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
;
136 if (xen_hvm_domain())
139 if (per_cpu(xen_irq_work
, cpu
).irq
>= 0) {
140 unbind_from_irqhandler(per_cpu(xen_irq_work
, cpu
).irq
, NULL
);
141 per_cpu(xen_irq_work
, cpu
).irq
= -1;
142 kfree(per_cpu(xen_irq_work
, cpu
).name
);
143 per_cpu(xen_irq_work
, cpu
).name
= NULL
;
146 if (per_cpu(xen_pmu_irq
, cpu
).irq
>= 0) {
147 unbind_from_irqhandler(per_cpu(xen_pmu_irq
, cpu
).irq
, NULL
);
148 per_cpu(xen_pmu_irq
, cpu
).irq
= -1;
149 kfree(per_cpu(xen_pmu_irq
, cpu
).name
);
150 per_cpu(xen_pmu_irq
, cpu
).name
= NULL
;
153 int xen_smp_intr_init(unsigned int cpu
)
156 char *resched_name
, *callfunc_name
, *debug_name
, *pmu_name
;
158 resched_name
= kasprintf(GFP_KERNEL
, "resched%d", cpu
);
159 rc
= bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR
,
161 xen_reschedule_interrupt
,
162 IRQF_PERCPU
|IRQF_NOBALANCING
,
167 per_cpu(xen_resched_irq
, cpu
).irq
= rc
;
168 per_cpu(xen_resched_irq
, cpu
).name
= resched_name
;
170 callfunc_name
= kasprintf(GFP_KERNEL
, "callfunc%d", cpu
);
171 rc
= bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR
,
173 xen_call_function_interrupt
,
174 IRQF_PERCPU
|IRQF_NOBALANCING
,
179 per_cpu(xen_callfunc_irq
, cpu
).irq
= rc
;
180 per_cpu(xen_callfunc_irq
, cpu
).name
= callfunc_name
;
182 debug_name
= kasprintf(GFP_KERNEL
, "debug%d", cpu
);
183 rc
= bind_virq_to_irqhandler(VIRQ_DEBUG
, cpu
, xen_debug_interrupt
,
184 IRQF_PERCPU
| IRQF_NOBALANCING
,
188 per_cpu(xen_debug_irq
, cpu
).irq
= rc
;
189 per_cpu(xen_debug_irq
, cpu
).name
= debug_name
;
191 callfunc_name
= kasprintf(GFP_KERNEL
, "callfuncsingle%d", cpu
);
192 rc
= bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR
,
194 xen_call_function_single_interrupt
,
195 IRQF_PERCPU
|IRQF_NOBALANCING
,
200 per_cpu(xen_callfuncsingle_irq
, cpu
).irq
= rc
;
201 per_cpu(xen_callfuncsingle_irq
, cpu
).name
= callfunc_name
;
204 * The IRQ worker on PVHVM goes through the native path and uses the
207 if (xen_hvm_domain())
210 callfunc_name
= kasprintf(GFP_KERNEL
, "irqwork%d", cpu
);
211 rc
= bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR
,
213 xen_irq_work_interrupt
,
214 IRQF_PERCPU
|IRQF_NOBALANCING
,
219 per_cpu(xen_irq_work
, cpu
).irq
= rc
;
220 per_cpu(xen_irq_work
, cpu
).name
= callfunc_name
;
222 if (is_xen_pmu(cpu
)) {
223 pmu_name
= kasprintf(GFP_KERNEL
, "pmu%d", cpu
);
224 rc
= bind_virq_to_irqhandler(VIRQ_XENPMU
, cpu
,
226 IRQF_PERCPU
|IRQF_NOBALANCING
,
230 per_cpu(xen_pmu_irq
, cpu
).irq
= rc
;
231 per_cpu(xen_pmu_irq
, cpu
).name
= pmu_name
;
237 xen_smp_intr_free(cpu
);
241 static void __init
xen_fill_possible_map(void)
245 if (xen_initial_domain())
248 for (i
= 0; i
< nr_cpu_ids
; i
++) {
249 rc
= HYPERVISOR_vcpu_op(VCPUOP_is_up
, i
, NULL
);
252 set_cpu_possible(i
, true);
257 static void __init
xen_filter_cpu_maps(void)
260 unsigned int subtract
= 0;
262 if (!xen_initial_domain())
267 for (i
= 0; i
< nr_cpu_ids
; i
++) {
268 rc
= HYPERVISOR_vcpu_op(VCPUOP_is_up
, i
, NULL
);
271 set_cpu_possible(i
, true);
273 set_cpu_possible(i
, false);
274 set_cpu_present(i
, false);
278 #ifdef CONFIG_HOTPLUG_CPU
279 /* This is akin to using 'nr_cpus' on the Linux command line.
