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/smp_lock.h>
  91 #include <linux/string.h>
  92 #include <linux/bootmem.h>
  93
  94 #include <asm/delay.h>
  95 #include <asm/hw_irq.h>
  96 #include <asm/io.h>
  97 #include <asm/iosapic.h>
  98 #include <asm/machvec.h>
  99 #include <asm/processor.h>
 100 #include <asm/ptrace.h>
 101 #include <asm/system.h>
 102
 103 #undef DEBUG_INTERRUPT_ROUTING
 104
 105 #ifdef DEBUG_INTERRUPT_ROUTING
 106 #define DBG(fmt...)     printk(fmt)
 107 #else
 108 #define DBG(fmt...)
 109 #endif
 110
 111 #define NR_PREALLOCATE_RTE_ENTRIES \
 112         (PAGE_SIZE / sizeof(struct iosapic_rte_info))
 113 #define RTE_PREALLOCATED        (1)
 114
 115 static DEFINE_SPINLOCK(iosapic_lock);
 116
 117 /*
 118  * These tables map IA-64 vectors to the IOSAPIC pin that generates this
 119  * vector.
 120  */
 121
 122 struct iosapic_rte_info {
 123         struct list_head rte_list;      /* node in list of RTEs sharing the
 124                                          * same vector */
 125         char __iomem    *addr;          /* base address of IOSAPIC */
 126         unsigned int    gsi_base;       /* first GSI assigned to this
 127                                          * IOSAPIC */
 128         char            rte_index;      /* IOSAPIC RTE index */
 129         int             refcnt;         /* reference counter */
 130         unsigned int    flags;          /* flags */
 131 } ____cacheline_aligned;
 132
 133 static struct iosapic_intr_info {
 134         struct list_head rtes;          /* RTEs using this vector (empty =>
 135                                          * not an IOSAPIC interrupt) */
 136         int             count;          /* # of RTEs that shares this vector */
 137         u32             low32;          /* current value of low word of
 138                                          * Redirection table entry */
 139         unsigned int    dest;           /* destination CPU physical ID */
 140         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
 141         unsigned char   polarity: 1;    /* interrupt polarity
 142                                          * (see iosapic.h) */
 143         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
 144 } iosapic_intr_info[IA64_NUM_VECTORS];
 145
 146 static struct iosapic {
 147         char __iomem    *addr;          /* base address of IOSAPIC */
 148         unsigned int    gsi_base;       /* first GSI assigned to this
 149                                          * IOSAPIC */
 150         unsigned short  num_rte;        /* # of RTEs on this IOSAPIC */
 151         int             rtes_inuse;     /* # of RTEs in use on this IOSAPIC */
 152 #ifdef CONFIG_NUMA
 153         unsigned short  node;           /* numa node association via pxm */
 154 #endif
 155 } iosapic_lists[NR_IOSAPICS];
 156
 157 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
 158
 159 static int iosapic_kmalloc_ok;
 160 static LIST_HEAD(free_rte_list);
 161
 162 /*
 163  * Find an IOSAPIC associated with a GSI
 164  */
 165 static inline int
 166 find_iosapic (unsigned int gsi)
 167 {
 168         int i;
 169
 170         for (i = 0; i < NR_IOSAPICS; i++) {
 171                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
 172                     iosapic_lists[i].num_rte)
 173                         return i;
 174         }
 175
 176         return -1;
 177 }
 178
 179 static inline int
 180 _gsi_to_vector (unsigned int gsi)
 181 {
 182         struct iosapic_intr_info *info;
 183         struct iosapic_rte_info *rte;
 184
 185         for (info = iosapic_intr_info; info <
 186                      iosapic_intr_info + IA64_NUM_VECTORS; ++info)
 187                 list_for_each_entry(rte, &info->rtes, rte_list)
 188                         if (rte->gsi_base + rte->rte_index == gsi)
 189                                 return info - iosapic_intr_info;
 190         return -1;
 191 }
 192
 193 /*
 194  * Translate GSI number to the corresponding IA-64 interrupt vector.  If no
 195  * entry exists, return -1.
 196  */
 197 inline int
 198 gsi_to_vector (unsigned int gsi)
 199 {
 200         return _gsi_to_vector(gsi);
 201 }
 202
 203 int
 204 gsi_to_irq (unsigned int gsi)
 205 {
 206         unsigned long flags;
 207         int irq;
 208         /*
 209          * XXX fix me: this assumes an identity mapping between IA-64 vector
 210          * and Linux irq numbers...
