lbus, mp_bus_not_pci) &&
 490                     !mp_irqs[i].mpc_irqtype &&
 491                     (bus == lbus) &&
 492                     (slot == ((mp_irqs[i].mpc_srcbusirq >> 2) & 0x1f))) {
 493                         int irq = pin_2_irq(i,apic,mp_irqs[i].mpc_dstirq);
 494
 495                         if (!(apic || IO_APIC_IRQ(irq)))
 496                                 continue;
 497
 498                         if (pin == (mp_irqs[i].mpc_srcbusirq & 3))
 499                                 return irq;
 500                         /*
 501                          * Use the first all-but-pin matching entry as a
 502                          * best-guess fuzzy result for broken mptables.
 503                          */
 504                         if (best_guess < 0)
 505                                 best_guess = irq;
 506                 }
 507         }
 508         BUG_ON(best_guess >= NR_IRQS);
 509         return best_guess;
 510 }
 511
 512 /* ISA interrupts are always polarity zero edge triggered,
 513  * when listed as conforming in the MP table. */
 514
 515 #define default_ISA_trigger(idx)        (0)
 516 #define default_ISA_polarity(idx)       (0)
 517
 518 /* PCI interrupts are always polarity one level triggered,
 519  * when listed as conforming in the MP table. */
 520
 521 #define default_PCI_trigger(idx)        (1)
 522 #define default_PCI_polarity(idx)       (1)
 523
 524 static int __init MPBIOS_polarity(int idx)
 525 {
 526         int bus = mp_irqs[idx].mpc_srcbus;
 527         int polarity;
 528
 529         /*
 530          * Determine IRQ line polarity (high active or low active):
 531          */
 532         switch (mp_irqs[idx].mpc_irqflag & 3)
 533         {
 534                 case 0: /* conforms, ie. bus-type dependent polarity */
 535                         if (test_bit(bus, mp_bus_not_pci))
 536                                 polarity = default_ISA_polarity(idx);
 537                         else
 538                                 polarity = default_PCI_polarity(idx);
 539                         break;
 540                 case 1: /* high active */
 541                 {
 542                         polarity = 0;
 543                         break;
 544                 }
 545                 case 2: /* reserved */
 546                 {
 547                         printk(KERN_WARNING "broken BIOS!!\n");
 548                         polarity = 1;
 549                         break;
 550                 }
 551                 case 3: /* low active */
 552                 {
 553                         polarity = 1;
 554                         break;
 555                 }
 556                 default: /* invalid */
 557                 {
 558                         printk(KERN_WARNING "broken BIOS!!\n");
 559                         polarity = 1;
 560                         break;
 561                 }
 562         }
 563         return polarity;
 564 }
 565
 566 static int MPBIOS_trigger(int idx)
 567 {
 568         int bus = mp_irqs[idx].mpc_srcbus;
 569         int trigger;
 570
 571         /*
 572          * Determine IRQ trigger mode (edge or level sensitive):
 573          */
 574         switch ((mp_irqs[idx].mpc_irqflag>>2) & 3)
 575         {
 576                 case 0: /* conforms, ie. bus-type dependent */
 577                         if (test_bit(bus, mp_bus_not_pci))
 578                                 trigger = default_ISA_trigger(idx);
 579                         else
 580                                 trigger = default_PCI_trigger(idx);
 581                         break;
 582                 case 1: /* edge */
 583                 {
 584                         trigger = 0;
 585                         break;
 586                 }
 587                 case 2: /* reserved */
 588                 {
 589                         printk(KERN_WARNING "broken BIOS!!\n");
 590                         trigger = 1;
 591                         break;
 592                 }
 593                 case 3: /* level */
 594                 {
 595                         trigger = 1;
 596                         break;
 597                 }
 598                 default: /* invalid */
 599                 {
 600                         printk(KERN_WARNING "broken BIOS!!\n");
 601                         trigger = 0;
 602                         break;
 603                 }
 604         }
 605         return trigger;
 606 }
 607
 608 static inline int irq_polarity(int idx)
 609 {
 610         return MPBIOS_polarity(idx);
 611 }
 612
 613 static inline int irq_trigger(int idx)
 614 {
 615         return MPBIOS_trigger(idx);
 616 }
 617
 618 static int pin_2_irq(int idx, int apic, int pin)
 619 {
 620         int irq, i;
 621         int bus = mp_irqs[idx].mpc_srcbus;
 622
 623         /*
 624          * Debugging check, we are in big trouble if this message pops up!
 625          */
 626         if (mp_irqs[idx].mpc_dstirq != pin)
 627                 printk(KERN_ERR "broken BIOS or MPTABLE parser, ayiee!!\n");
 628
 629         if (test_bit(bus, mp_bus_not_pci)) {
 630                 irq = mp_irqs[idx].mpc_srcbusirq;
 631         } else {
 632                 /*
 633                  * PCI IRQs are mapped in order
 634                  */
 635                 i = irq = 0;
 636                 while (i < apic)
 637                         irq += nr_ioapic_registers[i++];
 638                 irq += pin;
 639         }
 640         BUG_ON(irq >= NR_IRQS);
 641         return irq;
 642 }
 643
 644 static int __assign_irq_vector(int irq, cpumask_t mask)
 645 {
 646         /*
 647          * NOTE! The local APIC isn't very good at handling
 648          * multiple interrupts at the same interrupt level.
 649          * As the interrupt level is determined by taking the
 650          * vector number and shifting that right by 4, we
 651          * want to spread these out a bit so that they don't
 652          * all fall in the same interrupt level.
 653          *
 654          * Also, we've got to be careful not to trash gate
 655          * 0x80, because int 0x80 is hm, kind of importantish. ;)
 656          */
 657         static int current_vector = FIRST_DEVICE_VECTOR, current_offset = 0;
 658         unsigned int old_vector;
 659         int cpu;
 660         struct irq_cfg *cfg;
 661
 662         BUG_ON((unsigned)irq >= NR_IRQS);
 663         cfg = &irq_cfg[irq];
 664
 665         /* Only try and allocate irqs on cpus that are present */
 666         cpus_and(mask, mask, cpu_online_map);
 667
 668         if ((cfg->move_in_progress) || cfg->move_cleanup_count)
 669                 return -EBUSY;
 670
 671         old_vector = cfg->vector;
 672         if (old_vector) {
 673                 cpumask_t tmp;
 674                 cpus_and(tmp, cfg->domain, mask);
 675                 if (!cpus_empty(tmp))
 676                         return 0;
 677         }
 678
 679         for_each_cpu_mask(cpu, mask) {
 680                 cpumask_t domain, new_mask;
 681                 int new_cpu;
 682                 int vector, offset;
 683
 684                 domain = vector_allocation_domain(cpu);
 685                 cpus_and(new_mask, domain, cpu_online_map);
 686
 687                 vector = current_vector;
 688                 offset = current_offset;
 689 next:
 690                 vector += 8;
 691                 if (vector >= FIRST_SYSTEM_VECTOR) {
 692                         /* If we run out of vectors on large boxen, must share them. */
 693                         offset = (offset + 1) % 8;
 694                         vector = FIRST_DEVICE_VECTOR + offset;
 695                 }
 696                 if (unlikely(current_vector == vector))
 697                         continue;
 698                 if (vector == IA32_SYSCALL_VECTOR)
 699                         goto next;
 700                 for_each_cpu_mask(new_cpu, new_mask)
 701                         if (per_cpu(vector_irq, new_cpu)[vector] != -1)
 702                                 goto next;
 703                 /* Found one! */
 704                 current_vector = vector;
