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

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is recieved, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. Hardware interrupts. Not supported at present.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29
  30 #include <asm/ptrace.h>
  31 #include <asm/irq.h>
  32 #include <asm/sync_bitops.h>
  33 #include <asm/xen/hypercall.h>
  34 #include <asm/xen/hypervisor.h>
  35
  36 #include <xen/events.h>
  37 #include <xen/interface/xen.h>
  38 #include <xen/interface/event_channel.h>
  39
  40 #include "xen-ops.h"
  41
  42 /*
  43  * This lock protects updates to the following mapping and reference-count
  44  * arrays. The lock does not need to be acquired to read the mapping tables.
  45  */
  46 static DEFINE_SPINLOCK(irq_mapping_update_lock);
  47
  48 /* IRQ <-> VIRQ mapping. */
  49 static DEFINE_PER_CPU(int, virq_to_irq[NR_VIRQS]) = {[0 ... NR_VIRQS-1] = -1};
  50
  51 /* IRQ <-> IPI mapping */
  52 static DEFINE_PER_CPU(int, ipi_to_irq[XEN_NR_IPIS]) = {[0 ... XEN_NR_IPIS-1] = -1};
  53
  54 /* Packed IRQ information: binding type, sub-type index, and event channel. */
  55 struct packed_irq
  56 {
  57         unsigned short evtchn;
  58         unsigned char index;
  59         unsigned char type;
  60 };
  61
  62 static struct packed_irq irq_info[NR_IRQS];
  63
  64 /* Binding types. */
  65 enum {
  66         IRQT_UNBOUND,
  67         IRQT_PIRQ,
  68         IRQT_VIRQ,
  69         IRQT_IPI,
  70         IRQT_EVTCHN
  71 };
  72
  73 /* Convenient shorthand for packed representation of an unbound IRQ. */
  74 #define IRQ_UNBOUND     mk_irq_info(IRQT_UNBOUND, 0, 0)
  75
  76 static int evtchn_to_irq[NR_EVENT_CHANNELS] = {
  77         [0 ... NR_EVENT_CHANNELS-1] = -1
  78 };
  79 static unsigned long cpu_evtchn_mask[NR_CPUS][NR_EVENT_CHANNELS/BITS_PER_LONG];
  80 static u8 cpu_evtchn[NR_EVENT_CHANNELS];
  81
  82 /* Reference counts for bindings to IRQs. */
  83 static int irq_bindcount[NR_IRQS];
  84
  85 /* Xen will never allocate port zero for any purpose. */
  86 #define VALID_EVTCHN(chn)       ((chn) != 0)
  87
  88 /*
  89  * Force a proper event-channel callback from Xen after clearing the
  90  * callback mask. We do this in a very simple manner, by making a call
  91  * down into Xen. The pending flag will be checked by Xen on return.
  92  */
  93 void force_evtchn_callback(void)
  94 {
  95         (void)HYPERVISOR_xen_version(0, NULL);
  96 }
  97 EXPORT_SYMBOL_GPL(force_evtchn_callback);
  98
  99 static struct irq_chip xen_dynamic_chip;
 100
 101 /* Constructor for packed IRQ information. */
 102 static inline struct packed_irq mk_irq_info(u32 type, u32 index, u32 evtchn)
 103 {
 104         return (struct packed_irq) { evtchn, index, type };
 105 }
 106
 107 /*
 108  * Accessors for packed IRQ information.
