xref: /linux/arch/x86/xen/smp.c (revision 80d443e8876602be2c130f79c4de81e12e2a700d)
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 #include <xen/interface/xenpmu.h>
30 
31 #include <asm/xen/interface.h>
32 #include <asm/xen/hypercall.h>
33 
34 #include <xen/xen.h>
35 #include <xen/page.h>
36 #include <xen/events.h>
37 
38 #include <xen/hvc-console.h>
39 #include "xen-ops.h"
40 #include "mmu.h"
41 #include "smp.h"
42 #include "pmu.h"
43 
44 cpumask_var_t xen_cpu_initialized_map;
45 
46 struct xen_common_irq {
47 	int irq;
48 	char *name;
49 };
50 static DEFINE_PER_CPU(struct xen_common_irq, xen_resched_irq) = { .irq = -1 };
51 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfunc_irq) = { .irq = -1 };
52 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfuncsingle_irq) = { .irq = -1 };
53 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
54 static DEFINE_PER_CPU(struct xen_common_irq, xen_debug_irq) = { .irq = -1 };
55 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
56 
57 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
58 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
59 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
60 
61 /*
62  * Reschedule call back.
63  */
64 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
65 {
66 	inc_irq_stat(irq_resched_count);
67 	scheduler_ipi();
68 
69 	return IRQ_HANDLED;
70 }
71 
72 static void cpu_bringup(void)
73 {
74 	int cpu;
75 
76 	cpu_init();
77 	touch_softlockup_watchdog();
78 	preempt_disable();
79 
80 	/* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
81 	if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
82 		xen_enable_sysenter();
83 		xen_enable_syscall();
84 	}
85 	cpu = smp_processor_id();
86 	smp_store_cpu_info(cpu);
87 	cpu_data(cpu).x86_max_cores = 1;
88 	set_cpu_sibling_map(cpu);
89 
90 	xen_setup_cpu_clockevents();
91 
92 	notify_cpu_starting(cpu);
93 
94 	set_cpu_online(cpu, true);
95 
96 	cpu_set_state_online(cpu);  /* Implies full memory barrier. */
97 
98 	/* We can take interrupts now: we're officially "up". */
99 	local_irq_enable();
100 }
101 
102 /*
103  * Note: cpu parameter is only relevant for PVH. The reason for passing it
104  * is we can't do smp_processor_id until the percpu segments are loaded, for
105  * which we need the cpu number! So we pass it in rdi as first parameter.
106  */
107 asmlinkage __visible void cpu_bringup_and_idle(int cpu)
108 {
109 #ifdef CONFIG_XEN_PVH
110 	if (xen_feature(XENFEAT_auto_translated_physmap) &&
111 	    xen_feature(XENFEAT_supervisor_mode_kernel))
112 		xen_pvh_secondary_vcpu_init(cpu);
113 #endif
114 	cpu_bringup();
115 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
116 }
117 
118 void xen_smp_intr_free(unsigned int cpu)
119 {
120 	if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
121 		unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
122 		per_cpu(xen_resched_irq, cpu).irq = -1;
123 		kfree(per_cpu(xen_resched_irq, cpu).name);
124 		per_cpu(xen_resched_irq, cpu).name = NULL;
125 	}
126 	if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
127 		unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
128 		per_cpu(xen_callfunc_irq, cpu).irq = -1;
129 		kfree(per_cpu(xen_callfunc_irq, cpu).name);
130 		per_cpu(xen_callfunc_irq, cpu).name = NULL;
131 	}
132 	if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
133 		unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
134 		per_cpu(xen_debug_irq, cpu).irq = -1;
135 		kfree(per_cpu(xen_debug_irq, cpu).name);
136 		per_cpu(xen_debug_irq, cpu).name = NULL;
137 	}
138 	if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
139 		unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
140 				       NULL);
141 		per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
142 		kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
143 		per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
144 	}
145 	if (xen_hvm_domain())
146 		return;
147 
148 	if (per_cpu(xen_irq_work, cpu).irq >= 0) {
149 		unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
150 		per_cpu(xen_irq_work, cpu).irq = -1;