 705                 current_offset = offset;
 706                 if (old_vector) {
 707                         cfg->move_in_progress = 1;
 708                         cfg->old_domain = cfg->domain;
 709                 }
 710                 for_each_cpu_mask(new_cpu, new_mask)
 711                         per_cpu(vector_irq, new_cpu)[vector] = irq;
 712                 cfg->vector = vector;
 713                 cfg->domain = domain;
 714                 return 0;
 715         }
 716         return -ENOSPC;
 717 }
 718
 719 static int assign_irq_vector(int irq, cpumask_t mask)
 720 {
 721         int err;
 722         unsigned long flags;
 723
 724         spin_lock_irqsave(&vector_lock, flags);
 725         err = __assign_irq_vector(irq, mask);
 726         spin_unlock_irqrestore(&vector_lock, flags);
 727         return err;
 728 }
 729
 730 static void __clear_irq_vector(int irq)
 731 {
 732         struct irq_cfg *cfg;
 733         cpumask_t mask;
 734         int cpu, vector;
 735
 736         BUG_ON((unsigned)irq >= NR_IRQS);
 737         cfg = &irq_cfg[irq];
 738         BUG_ON(!cfg->vector);
 739
 740         vector = cfg->vector;
 741         cpus_and(mask, cfg->domain, cpu_online_map);
 742         for_each_cpu_mask(cpu, mask)
 743                 per_cpu(vector_irq, cpu)[vector] = -1;
 744
 745         cfg->vector = 0;
 746         cfg->domain = CPU_MASK_NONE;
 747 }
 748
 749 void __setup_vector_irq(int cpu)
 750 {
 751         /* Initialize vector_irq on a new cpu */
 752         /* This function must be called with vector_lock held */
 753         int irq, vector;
 754
 755         /* Mark the inuse vectors */
 756         for (irq = 0; irq < NR_IRQS; ++irq) {
 757                 if (!cpu_isset(cpu, irq_cfg[irq].domain))
 758                         continue;
 759                 vector = irq_cfg[irq].vector;
 760                 per_cpu(vector_irq, cpu)[vector] = irq;
 761         }
 762         /* Mark the free vectors */
 763         for (vector = 0; vector < NR_VECTORS; ++vector) {
 764                 irq = per_cpu(vector_irq, cpu)[vector];
 765                 if (irq < 0)
 766                         continue;
 767                 if (!cpu_isset(cpu, irq_cfg[irq].domain))
 768                         per_cpu(vector_irq, cpu)[vector] = -1;
 769         }
 770 }
 771
 772
 773 static struct irq_chip ioapic_chip;
 774
 775 static void ioapic_register_intr(int irq, unsigned long trigger)
 776 {
 777         if (trigger)
 778                 set_irq_chip_and_handler_name(irq, &ioapic_chip,
 779                                               handle_fasteoi_irq, "fasteoi");
 780         else
 781                 set_irq_chip_and_handler_name(irq, &ioapic_chip,
 782                                               handle_edge_irq, "edge");
 783 }
 784
 785 static void setup_IO_APIC_irq(int apic, int pin, unsigned int irq,
 786                               int trigger, int polarity)
 787 {
 788         struct irq_cfg *cfg = irq_cfg + irq;
 789         struct IO_APIC_route_entry entry;
 790         cpumask_t mask;
 791
 792         if (!IO_APIC_IRQ(irq))
 793                 return;
 794
 795         mask = TARGET_CPUS;
 796         if (assign_irq_vector(irq, mask))
 797                 return;
 798
 799         cpus_and(mask, cfg->domain, mask);
 800
 801         apic_printk(APIC_VERBOSE,KERN_DEBUG
 802                     "IOAPIC[%d]: Set routing entry (%d-%d -> 0x%x -> "
 803                     "IRQ %d Mode:%i Active:%i)\n",
 804                     apic, mp_ioapics[apic].mpc_apicid, pin, cfg->vector,
 805                     irq, trigger, polarity);
 806
 807         /*
 808          * add it to the IO-APIC irq-routing table:
 809          */
 810         memset(&entry,0,sizeof(entry));
 811
 812         entry.delivery_mode = INT_DELIVERY_MODE;
 813         entry.dest_mode = INT_DEST_MODE;
 814         entry.dest = cpu_mask_to_apicid(mask);
 815         entry.mask = 0;                         /* enable IRQ */
 816         entry.trigger = trigger;
 817         entry.polarity = polarity;
 818         entry.vector = cfg->vector;
 819
 820         /* Mask level triggered irqs.
 821          * Use IRQ_DELAYED_DISABLE for edge triggered irqs.
 822          */
 823         if (trigger)
 824                 entry.mask = 1;
 825
 826         ioapic_register_intr(irq, trigger);
 827         if (irq < 16)
 828                 disable_8259A_irq(irq);
 829
 830         ioapic_write_entry(apic, pin, entry);
 831 }
 832
 833 static void __init setup_IO_APIC_irqs(void)
 834 {
 835         int apic, pin, idx, irq, first_notcon = 1;
 836
 837         apic_printk(APIC_VERBOSE, KERN_DEBUG "init IO_APIC IRQs\n");
 838
 839         for (apic = 0; apic < nr_ioapics; apic++) {
 840         for (pin = 0; pin < nr_ioapic_registers[apic]; pin++) {
 841
 842                 idx = find_irq_entry(apic,pin,mp_INT);
 843                 if (idx == -1) {
 844                         if (first_notcon) {
 845                                 apic_printk(APIC_VERBOSE, KERN_DEBUG " IO-APIC (apicid-pin) %d-%d", mp_ioapics[apic].mpc_apicid, pin);
 846                                 first_notcon = 0;
 847                         } else
 848                                 apic_printk(APIC_VERBOSE, ", %d-%d", mp_ioapics[apic].mpc_apicid, pin);
 849                         continue;
 850                 }
 851
 852                 irq = pin_2_irq(idx, apic, pin);
 853                 add_pin_to_irq(irq, apic, pin);
 854
 855                 setup_IO_APIC_irq(apic, pin, irq,
 856                                   irq_trigger(idx), irq_polarity(idx));
 857         }
 858         }
 859
 860         if (!first_notcon)
 861                 apic_printk(APIC_VERBOSE," not connected.\n");
 862 }
 863
 864 /*
 865  * Set up the 8259A-master output pin as broadcast to all
 866  * CPUs.
 867  */
 868 static void __init setup_ExtINT_IRQ0_pin(unsigned int apic, unsigned int pin, int vector)
 869 {
 870         struct IO_APIC_route_entry entry;
 871         unsigned long flags;
 872
 873         memset(&entry,0,sizeof(entry));
 874
 875         disable_8259A_irq(0);
 876
 877         /* mask LVT0 */
 878         apic_write(APIC_LVT0, APIC_LVT_MASKED | APIC_DM_EXTINT);
 879
 880         /*
 881          * We use logical delivery to get the timer IRQ
 882          * to the first CPU.
 883          */
 884         entry.dest_mode = INT_DEST_MODE;
 885         entry.mask = 0;                                 /* unmask IRQ now */
 886         entry.dest = cpu_mask_to_apicid(TARGET_CPUS);
 887         entry.delivery_mode = INT_DELIVERY_MODE;
 888         entry.polarity = 0;
 889         entry.trigger = 0;
 890         entry.vector = vector;
 891
 892         /*
 893          * The timer IRQ doesn't have to know that behind the
 894          * scene we have a 8259A-master in AEOI mode ...
 895          */
 896         set_irq_chip_and_handler_name(0, &ioapic_chip, handle_edge_irq, "edge");
 897
 898         /*
 899          * Add it to the IO-APIC irq-routing table:
 900          */
 901         spin_lock_irqsave(&ioapic_lock, flags);
 902         io_apic_write(apic, 0x11+2*pin, *(((int *)&entry)+1));
 903         io_apic_write(apic, 0x10+2*pin, *(((int *)&entry)+0));
 904         spin_unlock_irqrestore(&ioapic_lock, flags);
 905
 906         enable_8259A_irq(0);
 907 }
 908
 909 void __apicdebuginit print_IO_APIC(void)
 910 {
 911         int apic, i;
 912         union IO_APIC_reg_00 reg_00;
 913         union IO_APIC_reg_01 reg_01;
 914         union IO_APIC_reg_02 reg_02;
 915         unsigned long flags;
 916
 917         if (apic_verbosity == APIC_QUIET)
 918                 return;
 919
 920         printk(KERN_DEBUG "number of MP IRQ sources: %d.\n", mp_irq_entries);
 921         for (i = 0; i < nr_ioapics; i++)
 922                 printk(KERN_DEBUG "number of IO-APIC #%d registers: %d.\n",
 923                        mp_ioapics[i].mpc_apicid, nr_ioapic_registers[i]);
 924
 925         /*
 926          * We are a bit conservative about what we expect.  We have to
 927          * know about every hardware change ASAP.
