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>
29 #include <xen/interface/xenpmu.h>
31 #include <asm/xen/interface.h>
32 #include <asm/xen/hypercall.h>
36 #include <xen/events.h>
38 #include <xen/hvc-console.h>
44 cpumask_var_t xen_cpu_initialized_map;
46 struct xen_common_irq {
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 };
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);
62 * Reschedule call back.
64 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
66 inc_irq_stat(irq_resched_count);
72 static void cpu_bringup(void)
77 touch_softlockup_watchdog();
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();
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);
90 xen_setup_cpu_clockevents();
92 notify_cpu_starting(cpu);
94 set_cpu_online(cpu, true);
96 cpu_set_state_online(cpu); /* Implies full memory barrier. */
98 /* We can take interrupts now: we're officially "up". */
102 asmlinkage __visible void cpu_bringup_and_idle(void)
105 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
108 void xen_smp_intr_free(unsigned int cpu)
110 if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
111 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
112 per_cpu(xen_resched_irq, cpu).irq = -1;
113 kfree(per_cpu(xen_resched_irq, cpu).name);
114 per_cpu(xen_resched_irq, cpu).name = NULL;
116 if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
117 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
118 per_cpu(xen_callfunc_irq, cpu).irq = -1;
119 kfree(per_cpu(xen_callfunc_irq, cpu).name);
120 per_cpu(xen_callfunc_irq, cpu).name = NULL;
122 if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
123 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
124 per_cpu(xen_debug_irq, cpu).irq = -1;
125 kfree(per_cpu(xen_debug_irq, cpu).name);
126 per_cpu(xen_debug_irq, cpu).name = NULL;
128 if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
129 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
131 per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
132 kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
133 per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
135 if (xen_hvm_domain())
138 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
139 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
140 per_cpu(xen_irq_work, cpu).irq = -1;
141 kfree(per_cpu(xen_irq_work, cpu).name);
142 per_cpu(xen_irq_work, cpu).name = NULL;
145 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
146 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
147 per_cpu(xen_pmu_irq, cpu).irq = -1;
148 kfree(per_cpu(xen_pmu_irq, cpu).name);
149 per_cpu(xen_pmu_irq, cpu).name = NULL;
152 int xen_smp_intr_init(unsigned int cpu)
155 char *resched_name, *callfunc_name, *debug_name, *pmu_name;
157 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
158 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
160 xen_reschedule_interrupt,
161 IRQF_PERCPU|IRQF_NOBALANCING,
166 per_cpu(xen_resched_irq, cpu).irq = rc;
167 per_cpu(xen_resched_irq, cpu).name = resched_name;
169 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
170 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
172 xen_call_function_interrupt,
173 IRQF_PERCPU|IRQF_NOBALANCING,
178 per_cpu(xen_callfunc_irq, cpu).irq = rc;
179 per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
181 debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
182 rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
183 IRQF_PERCPU | IRQF_NOBALANCING,
187 per_cpu(xen_debug_irq, cpu).irq = rc;
188 per_cpu(xen_debug_irq, cpu).name = debug_name;
190 callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
191 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
193 xen_call_function_single_interrupt,
194 IRQF_PERCPU|IRQF_NOBALANCING,
199 per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
200 per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
203 * The IRQ worker on PVHVM goes through the native path and uses the
206 if (xen_hvm_domain())
209 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
210 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
212 xen_irq_work_interrupt,
213 IRQF_PERCPU|IRQF_NOBALANCING,
218 per_cpu(xen_irq_work, cpu).irq = rc;
219 per_cpu(xen_irq_work, cpu).name = callfunc_name;
221 if (is_xen_pmu(cpu)) {
222 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
223 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
225 IRQF_PERCPU|IRQF_NOBALANCING,
229 per_cpu(xen_pmu_irq, cpu).irq = rc;
230 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
236 xen_smp_intr_free(cpu);
240 static void __init xen_fill_possible_map(void)
244 if (xen_initial_domain())
247 for (i = 0; i < nr_cpu_ids; i++) {
248 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
251 set_cpu_possible(i, true);
256 static void __init xen_filter_cpu_maps(void)
259 unsigned int subtract = 0;
261 if (!xen_initial_domain())
266 for (i = 0; i < nr_cpu_ids; i++) {
267 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
270 set_cpu_possible(i, true);
272 set_cpu_possible(i, false);
273 set_cpu_present(i, false);
277 #ifdef CONFIG_HOTPLUG_CPU
278 /* This is akin to using 'nr_cpus' on the Linux command line.