 211          */
 212         spin_lock_irqsave(&iosapic_lock, flags);
 213         {
 214                 irq = _gsi_to_vector(gsi);
 215         }
 216         spin_unlock_irqrestore(&iosapic_lock, flags);
 217
 218         return irq;
 219 }
 220
 221 static struct iosapic_rte_info *gsi_vector_to_rte(unsigned int gsi,
 222                                                   unsigned int vec)
 223 {
 224         struct iosapic_rte_info *rte;
 225
 226         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
 227                 if (rte->gsi_base + rte->rte_index == gsi)
 228                         return rte;
 229         return NULL;
 230 }
 231
 232 static void
 233 set_rte (unsigned int gsi, unsigned int vector, unsigned int dest, int mask)
 234 {
 235         unsigned long pol, trigger, dmode;
 236         u32 low32, high32;
 237         char __iomem *addr;
 238         int rte_index;
 239         char redir;
 240         struct iosapic_rte_info *rte;
 241
 242         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
 243
 244         rte = gsi_vector_to_rte(gsi, vector);
 245         if (!rte)
 246                 return;         /* not an IOSAPIC interrupt */
 247
 248         rte_index = rte->rte_index;
 249         addr    = rte->addr;
 250         pol     = iosapic_intr_info[vector].polarity;
 251         trigger = iosapic_intr_info[vector].trigger;
 252         dmode   = iosapic_intr_info[vector].dmode;
 253
 254         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
 255
 256 #ifdef CONFIG_SMP
 257         {
 258                 unsigned int irq;
 259
 260                 for (irq = 0; irq < NR_IRQS; ++irq)
 261                         if (irq_to_vector(irq) == vector) {
 262                                 set_irq_affinity_info(irq,
 263                                                       (int)(dest & 0xffff),
 264                                                       redir);
 265                                 break;
 266                         }
 267         }
 268 #endif
 269
 270         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
 271                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
 272                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
 273                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
 274                  vector);
 275
 276         /* dest contains both id and eid */
 277         high32 = (dest << IOSAPIC_DEST_SHIFT);
 278
 279         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index), high32);
 280         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
 281         iosapic_intr_info[vector].low32 = low32;
 282         iosapic_intr_info[vector].dest = dest;
 283 }
 284
 285 static void
 286 nop (unsigned int irq)
 287 {
 288         /* do nothing... */
 289 }
 290
 291
 292 #ifdef CONFIG_KEXEC
 293 void
 294 kexec_disable_iosapic(void)
 295 {
 296         struct iosapic_intr_info *info;
 297         struct iosapic_rte_info *rte;
 298         u8 vec = 0;
 299         for (info = iosapic_intr_info; info <
 300                         iosapic_intr_info + IA64_NUM_VECTORS; ++info, ++vec) {
 301                 list_for_each_entry(rte, &info->rtes,
 302                                 rte_list) {
 303                         iosapic_write(rte->addr,
 304                                         IOSAPIC_RTE_LOW(rte->rte_index),
 305                                         IOSAPIC_MASK|vec);
 306                         iosapic_eoi(rte->addr, vec);
 307                 }
 308         }
 309 }
 310 #endif
 311
 312 static void
 313 mask_irq (unsigned int irq)
 314 {
 315         unsigned long flags;
 316         char __iomem *addr;
 317         u32 low32;
 318         int rte_index;
 319         ia64_vector vec = irq_to_vector(irq);
 320         struct iosapic_rte_info *rte;
 321
 322         if (list_empty(&iosapic_intr_info[vec].rtes))
 323                 return;                 /* not an IOSAPIC interrupt! */
 324
 325         spin_lock_irqsave(&iosapic_lock, flags);
 326         {
 327                 /* set only the mask bit */
 328                 low32 = iosapic_intr_info[vec].low32 |= IOSAPIC_MASK;
 329                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes,
 330                                     rte_list) {
 331                         addr = rte->addr;
 332                         rte_index = rte->rte_index;
 333                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
 334                 }
 335         }
 336         spin_unlock_irqrestore(&iosapic_lock, flags);
 337 }
 338
 339 static void
 340 unmask_irq (unsigned int irq)
 341 {
 342         unsigned long flags;
 343         char __iomem *addr;
 344         u32 low32;
 345         int rte_index;
 346         ia64_vector vec = irq_to_vector(irq);
 347         struct iosapic_rte_info *rte;
 348
 349         if (list_empty(&iosapic_intr_info[vec].rtes))
