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.
8 * IPIs are handled through the Xen event mechanism.
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
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>
22 #include <asm/paravirt.h>
24 #include <asm/pgtable.h>
27 #include <xen/interface/xen.h>
28 #include <xen/interface/vcpu.h>
30 #include <asm/xen/interface.h>
31 #include <asm/xen/hypercall.h>
35 #include <xen/events.h>
37 #include <xen/hvc-console.h>
41 cpumask_var_t xen_cpu_initialized_map;
43 static DEFINE_PER_CPU(int, xen_resched_irq);
44 static DEFINE_PER_CPU(int, xen_callfunc_irq);
45 static DEFINE_PER_CPU(int, xen_callfuncsingle_irq);
46 static DEFINE_PER_CPU(int, xen_irq_work);
47 static DEFINE_PER_CPU(int, xen_debug_irq) = -1;
49 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
50 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
51 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
54 * Reschedule call back.
56 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
58 inc_irq_stat(irq_resched_count);
64 static void __cpuinit cpu_bringup(void)
69 touch_softlockup_watchdog();
72 xen_enable_sysenter();
75 cpu = smp_processor_id();
76 smp_store_cpu_info(cpu);
77 cpu_data(cpu).x86_max_cores = 1;
78 set_cpu_sibling_map(cpu);
80 xen_setup_cpu_clockevents();
82 notify_cpu_starting(cpu);
84 set_cpu_online(cpu, true);
86 this_cpu_write(cpu_state, CPU_ONLINE);
90 /* We can take interrupts now: we're officially "up". */
93 wmb(); /* make sure everything is out */
96 static void __cpuinit cpu_bringup_and_idle(void)
99 cpu_startup_entry(CPUHP_ONLINE);
102 static int xen_smp_intr_init(unsigned int cpu)
105 const char *resched_name, *callfunc_name, *debug_name;
107 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
108 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
110 xen_reschedule_interrupt,
111 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
116 per_cpu(xen_resched_irq, cpu) = rc;
118 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
119 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
121 xen_call_function_interrupt,
122 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
127 per_cpu(xen_callfunc_irq, cpu) = rc;
129 debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
130 rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
131 IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING,
135 per_cpu(xen_debug_irq, cpu) = rc;
137 callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
138 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
140 xen_call_function_single_interrupt,
141 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
146 per_cpu(xen_callfuncsingle_irq, cpu) = rc;
149 * The IRQ worker on PVHVM goes through the native path and uses the
152 if (xen_hvm_domain())
155 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
156 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
158 xen_irq_work_interrupt,
159 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
164 per_cpu(xen_irq_work, cpu) = rc;
169 if (per_cpu(xen_resched_irq, cpu) >= 0)
170 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
171 if (per_cpu(xen_callfunc_irq, cpu) >= 0)
172 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
173 if (per_cpu(xen_debug_irq, cpu) >= 0)
174 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
175 if (per_cpu(xen_callfuncsingle_irq, cpu) >= 0)
176 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu),
178 if (xen_hvm_domain())
181 if (per_cpu(xen_irq_work, cpu) >= 0)
182 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu), NULL);
187 static void __init xen_fill_possible_map(void)
191 if (xen_initial_domain())
194 for (i = 0; i < nr_cpu_ids; i++) {
195 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
198 set_cpu_possible(i, true);
203 static void __init xen_filter_cpu_maps(void)
206 unsigned int subtract = 0;
208 if (!xen_initial_domain())
213 for (i = 0; i < nr_cpu_ids; i++) {
214 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
217 set_cpu_possible(i, true);
219 set_cpu_possible(i, false);
220 set_cpu_present(i, false);
224 #ifdef CONFIG_HOTPLUG_CPU
225 /* This is akin to using 'nr_cpus' on the Linux command line.
226 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
227 * have up to X, while nr_cpu_ids is greater than X. This
228 * normally is not a problem, except when CPU hotplugging
229 * is involved and then there might be more than X CPUs
230 * in the guest - which will not work as there is no
231 * hypercall to expand the max number of VCPUs an already
232 * running guest has. So cap it up to X. */
234 nr_cpu_ids = nr_cpu_ids - subtract;
239 static void __init xen_smp_prepare_boot_cpu(void)
241 BUG_ON(smp_processor_id() != 0);
242 native_smp_prepare_boot_cpu();
244 /* We've switched to the "real" per-cpu gdt, so make sure the
245 old memory can be recycled */
246 make_lowmem_page_readwrite(xen_initial_gdt);
250 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
253 loadsegment(ds, __USER_DS);
254 loadsegment(es, __USER_DS);
257 xen_filter_cpu_maps();
258 xen_setup_vcpu_info_placement();
261 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
266 if (skip_ioapic_setup) {
267 char *m = (max_cpus == 0) ?
268 "The nosmp parameter is incompatible with Xen; " \
269 "use Xen dom0_max_vcpus=1 parameter" :
270 "The noapic parameter is incompatible with Xen";
275 xen_init_lock_cpu(0);
277 smp_store_boot_cpu_info();
278 cpu_data(0).x86_max_cores = 1;
280 for_each_possible_cpu(i) {
281 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
282 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
283 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
285 set_cpu_sibling_map(0);
287 if (xen_smp_intr_init(0))
290 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
291 panic("could not allocate xen_cpu_initialized_map\n");
293 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
295 /* Restrict the possible_map according to max_cpus. */
296 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
297 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
299 set_cpu_possible(cpu, false);
302 for_each_possible_cpu(cpu)
303 set_cpu_present(cpu, true);
307 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
309 struct vcpu_guest_context *ctxt;
310 struct desc_struct *gdt;
311 unsigned long gdt_mfn;
313 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
316 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
320 gdt = get_cpu_gdt_table(cpu);
322 ctxt->flags = VGCF_IN_KERNEL;
323 ctxt->user_regs.ss = __KERNEL_DS;
325 ctxt->user_regs.fs = __KERNEL_PERCPU;
326 ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
328 ctxt->gs_base_kernel = per_cpu_offset(cpu);
330 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
332 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
335 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
336 ctxt->user_regs.ds = __USER_DS;
337 ctxt->user_regs.es = __USER_DS;
339 xen_copy_trap_info(ctxt->trap_ctxt);
343 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
345 gdt_mfn = arbitrary_virt_to_mfn(gdt);
346 make_lowmem_page_readonly(gdt);
347 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
349 ctxt->gdt_frames[0] = gdt_mfn;
350 ctxt->gdt_ents = GDT_ENTRIES;
352 ctxt->kernel_ss = __KERNEL_DS;
353 ctxt->kernel_sp = idle->thread.sp0;
356 ctxt->event_callback_cs = __KERNEL_CS;
357 ctxt->failsafe_callback_cs = __KERNEL_CS;
359 ctxt->event_callback_eip =
360 (unsigned long)xen_hypervisor_callback;
361 ctxt->failsafe_callback_eip =
362 (unsigned long)xen_failsafe_callback;
364 ctxt->user_regs.cs = __KERNEL_CS;
365 ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
367 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
368 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
370 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
377 static int __cpuinit xen_cpu_up(unsigned int cpu, struct task_struct *idle)
381 per_cpu(current_task, cpu) = idle;
385 clear_tsk_thread_flag(idle, TIF_FORK);
386 per_cpu(kernel_stack, cpu) =
387 (unsigned long)task_stack_page(idle) -
388 KERNEL_STACK_OFFSET + THREAD_SIZE;
390 xen_setup_runstate_info(cpu);
391 xen_setup_timer(cpu);
392 xen_init_lock_cpu(cpu);
394 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
396 /* make sure interrupts start blocked */
397 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
399 rc = cpu_initialize_context(cpu, idle);
403 if (num_online_cpus() == 1)
404 /* Just in case we booted with a single CPU. */
405 alternatives_enable_smp();
407 rc = xen_smp_intr_init(cpu);
411 rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
414 while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
415 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
422 static void xen_smp_cpus_done(unsigned int max_cpus)
426 #ifdef CONFIG_HOTPLUG_CPU
427 static int xen_cpu_disable(void)
429 unsigned int cpu = smp_processor_id();
433 cpu_disable_common();
435 load_cr3(swapper_pg_dir);
439 static void xen_cpu_die(unsigned int cpu)
441 while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
442 current->state = TASK_UNINTERRUPTIBLE;
443 schedule_timeout(HZ/10);
445 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
446 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
447 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
448 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL);
449 if (!xen_hvm_domain())
450 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu), NULL);
451 xen_uninit_lock_cpu(cpu);
452 xen_teardown_timer(cpu);
455 static void __cpuinit xen_play_dead(void) /* used only with HOTPLUG_CPU */
458 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
461 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
462 * clears certain data that the cpu_idle loop (which called us
463 * and that we return from) expects. The only way to get that
464 * data back is to call:
466 tick_nohz_idle_enter();
469 #else /* !CONFIG_HOTPLUG_CPU */
470 static int xen_cpu_disable(void)
475 static void xen_cpu_die(unsigned int cpu)
480 static void xen_play_dead(void)
486 static void stop_self(void *v)
488 int cpu = smp_processor_id();
490 /* make sure we're not pinning something down */
491 load_cr3(swapper_pg_dir);
492 /* should set up a minimal gdt */
494 set_cpu_online(cpu, false);
496 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
500 static void xen_stop_other_cpus(int wait)
502 smp_call_function(stop_self, NULL, wait);
505 static void xen_smp_send_reschedule(int cpu)
507 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
510 static void __xen_send_IPI_mask(const struct cpumask *mask,
515 for_each_cpu_and(cpu, mask, cpu_online_mask)
516 xen_send_IPI_one(cpu, vector);
519 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
523 __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
525 /* Make sure other vcpus get a chance to run if they need to. */
526 for_each_cpu(cpu, mask) {
527 if (xen_vcpu_stolen(cpu)) {
528 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
534 static void xen_smp_send_call_function_single_ipi(int cpu)
536 __xen_send_IPI_mask(cpumask_of(cpu),
537 XEN_CALL_FUNCTION_SINGLE_VECTOR);
540 static inline int xen_map_vector(int vector)
545 case RESCHEDULE_VECTOR:
546 xen_vector = XEN_RESCHEDULE_VECTOR;
548 case CALL_FUNCTION_VECTOR:
549 xen_vector = XEN_CALL_FUNCTION_VECTOR;
551 case CALL_FUNCTION_SINGLE_VECTOR:
552 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
554 case IRQ_WORK_VECTOR:
555 xen_vector = XEN_IRQ_WORK_VECTOR;
559 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
566 void xen_send_IPI_mask(const struct cpumask *mask,
569 int xen_vector = xen_map_vector(vector);
572 __xen_send_IPI_mask(mask, xen_vector);
575 void xen_send_IPI_all(int vector)
577 int xen_vector = xen_map_vector(vector);
580 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
583 void xen_send_IPI_self(int vector)
585 int xen_vector = xen_map_vector(vector);
588 xen_send_IPI_one(smp_processor_id(), xen_vector);
591 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
595 unsigned int this_cpu = smp_processor_id();
596 int xen_vector = xen_map_vector(vector);
598 if (!(num_online_cpus() > 1) || (xen_vector < 0))
601 for_each_cpu_and(cpu, mask, cpu_online_mask) {
605 xen_send_IPI_one(cpu, xen_vector);
609 void xen_send_IPI_allbutself(int vector)
611 xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
614 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
617 generic_smp_call_function_interrupt();
618 inc_irq_stat(irq_call_count);
624 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
627 generic_smp_call_function_single_interrupt();
628 inc_irq_stat(irq_call_count);
634 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
638 inc_irq_stat(apic_irq_work_irqs);
644 static const struct smp_ops xen_smp_ops __initconst = {
645 .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
646 .smp_prepare_cpus = xen_smp_prepare_cpus,
647 .smp_cpus_done = xen_smp_cpus_done,
649 .cpu_up = xen_cpu_up,
650 .cpu_die = xen_cpu_die,
651 .cpu_disable = xen_cpu_disable,
652 .play_dead = xen_play_dead,
654 .stop_other_cpus = xen_stop_other_cpus,
655 .smp_send_reschedule = xen_smp_send_reschedule,
657 .send_call_func_ipi = xen_smp_send_call_function_ipi,
658 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
661 void __init xen_smp_init(void)
663 smp_ops = xen_smp_ops;
664 xen_fill_possible_map();
665 xen_init_spinlocks();
668 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
670 native_smp_prepare_cpus(max_cpus);
671 WARN_ON(xen_smp_intr_init(0));
673 xen_init_lock_cpu(0);
676 static int __cpuinit xen_hvm_cpu_up(unsigned int cpu, struct task_struct *tidle)
680 * xen_smp_intr_init() needs to run before native_cpu_up()
681 * so that IPI vectors are set up on the booting CPU before
682 * it is marked online in native_cpu_up().
684 rc = xen_smp_intr_init(cpu);
687 rc = native_cpu_up(cpu, tidle);
691 static void xen_hvm_cpu_die(unsigned int cpu)
697 void __init xen_hvm_smp_init(void)
699 if (!xen_have_vector_callback)
701 smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
702 smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
703 smp_ops.cpu_up = xen_hvm_cpu_up;
704 smp_ops.cpu_die = xen_hvm_cpu_die;
705 smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
706 smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;