Merge branch 'x86-debug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[~shefty/rdma-dev.git] / drivers / 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 received, 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. PIRQs - Hardware interrupts.
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 #include <linux/bootmem.h>
30 #include <linux/slab.h>
31 #include <linux/irqnr.h>
32 #include <linux/pci.h>
33
34 #ifdef CONFIG_X86
35 #include <asm/desc.h>
36 #include <asm/ptrace.h>
37 #include <asm/irq.h>
38 #include <asm/idle.h>
39 #include <asm/io_apic.h>
40 #include <asm/xen/page.h>
41 #include <asm/xen/pci.h>
42 #endif
43 #include <asm/sync_bitops.h>
44 #include <asm/xen/hypercall.h>
45 #include <asm/xen/hypervisor.h>
46
47 #include <xen/xen.h>
48 #include <xen/hvm.h>
49 #include <xen/xen-ops.h>
50 #include <xen/events.h>
51 #include <xen/interface/xen.h>
52 #include <xen/interface/event_channel.h>
53 #include <xen/interface/hvm/hvm_op.h>
54 #include <xen/interface/hvm/params.h>
55 #include <xen/interface/physdev.h>
56 #include <xen/interface/sched.h>
57 #include <asm/hw_irq.h>
58
59 /*
60  * This lock protects updates to the following mapping and reference-count
61  * arrays. The lock does not need to be acquired to read the mapping tables.
62  */
63 static DEFINE_MUTEX(irq_mapping_update_lock);
64
65 static LIST_HEAD(xen_irq_list_head);
66
67 /* IRQ <-> VIRQ mapping. */
68 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
69
70 /* IRQ <-> IPI mapping */
71 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
72
73 /* Interrupt types. */
74 enum xen_irq_type {
75         IRQT_UNBOUND = 0,
76         IRQT_PIRQ,
77         IRQT_VIRQ,
78         IRQT_IPI,
79         IRQT_EVTCHN
80 };
81
82 /*
83  * Packed IRQ information:
84  * type - enum xen_irq_type
85  * event channel - irq->event channel mapping
86  * cpu - cpu this event channel is bound to
87  * index - type-specific information:
88  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
89  *           guest, or GSI (real passthrough IRQ) of the device.
90  *    VIRQ - virq number
91  *    IPI - IPI vector
92  *    EVTCHN -
93  */
94 struct irq_info {
95         struct list_head list;
96         int refcnt;
97         enum xen_irq_type type; /* type */
98         unsigned irq;
99         unsigned short evtchn;  /* event channel */
100         unsigned short cpu;     /* cpu bound */
101
102         union {
103                 unsigned short virq;
104                 enum ipi_vector ipi;
105                 struct {
106                         unsigned short pirq;
107                         unsigned short gsi;
108                         unsigned char vector;
109                         unsigned char flags;
110                         uint16_t domid;
111                 } pirq;
112         } u;
113 };
114 #define PIRQ_NEEDS_EOI  (1 << 0)
115 #define PIRQ_SHAREABLE  (1 << 1)
116
117 static int *evtchn_to_irq;
118 #ifdef CONFIG_X86
119 static unsigned long *pirq_eoi_map;
120 #endif
121 static bool (*pirq_needs_eoi)(unsigned irq);
122
123 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
124                       cpu_evtchn_mask);
125
126 /* Xen will never allocate port zero for any purpose. */
127 #define VALID_EVTCHN(chn)       ((chn) != 0)
128
129 static struct irq_chip xen_dynamic_chip;
130 static struct irq_chip xen_percpu_chip;
131 static struct irq_chip xen_pirq_chip;
132 static void enable_dynirq(struct irq_data *data);
133 static void disable_dynirq(struct irq_data *data);
134
135 /* Get info for IRQ */
136 static struct irq_info *info_for_irq(unsigned irq)
137 {
138         return irq_get_handler_data(irq);
139 }
140
141 /* Constructors for packed IRQ information. */
142 static void xen_irq_info_common_init(struct irq_info *info,
143                                      unsigned irq,
144                                      enum xen_irq_type type,
145                                      unsigned short evtchn,
146                                      unsigned short cpu)
147 {
148
149         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
150
151         info->type = type;
152         info->irq = irq;
153         info->evtchn = evtchn;
154         info->cpu = cpu;
155
156         evtchn_to_irq[evtchn] = irq;
157 }
158
159 static void xen_irq_info_evtchn_init(unsigned irq,
160                                      unsigned short evtchn)
161 {
162         struct irq_info *info = info_for_irq(irq);
163
164         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
165 }
166
167 static void xen_irq_info_ipi_init(unsigned cpu,
168                                   unsigned irq,
169                                   unsigned short evtchn,
170                                   enum ipi_vector ipi)
171 {
172         struct irq_info *info = info_for_irq(irq);
173
174         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
175
176         info->u.ipi = ipi;
177
178         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
179 }
180
181 static void xen_irq_info_virq_init(unsigned cpu,
182                                    unsigned irq,
183                                    unsigned short evtchn,
184                                    unsigned short virq)
185 {
186         struct irq_info *info = info_for_irq(irq);
187
188         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
189
190         info->u.virq = virq;
191
192         per_cpu(virq_to_irq, cpu)[virq] = irq;
193 }
194
195 static void xen_irq_info_pirq_init(unsigned irq,
196                                    unsigned short evtchn,
197                                    unsigned short pirq,
198                                    unsigned short gsi,
199                                    unsigned short vector,
200                                    uint16_t domid,
201                                    unsigned char flags)
202 {
203         struct irq_info *info = info_for_irq(irq);
204
205         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
206
207         info->u.pirq.pirq = pirq;
208         info->u.pirq.gsi = gsi;
209         info->u.pirq.vector = vector;
210         info->u.pirq.domid = domid;
211         info->u.pirq.flags = flags;
212 }
213
214 /*
215  * Accessors for packed IRQ information.