 928          */
 929         printk(KERN_INFO "testing the IO APIC.......................\n");
 930
 931         for (apic = 0; apic < nr_ioapics; apic++) {
 932
 933         spin_lock_irqsave(&ioapic_lock, flags);
 934         reg_00.raw = io_apic_read(apic, 0);
 935         reg_01.raw = io_apic_read(apic, 1);
 936         if (reg_01.bits.version >= 0x10)
 937                 reg_02.raw = io_apic_read(apic, 2);
 938         spin_unlock_irqrestore(&ioapic_lock, flags);
 939
 940         printk("\n");
 941         printk(KERN_DEBUG "IO APIC #%d......\n", mp_ioapics[apic].mpc_apicid);
 942         printk(KERN_DEBUG ".... register #00: %08X\n", reg_00.raw);
 943         printk(KERN_DEBUG ".......    : physical APIC id: %02X\n", reg_00.bits.ID);
 944
 945         printk(KERN_DEBUG ".... register #01: %08X\n", *(int *)&reg_01);
 946         printk(KERN_DEBUG ".......     : max redirection entries: %04X\n", reg_01.bits.entries);
 947
 948         printk(KERN_DEBUG ".......     : PRQ implemented: %X\n", reg_01.bits.PRQ);
 949         printk(KERN_DEBUG ".......     : IO APIC version: %04X\n", reg_01.bits.version);
 950
 951         if (reg_01.bits.version >= 0x10) {
 952                 printk(KERN_DEBUG ".... register #02: %08X\n", reg_02.raw);
 953                 printk(KERN_DEBUG ".......     : arbitration: %02X\n", reg_02.bits.arbitration);
 954         }
 955
 956         printk(KERN_DEBUG ".... IRQ redirection table:\n");
 957
 958         printk(KERN_DEBUG " NR Dst Mask Trig IRR Pol"
 959                           " Stat Dmod Deli Vect:   \n");
 960
 961         for (i = 0; i <= reg_01.bits.entries; i++) {
 962                 struct IO_APIC_route_entry entry;
 963
 964                 entry = ioapic_read_entry(apic, i);
 965
 966                 printk(KERN_DEBUG " %02x %03X ",
 967                         i,
 968                         entry.dest
 969                 );
 970
 971                 printk("%1d    %1d    %1d   %1d   %1d    %1d    %1d    %02X\n",
 972                         entry.mask,
 973                         entry.trigger,
 974                         entry.irr,
 975                         entry.polarity,
 976                         entry.delivery_status,
 977                         entry.dest_mode,
 978                         entry.delivery_mode,
 979                         entry.vector
 980                 );
 981         }
 982         }
 983         printk(KERN_DEBUG "IRQ to pin mappings:\n");
 984         for (i = 0; i < NR_IRQS; i++) {
 985                 struct irq_pin_list *entry = irq_2_pin + i;
 986                 if (entry->pin < 0)
 987                         continue;
 988                 printk(KERN_DEBUG "IRQ%d ", i);
 989                 for (;;) {
 990                         printk("-> %d:%d", entry->apic, entry->pin);
 991                         if (!entry->next)
 992                                 break;
 993                         entry = irq_2_pin + entry->next;
 994                 }
 995                 printk("\n");
 996         }
 997
 998         printk(KERN_INFO ".................................... done.\n");
 999
1000         return;
1001 }
1002
1003 #if 0
1004
1005 static __apicdebuginit void print_APIC_bitfield (int base)
1006 {
1007         unsigned int v;
1008         int i, j;
1009
1010         if (apic_verbosity == APIC_QUIET)
1011                 return;
1012
1013         printk(KERN_DEBUG "0123456789abcdef0123456789abcdef\n" KERN_DEBUG);
1014         for (i = 0; i < 8; i++) {
1015                 v = apic_read(base + i*0x10);
1016                 for (j = 0; j < 32; j++) {
1017                         if (v & (1<<j))
1018                                 printk("1");
1019                         else
1020                                 printk("0");
1021                 }
1022                 printk("\n");
1023         }
1024 }
1025
1026 void __apicdebuginit print_local_APIC(void * dummy)
1027 {
1028         unsigned int v, ver, maxlvt;
1029
1030         if (apic_verbosity == APIC_QUIET)
1031                 return;
1032
1033         printk("\n" KERN_DEBUG "printing local APIC contents on CPU#%d/%d:\n",
1034                 smp_processor_id(), hard_smp_processor_id());
1035         v = apic_read(APIC_ID);
1036         printk(KERN_INFO "... APIC ID:      %08x (%01x)\n", v, GET_APIC_ID(v));
1037         v = apic_read(APIC_LVR);
1038         printk(KERN_INFO "... APIC VERSION: %08x\n", v);
1039         ver = GET_APIC_VERSION(v);
1040         maxlvt = get_maxlvt();
1041
1042         v = apic_read(APIC_TASKPRI);
1043         printk(KERN_DEBUG "... APIC TASKPRI: %08x (%02x)\n", v, v & APIC_TPRI_MASK);
1044
1045         v = apic_read(APIC_ARBPRI);
1046         printk(KERN_DEBUG "... APIC ARBPRI: %08x (%02x)\n", v,
1047                 v & APIC_ARBPRI_MASK);
1048         v = apic_read(APIC_PROCPRI);
1049         printk(KERN_DEBUG "... APIC PROCPRI: %08x\n", v);
1050
1051         v = apic_read(APIC_EOI);
1052         printk(KERN_DEBUG "... APIC EOI: %08x\n", v);
1053         v = apic_read(APIC_RRR);
1054         printk(KERN_DEBUG "... APIC RRR: %08x\n", v);
1055         v = apic_read(APIC_LDR);
1056         printk(KERN_DEBUG "... APIC LDR: %08x\n", v);
1057         v = apic_read(APIC_DFR);
1058         printk(KERN_DEBUG "... APIC DFR: %08x\n", v);
1059         v = apic_read(APIC_SPIV);
1060         printk(KERN_DEBUG "... APIC SPIV: %08x\n", v);
1061
1062         printk(KERN_DEBUG "... APIC ISR field:\n");
1063         print_APIC_bitfield(APIC_ISR);
1064         printk(KERN_DEBUG "... APIC TMR field:\n");
1065         print_APIC_bitfield(APIC_TMR);
1066         printk(KERN_DEBUG "... APIC IRR field:\n");
1067         print_APIC_bitfield(APIC_IRR);
1068
1069         v = apic_read(APIC_ESR);
1070         printk(KERN_DEBUG "... APIC ESR: %08x\n", v);
1071
1072         v = apic_read(APIC_ICR);
1073         printk(KERN_DEBUG "... APIC ICR: %08x\n", v);
1074         v = apic_read(APIC_ICR2);
1075         printk(KERN_DEBUG "... APIC ICR2: %08x\n", v);
1076
1077         v = apic_read(APIC_LVTT);
1078         printk(KERN_DEBUG "... APIC LVTT: %08x\n", v);
1079
1080         if (maxlvt > 3) {                       /* PC is LVT#4. */
1081                 v = apic_read(APIC_LVTPC);
1082                 printk(KERN_DEBUG "... APIC LVTPC: %08x\n", v);
1083         }
1084         v = apic_read(APIC_LVT0);
1085         printk(KERN_DEBUG "... APIC LVT0: %08x\n", v);
1086         v = apic_read(APIC_LVT1);
1087         printk(KERN_DEBUG "... APIC LVT1: %08x\n", v);
1088
1089         if (maxlvt > 2) {                       /* ERR is LVT#3. */
1090                 v = apic_read(APIC_LVTERR);
1091                 printk(KERN_DEBUG "... APIC LVTERR: %08x\n", v);
1092         }
1093
1094         v = apic_read(APIC_TMICT);
1095         printk(KERN_DEBUG "... APIC TMICT: %08x\n", v);
1096         v = apic_read(APIC_TMCCT);
1097         printk(KERN_DEBUG "... APIC TMCCT: %08x\n", v);
1098         v = apic_read(APIC_TDCR);
1099         printk(KERN_DEBUG "... APIC TDCR: %08x\n", v);
1100         printk("\n");
1101 }
1102
1103 void print_all_local_APICs (void)
1104 {
1105         on_each_cpu(print_local_APIC, NULL, 1, 1);
1106 }
1107
1108 void __apicdebuginit print_PIC(void)
1109 {
1110         unsigned int v;
1111         unsigned long flags;
1112
1113         if (apic_verbosity == APIC_QUIET)
1114                 return;
1115
1116         printk(KERN_DEBUG "\nprinting PIC contents\n");
1117
1118         spin_lock_irqsave(&i8259A_lock, flags);
1119
1120         v = inb(0xa1) << 8 | inb(0x21);
1121         printk(KERN_DEBUG "... PIC  IMR: %04x\n", v);
1122
1123         v = inb(0xa0) << 8 | inb(0x20);
1124         printk(KERN_DEBUG "... PIC  IRR: %04x\n", v);
1125
1126         outb(0x0b,0xa0);
1127         outb(0x0b,0x20);
1128         v = inb(0xa0) << 8 | inb(0x20);
1129         outb(0x0a,0xa0);
1130         outb(0x0a,0x20);
1131
1132         spin_unlock_irqrestore(&i8259A_lock, flags);
1133
1134         printk(KERN_DEBUG "... PIC  ISR: %04x\n", v);
1135
1136         v = inb(0x4d1) << 8 | inb(0x4d0);
1137         printk(KERN_DEBUG "... PIC ELCR: %04x\n", v);
1138 }
1139
1140 #endif  /*  0  */
1141
1142 static void __init enable_IO_APIC(void)
1143 {
1144         union IO_APIC_reg_01 reg_01;
1145         int i8259_apic, i8259_pin;
1146         int i, apic;
1147         unsigned long flags;
1148
1149         for (i = 0; i < PIN_MAP_SIZE; i++) {
1150                 irq_2_pin[i].pin = -1;
1151                 irq_2_pin[i].next = 0;
1152         }
1153
1154         /*
1155          * The number of IO-APIC IRQ registers (== #pins):
1156          */
1157         for (apic = 0; apic < nr_ioapics; apic++) {
1158                 spin_lock_irqsave(&ioapic_lock, flags);
1159                 reg_01.raw = io_apic_read(apic, 1);
1160                 spin_unlock_irqrestore(&ioapic_lock, flags);
1161                 nr_ioapic_registers[apic] = reg_01.bits.entries+1;
1162         }
1163         for(apic = 0; apic < nr_ioapics; apic++) {
1164                 int pin;
1165                 /* See if any of the pins is in ExtINT mode */
1166                 for (pin = 0; pin < nr_ioapic_registers[apic]; pin++) {
1167                         struct IO_APIC_route_entry entry;
1168                         entry = ioapic_read_entry(apic, pin);
1169
1170                         /* If the interrupt line is enabled and in ExtInt mode
1171                          * I have found the pin where the i8259 is connected.
