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>
21 #include <linux/nmi.h>
23 #include <asm/paravirt.h>
25 #include <asm/pgtable.h>
28 #include <xen/interface/xen.h>
29 #include <xen/interface/vcpu.h>
30 #include <xen/interface/xenpmu.h>
32 #include <asm/xen/interface.h>
33 #include <asm/xen/hypercall.h>
37 #include <xen/events.h>
39 #include <xen/hvc-console.h>
45 cpumask_var_t xen_cpu_initialized_map;
47 struct xen_common_irq {
51 static DEFINE_PER_CPU(struct xen_common_irq, xen_resched_irq) = { .irq = -1 };
52 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfunc_irq) = { .irq = -1 };
53 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfuncsingle_irq) = { .irq = -1 };
54 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
55 static DEFINE_PER_CPU(struct xen_common_irq, xen_debug_irq) = { .irq = -1 };
56 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
58 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
59 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
60 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
63 * Reschedule call back.
65 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
67 inc_irq_stat(irq_resched_count);
73 static void cpu_bringup(void)
78 touch_softlockup_watchdog();
81 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
82 if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
83 xen_enable_sysenter();
86 cpu = smp_processor_id();
87 smp_store_cpu_info(cpu);
88 cpu_data(cpu).x86_max_cores = 1;
89 set_cpu_sibling_map(cpu);
91 xen_setup_cpu_clockevents();
93 notify_cpu_starting(cpu);
95 set_cpu_online(cpu, true);
97 cpu_set_state_online(cpu); /* Implies full memory barrier. */
99 /* We can take interrupts now: we're officially "up". */
103 asmlinkage __visible void cpu_bringup_and_idle(void)
106 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
109 void xen_smp_intr_free(unsigned int cpu)
111 if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
112 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
113 per_cpu(xen_resched_irq, cpu).irq = -1;
114 kfree(per_cpu(xen_resched_irq, cpu).name);
115 per_cpu(xen_resched_irq, cpu).name = NULL;
117 if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
118 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
119 per_cpu(xen_callfunc_irq, cpu).irq = -1;
120 kfree(per_cpu(xen_callfunc_irq, cpu).name);
121 per_cpu(xen_callfunc_irq, cpu).name = NULL;
123 if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
124 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
125 per_cpu(xen_debug_irq, cpu).irq = -1;
126 kfree(per_cpu(xen_debug_irq, cpu).name);
127 per_cpu(xen_debug_irq, cpu).name = NULL;
129 if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
130 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
132 per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
133 kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
134 per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
138 void xen_smp_intr_free_pv(unsigned int cpu)
140 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
141 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
142 per_cpu(xen_irq_work, cpu).irq = -1;
143 kfree(per_cpu(xen_irq_work, cpu).name);
144 per_cpu(xen_irq_work, cpu).name = NULL;
147 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
148 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
149 per_cpu(xen_pmu_irq, cpu).irq = -1;
150 kfree(per_cpu(xen_pmu_irq, cpu).name);
151 per_cpu(xen_pmu_irq, cpu).name = NULL;
155 int xen_smp_intr_init(unsigned int cpu)
158 char *resched_name, *callfunc_name, *debug_name;
160 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
161 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
163 xen_reschedule_interrupt,
164 IRQF_PERCPU|IRQF_NOBALANCING,
169 per_cpu(xen_resched_irq, cpu).irq = rc;
170 per_cpu(xen_resched_irq, cpu).name = resched_name;
172 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
