arch/ia64/kernel/iosapic.c

   1 /*
   2  * I/O SAPIC support.
   3  *
   4  * Copyright (C) 1999 Intel Corp.
   5  * Copyright (C) 1999 Asit Mallick <asit.k.mallick@intel.com>
   6  * Copyright (C) 2000-2002 J.I. Lee <jung-ik.lee@intel.com>
   7  * Copyright (C) 1999-2000, 2002-2003 Hewlett-Packard Co.
   8  *      David Mosberger-Tang <davidm@hpl.hp.com>
   9  * Copyright (C) 1999 VA Linux Systems
  10  * Copyright (C) 1999,2000 Walt Drummond <drummond@valinux.com>
  11  *
  12  * 00/04/19     D. Mosberger    Rewritten to mirror more closely the x86 I/O
  13  *                              APIC code.  In particular, we now have separate
  14  *                              handlers for edge and level triggered
  15  *                              interrupts.
  16  * 00/10/27     Asit Mallick, Goutham Rao <goutham.rao@intel.com> IRQ vector
  17  *                              allocation PCI to vector mapping, shared PCI
  18  *                              interrupts.
  19  * 00/10/27     D. Mosberger    Document things a bit more to make them more
  20  *                              understandable.  Clean up much of the old
  21  *                              IOSAPIC cruft.
  22  * 01/07/27     J.I. Lee        PCI irq routing, Platform/Legacy interrupts
  23  *                              and fixes for ACPI S5(SoftOff) support.
  24  * 02/01/23     J.I. Lee        iosapic pgm fixes for PCI irq routing from _PRT
  25  * 02/01/07     E. Focht        <efocht@ess.nec.de> Redirectable interrupt
  26  *                              vectors in iosapic_set_affinity(),
  27  *                              initializations for /proc/irq/#/smp_affinity
  28  * 02/04/02     P. Diefenbaugh  Cleaned up ACPI PCI IRQ routing.
  29  * 02/04/18     J.I. Lee        bug fix in iosapic_init_pci_irq
  30  * 02/04/30     J.I. Lee        bug fix in find_iosapic to fix ACPI PCI IRQ to
  31  *                              IOSAPIC mapping error
  32  * 02/07/29     T. Kochi        Allocate interrupt vectors dynamically
  33  * 02/08/04     T. Kochi        Cleaned up terminology (irq, global system
  34  *                              interrupt, vector, etc.)
  35  * 02/09/20     D. Mosberger    Simplified by taking advantage of ACPI's
  36  *                              pci_irq code.
  37  * 03/02/19     B. Helgaas      Make pcat_compat system-wide, not per-IOSAPIC.
  38  *                              Remove iosapic_address & gsi_base from
  39  *                              external interfaces.  Rationalize
  40  *                              __init/__devinit attributes.
  41  * 04/12/04 Ashok Raj   <ashok.raj@intel.com> Intel Corporation 2004
  42  *                              Updated to work with irq migration necessary
  43  *                              for CPU Hotplug
  44  */
  45 /*
  46  * Here is what the interrupt logic between a PCI device and the kernel looks
  47  * like:
  48  *
  49  * (1) A PCI device raises one of the four interrupt pins (INTA, INTB, INTC,
  50  *     INTD).  The device is uniquely identified by its bus-, and slot-number
  51  *     (the function number does not matter here because all functions share
  52  *     the same interrupt lines).
  53  *
  54  * (2) The motherboard routes the interrupt line to a pin on a IOSAPIC
  55  *     controller.  Multiple interrupt lines may have to share the same
  56  *     IOSAPIC pin (if they're level triggered and use the same polarity).
  57  *     Each interrupt line has a unique Global System Interrupt (GSI) number
  58  *     which can be calculated as the sum of the controller's base GSI number
  59  *     and the IOSAPIC pin number to which the line connects.
  60  *
  61  * (3) The IOSAPIC uses an internal routing table entries (RTEs) to map the
  62  * IOSAPIC pin into the IA-64 interrupt vector.  This interrupt vector is then
  63  * sent to the CPU.
  64  *
  65  * (4) The kernel recognizes an interrupt as an IRQ.  The IRQ interface is
  66  *     used as architecture-independent interrupt handling mechanism in Linux.
  67  *     As an IRQ is a number, we have to have
  68  *     IA-64 interrupt vector number <-> IRQ number mapping.  On smaller
  69  *     systems, we use one-to-one mapping between IA-64 vector and IRQ.  A
  70  *     platform can implement platform_irq_to_vector(irq) and
  71  *     platform_local_vector_to_irq(vector) APIs to differentiate the mapping.
  72  *     Please see also include/asm-ia64/hw_irq.h for those APIs.
  73  *
  74  * To sum up, there are three levels of mappings involved:
  75  *
  76  *      PCI pin -> global system interrupt (GSI) -> IA-64 vector <-> IRQ
  77  *
  78  * Note: The term "IRQ" is loosely used everywhere in Linux kernel to
  79  * describeinterrupts.  Now we use "IRQ" only for Linux IRQ's.  ISA IRQ
  80  * (isa_irq) is the only exception in this source code.
  81  */
  82
  83 #include <linux/acpi.h>
  84 #include <linux/init.h>
  85 #include <linux/irq.h>
  86 #include <linux/kernel.h>
  87 #include <linux/list.h>
  88 #include <linux/pci.h>
  89 #include <linux/smp.h>
  90 #include <linux/string.h>
  91 #include <linux/bootmem.h>
  92
  93 #include <asm/delay.h>
  94 #include <asm/hw_irq.h>
  95 #include <asm/io.h>
  96 #include <asm/iosapic.h>
  97 #include <asm/machvec.h>
  98 #include <asm/processor.h>
  99 #include <asm/ptrace.h>
 100 #include <asm/system.h>
 101
 102 #undef DEBUG_INTERRUPT_ROUTING
 103
 104 #ifdef DEBUG_INTERRUPT_ROUTING
 105 #define DBG(fmt...)     printk(fmt)
 106 #else
 107 #define DBG(fmt...)
 108 #endif
 109
 110 #define NR_PREALLOCATE_RTE_ENTRIES \
 111         (PAGE_SIZE / sizeof(struct iosapic_rte_info))
 112 #define RTE_PREALLOCATED        (1)
 113
 114 static DEFINE_SPINLOCK(iosapic_lock);
 115
 116 /*
 117  * These tables map IA-64 vectors to the IOSAPIC pin that generates this
 118  * vector.
 119  */
 120
 121 #define NO_REF_RTE      0
 122
 123 static struct iosapic {
 124         char __iomem    *addr;          /* base address of IOSAPIC */
 125         unsigned int    gsi_base;       /* GSI base */
 126         unsigned short  num_rte;        /* # of RTEs on this IOSAPIC */
 127         int             rtes_inuse;     /* # of RTEs in use on this IOSAPIC */
 128 #ifdef CONFIG_NUMA
 129         unsigned short  node;           /* numa node association via pxm */
 130 #endif
 131         spinlock_t      lock;           /* lock for indirect reg access */
 132 } iosapic_lists[NR_IOSAPICS];
 133
 134 struct iosapic_rte_info {
 135         struct list_head rte_list;      /* RTEs sharing the same vector */
 136         char            rte_index;      /* IOSAPIC RTE index */
 137         int             refcnt;         /* reference counter */
 138         unsigned int    flags;          /* flags */
 139         struct iosapic  *iosapic;
 140 } ____cacheline_aligned;
 141
 142 static struct iosapic_intr_info {
 143         struct list_head rtes;          /* RTEs using this vector (empty =>
 144                                          * not an IOSAPIC interrupt) */
 145         int             count;          /* # of RTEs that shares this vector */
 146         u32             low32;          /* current value of low word of
 147                                          * Redirection table entry */
 148         unsigned int    dest;           /* destination CPU physical ID */
 149         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
 150         unsigned char   polarity: 1;    /* interrupt polarity
 151                                          * (see iosapic.h) */
 152         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
 153 } iosapic_intr_info[NR_IRQS];
 154
 155 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
 156
 157 static int iosapic_kmalloc_ok;
 158 static LIST_HEAD(free_rte_list);
 159
 160 static inline void
 161 iosapic_write(struct iosapic *iosapic, unsigned int reg, u32 val)
 162 {
 163         unsigned long flags;
 164
 165         spin_lock_irqsave(&iosapic->lock, flags);
 166         __iosapic_write(iosapic->addr, reg, val);
 167         spin_unlock_irqrestore(&iosapic->lock, flags);
 168 }
 169
 170 /*
 171  * Find an IOSAPIC associated with a GSI
 172  */
 173 static inline int
 174 find_iosapic (unsigned int gsi)
 175 {
 176         int i;
 177
 178         for (i = 0; i < NR_IOSAPICS; i++) {
 179                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
 180                     iosapic_lists[i].num_rte)
 181                         return i;
 182         }
 183
 184         return -1;
 185 }
 186
 187 static inline int __gsi_to_irq(unsigned int gsi)
 188 {
 189         int irq;
 190         struct iosapic_intr_info *info;
 191         struct iosapic_rte_info *rte;
 192
 193         for (irq = 0; irq < NR_IRQS; irq++) {
 194                 info = &iosapic_intr_info[irq];
 195                 list_for_each_entry(rte, &info->rtes, rte_list)
 196                         if (rte->iosapic->gsi_base + rte->rte_index == gsi)
 197                                 return irq;
 198         }
 199         return -1;
 200 }
 201
 202 /*
 203  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
 204  * entry exists, return -1.
 205  */
 206 inline int
 207 gsi_to_vector (unsigned int gsi)
 208 {
 209         int irq = __gsi_to_irq(gsi);
 210         if (check_irq_used(irq) < 0)
 211                 return -1;
 212         return irq_to_vector(irq);
 213 }
 214
 215 int
 216 gsi_to_irq (unsigned int gsi)
 217 {
 218         unsigned long flags;
 219         int irq;
 220
 221         spin_lock_irqsave(&iosapic_lock, flags);
 222         irq = __gsi_to_irq(gsi);
 223         spin_unlock_irqrestore(&iosapic_lock, flags);
 224         return irq;
 225 }
 226
 227 static struct iosapic_rte_info *find_rte(unsigned int irq, unsigned int gsi)
 228 {
 229         struct iosapic_rte_info *rte;
 230
 231         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
 232                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
 233                         return rte;
 234         return NULL;
 235 }
 236
 237 static void
 238 set_rte (unsigned int gsi, unsigned int irq, unsigned int dest, int mask)
 239 {
 240         unsigned long pol, trigger, dmode;
 241         u32 low32, high32;
 242         int rte_index;
 243         char redir;
 244         struct iosapic_rte_info *rte;
 245         ia64_vector vector = irq_to_vector(irq);
 246
 247         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
 248
 249         rte = find_rte(irq, gsi);
 250         if (!rte)
 251                 return;         /* not an IOSAPIC interrupt */
 252
 253         rte_index = rte->rte_index;
 254         pol     = iosapic_intr_info[irq].polarity;
 255         trigger = iosapic_intr_info[irq].trigger;
 256         dmode   = iosapic_intr_info[irq].dmode;
 257
 258         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
 259
 260 #ifdef CONFIG_SMP
 261         set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
 262 #endif
 263
 264         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
 265                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
 266                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
 267                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
 268                  vector);
 269
 270         /* dest contains both id and eid */
 271         high32 = (dest << IOSAPIC_DEST_SHIFT);
 272
 273         iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
 274         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 275         iosapic_intr_info[irq].low32 = low32;
 276         iosapic_intr_info[irq].dest = dest;
 277 }
 278
 279 static void
 280 nop (unsigned int irq)
 281 {
 282         /* do nothing... */
 283 }
 284
 285
 286 #ifdef CONFIG_KEXEC
 287 void
 288 kexec_disable_iosapic(void)
 289 {
 290         struct iosapic_intr_info *info;
 291         struct iosapic_rte_info *rte;
 292         ia64_vector vec;
 293         int irq;
 294
 295         for (irq = 0; irq < NR_IRQS; irq++) {
 296                 info = &iosapic_intr_info[irq];
 297                 vec = irq_to_vector(irq);
 298                 list_for_each_entry(rte, &info->rtes,
 299                                 rte_list) {
 300                         iosapic_write(rte->iosapic,
 301                                         IOSAPIC_RTE_LOW(rte->rte_index),
 302                                         IOSAPIC_MASK|vec);
 303                         iosapic_eoi(rte->iosapic->addr, vec);
 304                 }
 305         }
 306 }
 307 #endif
 308
 309 static void
 310 mask_irq (unsigned int irq)
 311 {
 312         u32 low32;
 313         int rte_index;
 314         struct iosapic_rte_info *rte;
 315
 316         if (list_empty(&iosapic_intr_info[irq].rtes))
 317                 return;                 /* not an IOSAPIC interrupt! */
 318
 319         /* set only the mask bit */
 320         low32 = iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
 321         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
 322                 rte_index = rte->rte_index;
 323                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 324         }
 325 }
 326
 327 static void
 328 unmask_irq (unsigned int irq)
 329 {
 330         u32 low32;
 331         int rte_index;
 332         struct iosapic_rte_info *rte;
 333
 334         if (list_empty(&iosapic_intr_info[irq].rtes))
 335                 return;                 /* not an IOSAPIC interrupt! */
 336
 337         low32 = iosapic_intr_info[irq].low32 &= ~IOSAPIC_MASK;
 338         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
 339                 rte_index = rte->rte_index;
 340                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 341         }
 342 }
 343
 344
 345 static void
 346 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
 347 {
 348 #ifdef CONFIG_SMP
 349         u32 high32, low32;
 350         int dest, rte_index;
 351         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
 352         struct iosapic_rte_info *rte;
 353         struct iosapic *iosapic;
 354
 355         irq &= (~IA64_IRQ_REDIRECTED);
 356
 357         cpus_and(mask, mask, cpu_online_map);
 358         if (cpus_empty(mask))
 359                 return;
 360
 361         if (reassign_irq_vector(irq, first_cpu(mask)))
 362                 return;
 363
 364         dest = cpu_physical_id(first_cpu(mask));
 365
 366         if (list_empty(&iosapic_intr_info[irq].rtes))
 367                 return;                 /* not an IOSAPIC interrupt */
 368
 369         set_irq_affinity_info(irq, dest, redir);
 370
 371         /* dest contains both id and eid */
 372         high32 = dest << IOSAPIC_DEST_SHIFT;
 373
 374         low32 = iosapic_intr_info[irq].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
 375         if (redir)
 376                 /* change delivery mode to lowest priority */
 377                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
 378         else
 379                 /* change delivery mode to fixed */
 380                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
 381         low32 &= IOSAPIC_VECTOR_MASK;
 382         low32 |= irq_to_vector(irq);
 383
 384         iosapic_intr_info[irq].low32 = low32;
 385         iosapic_intr_info[irq].dest = dest;
 386         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
 387                 iosapic = rte->iosapic;
 388                 rte_index = rte->rte_index;
 389                 iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
 390                 iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 391         }
 392 #endif
 393 }
 394
 395 /*
 396  * Handlers for level-triggered interrupts.
 397  */
 398
 399 static unsigned int
 400 iosapic_startup_level_irq (unsigned int irq)
 401 {
 402         unmask_irq(irq);
 403         return 0;
 404 }
 405
 406 static void
 407 iosapic_end_level_irq (unsigned int irq)
 408 {
 409         ia64_vector vec = irq_to_vector(irq);
 410         struct iosapic_rte_info *rte;
 411         int do_unmask_irq = 0;
 412
 413         if (unlikely(irq_desc[irq].status & IRQ_MOVE_PENDING)) {
 414                 do_unmask_irq = 1;
 415                 mask_irq(irq);
 416         }
 417
 418         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
 419                 iosapic_eoi(rte->iosapic->addr, vec);
 420
 421         if (unlikely(do_unmask_irq)) {
 422                 move_masked_irq(irq);
 423                 unmask_irq(irq);
 424         }
 425 }
 426
 427 #define iosapic_shutdown_level_irq      mask_irq
 428 #define iosapic_enable_level_irq        unmask_irq
 429 #define iosapic_disable_level_irq       mask_irq
 430 #define iosapic_ack_level_irq           nop
 431
 432 struct irq_chip irq_type_iosapic_level = {
 433         .name =         "IO-SAPIC-level",
 434         .startup =      iosapic_startup_level_irq,
 435         .shutdown =     iosapic_shutdown_level_irq,
 436         .enable =       iosapic_enable_level_irq,
 437         .disable =      iosapic_disable_level_irq,
 438         .ack =          iosapic_ack_level_irq,
 439         .end =          iosapic_end_level_irq,
 440         .mask =         mask_irq,
 441         .unmask =       unmask_irq,
 442         .set_affinity = iosapic_set_affinity
 443 };
 444
 445 /*
 446  * Handlers for edge-triggered interrupts.
 447  */
 448
 449 static unsigned int
 450 iosapic_startup_edge_irq (unsigned int irq)
 451 {
 452         unmask_irq(irq);
 453         /*
 454          * IOSAPIC simply drops interrupts pended while the
 455          * corresponding pin was masked, so we can't know if an
 456          * interrupt is pending already.  Let's hope not...
 457          */
 458         return 0;
 459 }
 460
 461 static void
 462 iosapic_ack_edge_irq (unsigned int irq)
 463 {
 464         irq_desc_t *idesc = irq_desc + irq;
 465
 466         move_native_irq(irq);
 467         /*
 468          * Once we have recorded IRQ_PENDING already, we can mask the
 469          * interrupt for real. This prevents IRQ storms from unhandled
 470          * devices.
 471          */
 472         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
 473             (IRQ_PENDING|IRQ_DISABLED))
 474                 mask_irq(irq);
 475 }
 476
 477 #define iosapic_enable_edge_irq         unmask_irq
 478 #define iosapic_disable_edge_irq        nop
 479 #define iosapic_end_edge_irq            nop
 480
 481 struct irq_chip irq_type_iosapic_edge = {
 482         .name =         "IO-SAPIC-edge",
 483         .startup =      iosapic_startup_edge_irq,
 484         .shutdown =     iosapic_disable_edge_irq,
 485         .enable =       iosapic_enable_edge_irq,
 486         .disable =      iosapic_disable_edge_irq,
 487         .ack =          iosapic_ack_edge_irq,
 488         .end =          iosapic_end_edge_irq,
 489         .mask =         mask_irq,
 490         .unmask =       unmask_irq,
 491         .set_affinity = iosapic_set_affinity
 492 };
 493
 494 unsigned int
 495 iosapic_version (char __iomem *addr)
 496 {
 497         /*
 498          * IOSAPIC Version Register return 32 bit structure like:
 499          * {
 500          *      unsigned int version   : 8;
 501          *      unsigned int reserved1 : 8;
 502          *      unsigned int max_redir : 8;
 503          *      unsigned int reserved2 : 8;
 504          * }
 505          */
 506         return __iosapic_read(addr, IOSAPIC_VERSION);
 507 }
 508
 509 static int iosapic_find_sharable_irq(unsigned long trigger, unsigned long pol)
 510 {
 511         int i, irq = -ENOSPC, min_count = -1;
 512         struct iosapic_intr_info *info;
 513
 514         /*
 515          * shared vectors for edge-triggered interrupts are not
 516          * supported yet
 517          */
 518         if (trigger == IOSAPIC_EDGE)
 519                 return -EINVAL;
 520
 521         for (i = 0; i <= NR_IRQS; i++) {
 522                 info = &iosapic_intr_info[i];
 523                 if (info->trigger == trigger && info->polarity == pol &&
 524                     (info->dmode == IOSAPIC_FIXED ||
 525                      info->dmode == IOSAPIC_LOWEST_PRIORITY) &&
 526                     can_request_irq(i, IRQF_SHARED)) {
 527                         if (min_count == -1 || info->count < min_count) {
 528                                 irq = i;
 529                                 min_count = info->count;
 530                         }
 531                 }
 532         }
 533         return irq;
 534 }
 535
 536 /*
 537  * if the given vector is already owned by other,
 538  *  assign a new vector for the other and make the vector available
 539  */
 540 static void __init
 541 iosapic_reassign_vector (int irq)
 542 {
 543         int new_irq;
 544
 545         if (!list_empty(&iosapic_intr_info[irq].rtes)) {
 546                 new_irq = create_irq();
 547                 if (new_irq < 0)
 548                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
 549                 printk(KERN_INFO "Reassigning vector %d to %d\n",
 550                        irq_to_vector(irq), irq_to_vector(new_irq));