280 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
281 * have up to X, while nr_cpu_ids is greater than X. This
282 * normally is not a problem, except when CPU hotplugging
283 * is involved and then there might be more than X CPUs
284 * in the guest - which will not work as there is no
285 * hypercall to expand the max number of VCPUs an already
286 * running guest has. So cap it up to X. */
288 nr_cpu_ids
= nr_cpu_ids
- subtract
;
293 static void __init
xen_smp_prepare_boot_cpu(void)
295 BUG_ON(smp_processor_id() != 0);
296 native_smp_prepare_boot_cpu();
298 if (xen_pv_domain()) {
299 if (!xen_feature(XENFEAT_writable_page_tables
))
300 /* We've switched to the "real" per-cpu gdt, so make
301 * sure the old memory can be recycled. */
302 make_lowmem_page_readwrite(xen_initial_gdt
);
306 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
309 loadsegment(ds
, __USER_DS
);
310 loadsegment(es
, __USER_DS
);
313 xen_filter_cpu_maps();
314 xen_setup_vcpu_info_placement();
318 * Setup vcpu_info for boot CPU.
320 if (xen_hvm_domain())
324 * The alternative logic (which patches the unlock/lock) runs before
325 * the smp bootup up code is activated. Hence we need to set this up
326 * the core kernel is being patched. Otherwise we will have only
327 * modules patched but not core code.
329 xen_init_spinlocks();
332 static void __init
xen_smp_prepare_cpus(unsigned int max_cpus
)
337 if (skip_ioapic_setup
) {
338 char *m
= (max_cpus
== 0) ?
339 "The nosmp parameter is incompatible with Xen; " \
340 "use Xen dom0_max_vcpus=1 parameter" :
341 "The noapic parameter is incompatible with Xen";
346 xen_init_lock_cpu(0);
348 smp_store_boot_cpu_info();
349 cpu_data(0).x86_max_cores
= 1;
351 for_each_possible_cpu(i
) {
352 zalloc_cpumask_var(&per_cpu(cpu_sibling_map
, i
), GFP_KERNEL
);
353 zalloc_cpumask_var(&per_cpu(cpu_core_map
, i
), GFP_KERNEL
);
354 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map
, i
), GFP_KERNEL
);
356 set_cpu_sibling_map(0);
360 if (xen_smp_intr_init(0))
363 if (!alloc_cpumask_var(&xen_cpu_initialized_map
, GFP_KERNEL
))
364 panic("could not allocate xen_cpu_initialized_map\n");
366 cpumask_copy(xen_cpu_initialized_map
, cpumask_of(0));
368 /* Restrict the possible_map according to max_cpus. */
369 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus
)) {
370 for (cpu
= nr_cpu_ids
- 1; !cpu_possible(cpu
); cpu
--)
372 set_cpu_possible(cpu
, false);
375 for_each_possible_cpu(cpu
)
376 set_cpu_present(cpu
, true);
380 cpu_initialize_context(unsigned int cpu
, struct task_struct
*idle
)
382 struct vcpu_guest_context
*ctxt
;
383 struct desc_struct
*gdt
;
384 unsigned long gdt_mfn
;
386 /* used to tell cpu_init() that it can proceed with initialization */
387 cpumask_set_cpu(cpu
, cpu_callout_mask
);
388 if (cpumask_test_and_set_cpu(cpu
, xen_cpu_initialized_map
))
391 ctxt
= kzalloc(sizeof(*ctxt
), GFP_KERNEL
);
395 gdt
= get_cpu_gdt_table(cpu
);
398 ctxt
->user_regs
.fs
= __KERNEL_PERCPU
;
399 ctxt
->user_regs
.gs
= __KERNEL_STACK_CANARY
;
401 memset(&ctxt
->fpu_ctxt
, 0, sizeof(ctxt
->fpu_ctxt
));
403 ctxt
->user_regs
.eip
= (unsigned long)cpu_bringup_and_idle
;
404 ctxt
->flags
= VGCF_IN_KERNEL
;
405 ctxt
->user_regs
.eflags
= 0x1000; /* IOPL_RING1 */
406 ctxt
->user_regs
.ds
= __USER_DS
;
407 ctxt
->user_regs
.es
= __USER_DS
;
408 ctxt
->user_regs
.ss
= __KERNEL_DS
;
410 xen_copy_trap_info(ctxt
->trap_ctxt
);
414 BUG_ON((unsigned long)gdt
& ~PAGE_MASK
);
416 gdt_mfn
= arbitrary_virt_to_mfn(gdt
);
417 make_lowmem_page_readonly(gdt
);
418 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn
));
420 ctxt
->gdt_frames
[0] = gdt_mfn
;
421 ctxt
->gdt_ents
= GDT_ENTRIES
;
423 ctxt
->kernel_ss
= __KERNEL_DS
;
424 ctxt
->kernel_sp
= idle
->thread
.sp0
;
427 ctxt
->event_callback_cs
= __KERNEL_CS
;
428 ctxt
->failsafe_callback_cs
= __KERNEL_CS
;
430 ctxt
->gs_base_kernel
= per_cpu_offset(cpu
);
432 ctxt
->event_callback_eip
=
433 (unsigned long)xen_hypervisor_callback
;
434 ctxt
->failsafe_callback_eip
=
435 (unsigned long)xen_failsafe_callback
;
436 ctxt
->user_regs
.cs
= __KERNEL_CS
;
437 per_cpu(xen_cr3
, cpu
) = __pa(swapper_pg_dir
);
439 ctxt
->user_regs
.esp
= idle
->thread
.sp0
- sizeof(struct pt_regs
);
440 ctxt
->ctrlreg
[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir
));
441 if (HYPERVISOR_vcpu_op(VCPUOP_initialise
, xen_vcpu_nr(cpu
), ctxt
))
448 static int xen_cpu_up(unsigned int cpu
, struct task_struct
*idle
)
452 common_cpu_up(cpu
, idle
);
454 xen_setup_runstate_info(cpu
);
457 * PV VCPUs are always successfully taken down (see 'while' loop
458 * in xen_cpu_die()), so -EBUSY is an error.
460 rc
= cpu_check_up_prepare(cpu
);
464 /* make sure interrupts start blocked */
465 per_cpu(xen_vcpu
, cpu
)->evtchn_upcall_mask
= 1;
467 rc
= cpu_initialize_context(cpu
, idle
);
473 rc
= HYPERVISOR_vcpu_op(VCPUOP_up
, xen_vcpu_nr(cpu
), NULL
);
476 while (cpu_report_state(cpu
) != CPU_ONLINE
)
477 HYPERVISOR_sched_op(SCHEDOP_yield
, NULL
);
482 static void xen_smp_cpus_done(unsigned int max_cpus
)
486 #ifdef CONFIG_HOTPLUG_CPU
487 static int xen_cpu_disable(void)
489 unsigned int cpu
= smp_processor_id();
493 cpu_disable_common();
495 load_cr3(swapper_pg_dir
);
499 static void xen_cpu_die(unsigned int cpu
)
501 while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up
,
502 xen_vcpu_nr(cpu
), NULL
)) {
503 __set_current_state(TASK_UNINTERRUPTIBLE
);
504 schedule_timeout(HZ
/10);
507 if (common_cpu_die(cpu
) == 0) {
508 xen_smp_intr_free(cpu
);
509 xen_uninit_lock_cpu(cpu
);
510 xen_teardown_timer(cpu
);
515 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
518 HYPERVISOR_vcpu_op(VCPUOP_down
, xen_vcpu_nr(smp_processor_id()), NULL
);
521 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
522 * clears certain data that the cpu_idle loop (which called us
523 * and that we return from) expects. The only way to get that
524 * data back is to call:
526 tick_nohz_idle_enter();
528 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE
);
531 #else /* !CONFIG_HOTPLUG_CPU */
532 static int xen_cpu_disable(void)
537 static void xen_cpu_die(unsigned int cpu
)
542 static void xen_play_dead(void)
548 static void stop_self(void *v
)
550 int cpu
= smp_processor_id();
552 /* make sure we're not pinning something down */
553 load_cr3(swapper_pg_dir
);
554 /* should set up a minimal gdt */
556 set_cpu_online(cpu
, false);
558 HYPERVISOR_vcpu_op(VCPUOP_down
, xen_vcpu_nr(cpu
), NULL
);
562 static void xen_stop_other_cpus(int wait
)
564 smp_call_function(stop_self
, NULL
, wait
);
567 static void xen_smp_send_reschedule(int cpu
)
569 xen_send_IPI_one(cpu
, XEN_RESCHEDULE_VECTOR
);
572 static void __xen_send_IPI_mask(const struct cpumask
*mask
,
577 for_each_cpu_and(cpu
, mask
, cpu_online_mask
)
578 xen_send_IPI_one(cpu
, vector
);
581 static void xen_smp_send_call_function_ipi(const struct cpumask