 109  */
 110 static inline unsigned int evtchn_from_irq(int irq)
 111 {
 112         return irq_info[irq].evtchn;
 113 }
 114
 115 static inline unsigned int index_from_irq(int irq)
 116 {
 117         return irq_info[irq].index;
 118 }
 119
 120 static inline unsigned int type_from_irq(int irq)
 121 {
 122         return irq_info[irq].type;
 123 }
 124
 125 static inline unsigned long active_evtchns(unsigned int cpu,
 126                                            struct shared_info *sh,
 127                                            unsigned int idx)
 128 {
 129         return (sh->evtchn_pending[idx] &
 130                 cpu_evtchn_mask[cpu][idx] &
 131                 ~sh->evtchn_mask[idx]);
 132 }
 133
 134 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 135 {
 136         int irq = evtchn_to_irq[chn];
 137
 138         BUG_ON(irq == -1);
 139 #ifdef CONFIG_SMP
 140         irq_desc[irq].affinity = cpumask_of_cpu(cpu);
 141 #endif
 142
 143         __clear_bit(chn, cpu_evtchn_mask[cpu_evtchn[chn]]);
 144         __set_bit(chn, cpu_evtchn_mask[cpu]);
 145
 146         cpu_evtchn[chn] = cpu;
 147 }
 148
 149 static void init_evtchn_cpu_bindings(void)
 150 {
 151 #ifdef CONFIG_SMP
 152         int i;
 153         /* By default all event channels notify CPU#0. */
 154         for (i = 0; i < NR_IRQS; i++)
 155                 irq_desc[i].affinity = cpumask_of_cpu(0);
 156 #endif
 157
 158         memset(cpu_evtchn, 0, sizeof(cpu_evtchn));
 159         memset(cpu_evtchn_mask[0], ~0, sizeof(cpu_evtchn_mask[0]));
 160 }
 161
 162 static inline unsigned int cpu_from_evtchn(unsigned int evtchn)
 163 {
 164         return cpu_evtchn[evtchn];
 165 }
 166
 167 static inline void clear_evtchn(int port)
 168 {
 169         struct shared_info *s = HYPERVISOR_shared_info;
 170         sync_clear_bit(port, &s->evtchn_pending[0]);
 171 }
 172
 173 static inline void set_evtchn(int port)
 174 {
 175         struct shared_info *s = HYPERVISOR_shared_info;
 176         sync_set_bit(port, &s->evtchn_pending[0]);
 177 }
 178
 179
 180 /**
 181  * notify_remote_via_irq - send event to remote end of event channel via irq
 182  * @irq: irq of event channel to send event to
 183  *
 184  * Unlike notify_remote_via_evtchn(), this is safe to use across
 185  * save/restore. Notifications on a broken connection are silently
 186  * dropped.
 187  */
 188 void notify_remote_via_irq(int irq)
 189 {
 190         int evtchn = evtchn_from_irq(irq);
 191
 192         if (VALID_EVTCHN(evtchn))
 193                 notify_remote_via_evtchn(evtchn);
 194 }
 195 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 196
 197 static void mask_evtchn(int port)
 198 {
 199         struct shared_info *s = HYPERVISOR_shared_info;
 200         sync_set_bit(port, &s->evtchn_mask[0]);
 201 }
 202
 203 static void unmask_evtchn(int port)
 204 {
 205         struct shared_info *s = HYPERVISOR_shared_info;
 206         unsigned int cpu = get_cpu();
 207
 208         BUG_ON(!irqs_disabled());
 209
 210         /* Slow path (hypercall) if this is a non-local port. */
 211         if (unlikely(cpu != cpu_from_evtchn(port))) {
 212                 struct evtchn_unmask unmask = { .port = port };
 213                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 214         } else {
 215                 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 216
 217                 sync_clear_bit(port, &s->evtchn_mask[0]);
 218
 219                 /*
 220                  * The following is basically the equivalent of
 221                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 222                  * the interrupt edge' if the channel is masked.
 223                  */
 224                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 225                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 226                                            &vcpu_info->evtchn_pending_sel))
 227                         vcpu_info->evtchn_upcall_pending = 1;
 228         }
 229
 230         put_cpu();
 231 }
 232
 233 static int find_unbound_irq(void)
 234 {
 235         int irq;
 236
 237         /* Only allocate from dynirq range */
 238         for (irq = 0; irq < NR_IRQS; irq++)
 239                 if (irq_bindcount[irq] == 0)
 240                         break;
 241
 242         if (irq == NR_IRQS)
 243                 panic("No available IRQ to bind to: increase NR_IRQS!\n");
 244
 245         return irq;
 246 }
 247
 248 int bind_evtchn_to_irq(unsigned int evtchn)
 249 {
 250         int irq;
 251
 252         spin_lock(&irq_mapping_update_lock);
 253
 254         irq = evtchn_to_irq[evtchn];
 255
 256         if (irq == -1) {
 257                 irq = find_unbound_irq();
 258
 259                 dynamic_irq_init(irq);
 260                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
 261                                               handle_level_irq, "event");
 262
 263                 evtchn_to_irq[evtchn] = irq;
 264                 irq_info[irq] = mk_irq_info(IRQT_EVTCHN, 0, evtchn);
 265         }
 266
 267         irq_bindcount[irq]++;
 268
 269         spin_unlock(&irq_mapping_update_lock);
 270
 271         return irq;
 272 }
 273 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 274
 275 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 276 {
 277         struct evtchn_bind_ipi bind_ipi;
 278         int evtchn, irq;
 279
 280         spin_lock(&irq_mapping_update_lock);
 281
 282         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 283         if (irq == -1) {
 284                 irq = find_unbound_irq();
 285                 if (irq < 0)
 286                         goto out;
 287
 288                 dynamic_irq_init(irq);
 289                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
 290                                               handle_level_irq, "ipi");
 291
 292                 bind_ipi.vcpu = cpu;
 293                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 294                                                 &bind_ipi) != 0)
 295                         BUG();
 296                 evtchn = bind_ipi.port;
 297
 298                 evtchn_to_irq[evtchn] = irq;
 299                 irq_info[irq] = mk_irq_info(IRQT_IPI, ipi, evtchn);
 300
 301                 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 302
 303                 bind_evtchn_to_cpu(evtchn, cpu);
 304         }
 305
 306         irq_bindcount[irq]++;
 307
 308  out:
 309         spin_unlock(&irq_mapping_update_lock);
 310         return irq;
 311 }
 312
 313
 314 static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 315 {
 316         struct evtchn_bind_virq bind_virq;
 317         int evtchn, irq;
 318
 319         spin_lock(&irq_mapping_update_lock);
 320
 321         irq = per_cpu(virq_to_irq, cpu)[virq];
 322
 323         if (irq == -1) {
 324                 bind_virq.virq = virq;
 325                 bind_virq.vcpu = cpu;
 326                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 327                                                 &bind_virq) != 0)
 328                         BUG();
 329                 evtchn = bind_virq.port;
 330
 331                 irq = find_unbound_irq();
 332
 333                 dynamic_irq_init(irq);
 334                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
 335                                               handle_level_irq, "virq");
 336
 337                 evtchn_to_irq[evtchn] = irq;
 338                 irq_info[irq] = mk_irq_info(IRQT_VIRQ, virq, evtchn);
 339
 340                 per_cpu(virq_to_irq, cpu)[virq] = irq;
 341
 342                 bind_evtchn_to_cpu(evtchn, cpu);
 343         }
 344
 345         irq_bindcount[irq]++;
 346
 347         spin_unlock(&irq_mapping_update_lock);
 348
 349         return irq;
 350 }
 351
 352 static void unbind_from_irq(unsigned int irq)
 353 {
 354         struct evtchn_close close;
 355         int evtchn = evtchn_from_irq(irq);
 356
 357         spin_lock(&irq_mapping_update_lock);
 358
 359         if (VALID_EVTCHN(evtchn) && (--irq_bindcount[irq] == 0)) {
 360                 close.port = evtchn;
 361                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 362                         BUG();
 363
 364                 switch (type_from_irq(irq)) {
 365                 case IRQT_VIRQ:
 366                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 367                                 [index_from_irq(irq)] = -1;
 368                         break;
 369                 default:
 370                         break;
 371                 }
 372
 373                 /* Closed ports are implicitly re-bound to VCPU0. */
 374                 bind_evtchn_to_cpu(evtchn, 0);
 375
 376                 evtchn_to_irq[evtchn] = -1;
 377                 irq_info[irq] = IRQ_UNBOUND;
 378
 379                 dynamic_irq_init(irq);
 380         }
 381
 382         spin_unlock(&irq_mapping_update_lock);
 383 }
 384
 385 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 386                               irq_handler_t handler,
 387                               unsigned long irqflags,
 388                               const char *devname, void *dev_id)
 389 {
 390         unsigned int irq;
 391         int retval;
 392
 393         irq = bind_evtchn_to_irq(evtchn);
 394         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 395         if (retval != 0) {
 396                 unbind_from_irq(irq);
 397                 return retval;
 398         }
 399
 400         return irq;
 401 }
 402 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 403
 404 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
 405                             irq_handler_t handler,
 406                             unsigned long irqflags, const char *devname, void *dev_id)
 407 {
 408         unsigned int irq;
 409         int retval;
 410
 411         irq = bind_virq_to_irq(virq, cpu);
 412         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 413         if (retval != 0) {
 414                 unbind_from_irq(irq);
 415                 return retval;
 416         }
 417
 418         return irq;
 419 }
 420 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
 421
 422 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
 423                            unsigned int cpu,
 424                            irq_handler_t handler,
 425                            unsigned long irqflags,
 426                            const char *devname,
 427                            void *dev_id)
 428 {
 429         int irq, retval;
 430
 431         irq = bind_ipi_to_irq(ipi, cpu);
 432         if (irq < 0)
 433                 return irq;
 434
 435         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 436         if (retval != 0) {
 437                 unbind_from_irq(irq);
 438                 return retval;
 439         }
 440
 441         return irq;
 442 }
 443
 444 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
 445 {
 446         free_irq(irq, dev_id);
 447         unbind_from_irq(irq);
 448 }
 449 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
 450
 451 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
 452 {
 453         int irq = per_cpu(ipi_to_irq, cpu)[vector];
 454         BUG_ON(irq < 0);
 455         notify_remote_via_irq(irq);
 456 }
 457
 458 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
 459 {
 460         struct shared_info *sh = HYPERVISOR_shared_info;
 461         int cpu = smp_processor_id();
 462         int i;
 463         unsigned long flags;
 464         static DEFINE_SPINLOCK(debug_lock);
 465
 466         spin_lock_irqsave(&debug_lock, flags);
 467
 468         printk("vcpu %d\n  ", cpu);
 469
 470         for_each_online_cpu(i) {
 471                 struct vcpu_info *v = per_cpu(xen_vcpu, i);
 472                 printk("%d: masked=%d pending=%d event_sel %08lx\n  ", i,
 473                         (get_irq_regs() && i == cpu) ? !(get_irq_regs()->flags & X86_EFLAGS_IF) : v->evtchn_upcall_mask,
 474                         v->evtchn_upcall_pending,
 475                         v->evtchn_pending_sel);
 476         }
 477         printk("pending:\n   ");
 478         for(i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
 479                 printk("%08lx%s", sh->evtchn_pending[i],
 480                         i % 8 == 0 ? "\n   " : " ");
 481         printk("\nmasks:\n   ");
 482         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 483                 printk("%08lx%s", sh->evtchn_mask[i],
 484                         i % 8 == 0 ? "\n   " : " ");
 485
 486         printk("\nunmasked:\n   ");
 487         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 488                 printk("%08lx%s", sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
 489                         i % 8 == 0 ? "\n   " : " ");
 490
 491         printk("\npending list:\n");
 492         for(i = 0; i < NR_EVENT_CHANNELS; i++) {
 493                 if (sync_test_bit(i, sh->evtchn_pending)) {
 494                         printk("  %d: event %d -> irq %d\n",
 495                                 cpu_evtchn[i], i,
 496                                 evtchn_to_irq[i]);
 497                 }
 498         }
 499
 500         spin_unlock_irqrestore(&debug_lock, flags);
 501
 502         return IRQ_HANDLED;
 503 }
 504
 505
 506 /*
 507  * Search the CPUs pending events bitmasks.  For each one found, map
 508  * the event number to an irq, and feed it into do_IRQ() for
 509  * handling.
 510  *
 511  * Xen uses a two-level bitmap to speed searching.  The first level is
 512  * a bitset of words which contain pending event bits.  The second
 513  * level is a bitset of pending events themselves.
 514  */
 515 void xen_evtchn_do_upcall(struct pt_regs *regs)
 516 {
 517         int cpu = get_cpu();
 518         struct shared_info *s = HYPERVISOR_shared_info;
 519         struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 520         static DEFINE_PER_CPU(unsigned, nesting_count);
 521         unsigned count;
 522
 523         do {
 524                 unsigned long pending_words;
 525
 526                 vcpu_info->evtchn_upcall_pending = 0;
 527
 528                 if (__get_cpu_var(nesting_count)++)
 529                         goto out;
 530
 531                 /* NB. No need for a barrier here -- XCHG is a barrier on x86. */
 532                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
 533                 while (pending_words != 0) {
 534                         unsigned long pending_bits;
 535                         int word_idx = __ffs(pending_words);
 536                         pending_words &= ~(1UL << word_idx);
 537
 538                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
 539                                 int bit_idx = __ffs(pending_bits);
 540                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
 541                                 int irq = evtchn_to_irq[port];
 542
 543                                 if (irq != -1) {
 544                                         regs->orig_ax = ~irq;
 545                                         do_IRQ(regs);
 546                                 }
 547                         }
 548                 }
 549
 550                 BUG_ON(!irqs_disabled());
 551
 552                 count = __get_cpu_var(nesting_count);
 553                 __get_cpu_var(nesting_count) = 0;
 554         } while(count != 1);
 555
 556 out:
 557         put_cpu();
 558 }
 559
 560 /* Rebind an evtchn so that it gets delivered to a specific cpu */
 561 static void rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
 562 {
 563         struct evtchn_bind_vcpu bind_vcpu;
 564         int evtchn = evtchn_from_irq(irq);
 565
 566         if (!VALID_EVTCHN(evtchn))
 567                 return;
 568
 569         /* Send future instances of this interrupt to other vcpu. */
 570         bind_vcpu.port = evtchn;
 571         bind_vcpu.vcpu = tcpu;
 572
 573         /*
 574          * If this fails, it usually just indicates that we're dealing with a
 575          * virq or IPI channel, which don't actually need to be rebound. Ignore
 576          * it, but don't do the xenlinux-level rebind in that case.
 577          */
 578         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
 579                 bind_evtchn_to_cpu(evtchn, tcpu);
 580 }
 581
 582
 583 static void set_affinity_irq(unsigned irq, cpumask_t dest)
 584 {
 585         unsigned tcpu = first_cpu(dest);
 586         rebind_irq_to_cpu(irq, tcpu);
 587 }
 588
 589 static void enable_dynirq(unsigned int irq)
 590 {
 591         int evtchn = evtchn_from_irq(irq);
 592
 593         if (VALID_EVTCHN(evtchn))
 594                 unmask_evtchn(evtchn);
 595 }
 596
 597 static void disable_dynirq(unsigned int irq)
 598 {
 599         int evtchn = evtchn_from_irq(irq);
 600
 601         if (VALID_EVTCHN(evtchn))
 602                 mask_evtchn(evtchn);
 603 }
 604
 605 static void ack_dynirq(unsigned int irq)
 606 {
 607         int evtchn = evtchn_from_irq(irq);
 608
 609         move_native_irq(irq);
 610
 611         if (VALID_EVTCHN(evtchn))
 612                 clear_evtchn(evtchn);
 613 }
 614
 615 static int retrigger_dynirq(unsigned int irq)
 616 {
 617         int evtchn = evtchn_from_irq(irq);
 618         struct shared_info *sh = HYPERVISOR_shared_info;
 619         int ret = 0;
 620
 621         if (VALID_EVTCHN(evtchn)) {
 622                 int masked;
 623
 624                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
 625                 sync_set_bit(evtchn, sh->evtchn_pending);
 626                 if (!masked)
 627                         unmask_evtchn(evtchn);
 628                 ret = 1;
 629         }
 630
 631         return ret;
 632 }
 633
 634 static struct irq_chip xen_dynamic_chip __read_mostly = {
 635         .name           = "xen-dyn",
 636         .mask           = disable_dynirq,
 637         .unmask         = enable_dynirq,
 638         .ack            = ack_dynirq,
 639         .set_affinity   = set_affinity_irq,
 640         .retrigger      = retrigger_dynirq,
 641 };
 642
 643 void __init xen_init_IRQ(void)
 644 {
 645         int i;
 646
 647         init_evtchn_cpu_bindings();
 648
 649         /* No event channels are 'live' right now. */
 650         for (i = 0; i < NR_EVENT_CHANNELS; i++)
 651                 mask_evtchn(i);
 652
 653         /* Dynamic IRQ space is currently unbound. Zero the refcnts. */
 654         for (i = 0; i < NR_IRQS; i++)
 655                 irq_bindcount[i] = 0;
 656
 657         irq_ctx_init(smp_processor_id());
 658 }