173 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
175 xen_call_function_interrupt,
176 IRQF_PERCPU|IRQF_NOBALANCING,
181 per_cpu(xen_callfunc_irq, cpu).irq = rc;
182 per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
184 debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
185 rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
186 IRQF_PERCPU | IRQF_NOBALANCING,
190 per_cpu(xen_debug_irq, cpu).irq = rc;
191 per_cpu(xen_debug_irq, cpu).name = debug_name;
193 callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
194 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
196 xen_call_function_single_interrupt,
197 IRQF_PERCPU|IRQF_NOBALANCING,
202 per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
203 per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
208 xen_smp_intr_free(cpu);
212 int xen_smp_intr_init_pv(unsigned int cpu)
215 char *callfunc_name, *pmu_name;
217 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
218 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
220 xen_irq_work_interrupt,
221 IRQF_PERCPU|IRQF_NOBALANCING,
226 per_cpu(xen_irq_work, cpu).irq = rc;
227 per_cpu(xen_irq_work, cpu).name = callfunc_name;
229 if (is_xen_pmu(cpu)) {
230 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
231 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
233 IRQF_PERCPU|IRQF_NOBALANCING,
237 per_cpu(xen_pmu_irq, cpu).irq = rc;
238 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
244 xen_smp_intr_free_pv(cpu);
248 static void __init xen_fill_possible_map(void)
252 if (xen_initial_domain())
255 for (i = 0; i < nr_cpu_ids; i++) {
256 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
259 set_cpu_possible(i, true);
264 static void __init xen_filter_cpu_maps(void)
267 unsigned int subtract = 0;
269 if (!xen_initial_domain())
274 for (i = 0; i < nr_cpu_ids; i++) {
275 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
278 set_cpu_possible(i, true);
280 set_cpu_possible(i, false);
281 set_cpu_present(i, false);
285 #ifdef CONFIG_HOTPLUG_CPU
286 /* This is akin to using 'nr_cpus' on the Linux command line.
287 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
288 * have up to X, while nr_cpu_ids is greater than X. This
289 * normally is not a problem, except when CPU hotplugging
290 * is involved and then there might be more than X CPUs
291 * in the guest - which will not work as there is no
292 * hypercall to expand the max number of VCPUs an already
293 * running guest has. So cap it up to X. */
295 nr_cpu_ids = nr_cpu_ids - subtract;
300 static void __init xen_pv_smp_prepare_boot_cpu(void)
302 BUG_ON(smp_processor_id() != 0);
303 native_smp_prepare_boot_cpu();
305 if (!xen_feature(XENFEAT_writable_page_tables))
306 /* We've switched to the "real" per-cpu gdt, so make
307 * sure the old memory can be recycled. */
308 make_lowmem_page_readwrite(xen_initial_gdt);
312 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
315 loadsegment(ds, __USER_DS);
316 loadsegment(es, __USER_DS);
319 xen_filter_cpu_maps();
320 xen_setup_vcpu_info_placement();
323 * The alternative logic (which patches the unlock/lock) runs before
324 * the smp bootup up code is activated. Hence we need to set this up
325 * the core kernel is being patched. Otherwise we will have only
326 * modules patched but not core code.
328 xen_init_spinlocks();
331 static void __init xen_hvm_smp_prepare_boot_cpu(void)
333 BUG_ON(smp_processor_id() != 0);
334 native_smp_prepare_boot_cpu();
337 * Setup vcpu_info for boot CPU.
342 * The alternative logic (which patches the unlock/lock) runs before
343 * the smp bootup up code is activated. Hence we need to set this up
344 * the core kernel is being patched. Otherwise we will have only
345 * modules patched but not core code.
347 xen_init_spinlocks();
350 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
355 if (skip_ioapic_setup) {
356 char *m = (max_cpus == 0) ?
357 "The nosmp parameter is incompatible with Xen; " \
358 "use Xen dom0_max_vcpus=1 parameter" :
359 "The noapic parameter is incompatible with Xen";
364 xen_init_lock_cpu(0);
366 smp_store_boot_cpu_info();
367 cpu_data(0).x86_max_cores = 1;
369 for_each_possible_cpu(i) {
370 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
371 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
372 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
374 set_cpu_sibling_map(0);
378 if (xen_smp_intr_init(0))
381 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
382 panic("could not allocate xen_cpu_initialized_map\n");
384 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
386 /* Restrict the possible_map according to max_cpus. */
387 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
388 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
390 set_cpu_possible(cpu, false);
393 for_each_possible_cpu(cpu)
394 set_cpu_present(cpu, true);
398 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
400 struct vcpu_guest_context *ctxt;
401 struct desc_struct *gdt;
402 unsigned long gdt_mfn;
404 /* used to tell cpu_init() that it can proceed with initialization */
405 cpumask_set_cpu(cpu, cpu_callout_mask);
406 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
409 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
413 gdt = get_cpu_gdt_rw(cpu);
416 ctxt->user_regs.fs = __KERNEL_PERCPU;
417 ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
419 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
421 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
422 ctxt->flags = VGCF_IN_KERNEL;
423 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
424 ctxt->user_regs.ds = __USER_DS;
425 ctxt->user_regs.es = __USER_DS;
426 ctxt->user_regs.ss = __KERNEL_DS;
428 xen_copy_trap_info(ctxt->trap_ctxt);
432 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
434 gdt_mfn = arbitrary_virt_to_mfn(gdt);
435 make_lowmem_page_readonly(gdt);
436 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
438 ctxt->gdt_frames[0] = gdt_mfn;
439 ctxt->gdt_ents = GDT_ENTRIES;
441 ctxt->kernel_ss = __KERNEL_DS;
442 ctxt->kernel_sp = idle->thread.sp0;
445 ctxt->event_callback_cs = __KERNEL_CS;
446 ctxt->failsafe_callback_cs = __KERNEL_CS;
448 ctxt->gs_base_kernel = per_cpu_offset(cpu);
450 ctxt->event_callback_eip =
451 (unsigned long)xen_hypervisor_callback;
452 ctxt->failsafe_callback_eip =
453 (unsigned long)xen_failsafe_callback;
454 ctxt->user_regs.cs = __KERNEL_CS;
455 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
457 ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
458 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
459 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
466 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
470 common_cpu_up(cpu, idle);
472 xen_setup_runstate_info(cpu);
475 * PV VCPUs are always successfully taken down (see 'while' loop
476 * in xen_cpu_die()), so -EBUSY is an error.
478 rc = cpu_check_up_prepare(cpu);
482 /* make sure interrupts start blocked */
483 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
485 rc = cpu_initialize_context(cpu, idle);
491 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
494 while (cpu_report_state(cpu) != CPU_ONLINE)
495 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
500 static void xen_smp_cpus_done(unsigned int max_cpus)
504 #ifdef CONFIG_HOTPLUG_CPU
505 static int xen_cpu_disable(void)
507 unsigned int cpu = smp_processor_id();
511 cpu_disable_common();
513 load_cr3(swapper_pg_dir);
517 static void xen_pv_cpu_die(unsigned int cpu)
519 while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
520 xen_vcpu_nr(cpu), NULL)) {
521 __set_current_state(TASK_UNINTERRUPTIBLE);
522 schedule_timeout(HZ/10);
525 if (common_cpu_die(cpu) == 0) {
526 xen_smp_intr_free(cpu);
527 xen_uninit_lock_cpu(cpu);
528 xen_teardown_timer(cpu);
533 static void xen_hvm_cpu_die(unsigned int cpu)
535 if (common_cpu_die(cpu) == 0) {
536 xen_smp_intr_free(cpu);
537 xen_uninit_lock_cpu(cpu);
538 xen_teardown_timer(cpu);
542 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
545 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
548 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
549 * clears certain data that the cpu_idle loop (which called us