*mask
)
585 __xen_send_IPI_mask(mask
, XEN_CALL_FUNCTION_VECTOR
);
587 /* Make sure other vcpus get a chance to run if they need to. */
588 for_each_cpu(cpu
, mask
) {
589 if (xen_vcpu_stolen(cpu
)) {
590 HYPERVISOR_sched_op(SCHEDOP_yield
, NULL
);
596 static void xen_smp_send_call_function_single_ipi(int cpu
)
598 __xen_send_IPI_mask(cpumask_of(cpu
),
599 XEN_CALL_FUNCTION_SINGLE_VECTOR
);
602 static inline int xen_map_vector(int vector
)
607 case RESCHEDULE_VECTOR
:
608 xen_vector
= XEN_RESCHEDULE_VECTOR
;
610 case CALL_FUNCTION_VECTOR
:
611 xen_vector
= XEN_CALL_FUNCTION_VECTOR
;
613 case CALL_FUNCTION_SINGLE_VECTOR
:
614 xen_vector
= XEN_CALL_FUNCTION_SINGLE_VECTOR
;
616 case IRQ_WORK_VECTOR
:
617 xen_vector
= XEN_IRQ_WORK_VECTOR
;
621 case APIC_DM_NMI
: /* Some use that instead of NMI_VECTOR */
622 xen_vector
= XEN_NMI_VECTOR
;
627 printk(KERN_ERR
"xen: vector 0x%x is not implemented\n",
634 void xen_send_IPI_mask(const struct cpumask
*mask
,
637 int xen_vector
= xen_map_vector(vector
);
640 __xen_send_IPI_mask(mask
, xen_vector
);
643 void xen_send_IPI_all(int vector
)
645 int xen_vector
= xen_map_vector(vector
);
648 __xen_send_IPI_mask(cpu_online_mask
, xen_vector
);
651 void xen_send_IPI_self(int vector
)
653 int xen_vector
= xen_map_vector(vector
);
656 xen_send_IPI_one(smp_processor_id(), xen_vector
);
659 void xen_send_IPI_mask_allbutself(const struct cpumask
*mask
,
663 unsigned int this_cpu
= smp_processor_id();
664 int xen_vector
= xen_map_vector(vector
);
666 if (!(num_online_cpus() > 1) || (xen_vector
< 0))
669 for_each_cpu_and(cpu
, mask
, cpu_online_mask
) {
673 xen_send_IPI_one(cpu
, xen_vector
);
677 void xen_send_IPI_allbutself(int vector
)
679 xen_send_IPI_mask_allbutself(cpu_online_mask
, vector
);
682 static irqreturn_t
xen_call_function_interrupt(int irq
, void *dev_id
)
685 generic_smp_call_function_interrupt();
686 inc_irq_stat(irq_call_count
);
692 static irqreturn_t
xen_call_function_single_interrupt(int irq
, void *dev_id
)
695 generic_smp_call_function_single_interrupt();
696 inc_irq_stat(irq_call_count
);
702 static irqreturn_t
xen_irq_work_interrupt(int irq
, void *dev_id
)
706 inc_irq_stat(apic_irq_work_irqs
);
712 static const struct smp_ops xen_smp_ops __initconst
= {
713 .smp_prepare_boot_cpu
= xen_smp_prepare_boot_cpu
,
714 .smp_prepare_cpus
= xen_smp_prepare_cpus
,
715 .smp_cpus_done
= xen_smp_cpus_done
,
717 .cpu_up
= xen_cpu_up
,
718 .cpu_die
= xen_cpu_die
,
719 .cpu_disable
= xen_cpu_disable
,
720 .play_dead
= xen_play_dead
,
722 .stop_other_cpus
= xen_stop_other_cpus
,
723 .smp_send_reschedule
= xen_smp_send_reschedule
,
725 .send_call_func_ipi
= xen_smp_send_call_function_ipi
,
726 .send_call_func_single_ipi
= xen_smp_send_call_function_single_ipi
,
729 void __init
xen_smp_init(void)
731 smp_ops
= xen_smp_ops
;
732 xen_fill_possible_map();
735 static void __init
xen_hvm_smp_prepare_cpus(unsigned int max_cpus
)
737 native_smp_prepare_cpus(max_cpus
);
738 WARN_ON(xen_smp_intr_init(0));
740 xen_init_lock_cpu(0);
743 void __init
xen_hvm_smp_init(void)
745 smp_ops
.smp_prepare_cpus
= xen_hvm_smp_prepare_cpus
;
746 smp_ops
.smp_send_reschedule
= xen_smp_send_reschedule
;
747 smp_ops
.cpu_die
= xen_cpu_die
;
748 smp_ops
.send_call_func_ipi
= xen_smp_send_call_function_ipi
;
749 smp_ops
.send_call_func_single_ipi
= xen_smp_send_call_function_single_ipi
;
750 smp_ops
.smp_prepare_boot_cpu
= xen_smp_prepare_boot_cpu
;