1172                          */
1173                         if ((entry.mask == 0) && (entry.delivery_mode == dest_ExtINT)) {
1174                                 ioapic_i8259.apic = apic;
1175                                 ioapic_i8259.pin  = pin;
1176                                 goto found_i8259;
1177                         }
1178                 }
1179         }
1180  found_i8259:
1181         /* Look to see what if the MP table has reported the ExtINT */
1182         i8259_pin  = find_isa_irq_pin(0, mp_ExtINT);
1183         i8259_apic = find_isa_irq_apic(0, mp_ExtINT);
1184         /* Trust the MP table if nothing is setup in the hardware */
1185         if ((ioapic_i8259.pin == -1) && (i8259_pin >= 0)) {
1186                 printk(KERN_WARNING "ExtINT not setup in hardware but reported by MP table\n");
1187                 ioapic_i8259.pin  = i8259_pin;
1188                 ioapic_i8259.apic = i8259_apic;
1189         }
1190         /* Complain if the MP table and the hardware disagree */
1191         if (((ioapic_i8259.apic != i8259_apic) || (ioapic_i8259.pin != i8259_pin)) &&
1192                 (i8259_pin >= 0) && (ioapic_i8259.pin >= 0))
1193         {
1194                 printk(KERN_WARNING "ExtINT in hardware and MP table differ\n");
1195         }
1196
1197         /*
1198          * Do not trust the IO-APIC being empty at bootup
1199          */
1200         clear_IO_APIC();
1201 }
1202
1203 /*
1204  * Not an __init, needed by the reboot code
1205  */
1206 void disable_IO_APIC(void)
1207 {
1208         /*
1209          * Clear the IO-APIC before rebooting:
1210          */
1211         clear_IO_APIC();
1212
1213         /*
1214          * If the i8259 is routed through an IOAPIC
1215          * Put that IOAPIC in virtual wire mode
1216          * so legacy interrupts can be delivered.
1217          */
1218         if (ioapic_i8259.pin != -1) {
1219                 struct IO_APIC_route_entry entry;
1220
1221                 memset(&entry, 0, sizeof(entry));
1222                 entry.mask            = 0; /* Enabled */
1223                 entry.trigger         = 0; /* Edge */
1224                 entry.irr             = 0;
1225                 entry.polarity        = 0; /* High */
1226                 entry.delivery_status = 0;
1227                 entry.dest_mode       = 0; /* Physical */
1228                 entry.delivery_mode   = dest_ExtINT; /* ExtInt */
1229                 entry.vector          = 0;
1230                 entry.dest          = GET_APIC_ID(apic_read(APIC_ID));
1231
1232                 /*
1233                  * Add it to the IO-APIC irq-routing table:
1234                  */
1235                 ioapic_write_entry(ioapic_i8259.apic, ioapic_i8259.pin, entry);
1236         }
1237
1238         disconnect_bsp_APIC(ioapic_i8259.pin != -1);
1239 }
1240
1241 /*
1242  * There is a nasty bug in some older SMP boards, their mptable lies
1243  * about the timer IRQ. We do the following to work around the situation:
1244  *
1245  *      - timer IRQ defaults to IO-APIC IRQ
1246  *      - if this function detects that timer IRQs are defunct, then we fall
1247  *        back to ISA timer IRQs
1248  */
1249 static int __init timer_irq_works(void)
1250 {
1251         unsigned long t1 = jiffies;
1252
1253         local_irq_enable();
1254         /* Let ten ticks pass... */
1255         mdelay((10 * 1000) / HZ);
1256
1257         /*
1258          * Expect a few ticks at least, to be sure some possible
1259          * glue logic does not lock up after one or two first
1260          * ticks in a non-ExtINT mode.  Also the local APIC
1261          * might have cached one ExtINT interrupt.  Finally, at
1262          * least one tick may be lost due to delays.
1263          */
1264
1265         /* jiffies wrap? */
1266         if (jiffies - t1 > 4)
1267                 return 1;
1268         return 0;
1269 }
1270
1271 /*
1272  * In the SMP+IOAPIC case it might happen that there are an unspecified
1273  * number of pending IRQ events unhandled. These cases are very rare,
1274  * so we 'resend' these IRQs via IPIs, to the same CPU. It's much
1275  * better to do it this way as thus we do not have to be aware of
1276  * 'pending' interrupts in the IRQ path, except at this point.
1277  */
1278 /*
1279  * Edge triggered needs to resend any interrupt
1280  * that was delayed but this is now handled in the device
1281  * independent code.
1282  */
1283
1284 /*
1285  * Starting up a edge-triggered IO-APIC interrupt is
1286  * nasty - we need to make sure that we get the edge.
1287  * If it is already asserted for some reason, we need
1288  * return 1 to indicate that is was pending.
1289  *
1290  * This is not complete - we should be able to fake
1291  * an edge even if it isn't on the 8259A...
1292  */
1293
1294 static unsigned int startup_ioapic_irq(unsigned int irq)
1295 {
1296         int was_pending = 0;
1297         unsigned long flags;
1298
1299         spin_lock_irqsave(&ioapic_lock, flags);
1300         if (irq < 16) {
1301                 disable_8259A_irq(irq);
1302                 if (i8259A_irq_pending(irq))
1303                         was_pending = 1;
1304         }
1305         __unmask_IO_APIC_irq(irq);
1306         spin_unlock_irqrestore(&ioapic_lock, flags);
1307
1308         return was_pending;
1309 }
1310
1311 static int ioapic_retrigger_irq(unsigned int irq)
1312 {
1313         struct irq_cfg *cfg = &irq_cfg[irq];
1314         cpumask_t mask;
1315         unsigned long flags;
1316
1317         spin_lock_irqsave(&vector_lock, flags);
1318         cpus_clear(mask);
1319         cpu_set(first_cpu(cfg->domain), mask);
1320
1321         send_IPI_mask(mask, cfg->vector);
1322         spin_unlock_irqrestore(&vector_lock, flags);
1323
1324         return 1;
1325 }
1326
1327 /*
1328  * Level and edge triggered IO-APIC interrupts need different handling,
1329  * so we use two separate IRQ descriptors. Edge triggered IRQs can be
1330  * handled with the level-triggered descriptor, but that one has slightly
1331  * more overhead. Level-triggered interrupts cannot be handled with the
1332  * edge-triggered handler, without risking IRQ storms and other ugly
1333  * races.
