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". */
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.
107 asmlinkage __visible void cpu_bringup_and_idle(int cpu)
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);
115 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
118 void xen_smp_intr_free(unsigned int cpu)
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;
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;
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;
138 if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
139 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
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;
145 if (xen_hvm_domain())
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;
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;
162 int xen_smp_intr_init(unsigned int cpu)
165 char *resched_name, *callfunc_name, *debug_name, *pmu_name;
167 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
168 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
170 xen_reschedule_interrupt,
171 IRQF_PERCPU|IRQF_NOBALANCING,
176 per_cpu(xen_resched_irq, cpu).irq = rc;
177 per_cpu(xen_resched_irq, cpu).name = resched_name;
179 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
180 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
182 xen_call_function_interrupt,
183 IRQF_PERCPU|IRQF_NOBALANCING,
188 per_cpu(xen_callfunc_irq, cpu).irq = rc;
189 per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
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,
197 per_cpu(xen_debug_irq, cpu).irq = rc;
198 per_cpu(xen_debug_irq, cpu).name = debug_name;
200 callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
201 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
203 xen_call_function_single_interrupt,
204 IRQF_PERCPU|IRQF_NOBALANCING,
209 per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
210 per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
213 * The IRQ worker on PVHVM goes through the native path and uses the
216 if (xen_hvm_domain())
219 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
220 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
222 xen_irq_work_interrupt,
223 IRQF_PERCPU|IRQF_NOBALANCING,
228 per_cpu(xen_irq_work, cpu).irq = rc;
229 per_cpu(xen_irq_work, cpu).name = callfunc_name;
231 if (is_xen_pmu(cpu)) {
232 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
233 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
235 IRQF_PERCPU|IRQF_NOBALANCING,
239 per_cpu(xen_pmu_irq, cpu).irq = rc;
240 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
246 xen_smp_intr_free(cpu);
250 static void __init xen_fill_possible_map(void)
254 if (xen_initial_domain())
257 for (i = 0; i < nr_cpu_ids; i++) {
258 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
261 set_cpu_possible(i, true);
266 static void __init xen_filter_cpu_maps(void)
269 unsigned int subtract = 0;
271 if (!xen_initial_domain())
276 for (i = 0; i < nr_cpu_ids; i++) {
277 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
280 set_cpu_possible(i, true);
282 set_cpu_possible(i, false);
283 set_cpu_present(i, false);
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. */
297 nr_cpu_ids = nr_cpu_ids - subtract;
302 static void __init xen_smp_prepare_boot_cpu(void)
304 BUG_ON(smp_processor_id() != 0);
305 native_smp_prepare_boot_cpu();
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);
315 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
318 loadsegment(ds, __USER_DS);
319 loadsegment(es, __USER_DS);
322 xen_filter_cpu_maps();
323 xen_setup_vcpu_info_placement();
327 * Setup vcpu_info for boot CPU.
329 if (xen_hvm_domain())
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.
338 xen_init_spinlocks();
341 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
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";
355 xen_init_lock_cpu(0);
357 smp_store_boot_cpu_info();
358 cpu_data(0).x86_max_cores = 1;
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);
365 set_cpu_sibling_map(0);
369 if (xen_smp_intr_init(0))
372 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
373 panic("could not allocate xen_cpu_initialized_map\n");
375 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
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--)
381 set_cpu_possible(cpu, false);
384 for_each_possible_cpu(cpu)
385 set_cpu_present(cpu, true);
389 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
391 struct vcpu_guest_context *ctxt;
392 struct desc_struct *gdt;
393 unsigned long gdt_mfn;
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))
400 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
404 gdt = get_cpu_gdt_table(cpu);
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;
411 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
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;
421 xen_copy_trap_info(ctxt->trap_ctxt);
425 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
427 gdt_mfn = arbitrary_virt_to_mfn(gdt);
428 make_lowmem_page_readonly(gdt);
429 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
431 ctxt->gdt_frames[0] = gdt_mfn;
432 ctxt->gdt_ents = GDT_ENTRIES;
434 ctxt->kernel_ss = __KERNEL_DS;
435 ctxt->kernel_sp = idle->thread.sp0;
438 ctxt->event_callback_cs = __KERNEL_CS;
439 ctxt->failsafe_callback_cs = __KERNEL_CS;
441 ctxt->gs_base_kernel = per_cpu_offset(cpu);
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);
450 #ifdef CONFIG_XEN_PVH
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.
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 */
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))
471 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
475 common_cpu_up(cpu, idle);
477 xen_setup_runstate_info(cpu);
480 * PV VCPUs are always successfully taken down (see 'while' loop
481 * in xen_cpu_die()), so -EBUSY is an error.