216  */
217 static unsigned int evtchn_from_irq(unsigned irq)
218 {
219         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
220                 return 0;
221
222         return info_for_irq(irq)->evtchn;
223 }
224
225 unsigned irq_from_evtchn(unsigned int evtchn)
226 {
227         return evtchn_to_irq[evtchn];
228 }
229 EXPORT_SYMBOL_GPL(irq_from_evtchn);
230
231 static enum ipi_vector ipi_from_irq(unsigned irq)
232 {
233         struct irq_info *info = info_for_irq(irq);
234
235         BUG_ON(info == NULL);
236         BUG_ON(info->type != IRQT_IPI);
237
238         return info->u.ipi;
239 }
240
241 static unsigned virq_from_irq(unsigned irq)
242 {
243         struct irq_info *info = info_for_irq(irq);
244
245         BUG_ON(info == NULL);
246         BUG_ON(info->type != IRQT_VIRQ);
247
248         return info->u.virq;
249 }
250
251 static unsigned pirq_from_irq(unsigned irq)
252 {
253         struct irq_info *info = info_for_irq(irq);
254
255         BUG_ON(info == NULL);
256         BUG_ON(info->type != IRQT_PIRQ);
257
258         return info->u.pirq.pirq;
259 }
260
261 static enum xen_irq_type type_from_irq(unsigned irq)
262 {
263         return info_for_irq(irq)->type;
264 }
265
266 static unsigned cpu_from_irq(unsigned irq)
267 {
268         return info_for_irq(irq)->cpu;
269 }
270
271 static unsigned int cpu_from_evtchn(unsigned int evtchn)
272 {
273         int irq = evtchn_to_irq[evtchn];
274         unsigned ret = 0;
275
276         if (irq != -1)
277                 ret = cpu_from_irq(irq);
278
279         return ret;
280 }
281
282 #ifdef CONFIG_X86
283 static bool pirq_check_eoi_map(unsigned irq)
284 {
285         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
286 }
287 #endif
288
289 static bool pirq_needs_eoi_flag(unsigned irq)
290 {
291         struct irq_info *info = info_for_irq(irq);
292         BUG_ON(info->type != IRQT_PIRQ);
293
294         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
295 }
296
297 static inline unsigned long active_evtchns(unsigned int cpu,
298                                            struct shared_info *sh,
299                                            unsigned int idx)
300 {
301         return sh->evtchn_pending[idx] &
302                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
303                 ~sh->evtchn_mask[idx];
304 }
305
306 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
307 {
308         int irq = evtchn_to_irq[chn];
309
310         BUG_ON(irq == -1);
311 #ifdef CONFIG_SMP
312         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
313 #endif
314
315         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
316         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
317
318         info_for_irq(irq)->cpu = cpu;
319 }
320
321 static void init_evtchn_cpu_bindings(void)
322 {
323         int i;
324 #ifdef CONFIG_SMP
325         struct irq_info *info;
326
327         /* By default all event channels notify CPU#0. */
328         list_for_each_entry(info, &xen_irq_list_head, list) {
329                 struct irq_desc *desc = irq_to_desc(info->irq);
330                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
331         }
332 #endif
333
334         for_each_possible_cpu(i)
335                 memset(per_cpu(cpu_evtchn_mask, i),
336                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
337 }
338
339 static inline void clear_evtchn(int port)
340 {
341         struct shared_info *s = HYPERVISOR_shared_info;
342         sync_clear_bit(port, &s->evtchn_pending[0]);
343 }
344
345 static inline void set_evtchn(int port)
346 {
347         struct shared_info *s = HYPERVISOR_shared_info;
348         sync_set_bit(port, &s->evtchn_pending[0]);
349 }
350
351 static inline int test_evtchn(int port)
352 {
353         struct shared_info *s = HYPERVISOR_shared_info;
354         return sync_test_bit(port, &s->evtchn_pending[0]);
355 }
356
357
358 /**
359  * notify_remote_via_irq - send event to remote end of event channel via irq
360  * @irq: irq of event channel to send event to
361  *
362  * Unlike notify_remote_via_evtchn(), this is safe to use across
363  * save/restore. Notifications on a broken connection are silently
364  * dropped.
365  */
366 void notify_remote_via_irq(int irq)
367 {
368         int evtchn = evtchn_from_irq(irq);
369
370         if (VALID_EVTCHN(evtchn))
371                 notify_remote_via_evtchn(evtchn);
372 }
373 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
374
375 static void mask_evtchn(int port)
376 {
377         struct shared_info *s = HYPERVISOR_shared_info;
378         sync_set_bit(port, &s->evtchn_mask[0]);
379 }
380
381 static void unmask_evtchn(int port)
382 {
383         struct shared_info *s = HYPERVISOR_shared_info;
384         unsigned int cpu = get_cpu();
385         int do_hypercall = 0, evtchn_pending = 0;
386
387         BUG_ON(!irqs_disabled());
388
389         if (unlikely((cpu != cpu_from_evtchn(port))))
390                 do_hypercall = 1;
391         else
392                 evtchn_pending = sync_test_bit(port, &s->evtchn_pending[0]);
393
394         if (unlikely(evtchn_pending && xen_hvm_domain()))
395                 do_hypercall = 1;
396
397         /* Slow path (hypercall) if this is a non-local port or if this is
398          * an hvm domain and an event is pending (hvm domains don't have
399          * their own implementation of irq_enable). */
400         if (do_hypercall) {
401                 struct evtchn_unmask unmask = { .port = port };
402                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
403         } else {
404                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
405
406                 sync_clear_bit(port, &s->evtchn_mask[0]);
407
408                 /*
409                  * The following is basically the equivalent of
410                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
411                  * the interrupt edge' if the channel is masked.
412                  */
413                 if (evtchn_pending &&
414                     !sync_test_and_set_bit(port / BITS_PER_LONG,
415                                            &vcpu_info->evtchn_pending_sel))
416                         vcpu_info->evtchn_upcall_pending = 1;
417         }
418
419         put_cpu();
420 }
421
422 static void xen_irq_init(unsigned irq)
423 {
424         struct irq_info *info;
425 #ifdef CONFIG_SMP
426         struct irq_desc *desc = irq_to_desc(irq);
427
428         /* By default all event channels notify CPU#0. */
429         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
430 #endif
431
432         info = kzalloc(sizeof(*info), GFP_KERNEL);
433         if (info == NULL)
434                 panic("Unable to allocate metadata for IRQ%d\n", irq);
435
436         info->type = IRQT_UNBOUND;
437         info->refcnt = -1;
438
439         irq_set_handler_data(irq, info);
440
441         list_add_tail(&info->list, &xen_irq_list_head);
442 }
443
444 static int __must_check xen_allocate_irq_dynamic(void)
445 {
446         int first = 0;
447         int irq;
448
449 #ifdef CONFIG_X86_IO_APIC
450         /*
451          * For an HVM guest or domain 0 which see "real" (emulated or
452          * actual respectively) GSIs we allocate dynamic IRQs
453          * e.g. those corresponding to event channels or MSIs
454          * etc. from the range above those "real" GSIs to avoid
455          * collisions.
456          */
457         if (xen_initial_domain() || xen_hvm_domain())
458                 first = get_nr_irqs_gsi();
459 #endif
460
461         irq = irq_alloc_desc_from(first, -1);
462
463         if (irq >= 0)
464                 xen_irq_init(irq);
465
466         return irq;
467 }
468
469 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
470 {
471         int irq;
472
473         /*
474          * A PV guest has no concept of a GSI (since it has no ACPI
475          * nor access to/knowledge of the physical APICs). Therefore
476          * all IRQs are dynamically allocated from the entire IRQ
477          * space.