550 * and that we return from) expects. The only way to get that
551 * data back is to call:
553 tick_nohz_idle_enter();
555 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
558 #else /* !CONFIG_HOTPLUG_CPU */
559 static int xen_cpu_disable(void)
564 static void xen_pv_cpu_die(unsigned int cpu)
569 static void xen_hvm_cpu_die(unsigned int cpu)
574 static void xen_play_dead(void)
580 static void stop_self(void *v)
582 int cpu = smp_processor_id();
584 /* make sure we're not pinning something down */
585 load_cr3(swapper_pg_dir);
586 /* should set up a minimal gdt */
588 set_cpu_online(cpu, false);
590 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
594 static void xen_stop_other_cpus(int wait)
596 smp_call_function(stop_self, NULL, wait);
599 static void xen_smp_send_reschedule(int cpu)
601 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
604 static void __xen_send_IPI_mask(const struct cpumask *mask,
609 for_each_cpu_and(cpu, mask, cpu_online_mask)
610 xen_send_IPI_one(cpu, vector);
613 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
617 __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
619 /* Make sure other vcpus get a chance to run if they need to. */
620 for_each_cpu(cpu, mask) {
621 if (xen_vcpu_stolen(cpu)) {
622 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
628 static void xen_smp_send_call_function_single_ipi(int cpu)
630 __xen_send_IPI_mask(cpumask_of(cpu),
631 XEN_CALL_FUNCTION_SINGLE_VECTOR);
634 static inline int xen_map_vector(int vector)
639 case RESCHEDULE_VECTOR:
640 xen_vector = XEN_RESCHEDULE_VECTOR;
642 case CALL_FUNCTION_VECTOR:
643 xen_vector = XEN_CALL_FUNCTION_VECTOR;
645 case CALL_FUNCTION_SINGLE_VECTOR:
646 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
648 case IRQ_WORK_VECTOR:
649 xen_vector = XEN_IRQ_WORK_VECTOR;
653 case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
654 xen_vector = XEN_NMI_VECTOR;
659 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
666 void xen_send_IPI_mask(const struct cpumask *mask,
669 int xen_vector = xen_map_vector(vector);
672 __xen_send_IPI_mask(mask, xen_vector);
675 void xen_send_IPI_all(int vector)
677 int xen_vector = xen_map_vector(vector);
680 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
683 void xen_send_IPI_self(int vector)
685 int xen_vector = xen_map_vector(vector);
688 xen_send_IPI_one(smp_processor_id(), xen_vector);
691 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
695 unsigned int this_cpu = smp_processor_id();
696 int xen_vector = xen_map_vector(vector);
698 if (!(num_online_cpus() > 1) || (xen_vector < 0))
701 for_each_cpu_and(cpu, mask, cpu_online_mask) {
705 xen_send_IPI_one(cpu, xen_vector);
709 void xen_send_IPI_allbutself(int vector)
711 xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
714 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
717 generic_smp_call_function_interrupt();
718 inc_irq_stat(irq_call_count);
724 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
727 generic_smp_call_function_single_interrupt();
728 inc_irq_stat(irq_call_count);
734 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
738 inc_irq_stat(apic_irq_work_irqs);
744 static const struct smp_ops xen_smp_ops __initconst = {
745 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
746 .smp_prepare_cpus = xen_smp_prepare_cpus,
747 .smp_cpus_done = xen_smp_cpus_done,
749 .cpu_up = xen_cpu_up,
750 .cpu_die = xen_pv_cpu_die,
751 .cpu_disable = xen_cpu_disable,
752 .play_dead = xen_play_dead,
754 .stop_other_cpus = xen_stop_other_cpus,
755 .smp_send_reschedule = xen_smp_send_reschedule,
757 .send_call_func_ipi = xen_smp_send_call_function_ipi,
758 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
761 void __init xen_smp_init(void)
763 smp_ops = xen_smp_ops;
764 xen_fill_possible_map();
767 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
769 native_smp_prepare_cpus(max_cpus);
770 WARN_ON(xen_smp_intr_init(0));
772 xen_init_lock_cpu(0);
775 void __init xen_hvm_smp_init(void)
777 smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
778 smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
779 smp_ops.cpu_die = xen_hvm_cpu_die;
780 smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
781 smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
782 smp_ops.smp_prepare_boot_cpu = xen_hvm_smp_prepare_boot_cpu;