 350                 return;                 /* not an IOSAPIC interrupt! */
 351
 352         spin_lock_irqsave(&iosapic_lock, flags);
 353         {
 354                 low32 = iosapic_intr_info[vec].low32 &= ~IOSAPIC_MASK;
 355                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes,
 356                                     rte_list) {
 357                         addr = rte->addr;
 358                         rte_index = rte->rte_index;
 359                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
 360                 }
 361         }
 362         spin_unlock_irqrestore(&iosapic_lock, flags);
 363 }
 364
 365
 366 static void
 367 iosapic_set_affinity (unsigned int irq, cpumask_t mask)
 368 {
 369 #ifdef CONFIG_SMP
 370         unsigned long flags;
 371         u32 high32, low32;
 372         int dest, rte_index;
 373         char __iomem *addr;
 374         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
 375         ia64_vector vec;
 376         struct iosapic_rte_info *rte;
 377
 378         irq &= (~IA64_IRQ_REDIRECTED);
 379         vec = irq_to_vector(irq);
 380
 381         if (cpus_empty(mask))
 382                 return;
 383
 384         dest = cpu_physical_id(first_cpu(mask));
 385
 386         if (list_empty(&iosapic_intr_info[vec].rtes))
 387                 return;                 /* not an IOSAPIC interrupt */
 388
 389         set_irq_affinity_info(irq, dest, redir);
 390
 391         /* dest contains both id and eid */
 392         high32 = dest << IOSAPIC_DEST_SHIFT;
 393
 394         spin_lock_irqsave(&iosapic_lock, flags);
 395         {
 396                 low32 = iosapic_intr_info[vec].low32 &
 397                         ~(7 << IOSAPIC_DELIVERY_SHIFT);
 398
 399                 if (redir)
 400                         /* change delivery mode to lowest priority */
 401                         low32 |= (IOSAPIC_LOWEST_PRIORITY <<
 402                                   IOSAPIC_DELIVERY_SHIFT);
 403                 else
 404                         /* change delivery mode to fixed */
 405                         low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
 406
 407                 iosapic_intr_info[vec].low32 = low32;
 408                 iosapic_intr_info[vec].dest = dest;
 409                 list_for_each_entry(rte, &iosapic_intr_info[vec].rtes,
 410                                     rte_list) {
 411                         addr = rte->addr;
 412                         rte_index = rte->rte_index;
 413                         iosapic_write(addr, IOSAPIC_RTE_HIGH(rte_index),
 414                                       high32);
 415                         iosapic_write(addr, IOSAPIC_RTE_LOW(rte_index), low32);
 416                 }
 417         }
 418         spin_unlock_irqrestore(&iosapic_lock, flags);
 419 #endif
 420 }
 421
 422 /*
 423  * Handlers for level-triggered interrupts.
 424  */
 425
 426 static unsigned int
 427 iosapic_startup_level_irq (unsigned int irq)
 428 {
 429         unmask_irq(irq);
 430         return 0;
 431 }
 432
 433 static void
 434 iosapic_end_level_irq (unsigned int irq)
 435 {
 436         ia64_vector vec = irq_to_vector(irq);
 437         struct iosapic_rte_info *rte;
 438
 439         move_native_irq(irq);
 440         list_for_each_entry(rte, &iosapic_intr_info[vec].rtes, rte_list)
 441                 iosapic_eoi(rte->addr, vec);
 442 }
 443
 444 #define iosapic_shutdown_level_irq      mask_irq
 445 #define iosapic_enable_level_irq        unmask_irq
 446 #define iosapic_disable_level_irq       mask_irq
 447 #define iosapic_ack_level_irq           nop
 448
 449 struct hw_interrupt_type irq_type_iosapic_level = {
 450         .name =         "IO-SAPIC-level",
 451         .startup =      iosapic_startup_level_irq,
 452         .shutdown =     iosapic_shutdown_level_irq,
 453         .enable =       iosapic_enable_level_irq,
 454         .disable =      iosapic_disable_level_irq,
 455         .ack =          iosapic_ack_level_irq,
 456         .end =          iosapic_end_level_irq,
 457         .set_affinity = iosapic_set_affinity
 458 };
 459
 460 /*
 461  * Handlers for edge-triggered interrupts.
 462  */
 463
 464 static unsigned int
 465 iosapic_startup_edge_irq (unsigned int irq)
 466 {
 467         unmask_irq(irq);
 468         /*
 469          * IOSAPIC simply drops interrupts pended while the
 470          * corresponding pin was masked, so we can't know if an
 471          * interrupt is pending already.  Let's hope not...
 472          */
 473         return 0;
 474 }
 475
 476 static void
 477 iosapic_ack_edge_irq (unsigned int irq)
 478 {
 479         irq_desc_t *idesc = irq_desc + irq;
 480
 481         move_native_irq(irq);
 482         /*
 483          * Once we have recorded IRQ_PENDING already, we can mask the
 484          * interrupt for real. This prevents IRQ storms from unhandled
 485          * devices.