478          */
479         if (xen_pv_domain() && !xen_initial_domain())
480                 return xen_allocate_irq_dynamic();
481
482         /* Legacy IRQ descriptors are already allocated by the arch. */
483         if (gsi < NR_IRQS_LEGACY)
484                 irq = gsi;
485         else
486                 irq = irq_alloc_desc_at(gsi, -1);
487
488         xen_irq_init(irq);
489
490         return irq;
491 }
492
493 static void xen_free_irq(unsigned irq)
494 {
495         struct irq_info *info = irq_get_handler_data(irq);
496
497         list_del(&info->list);
498
499         irq_set_handler_data(irq, NULL);
500
501         WARN_ON(info->refcnt > 0);
502
503         kfree(info);
504
505         /* Legacy IRQ descriptors are managed by the arch. */
506         if (irq < NR_IRQS_LEGACY)
507                 return;
508
509         irq_free_desc(irq);
510 }
511
512 static void pirq_query_unmask(int irq)
513 {
514         struct physdev_irq_status_query irq_status;
515         struct irq_info *info = info_for_irq(irq);
516
517         BUG_ON(info->type != IRQT_PIRQ);
518
519         irq_status.irq = pirq_from_irq(irq);
520         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
521                 irq_status.flags = 0;
522
523         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
524         if (irq_status.flags & XENIRQSTAT_needs_eoi)
525                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
526 }
527
528 static bool probing_irq(int irq)
529 {
530         struct irq_desc *desc = irq_to_desc(irq);
531
532         return desc && desc->action == NULL;
533 }
534
535 static void eoi_pirq(struct irq_data *data)
536 {
537         int evtchn = evtchn_from_irq(data->irq);
538         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
539         int rc = 0;
540
541         irq_move_irq(data);
542
543         if (VALID_EVTCHN(evtchn))
544                 clear_evtchn(evtchn);
545
546         if (pirq_needs_eoi(data->irq)) {
547                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
548                 WARN_ON(rc);
549         }
550 }
551
552 static void mask_ack_pirq(struct irq_data *data)
553 {
554         disable_dynirq(data);
555         eoi_pirq(data);
556 }
557
558 static unsigned int __startup_pirq(unsigned int irq)
559 {
560         struct evtchn_bind_pirq bind_pirq;
561         struct irq_info *info = info_for_irq(irq);
562         int evtchn = evtchn_from_irq(irq);
563         int rc;
564
565         BUG_ON(info->type != IRQT_PIRQ);
566
567         if (VALID_EVTCHN(evtchn))
568                 goto out;
569
570         bind_pirq.pirq = pirq_from_irq(irq);
571         /* NB. We are happy to share unless we are probing. */
572         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
573                                         BIND_PIRQ__WILL_SHARE : 0;
574         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
575         if (rc != 0) {
576                 if (!probing_irq(irq))
577                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
578                                irq);
579                 return 0;
580         }
581         evtchn = bind_pirq.port;
582
583         pirq_query_unmask(irq);
584
585         evtchn_to_irq[evtchn] = irq;
586         bind_evtchn_to_cpu(evtchn, 0);
587         info->evtchn = evtchn;
588
589 out:
590         unmask_evtchn(evtchn);
591         eoi_pirq(irq_get_irq_data(irq));
592
593         return 0;
594 }
595
596 static unsigned int startup_pirq(struct irq_data *data)
597 {
598         return __startup_pirq(data->irq);
599 }
600
601 static void shutdown_pirq(struct irq_data *data)
602 {
603         struct evtchn_close close;
604         unsigned int irq = data->irq;
605         struct irq_info *info = info_for_irq(irq);
606         int evtchn = evtchn_from_irq(irq);
607
608         BUG_ON(info->type != IRQT_PIRQ);
609
610         if (!VALID_EVTCHN(evtchn))
611                 return;
612
613         mask_evtchn(evtchn);
614
615         close.port = evtchn;
616         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
617                 BUG();
618
619         bind_evtchn_to_cpu(evtchn, 0);
620         evtchn_to_irq[evtchn] = -1;
621         info->evtchn = 0;
622 }
623
624 static void enable_pirq(struct irq_data *data)
625 {
626         startup_pirq(data);
627 }
628
629 static void disable_pirq(struct irq_data *data)
630 {
631         disable_dynirq(data);
632 }
633
634 int xen_irq_from_gsi(unsigned gsi)
635 {
636         struct irq_info *info;
637
638         list_for_each_entry(info, &xen_irq_list_head, list) {
639                 if (info->type != IRQT_PIRQ)
640                         continue;
641
642                 if (info->u.pirq.gsi == gsi)
643                         return info->irq;
644         }
645
646         return -1;
647 }
648 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
649
650 /*
651  * Do not make any assumptions regarding the relationship between the
652  * IRQ number returned here and the Xen pirq argument.
653  *
654  * Note: We don't assign an event channel until the irq actually started
655  * up.  Return an existing irq if we've already got one for the gsi.
656  *
657  * Shareable implies level triggered, not shareable implies edge
658  * triggered here.
659  */
660 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
661                              unsigned pirq, int shareable, char *name)
662 {
663         int irq = -1;
664         struct physdev_irq irq_op;
665
666         mutex_lock(&irq_mapping_update_lock);
667
668         irq = xen_irq_from_gsi(gsi);
669         if (irq != -1) {
670                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
671                        irq, gsi);
672                 goto out;
673         }
674
675         irq = xen_allocate_irq_gsi(gsi);
676         if (irq < 0)
677                 goto out;
678
679         irq_op.irq = irq;
680         irq_op.vector = 0;
681
682         /* Only the privileged domain can do this. For non-priv, the pcifront
683          * driver provides a PCI bus that does the call to do exactly
684          * this in the priv domain. */
685         if (xen_initial_domain() &&
686             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
687                 xen_free_irq(irq);
688                 irq = -ENOSPC;
689                 goto out;
690         }
691
692         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
693                                shareable ? PIRQ_SHAREABLE : 0);
694
695         pirq_query_unmask(irq);
696         /* We try to use the handler with the appropriate semantic for the
697          * type of interrupt: if the interrupt is an edge triggered
698          * interrupt we use handle_edge_irq.
699          *
700          * On the other hand if the interrupt is level triggered we use
701          * handle_fasteoi_irq like the native code does for this kind of
702          * interrupts.
703          *
704          * Depending on the Xen version, pirq_needs_eoi might return true
705          * not only for level triggered interrupts but for edge triggered
706          * interrupts too. In any case Xen always honors the eoi mechanism,
707          * not injecting any more pirqs of the same kind if the first one
708          * hasn't received an eoi yet. Therefore using the fasteoi handler
709          * is the right choice either way.