1334  */
1335
1336 #ifdef CONFIG_SMP
1337 asmlinkage void smp_irq_move_cleanup_interrupt(void)
1338 {
1339         unsigned vector, me;
1340         ack_APIC_irq();
1341         exit_idle();
1342         irq_enter();
1343
1344         me = smp_processor_id();
1345         for (vector = FIRST_EXTERNAL_VECTOR; vector < NR_VECTORS; vector++) {
1346                 unsigned int irq;
1347                 struct irq_desc *desc;
1348                 struct irq_cfg *cfg;
1349                 irq = __get_cpu_var(vector_irq)[vector];
1350                 if (irq >= NR_IRQS)
1351                         continue;
1352
1353                 desc = irq_desc + irq;
1354                 cfg = irq_cfg + irq;
1355                 spin_lock(&desc->lock);
1356                 if (!cfg->move_cleanup_count)
1357                         goto unlock;
1358
1359                 if ((vector == cfg->vector) && cpu_isset(me, cfg->domain))
1360                         goto unlock;
1361
1362                 __get_cpu_var(vector_irq)[vector] = -1;
1363                 cfg->move_cleanup_count--;
1364 unlock:
1365                 spin_unlock(&desc->lock);
1366         }
1367
1368         irq_exit();
1369 }
1370
1371 static void irq_complete_move(unsigned int irq)
1372 {
1373         struct irq_cfg *cfg = irq_cfg + irq;
1374         unsigned vector, me;
1375
1376         if (likely(!cfg->move_in_progress))
1377                 return;
1378
1379         vector = ~get_irq_regs()->orig_rax;
1380         me = smp_processor_id();
1381         if ((vector == cfg->vector) && cpu_isset(me, cfg->domain)) {
1382                 cpumask_t cleanup_mask;
1383
1384                 cpus_and(cleanup_mask, cfg->old_domain, cpu_online_map);
1385                 cfg->move_cleanup_count = cpus_weight(cleanup_mask);
1386                 send_IPI_mask(cleanup_mask, IRQ_MOVE_CLEANUP_VECTOR);
1387                 cfg->move_in_progress = 0;
1388         }
1389 }
1390 #else
1391 static inline void irq_complete_move(unsigned int irq) {}
1392 #endif
1393
1394 static void ack_apic_edge(unsigned int irq)
1395 {
1396         irq_complete_move(irq);
1397         move_native_irq(irq);
1398         ack_APIC_irq();
1399 }
1400
1401 static void ack_apic_level(unsigned int irq)
1402 {
1403         int do_unmask_irq = 0;
1404
1405         irq_complete_move(irq);
1406 #if defined(CONFIG_GENERIC_PENDING_IRQ) || defined(CONFIG_IRQBALANCE)
1407         /* If we are moving the irq we need to mask it */
1408         if (unlikely(irq_desc[irq].status & IRQ_MOVE_PENDING)) {
1409                 do_unmask_irq = 1;
1410                 mask_IO_APIC_irq(irq);
1411         }
1412 #endif
1413
1414         /*
1415          * We must acknowledge the irq before we move it or the acknowledge will
1416          * not propogate properly.
1417          */
1418         ack_APIC_irq();
1419
1420         /* Now we can move and renable the irq */
1421         move_masked_irq(irq);
1422         if (unlikely(do_unmask_irq))
1423                 unmask_IO_APIC_irq(irq);
1424 }
1425
1426 static struct irq_chip ioapic_chip __read_mostly = {
1427         .name           = "IO-APIC",
1428         .startup        = startup_ioapic_irq,
1429         .mask           = mask_IO_APIC_irq,
1430         .unmask         = unmask_IO_APIC_irq,
1431         .ack            = ack_apic_edge,
1432         .eoi            = ack_apic_level,
1433 #ifdef CONFIG_SMP
1434         .set_affinity   = set_ioapic_affinity_irq,
1435 #endif
1436         .retrigger      = ioapic_retrigger_irq,
1437 };
1438
1439 static inline void init_IO_APIC_traps(void)
1440 {
1441         int irq;
1442
1443         /*
1444          * NOTE! The local APIC isn't very good at handling
1445          * multiple interrupts at the same interrupt level.
1446          * As the interrupt level is determined by taking the
1447          * vector number and shifting that right by 4, we
1448          * want to spread these out a bit so that they don't
1449          * all fall in the same interrupt level.
1450          *
1451          * Also, we've got to be careful not to trash gate
1452          * 0x80, because int 0x80 is hm, kind of importantish. ;)
1453          */
1454         for (irq = 0; irq < NR_IRQS ; irq++) {
1455                 int tmp = irq;
1456                 if (IO_APIC_IRQ(tmp) && !irq_cfg[tmp].vector) {
1457                         /*
1458                          * Hmm.. We don't have an entry for this,
1459                          * so default to an old-fashioned 8259
1460                          * interrupt if we can..
1461                          */
1462                         if (irq < 16)
1463                                 make_8259A_irq(irq);
1464                         else
1465                                 /* Strange. Oh, well.. */
1466                                 irq_desc[irq].chip = &no_irq_chip;
1467                 }
1468         }
1469 }
1470
1471 static void enable_lapic_irq (unsigned int irq)
1472 {
1473         unsigned long v;
1474
1475         v = apic_read(APIC_LVT0);
1476         apic_write(APIC_LVT0, v & ~APIC_LVT_MASKED);
1477 }
1478
1479 static void disable_lapic_irq (unsigned int irq)
1480 {
1481         unsigned long v;
1482
1483         v = apic_read(APIC_LVT0);
1484         apic_write(APIC_LVT0, v | APIC_LVT_MASKED);
1485 }
1486
1487 static void ack_lapic_irq (unsigned int irq)
1488 {
1489         ack_APIC_irq();
1490 }
1491
1492 static void end_lapic_irq (unsigned int i) { /* nothing */ }
1493
1494 static struct hw_interrupt_type lapic_irq_type __read_mostly = {
1495         .typename = "local-APIC-edge",
1496         .startup = NULL, /* startup_irq() not used for IRQ0 */
1497         .shutdown = NULL, /* shutdown_irq() not used for IRQ0 */
1498         .enable = enable_lapic_irq,
1499         .disable = disable_lapic_irq,
1500         .ack = ack_lapic_irq,
1501         .end = end_lapic_irq,
1502 };
1503
1504 static void setup_nmi (void)
1505 {
1506         /*
1507          * Dirty trick to enable the NMI watchdog ...
1508          * We put the 8259A master into AEOI mode and
1509          * unmask on all local APICs LVT0 as NMI.
1510          *
1511          * The idea to use the 8259A in AEOI mode ('8259A Virtual Wire')
1512          * is from Maciej W. Rozycki - so we do not have to EOI from
1513          * the NMI handler or the timer interrupt.
