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/spec-ctrl.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 speculative_store_bypass_ht_init();
93 xen_setup_cpu_clockevents();
95 notify_cpu_starting(cpu);
97 set_cpu_online(cpu, true);
99 cpu_set_state_online(cpu); /* Implies full memory barrier. */
101 /* We can take interrupts now: we're officially "up". */
106 * Note: cpu parameter is only relevant for PVH. The reason for passing it
107 * is we can't do smp_processor_id until the percpu segments are loaded, for
108 * which we need the cpu number! So we pass it in rdi as first parameter.
110 asmlinkage __visible void cpu_bringup_and_idle(int cpu)
112 #ifdef CONFIG_XEN_PVH
113 if (xen_feature(XENFEAT_auto_translated_physmap) &&
114 xen_feature(XENFEAT_supervisor_mode_kernel))
115 xen_pvh_secondary_vcpu_init(cpu);
118 cpu_startup_entry(CPUHP_ONLINE);
119 prevent_tail_call_optimization();
122 static void xen_smp_intr_free(unsigned int cpu)
124 if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
125 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
126 per_cpu(xen_resched_irq, cpu).irq = -1;
127 kfree(per_cpu(xen_resched_irq, cpu).name);
128 per_cpu(xen_resched_irq, cpu).name = NULL;
130 if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
131 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
132 per_cpu(xen_callfunc_irq, cpu).irq = -1;
133 kfree(per_cpu(xen_callfunc_irq, cpu).name);
134 per_cpu(xen_callfunc_irq, cpu).name = NULL;
136 if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
137 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
138 per_cpu(xen_debug_irq, cpu).irq = -1;
139 kfree(per_cpu(xen_debug_irq, cpu).name);
140 per_cpu(xen_debug_irq, cpu).name = NULL;
142 if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
143 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
145 per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
146 kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
147 per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
149 if (xen_hvm_domain())
152 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
153 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
154 per_cpu(xen_irq_work, cpu).irq = -1;
155 kfree(per_cpu(xen_irq_work, cpu).name);
156 per_cpu(xen_irq_work, cpu).name = NULL;
159 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
160 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
161 per_cpu(xen_pmu_irq, cpu).irq = -1;
162 kfree(per_cpu(xen_pmu_irq, cpu).name);
163 per_cpu(xen_pmu_irq, cpu).name = NULL;
166 static int xen_smp_intr_init(unsigned int cpu)
169 char *resched_name, *callfunc_name, *debug_name, *pmu_name;
171 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
172 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
174 xen_reschedule_interrupt,
175 IRQF_PERCPU|IRQF_NOBALANCING,
180 per_cpu(xen_resched_irq, cpu).irq = rc;
181 per_cpu(xen_resched_irq, cpu).name = resched_name;
183 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
184 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
186 xen_call_function_interrupt,
187 IRQF_PERCPU|IRQF_NOBALANCING,
192 per_cpu(xen_callfunc_irq, cpu).irq = rc;
193 per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
195 debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
196 rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
197 IRQF_PERCPU | IRQF_NOBALANCING,
201 per_cpu(xen_debug_irq, cpu).irq = rc;
202 per_cpu(xen_debug_irq, cpu).name = debug_name;
204 callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
205 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
207 xen_call_function_single_interrupt,
208 IRQF_PERCPU|IRQF_NOBALANCING,
213 per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
214 per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
217 * The IRQ worker on PVHVM goes through the native path and uses the
220 if (xen_hvm_domain())
223 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
224 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
226 xen_irq_work_interrupt,
227 IRQF_PERCPU|IRQF_NOBALANCING,
232 per_cpu(xen_irq_work, cpu).irq = rc;
233 per_cpu(xen_irq_work, cpu).name = callfunc_name;
235 if (is_xen_pmu(cpu)) {
236 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
237 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
239 IRQF_PERCPU|IRQF_NOBALANCING,
243 per_cpu(xen_pmu_irq, cpu).irq = rc;
244 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
250 xen_smp_intr_free(cpu);
254 static void __init xen_fill_possible_map(void)
258 if (xen_initial_domain())
261 for (i = 0; i < nr_cpu_ids; i++) {
262 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
265 set_cpu_possible(i, true);
270 static void __init xen_filter_cpu_maps(void)
273 unsigned int subtract = 0;
275 if (!xen_initial_domain())
280 for (i = 0; i < nr_cpu_ids; i++) {
281 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
284 set_cpu_possible(i, true);
286 set_cpu_possible(i, false);
287 set_cpu_present(i, false);
291 #ifdef CONFIG_HOTPLUG_CPU
292 /* This is akin to using 'nr_cpus' on the Linux command line.