 486          */
 487         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
 488             (IRQ_PENDING|IRQ_DISABLED))
 489                 mask_irq(irq);
 490 }
 491
 492 #define iosapic_enable_edge_irq         unmask_irq
 493 #define iosapic_disable_edge_irq        nop
 494 #define iosapic_end_edge_irq            nop
 495
 496 struct hw_interrupt_type irq_type_iosapic_edge = {
 497         .name =         "IO-SAPIC-edge",
 498         .startup =      iosapic_startup_edge_irq,
 499         .shutdown =     iosapic_disable_edge_irq,
 500         .enable =       iosapic_enable_edge_irq,
 501         .disable =      iosapic_disable_edge_irq,
 502         .ack =          iosapic_ack_edge_irq,
 503         .end =          iosapic_end_edge_irq,
 504         .set_affinity = iosapic_set_affinity
 505 };
 506
 507 unsigned int
 508 iosapic_version (char __iomem *addr)
 509 {
 510         /*
 511          * IOSAPIC Version Register return 32 bit structure like:
 512          * {
 513          *      unsigned int version   : 8;
 514          *      unsigned int reserved1 : 8;
 515          *      unsigned int max_redir : 8;
 516          *      unsigned int reserved2 : 8;
 517          * }
 518          */
 519         return iosapic_read(addr, IOSAPIC_VERSION);
 520 }
 521
 522 static int iosapic_find_sharable_vector (unsigned long trigger,
 523                                          unsigned long pol)
 524 {
 525         int i, vector = -1, min_count = -1;
 526         struct iosapic_intr_info *info;
 527
 528         /*
 529          * shared vectors for edge-triggered interrupts are not
 530          * supported yet
 531          */
 532         if (trigger == IOSAPIC_EDGE)
 533                 return -1;
 534
 535         for (i = IA64_FIRST_DEVICE_VECTOR; i <= IA64_LAST_DEVICE_VECTOR; i++) {
 536                 info = &iosapic_intr_info[i];
 537                 if (info->trigger == trigger && info->polarity == pol &&
 538                     (info->dmode == IOSAPIC_FIXED || info->dmode ==
 539                      IOSAPIC_LOWEST_PRIORITY)) {
 540                         if (min_count == -1 || info->count < min_count) {
 541                                 vector = i;
 542                                 min_count = info->count;
 543                         }
 544                 }
 545         }
 546
 547         return vector;
 548 }
 549
 550 /*
 551  * if the given vector is already owned by other,
 552  *  assign a new vector for the other and make the vector available
 553  */
 554 static void __init
 555 iosapic_reassign_vector (int vector)
 556 {
 557         int new_vector;
 558
 559         if (!list_empty(&iosapic_intr_info[vector].rtes)) {
 560                 new_vector = assign_irq_vector(AUTO_ASSIGN);
 561                 if (new_vector < 0)
 562                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
 563                 printk(KERN_INFO "Reassigning vector %d to %d\n",
 564                        vector, new_vector);
 565                 memcpy(&iosapic_intr_info[new_vector], &iosapic_intr_info[vector],
 566                        sizeof(struct iosapic_intr_info));
 567                 INIT_LIST_HEAD(&iosapic_intr_info[new_vector].rtes);
 568                 list_move(iosapic_intr_info[vector].rtes.next,
 569                           &iosapic_intr_info[new_vector].rtes);
 570                 memset(&iosapic_intr_info[vector], 0,
 571                        sizeof(struct iosapic_intr_info));
 572                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
 573                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
 574         }
 575 }
 576
 577 static struct iosapic_rte_info *iosapic_alloc_rte (void)
 578 {
 579         int i;
 580         struct iosapic_rte_info *rte;
 581         int preallocated = 0;
 582
 583         if (!iosapic_kmalloc_ok && list_empty(&free_rte_list)) {
 584                 rte = alloc_bootmem(sizeof(struct iosapic_rte_info) *
 585                                     NR_PREALLOCATE_RTE_ENTRIES);
 586                 if (!rte)
 587                         return NULL;
 588                 for (i = 0; i < NR_PREALLOCATE_RTE_ENTRIES; i++, rte++)
 589                         list_add(&rte->rte_list, &free_rte_list);
 590         }
 591
 592         if (!list_empty(&free_rte_list)) {
 593                 rte = list_entry(free_rte_list.next, struct iosapic_rte_info,
 594                                  rte_list);
 595                 list_del(&rte->rte_list);
 596                 preallocated++;
 597         } else {
 598                 rte = kmalloc(sizeof(struct iosapic_rte_info), GFP_ATOMIC);
 599                 if (!rte)
 600                         return NULL;
 601         }
 602
 603         memset(rte, 0, sizeof(struct iosapic_rte_info));
 604         if (preallocated)
 605                 rte->flags |= RTE_PREALLOCATED;
 606
 607         return rte;
 608 }
 609
 610 static void iosapic_free_rte (struct iosapic_rte_info *rte)
 611 {
 612         if (rte->flags & RTE_PREALLOCATED)
 613                 list_add_tail(&rte->rte_list, &free_rte_list);
 614         else
 615                 kfree(rte);
 616 }
 617
 618 static inline int vector_is_shared (int vector)
 619 {
 620         return (iosapic_intr_info[vector].count > 1);
 621 }
 622
 623 static int
 624 register_intr (unsigned int gsi, int vector, unsigned char delivery,
 625                unsigned long polarity, unsigned long trigger)