279 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
280 * have up to X, while nr_cpu_ids is greater than X. This
281 * normally is not a problem, except when CPU hotplugging
282 * is involved and then there might be more than X CPUs
283 * in the guest - which will not work as there is no
284 * hypercall to expand the max number of VCPUs an already
285 * running guest has. So cap it up to X. */
287 nr_cpu_ids = nr_cpu_ids - subtract;
292 static void __init xen_smp_prepare_boot_cpu(void)
294 BUG_ON(smp_processor_id() != 0);
295 native_smp_prepare_boot_cpu();
297 if (xen_pv_domain()) {
298 if (!xen_feature(XENFEAT_writable_page_tables))
299 /* We've switched to the "real" per-cpu gdt, so make
300 * sure the old memory can be recycled. */
301 make_lowmem_page_readwrite(xen_initial_gdt);
305 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
308 loadsegment(ds, __USER_DS);
309 loadsegment(es, __USER_DS);
312 xen_filter_cpu_maps();
313 xen_setup_vcpu_info_placement();
317 * Setup vcpu_info for boot CPU.
319 if (xen_hvm_domain())
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_smp_prepare_cpus(unsigned int max_cpus)
336 if (skip_ioapic_setup) {
337 char *m = (max_cpus == 0) ?
338 "The nosmp parameter is incompatible with Xen; " \
339 "use Xen dom0_max_vcpus=1 parameter" :
340 "The noapic parameter is incompatible with Xen";
345 xen_init_lock_cpu(0);
347 smp_store_boot_cpu_info();
348 cpu_data(0).x86_max_cores = 1;
350 for_each_possible_cpu(i) {
351 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
352 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
353 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
355 set_cpu_sibling_map(0);
359 if (xen_smp_intr_init(0))
362 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
363 panic("could not allocate xen_cpu_initialized_map\n");
365 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
367 /* Restrict the possible_map according to max_cpus. */
368 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
369 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
371 set_cpu_possible(cpu, false);
374 for_each_possible_cpu(cpu)
375 set_cpu_present(cpu, true);
379 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
381 struct vcpu_guest_context *ctxt;
382 struct desc_struct *gdt;
383 unsigned long gdt_mfn;
385 /* used to tell cpu_init() that it can proceed with initialization */
386 cpumask_set_cpu(cpu, cpu_callout_mask);
387 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
390 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
394 gdt = get_cpu_gdt_table(cpu);
397 ctxt->user_regs.fs = __KERNEL_PERCPU;
398 ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
400 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
402 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
403 ctxt->flags = VGCF_IN_KERNEL;
404 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
405 ctxt->user_regs.ds = __USER_DS;
406 ctxt->user_regs.es = __USER_DS;
407 ctxt->user_regs.ss = __KERNEL_DS;
409 xen_copy_trap_info(ctxt->trap_ctxt);
413 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
415 gdt_mfn = arbitrary_virt_to_mfn(gdt);
416 make_lowmem_page_readonly(gdt);
417 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
419 ctxt->gdt_frames[0] = gdt_mfn;
420 ctxt->gdt_ents = GDT_ENTRIES;
422 ctxt->kernel_ss = __KERNEL_DS;
423 ctxt->kernel_sp = idle->thread.sp0;
426 ctxt->event_callback_cs = __KERNEL_CS;
427 ctxt->failsafe_callback_cs = __KERNEL_CS;
429 ctxt->gs_base_kernel = per_cpu_offset(cpu);
431 ctxt->event_callback_eip =
432 (unsigned long)xen_hypervisor_callback;
433 ctxt->failsafe_callback_eip =
434 (unsigned long)xen_failsafe_callback;
435 ctxt->user_regs.cs = __KERNEL_CS;
436 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
438 ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
439 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
440 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
447 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
451 common_cpu_up(cpu, idle);
453 xen_setup_runstate_info(cpu);
456 * PV VCPUs are always successfully taken down (see 'while' loop
457 * in xen_cpu_die()), so -EBUSY is an error.