293 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
294 * have up to X, while nr_cpu_ids is greater than X. This
295 * normally is not a problem, except when CPU hotplugging
296 * is involved and then there might be more than X CPUs
297 * in the guest - which will not work as there is no
298 * hypercall to expand the max number of VCPUs an already
299 * running guest has. So cap it up to X. */
301 nr_cpu_ids = nr_cpu_ids - subtract;
306 static void __init xen_smp_prepare_boot_cpu(void)
308 BUG_ON(smp_processor_id() != 0);
309 native_smp_prepare_boot_cpu();
311 if (xen_pv_domain()) {
312 if (!xen_feature(XENFEAT_writable_page_tables))
313 /* We've switched to the "real" per-cpu gdt, so make
314 * sure the old memory can be recycled. */
315 make_lowmem_page_readwrite(xen_initial_gdt);
319 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
322 loadsegment(ds, __USER_DS);
323 loadsegment(es, __USER_DS);
326 xen_filter_cpu_maps();
327 xen_setup_vcpu_info_placement();
330 * The alternative logic (which patches the unlock/lock) runs before
331 * the smp bootup up code is activated. Hence we need to set this up
332 * the core kernel is being patched. Otherwise we will have only
333 * modules patched but not core code.
335 xen_init_spinlocks();
338 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
343 if (skip_ioapic_setup) {
344 char *m = (max_cpus == 0) ?
345 "The nosmp parameter is incompatible with Xen; " \
346 "use Xen dom0_max_vcpus=1 parameter" :
347 "The noapic parameter is incompatible with Xen";
352 xen_init_lock_cpu(0);
354 smp_store_boot_cpu_info();
355 cpu_data(0).x86_max_cores = 1;
357 for_each_possible_cpu(i) {
358 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
359 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
360 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
362 set_cpu_sibling_map(0);
364 speculative_store_bypass_ht_init();
368 if (xen_smp_intr_init(0))
371 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
372 panic("could not allocate xen_cpu_initialized_map\n");
374 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
376 /* Restrict the possible_map according to max_cpus. */
377 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
378 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
380 set_cpu_possible(cpu, false);
383 for_each_possible_cpu(cpu)
384 set_cpu_present(cpu, true);
388 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
390 struct vcpu_guest_context *ctxt;
391 struct desc_struct *gdt;
392 unsigned long gdt_mfn;
394 /* used to tell cpu_init() that it can proceed with initialization */
395 cpumask_set_cpu(cpu, cpu_callout_mask);
396 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
399 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
403 gdt = get_cpu_gdt_table(cpu);
406 /* Note: PVH is not yet supported on x86_32. */
407 ctxt->user_regs.fs = __KERNEL_PERCPU;
408 ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
410 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
412 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
413 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
414 ctxt->flags = VGCF_IN_KERNEL;
415 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
416 ctxt->user_regs.ds = __USER_DS;
417 ctxt->user_regs.es = __USER_DS;
418 ctxt->user_regs.ss = __KERNEL_DS;
420 xen_copy_trap_info(ctxt->trap_ctxt);
424 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
426 gdt_mfn = arbitrary_virt_to_mfn(gdt);
427 make_lowmem_page_readonly(gdt);
428 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
430 ctxt->gdt_frames[0] = gdt_mfn;
431 ctxt->gdt_ents = GDT_ENTRIES;
433 ctxt->kernel_ss = __KERNEL_DS;
434 ctxt->kernel_sp = idle->thread.sp0;
437 ctxt->event_callback_cs = __KERNEL_CS;
438 ctxt->failsafe_callback_cs = __KERNEL_CS;
440 ctxt->gs_base_kernel = per_cpu_offset(cpu);
442 ctxt->event_callback_eip =
443 (unsigned long)xen_hypervisor_callback;
444 ctxt->failsafe_callback_eip =
445 (unsigned long)xen_failsafe_callback;
446 ctxt->user_regs.cs = __KERNEL_CS;
447 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
449 #ifdef CONFIG_XEN_PVH
452 * The vcpu comes on kernel page tables which have the NX pte
453 * bit set. This means before DS/SS is touched, NX in
454 * EFER must be set. Hence the following assembly glue code.
456 ctxt->user_regs.eip = (unsigned long)xen_pvh_early_cpu_init;
457 ctxt->user_regs.rdi = cpu;
458 ctxt->user_regs.rsi = true; /* entry == true */
461 ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
462 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
463 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
470 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
474 common_cpu_up(cpu, idle);
476 xen_setup_runstate_info(cpu);
477 xen_setup_timer(cpu);
478 xen_init_lock_cpu(cpu);
481 * PV VCPUs are always successfully taken down (see 'while' loop
482 * in xen_cpu_die()), so -EBUSY is an error.
484 rc = cpu_check_up_prepare(cpu);
488 /* make sure interrupts start blocked */
489 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
491 rc = cpu_initialize_context(cpu, idle);
497 rc = xen_smp_intr_init(cpu);
501 rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
504 while (cpu_report_state(cpu) != CPU_ONLINE)
505 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
510 static void xen_smp_cpus_done(unsigned int max_cpus)
514 #ifdef CONFIG_HOTPLUG_CPU
515 static int xen_cpu_disable(void)
517 unsigned int cpu = smp_processor_id();
521 cpu_disable_common();
523 load_cr3(swapper_pg_dir);
527 static void xen_cpu_die(unsigned int cpu)
529 while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
530 __set_current_state(TASK_UNINTERRUPTIBLE);
531 schedule_timeout(HZ/10);
534 if (common_cpu_die(cpu) == 0) {
535 xen_smp_intr_free(cpu);
536 xen_uninit_lock_cpu(cpu);
537 xen_teardown_timer(cpu);
542 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
545 HYPERVISOR_vcpu_op(VCPUOP_down, 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();
556 #else /* !CONFIG_HOTPLUG_CPU */
557 static int xen_cpu_disable(void)
562 static void xen_cpu_die(unsigned int cpu)
567 static void xen_play_dead(void)
573 static void stop_self(void *v)
575 int cpu = smp_processor_id();
577 /* make sure we're not pinning something down */
578 load_cr3(swapper_pg_dir);
579 /* should set up a minimal gdt */
581 set_cpu_online(cpu, false);
583 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
587 static void xen_stop_other_cpus(int wait)
589 smp_call_function(stop_self, NULL, wait);
592 static void xen_smp_send_reschedule(int cpu)
594 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
597 static void __xen_send_IPI_mask(const struct cpumask *mask,
602 for_each_cpu_and(cpu, mask, cpu_online_mask)
603 xen_send_IPI_one(cpu, vector);
606 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
610 __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
612 /* Make sure other vcpus get a chance to run if they need to. */
613 for_each_cpu(cpu, mask) {
614 if (xen_vcpu_stolen(cpu)) {
615 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
621 static void xen_smp_send_call_function_single_ipi(int cpu)
623 __xen_send_IPI_mask(cpumask_of(cpu),
624 XEN_CALL_FUNCTION_SINGLE_VECTOR);
627 static inline int xen_map_vector(int vector)
632 case RESCHEDULE_VECTOR:
633 xen_vector = XEN_RESCHEDULE_VECTOR;
635 case CALL_FUNCTION_VECTOR:
636 xen_vector = XEN_CALL_FUNCTION_VECTOR;
638 case CALL_FUNCTION_SINGLE_VECTOR:
639 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
641 case IRQ_WORK_VECTOR:
642 xen_vector = XEN_IRQ_WORK_VECTOR;
646 case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
647 xen_vector = XEN_NMI_VECTOR;