151 		kfree(per_cpu(xen_irq_work, cpu).name);
152 		per_cpu(xen_irq_work, cpu).name = NULL;
153 	}
154 
155 	if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
156 		unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
157 		per_cpu(xen_pmu_irq, cpu).irq = -1;
158 		kfree(per_cpu(xen_pmu_irq, cpu).name);
159 		per_cpu(xen_pmu_irq, cpu).name = NULL;
160 	}
161 };
162 int xen_smp_intr_init(unsigned int cpu)
163 {
164 	int rc;
165 	char *resched_name, *callfunc_name, *debug_name, *pmu_name;
166 
167 	resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
168 	rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
169 				    cpu,
170 				    xen_reschedule_interrupt,
171 				    IRQF_PERCPU|IRQF_NOBALANCING,
172 				    resched_name,
173 				    NULL);
174 	if (rc < 0)
175 		goto fail;
176 	per_cpu(xen_resched_irq, cpu).irq = rc;
177 	per_cpu(xen_resched_irq, cpu).name = resched_name;
178 
179 	callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
180 	rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
181 				    cpu,
182 				    xen_call_function_interrupt,
183 				    IRQF_PERCPU|IRQF_NOBALANCING,
184 				    callfunc_name,
185 				    NULL);
186 	if (rc < 0)
187 		goto fail;
188 	per_cpu(xen_callfunc_irq, cpu).irq = rc;
189 	per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
190 
191 	debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
192 	rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
193 				     IRQF_PERCPU | IRQF_NOBALANCING,
194 				     debug_name, NULL);
195 	if (rc < 0)
196 		goto fail;
197 	per_cpu(xen_debug_irq, cpu).irq = rc;
198 	per_cpu(xen_debug_irq, cpu).name = debug_name;
199 
200 	callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
201 	rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
202 				    cpu,
203 				    xen_call_function_single_interrupt,
204 				    IRQF_PERCPU|IRQF_NOBALANCING,
205 				    callfunc_name,
206 				    NULL);
207 	if (rc < 0)
208 		goto fail;
209 	per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
210 	per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
211 
212 	/*
213 	 * The IRQ worker on PVHVM goes through the native path and uses the
214 	 * IPI mechanism.
215 	 */
216 	if (xen_hvm_domain())
217 		return 0;
218 
219 	callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
220 	rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
221 				    cpu,
222 				    xen_irq_work_interrupt,
223 				    IRQF_PERCPU|IRQF_NOBALANCING,
224 				    callfunc_name,
225 				    NULL);
226 	if (rc < 0)
227 		goto fail;
228 	per_cpu(xen_irq_work, cpu).irq = rc;
229 	per_cpu(xen_irq_work, cpu).name = callfunc_name;
230 
231 	if (is_xen_pmu(cpu)) {
232 		pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
233 		rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
234 					     xen_pmu_irq_handler,
235 					     IRQF_PERCPU|IRQF_NOBALANCING,
236 					     pmu_name, NULL);
237 		if (rc < 0)
238 			goto fail;
239 		per_cpu(xen_pmu_irq, cpu).irq = rc;
240 		per_cpu(xen_pmu_irq, cpu).name = pmu_name;
241 	}
242 
243 	return 0;
244 
245  fail:
246 	xen_smp_intr_free(cpu);
247 	return rc;
248 }
249 
250 static void __init xen_fill_possible_map(void)
251 {
252 	int i, rc;
253 
254 	if (xen_initial_domain())
255 		return;
256 
257 	for (i = 0; i < nr_cpu_ids; i++) {
258 		rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
259 		if (rc >= 0) {
260 			num_processors++;
261 			set_cpu_possible(i, true);
262 		}
263 	}
264 }
265 
266 static void __init xen_filter_cpu_maps(void)
267 {
268 	int i, rc;
269 	unsigned int subtract = 0;
270 
271 	if (!xen_initial_domain())
272 		return;
273 
274 	num_processors = 0;
275 	disabled_cpus = 0;
276 	for (i = 0; i < nr_cpu_ids; i++) {
277 		rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
278 		if (rc >= 0) {
279 			num_processors++;
280 			set_cpu_possible(i, true);
281 		} else {
282 			set_cpu_possible(i, false);
283 			set_cpu_present(i, false);
284 			subtract++;
285 		}
286 	}
287 #ifdef CONFIG_HOTPLUG_CPU
288 	/* This is akin to using 'nr_cpus' on the Linux command line.
289 	 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
290 	 * have up to X, while nr_cpu_ids is greater than X. This
291 	 * normally is not a problem, except when CPU hotplugging
292 	 * is involved and then there might be more than X CPUs
293 	 * in the guest - which will not work as there is no
294 	 * hypercall to expand the max number of VCPUs an already
295 	 * running guest has. So cap it up to X. */
296 	if (subtract)
297 		nr_cpu_ids = nr_cpu_ids - subtract;
298 #endif
299 
300 }
301 
302 static void __init xen_smp_prepare_boot_cpu(void)
303 {
304 	BUG_ON(smp_processor_id() != 0);
305 	native_smp_prepare_boot_cpu();
306 
307 	if (xen_pv_domain()) {
308 		if (!xen_feature(XENFEAT_writable_page_tables))
309 			/* We've switched to the "real" per-cpu gdt, so make
310 			 * sure the old memory can be recycled. */
311 			make_lowmem_page_readwrite(xen_initial_gdt);
312 
313 #ifdef CONFIG_X86_32
314 		/*
315 		 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
316 		 * expects __USER_DS
317 		 */
318 		loadsegment(ds, __USER_DS);
319 		loadsegment(es, __USER_DS);
320 #endif
321 
322 		xen_filter_cpu_maps();
323 		xen_setup_vcpu_info_placement();
324 	}
325 
326 	/*
327 	 * Setup vcpu_info for boot CPU.
328 	 */
329 	if (xen_hvm_domain())
330 		xen_vcpu_setup(0);
331 
332 	/*
333 	 * The alternative logic (which patches the unlock/lock) runs before
334 	 * the smp bootup up code is activated. Hence we need to set this up
335 	 * the core kernel is being patched. Otherwise we will have only
336 	 * modules patched but not core code.
337 	 */
338 	xen_init_spinlocks();
339 }
340 
341 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
342 {
343 	unsigned cpu;
344 	unsigned int i;
345 
346 	if (skip_ioapic_setup) {
347 		char *m = (max_cpus == 0) ?
348 			"The nosmp parameter is incompatible with Xen; " \
349 			"use Xen dom0_max_vcpus=1 parameter" :
350 			"The noapic parameter is incompatible with Xen";
351 
352 		xen_raw_printk(m);
353 		panic(m);
354 	}
355 	xen_init_lock_cpu(0);
356 
357 	smp_store_boot_cpu_info();
358 	cpu_data(0).x86_max_cores = 1;
359 
360 	for_each_possible_cpu(i) {
361 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
362 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
363 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
364 	}
365 	set_cpu_sibling_map(0);
366 
367 	xen_pmu_init(0);
368 
369 	if (xen_smp_intr_init(0))
370 		BUG();
371 
372 	if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
373 		panic("could not allocate xen_cpu_initialized_map\n");
374 
375 	cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
376 
377 	/* Restrict the possible_map according to max_cpus. */
378 	while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
379 		for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
380 			continue;
381 		set_cpu_possible(cpu, false);
382 	}
383 
384 	for_each_possible_cpu(cpu)
385 		set_cpu_present(cpu, true);
386 }
387 
388 static int
389 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
390 {
391 	struct vcpu_guest_context *ctxt;
392 	struct desc_struct *gdt;
393 	unsigned long gdt_mfn;
394 
395 	/* used to tell cpu_init() that it can proceed with initialization */
396 	cpumask_set_cpu(cpu, cpu_callout_mask);
397 	if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
398 		return 0;
399 
400 	ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
401 	if (ctxt == NULL)
402 		return -ENOMEM;
403 
404 	gdt = get_cpu_gdt_table(cpu);
405 
406 #ifdef CONFIG_X86_32
407 	/* Note: PVH is not yet supported on x86_32. */
408 	ctxt->user_regs.fs = __KERNEL_PERCPU;
409 	ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
410 #endif
411 	memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
412 
413 	if (!xen_feature(XENFEAT_auto_translated_physmap)) {
414 		ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
415 		ctxt->flags = VGCF_IN_KERNEL;
416 		ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
417 		ctxt->user_regs.ds = __USER_DS;
418 		ctxt->user_regs.es = __USER_DS;
419 		ctxt->user_regs.ss = __KERNEL_DS;
420 
421 		xen_copy_trap_info(ctxt->trap_ctxt);
422 
423 		ctxt->ldt_ents = 0;
424 
425 		BUG_ON((unsigned long)gdt & ~PAGE_MASK);
426 
427 		gdt_mfn = arbitrary_virt_to_mfn(gdt);
428 		make_lowmem_page_readonly(gdt);
429 		make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
430 
431 		ctxt->gdt_frames[0] = gdt_mfn;
432 		ctxt->gdt_ents      = GDT_ENTRIES;
433 
434 		ctxt->kernel_ss = __KERNEL_DS;
435 		ctxt->kernel_sp = idle->thread.sp0;
436 
437 #ifdef CONFIG_X86_32
438 		ctxt->event_callback_cs     = __KERNEL_CS;
439 		ctxt->failsafe_callback_cs  = __KERNEL_CS;
440 #else
441 		ctxt->gs_base_kernel = per_cpu_offset(cpu);
442 #endif
443 		ctxt->event_callback_eip    =
444 					(unsigned long)xen_hypervisor_callback;
445 		ctxt->failsafe_callback_eip =
446 					(unsigned long)xen_failsafe_callback;
447 		ctxt->user_regs.cs = __KERNEL_CS;
448 		per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
449 	}
450 #ifdef CONFIG_XEN_PVH
451 	else {
452 		/*
453 		 * The vcpu comes on kernel page tables which have the NX pte
454 		 * bit set. This means before DS/SS is touched, NX in
455 		 * EFER must be set. Hence the following assembly glue code.
456 		 */
457 		ctxt->user_regs.eip = (unsigned long)xen_pvh_early_cpu_init;
458 		ctxt->user_regs.rdi = cpu;
459 		ctxt->user_regs.rsi = true;  /* entry == true */
460 	}
461 #endif
462 	ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
463 	ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
464 	if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
465 		BUG();
466 
467 	kfree(ctxt);
468 	return 0;
469 }
470 
471 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
472 {
473 	int rc;
474 
475 	common_cpu_up(cpu, idle);
476 
477 	xen_setup_runstate_info(cpu);
478 
479 	/*
480 	 * PV VCPUs are always successfully taken down (see 'while' loop
481 	 * in xen_cpu_die()), so -EBUSY is an error.
482 	 */
483 	rc = cpu_check_up_prepare(cpu);
484 	if (rc)
485 		return rc;
486 
487 	/* make sure interrupts start blocked */
488 	per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
489 
490 	rc = cpu_initialize_context(cpu, idle);
491 	if (rc)
492 		return rc;
493 
494 	xen_pmu_init(cpu);
495 
496 	rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
497 	BUG_ON(rc);
498 
499 	while (cpu_report_state(cpu) != CPU_ONLINE)
500 		HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
501 
502 	return 0;
503 }
504 
505 static void xen_smp_cpus_done(unsigned int max_cpus)
506 {
507 }
508 
509 #ifdef CONFIG_HOTPLUG_CPU
510 static int xen_cpu_disable(void)
511 {
512 	unsigned int cpu = smp_processor_id();
513 	if (cpu == 0)
514 		return -EBUSY;
515 
516 	cpu_disable_common();
517 
518 	load_cr3(swapper_pg_dir);
519 	return 0;
520 }
521 
522 static void xen_cpu_die(unsigned int cpu)
523 {
524 	while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up,
525 						     xen_vcpu_nr(cpu), NULL)) {
526 		__set_current_state(TASK_UNINTERRUPTIBLE);
527 		schedule_timeout(HZ/10);
528 	}
529 
530 	if (common_cpu_die(cpu) == 0) {
531 		xen_smp_intr_free(cpu);
532 		xen_uninit_lock_cpu(cpu);
533 		xen_teardown_timer(cpu);
534 		xen_pmu_finish(cpu);
535 	}
536 }
537 
538 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
539 {
540 	play_dead_common();
541 	HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
542 	cpu_bringup();
543 	/*
544 	 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
545 	 * clears certain data that the cpu_idle loop (which called us
546 	 * and that we return from) expects. The only way to get that
547 	 * data back is to call:
548 	 */
549 	tick_nohz_idle_enter();
550 
551 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
552 }
553 
554 #else /* !CONFIG_HOTPLUG_CPU */
555 static int xen_cpu_disable(void)
556 {
557 	return -ENOSYS;
558 }
559 
560 static void xen_cpu_die(unsigned int cpu)
561 {
562 	BUG();
563 }
564 
565 static void xen_play_dead(void)
566 {
567 	BUG();
568 }
569 
570 #endif
571 static void stop_self(void *v)
572 {
573 	int cpu = smp_processor_id();
574 
575 	/* make sure we're not pinning something down */
576 	load_cr3(swapper_pg_dir);
577 	/* should set up a minimal gdt */
578 
579 	set_cpu_online(cpu, false);
580 
581 	HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
582 	BUG();
583 }
584 
585 static void xen_stop_other_cpus(int wait)
586 {
587 	smp_call_function(stop_self, NULL, wait);
588 }
589 
590 static void xen_smp_send_reschedule(int cpu)
591 {
592 	xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
593 }
594 
595 static void __xen_send_IPI_mask(const struct cpumask *mask,
596 			      int vector)
597 {
598 	unsigned cpu;
599 
600 	for_each_cpu_and(cpu, mask, cpu_online_mask)
601 		xen_send_IPI_one(cpu, vector);
602 }
603 
604 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
605 {
606 	int cpu;
607 
608 	__xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
609 
610 	/* Make sure other vcpus get a chance to run if they need to. */
611 	for_each_cpu(cpu, mask) {
612 		if (xen_vcpu_stolen(cpu)) {
613 			HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
614 			break;
615 		}
616 	}
617 }
618 
619 static void xen_smp_send_call_function_single_ipi(int cpu)
620 {
621 	__xen_send_IPI_mask(cpumask_of(cpu),
622 			  XEN_CALL_FUNCTION_SINGLE_VECTOR);
623 }
624 
625 static inline int xen_map_vector(int vector)
626 {
627 	int xen_vector;
628 
629 	switch (vector) {
630 	case RESCHEDULE_VECTOR:
631 		xen_vector = XEN_RESCHEDULE_VECTOR;
632 		break;
633 	case CALL_FUNCTION_VECTOR:
634 		xen_vector = XEN_CALL_FUNCTION_VECTOR;
