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 arch/ia64/include/asm/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/slab.h>
  90 #include <linux/smp.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 static DEFINE_SPINLOCK(iosapic_lock);
 112
 113 /*
 114  * These tables map IA-64 vectors to the IOSAPIC pin that generates this
 115  * vector.
 116  */
 117
 118 #define NO_REF_RTE      0
 119
 120 static struct iosapic {
 121         char __iomem    *addr;          /* base address of IOSAPIC */
 122         unsigned int    gsi_base;       /* GSI base */
 123         unsigned short  num_rte;        /* # of RTEs on this IOSAPIC */
 124         int             rtes_inuse;     /* # of RTEs in use on this IOSAPIC */
 125 #ifdef CONFIG_NUMA
 126         unsigned short  node;           /* numa node association via pxm */
 127 #endif
 128         spinlock_t      lock;           /* lock for indirect reg access */
 129 } iosapic_lists[NR_IOSAPICS];
 130
 131 struct iosapic_rte_info {
 132         struct list_head rte_list;      /* RTEs sharing the same vector */
 133         char            rte_index;      /* IOSAPIC RTE index */
 134         int             refcnt;         /* reference counter */
 135         struct iosapic  *iosapic;
 136 } ____cacheline_aligned;
 137
 138 static struct iosapic_intr_info {
 139         struct list_head rtes;          /* RTEs using this vector (empty =>
 140                                          * not an IOSAPIC interrupt) */
 141         int             count;          /* # of registered RTEs */
 142         u32             low32;          /* current value of low word of
 143                                          * Redirection table entry */
 144         unsigned int    dest;           /* destination CPU physical ID */
 145         unsigned char   dmode   : 3;    /* delivery mode (see iosapic.h) */
 146         unsigned char   polarity: 1;    /* interrupt polarity
 147                                          * (see iosapic.h) */
 148         unsigned char   trigger : 1;    /* trigger mode (see iosapic.h) */
 149 } iosapic_intr_info[NR_IRQS];
 150
 151 static unsigned char pcat_compat __devinitdata; /* 8259 compatibility flag */
 152
 153 static inline void
 154 iosapic_write(struct iosapic *iosapic, unsigned int reg, u32 val)
 155 {
 156         unsigned long flags;
 157
 158         spin_lock_irqsave(&iosapic->lock, flags);
 159         __iosapic_write(iosapic->addr, reg, val);
 160         spin_unlock_irqrestore(&iosapic->lock, flags);
 161 }
 162
 163 /*
 164  * Find an IOSAPIC associated with a GSI
 165  */
 166 static inline int
 167 find_iosapic (unsigned int gsi)
 168 {
 169         int i;
 170
 171         for (i = 0; i < NR_IOSAPICS; i++) {
 172                 if ((unsigned) (gsi - iosapic_lists[i].gsi_base) <
 173                     iosapic_lists[i].num_rte)
 174                         return i;
 175         }
 176
 177         return -1;
 178 }
 179
 180 static inline int __gsi_to_irq(unsigned int gsi)
 181 {
 182         int irq;
 183         struct iosapic_intr_info *info;
 184         struct iosapic_rte_info *rte;
 185
 186         for (irq = 0; irq < NR_IRQS; irq++) {
 187                 info = &iosapic_intr_info[irq];
 188                 list_for_each_entry(rte, &info->rtes, rte_list)
 189                         if (rte->iosapic->gsi_base + rte->rte_index == gsi)
 190                                 return irq;
 191         }
 192         return -1;
 193 }
 194
 195 int
 196 gsi_to_irq (unsigned int gsi)
 197 {
 198         unsigned long flags;
 199         int irq;
 200
 201         spin_lock_irqsave(&iosapic_lock, flags);
 202         irq = __gsi_to_irq(gsi);
 203         spin_unlock_irqrestore(&iosapic_lock, flags);
 204         return irq;
 205 }
 206
 207 static struct iosapic_rte_info *find_rte(unsigned int irq, unsigned int gsi)
 208 {
 209         struct iosapic_rte_info *rte;
 210
 211         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
 212                 if (rte->iosapic->gsi_base + rte->rte_index == gsi)
 213                         return rte;
 214         return NULL;
 215 }
 216
 217 static void
 218 set_rte (unsigned int gsi, unsigned int irq, unsigned int dest, int mask)
 219 {
 220         unsigned long pol, trigger, dmode;
 221         u32 low32, high32;
 222         int rte_index;
 223         char redir;
 224         struct iosapic_rte_info *rte;
 225         ia64_vector vector = irq_to_vector(irq);
 226
 227         DBG(KERN_DEBUG"IOSAPIC: routing vector %d to 0x%x\n", vector, dest);
 228
 229         rte = find_rte(irq, gsi);
 230         if (!rte)
 231                 return;         /* not an IOSAPIC interrupt */
 232
 233         rte_index = rte->rte_index;
 234         pol     = iosapic_intr_info[irq].polarity;
 235         trigger = iosapic_intr_info[irq].trigger;
 236         dmode   = iosapic_intr_info[irq].dmode;
 237
 238         redir = (dmode == IOSAPIC_LOWEST_PRIORITY) ? 1 : 0;
 239
 240 #ifdef CONFIG_SMP