483 rc = cpu_check_up_prepare(cpu);
487 /* make sure interrupts start blocked */
488 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
490 rc = cpu_initialize_context(cpu, idle);
496 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
499 while (cpu_report_state(cpu) != CPU_ONLINE)
500 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
505 static void xen_smp_cpus_done(unsigned int max_cpus)
509 #ifdef CONFIG_HOTPLUG_CPU
510 static int xen_cpu_disable(void)
512 unsigned int cpu = smp_processor_id();
516 cpu_disable_common();
518 load_cr3(swapper_pg_dir);
522 static void xen_cpu_die(unsigned int cpu)
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);
530 if (common_cpu_die(cpu) == 0) {
531 xen_smp_intr_free(cpu);
532 xen_uninit_lock_cpu(cpu);
533 xen_teardown_timer(cpu);
538 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
541 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
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:
549 tick_nohz_idle_enter();
551 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
554 #else /* !CONFIG_HOTPLUG_CPU */
555 static int xen_cpu_disable(void)
560 static void xen_cpu_die(unsigned int cpu)
565 static void xen_play_dead(void)
571 static void stop_self(void *v)
573 int cpu = smp_processor_id();
575 /* make sure we're not pinning something down */
576 load_cr3(swapper_pg_dir);
577 /* should set up a minimal gdt */
579 set_cpu_online(cpu, false);
581 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
585 static void xen_stop_other_cpus(int wait)
587 smp_call_function(stop_self, NULL, wait);
590 static void xen_smp_send_reschedule(int cpu)
592 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
595 static void __xen_send_IPI_mask(const struct cpumask *mask,
600 for_each_cpu_and(cpu, mask, cpu_online_mask)
601 xen_send_IPI_one(cpu, vector);
604 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
608 __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
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);
619 static void xen_smp_send_call_function_single_ipi(int cpu)
621 __xen_send_IPI_mask(cpumask_of(cpu),
622 XEN_CALL_FUNCTION_SINGLE_VECTOR);
625 static inline int xen_map_vector(int vector)
630 case RESCHEDULE_VECTOR:
631 xen_vector = XEN_RESCHEDULE_VECTOR;
633 case CALL_FUNCTION_VECTOR:
634 xen_vector = XEN_CALL_FUNCTION_VECTOR;
636 case CALL_FUNCTION_SINGLE_VECTOR:
637 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
639 case IRQ_WORK_VECTOR:
640 xen_vector = XEN_IRQ_WORK_VECTOR;
644 case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
645 xen_vector = XEN_NMI_VECTOR;
650 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
657 void xen_send_IPI_mask(const struct cpumask *mask,
660 int xen_vector = xen_map_vector(vector);
663 __xen_send_IPI_mask(mask, xen_vector);
666 void xen_send_IPI_all(int vector)
668 int xen_vector = xen_map_vector(vector);
671 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
674 void xen_send_IPI_self(int vector)
676 int xen_vector = xen_map_vector(vector);
679 xen_send_IPI_one(smp_processor_id(), xen_vector);
682 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
686 unsigned int this_cpu = smp_processor_id();
687 int xen_vector = xen_map_vector(vector);
689 if (!(num_online_cpus() > 1) || (xen_vector < 0))
692 for_each_cpu_and(cpu, mask, cpu_online_mask) {
696 xen_send_IPI_one(cpu, xen_vector);
700 void xen_send_IPI_allbutself(int vector)
702 xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
705 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
708 generic_smp_call_function_interrupt();
709 inc_irq_stat(irq_call_count);
715 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
718 generic_smp_call_function_single_interrupt();
719 inc_irq_stat(irq_call_count);
725 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
729 inc_irq_stat(apic_irq_work_irqs);
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,
740 .cpu_up = xen_cpu_up,
741 .cpu_die = xen_cpu_die,
742 .cpu_disable = xen_cpu_disable,
743 .play_dead = xen_play_dead,
745 .stop_other_cpus = xen_stop_other_cpus,
746 .smp_send_reschedule = xen_smp_send_reschedule,
748 .send_call_func_ipi = xen_smp_send_call_function_ipi,
749 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
752 void __init xen_smp_init(void)
754 smp_ops = xen_smp_ops;
755 xen_fill_possible_map();
758 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
760 native_smp_prepare_cpus(max_cpus);
761 WARN_ON(xen_smp_intr_init(0));
763 xen_init_lock_cpu(0);
766 void __init xen_hvm_smp_init(void)
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;