635 		break;
636 	case CALL_FUNCTION_SINGLE_VECTOR:
637 		xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
638 		break;
639 	case IRQ_WORK_VECTOR:
640 		xen_vector = XEN_IRQ_WORK_VECTOR;
641 		break;
642 #ifdef CONFIG_X86_64
643 	case NMI_VECTOR:
644 	case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
645 		xen_vector = XEN_NMI_VECTOR;
646 		break;
647 #endif
648 	default:
649 		xen_vector = -1;
650 		printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
651 			vector);
652 	}
653 
654 	return xen_vector;
655 }
656 
657 void xen_send_IPI_mask(const struct cpumask *mask,
658 			      int vector)
659 {
660 	int xen_vector = xen_map_vector(vector);
661 
662 	if (xen_vector >= 0)
663 		__xen_send_IPI_mask(mask, xen_vector);
664 }
665 
666 void xen_send_IPI_all(int vector)
667 {
668 	int xen_vector = xen_map_vector(vector);
669 
670 	if (xen_vector >= 0)
671 		__xen_send_IPI_mask(cpu_online_mask, xen_vector);
672 }
673 
674 void xen_send_IPI_self(int vector)
675 {
676 	int xen_vector = xen_map_vector(vector);
677 
678 	if (xen_vector >= 0)
679 		xen_send_IPI_one(smp_processor_id(), xen_vector);
680 }
681 
682 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
683 				int vector)
684 {
685 	unsigned cpu;
686 	unsigned int this_cpu = smp_processor_id();
687 	int xen_vector = xen_map_vector(vector);
688 
689 	if (!(num_online_cpus() > 1) || (xen_vector < 0))
690 		return;
691 
692 	for_each_cpu_and(cpu, mask, cpu_online_mask) {
693 		if (this_cpu == cpu)
694 			continue;
695 
696 		xen_send_IPI_one(cpu, xen_vector);
697 	}
698 }
699 
700 void xen_send_IPI_allbutself(int vector)
701 {
702 	xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
703 }
704 
705 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
706 {
707 	irq_enter();
708 	generic_smp_call_function_interrupt();
709 	inc_irq_stat(irq_call_count);
710 	irq_exit();
711 
712 	return IRQ_HANDLED;
713 }
714 
715 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
716 {
717 	irq_enter();
718 	generic_smp_call_function_single_interrupt();
719 	inc_irq_stat(irq_call_count);
720 	irq_exit();
721 
722 	return IRQ_HANDLED;
723 }
724 
725 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
726 {
727 	irq_enter();
728 	irq_work_run();
729 	inc_irq_stat(apic_irq_work_irqs);
730 	irq_exit();
731 
732 	return IRQ_HANDLED;
733 }
734 
735 static const struct smp_ops xen_smp_ops __initconst = {
736 	.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
737 	.smp_prepare_cpus = xen_smp_prepare_cpus,
738 	.smp_cpus_done = xen_smp_cpus_done,
739 
740 	.cpu_up = xen_cpu_up,
741 	.cpu_die = xen_cpu_die,
742 	.cpu_disable = xen_cpu_disable,
743 	.play_dead = xen_play_dead,
744 
745 	.stop_other_cpus = xen_stop_other_cpus,
746 	.smp_send_reschedule = xen_smp_send_reschedule,
747 
748 	.send_call_func_ipi = xen_smp_send_call_function_ipi,
749 	.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
750 };
751 
752 void __init xen_smp_init(void)
753 {
754 	smp_ops = xen_smp_ops;
755 	xen_fill_possible_map();
756 }
757 
758 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
759 {
760 	native_smp_prepare_cpus(max_cpus);
761 	WARN_ON(xen_smp_intr_init(0));
762 
763 	xen_init_lock_cpu(0);
764 }
765 
766 void __init xen_hvm_smp_init(void)
767 {
768 	smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
769 	smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
770 	smp_ops.cpu_die = xen_cpu_die;
771 	smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
772 	smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
773 	smp_ops.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu;
774 }
775