710          */
711         if (shareable)
712                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
713                                 handle_fasteoi_irq, name);
714         else
715                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
716                                 handle_edge_irq, name);
717
718 out:
719         mutex_unlock(&irq_mapping_update_lock);
720
721         return irq;
722 }
723
724 #ifdef CONFIG_PCI_MSI
725 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
726 {
727         int rc;
728         struct physdev_get_free_pirq op_get_free_pirq;
729
730         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
731         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
732
733         WARN_ONCE(rc == -ENOSYS,
734                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
735
736         return rc ? -1 : op_get_free_pirq.pirq;
737 }
738
739 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
740                              int pirq, int vector, const char *name,
741                              domid_t domid)
742 {
743         int irq, ret;
744
745         mutex_lock(&irq_mapping_update_lock);
746
747         irq = xen_allocate_irq_dynamic();
748         if (irq < 0)
749                 goto out;
750
751         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
752                         name);
753
754         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
755         ret = irq_set_msi_desc(irq, msidesc);
756         if (ret < 0)
757                 goto error_irq;
758 out:
759         mutex_unlock(&irq_mapping_update_lock);
760         return irq;
761 error_irq:
762         mutex_unlock(&irq_mapping_update_lock);
763         xen_free_irq(irq);
764         return ret;
765 }
766 #endif
767
768 int xen_destroy_irq(int irq)
769 {
770         struct irq_desc *desc;
771         struct physdev_unmap_pirq unmap_irq;
772         struct irq_info *info = info_for_irq(irq);
773         int rc = -ENOENT;
774
775         mutex_lock(&irq_mapping_update_lock);
776
777         desc = irq_to_desc(irq);
778         if (!desc)
779                 goto out;
780
781         if (xen_initial_domain()) {
782                 unmap_irq.pirq = info->u.pirq.pirq;
783                 unmap_irq.domid = info->u.pirq.domid;
784                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
785                 /* If another domain quits without making the pci_disable_msix
786                  * call, the Xen hypervisor takes care of freeing the PIRQs
787                  * (free_domain_pirqs).
788                  */
789                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
790                         printk(KERN_INFO "domain %d does not have %d anymore\n",
791                                 info->u.pirq.domid, info->u.pirq.pirq);
792                 else if (rc) {
793                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
794                         goto out;
795                 }
796         }
797
798         xen_free_irq(irq);
799
800 out:
801         mutex_unlock(&irq_mapping_update_lock);
802         return rc;
803 }
804
805 int xen_irq_from_pirq(unsigned pirq)
806 {
807         int irq;
808
809         struct irq_info *info;
810
811         mutex_lock(&irq_mapping_update_lock);
812
813         list_for_each_entry(info, &xen_irq_list_head, list) {
814                 if (info->type != IRQT_PIRQ)
815                         continue;
816                 irq = info->irq;
817                 if (info->u.pirq.pirq == pirq)
818                         goto out;
819         }
820         irq = -1;
821 out:
822         mutex_unlock(&irq_mapping_update_lock);
823
824         return irq;
825 }
826
827
828 int xen_pirq_from_irq(unsigned irq)
829 {
830         return pirq_from_irq(irq);
831 }
832 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
833 int bind_evtchn_to_irq(unsigned int evtchn)
834 {
835         int irq;
836
837         mutex_lock(&irq_mapping_update_lock);
838
839         irq = evtchn_to_irq[evtchn];
840
841         if (irq == -1) {
842                 irq = xen_allocate_irq_dynamic();
843                 if (irq < 0)
844                         goto out;
845
846                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
847                                               handle_edge_irq, "event");
848
849                 xen_irq_info_evtchn_init(irq, evtchn);
850         } else {
851                 struct irq_info *info = info_for_irq(irq);
852                 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
853         }
854         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
855
856 out:
857         mutex_unlock(&irq_mapping_update_lock);
858
859         return irq;
860 }
861 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
862
863 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
864 {
865         struct evtchn_bind_ipi bind_ipi;
866         int evtchn, irq;
867
868         mutex_lock(&irq_mapping_update_lock);
869
870         irq = per_cpu(ipi_to_irq, cpu)[ipi];
871
872         if (irq == -1) {
873                 irq = xen_allocate_irq_dynamic();
874                 if (irq < 0)
875                         goto out;
876
877                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
878                                               handle_percpu_irq, "ipi");
879
880                 bind_ipi.vcpu = cpu;
881                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
882                                                 &bind_ipi) != 0)
883                         BUG();
884                 evtchn = bind_ipi.port;
885
886                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
887
888                 bind_evtchn_to_cpu(evtchn, cpu);
889         } else {
890                 struct irq_info *info = info_for_irq(irq);
891                 WARN_ON(info == NULL || info->type != IRQT_IPI);
892         }
893
894  out:
895         mutex_unlock(&irq_mapping_update_lock);
896         return irq;
897 }
898
899 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
900                                           unsigned int remote_port)
901 {
902         struct evtchn_bind_interdomain bind_interdomain;
903         int err;
904
905         bind_interdomain.remote_dom  = remote_domain;
906         bind_interdomain.remote_port = remote_port;
907
908         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
909                                           &bind_interdomain);
910
911         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
912 }
913
914 static int find_virq(unsigned int virq, unsigned int cpu)
915 {
916         struct evtchn_status status;
917         int port, rc = -ENOENT;
918
919         memset(&status, 0, sizeof(status));
920         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
921                 status.dom = DOMID_SELF;
922                 status.port = port;
923                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
924                 if (rc < 0)
925                         continue;
926                 if (status.status != EVTCHNSTAT_virq)
927                         continue;
928                 if (status.u.virq == virq && status.vcpu == cpu) {
929                         rc = port;
930                         break;
931                 }
932         }
933         return rc;
934 }
935
936 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
937 {
938         struct evtchn_bind_virq bind_virq;
939         int evtchn, irq, ret;
940
941         mutex_lock(&irq_mapping_update_lock);
942
943         irq = per_cpu(virq_to_irq, cpu)[virq];
944
945         if (irq == -1) {
946                 irq = xen_allocate_irq_dynamic();
947                 if (irq < 0)
948                         goto out;
949
950                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
951                                               handle_percpu_irq, "virq");
952
953                 bind_virq.virq = virq;
954                 bind_virq.vcpu = cpu;
955                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
956                                                 &bind_virq);
957                 if (ret == 0)
958                         evtchn = bind_virq.port;
959                 else {
960                         if (ret == -EEXIST)
961                                 ret = find_virq(virq, cpu);
962                         BUG_ON(ret < 0);
963                         evtchn = ret;
964                 }
965
966                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
967
968                 bind_evtchn_to_cpu(evtchn, cpu);
969         } else {
970                 struct irq_info *info = info_for_irq(irq);
971                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
972         }
973
974 out:
975         mutex_unlock(&irq_mapping_update_lock);
976
977         return irq;
978 }
979
980 static void unbind_from_irq(unsigned int irq)
981 {
982         struct evtchn_close close;
983         int evtchn = evtchn_from_irq(irq);
984         struct irq_info *info = irq_get_handler_data(irq);
985
986         mutex_lock(&irq_mapping_update_lock);
987
988         if (info->refcnt > 0) {
989                 info->refcnt--;
990                 if (info->refcnt != 0)
991                         goto done;
992         }
993
994         if (VALID_EVTCHN(evtchn)) {
995                 close.port = evtchn;
996                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
997                         BUG();
998
999                 switch (type_from_irq(irq)) {
1000                 case IRQT_VIRQ:
1001                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
1002                                 [virq_from_irq(irq)] = -1;
1003                         break;
1004                 case IRQT_IPI:
1005                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
1006                                 [ipi_from_irq(irq)] = -1;
1007                         break;
1008                 default:
1009                         break;
1010                 }
1011
1012                 /* Closed ports are implicitly re-bound to VCPU0. */
1013                 bind_evtchn_to_cpu(evtchn, 0);
1014
1015                 evtchn_to_irq[evtchn] = -1;
1016         }
1017
1018         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1019
1020         xen_free_irq(irq);
1021
1022  done:
1023         mutex_unlock(&irq_mapping_update_lock);
1024 }
1025
1026 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1027                               irq_handler_t handler,
1028                               unsigned long irqflags,
1029                               const char *devname, void *dev_id)
1030 {
1031         int irq, retval;
1032
1033         irq = bind_evtchn_to_irq(evtchn);
1034         if (irq < 0)
1035                 return irq;
1036         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1037         if (retval != 0) {
1038                 unbind_from_irq(irq);
1039                 return retval;
1040         }
1041
1042         return irq;
1043 }
1044 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1045
1046 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1047                                           unsigned int remote_port,
1048                                           irq_handler_t handler,
1049                                           unsigned long irqflags,
1050                                           const char *devname,
1051                                           void *dev_id)
1052 {
1053         int irq, retval;
1054
1055         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1056         if (irq < 0)
1057                 return irq;
1058
1059         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1060         if (retval != 0) {
1061                 unbind_from_irq(irq);
1062                 return retval;
1063         }
1064
1065         return irq;
1066 }
1067 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1068
1069 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1070                             irq_handler_t handler,
1071                             unsigned long irqflags, const char *devname, void *dev_id)
1072 {
1073         int irq, retval;
1074
1075         irq = bind_virq_to_irq(virq, cpu);
1076         if (irq < 0)
1077                 return irq;
1078         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1079         if (retval != 0) {
1080                 unbind_from_irq(irq);
1081                 return retval;
1082         }
1083
1084         return irq;
1085 }
1086 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1087
1088 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1089                            unsigned int cpu,
1090                            irq_handler_t handler,
1091                            unsigned long irqflags,
1092                            const char *devname,
1093                            void *dev_id)
1094 {
1095         int irq, retval;
1096
1097         irq = bind_ipi_to_irq(ipi, cpu);
1098         if (irq < 0)
1099                 return irq;
1100
1101         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1102         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1103         if (retval != 0) {
1104                 unbind_from_irq(irq);
1105                 return retval;
1106         }
1107
1108         return irq;
1109 }
1110
1111 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1112 {
1113         free_irq(irq, dev_id);
1114         unbind_from_irq(irq);
1115 }
1116 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1117
1118 int evtchn_make_refcounted(unsigned int evtchn)
1119 {
1120         int irq = evtchn_to_irq[evtchn];
1121         struct irq_info *info;
1122
1123         if (irq == -1)
1124                 return -ENOENT;
1125
1126         info = irq_get_handler_data(irq);
1127
1128         if (!info)
1129                 return -ENOENT;
1130
1131         WARN_ON(info->refcnt != -1);
1132
1133         info->refcnt = 1;
1134
1135         return 0;
1136 }
1137 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1138
1139 int evtchn_get(unsigned int evtchn)
1140 {
1141         int irq;
1142         struct irq_info *info;
1143         int err = -ENOENT;
1144
1145         if (evtchn >= NR_EVENT_CHANNELS)
1146                 return -EINVAL;
1147
1148         mutex_lock(&irq_mapping_update_lock);
1149
1150         irq = evtchn_to_irq[evtchn];
1151         if (irq == -1)
1152                 goto done;
1153
1154         info = irq_get_handler_data(irq);
1155
1156         if (!info)
1157                 goto done;
1158
1159         err = -EINVAL;
1160         if (info->refcnt <= 0)
1161                 goto done;
1162
1163         info->refcnt++;
1164         err = 0;
1165  done:
1166         mutex_unlock(&irq_mapping_update_lock);
1167
1168         return err;
1169 }
1170 EXPORT_SYMBOL_GPL(evtchn_get);
1171
1172 void evtchn_put(unsigned int evtchn)
1173 {
1174         int irq = evtchn_to_irq[evtchn];
1175         if (WARN_ON(irq == -1))
1176                 return;
1177         unbind_from_irq(irq);
1178 }
1179 EXPORT_SYMBOL_GPL(evtchn_put);
1180
1181 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1182 {
1183         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1184         BUG_ON(irq < 0);
1185         notify_remote_via_irq(irq);
1186 }
1187
1188 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1189 {
1190         struct shared_info *sh = HYPERVISOR_shared_info;
1191         int cpu = smp_processor_id();
1192         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1193         int i;
1194         unsigned long flags;
1195         static DEFINE_SPINLOCK(debug_lock);
1196         struct vcpu_info *v;
1197
1198         spin_lock_irqsave(&debug_lock, flags);
1199
1200         printk("\nvcpu %d\n  ", cpu);
1201
1202         for_each_online_cpu(i) {
1203                 int pending;
1204                 v = per_cpu(xen_vcpu, i);
1205                 pending = (get_irq_regs() && i == cpu)
1206                         ? xen_irqs_disabled(get_irq_regs())
1207                         : v->evtchn_upcall_mask;
1208                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1209                        pending, v->evtchn_upcall_pending,
1210                        (int)(sizeof(v->evtchn_pending_sel)*2),
1211                        v->evtchn_pending_sel);
1212         }
1213         v = per_cpu(xen_vcpu, cpu);
1214
1215         printk("\npending:\n   ");
1216         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1217                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1218                        sh->evtchn_pending[i],
1219                        i % 8 == 0 ? "\n   " : " ");
1220         printk("\nglobal mask:\n   ");
1221         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1222                 printk("%0*lx%s",
1223                        (int)(sizeof(sh->evtchn_mask[0])*2),
1224                        sh->evtchn_mask[i],
1225                        i % 8 == 0 ? "\n   " : " ");
1226
1227         printk("\nglobally unmasked:\n   ");
1228         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1229                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1230                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1231                        i % 8 == 0 ? "\n   " : " ");
1232
1233         printk("\nlocal cpu%d mask:\n   ", cpu);
1234         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1235                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1236                        cpu_evtchn[i],
1237                        i % 8 == 0 ? "\n   " : " ");
1238
1239         printk("\nlocally unmasked:\n   ");
1240         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1241                 unsigned long pending = sh->evtchn_pending[i]
1242                         & ~sh->evtchn_mask[i]
1243                         & cpu_evtchn[i];
1244                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1245                        pending, i % 8 == 0 ? "\n   " : " ");
1246         }
1247
1248         printk("\npending list:\n");
1249         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1250                 if (sync_test_bit(i, sh->evtchn_pending)) {
1251                         int word_idx = i / BITS_PER_LONG;
1252                         printk("  %d: event %d -> irq %d%s%s%s\n",
1253                                cpu_from_evtchn(i), i,
1254                                evtchn_to_irq[i],
1255                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1256                                              ? "" : " l2-clear",
1257                                !sync_test_bit(i, sh->evtchn_mask)
1258                                              ? "" : " globally-masked",
1259                                sync_test_bit(i, cpu_evtchn)
1260                                              ? "" : " locally-masked");
1261                 }
1262         }
1263
1264         spin_unlock_irqrestore(&debug_lock, flags);
1265
1266         return IRQ_HANDLED;
1267 }
1268
1269 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1270 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1271 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1272
1273 /*
1274  * Mask out the i least significant bits of w
1275  */
1276 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1277
1278 /*
1279  * Search the CPUs pending events bitmasks.  For each one found, map
1280  * the event number to an irq, and feed it into do_IRQ() for
1281  * handling.
1282  *
1283  * Xen uses a two-level bitmap to speed searching.  The first level is
1284  * a bitset of words which contain pending event bits.  The second
1285  * level is a bitset of pending events themselves.
1286  */
1287 static void __xen_evtchn_do_upcall(void)
1288 {
1289         int start_word_idx, start_bit_idx;
1290         int word_idx, bit_idx;
1291         int i;
1292         int cpu = get_cpu();
1293         struct shared_info *s = HYPERVISOR_shared_info;
1294         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1295         unsigned count;
1296
1297         do {
1298                 unsigned long pending_words;
1299
1300                 vcpu_info->evtchn_upcall_pending = 0;
1301
1302                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1303                         goto out;
1304
1305 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1306                 /* Clear master flag /before/ clearing selector flag. */
1307                 wmb();
1308 #endif
1309                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1310
1311                 start_word_idx = __this_cpu_read(current_word_idx);
1312                 start_bit_idx = __this_cpu_read(current_bit_idx);
1313
1314                 word_idx = start_word_idx;
1315
1316                 for (i = 0; pending_words != 0; i++) {
1317                         unsigned long pending_bits;
1318                         unsigned long words;
1319
1320                         words = MASK_LSBS(pending_words, word_idx);
1321
1322                         /*
1323                          * If we masked out all events, wrap to beginning.
1324                          */
1325                         if (words == 0) {
1326                                 word_idx = 0;
1327                                 bit_idx = 0;
1328                                 continue;
1329                         }
1330                         word_idx = __ffs(words);
1331
1332                         pending_bits = active_evtchns(cpu, s, word_idx);
1333                         bit_idx = 0; /* usually scan entire word from start */
1334                         if (word_idx == start_word_idx) {
1335                                 /* We scan the starting word in two parts */
1336                                 if (i == 0)
1337                                         /* 1st time: start in the middle */
1338                                         bit_idx = start_bit_idx;
1339                                 else
1340                                         /* 2nd time: mask bits done already */
1341                                         bit_idx &= (1UL << start_bit_idx) - 1;
1342                         }
1343
1344                         do {
1345                                 unsigned long bits;
1346                                 int port, irq;
1347                                 struct irq_desc *desc;
1348
1349                                 bits = MASK_LSBS(pending_bits, bit_idx);
1350
1351                                 /* If we masked out all events, move on. */
1352                                 if (bits == 0)
1353                                         break;
1354
1355                                 bit_idx = __ffs(bits);
1356
1357                                 /* Process port. */
1358                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1359                                 irq = evtchn_to_irq[port];
1360
1361                                 if (irq != -1) {
1362                                         desc = irq_to_desc(irq);
1363                                         if (desc)
1364                                                 generic_handle_irq_desc(irq, desc);
1365                                 }
1366
1367                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1368
1369                                 /* Next caller starts at last processed + 1 */
1370                                 __this_cpu_write(current_word_idx,
1371                                                  bit_idx ? word_idx :
1372                                                  (word_idx+1) % BITS_PER_LONG);
1373                                 __this_cpu_write(current_bit_idx, bit_idx);
1374                         } while (bit_idx != 0);
1375
1376                         /* Scan start_l1i twice; all others once. */
1377                         if ((word_idx != start_word_idx) || (i != 0))
1378                                 pending_words &= ~(1UL << word_idx);
1379
1380                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1381                 }
1382
1383                 BUG_ON(!irqs_disabled());
1384
1385                 count = __this_cpu_read(xed_nesting_count);
1386                 __this_cpu_write(xed_nesting_count, 0);
1387         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1388
1389 out:
1390
1391         put_cpu();
1392 }
1393
1394 void xen_evtchn_do_upcall(struct pt_regs *regs)
1395 {
1396         struct pt_regs *old_regs = set_irq_regs(regs);
1397
1398         irq_enter();
1399 #ifdef CONFIG_X86
1400         exit_idle();
1401 #endif
1402
1403         __xen_evtchn_do_upcall();
1404
1405         irq_exit();
1406         set_irq_regs(old_regs);
1407 }
1408
1409 void xen_hvm_evtchn_do_upcall(void)
1410 {
1411         __xen_evtchn_do_upcall();
1412 }
1413 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1414
1415 /* Rebind a new event channel to an existing irq. */
1416 void rebind_evtchn_irq(int evtchn, int irq)
1417 {
1418         struct irq_info *info = info_for_irq(irq);
1419
1420         /* Make sure the irq is masked, since the new event channel
1421            will also be masked. */
1422         disable_irq(irq);
1423
1424         mutex_lock(&irq_mapping_update_lock);
1425
1426         /* After resume the irq<->evtchn mappings are all cleared out */
1427         BUG_ON(evtchn_to_irq[evtchn] != -1);
1428         /* Expect irq to have been bound before,
1429            so there should be a proper type */
1430         BUG_ON(info->type == IRQT_UNBOUND);
1431
1432         xen_irq_info_evtchn_init(irq, evtchn);
1433
1434         mutex_unlock(&irq_mapping_update_lock);
1435
1436         /* new event channels are always bound to cpu 0 */
1437         irq_set_affinity(irq, cpumask_of(0));
1438
1439         /* Unmask the event channel. */
1440         enable_irq(irq);
1441 }
1442
1443 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1444 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1445 {
1446         struct evtchn_bind_vcpu bind_vcpu;
1447         int evtchn = evtchn_from_irq(irq);
1448
1449         if (!VALID_EVTCHN(evtchn))
1450                 return -1;
1451
1452         /*
1453          * Events delivered via platform PCI interrupts are always
1454          * routed to vcpu 0 and hence cannot be rebound.
1455          */
1456         if (xen_hvm_domain() && !xen_have_vector_callback)
1457                 return -1;
1458
1459         /* Send future instances of this interrupt to other vcpu. */
1460         bind_vcpu.port = evtchn;
1461         bind_vcpu.vcpu = tcpu;
1462
1463         /*
1464          * If this fails, it usually just indicates that we're dealing with a
1465          * virq or IPI channel, which don't actually need to be rebound. Ignore
1466          * it, but don't do the xenlinux-level rebind in that case.
1467          */
1468         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1469                 bind_evtchn_to_cpu(evtchn, tcpu);
1470
1471         return 0;
1472 }
1473
1474 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1475                             bool force)
1476 {
1477         unsigned tcpu = cpumask_first(dest);
1478
1479         return rebind_irq_to_cpu(data->irq, tcpu);
1480 }
1481
1482 int resend_irq_on_evtchn(unsigned int irq)
1483 {
1484         int masked, evtchn = evtchn_from_irq(irq);
1485         struct shared_info *s = HYPERVISOR_shared_info;
1486
1487         if (!VALID_EVTCHN(evtchn))
1488                 return 1;
1489
1490         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1491         sync_set_bit(evtchn, s->evtchn_pending);
1492         if (!masked)
1493                 unmask_evtchn(evtchn);
1494
1495         return 1;
1496 }
1497
1498 static void enable_dynirq(struct irq_data *data)
1499 {
1500         int evtchn = evtchn_from_irq(data->irq);
1501
1502         if (VALID_EVTCHN(evtchn))
1503                 unmask_evtchn(evtchn);
1504 }
1505
1506 static void disable_dynirq(struct irq_data *data)
1507 {
1508         int evtchn = evtchn_from_irq(data->irq);
1509
1510         if (VALID_EVTCHN(evtchn))
1511                 mask_evtchn(evtchn);
1512 }
1513
1514 static void ack_dynirq(struct irq_data *data)
1515 {
1516         int evtchn = evtchn_from_irq(data->irq);
1517
1518         irq_move_irq(data);
1519
1520         if (VALID_EVTCHN(evtchn))
1521                 clear_evtchn(evtchn);
1522 }
1523
1524 static void mask_ack_dynirq(struct irq_data *data)
1525 {
1526         disable_dynirq(data);
1527         ack_dynirq(data);
1528 }
1529
1530 static int retrigger_dynirq(struct irq_data *data)
1531 {
1532         int evtchn = evtchn_from_irq(data->irq);
1533         struct shared_info *sh = HYPERVISOR_shared_info;
1534         int ret = 0;
1535
1536         if (VALID_EVTCHN(evtchn)) {
1537                 int masked;
1538
1539                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1540                 sync_set_bit(evtchn, sh->evtchn_pending);
1541                 if (!masked)
1542                         unmask_evtchn(evtchn);
1543                 ret = 1;
1544         }
1545
1546         return ret;
1547 }
1548
1549 static void restore_pirqs(void)
1550 {
1551         int pirq, rc, irq, gsi;
1552         struct physdev_map_pirq map_irq;
1553         struct irq_info *info;
1554
1555         list_for_each_entry(info, &xen_irq_list_head, list) {
1556                 if (info->type != IRQT_PIRQ)
1557                         continue;
1558
1559                 pirq = info->u.pirq.pirq;
1560                 gsi = info->u.pirq.gsi;
1561                 irq = info->irq;
1562
1563                 /* save/restore of PT devices doesn't work, so at this point the
1564                  * only devices present are GSI based emulated devices */
1565                 if (!gsi)
1566                         continue;
1567
1568                 map_irq.domid = DOMID_SELF;
1569                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1570                 map_irq.index = gsi;
1571                 map_irq.pirq = pirq;
1572
1573                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1574                 if (rc) {
1575                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1576                                         gsi, irq, pirq, rc);
1577                         xen_free_irq(irq);
1578                         continue;
1579                 }
1580
1581                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1582
1583                 __startup_pirq(irq);
1584         }
1585 }
1586
1587 static void restore_cpu_virqs(unsigned int cpu)
1588 {
1589         struct evtchn_bind_virq bind_virq;
1590         int virq, irq, evtchn;
1591
1592         for (virq = 0; virq < NR_VIRQS; virq++) {
1593                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1594                         continue;
1595
1596                 BUG_ON(virq_from_irq(irq) != virq);
1597
1598                 /* Get a new binding from Xen. */
1599                 bind_virq.virq = virq;
1600                 bind_virq.vcpu = cpu;
1601                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1602                                                 &bind_virq) != 0)
1603                         BUG();
1604                 evtchn = bind_virq.port;
1605
1606                 /* Record the new mapping. */
1607                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1608                 bind_evtchn_to_cpu(evtchn, cpu);
1609         }
1610 }
1611
1612 static void restore_cpu_ipis(unsigned int cpu)
1613 {
1614         struct evtchn_bind_ipi bind_ipi;
1615         int ipi, irq, evtchn;
1616
1617         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1618                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1619                         continue;
1620
1621                 BUG_ON(ipi_from_irq(irq) != ipi);
1622
1623                 /* Get a new binding from Xen. */
1624                 bind_ipi.vcpu = cpu;
1625                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1626                                                 &bind_ipi) != 0)
1627                         BUG();
1628                 evtchn = bind_ipi.port;
1629
1630                 /* Record the new mapping. */
1631                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1632                 bind_evtchn_to_cpu(evtchn, cpu);
1633         }
1634 }
1635
1636 /* Clear an irq's pending state, in preparation for polling on it */
1637 void xen_clear_irq_pending(int irq)
1638 {
1639         int evtchn = evtchn_from_irq(irq);
1640
1641         if (VALID_EVTCHN(evtchn))
1642                 clear_evtchn(evtchn);
1643 }
1644 EXPORT_SYMBOL(xen_clear_irq_pending);
1645 void xen_set_irq_pending(int irq)
1646 {
1647         int evtchn = evtchn_from_irq(irq);
1648
1649         if (VALID_EVTCHN(evtchn))
1650                 set_evtchn(evtchn);
1651 }
1652
1653 bool xen_test_irq_pending(int irq)
1654 {
1655         int evtchn = evtchn_from_irq(irq);
1656         bool ret = false;
1657
1658         if (VALID_EVTCHN(evtchn))
1659                 ret = test_evtchn(evtchn);
1660
1661         return ret;
1662 }
1663
1664 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1665  * the irq will be disabled so it won't deliver an interrupt. */