1514          */
1515         printk(KERN_INFO "activating NMI Watchdog ...");
1516
1517         enable_NMI_through_LVT0(NULL);
1518
1519         printk(" done.\n");
1520 }
1521
1522 /*
1523  * This looks a bit hackish but it's about the only one way of sending
1524  * a few INTA cycles to 8259As and any associated glue logic.  ICR does
1525  * not support the ExtINT mode, unfortunately.  We need to send these
1526  * cycles as some i82489DX-based boards have glue logic that keeps the
1527  * 8259A interrupt line asserted until INTA.  --macro
1528  */
1529 static inline void unlock_ExtINT_logic(void)
1530 {
1531         int apic, pin, i;
1532         struct IO_APIC_route_entry entry0, entry1;
1533         unsigned char save_control, save_freq_select;
1534         unsigned long flags;
1535
1536         pin  = find_isa_irq_pin(8, mp_INT);
1537         apic = find_isa_irq_apic(8, mp_INT);
1538         if (pin == -1)
1539                 return;
1540
1541         spin_lock_irqsave(&ioapic_lock, flags);
1542         *(((int *)&entry0) + 1) = io_apic_read(apic, 0x11 + 2 * pin);
1543         *(((int *)&entry0) + 0) = io_apic_read(apic, 0x10 + 2 * pin);
1544         spin_unlock_irqrestore(&ioapic_lock, flags);
1545         clear_IO_APIC_pin(apic, pin);
1546
1547         memset(&entry1, 0, sizeof(entry1));
1548
1549         entry1.dest_mode = 0;                   /* physical delivery */
1550         entry1.mask = 0;                        /* unmask IRQ now */
1551         entry1.dest = hard_smp_processor_id();
1552         entry1.delivery_mode = dest_ExtINT;
1553         entry1.polarity = entry0.polarity;
1554         entry1.trigger = 0;
1555         entry1.vector = 0;
1556
1557         spin_lock_irqsave(&ioapic_lock, flags);
1558         io_apic_write(apic, 0x11 + 2 * pin, *(((int *)&entry1) + 1));
1559         io_apic_write(apic, 0x10 + 2 * pin, *(((int *)&entry1) + 0));
1560         spin_unlock_irqrestore(&ioapic_lock, flags);
1561
1562         save_control = CMOS_READ(RTC_CONTROL);
1563         save_freq_select = CMOS_READ(RTC_FREQ_SELECT);
1564         CMOS_WRITE((save_freq_select & ~RTC_RATE_SELECT) | 0x6,
1565                    RTC_FREQ_SELECT);
1566         CMOS_WRITE(save_control | RTC_PIE, RTC_CONTROL);
1567
1568         i = 100;
1569         while (i-- > 0) {
1570                 mdelay(10);
1571                 if ((CMOS_READ(RTC_INTR_FLAGS) & RTC_PF) == RTC_PF)
1572                         i -= 10;
1573         }
1574
1575         CMOS_WRITE(save_control, RTC_CONTROL);
1576         CMOS_WRITE(save_freq_select, RTC_FREQ_SELECT);
1577         clear_IO_APIC_pin(apic, pin);
1578
1579         spin_lock_irqsave(&ioapic_lock, flags);
1580         io_apic_write(apic, 0x11 + 2 * pin, *(((int *)&entry0) + 1));
1581         io_apic_write(apic, 0x10 + 2 * pin, *(((int *)&entry0) + 0));
1582         spin_unlock_irqrestore(&ioapic_lock, flags);
1583 }
1584
1585 /*
1586  * This code may look a bit paranoid, but it's supposed to cooperate with
1587  * a wide range of boards and BIOS bugs.  Fortunately only the timer IRQ
1588  * is so screwy.  Thanks to Brian Perkins for testing/hacking this beast
1589  * fanatically on his truly buggy board.
1590  *
1591  * FIXME: really need to revamp this for modern platforms only.
1592  */
1593 static inline void check_timer(void)
1594 {
1595         struct irq_cfg *cfg = irq_cfg + 0;
1596         int apic1, pin1, apic2, pin2;
1597
1598         /*
1599          * get/set the timer IRQ vector:
1600          */
1601         disable_8259A_irq(0);
1602         assign_irq_vector(0, TARGET_CPUS);
1603
1604         /*
1605          * Subtle, code in do_timer_interrupt() expects an AEOI
1606          * mode for the 8259A whenever interrupts are routed
1607          * through I/O APICs.  Also IRQ0 has to be enabled in
1608          * the 8259A which implies the virtual wire has to be
1609          * disabled in the local APIC.
1610          */
1611         apic_write(APIC_LVT0, APIC_LVT_MASKED | APIC_DM_EXTINT);
1612         init_8259A(1);
1613         if (timer_over_8254 > 0)
1614                 enable_8259A_irq(0);
1615
1616         pin1  = find_isa_irq_pin(0, mp_INT);
1617         apic1 = find_isa_irq_apic(0, mp_INT);
1618         pin2  = ioapic_i8259.pin;
1619         apic2 = ioapic_i8259.apic;
1620
1621         apic_printk(APIC_VERBOSE,KERN_INFO "..TIMER: vector=0x%02X apic1=%d pin1=%d apic2=%d pin2=%d\n",
1622                 cfg->vector, apic1, pin1, apic2, pin2);
1623
1624         if (pin1 != -1) {
1625                 /*
1626                  * Ok, does IRQ0 through the IOAPIC work?
1627                  */
1628                 unmask_IO_APIC_irq(0);
1629                 if (!no_timer_check && timer_irq_works()) {
1630                         nmi_watchdog_default();
1631                         if (nmi_watchdog == NMI_IO_APIC) {
1632                                 disable_8259A_irq(0);
1633                                 setup_nmi();
1634                                 enable_8259A_irq(0);
1635                         }
1636                         if (disable_timer_pin_1 > 0)
1637                                 clear_IO_APIC_pin(0, pin1);
1638                         return;
1639                 }
1640                 clear_IO_APIC_pin(apic1, pin1);
1641                 apic_printk(APIC_QUIET,KERN_ERR "..MP-BIOS bug: 8254 timer not "
1642                                 "connected to IO-APIC\n");
1643         }
1644
1645         apic_printk(APIC_VERBOSE,KERN_INFO "...trying to set up timer (IRQ0) "
1646                                 "through the 8259A ... ");
1647         if (pin2 != -1) {
1648                 apic_printk(APIC_VERBOSE,"\n..... (found apic %d pin %d) ...",
1649                         apic2, pin2);
1650                 /*
1651                  * legacy devices should be connected to IO APIC #0
1652                  */
1653                 setup_ExtINT_IRQ0_pin(apic2, pin2, cfg->vector);
1654                 if (timer_irq_works()) {
1655                         apic_printk(APIC_VERBOSE," works.\n");
1656                         nmi_watchdog_default();
1657                         if (nmi_watchdog == NMI_IO_APIC) {
1658                                 setup_nmi();
1659                         }
1660                         return;
1661                 }
1662                 /*
1663                  * Cleanup, just in case ...
1664                  */
1665                 clear_IO_APIC_pin(apic2, pin2);
1666         }
1667         apic_printk(APIC_VERBOSE," failed.\n");
1668
1669         if (nmi_watchdog == NMI_IO_APIC) {
1670                 printk(KERN_WARNING "timer doesn't work through the IO-APIC - disabling NMI Watchdog!\n");
1671                 nmi_watchdog = 0;
1672         }
1673
1674         apic_printk(APIC_VERBOSE, KERN_INFO "...trying to set up timer as Virtual Wire IRQ...");
1675
1676         disable_8259A_irq(0);
1677         irq_desc[0].chip = &lapic_irq_type;
1678         apic_write(APIC_LVT0, APIC_DM_FIXED | cfg->vector);     /* Fixed mode */
1679         enable_8259A_irq(0);
1680
1681         if (timer_irq_works()) {
1682                 apic_printk(APIC_VERBOSE," works.\n");
1683                 return;
1684         }
1685         apic_write(APIC_LVT0, APIC_LVT_MASKED | APIC_DM_FIXED | cfg->vector);
1686         apic_printk(APIC_VERBOSE," failed.\n");
1687
1688         apic_printk(APIC_VERBOSE, KERN_INFO "...trying to set up timer as ExtINT IRQ...");
1689
1690         init_8259A(0);
1691         make_8259A_irq(0);
1692         apic_write(APIC_LVT0, APIC_DM_EXTINT);
1693
1694         unlock_ExtINT_logic();
1695
1696         if (timer_irq_works()) {
1697                 apic_printk(APIC_VERBOSE," works.\n");
1698                 return;
1699         }
1700         apic_printk(APIC_VERBOSE," failed :(.\n");
1701         panic("IO-APIC + timer doesn't work! Try using the 'noapic' kernel parameter\n");
1702 }
1703
1704 static int __init notimercheck(char *s)
1705 {
1706         no_timer_check = 1;
1707         return 1;
1708 }
1709 __setup("no_timer_check", notimercheck);
1710
1711 /*
1712  *
1713  * IRQ's that are handled by the PIC in the MPS IOAPIC case.
1714  * - IRQ2 is the cascade IRQ, and cannot be a io-apic IRQ.
1715  *   Linux doesn't really care, as it's not actually used
1716  *   for any interrupt handling anyway.