 626 {
 627         irq_desc_t *idesc;
 628         struct hw_interrupt_type *irq_type;
 629         int rte_index;
 630         int index;
 631         unsigned long gsi_base;
 632         void __iomem *iosapic_address;
 633         struct iosapic_rte_info *rte;
 634
 635         index = find_iosapic(gsi);
 636         if (index < 0) {
 637                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
 638                        __FUNCTION__, gsi);
 639                 return -ENODEV;
 640         }
 641
 642         iosapic_address = iosapic_lists[index].addr;
 643         gsi_base = iosapic_lists[index].gsi_base;
 644
 645         rte = gsi_vector_to_rte(gsi, vector);
 646         if (!rte) {
 647                 rte = iosapic_alloc_rte();
 648                 if (!rte) {
 649                         printk(KERN_WARNING "%s: cannot allocate memory\n",
 650                                __FUNCTION__);
 651                         return -ENOMEM;
 652                 }
 653
 654                 rte_index = gsi - gsi_base;
 655                 rte->rte_index  = rte_index;
 656                 rte->addr       = iosapic_address;
 657                 rte->gsi_base   = gsi_base;
 658                 rte->refcnt++;
 659                 list_add_tail(&rte->rte_list, &iosapic_intr_info[vector].rtes);
 660                 iosapic_intr_info[vector].count++;
 661                 iosapic_lists[index].rtes_inuse++;
 662         }
 663         else if (vector_is_shared(vector)) {
 664                 struct iosapic_intr_info *info = &iosapic_intr_info[vector];
 665                 if (info->trigger != trigger || info->polarity != polarity) {
 666                         printk (KERN_WARNING
 667                                 "%s: cannot override the interrupt\n",
 668                                 __FUNCTION__);
 669                         return -EINVAL;
 670                 }
 671         }
 672
 673         iosapic_intr_info[vector].polarity = polarity;
 674         iosapic_intr_info[vector].dmode    = delivery;
 675         iosapic_intr_info[vector].trigger  = trigger;
 676
 677         if (trigger == IOSAPIC_EDGE)
 678                 irq_type = &irq_type_iosapic_edge;
 679         else
 680                 irq_type = &irq_type_iosapic_level;
 681
 682         idesc = irq_desc + vector;
 683         if (idesc->chip != irq_type) {
 684                 if (idesc->chip != &no_irq_type)
 685                         printk(KERN_WARNING
 686                                "%s: changing vector %d from %s to %s\n",
 687                                __FUNCTION__, vector,
 688                                idesc->chip->name, irq_type->name);
 689                 idesc->chip = irq_type;
 690         }
 691         return 0;
 692 }
 693
 694 static unsigned int
 695 get_target_cpu (unsigned int gsi, int vector)
 696 {
 697 #ifdef CONFIG_SMP
 698         static int cpu = -1;
 699         extern int cpe_vector;
 700
 701         /*
 702          * In case of vector shared by multiple RTEs, all RTEs that
 703          * share the vector need to use the same destination CPU.
 704          */
 705         if (!list_empty(&iosapic_intr_info[vector].rtes))
 706                 return iosapic_intr_info[vector].dest;
 707
 708         /*
 709          * If the platform supports redirection via XTP, let it
 710          * distribute interrupts.
 711          */
 712         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
 713                 return cpu_physical_id(smp_processor_id());
 714
 715         /*
 716          * Some interrupts (ACPI SCI, for instance) are registered
 717          * before the BSP is marked as online.
 718          */
 719         if (!cpu_online(smp_processor_id()))
 720                 return cpu_physical_id(smp_processor_id());
 721
 722 #ifdef CONFIG_ACPI
 723         if (cpe_vector > 0 && vector == IA64_CPEP_VECTOR)
 724                 return get_cpei_target_cpu();
 725 #endif
 726
 727 #ifdef CONFIG_NUMA
 728         {
 729                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
 730                 cpumask_t cpu_mask;
 731
 732                 iosapic_index = find_iosapic(gsi);
 733                 if (iosapic_index < 0 ||
 734                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
 735                         goto skip_numa_setup;
 736
 737                 cpu_mask = node_to_cpumask(iosapic_lists[iosapic_index].node);
 738
 739                 for_each_cpu_mask(numa_cpu, cpu_mask) {
 740                         if (!cpu_online(numa_cpu))
 741                                 cpu_clear(numa_cpu, cpu_mask);
 742                 }
 743
 744                 num_cpus = cpus_weight(cpu_mask);
 745
 746                 if (!num_cpus)
 747                         goto skip_numa_setup;
 748
 749                 /* Use vector assignment to distribute across cpus in node */
 750                 cpu_index = vector % num_cpus;
 751
 752                 for (numa_cpu = first_cpu(cpu_mask) ; i < cpu_index ; i++)
 753                         numa_cpu = next_cpu(numa_cpu, cpu_mask);
 754
 755                 if (numa_cpu != NR_CPUS)
 756                         return cpu_physical_id(numa_cpu);
 757         }
 758 skip_numa_setup:
 759 #endif
 760         /*
 761          * Otherwise, round-robin interrupt vectors across all the
 762          * processors.  (It'd be nice if we could be smarter in the
 763          * case of NUMA.)
 764          */
 765         do {
 766                 if (++cpu >= NR_CPUS)
 767                         cpu = 0;
 768         } while (!cpu_online(cpu));
 769
 770         return cpu_physical_id(cpu);
 771 #else  /* CONFIG_SMP */
 772         return cpu_physical_id(smp_processor_id());
 773 #endif
 774 }
 775
 776 /*
 777  * ACPI can describe IOSAPIC interrupts via static tables and namespace
 778  * methods.  This provides an interface to register those interrupts and
 779  * program the IOSAPIC RTE.
 780  */
 781 int
 782 iosapic_register_intr (unsigned int gsi,
 783                        unsigned long polarity, unsigned long trigger)
 784 {
 785         int vector, mask = 1, err;
 786         unsigned int dest;
 787         unsigned long flags;
 788         struct iosapic_rte_info *rte;
 789         u32 low32;
 790 again:
 791         /*
 792          * If this GSI has already been registered (i.e., it's a
 793          * shared interrupt, or we lost a race to register it),
 794          * don't touch the RTE.
 795          */
 796         spin_lock_irqsave(&iosapic_lock, flags);
 797         {
 798                 vector = gsi_to_vector(gsi);
 799                 if (vector > 0) {
 800                         rte = gsi_vector_to_rte(gsi, vector);
 801                         rte->refcnt++;
 802                         spin_unlock_irqrestore(&iosapic_lock, flags);
 803                         return vector;
 804                 }
 805         }
 806         spin_unlock_irqrestore(&iosapic_lock, flags);
 807
 808         /* If vector is running out, we try to find a sharable vector */
 809         vector = assign_irq_vector(AUTO_ASSIGN);
 810         if (vector < 0) {
 811                 vector = iosapic_find_sharable_vector(trigger, polarity);
 812                 if (vector < 0)
 813                         return -ENOSPC;
 814         }
 815
 816         spin_lock_irqsave(&irq_desc[vector].lock, flags);
 817         spin_lock(&iosapic_lock);
 818         {
 819                 if (gsi_to_vector(gsi) > 0) {
 820                         if (list_empty(&iosapic_intr_info[vector].rtes))
 821                                 free_irq_vector(vector);
 822                         spin_unlock(&iosapic_lock);
 823                         spin_unlock_irqrestore(&irq_desc[vector].lock,
 824                                                flags);
 825                         goto again;
 826                 }
 827
 828                 dest = get_target_cpu(gsi, vector);
 829                 err = register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY,
 830                               polarity, trigger);
 831                 if (err < 0) {
 832                         spin_unlock(&iosapic_lock);
 833                         spin_unlock_irqrestore(&irq_desc[vector].lock,
 834                                                flags);
 835                         return err;
 836                 }
 837
 838                 /*
 839                  * If the vector is shared and already unmasked for
 840                  * other interrupt sources, don't mask it.
 841                  */
 842                 low32 = iosapic_intr_info[vector].low32;
 843                 if (vector_is_shared(vector) && !(low32 & IOSAPIC_MASK))
 844                         mask = 0;
 845                 set_rte(gsi, vector, dest, mask);
 846         }
 847         spin_unlock(&iosapic_lock);
 848         spin_unlock_irqrestore(&irq_desc[vector].lock, flags);
 849
 850         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
 851                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 852                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 853                cpu_logical_id(dest), dest, vector);
 854
 855         return vector;
 856 }
 857
 858 void
 859 iosapic_unregister_intr (unsigned int gsi)
 860 {
 861         unsigned long flags;
 862         int irq, vector, index;
 863         irq_desc_t *idesc;
 864         u32 low32;
 865         unsigned long trigger, polarity;
 866         unsigned int dest;
 867         struct iosapic_rte_info *rte;
 868
 869         /*
 870          * If the irq associated with the gsi is not found,
 871          * iosapic_unregister_intr() is unbalanced. We need to check
 872          * this again after getting locks.
 873          */
 874         irq = gsi_to_irq(gsi);
 875         if (irq < 0) {
 876                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
 877                        gsi);
 878                 WARN_ON(1);
 879                 return;
 880         }
 881         vector = irq_to_vector(irq);
 882
 883         idesc = irq_desc + irq;
 884         spin_lock_irqsave(&idesc->lock, flags);
 885         spin_lock(&iosapic_lock);
 886         {
 887                 if ((rte = gsi_vector_to_rte(gsi, vector)) == NULL) {
 888                         printk(KERN_ERR
 889                                "iosapic_unregister_intr(%u) unbalanced\n",
 890                                gsi);
 891                         WARN_ON(1);
 892                         goto out;
 893                 }
 894
 895                 if (--rte->refcnt > 0)
 896                         goto out;
 897
 898                 /* Mask the interrupt */
 899                 low32 = iosapic_intr_info[vector].low32 | IOSAPIC_MASK;
 900                 iosapic_write(rte->addr, IOSAPIC_RTE_LOW(rte->rte_index),
 901                               low32);
 902
 903                 /* Remove the rte entry from the list */
 904                 list_del(&rte->rte_list);
 905                 iosapic_intr_info[vector].count--;
 906                 iosapic_free_rte(rte);
 907                 index = find_iosapic(gsi);
 908                 iosapic_lists[index].rtes_inuse--;
 909                 WARN_ON(iosapic_lists[index].rtes_inuse < 0);
 910
 911                 trigger  = iosapic_intr_info[vector].trigger;
 912                 polarity = iosapic_intr_info[vector].polarity;
 913                 dest     = iosapic_intr_info[vector].dest;
 914                 printk(KERN_INFO
 915                        "GSI %u (%s, %s) -> CPU %d (0x%04x)"
 916                        " vector %d unregistered\n",
 917                        gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 918                        (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 919                        cpu_logical_id(dest), dest, vector);
 920
 921                 if (list_empty(&iosapic_intr_info[vector].rtes)) {
 922                         /* Sanity check */
 923                         BUG_ON(iosapic_intr_info[vector].count);
 924
 925                         /* Clear the interrupt controller descriptor */
 926                         idesc->chip = &no_irq_type;
 927
 928                         /* Clear the interrupt information */
 929                         memset(&iosapic_intr_info[vector], 0,
 930                                sizeof(struct iosapic_intr_info));
 931                         iosapic_intr_info[vector].low32 |= IOSAPIC_MASK;
 932                         INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
 933
 934                         if (idesc->action) {
 935                                 printk(KERN_ERR
 936                                        "interrupt handlers still exist on"
 937                                        "IRQ %u\n", irq);
 938                                 WARN_ON(1);
 939                         }
 940
 941                         /* Free the interrupt vector */
 942                         free_irq_vector(vector);
 943                 }
 944         }
 945  out:
 946         spin_unlock(&iosapic_lock);
 947         spin_unlock_irqrestore(&idesc->lock, flags);
 948 }
 949
 950 /*
 951  * ACPI calls this when it finds an entry for a platform interrupt.
 952  */
 953 int __init
 954 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
 955                                 int iosapic_vector, u16 eid, u16 id,
 956                                 unsigned long polarity, unsigned long trigger)
 957 {
 958         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
 959         unsigned char delivery;
 960         int vector, mask = 0;
 961         unsigned int dest = ((id << 8) | eid) & 0xffff;
 962
 963         switch (int_type) {
 964               case ACPI_INTERRUPT_PMI:
 965                 vector = iosapic_vector;
 966                 /*
 967                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
 968                  * we need to make sure the vector is available
 969                  */
 970                 iosapic_reassign_vector(vector);
 971                 delivery = IOSAPIC_PMI;
 972                 break;
 973               case ACPI_INTERRUPT_INIT:
 974                 vector = assign_irq_vector(AUTO_ASSIGN);
 975                 if (vector < 0)
 976                         panic("%s: out of interrupt vectors!\n", __FUNCTION__);
 977                 delivery = IOSAPIC_INIT;
 978                 break;
 979               case ACPI_INTERRUPT_CPEI:
 980                 vector = IA64_CPE_VECTOR;
 981                 delivery = IOSAPIC_LOWEST_PRIORITY;
 982                 mask = 1;
 983                 break;
 984               default:
 985                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __FUNCTION__,
 986                        int_type);
 987                 return -1;
 988         }
 989
 990         register_intr(gsi, vector, delivery, polarity, trigger);
 991
 992         printk(KERN_INFO
 993                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
 994                " vector %d\n",
 995                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
 996                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 997                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 998                cpu_logical_id(dest), dest, vector);
 999
1000         set_rte(gsi, vector, dest, mask);
1001         return vector;
1002 }
1003
1004 /*
1005  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
1006  */
1007 void __init
1008 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
1009                           unsigned long polarity,
1010                           unsigned long trigger)
1011 {
1012         int vector;
1013         unsigned int dest = cpu_physical_id(smp_processor_id());
1014
1015         vector = isa_irq_to_vector(isa_irq);
1016
1017         register_intr(gsi, vector, IOSAPIC_LOWEST_PRIORITY, polarity, trigger);
1018
1019         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
1020             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
1021             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
1022             cpu_logical_id(dest), dest, vector);
1023
1024         set_rte(gsi, vector, dest, 1);
1025 }
1026
1027 void __init
1028 iosapic_system_init (int system_pcat_compat)
1029 {
1030         int vector;
1031
1032         for (vector = 0; vector < IA64_NUM_VECTORS; ++vector) {
1033                 iosapic_intr_info[vector].low32 = IOSAPIC_MASK;
1034                 /* mark as unused */
1035                 INIT_LIST_HEAD(&iosapic_intr_info[vector].rtes);
1036         }
1037
1038         pcat_compat = system_pcat_compat;
1039         if (pcat_compat) {
1040                 /*
1041                  * Disable the compatibility mode interrupts (8259 style),
1042                  * needs IN/OUT support enabled.
1043                  */
1044                 printk(KERN_INFO
1045                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
1046                        __FUNCTION__);
1047                 outb(0xff, 0xA1);
1048                 outb(0xff, 0x21);
1049         }
1050 }
1051
1052 static inline int
1053 iosapic_alloc (void)
1054 {
1055         int index;
1056
1057         for (index = 0; index < NR_IOSAPICS; index++)
1058                 if (!iosapic_lists[index].addr)
1059                         return index;
1060
1061         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __FUNCTION__);
1062         return -1;
1063 }
1064
1065 static inline void
1066 iosapic_free (int index)
1067 {
1068         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1069 }
1070
1071 static inline int
1072 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1073 {
1074         int index;
1075         unsigned int gsi_end, base, end;
1076
1077         /* check gsi range */
1078         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1079         for (index = 0; index < NR_IOSAPICS; index++) {
1080                 if (!iosapic_lists[index].addr)
1081                         continue;
1082
1083                 base = iosapic_lists[index].gsi_base;
1084                 end  = base + iosapic_lists[index].num_rte - 1;
1085
1086                 if (gsi_end < base || end < gsi_base)
1087                         continue; /* OK */
1088
1089                 return -EBUSY;
1090         }
1091         return 0;
1092 }
1093
1094 int __devinit
1095 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1096 {
1097         int num_rte, err, index;
1098         unsigned int isa_irq, ver;
1099         char __iomem *addr;
1100         unsigned long flags;
1101
1102         spin_lock_irqsave(&iosapic_lock, flags);
1103         {
1104                 addr = ioremap(phys_addr, 0);
1105                 ver = iosapic_version(addr);
1106
1107                 if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1108                         iounmap(addr);
1109                         spin_unlock_irqrestore(&iosapic_lock, flags);
1110                         return err;
1111                 }
1112
1113                 /*
1114                  * The MAX_REDIR register holds the highest input pin
1115                  * number (starting from 0).
1116                  * We add 1 so that we can use it for number of pins (= RTEs)
1117                  */
1118                 num_rte = ((ver >> 16) & 0xff) + 1;
1119
1120                 index = iosapic_alloc();
1121                 iosapic_lists[index].addr = addr;
1122                 iosapic_lists[index].gsi_base = gsi_base;
1123                 iosapic_lists[index].num_rte = num_rte;
1124 #ifdef CONFIG_NUMA
1125                 iosapic_lists[index].node = MAX_NUMNODES;
1126 #endif
1127         }
1128         spin_unlock_irqrestore(&iosapic_lock, flags);
1129
1130         if ((gsi_base == 0) && pcat_compat) {
1131                 /*
1132                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1133                  * these may get reprogrammed later on with data from the ACPI
1134                  * Interrupt Source Override table.
1135                  */
1136                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1137                         iosapic_override_isa_irq(isa_irq, isa_irq,
1138                                                  IOSAPIC_POL_HIGH,
1139                                                  IOSAPIC_EDGE);
1140         }
1141         return 0;
1142 }
1143
1144 #ifdef CONFIG_HOTPLUG
1145 int
1146 iosapic_remove (unsigned int gsi_base)
1147 {
1148         int index, err = 0;
1149         unsigned long flags;
1150
1151         spin_lock_irqsave(&iosapic_lock, flags);
1152         {
1153                 index = find_iosapic(gsi_base);
1154                 if (index < 0) {
1155                         printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1156                                __FUNCTION__, gsi_base);
1157                         goto out;
1158                 }
1159
1160                 if (iosapic_lists[index].rtes_inuse) {
1161                         err = -EBUSY;
1162                         printk(KERN_WARNING
1163                                "%s: IOSAPIC for GSI base %u is busy\n",
1164                                __FUNCTION__, gsi_base);
1165                         goto out;
1166                 }
1167
1168                 iounmap(iosapic_lists[index].addr);
1169                 iosapic_free(index);
1170         }
1171  out:
1172         spin_unlock_irqrestore(&iosapic_lock, flags);
1173         return err;
1174 }
1175 #endif /* CONFIG_HOTPLUG */
1176
1177 #ifdef CONFIG_NUMA
1178 void __devinit
1179 map_iosapic_to_node(unsigned int gsi_base, int node)
1180 {
1181         int index;
1182
1183         index = find_iosapic(gsi_base);
1184         if (index < 0) {
1185                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1186                        __FUNCTION__, gsi_base);
1187                 return;
1188         }
1189         iosapic_lists[index].node = node;
1190         return;
1191 }
1192 #endif
1193
1194 static int __init iosapic_enable_kmalloc (void)
1195 {
1196         iosapic_kmalloc_ok = 1;
1197         return 0;
1198 }
1199 core_initcall (iosapic_enable_kmalloc);