459 rc = cpu_check_up_prepare(cpu);
463 /* make sure interrupts start blocked */
464 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
466 rc = cpu_initialize_context(cpu, idle);
472 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
475 while (cpu_report_state(cpu) != CPU_ONLINE)
476 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
481 static void xen_smp_cpus_done(unsigned int max_cpus)
485 #ifdef CONFIG_HOTPLUG_CPU
486 static int xen_cpu_disable(void)
488 unsigned int cpu = smp_processor_id();
492 cpu_disable_common();
494 load_cr3(swapper_pg_dir);
498 static void xen_cpu_die(unsigned int cpu)
500 while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up,
501 xen_vcpu_nr(cpu), NULL)) {
502 __set_current_state(TASK_UNINTERRUPTIBLE);
503 schedule_timeout(HZ/10);
506 if (common_cpu_die(cpu) == 0) {
507 xen_smp_intr_free(cpu);
508 xen_uninit_lock_cpu(cpu);
509 xen_teardown_timer(cpu);
514 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
517 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
520 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
521 * clears certain data that the cpu_idle loop (which called us
522 * and that we return from) expects. The only way to get that
523 * data back is to call:
525 tick_nohz_idle_enter();
527 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
530 #else /* !CONFIG_HOTPLUG_CPU */
531 static int xen_cpu_disable(void)
536 static void xen_cpu_die(unsigned int cpu)
541 static void xen_play_dead(void)
547 static void stop_self(void *v)
549 int cpu = smp_processor_id();
551 /* make sure we're not pinning something down */
552 load_cr3(swapper_pg_dir);
553 /* should set up a minimal gdt */
555 set_cpu_online(cpu, false);
557 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
561 static void xen_stop_other_cpus(int wait)
563 smp_call_function(stop_self, NULL, wait);
566 static void xen_smp_send_reschedule(int cpu)
568 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
571 static void __xen_send_IPI_mask(const struct cpumask *mask,
576 for_each_cpu_and(cpu, mask, cpu_online_mask)
577 xen_send_IPI_one(cpu, vector);
580 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
584 __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
586 /* Make sure other vcpus get a chance to run if they need to. */
587 for_each_cpu(cpu, mask) {
588 if (xen_vcpu_stolen(cpu)) {
589 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
595 static void xen_smp_send_call_function_single_ipi(int cpu)
597 __xen_send_IPI_mask(cpumask_of(cpu),
598 XEN_CALL_FUNCTION_SINGLE_VECTOR);
601 static inline int xen_map_vector(int vector)
606 case RESCHEDULE_VECTOR:
607 xen_vector = XEN_RESCHEDULE_VECTOR;
609 case CALL_FUNCTION_VECTOR:
610 xen_vector = XEN_CALL_FUNCTION_VECTOR;
612 case CALL_FUNCTION_SINGLE_VECTOR:
613 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
615 case IRQ_WORK_VECTOR:
616 xen_vector = XEN_IRQ_WORK_VECTOR;
620 case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
621 xen_vector = XEN_NMI_VECTOR;
626 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
633 void xen_send_IPI_mask(const struct cpumask *mask,
636 int xen_vector = xen_map_vector(vector);
639 __xen_send_IPI_mask(mask, xen_vector);
642 void xen_send_IPI_all(int vector)
644 int xen_vector = xen_map_vector(vector);
647 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
650 void xen_send_IPI_self(int vector)
652 int xen_vector = xen_map_vector(vector);
655 xen_send_IPI_one(smp_processor_id(), xen_vector);
658 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
662 unsigned int this_cpu = smp_processor_id();
663 int xen_vector = xen_map_vector(vector);
665 if (!(num_online_cpus() > 1) || (xen_vector < 0))
668 for_each_cpu_and(cpu, mask, cpu_online_mask) {
672 xen_send_IPI_one(cpu, xen_vector);
676 void xen_send_IPI_allbutself(int vector)
678 xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
681 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
684 generic_smp_call_function_interrupt();
685 inc_irq_stat(irq_call_count);
691 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
694 generic_smp_call_function_single_interrupt();
695 inc_irq_stat(irq_call_count);
701 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
705 inc_irq_stat(apic_irq_work_irqs);
711 static const struct smp_ops xen_smp_ops __initconst = {
712 .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
713 .smp_prepare_cpus = xen_smp_prepare_cpus,
714 .smp_cpus_done = xen_smp_cpus_done,
716 .cpu_up = xen_cpu_up,
717 .cpu_die = xen_cpu_die,
718 .cpu_disable = xen_cpu_disable,
719 .play_dead = xen_play_dead,
721 .stop_other_cpus = xen_stop_other_cpus,
722 .smp_send_reschedule = xen_smp_send_reschedule,
724 .send_call_func_ipi = xen_smp_send_call_function_ipi,
725 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
728 void __init xen_smp_init(void)
730 smp_ops = xen_smp_ops;
731 xen_fill_possible_map();
734 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
736 native_smp_prepare_cpus(max_cpus);
737 WARN_ON(xen_smp_intr_init(0));
739 xen_init_lock_cpu(0);
742 void __init xen_hvm_smp_init(void)
744 smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
745 smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
746 smp_ops.cpu_die = xen_cpu_die;
747 smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
748 smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
749 smp_ops.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu;