Merge tag 'hsi-for-4.1' of git://git.kernel.org/pub/scm/linux/kernel/git/sre/linux-hsi
[sfrench/cifs-2.6.git] / arch / x86 / xen / smp.c
1 /*
2  * Xen SMP support
3  *
4  * This file implements the Xen versions of smp_ops.  SMP under Xen is
5  * very straightforward.  Bringing a CPU up is simply a matter of
6  * loading its initial context and setting it running.
7  *
8  * IPIs are handled through the Xen event mechanism.
9  *
10  * Because virtual CPUs can be scheduled onto any real CPU, there's no
11  * useful topology information for the kernel to make use of.  As a
12  * result, all CPUs are treated as if they're single-core and
13  * single-threaded.
14  */
15 #include <linux/sched.h>
16 #include <linux/err.h>
17 #include <linux/slab.h>
18 #include <linux/smp.h>
19 #include <linux/irq_work.h>
20 #include <linux/tick.h>
21
22 #include <asm/paravirt.h>
23 #include <asm/desc.h>
24 #include <asm/pgtable.h>
25 #include <asm/cpu.h>
26
27 #include <xen/interface/xen.h>
28 #include <xen/interface/vcpu.h>
29
30 #include <asm/xen/interface.h>
31 #include <asm/xen/hypercall.h>
32
33 #include <xen/xen.h>
34 #include <xen/page.h>
35 #include <xen/events.h>
36
37 #include <xen/hvc-console.h>
38 #include "xen-ops.h"
39 #include "mmu.h"
40 #include "smp.h"
41
42 cpumask_var_t xen_cpu_initialized_map;
43
44 struct xen_common_irq {
45         int irq;
46         char *name;
47 };
48 static DEFINE_PER_CPU(struct xen_common_irq, xen_resched_irq) = { .irq = -1 };
49 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfunc_irq) = { .irq = -1 };
50 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfuncsingle_irq) = { .irq = -1 };
51 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
52 static DEFINE_PER_CPU(struct xen_common_irq, xen_debug_irq) = { .irq = -1 };
53
54 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
55 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
56 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
57
58 /*
59  * Reschedule call back.
60  */
61 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
62 {
63         inc_irq_stat(irq_resched_count);
64         scheduler_ipi();
65
66         return IRQ_HANDLED;
67 }
68
69 static void cpu_bringup(void)
70 {
71         int cpu;
72
73         cpu_init();
74         touch_softlockup_watchdog();
75         preempt_disable();
76
77         /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
78         if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
79                 xen_enable_sysenter();
80                 xen_enable_syscall();
81         }
82         cpu = smp_processor_id();
83         smp_store_cpu_info(cpu);
84         cpu_data(cpu).x86_max_cores = 1;
85         set_cpu_sibling_map(cpu);
86
87         xen_setup_cpu_clockevents();
88
89         notify_cpu_starting(cpu);
90
91         set_cpu_online(cpu, true);
92
93         this_cpu_write(cpu_state, CPU_ONLINE);
94
95         wmb();
96
97         /* We can take interrupts now: we're officially "up". */
98         local_irq_enable();
99
100         wmb();                  /* make sure everything is out */
101 }
102
103 /*
104  * Note: cpu parameter is only relevant for PVH. The reason for passing it
105  * is we can't do smp_processor_id until the percpu segments are loaded, for
106  * which we need the cpu number! So we pass it in rdi as first parameter.
107  */
108 asmlinkage __visible void cpu_bringup_and_idle(int cpu)
109 {
110 #ifdef CONFIG_XEN_PVH
111         if (xen_feature(XENFEAT_auto_translated_physmap) &&
112             xen_feature(XENFEAT_supervisor_mode_kernel))
113                 xen_pvh_secondary_vcpu_init(cpu);
114 #endif
115         cpu_bringup();
116         cpu_startup_entry(CPUHP_ONLINE);
117 }
118
119 static void xen_smp_intr_free(unsigned int cpu)
120 {
121         if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
122                 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
123                 per_cpu(xen_resched_irq, cpu).irq = -1;
124                 kfree(per_cpu(xen_resched_irq, cpu).name);
125                 per_cpu(xen_resched_irq, cpu).name = NULL;
126         }
127         if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
128                 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
129                 per_cpu(xen_callfunc_irq, cpu).irq = -1;
130                 kfree(per_cpu(xen_callfunc_irq, cpu).name);
131                 per_cpu(xen_callfunc_irq, cpu).name = NULL;
132         }
133         if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
134                 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
135                 per_cpu(xen_debug_irq, cpu).irq = -1;
136                 kfree(per_cpu(xen_debug_irq, cpu).name);
137                 per_cpu(xen_debug_irq, cpu).name = NULL;
138         }
139         if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
140                 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
141                                        NULL);
142                 per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
143                 kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
144                 per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
145         }
146         if (xen_hvm_domain())
147                 return;
148
149         if (per_cpu(xen_irq_work, cpu).irq >= 0) {
150                 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
151                 per_cpu(xen_irq_work, cpu).irq = -1;
152                 kfree(per_cpu(xen_irq_work, cpu).name);
153                 per_cpu(xen_irq_work, cpu).name = NULL;
154         }
155 };
156 static int xen_smp_intr_init(unsigned int cpu)
157 {
158         int rc;
159         char *resched_name, *callfunc_name, *debug_name;
160
161         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
162         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
163                                     cpu,
164                                     xen_reschedule_interrupt,
165                                     IRQF_PERCPU|IRQF_NOBALANCING,
166                                     resched_name,
167                                     NULL);
168         if (rc < 0)
169                 goto fail;
170         per_cpu(xen_resched_irq, cpu).irq = rc;
171         per_cpu(xen_resched_irq, cpu).name = resched_name;
172
173         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
174         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
175                                     cpu,
176                                     xen_call_function_interrupt,
177                                     IRQF_PERCPU|IRQF_NOBALANCING,
178                                     callfunc_name,
179                                     NULL);
180         if (rc < 0)
181                 goto fail;
182         per_cpu(xen_callfunc_irq, cpu).irq = rc;
183         per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
184
185         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
186         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
187                                      IRQF_PERCPU | IRQF_NOBALANCING,
188                                      debug_name, NULL);
189         if (rc < 0)
190                 goto fail;
191         per_cpu(xen_debug_irq, cpu).irq = rc;
192         per_cpu(xen_debug_irq, cpu).name = debug_name;
193
194         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
195         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
196                                     cpu,
197                                     xen_call_function_single_interrupt,
198                                     IRQF_PERCPU|IRQF_NOBALANCING,
199                                     callfunc_name,
200                                     NULL);
201         if (rc < 0)
202                 goto fail;
203         per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
204         per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
205
206         /*
207          * The IRQ worker on PVHVM goes through the native path and uses the
208          * IPI mechanism.
209          */
210         if (xen_hvm_domain())
211                 return 0;
212
213         callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
214         rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
215                                     cpu,
216                                     xen_irq_work_interrupt,
217                                     IRQF_PERCPU|IRQF_NOBALANCING,
218                                     callfunc_name,
219                                     NULL);
220         if (rc < 0)
221                 goto fail;
222         per_cpu(xen_irq_work, cpu).irq = rc;
223         per_cpu(xen_irq_work, cpu).name = callfunc_name;
224
225         return 0;
226
227  fail:
228         xen_smp_intr_free(cpu);
229         return rc;
230 }
231
232 static void __init xen_fill_possible_map(void)
233 {
234         int i, rc;
235
236         if (xen_initial_domain())
237                 return;
238
239         for (i = 0; i < nr_cpu_ids; i++) {
240                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
241                 if (rc >= 0) {
242                         num_processors++;
243                         set_cpu_possible(i, true);
244                 }
245         }
246 }
247
248 static void __init xen_filter_cpu_maps(void)
249 {
250         int i, rc;
251         unsigned int subtract = 0;
252
253         if (!xen_initial_domain())
254                 return;
255
256         num_processors = 0;
257         disabled_cpus = 0;
258         for (i = 0; i < nr_cpu_ids; i++) {
259                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
260                 if (rc >= 0) {
261                         num_processors++;
262                         set_cpu_possible(i, true);
263                 } else {
264                         set_cpu_possible(i, false);
265                         set_cpu_present(i, false);
266                         subtract++;
267                 }
268         }
269 #ifdef CONFIG_HOTPLUG_CPU
270         /* This is akin to using 'nr_cpus' on the Linux command line.
271          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
272          * have up to X, while nr_cpu_ids is greater than X. This
273          * normally is not a problem, except when CPU hotplugging
274          * is involved and then there might be more than X CPUs
275          * in the guest - which will not work as there is no
276          * hypercall to expand the max number of VCPUs an already
277          * running guest has. So cap it up to X. */
278         if (subtract)
279                 nr_cpu_ids = nr_cpu_ids - subtract;
280 #endif
281
282 }
283
284 static void __init xen_smp_prepare_boot_cpu(void)
285 {
286         BUG_ON(smp_processor_id() != 0);
287         native_smp_prepare_boot_cpu();
288
289         if (xen_pv_domain()) {
290                 if (!xen_feature(XENFEAT_writable_page_tables))
291                         /* We've switched to the "real" per-cpu gdt, so make
292                          * sure the old memory can be recycled. */
293                         make_lowmem_page_readwrite(xen_initial_gdt);
294
295 #ifdef CONFIG_X86_32
296                 /*
297                  * Xen starts us with XEN_FLAT_RING1_DS, but linux code
298                  * expects __USER_DS
299                  */
300                 loadsegment(ds, __USER_DS);
301                 loadsegment(es, __USER_DS);
302 #endif
303
304                 xen_filter_cpu_maps();
305                 xen_setup_vcpu_info_placement();
306         }
307         /*
308          * The alternative logic (which patches the unlock/lock) runs before
309          * the smp bootup up code is activated. Hence we need to set this up
310          * the core kernel is being patched. Otherwise we will have only
311          * modules patched but not core code.
312          */
313         xen_init_spinlocks();
314 }
315
316 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
317 {
318         unsigned cpu;
319         unsigned int i;
320
321         if (skip_ioapic_setup) {
322                 char *m = (max_cpus == 0) ?
323                         "The nosmp parameter is incompatible with Xen; " \
324                         "use Xen dom0_max_vcpus=1 parameter" :
325                         "The noapic parameter is incompatible with Xen";
326
327                 xen_raw_printk(m);
328                 panic(m);
329         }
330         xen_init_lock_cpu(0);
331
332         smp_store_boot_cpu_info();
333         cpu_data(0).x86_max_cores = 1;
334
335         for_each_possible_cpu(i) {
336                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
337                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
338                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
339         }
340         set_cpu_sibling_map(0);
341
342         if (xen_smp_intr_init(0))
343                 BUG();
344
345         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
346                 panic("could not allocate xen_cpu_initialized_map\n");
347
348         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
349
350         /* Restrict the possible_map according to max_cpus. */
351         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
352                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
353                         continue;
354                 set_cpu_possible(cpu, false);
355         }
356
357         for_each_possible_cpu(cpu)
358                 set_cpu_present(cpu, true);
359 }
360
361 static int
362 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
363 {
364         struct vcpu_guest_context *ctxt;
365         struct desc_struct *gdt;
366         unsigned long gdt_mfn;
367
368         /* used to tell cpu_init() that it can proceed with initialization */
369         cpumask_set_cpu(cpu, cpu_callout_mask);
370         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
371                 return 0;
372
373         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
374         if (ctxt == NULL)
375                 return -ENOMEM;
376
377         gdt = get_cpu_gdt_table(cpu);
378
379 #ifdef CONFIG_X86_32
380         /* Note: PVH is not yet supported on x86_32. */
381         ctxt->user_regs.fs = __KERNEL_PERCPU;
382         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
383 #endif
384         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
385
386         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
387                 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
388                 ctxt->flags = VGCF_IN_KERNEL;
389                 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
390                 ctxt->user_regs.ds = __USER_DS;
391                 ctxt->user_regs.es = __USER_DS;
392                 ctxt->user_regs.ss = __KERNEL_DS;
393
394                 xen_copy_trap_info(ctxt->trap_ctxt);
395
396                 ctxt->ldt_ents = 0;
397
398                 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
399
400                 gdt_mfn = arbitrary_virt_to_mfn(gdt);
401                 make_lowmem_page_readonly(gdt);
402                 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
403
404                 ctxt->gdt_frames[0] = gdt_mfn;
405                 ctxt->gdt_ents      = GDT_ENTRIES;
406
407                 ctxt->kernel_ss = __KERNEL_DS;
408                 ctxt->kernel_sp = idle->thread.sp0;
409
410 #ifdef CONFIG_X86_32
411                 ctxt->event_callback_cs     = __KERNEL_CS;
412                 ctxt->failsafe_callback_cs  = __KERNEL_CS;
413 #else
414                 ctxt->gs_base_kernel = per_cpu_offset(cpu);
415 #endif
416                 ctxt->event_callback_eip    =
417                                         (unsigned long)xen_hypervisor_callback;
418                 ctxt->failsafe_callback_eip =
419                                         (unsigned long)xen_failsafe_callback;
420                 ctxt->user_regs.cs = __KERNEL_CS;
421                 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
422         }
423 #ifdef CONFIG_XEN_PVH
424         else {
425                 /*
426                  * The vcpu comes on kernel page tables which have the NX pte
427                  * bit set. This means before DS/SS is touched, NX in
428                  * EFER must be set. Hence the following assembly glue code.
429                  */
430                 ctxt->user_regs.eip = (unsigned long)xen_pvh_early_cpu_init;
431                 ctxt->user_regs.rdi = cpu;
432                 ctxt->user_regs.rsi = true;  /* entry == true */
433         }
434 #endif
435         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
436         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
437         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
438                 BUG();
439
440         kfree(ctxt);
441         return 0;
442 }
443
444 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
445 {
446         int rc;
447
448         common_cpu_up(cpu, idle);
449
450         xen_setup_runstate_info(cpu);
451         xen_setup_timer(cpu);
452         xen_init_lock_cpu(cpu);
453
454         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
455
456         /* make sure interrupts start blocked */
457         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
458
459         rc = cpu_initialize_context(cpu, idle);
460         if (rc)
461                 return rc;
462
463         rc = xen_smp_intr_init(cpu);
464         if (rc)
465                 return rc;
466
467         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
468         BUG_ON(rc);
469
470         while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
471                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
472                 barrier();
473         }
474
475         return 0;
476 }
477
478 static void xen_smp_cpus_done(unsigned int max_cpus)
479 {
480 }
481
482 #ifdef CONFIG_HOTPLUG_CPU
483 static int xen_cpu_disable(void)
484 {
485         unsigned int cpu = smp_processor_id();
486         if (cpu == 0)
487                 return -EBUSY;
488
489         cpu_disable_common();
490
491         load_cr3(swapper_pg_dir);
492         return 0;
493 }
494
495 static void xen_cpu_die(unsigned int cpu)
496 {
497         while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
498                 __set_current_state(TASK_UNINTERRUPTIBLE);
499                 schedule_timeout(HZ/10);
500         }
501
502         cpu_die_common(cpu);
503
504         xen_smp_intr_free(cpu);
505         xen_uninit_lock_cpu(cpu);
506         xen_teardown_timer(cpu);
507 }
508
509 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
510 {
511         play_dead_common();
512         HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
513         cpu_bringup();
514         /*
515          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
516          * clears certain data that the cpu_idle loop (which called us
517          * and that we return from) expects. The only way to get that
518          * data back is to call:
519          */
520         tick_nohz_idle_enter();
521 }
522
523 #else /* !CONFIG_HOTPLUG_CPU */
524 static int xen_cpu_disable(void)
525 {
526         return -ENOSYS;
527 }
528
529 static void xen_cpu_die(unsigned int cpu)
530 {
531         BUG();
532 }
533
534 static void xen_play_dead(void)
535 {
536         BUG();
537 }
538
539 #endif
540 static void stop_self(void *v)
541 {
542         int cpu = smp_processor_id();
543
544         /* make sure we're not pinning something down */
545         load_cr3(swapper_pg_dir);
546         /* should set up a minimal gdt */
547
548         set_cpu_online(cpu, false);
549
550         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
551         BUG();
552 }
553
554 static void xen_stop_other_cpus(int wait)
555 {
556         smp_call_function(stop_self, NULL, wait);
557 }
558
559 static void xen_smp_send_reschedule(int cpu)
560 {
561         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
562 }
563
564 static void __xen_send_IPI_mask(const struct cpumask *mask,
565                               int vector)
566 {
567         unsigned cpu;
568
569         for_each_cpu_and(cpu, mask, cpu_online_mask)
570                 xen_send_IPI_one(cpu, vector);
571 }
572
573 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
574 {
575         int cpu;
576
577         __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
578
579         /* Make sure other vcpus get a chance to run if they need to. */
580         for_each_cpu(cpu, mask) {
581                 if (xen_vcpu_stolen(cpu)) {
582                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
583                         break;
584                 }
585         }
586 }
587
588 static void xen_smp_send_call_function_single_ipi(int cpu)
589 {
590         __xen_send_IPI_mask(cpumask_of(cpu),
591                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
592 }
593
594 static inline int xen_map_vector(int vector)
595 {
596         int xen_vector;
597
598         switch (vector) {
599         case RESCHEDULE_VECTOR:
600                 xen_vector = XEN_RESCHEDULE_VECTOR;
601                 break;
602         case CALL_FUNCTION_VECTOR:
603                 xen_vector = XEN_CALL_FUNCTION_VECTOR;
604                 break;
605         case CALL_FUNCTION_SINGLE_VECTOR:
606                 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
607                 break;
608         case IRQ_WORK_VECTOR:
609                 xen_vector = XEN_IRQ_WORK_VECTOR;
610                 break;
611 #ifdef CONFIG_X86_64
612         case NMI_VECTOR:
613         case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
614                 xen_vector = XEN_NMI_VECTOR;
615                 break;
616 #endif
617         default:
618                 xen_vector = -1;
619                 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
620                         vector);
621         }
622
623         return xen_vector;
624 }
625
626 void xen_send_IPI_mask(const struct cpumask *mask,
627                               int vector)
628 {
629         int xen_vector = xen_map_vector(vector);
630
631         if (xen_vector >= 0)
632                 __xen_send_IPI_mask(mask, xen_vector);
633 }
634
635 void xen_send_IPI_all(int vector)
636 {
637         int xen_vector = xen_map_vector(vector);
638
639         if (xen_vector >= 0)
640                 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
641 }
642
643 void xen_send_IPI_self(int vector)
644 {
645         int xen_vector = xen_map_vector(vector);
646
647         if (xen_vector >= 0)
648                 xen_send_IPI_one(smp_processor_id(), xen_vector);
649 }
650
651 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
652                                 int vector)
653 {
654         unsigned cpu;
655         unsigned int this_cpu = smp_processor_id();
656         int xen_vector = xen_map_vector(vector);
657
658         if (!(num_online_cpus() > 1) || (xen_vector < 0))
659                 return;
660
661         for_each_cpu_and(cpu, mask, cpu_online_mask) {
662                 if (this_cpu == cpu)
663                         continue;
664
665                 xen_send_IPI_one(cpu, xen_vector);
666         }
667 }
668
669 void xen_send_IPI_allbutself(int vector)
670 {
671         xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
672 }
673
674 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
675 {
676         irq_enter();
677         generic_smp_call_function_interrupt();
678         inc_irq_stat(irq_call_count);
679         irq_exit();
680
681         return IRQ_HANDLED;
682 }
683
684 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
685 {
686         irq_enter();
687         generic_smp_call_function_single_interrupt();
688         inc_irq_stat(irq_call_count);
689         irq_exit();
690
691         return IRQ_HANDLED;
692 }
693
694 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
695 {
696         irq_enter();
697         irq_work_run();
698         inc_irq_stat(apic_irq_work_irqs);
699         irq_exit();
700
701         return IRQ_HANDLED;
702 }
703
704 static const struct smp_ops xen_smp_ops __initconst = {
705         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
706         .smp_prepare_cpus = xen_smp_prepare_cpus,
707         .smp_cpus_done = xen_smp_cpus_done,
708
709         .cpu_up = xen_cpu_up,
710         .cpu_die = xen_cpu_die,
711         .cpu_disable = xen_cpu_disable,
712         .play_dead = xen_play_dead,
713
714         .stop_other_cpus = xen_stop_other_cpus,
715         .smp_send_reschedule = xen_smp_send_reschedule,
716
717         .send_call_func_ipi = xen_smp_send_call_function_ipi,
718         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
719 };
720
721 void __init xen_smp_init(void)
722 {
723         smp_ops = xen_smp_ops;
724         xen_fill_possible_map();
725 }
726
727 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
728 {
729         native_smp_prepare_cpus(max_cpus);
730         WARN_ON(xen_smp_intr_init(0));
731
732         xen_init_lock_cpu(0);
733 }
734
735 static int xen_hvm_cpu_up(unsigned int cpu, struct task_struct *tidle)
736 {
737         int rc;
738         /*
739          * xen_smp_intr_init() needs to run before native_cpu_up()
740          * so that IPI vectors are set up on the booting CPU before
741          * it is marked online in native_cpu_up().
742         */
743         rc = xen_smp_intr_init(cpu);
744         WARN_ON(rc);
745         if (!rc)
746                 rc =  native_cpu_up(cpu, tidle);
747
748         /*
749          * We must initialize the slowpath CPU kicker _after_ the native
750          * path has executed. If we initialized it before none of the
751          * unlocker IPI kicks would reach the booting CPU as the booting
752          * CPU had not set itself 'online' in cpu_online_mask. That mask
753          * is checked when IPIs are sent (on HVM at least).
754          */
755         xen_init_lock_cpu(cpu);
756         return rc;
757 }
758
759 static void xen_hvm_cpu_die(unsigned int cpu)
760 {
761         xen_cpu_die(cpu);
762         native_cpu_die(cpu);
763 }
764
765 void __init xen_hvm_smp_init(void)
766 {
767         if (!xen_have_vector_callback)
768                 return;
769         smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
770         smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
771         smp_ops.cpu_up = xen_hvm_cpu_up;
772         smp_ops.cpu_die = xen_hvm_cpu_die;
773         smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
774         smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
775         smp_ops.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu;
776 }