1717  */
1718 #define PIC_IRQS        (1<<2)
1719
1720 void __init setup_IO_APIC(void)
1721 {
1722         enable_IO_APIC();
1723
1724         if (acpi_ioapic)
1725                 io_apic_irqs = ~0;      /* all IRQs go through IOAPIC */
1726         else
1727                 io_apic_irqs = ~PIC_IRQS;
1728
1729         apic_printk(APIC_VERBOSE, "ENABLING IO-APIC IRQs\n");
1730
1731         sync_Arb_IDs();
1732         setup_IO_APIC_irqs();
1733         init_IO_APIC_traps();
1734         check_timer();
1735         if (!acpi_ioapic)
1736                 print_IO_APIC();
1737 }
1738
1739 struct sysfs_ioapic_data {
1740         struct sys_device dev;
1741         struct IO_APIC_route_entry entry[0];
1742 };
1743 static struct sysfs_ioapic_data * mp_ioapic_data[MAX_IO_APICS];
1744
1745 static int ioapic_suspend(struct sys_device *dev, pm_message_t state)
1746 {
1747         struct IO_APIC_route_entry *entry;
1748         struct sysfs_ioapic_data *data;
1749         int i;
1750
1751         data = container_of(dev, struct sysfs_ioapic_data, dev);
1752         entry = data->entry;
1753         for (i = 0; i < nr_ioapic_registers[dev->id]; i ++, entry ++ )
1754                 *entry = ioapic_read_entry(dev->id, i);
1755
1756         return 0;
1757 }
1758
1759 static int ioapic_resume(struct sys_device *dev)
1760 {
1761         struct IO_APIC_route_entry *entry;
1762         struct sysfs_ioapic_data *data;
1763         unsigned long flags;
1764         union IO_APIC_reg_00 reg_00;
1765         int i;
1766
1767         data = container_of(dev, struct sysfs_ioapic_data, dev);
1768         entry = data->entry;
1769
1770         spin_lock_irqsave(&ioapic_lock, flags);
1771         reg_00.raw = io_apic_read(dev->id, 0);
1772         if (reg_00.bits.ID != mp_ioapics[dev->id].mpc_apicid) {
1773                 reg_00.bits.ID = mp_ioapics[dev->id].mpc_apicid;
1774                 io_apic_write(dev->id, 0, reg_00.raw);
1775         }
1776         spin_unlock_irqrestore(&ioapic_lock, flags);
1777         for (i = 0; i < nr_ioapic_registers[dev->id]; i++)
1778                 ioapic_write_entry(dev->id, i, entry[i]);
1779
1780         return 0;
1781 }
1782
1783 static struct sysdev_class ioapic_sysdev_class = {
1784         set_kset_name("ioapic"),
1785         .suspend = ioapic_suspend,
1786         .resume = ioapic_resume,
1787 };
1788
1789 static int __init ioapic_init_sysfs(void)
1790 {
1791         struct sys_device * dev;
1792         int i, size, error = 0;
1793
1794         error = sysdev_class_register(&ioapic_sysdev_class);
1795         if (error)
1796                 return error;
1797
1798         for (i = 0; i < nr_ioapics; i++ ) {
1799                 size = sizeof(struct sys_device) + nr_ioapic_registers[i]
1800                         * sizeof(struct IO_APIC_route_entry);
1801                 mp_ioapic_data[i] = kmalloc(size, GFP_KERNEL);
1802                 if (!mp_ioapic_data[i]) {
1803                         printk(KERN_ERR "Can't suspend/resume IOAPIC %d\n", i);
1804                         continue;
1805                 }
1806                 memset(mp_ioapic_data[i], 0, size);
1807                 dev = &mp_ioapic_data[i]->dev;
1808                 dev->id = i;
1809                 dev->cls = &ioapic_sysdev_class;
1810                 error = sysdev_register(dev);
1811                 if (error) {
1812                         kfree(mp_ioapic_data[i]);
1813                         mp_ioapic_data[i] = NULL;
1814                         printk(KERN_ERR "Can't suspend/resume IOAPIC %d\n", i);
1815                         continue;
1816                 }
1817         }
1818
1819         return 0;
1820 }
1821
1822 device_initcall(ioapic_init_sysfs);
1823
1824 /*
1825  * Dynamic irq allocate and deallocation
1826  */
1827 int create_irq(void)
1828 {
1829         /* Allocate an unused irq */
1830         int irq;
1831         int new;
1832         unsigned long flags;
1833
1834         irq = -ENOSPC;
1835         spin_lock_irqsave(&vector_lock, flags);
1836         for (new = (NR_IRQS - 1); new >= 0; new--) {
1837                 if (platform_legacy_irq(new))
1838                         continue;
1839                 if (irq_cfg[new].vector != 0)
1840                         continue;
1841                 if (__assign_irq_vector(new, TARGET_CPUS) == 0)
1842                         irq = new;
1843                 break;
1844         }
1845         spin_unlock_irqrestore(&vector_lock, flags);
1846
1847         if (irq >= 0) {
1848                 dynamic_irq_init(irq);
1849         }
1850         return irq;
1851 }
1852
1853 void destroy_irq(unsigned int irq)
1854 {
1855         unsigned long flags;
1856
1857         dynamic_irq_cleanup(irq);
1858
1859         spin_lock_irqsave(&vector_lock, flags);
1860         __clear_irq_vector(irq);
1861         spin_unlock_irqrestore(&vector_lock, flags);
1862 }
1863
1864 /*
1865  * MSI mesage composition
1866  */
1867 #ifdef CONFIG_PCI_MSI
1868 static int msi_compose_msg(struct pci_dev *pdev, unsigned int irq, struct msi_msg *msg)
1869 {
1870         struct irq_cfg *cfg = irq_cfg + irq;
1871         int err;
1872         unsigned dest;
1873         cpumask_t tmp;
1874
1875         tmp = TARGET_CPUS;
1876         err = assign_irq_vector(irq, tmp);
1877         if (!err) {
1878                 cpus_and(tmp, cfg->domain, tmp);
1879                 dest = cpu_mask_to_apicid(tmp);
1880
1881                 msg->address_hi = MSI_ADDR_BASE_HI;
1882                 msg->address_lo =
1883                         MSI_ADDR_BASE_LO |
1884                         ((INT_DEST_MODE == 0) ?
1885                                 MSI_ADDR_DEST_MODE_PHYSICAL:
1886                                 MSI_ADDR_DEST_MODE_LOGICAL) |
1887                         ((INT_DELIVERY_MODE != dest_LowestPrio) ?
1888                                 MSI_ADDR_REDIRECTION_CPU:
1889                                 MSI_ADDR_REDIRECTION_LOWPRI) |
1890                         MSI_ADDR_DEST_ID(dest);
1891
1892                 msg->data =
1893                         MSI_DATA_TRIGGER_EDGE |
1894                         MSI_DATA_LEVEL_ASSERT |
1895                         ((INT_DELIVERY_MODE != dest_LowestPrio) ?
1896                                 MSI_DATA_DELIVERY_FIXED:
1897                                 MSI_DATA_DELIVERY_LOWPRI) |
1898                         MSI_DATA_VECTOR(cfg->vector);
1899         }
1900         return err;
1901 }
1902
1903 #ifdef CONFIG_SMP
1904 static void set_msi_irq_affinity(unsigned int irq, cpumask_t mask)
1905 {
1906         struct irq_cfg *cfg = irq_cfg + irq;
1907         struct msi_msg msg;
1908         unsigned int dest;
1909         cpumask_t tmp;
1910
1911         cpus_and(tmp, mask, cpu_online_map);
1912         if (cpus_empty(tmp))
1913                 return;
1914
1915         if (assign_irq_vector(irq, mask))
1916                 return;
1917
1918         cpus_and(tmp, cfg->domain, mask);
1919         dest = cpu_mask_to_apicid(tmp);
1920
1921         read_msi_msg(irq, &msg);
1922
1923         msg.data &= ~MSI_DATA_VECTOR_MASK;
1924         msg.data |= MSI_DATA_VECTOR(cfg->vector);
1925         msg.address_lo &= ~MSI_ADDR_DEST_ID_MASK;
1926         msg.address_lo |= MSI_ADDR_DEST_ID(dest);
1927
1928         write_msi_msg(irq, &msg);
1929         irq_desc[irq].affinity = mask;
1930 }
1931 #endif /* CONFIG_SMP */
1932
1933 /*
1934  * IRQ Chip for MSI PCI/PCI-X/PCI-Express Devices,
1935  * which implement the MSI or MSI-X Capability Structure.
1936  */
1937 static struct irq_chip msi_chip = {
1938         .name           = "PCI-MSI",
1939         .unmask         = unmask_msi_irq,
1940         .mask           = mask_msi_irq,
1941         .ack            = ack_apic_edge,
1942 #ifdef CONFIG_SMP
1943         .set_affinity   = set_msi_irq_affinity,
1944 #endif
1945         .retrigger      = ioapic_retrigger_irq,
1946 };
1947
1948 int arch_setup_msi_irq(struct pci_dev *dev, struct msi_desc *desc)
1949 {
1950         struct msi_msg msg;
1951         int irq, ret;
1952         irq = create_irq();
1953         if (irq < 0)
1954                 return irq;
1955
1956         ret = msi_compose_msg(dev, irq, &msg);
1957         if (ret < 0) {
1958                 destroy_irq(irq);
1959                 return ret;
1960         }
1961
1962         set_irq_msi(irq, desc);
1963         write_msi_msg(irq, &msg);
1964
1965         set_irq_chip_and_handler_name(irq, &msi_chip, handle_edge_irq, "edge");
1966
1967         return 0;
1968 }
1969
1970 void arch_teardown_msi_irq(unsigned int irq)
1971 {
1972         destroy_irq(irq);
1973 }
1974
1975 #endif /* CONFIG_PCI_MSI */
1976
1977 /*
1978  * Hypertransport interrupt support
1979  */
1980 #ifdef CONFIG_HT_IRQ
1981
1982 #ifdef CONFIG_SMP
1983
1984 static void target_ht_irq(unsigned int irq, unsigned int dest, u8 vector)
1985 {
1986         struct ht_irq_msg msg;
1987         fetch_ht_irq_msg(irq, &msg);
1988
1989         msg.address_lo &= ~(HT_IRQ_LOW_VECTOR_MASK | HT_IRQ_LOW_DEST_ID_MASK);
1990         msg.address_hi &= ~(HT_IRQ_HIGH_DEST_ID_MASK);
1991
1992         msg.address_lo |= HT_IRQ_LOW_VECTOR(vector) | HT_IRQ_LOW_DEST_ID(dest);
1993         msg.address_hi |= HT_IRQ_HIGH_DEST_ID(dest);
1994
1995         write_ht_irq_msg(irq, &msg);
1996 }
1997
1998 static void set_ht_irq_affinity(unsigned int irq, cpumask_t mask)
1999 {
2000         struct irq_cfg *cfg = irq_cfg + irq;
2001         unsigned int dest;
2002         cpumask_t tmp;
2003
2004         cpus_and(tmp, mask, cpu_online_map);
2005         if (cpus_empty(tmp))
2006                 return;
2007
2008         if (assign_irq_vector(irq, mask))
2009                 return;
2010
2011         cpus_and(tmp, cfg->domain, mask);
2012         dest = cpu_mask_to_apicid(tmp);
2013
2014         target_ht_irq(irq, dest, cfg->vector);
2015         irq_desc[irq].affinity = mask;
2016 }
2017 #endif
2018
2019 static struct irq_chip ht_irq_chip = {
2020         .name           = "PCI-HT",
2021         .mask           = mask_ht_irq,
2022         .unmask         = unmask_ht_irq,
2023         .ack            = ack_apic_edge,
2024 #ifdef CONFIG_SMP
2025         .set_affinity   = set_ht_irq_affinity,
2026 #endif
2027         .retrigger      = ioapic_retrigger_irq,
2028 };
2029
2030 int arch_setup_ht_irq(unsigned int irq, struct pci_dev *dev)
2031 {
2032         struct irq_cfg *cfg = irq_cfg + irq;
2033         int err;
2034         cpumask_t tmp;
2035
2036         tmp = TARGET_CPUS;
2037         err = assign_irq_vector(irq, tmp);
2038         if (!err) {
2039                 struct ht_irq_msg msg;
2040                 unsigned dest;
2041
2042                 cpus_and(tmp, cfg->domain, tmp);
2043                 dest = cpu_mask_to_apicid(tmp);
2044
2045                 msg.address_hi = HT_IRQ_HIGH_DEST_ID(dest);
2046
2047                 msg.address_lo =
2048                         HT_IRQ_LOW_BASE |
2049                         HT_IRQ_LOW_DEST_ID(dest) |
2050                         HT_IRQ_LOW_VECTOR(cfg->vector) |
2051                         ((INT_DEST_MODE == 0) ?
2052                                 HT_IRQ_LOW_DM_PHYSICAL :
2053                                 HT_IRQ_LOW_DM_LOGICAL) |
2054                         HT_IRQ_LOW_RQEOI_EDGE |
2055                         ((INT_DELIVERY_MODE != dest_LowestPrio) ?
2056                                 HT_IRQ_LOW_MT_FIXED :
2057                                 HT_IRQ_LOW_MT_ARBITRATED) |
2058                         HT_IRQ_LOW_IRQ_MASKED;
2059
2060                 write_ht_irq_msg(irq, &msg);
2061
2062                 set_irq_chip_and_handler_name(irq, &ht_irq_chip,
2063                                               handle_edge_irq, "edge");
2064         }
2065         return err;
2066 }
2067 #endif /* CONFIG_HT_IRQ */
2068
2069 /* --------------------------------------------------------------------------
2070                           ACPI-based IOAPIC Configuration
2071    -------------------------------------------------------------------------- */
2072
2073 #ifdef CONFIG_ACPI
2074
2075 #define IO_APIC_MAX_ID          0xFE
2076
2077 int __init io_apic_get_redir_entries (int ioapic)
2078 {
2079         union IO_APIC_reg_01    reg_01;
2080         unsigned long flags;
2081
2082         spin_lock_irqsave(&ioapic_lock, flags);
2083         reg_01.raw = io_apic_read(ioapic, 1);
2084         spin_unlock_irqrestore(&ioapic_lock, flags);
2085
2086         return reg_01.bits.entries;
2087 }
2088
2089
2090 int io_apic_set_pci_routing (int ioapic, int pin, int irq, int triggering, int polarity)
2091 {
2092         if (!IO_APIC_IRQ(irq)) {
2093                 apic_printk(APIC_QUIET,KERN_ERR "IOAPIC[%d]: Invalid reference to IRQ 0\n",
2094                         ioapic);
2095                 return -EINVAL;
2096         }
2097
2098         /*
2099          * IRQs < 16 are already in the irq_2_pin[] map
2100          */
2101         if (irq >= 16)
2102                 add_pin_to_irq(irq, ioapic, pin);
2103
2104         setup_IO_APIC_irq(ioapic, pin, irq, triggering, polarity);
2105
2106         return 0;
2107 }
2108
2109 #endif /* CONFIG_ACPI */
2110
2111
2112 /*
2113  * This function currently is only a helper for the i386 smp boot process where
2114  * we need to reprogram the ioredtbls to cater for the cpus which have come online
2115  * so mask in all cases should simply be TARGET_CPUS
2116  */
2117 #ifdef CONFIG_SMP
2118 void __init setup_ioapic_dest(void)
2119 {
2120         int pin, ioapic, irq, irq_entry;
2121
2122         if (skip_ioapic_setup == 1)
2123                 return;
2124
2125         for (ioapic = 0; ioapic < nr_ioapics; ioapic++) {
2126                 for (pin = 0; pin < nr_ioapic_registers[ioapic]; pin++) {
2127                         irq_entry = find_irq_entry(ioapic, pin, mp_INT);
2128                         if (irq_entry == -1)
2129                                 continue;
2130                         irq = pin_2_irq(irq_entry, ioapic, pin);
2131
2132                         /* setup_IO_APIC_irqs could fail to get vector for some device
2133                          * when you have too many devices, because at that time only boot
2134                          * cpu is online.
2135                          */
2136                         if (!irq_cfg[irq].vector)
2137                                 setup_IO_APIC_irq(ioapic, pin, irq,
2138                                                   irq_trigger(irq_entry),
2139                                                   irq_polarity(irq_entry));
2140                         else
2141                                 set_ioapic_affinity_irq(irq, TARGET_CPUS);
2142                 }
2143
2144         }
2145 }
2146 #endif