 551                 memcpy(&iosapic_intr_info[new_irq], &iosapic_intr_info[irq],
 552                        sizeof(struct iosapic_intr_info));
 553                 INIT_LIST_HEAD(&iosapic_intr_info[new_irq].rtes);
 554                 list_move(iosapic_intr_info[irq].rtes.next,
 555                           &iosapic_intr_info[new_irq].rtes);
 556                 memset(&iosapic_intr_info[irq], 0,
 557                        sizeof(struct iosapic_intr_info));
 558                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
 559                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
 560         }
 561 }
 562
 563 static struct iosapic_rte_info *iosapic_alloc_rte (void)
 564 {
 565         int i;
 566         struct iosapic_rte_info *rte;
 567         int preallocated = 0;
 568
 569         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
 570                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
 571                                     NR_PREALLOCATE_RTE_ENTRIES);
 572                 if (!rte)
 573                         return NULL;
 574                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
 575                         list_add(&rte->rte_list, &free_rte_list);
 576         }
 577
 578         if (!list_empty(&free_rte_list)) {
 579                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
 580                                  rte_list);
 581                 list_del(&rte->rte_list);
 582                 preallocated++;
 583         } else {
 584                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
 585                 if (!rte)
 586                         return NULL;
 587         }
 588
 589         memset(rte, 0, sizeof(struct iosapic_rte_info));
 590         if (preallocated)
 591                 rte->flags |= RTE_PREALLOCATED;
 592
 593         return rte;
 594 }
 595
 596 static void iosapic_free_rte (struct iosapic_rte_info *rte)
 597 {
 598         if (rte->flags & RTE_PREALLOCATED)
 599                 list_add_tail(&rte->rte_list, &free_rte_list);
 600         else
 601                 kfree(rte);
 602 }
 603
 604 static inline int irq_is_shared (int irq)
 605 {
 606         return (iosapic_intr_info[irq].count > 1);
 607 }
 608
 609 static int
 610 register_intr (unsigned int gsi, int irq, unsigned char delivery,
 611                unsigned long polarity, unsigned long trigger)
 612 {
 613         irq_desc_t *idesc;
 614         struct hw_interrupt_type *irq_type;
 615         int index;
 616         struct iosapic_rte_info *rte;
 617
 618         index = find_iosapic(gsi);
 619         if (index < 0) {
 620                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
 621                        __FUNCTION__, gsi);
 622                 return -ENODEV;
 623         }
 624
 625         rte = find_rte(irq, gsi);
 626         if (!rte) {
 627                 rte = iosapic_alloc_rte();
 628                 if (!rte) {
 629                         printk(KERN_WARNING "%s: cannot allocate memory\n",
 630                                __FUNCTION__);
 631                         return -ENOMEM;
 632                 }
 633
 634                 rte->iosapic    = &iosapic_lists[index];
 635                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
 636                 rte->refcnt++;
 637                 list_add_tail(&rte->rte_list, &iosapic_intr_info[irq].rtes);
 638                 iosapic_intr_info[irq].count++;
 639                 iosapic_lists[index].rtes_inuse++;
 640         }
 641         else if (rte->refcnt == NO_REF_RTE) {
 642                 struct iosapic_intr_info *info = &iosapic_intr_info[irq];
 643                 if (info->count > 0 &&
 644                     (info->trigger != trigger || info->polarity != polarity)){
 645                         printk (KERN_WARNING
 646                                 "%s: cannot override the interrupt\n",
 647                                 __FUNCTION__);
 648                         return -EINVAL;
 649                 }
 650                 rte->refcnt++;
 651                 iosapic_intr_info[irq].count++;
 652                 iosapic_lists[index].rtes_inuse++;
 653         }
 654
 655         iosapic_intr_info[irq].polarity = polarity;
 656         iosapic_intr_info[irq].dmode    = delivery;
 657         iosapic_intr_info[irq].trigger  = trigger;
 658
 659         if (trigger == IOSAPIC_EDGE)
 660                 irq_type = &irq_type_iosapic_edge;
 661         else
 662                 irq_type = &irq_type_iosapic_level;
 663
 664         idesc = irq_desc + irq;
 665         if (idesc->chip != irq_type) {
 666                 if (idesc->chip != &no_irq_type)
 667                         printk(KERN_WARNING
 668                                "%s: changing vector %d from %s to %s\n",
 669                                __FUNCTION__, irq_to_vector(irq),
 670                                idesc->chip->name, irq_type->name);
 671                 idesc->chip = irq_type;
 672         }
 673         return 0;
 674 }
 675
 676 static unsigned int
 677 get_target_cpu (unsigned int gsi, int irq)
 678 {
 679 #ifdef CONFIG_SMP
 680         static int cpu = -1;
 681         extern int cpe_vector;
 682         cpumask_t domain = irq_to_domain(irq);
 683
 684         /*
 685          * In case of vector shared by multiple RTEs, all RTEs that
 686          * share the vector need to use the same destination CPU.
 687          */
 688         if (!list_empty(&iosapic_intr_info[irq].rtes))
 689                 return iosapic_intr_info[irq].dest;
 690
 691         /*
 692          * If the platform supports redirection via XTP, let it
 693          * distribute interrupts.
 694          */
 695         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
 696                 return cpu_physical_id(smp_processor_id());
 697
 698         /*
 699          * Some interrupts (ACPI SCI, for instance) are registered
 700          * before the BSP is marked as online.
 701          */
 702         if (!cpu_online(smp_processor_id()))
 703                 return cpu_physical_id(smp_processor_id());
 704
 705 #ifdef CONFIG_ACPI
 706         if (cpe_vector > 0 && irq_to_vector(irq) == IA64_CPEP_VECTOR)
 707                 return get_cpei_target_cpu();
 708 #endif
 709
 710 #ifdef CONFIG_NUMA
 711         {
 712                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
 713                 cpumask_t cpu_mask;
 714
 715                 iosapic_index = find_iosapic(gsi);
 716                 if (iosapic_index < 0 ||
 717                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
 718                         goto skip_numa_setup;
 719
 720                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
 721                 cpus_and(cpu_mask, cpu_mask, domain);
 722                 for_each_cpu_mask(numa_cpu, cpu_mask) {
 723                         if (!cpu_online(numa_cpu))
 724                                 cpu_clear(numa_cpu, cpu_mask);
 725                 }
 726
 727                 num_cpus = cpus_weight(cpu_mask);
 728
 729                 if (!num_cpus)
 730                         goto skip_numa_setup;
 731
 732                 /* Use irq assignment to distribute across cpus in node */
 733                 cpu_index = irq % num_cpus;
 734
 735                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
 736                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
 737
 738                 if (numa_cpu != NR_CPUS)
 739                         return cpu_physical_id(numa_cpu);
 740         }
 741 skip_numa_setup:
 742 #endif
 743         /*
 744          * Otherwise, round-robin interrupt vectors across all the
 745          * processors.  (It'd be nice if we could be smarter in the
 746          * case of NUMA.)
 747          */
 748         do {
 749                 if (++cpu >= NR_CPUS)
 750                         cpu = 0;
 751         } while (!cpu_online(cpu) || !cpu_isset(cpu, domain));
 752
 753         return cpu_physical_id(cpu);
 754 #else  /* CONFIG_SMP */
 755         return cpu_physical_id(smp_processor_id());
 756 #endif
 757 }
 758
 759 /*
 760  * ACPI can describe IOSAPIC interrupts via static tables and namespace
 761  * methods.  This provides an interface to register those interrupts and
 762  * program the IOSAPIC RTE.
 763  */
 764 int
 765 iosapic_register_intr (unsigned int gsi,
 766                        unsigned long polarity, unsigned long trigger)
 767 {
 768         int irq, mask = 1, err;
 769         unsigned int dest;
 770         unsigned long flags;
 771         struct iosapic_rte_info *rte;
 772         u32 low32;
 773
 774         /*
 775          * If this GSI has already been registered (i.e., it's a
 776          * shared interrupt, or we lost a race to register it),
 777          * don't touch the RTE.
 778          */
 779         spin_lock_irqsave(&iosapic_lock, flags);
 780         irq = __gsi_to_irq(gsi);
 781         if (irq > 0) {
 782                 rte = find_rte(irq, gsi);
 783                 if(iosapic_intr_info[irq].count == 0) {
 784                         assign_irq_vector(irq);
 785                         dynamic_irq_init(irq);
 786                 } else if (rte->refcnt != NO_REF_RTE) {
 787                         rte->refcnt++;
 788                         goto unlock_iosapic_lock;
 789                 }
 790         } else
 791                 irq = create_irq();
 792
 793         /* If vector is running out, we try to find a sharable vector */
 794         if (irq < 0) {
 795                 irq = iosapic_find_sharable_irq(trigger, polarity);
 796                 if (irq < 0)
 797                         goto unlock_iosapic_lock;
 798         }
 799
 800         spin_lock(&irq_desc[irq].lock);
 801         dest = get_target_cpu(gsi, irq);
 802         err = register_intr(gsi, irq, IOSAPIC_LOWEST_PRIORITY,
 803                             polarity, trigger);
 804         if (err < 0) {
 805                 irq = err;
 806                 goto unlock_all;
 807         }
 808
 809         /*
 810          * If the vector is shared and already unmasked for other
 811          * interrupt sources, don't mask it.
 812          */
 813         low32 = iosapic_intr_info[irq].low32;
 814         if (irq_is_shared(irq) && !(low32 & IOSAPIC_MASK))
 815                 mask = 0;
 816         set_rte(gsi, irq, dest, mask);
 817
 818         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
 819                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 820                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 821                cpu_logical_id(dest), dest, irq_to_vector(irq));
 822  unlock_all:
 823         spin_unlock(&irq_desc[irq].lock);
 824  unlock_iosapic_lock:
 825         spin_unlock_irqrestore(&iosapic_lock, flags);
 826         return irq;
 827 }
 828
 829 void
 830 iosapic_unregister_intr (unsigned int gsi)
 831 {
 832         unsigned long flags;
 833         int irq, index;
 834         irq_desc_t *idesc;
 835         u32 low32;
 836         unsigned long trigger, polarity;
 837         unsigned int dest;
 838         struct iosapic_rte_info *rte;
 839
 840         /*
 841          * If the irq associated with the gsi is not found,
 842          * iosapic_unregister_intr() is unbalanced. We need to check
 843          * this again after getting locks.
 844          */
 845         irq = gsi_to_irq(gsi);
 846         if (irq < 0) {
 847                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
 848                        gsi);
 849                 WARN_ON(1);
 850                 return;
 851         }
 852
 853         spin_lock_irqsave(&iosapic_lock, flags);
 854         if ((rte = find_rte(irq, gsi)) == NULL) {
 855                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
 856                        gsi);
 857                 WARN_ON(1);
 858                 goto out;
 859         }
 860
 861         if (--rte->refcnt > 0)
 862                 goto out;
 863
 864         idesc = irq_desc + irq;
 865         rte->refcnt = NO_REF_RTE;
 866
 867         /* Mask the interrupt */
 868         low32 = iosapic_intr_info[irq].low32 | IOSAPIC_MASK;
 869         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
 870
 871         iosapic_intr_info[irq].count--;
 872         index = find_iosapic(gsi);
 873         iosapic_lists[index].rtes_inuse--;
 874         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
 875
 876         trigger  = iosapic_intr_info[irq].trigger;
 877         polarity = iosapic_intr_info[irq].polarity;
 878         dest     = iosapic_intr_info[irq].dest;
 879         printk(KERN_INFO
 880                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
 881                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 882                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 883                cpu_logical_id(dest), dest, irq_to_vector(irq));
 884
 885         if (iosapic_intr_info[irq].count == 0) {
 886 #ifdef CONFIG_SMP
 887                 /* Clear affinity */
 888                 cpus_setall(idesc->affinity);
 889 #endif
 890                 /* Clear the interrupt information */
 891                 iosapic_intr_info[irq].dest = 0;
 892                 iosapic_intr_info[irq].dmode = 0;
 893                 iosapic_intr_info[irq].polarity = 0;
 894                 iosapic_intr_info[irq].trigger = 0;
 895                 iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
 896
 897                 /* Destroy and reserve IRQ */
 898                 destroy_and_reserve_irq(irq);
 899         }
 900  out:
 901         spin_unlock_irqrestore(&iosapic_lock, flags);
 902 }
 903
 904 /*
 905  * ACPI calls this when it finds an entry for a platform interrupt.
 906  */
 907 int __init
 908 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
 909                                 int iosapic_vector, u16 eid, u16 id,
 910                                 unsigned long polarity, unsigned long trigger)
 911 {
 912         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
 913         unsigned char delivery;
 914         int irq, vector, mask = 0;
 915         unsigned int dest = ((id << 8) | eid) & 0xffff;
 916
 917         switch (int_type) {
 918               case ACPI_INTERRUPT_PMI:
 919                 irq = vector = iosapic_vector;
 920                 bind_irq_vector(irq, vector, CPU_MASK_ALL);
 921                 /*
 922                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
 923                  * we need to make sure the vector is available
 924                  */
 925                 iosapic_reassign_vector(irq);
 926                 delivery = IOSAPIC_PMI;
 927                 break;
 928               case ACPI_INTERRUPT_INIT:
 929                 irq = create_irq();
 930                 if (irq < 0)
 931                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
 932                 vector = irq_to_vector(irq);
 933                 delivery = IOSAPIC_INIT;
 934                 break;
 935               case ACPI_INTERRUPT_CPEI:
 936                 irq = vector = IA64_CPE_VECTOR;
 937                 BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
 938                 delivery = IOSAPIC_LOWEST_PRIORITY;
 939                 mask = 1;
 940                 break;
 941               default:
 942                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __FUNCTION__,
 943                        int_type);
 944                 return -1;
 945         }
 946
 947         register_intr(gsi, irq, delivery, polarity, trigger);
 948
 949         printk(KERN_INFO
 950                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
 951                " vector %d\n",
 952                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
 953                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 954                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 955                cpu_logical_id(dest), dest, vector);
 956
 957         set_rte(gsi, irq, dest, mask);
 958         return vector;
 959 }
 960
 961 /*
 962  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
 963  */
 964 void __devinit
 965 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
 966                           unsigned long polarity,
 967                           unsigned long trigger)
 968 {
 969         int vector, irq;
 970         unsigned int dest = cpu_physical_id(smp_processor_id());
 971
 972         irq = vector = isa_irq_to_vector(isa_irq);
 973         BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
 974         register_intr(gsi, irq, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
 975
 976         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
 977             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
 978             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
 979             cpu_logical_id(dest), dest, vector);
 980
 981         set_rte(gsi, irq, dest, 1);
 982 }
 983
 984 void __init
 985 iosapic_system_init (int system_pcat_compat)
 986 {
 987         int irq;
 988
 989         for (irq = 0; irq < NR_IRQS; ++irq) {
 990                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
 991                 /* mark as unused */
 992                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
 993
 994                 iosapic_intr_info[irq].count = 0;
 995         }
 996
 997         pcat_compat = system_pcat_compat;
 998         if (pcat_compat) {
 999                 /*
1000                  * Disable the compatibility mode interrupts (8259 style),
1001                  * needs IN/OUT support enabled.
1002                  */
1003                 printk(KERN_INFO
1004                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
1005                        __FUNCTION__);
1006                 outb(0xff, 0xA1);
1007                 outb(0xff, 0x21);
1008         }
1009 }
1010
1011 static inline int
1012 iosapic_alloc (void)
1013 {
1014         int index;
1015
1016         for (index = 0; index < NR_IOSAPICS; index++)
1017                 if (!iosapic_lists[index].addr)
1018                         return index;
1019
1020         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
1021         return -1;
1022 }
1023
1024 static inline void
1025 iosapic_free (int index)
1026 {
1027         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1028 }
1029
1030 static inline int
1031 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1032 {
1033         int index;
1034         unsigned int gsi_end, base, end;
1035
1036         /* check gsi range */
1037         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1038         for (index = 0; index < NR_IOSAPICS; index++) {
1039                 if (!iosapic_lists[index].addr)
1040                         continue;
1041
1042                 base = iosapic_lists[index].gsi_base;
1043                 end  = base + iosapic_lists[index].num_rte - 1;
1044
1045                 if (gsi_end < base || end < gsi_base)
1046                         continue; /* OK */
1047
1048                 return -EBUSY;
1049         }
1050         return 0;
1051 }
1052
1053 int __devinit
1054 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1055 {
1056         int num_rte, err, index;
1057         unsigned int isa_irq, ver;
1058         char __iomem *addr;
1059         unsigned long flags;
1060
1061         spin_lock_irqsave(&iosapic_lock, flags);
1062         index = find_iosapic(gsi_base);
1063         if (index >= 0) {
1064                 spin_unlock_irqrestore(&iosapic_lock, flags);
1065                 return -EBUSY;
1066         }
1067
1068         addr = ioremap(phys_addr, 0);
1069         ver = iosapic_version(addr);
1070         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1071                 iounmap(addr);
1072                 spin_unlock_irqrestore(&iosapic_lock, flags);
1073                 return err;
1074         }
1075
1076         /*
1077          * The MAX_REDIR register holds the highest input pin number
1078          * (starting from 0).  We add 1 so that we can use it for
1079          * number of pins (= RTEs)
1080          */
1081         num_rte = ((ver >> 16) & 0xff) + 1;
1082
1083         index = iosapic_alloc();
1084         iosapic_lists[index].addr = addr;
1085         iosapic_lists[index].gsi_base = gsi_base;
1086         iosapic_lists[index].num_rte = num_rte;
1087 #ifdef CONFIG_NUMA
1088         iosapic_lists[index].node = MAX_NUMNODES;
1089 #endif
1090         spin_lock_init(&iosapic_lists[index].lock);
1091         spin_unlock_irqrestore(&iosapic_lock, flags);
1092
1093         if ((gsi_base == 0) && pcat_compat) {
1094                 /*
1095                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1096                  * these may get reprogrammed later on with data from the ACPI
1097                  * Interrupt Source Override table.
1098                  */
1099                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1100                         iosapic_override_isa_irq(isa_irq, isa_irq,
1101                                                  IOSAPIC_POL_HIGH,
1102                                                  IOSAPIC_EDGE);
1103         }
1104         return 0;
1105 }
1106
1107 #ifdef CONFIG_HOTPLUG
1108 int
1109 iosapic_remove (unsigned int gsi_base)
1110 {
1111         int index, err = 0;
1112         unsigned long flags;
1113
1114         spin_lock_irqsave(&iosapic_lock, flags);
1115         index = find_iosapic(gsi_base);
1116         if (index < 0) {
1117                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1118                        __FUNCTION__, gsi_base);
1119                 goto out;
1120         }
1121
1122         if (iosapic_lists[index].rtes_inuse) {
1123                 err = -EBUSY;
1124                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1125                        __FUNCTION__, gsi_base);
1126                 goto out;
1127         }
1128
1129         iounmap(iosapic_lists[index].addr);
1130         iosapic_free(index);
1131  out:
1132         spin_unlock_irqrestore(&iosapic_lock, flags);
1133         return err;
1134 }
1135 #endif /* CONFIG_HOTPLUG */
1136
1137 #ifdef CONFIG_NUMA
1138 void __devinit
1139 map_iosapic_to_node(unsigned int gsi_base, int node)
1140 {
1141         int index;
1142
1143         index = find_iosapic(gsi_base);
1144         if (index < 0) {
1145                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1146                        __FUNCTION__, gsi_base);
1147                 return;
1148         }
1149         iosapic_lists[index].node = node;
1150         return;
1151 }
1152 #endif
1153
1154 static int __init iosapic_enable_kmalloc (void)
1155 {
1156         iosapic_kmalloc_ok = 1;
1157         return 0;
1158 }
1159 core_initcall (iosapic_enable_kmalloc);