 241         set_irq_affinity_info(irq, (int)(dest & 0xffff), redir);
 242 #endif
 243
 244         low32 = ((pol << IOSAPIC_POLARITY_SHIFT) |
 245                  (trigger << IOSAPIC_TRIGGER_SHIFT) |
 246                  (dmode << IOSAPIC_DELIVERY_SHIFT) |
 247                  ((mask ? 1 : 0) << IOSAPIC_MASK_SHIFT) |
 248                  vector);
 249
 250         /* dest contains both id and eid */
 251         high32 = (dest << IOSAPIC_DEST_SHIFT);
 252
 253         iosapic_write(rte->iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
 254         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 255         iosapic_intr_info[irq].low32 = low32;
 256         iosapic_intr_info[irq].dest = dest;
 257 }
 258
 259 static void
 260 nop (struct irq_data *data)
 261 {
 262         /* do nothing... */
 263 }
 264
 265
 266 #ifdef CONFIG_KEXEC
 267 void
 268 kexec_disable_iosapic(void)
 269 {
 270         struct iosapic_intr_info *info;
 271         struct iosapic_rte_info *rte;
 272         ia64_vector vec;
 273         int irq;
 274
 275         for (irq = 0; irq < NR_IRQS; irq++) {
 276                 info = &iosapic_intr_info[irq];
 277                 vec = irq_to_vector(irq);
 278                 list_for_each_entry(rte, &info->rtes,
 279                                 rte_list) {
 280                         iosapic_write(rte->iosapic,
 281                                         IOSAPIC_RTE_LOW(rte->rte_index),
 282                                         IOSAPIC_MASK|vec);
 283                         iosapic_eoi(rte->iosapic->addr, vec);
 284                 }
 285         }
 286 }
 287 #endif
 288
 289 static void
 290 mask_irq (struct irq_data *data)
 291 {
 292         unsigned int irq = data->irq;
 293         u32 low32;
 294         int rte_index;
 295         struct iosapic_rte_info *rte;
 296
 297         if (!iosapic_intr_info[irq].count)
 298                 return;                 /* not an IOSAPIC interrupt! */
 299
 300         /* set only the mask bit */
 301         low32 = iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
 302         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
 303                 rte_index = rte->rte_index;
 304                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 305         }
 306 }
 307
 308 static void
 309 unmask_irq (struct irq_data *data)
 310 {
 311         unsigned int irq = data->irq;
 312         u32 low32;
 313         int rte_index;
 314         struct iosapic_rte_info *rte;
 315
 316         if (!iosapic_intr_info[irq].count)
 317                 return;                 /* not an IOSAPIC interrupt! */
 318
 319         low32 = iosapic_intr_info[irq].low32 &= ~IOSAPIC_MASK;
 320         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
 321                 rte_index = rte->rte_index;
 322                 iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 323         }
 324 }
 325
 326
 327 static int
 328 iosapic_set_affinity(struct irq_data *data, const struct cpumask *mask,
 329                      bool force)
 330 {
 331 #ifdef CONFIG_SMP
 332         unsigned int irq = data->irq;
 333         u32 high32, low32;
 334         int cpu, dest, rte_index;
 335         int redir = (irq & IA64_IRQ_REDIRECTED) ? 1 : 0;
 336         struct iosapic_rte_info *rte;
 337         struct iosapic *iosapic;
 338
 339         irq &= (~IA64_IRQ_REDIRECTED);
 340
 341         cpu = cpumask_first_and(cpu_online_mask, mask);
 342         if (cpu >= nr_cpu_ids)
 343                 return -1;
 344
 345         if (irq_prepare_move(irq, cpu))
 346                 return -1;
 347
 348         dest = cpu_physical_id(cpu);
 349
 350         if (!iosapic_intr_info[irq].count)
 351                 return -1;                      /* not an IOSAPIC interrupt */
 352
 353         set_irq_affinity_info(irq, dest, redir);
 354
 355         /* dest contains both id and eid */
 356         high32 = dest << IOSAPIC_DEST_SHIFT;
 357
 358         low32 = iosapic_intr_info[irq].low32 & ~(7 << IOSAPIC_DELIVERY_SHIFT);
 359         if (redir)
 360                 /* change delivery mode to lowest priority */
 361                 low32 |= (IOSAPIC_LOWEST_PRIORITY << IOSAPIC_DELIVERY_SHIFT);
 362         else
 363                 /* change delivery mode to fixed */
 364                 low32 |= (IOSAPIC_FIXED << IOSAPIC_DELIVERY_SHIFT);
 365         low32 &= IOSAPIC_VECTOR_MASK;
 366         low32 |= irq_to_vector(irq);
 367
 368         iosapic_intr_info[irq].low32 = low32;
 369         iosapic_intr_info[irq].dest = dest;
 370         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list) {
 371                 iosapic = rte->iosapic;
 372                 rte_index = rte->rte_index;
 373                 iosapic_write(iosapic, IOSAPIC_RTE_HIGH(rte_index), high32);
 374                 iosapic_write(iosapic, IOSAPIC_RTE_LOW(rte_index), low32);
 375         }
 376
 377 #endif
 378         return 0;
 379 }
 380
 381 /*
 382  * Handlers for level-triggered interrupts.
 383  */
 384
 385 static unsigned int
 386 iosapic_startup_level_irq (struct irq_data *data)
 387 {
 388         unmask_irq(data);
 389         return 0;
 390 }
 391
 392 static void
 393 iosapic_unmask_level_irq (struct irq_data *data)
 394 {
 395         unsigned int irq = data->irq;
 396         ia64_vector vec = irq_to_vector(irq);
 397         struct iosapic_rte_info *rte;
 398         int do_unmask_irq = 0;
 399
 400         irq_complete_move(irq);
 401         if (unlikely(irq_desc[irq].status & IRQ_MOVE_PENDING)) {
 402                 do_unmask_irq = 1;
 403                 mask_irq(data);
 404         } else
 405                 unmask_irq(data);
 406
 407         list_for_each_entry(rte, &iosapic_intr_info[irq].rtes, rte_list)
 408                 iosapic_eoi(rte->iosapic->addr, vec);
 409
 410         if (unlikely(do_unmask_irq)) {
 411                 move_masked_irq(irq);
 412                 unmask_irq(data);
 413         }
 414 }
 415
 416 #define iosapic_shutdown_level_irq      mask_irq
 417 #define iosapic_enable_level_irq        unmask_irq
 418 #define iosapic_disable_level_irq       mask_irq
 419 #define iosapic_ack_level_irq           nop
 420
 421 static struct irq_chip irq_type_iosapic_level = {
 422         .name =                 "IO-SAPIC-level",
 423         .irq_startup =          iosapic_startup_level_irq,
 424         .irq_shutdown =         iosapic_shutdown_level_irq,
 425         .irq_enable =           iosapic_enable_level_irq,
 426         .irq_disable =          iosapic_disable_level_irq,
 427         .irq_ack =              iosapic_ack_level_irq,
 428         .irq_mask =             mask_irq,
 429         .irq_unmask =           iosapic_unmask_level_irq,
 430         .irq_set_affinity =     iosapic_set_affinity
 431 };
 432
 433 /*
 434  * Handlers for edge-triggered interrupts.
 435  */
 436
 437 static unsigned int
 438 iosapic_startup_edge_irq (struct irq_data *data)
 439 {
 440         unmask_irq(data);
 441         /*
 442          * IOSAPIC simply drops interrupts pended while the
 443          * corresponding pin was masked, so we can't know if an
 444          * interrupt is pending already.  Let's hope not...
 445          */
 446         return 0;
 447 }
 448
 449 static void
 450 iosapic_ack_edge_irq (struct irq_data *data)
 451 {
 452         unsigned int irq = data->irq;
 453         struct irq_desc *idesc = irq_desc + irq;
 454
 455         irq_complete_move(irq);
 456         move_native_irq(irq);
 457         /*
 458          * Once we have recorded IRQ_PENDING already, we can mask the
 459          * interrupt for real. This prevents IRQ storms from unhandled
 460          * devices.
 461          */
 462         if ((idesc->status & (IRQ_PENDING|IRQ_DISABLED)) ==
 463             (IRQ_PENDING|IRQ_DISABLED))
 464                 mask_irq(data);
 465 }
 466
 467 #define iosapic_enable_edge_irq         unmask_irq
 468 #define iosapic_disable_edge_irq        nop
 469
 470 static struct irq_chip irq_type_iosapic_edge = {
 471         .name =                 "IO-SAPIC-edge",
 472         .irq_startup =          iosapic_startup_edge_irq,
 473         .irq_shutdown =         iosapic_disable_edge_irq,
 474         .irq_enable =           iosapic_enable_edge_irq,
 475         .irq_disable =          iosapic_disable_edge_irq,
 476         .irq_ack =              iosapic_ack_edge_irq,
 477         .irq_mask =             mask_irq,
 478         .irq_unmask =           unmask_irq,
 479         .irq_set_affinity =     iosapic_set_affinity
 480 };
 481
 482 static unsigned int
 483 iosapic_version (char __iomem *addr)
 484 {
 485         /*
 486          * IOSAPIC Version Register return 32 bit structure like:
 487          * {
 488          *      unsigned int version   : 8;
 489          *      unsigned int reserved1 : 8;
 490          *      unsigned int max_redir : 8;
 491          *      unsigned int reserved2 : 8;
 492          * }
 493          */
 494         return __iosapic_read(addr, IOSAPIC_VERSION);
 495 }
 496
 497 static int iosapic_find_sharable_irq(unsigned long trigger, unsigned long pol)
 498 {
 499         int i, irq = -ENOSPC, min_count = -1;
 500         struct iosapic_intr_info *info;
 501
 502         /*
 503          * shared vectors for edge-triggered interrupts are not
 504          * supported yet
 505          */
 506         if (trigger == IOSAPIC_EDGE)
 507                 return -EINVAL;
 508
 509         for (i = 0; i < NR_IRQS; i++) {
 510                 info = &iosapic_intr_info[i];
 511                 if (info->trigger == trigger && info->polarity == pol &&
 512                     (info->dmode == IOSAPIC_FIXED ||
 513                      info->dmode == IOSAPIC_LOWEST_PRIORITY) &&
 514                     can_request_irq(i, IRQF_SHARED)) {
 515                         if (min_count == -1 || info->count < min_count) {
 516                                 irq = i;
 517                                 min_count = info->count;
 518                         }
 519                 }
 520         }
 521         return irq;
 522 }
 523
 524 /*
 525  * if the given vector is already owned by other,
 526  *  assign a new vector for the other and make the vector available
 527  */
 528 static void __init
 529 iosapic_reassign_vector (int irq)
 530 {
 531         int new_irq;
 532
 533         if (iosapic_intr_info[irq].count) {
 534                 new_irq = create_irq();
 535                 if (new_irq < 0)
 536                         panic("%s: out of interrupt vectors!\n", __func__);
 537                 printk(KERN_INFO "Reassigning vector %d to %d\n",
 538                        irq_to_vector(irq), irq_to_vector(new_irq));
 539                 memcpy(&iosapic_intr_info[new_irq], &iosapic_intr_info[irq],
 540                        sizeof(struct iosapic_intr_info));
 541                 INIT_LIST_HEAD(&iosapic_intr_info[new_irq].rtes);
 542                 list_move(iosapic_intr_info[irq].rtes.next,
 543                           &iosapic_intr_info[new_irq].rtes);
 544                 memset(&iosapic_intr_info[irq], 0,
 545                        sizeof(struct iosapic_intr_info));
 546                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
 547                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
 548         }
 549 }
 550
 551 static inline int irq_is_shared (int irq)
 552 {
 553         return (iosapic_intr_info[irq].count > 1);
 554 }
 555
 556 struct irq_chip*
 557 ia64_native_iosapic_get_irq_chip(unsigned long trigger)
 558 {
 559         if (trigger == IOSAPIC_EDGE)
 560                 return &irq_type_iosapic_edge;
 561         else
 562                 return &irq_type_iosapic_level;
 563 }
 564
 565 static int
 566 register_intr (unsigned int gsi, int irq, unsigned char delivery,
 567                unsigned long polarity, unsigned long trigger)
 568 {
 569         struct irq_desc *idesc;
 570         struct irq_chip *irq_type;
 571         int index;
 572         struct iosapic_rte_info *rte;
 573
 574         index = find_iosapic(gsi);
 575         if (index < 0) {
 576                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
 577                        __func__, gsi);
 578                 return -ENODEV;
 579         }
 580
 581         rte = find_rte(irq, gsi);
 582         if (!rte) {
 583                 rte = kzalloc(sizeof (*rte), GFP_ATOMIC);
 584                 if (!rte) {
 585                         printk(KERN_WARNING "%s: cannot allocate memory\n",
 586                                __func__);
 587                         return -ENOMEM;
 588                 }
 589
 590                 rte->iosapic    = &iosapic_lists[index];
 591                 rte->rte_index  = gsi - rte->iosapic->gsi_base;
 592                 rte->refcnt++;
 593                 list_add_tail(&rte->rte_list, &iosapic_intr_info[irq].rtes);
 594                 iosapic_intr_info[irq].count++;
 595                 iosapic_lists[index].rtes_inuse++;
 596         }
 597         else if (rte->refcnt == NO_REF_RTE) {
 598                 struct iosapic_intr_info *info = &iosapic_intr_info[irq];
 599                 if (info->count > 0 &&
 600                     (info->trigger != trigger || info->polarity != polarity)){
 601                         printk (KERN_WARNING
 602                                 "%s: cannot override the interrupt\n",
 603                                 __func__);
 604                         return -EINVAL;
 605                 }
 606                 rte->refcnt++;
 607                 iosapic_intr_info[irq].count++;
 608                 iosapic_lists[index].rtes_inuse++;
 609         }
 610
 611         iosapic_intr_info[irq].polarity = polarity;
 612         iosapic_intr_info[irq].dmode    = delivery;
 613         iosapic_intr_info[irq].trigger  = trigger;
 614
 615         irq_type = iosapic_get_irq_chip(trigger);
 616
 617         idesc = irq_desc + irq;
 618         if (irq_type != NULL && idesc->chip != irq_type) {
 619                 if (idesc->chip != &no_irq_chip)
 620                         printk(KERN_WARNING
 621                                "%s: changing vector %d from %s to %s\n",
 622                                __func__, irq_to_vector(irq),
 623                                idesc->chip->name, irq_type->name);
 624                 idesc->chip = irq_type;
 625         }
 626         if (trigger == IOSAPIC_EDGE)
 627                 __set_irq_handler_unlocked(irq, handle_edge_irq);
 628         else
 629                 __set_irq_handler_unlocked(irq, handle_level_irq);
 630         return 0;
 631 }
 632
 633 static unsigned int
 634 get_target_cpu (unsigned int gsi, int irq)
 635 {
 636 #ifdef CONFIG_SMP
 637         static int cpu = -1;
 638         extern int cpe_vector;
 639         cpumask_t domain = irq_to_domain(irq);
 640
 641         /*
 642          * In case of vector shared by multiple RTEs, all RTEs that
 643          * share the vector need to use the same destination CPU.
 644          */
 645         if (iosapic_intr_info[irq].count)
 646                 return iosapic_intr_info[irq].dest;
 647
 648         /*
 649          * If the platform supports redirection via XTP, let it
 650          * distribute interrupts.
 651          */
 652         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
 653                 return cpu_physical_id(smp_processor_id());
 654
 655         /*
 656          * Some interrupts (ACPI SCI, for instance) are registered
 657          * before the BSP is marked as online.
 658          */
 659         if (!cpu_online(smp_processor_id()))
 660                 return cpu_physical_id(smp_processor_id());
 661
 662 #ifdef CONFIG_ACPI
 663         if (cpe_vector > 0 && irq_to_vector(irq) == IA64_CPEP_VECTOR)
 664                 return get_cpei_target_cpu();
 665 #endif
 666
 667 #ifdef CONFIG_NUMA
 668         {
 669                 int num_cpus, cpu_index, iosapic_index, numa_cpu, i = 0;
 670                 const struct cpumask *cpu_mask;
 671
 672                 iosapic_index = find_iosapic(gsi);
 673                 if (iosapic_index < 0 ||
 674                     iosapic_lists[iosapic_index].node == MAX_NUMNODES)
 675                         goto skip_numa_setup;
 676
 677                 cpu_mask = cpumask_of_node(iosapic_lists[iosapic_index].node);
 678                 num_cpus = 0;
 679                 for_each_cpu_and(numa_cpu, cpu_mask, &domain) {
 680                         if (cpu_online(numa_cpu))
 681                                 num_cpus++;
 682                 }
 683
 684                 if (!num_cpus)
 685                         goto skip_numa_setup;
 686
 687                 /* Use irq assignment to distribute across cpus in node */
 688                 cpu_index = irq % num_cpus;
 689
 690                 for_each_cpu_and(numa_cpu, cpu_mask, &domain)
 691                         if (cpu_online(numa_cpu) && i++ >= cpu_index)
 692                                 break;
 693
 694                 if (numa_cpu < nr_cpu_ids)
 695                         return cpu_physical_id(numa_cpu);
 696         }
 697 skip_numa_setup:
 698 #endif
 699         /*
 700          * Otherwise, round-robin interrupt vectors across all the
 701          * processors.  (It'd be nice if we could be smarter in the
 702          * case of NUMA.)
 703          */
 704         do {
 705                 if (++cpu >= nr_cpu_ids)
 706                         cpu = 0;
 707         } while (!cpu_online(cpu) || !cpu_isset(cpu, domain));
 708
 709         return cpu_physical_id(cpu);
 710 #else  /* CONFIG_SMP */
 711         return cpu_physical_id(smp_processor_id());
 712 #endif
 713 }
 714
 715 static inline unsigned char choose_dmode(void)
 716 {
 717 #ifdef CONFIG_SMP
 718         if (smp_int_redirect & SMP_IRQ_REDIRECTION)
 719                 return IOSAPIC_LOWEST_PRIORITY;
 720 #endif
 721         return IOSAPIC_FIXED;
 722 }
 723
 724 /*
 725  * ACPI can describe IOSAPIC interrupts via static tables and namespace
 726  * methods.  This provides an interface to register those interrupts and
 727  * program the IOSAPIC RTE.
 728  */
 729 int
 730 iosapic_register_intr (unsigned int gsi,
 731                        unsigned long polarity, unsigned long trigger)
 732 {
 733         int irq, mask = 1, err;
 734         unsigned int dest;
 735         unsigned long flags;
 736         struct iosapic_rte_info *rte;
 737         u32 low32;
 738         unsigned char dmode;
 739
 740         /*
 741          * If this GSI has already been registered (i.e., it's a
 742          * shared interrupt, or we lost a race to register it),
 743          * don't touch the RTE.
 744          */
 745         spin_lock_irqsave(&iosapic_lock, flags);
 746         irq = __gsi_to_irq(gsi);
 747         if (irq > 0) {
 748                 rte = find_rte(irq, gsi);
 749                 if(iosapic_intr_info[irq].count == 0) {
 750                         assign_irq_vector(irq);
 751                         dynamic_irq_init(irq);
 752                 } else if (rte->refcnt != NO_REF_RTE) {
 753                         rte->refcnt++;
 754                         goto unlock_iosapic_lock;
 755                 }
 756         } else
 757                 irq = create_irq();
 758
 759         /* If vector is running out, we try to find a sharable vector */
 760         if (irq < 0) {
 761                 irq = iosapic_find_sharable_irq(trigger, polarity);
 762                 if (irq < 0)
 763                         goto unlock_iosapic_lock;
 764         }
 765
 766         raw_spin_lock(&irq_desc[irq].lock);
 767         dest = get_target_cpu(gsi, irq);
 768         dmode = choose_dmode();
 769         err = register_intr(gsi, irq, dmode, polarity, trigger);
 770         if (err < 0) {
 771                 raw_spin_unlock(&irq_desc[irq].lock);
 772                 irq = err;
 773                 goto unlock_iosapic_lock;
 774         }
 775
 776         /*
 777          * If the vector is shared and already unmasked for other
 778          * interrupt sources, don't mask it.
 779          */
 780         low32 = iosapic_intr_info[irq].low32;
 781         if (irq_is_shared(irq) && !(low32 & IOSAPIC_MASK))
 782                 mask = 0;
 783         set_rte(gsi, irq, dest, mask);
 784
 785         printk(KERN_INFO "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d\n",
 786                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 787                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 788                cpu_logical_id(dest), dest, irq_to_vector(irq));
 789
 790         raw_spin_unlock(&irq_desc[irq].lock);
 791  unlock_iosapic_lock:
 792         spin_unlock_irqrestore(&iosapic_lock, flags);
 793         return irq;
 794 }
 795
 796 void
 797 iosapic_unregister_intr (unsigned int gsi)
 798 {
 799         unsigned long flags;
 800         int irq, index;
 801         struct irq_desc *idesc;
 802         u32 low32;
 803         unsigned long trigger, polarity;
 804         unsigned int dest;
 805         struct iosapic_rte_info *rte;
 806
 807         /*
 808          * If the irq associated with the gsi is not found,
 809          * iosapic_unregister_intr() is unbalanced. We need to check
 810          * this again after getting locks.
 811          */
 812         irq = gsi_to_irq(gsi);
 813         if (irq < 0) {
 814                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
 815                        gsi);
 816                 WARN_ON(1);
 817                 return;
 818         }
 819
 820         spin_lock_irqsave(&iosapic_lock, flags);
 821         if ((rte = find_rte(irq, gsi)) == NULL) {
 822                 printk(KERN_ERR "iosapic_unregister_intr(%u) unbalanced\n",
 823                        gsi);
 824                 WARN_ON(1);
 825                 goto out;
 826         }
 827
 828         if (--rte->refcnt > 0)
 829                 goto out;
 830
 831         idesc = irq_desc + irq;
 832         rte->refcnt = NO_REF_RTE;
 833
 834         /* Mask the interrupt */
 835         low32 = iosapic_intr_info[irq].low32 | IOSAPIC_MASK;
 836         iosapic_write(rte->iosapic, IOSAPIC_RTE_LOW(rte->rte_index), low32);
 837
 838         iosapic_intr_info[irq].count--;
 839         index = find_iosapic(gsi);
 840         iosapic_lists[index].rtes_inuse--;
 841         WARN_ON(iosapic_lists[index].rtes_inuse < 0);
 842
 843         trigger  = iosapic_intr_info[irq].trigger;
 844         polarity = iosapic_intr_info[irq].polarity;
 845         dest     = iosapic_intr_info[irq].dest;
 846         printk(KERN_INFO
 847                "GSI %u (%s, %s) -> CPU %d (0x%04x) vector %d unregistered\n",
 848                gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 849                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 850                cpu_logical_id(dest), dest, irq_to_vector(irq));
 851
 852         if (iosapic_intr_info[irq].count == 0) {
 853 #ifdef CONFIG_SMP
 854                 /* Clear affinity */
 855                 cpumask_setall(idesc->affinity);
 856 #endif
 857                 /* Clear the interrupt information */
 858                 iosapic_intr_info[irq].dest = 0;
 859                 iosapic_intr_info[irq].dmode = 0;
 860                 iosapic_intr_info[irq].polarity = 0;
 861                 iosapic_intr_info[irq].trigger = 0;
 862                 iosapic_intr_info[irq].low32 |= IOSAPIC_MASK;
 863
 864                 /* Destroy and reserve IRQ */
 865                 destroy_and_reserve_irq(irq);
 866         }
 867  out:
 868         spin_unlock_irqrestore(&iosapic_lock, flags);
 869 }
 870
 871 /*
 872  * ACPI calls this when it finds an entry for a platform interrupt.
 873  */
 874 int __init
 875 iosapic_register_platform_intr (u32 int_type, unsigned int gsi,
 876                                 int iosapic_vector, u16 eid, u16 id,
 877                                 unsigned long polarity, unsigned long trigger)
 878 {
 879         static const char * const name[] = {"unknown", "PMI", "INIT", "CPEI"};
 880         unsigned char delivery;
 881         int irq, vector, mask = 0;
 882         unsigned int dest = ((id << 8) | eid) & 0xffff;
 883
 884         switch (int_type) {
 885               case ACPI_INTERRUPT_PMI:
 886                 irq = vector = iosapic_vector;
 887                 bind_irq_vector(irq, vector, CPU_MASK_ALL);
 888                 /*
 889                  * since PMI vector is alloc'd by FW(ACPI) not by kernel,
 890                  * we need to make sure the vector is available
 891                  */
 892                 iosapic_reassign_vector(irq);
 893                 delivery = IOSAPIC_PMI;
 894                 break;
 895               case ACPI_INTERRUPT_INIT:
 896                 irq = create_irq();
 897                 if (irq < 0)
 898                         panic("%s: out of interrupt vectors!\n", __func__);
 899                 vector = irq_to_vector(irq);
 900                 delivery = IOSAPIC_INIT;
 901                 break;
 902               case ACPI_INTERRUPT_CPEI:
 903                 irq = vector = IA64_CPE_VECTOR;
 904                 BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
 905                 delivery = IOSAPIC_FIXED;
 906                 mask = 1;
 907                 break;
 908               default:
 909                 printk(KERN_ERR "%s: invalid int type 0x%x\n", __func__,
 910                        int_type);
 911                 return -1;
 912         }
 913
 914         register_intr(gsi, irq, delivery, polarity, trigger);
 915
 916         printk(KERN_INFO
 917                "PLATFORM int %s (0x%x): GSI %u (%s, %s) -> CPU %d (0x%04x)"
 918                " vector %d\n",
 919                int_type < ARRAY_SIZE(name) ? name[int_type] : "unknown",
 920                int_type, gsi, (trigger == IOSAPIC_EDGE ? "edge" : "level"),
 921                (polarity == IOSAPIC_POL_HIGH ? "high" : "low"),
 922                cpu_logical_id(dest), dest, vector);
 923
 924         set_rte(gsi, irq, dest, mask);
 925         return vector;
 926 }
 927
 928 /*
 929  * ACPI calls this when it finds an entry for a legacy ISA IRQ override.
 930  */
 931 void __devinit
 932 iosapic_override_isa_irq (unsigned int isa_irq, unsigned int gsi,
 933                           unsigned long polarity,
 934                           unsigned long trigger)
 935 {
 936         int vector, irq;
 937         unsigned int dest = cpu_physical_id(smp_processor_id());
 938         unsigned char dmode;
 939
 940         irq = vector = isa_irq_to_vector(isa_irq);
 941         BUG_ON(bind_irq_vector(irq, vector, CPU_MASK_ALL));
 942         dmode = choose_dmode();
 943         register_intr(gsi, irq, dmode, polarity, trigger);
 944
 945         DBG("ISA: IRQ %u -> GSI %u (%s,%s) -> CPU %d (0x%04x) vector %d\n",
 946             isa_irq, gsi, trigger == IOSAPIC_EDGE ? "edge" : "level",
 947             polarity == IOSAPIC_POL_HIGH ? "high" : "low",
 948             cpu_logical_id(dest), dest, vector);
 949
 950         set_rte(gsi, irq, dest, 1);
 951 }
 952
 953 void __init
 954 ia64_native_iosapic_pcat_compat_init(void)
 955 {
 956         if (pcat_compat) {
 957                 /*
 958                  * Disable the compatibility mode interrupts (8259 style),
 959                  * needs IN/OUT support enabled.
 960                  */
 961                 printk(KERN_INFO
 962                        "%s: Disabling PC-AT compatible 8259 interrupts\n",
 963                        __func__);
 964                 outb(0xff, 0xA1);
 965                 outb(0xff, 0x21);
 966         }
 967 }
 968
 969 void __init
 970 iosapic_system_init (int system_pcat_compat)
 971 {
 972         int irq;
 973
 974         for (irq = 0; irq < NR_IRQS; ++irq) {
 975                 iosapic_intr_info[irq].low32 = IOSAPIC_MASK;
 976                 /* mark as unused */
 977                 INIT_LIST_HEAD(&iosapic_intr_info[irq].rtes);
 978
 979                 iosapic_intr_info[irq].count = 0;
 980         }
 981
 982         pcat_compat = system_pcat_compat;
 983         if (pcat_compat)
 984                 iosapic_pcat_compat_init();
 985 }
 986
 987 static inline int
 988 iosapic_alloc (void)
 989 {
 990         int index;
 991
 992         for (index = 0; index < NR_IOSAPICS; index++)
 993                 if (!iosapic_lists[index].addr)
 994                         return index;
 995
 996         printk(KERN_WARNING "%s: failed to allocate iosapic\n", __func__);
 997         return -1;
 998 }
 999
1000 static inline void
1001 iosapic_free (int index)
1002 {
1003         memset(&iosapic_lists[index], 0, sizeof(iosapic_lists[0]));
1004 }
1005
1006 static inline int
1007 iosapic_check_gsi_range (unsigned int gsi_base, unsigned int ver)
1008 {
1009         int index;
1010         unsigned int gsi_end, base, end;
1011
1012         /* check gsi range */
1013         gsi_end = gsi_base + ((ver >> 16) & 0xff);
1014         for (index = 0; index < NR_IOSAPICS; index++) {
1015                 if (!iosapic_lists[index].addr)
1016                         continue;
1017
1018                 base = iosapic_lists[index].gsi_base;
1019                 end  = base + iosapic_lists[index].num_rte - 1;
1020
1021                 if (gsi_end < base || end < gsi_base)
1022                         continue; /* OK */
1023
1024                 return -EBUSY;
1025         }
1026         return 0;
1027 }
1028
1029 int __devinit
1030 iosapic_init (unsigned long phys_addr, unsigned int gsi_base)
1031 {
1032         int num_rte, err, index;
1033         unsigned int isa_irq, ver;
1034         char __iomem *addr;
1035         unsigned long flags;
1036
1037         spin_lock_irqsave(&iosapic_lock, flags);
1038         index = find_iosapic(gsi_base);
1039         if (index >= 0) {
1040                 spin_unlock_irqrestore(&iosapic_lock, flags);
1041                 return -EBUSY;
1042         }
1043
1044         addr = ioremap(phys_addr, 0);
1045         if (addr == NULL) {
1046                 spin_unlock_irqrestore(&iosapic_lock, flags);
1047                 return -ENOMEM;
1048         }
1049         ver = iosapic_version(addr);
1050         if ((err = iosapic_check_gsi_range(gsi_base, ver))) {
1051                 iounmap(addr);
1052                 spin_unlock_irqrestore(&iosapic_lock, flags);
1053                 return err;
1054         }
1055
1056         /*
1057          * The MAX_REDIR register holds the highest input pin number
1058          * (starting from 0).  We add 1 so that we can use it for
1059          * number of pins (= RTEs)
1060          */
1061         num_rte = ((ver >> 16) & 0xff) + 1;
1062
1063         index = iosapic_alloc();
1064         iosapic_lists[index].addr = addr;
1065         iosapic_lists[index].gsi_base = gsi_base;
1066         iosapic_lists[index].num_rte = num_rte;
1067 #ifdef CONFIG_NUMA
1068         iosapic_lists[index].node = MAX_NUMNODES;
1069 #endif
1070         spin_lock_init(&iosapic_lists[index].lock);
1071         spin_unlock_irqrestore(&iosapic_lock, flags);
1072
1073         if ((gsi_base == 0) && pcat_compat) {
1074                 /*
1075                  * Map the legacy ISA devices into the IOSAPIC data.  Some of
1076                  * these may get reprogrammed later on with data from the ACPI
1077                  * Interrupt Source Override table.
1078                  */
1079                 for (isa_irq = 0; isa_irq < 16; ++isa_irq)
1080                         iosapic_override_isa_irq(isa_irq, isa_irq,
1081                                                  IOSAPIC_POL_HIGH,
1082                                                  IOSAPIC_EDGE);
1083         }
1084         return 0;
1085 }
1086
1087 #ifdef CONFIG_HOTPLUG
1088 int
1089 iosapic_remove (unsigned int gsi_base)
1090 {
1091         int index, err = 0;
1092         unsigned long flags;
1093
1094         spin_lock_irqsave(&iosapic_lock, flags);
1095         index = find_iosapic(gsi_base);
1096         if (index < 0) {
1097                 printk(KERN_WARNING "%s: No IOSAPIC for GSI base %u\n",
1098                        __func__, gsi_base);
1099                 goto out;
1100         }
1101
1102         if (iosapic_lists[index].rtes_inuse) {
1103                 err = -EBUSY;
1104                 printk(KERN_WARNING "%s: IOSAPIC for GSI base %u is busy\n",
1105                        __func__, gsi_base);
1106                 goto out;
1107         }
1108
1109         iounmap(iosapic_lists[index].addr);
1110         iosapic_free(index);
1111  out:
1112         spin_unlock_irqrestore(&iosapic_lock, flags);
1113         return err;
1114 }
1115 #endif /* CONFIG_HOTPLUG */
1116
1117 #ifdef CONFIG_NUMA
1118 void __devinit
1119 map_iosapic_to_node(unsigned int gsi_base, int node)
1120 {
1121         int index;
1122
1123         index = find_iosapic(gsi_base);
1124         if (index < 0) {
1125                 printk(KERN_WARNING "%s: No IOSAPIC for GSI %u\n",
1126                        __func__, gsi_base);
1127                 return;
1128         }
1129         iosapic_lists[index].node = node;
1130         return;
1131 }
1132 #endif