1666 void xen_poll_irq_timeout(int irq, u64 timeout)
1667 {
1668         evtchn_port_t evtchn = evtchn_from_irq(irq);
1669
1670         if (VALID_EVTCHN(evtchn)) {
1671                 struct sched_poll poll;
1672
1673                 poll.nr_ports = 1;
1674                 poll.timeout = timeout;
1675                 set_xen_guest_handle(poll.ports, &evtchn);
1676
1677                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1678                         BUG();
1679         }
1680 }
1681 EXPORT_SYMBOL(xen_poll_irq_timeout);
1682 /* Poll waiting for an irq to become pending.  In the usual case, the
1683  * irq will be disabled so it won't deliver an interrupt. */
1684 void xen_poll_irq(int irq)
1685 {
1686         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1687 }
1688
1689 /* Check whether the IRQ line is shared with other guests. */
1690 int xen_test_irq_shared(int irq)
1691 {
1692         struct irq_info *info = info_for_irq(irq);
1693         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1694
1695         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1696                 return 0;
1697         return !(irq_status.flags & XENIRQSTAT_shared);
1698 }
1699 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1700
1701 void xen_irq_resume(void)
1702 {
1703         unsigned int cpu, evtchn;
1704         struct irq_info *info;
1705
1706         init_evtchn_cpu_bindings();
1707
1708         /* New event-channel space is not 'live' yet. */
1709         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1710                 mask_evtchn(evtchn);
1711
1712         /* No IRQ <-> event-channel mappings. */
1713         list_for_each_entry(info, &xen_irq_list_head, list)
1714                 info->evtchn = 0; /* zap event-channel binding */
1715
1716         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1717                 evtchn_to_irq[evtchn] = -1;
1718
1719         for_each_possible_cpu(cpu) {
1720                 restore_cpu_virqs(cpu);
1721                 restore_cpu_ipis(cpu);
1722         }
1723
1724         restore_pirqs();
1725 }
1726
1727 static struct irq_chip xen_dynamic_chip __read_mostly = {
1728         .name                   = "xen-dyn",
1729
1730         .irq_disable            = disable_dynirq,
1731         .irq_mask               = disable_dynirq,
1732         .irq_unmask             = enable_dynirq,
1733
1734         .irq_ack                = ack_dynirq,
1735         .irq_mask_ack           = mask_ack_dynirq,
1736
1737         .irq_set_affinity       = set_affinity_irq,
1738         .irq_retrigger          = retrigger_dynirq,
1739 };
1740
1741 static struct irq_chip xen_pirq_chip __read_mostly = {
1742         .name                   = "xen-pirq",
1743
1744         .irq_startup            = startup_pirq,
1745         .irq_shutdown           = shutdown_pirq,
1746         .irq_enable             = enable_pirq,
1747         .irq_disable            = disable_pirq,
1748
1749         .irq_mask               = disable_dynirq,
1750         .irq_unmask             = enable_dynirq,
1751
1752         .irq_ack                = eoi_pirq,
1753         .irq_eoi                = eoi_pirq,
1754         .irq_mask_ack           = mask_ack_pirq,
1755
1756         .irq_set_affinity       = set_affinity_irq,
1757
1758         .irq_retrigger          = retrigger_dynirq,
1759 };
1760
1761 static struct irq_chip xen_percpu_chip __read_mostly = {
1762         .name                   = "xen-percpu",
1763
1764         .irq_disable            = disable_dynirq,
1765         .irq_mask               = disable_dynirq,
1766         .irq_unmask             = enable_dynirq,
1767
1768         .irq_ack                = ack_dynirq,
1769 };
1770
1771 int xen_set_callback_via(uint64_t via)
1772 {
1773         struct xen_hvm_param a;
1774         a.domid = DOMID_SELF;
1775         a.index = HVM_PARAM_CALLBACK_IRQ;
1776         a.value = via;
1777         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1778 }
1779 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1780
1781 #ifdef CONFIG_XEN_PVHVM
1782 /* Vector callbacks are better than PCI interrupts to receive event
1783  * channel notifications because we can receive vector callbacks on any
1784  * vcpu and we don't need PCI support or APIC interactions. */
1785 void xen_callback_vector(void)
1786 {
1787         int rc;
1788         uint64_t callback_via;
1789         if (xen_have_vector_callback) {
1790                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1791                 rc = xen_set_callback_via(callback_via);
1792                 if (rc) {
1793                         printk(KERN_ERR "Request for Xen HVM callback vector"
1794                                         " failed.\n");
1795                         xen_have_vector_callback = 0;
1796                         return;
1797                 }
1798                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1799                                 "enabled\n");
1800                 /* in the restore case the vector has already been allocated */
1801                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1802                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1803         }
1804 }
1805 #else
1806 void xen_callback_vector(void) {}
1807 #endif
1808
1809 void __init xen_init_IRQ(void)
1810 {
1811         int i;
1812
1813         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1814                                     GFP_KERNEL);
1815         BUG_ON(!evtchn_to_irq);
1816         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1817                 evtchn_to_irq[i] = -1;
1818
1819         init_evtchn_cpu_bindings();
1820
1821         /* No event channels are 'live' right now. */
1822         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1823                 mask_evtchn(i);
1824
1825         pirq_needs_eoi = pirq_needs_eoi_flag;
1826
1827 #ifdef CONFIG_X86
1828         if (xen_hvm_domain()) {
1829                 xen_callback_vector();
1830                 native_init_IRQ();
1831                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1832                  * __acpi_register_gsi can point at the right function */
1833                 pci_xen_hvm_init();
1834         } else {
1835                 int rc;
1836                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1837
1838                 irq_ctx_init(smp_processor_id());
1839                 if (xen_initial_domain())
1840                         pci_xen_initial_domain();
1841
1842                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1843                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1844                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1845                 if (rc != 0) {
1846                         free_page((unsigned long) pirq_eoi_map);
1847                         pirq_eoi_map = NULL;
1848                 } else
1849                         pirq_needs_eoi = pirq_check_eoi_map;
1850         }
1851 #endif
1852 }