652 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
659 void xen_send_IPI_mask(const struct cpumask *mask,
662 int xen_vector = xen_map_vector(vector);
665 __xen_send_IPI_mask(mask, xen_vector);
668 void xen_send_IPI_all(int vector)
670 int xen_vector = xen_map_vector(vector);
673 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
676 void xen_send_IPI_self(int vector)
678 int xen_vector = xen_map_vector(vector);
681 xen_send_IPI_one(smp_processor_id(), xen_vector);
684 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
688 unsigned int this_cpu = smp_processor_id();
689 int xen_vector = xen_map_vector(vector);
691 if (!(num_online_cpus() > 1) || (xen_vector < 0))
694 for_each_cpu_and(cpu, mask, cpu_online_mask) {
698 xen_send_IPI_one(cpu, xen_vector);
702 void xen_send_IPI_allbutself(int vector)
704 xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
707 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
710 generic_smp_call_function_interrupt();
711 inc_irq_stat(irq_call_count);
717 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
720 generic_smp_call_function_single_interrupt();
721 inc_irq_stat(irq_call_count);
727 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
731 inc_irq_stat(apic_irq_work_irqs);
737 static const struct smp_ops xen_smp_ops __initconst = {
738 .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
739 .smp_prepare_cpus = xen_smp_prepare_cpus,
740 .smp_cpus_done = xen_smp_cpus_done,
742 .cpu_up = xen_cpu_up,
743 .cpu_die = xen_cpu_die,
744 .cpu_disable = xen_cpu_disable,
745 .play_dead = xen_play_dead,
747 .stop_other_cpus = xen_stop_other_cpus,
748 .smp_send_reschedule = xen_smp_send_reschedule,
750 .send_call_func_ipi = xen_smp_send_call_function_ipi,
751 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
754 void __init xen_smp_init(void)
756 smp_ops = xen_smp_ops;
757 xen_fill_possible_map();
760 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
762 native_smp_prepare_cpus(max_cpus);
763 WARN_ON(xen_smp_intr_init(0));
765 xen_init_lock_cpu(0);
768 static int xen_hvm_cpu_up(unsigned int cpu, struct task_struct *tidle)
773 * This can happen if CPU was offlined earlier and
774 * offlining timed out in common_cpu_die().
776 if (cpu_report_state(cpu) == CPU_DEAD_FROZEN) {
777 xen_smp_intr_free(cpu);
778 xen_uninit_lock_cpu(cpu);
782 * xen_smp_intr_init() needs to run before native_cpu_up()
783 * so that IPI vectors are set up on the booting CPU before
784 * it is marked online in native_cpu_up().
786 rc = xen_smp_intr_init(cpu);
789 rc = native_cpu_up(cpu, tidle);
792 * We must initialize the slowpath CPU kicker _after_ the native
793 * path has executed. If we initialized it before none of the
794 * unlocker IPI kicks would reach the booting CPU as the booting
795 * CPU had not set itself 'online' in cpu_online_mask. That mask
796 * is checked when IPIs are sent (on HVM at least).
798 xen_init_lock_cpu(cpu);
802 void __init xen_hvm_smp_init(void)
804 if (!xen_have_vector_callback)
806 smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
807 smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
808 smp_ops.cpu_up = xen_hvm_cpu_up;
809 smp_ops.cpu_die = xen_cpu_die;
810 smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
811 smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
812 smp_ops.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu;