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>
30 #include <asm/xen/interface.h>
31 #include <asm/xen/hypercall.h>
35 #include <xen/events.h>
37 #include <xen/hvc-console.h>
41 cpumask_var_t xen_cpu_initialized_map
;
43 struct xen_common_irq
{
47 static DEFINE_PER_CPU(struct xen_common_irq
, xen_resched_irq
) = { .irq
= -1 };
48 static DEFINE_PER_CPU(struct xen_common_irq
, xen_callfunc_irq
) = { .irq
= -1 };
49 static DEFINE_PER_CPU(struct xen_common_irq
, xen_callfuncsingle_irq
) = { .irq
= -1 };
50 static DEFINE_PER_CPU(struct xen_common_irq
, xen_irq_work
) = { .irq
= -1 };
51 static DEFINE_PER_CPU(struct xen_common_irq
, xen_debug_irq
) = { .irq
= -1 };
53 static irqreturn_t
xen_call_function_interrupt(int irq
, void *dev_id
);
54 static irqreturn_t
xen_call_function_single_interrupt(int irq
, void *dev_id
);
55 static irqreturn_t
xen_irq_work_interrupt(int irq
, void *dev_id
);
58 * Reschedule call back.
60 static irqreturn_t
xen_reschedule_interrupt(int irq
, void *dev_id
)
62 inc_irq_stat(irq_resched_count
);
68 static void cpu_bringup(void)
73 touch_softlockup_watchdog();
76 xen_enable_sysenter();
79 cpu
= smp_processor_id();
80 smp_store_cpu_info(cpu
);
81 cpu_data(cpu
).x86_max_cores
= 1;
82 set_cpu_sibling_map(cpu
);
84 xen_setup_cpu_clockevents();
86 notify_cpu_starting(cpu
);
88 set_cpu_online(cpu
, true);
90 this_cpu_write(cpu_state
, CPU_ONLINE
);
94 /* We can take interrupts now: we're officially "up". */
97 wmb(); /* make sure everything is out */
100 static void cpu_bringup_and_idle(void)
103 cpu_startup_entry(CPUHP_ONLINE
);
106 static void xen_smp_intr_free(unsigned int cpu
)
108 if (per_cpu(xen_resched_irq
, cpu
).irq
>= 0) {
109 unbind_from_irqhandler(per_cpu(xen_resched_irq
, cpu
).irq
, NULL
);
110 per_cpu(xen_resched_irq
, cpu
).irq
= -1;
111 kfree(per_cpu(xen_resched_irq
, cpu
).name
);
112 per_cpu(xen_resched_irq
, cpu
).name
= NULL
;
114 if (per_cpu(xen_callfunc_irq
, cpu
).irq
>= 0) {
115 unbind_from_irqhandler(per_cpu(xen_callfunc_irq
, cpu
).irq
, NULL
);
116 per_cpu(xen_callfunc_irq
, cpu
).irq
= -1;
117 kfree(per_cpu(xen_callfunc_irq
, cpu
).name
);
118 per_cpu(xen_callfunc_irq
, cpu
).name
= NULL
;
120 if (per_cpu(xen_debug_irq
, cpu
).irq
>= 0) {
121 unbind_from_irqhandler(per_cpu(xen_debug_irq
, cpu
).irq
, NULL
);
122 per_cpu(xen_debug_irq
, cpu
).irq
= -1;
123 kfree(per_cpu(xen_debug_irq
, cpu
).name
);
124 per_cpu(xen_debug_irq
, cpu
).name
= NULL
;
126 if (per_cpu(xen_callfuncsingle_irq
, cpu
).irq
>= 0) {
127 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq
, cpu
).irq
,
129 per_cpu(xen_callfuncsingle_irq
, cpu
).irq
= -1;
130 kfree(per_cpu(xen_callfuncsingle_irq
, cpu
).name
);
131 per_cpu(xen_callfuncsingle_irq
, cpu
).name
= NULL
;
133 if (xen_hvm_domain())
136 if (per_cpu(xen_irq_work
, cpu
).irq
>= 0) {
137 unbind_from_irqhandler(per_cpu(xen_irq_work
, cpu
).irq
, NULL
);
138 per_cpu(xen_irq_work
, cpu
).irq
= -1;
139 kfree(per_cpu(xen_irq_work
, cpu
).name
);
140 per_cpu(xen_irq_work
, cpu
).name
= NULL
;
143 static int xen_smp_intr_init(unsigned int cpu
)
146 char *resched_name
, *callfunc_name
, *debug_name
;
148 resched_name
= kasprintf(GFP_KERNEL
, "resched%d", cpu
);
149 rc
= bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR
,
151 xen_reschedule_interrupt
,
152 IRQF_DISABLED
|IRQF_PERCPU
|IRQF_NOBALANCING
,
157 per_cpu(xen_resched_irq
, cpu
).irq
= rc
;
158 per_cpu(xen_resched_irq
, cpu
).name
= resched_name
;
160 callfunc_name
= kasprintf(GFP_KERNEL
, "callfunc%d", cpu
);
161 rc
= bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR
,
163 xen_call_function_interrupt
,
164 IRQF_DISABLED
|IRQF_PERCPU
|IRQF_NOBALANCING
,
169 per_cpu(xen_callfunc_irq
, cpu
).irq
= rc
;
170 per_cpu(xen_callfunc_irq
, cpu
).name
= callfunc_name
;
172 debug_name
= kasprintf(GFP_KERNEL
, "debug%d", cpu
);
173 rc
= bind_virq_to_irqhandler(VIRQ_DEBUG
, cpu
, xen_debug_interrupt
,
174 IRQF_DISABLED
| IRQF_PERCPU
| IRQF_NOBALANCING
,
178 per_cpu(xen_debug_irq
, cpu
).irq
= rc
;
179 per_cpu(xen_debug_irq
, cpu
).name
= debug_name
;
181 callfunc_name
= kasprintf(GFP_KERNEL
, "callfuncsingle%d", cpu
);
182 rc
= bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR
,
184 xen_call_function_single_interrupt
,
185 IRQF_DISABLED
|IRQF_PERCPU
|IRQF_NOBALANCING
,
190 per_cpu(xen_callfuncsingle_irq
, cpu
).irq
= rc
;
191 per_cpu(xen_callfuncsingle_irq
, cpu
).name
= callfunc_name
;
194 * The IRQ worker on PVHVM goes through the native path and uses the
197 if (xen_hvm_domain())
200 callfunc_name
= kasprintf(GFP_KERNEL
, "irqwork%d", cpu
);
201 rc
= bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR
,
203 xen_irq_work_interrupt
,
204 IRQF_DISABLED
|IRQF_PERCPU
|IRQF_NOBALANCING
,
209 per_cpu(xen_irq_work
, cpu
).irq
= rc
;
210 per_cpu(xen_irq_work
, cpu
).name
= callfunc_name
;
215 xen_smp_intr_free(cpu
);
219 static void __init
xen_fill_possible_map(void)
223 if (xen_initial_domain())
226 for (i
= 0; i
< nr_cpu_ids
; i
++) {
227 rc
= HYPERVISOR_vcpu_op(VCPUOP_is_up
, i
, NULL
);
230 set_cpu_possible(i
, true);
235 static void __init
xen_filter_cpu_maps(void)
238 unsigned int subtract
= 0;
240 if (!xen_initial_domain())
245 for (i
= 0; i
< nr_cpu_ids
; i
++) {
246 rc
= HYPERVISOR_vcpu_op(VCPUOP_is_up
, i
, NULL
);
249 set_cpu_possible(i
, true);
251 set_cpu_possible(i
, false);
252 set_cpu_present(i
, false);
256 #ifdef CONFIG_HOTPLUG_CPU
257 /* This is akin to using 'nr_cpus' on the Linux command line.
258 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
259 * have up to X, while nr_cpu_ids is greater than X. This
260 * normally is not a problem, except when CPU hotplugging
261 * is involved and then there might be more than X CPUs
262 * in the guest - which will not work as there is no
263 * hypercall to expand the max number of VCPUs an already
264 * running guest has. So cap it up to X. */
266 nr_cpu_ids
= nr_cpu_ids
- subtract
;
271 static void __init
xen_smp_prepare_boot_cpu(void)
273 BUG_ON(smp_processor_id() != 0);
274 native_smp_prepare_boot_cpu();
276 if (xen_pv_domain()) {
277 /* We've switched to the "real" per-cpu gdt, so make sure the
278 old memory can be recycled */
279 make_lowmem_page_readwrite(xen_initial_gdt
);
283 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
286 loadsegment(ds
, __USER_DS
);
287 loadsegment(es
, __USER_DS
);
290 xen_filter_cpu_maps();
291 xen_setup_vcpu_info_placement();
294 * The alternative logic (which patches the unlock/lock) runs before
295 * the smp bootup up code is activated. Hence we need to set this up
296 * the core kernel is being patched. Otherwise we will have only
297 * modules patched but not core code.
299 xen_init_spinlocks();
302 static void __init
xen_smp_prepare_cpus(unsigned int max_cpus
)
307 if (skip_ioapic_setup
) {
308 char *m
= (max_cpus
== 0) ?
309 "The nosmp parameter is incompatible with Xen; " \
310 "use Xen dom0_max_vcpus=1 parameter" :
311 "The noapic parameter is incompatible with Xen";
316 xen_init_lock_cpu(0);
318 smp_store_boot_cpu_info();
319 cpu_data(0).x86_max_cores
= 1;
321 for_each_possible_cpu(i
) {
322 zalloc_cpumask_var(&per_cpu(cpu_sibling_map
, i
), GFP_KERNEL
);
323 zalloc_cpumask_var(&per_cpu(cpu_core_map
, i
), GFP_KERNEL
);
324 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map
, i
), GFP_KERNEL
);
326 set_cpu_sibling_map(0);
328 if (xen_smp_intr_init(0))
331 if (!alloc_cpumask_var(&xen_cpu_initialized_map
, GFP_KERNEL
))
332 panic("could not allocate xen_cpu_initialized_map\n");
334 cpumask_copy(xen_cpu_initialized_map
, cpumask_of(0));
336 /* Restrict the possible_map according to max_cpus. */
337 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus
)) {
338 for (cpu
= nr_cpu_ids
- 1; !cpu_possible(cpu
); cpu
--)
340 set_cpu_possible(cpu
, false);
343 for_each_possible_cpu(cpu
)
344 set_cpu_present(cpu
, true);
348 cpu_initialize_context(unsigned int cpu
, struct task_struct
*idle
)
350 struct vcpu_guest_context
*ctxt
;
351 struct desc_struct
*gdt
;
352 unsigned long gdt_mfn
;
354 if (cpumask_test_and_set_cpu(cpu
, xen_cpu_initialized_map
))
357 ctxt
= kzalloc(sizeof(*ctxt
), GFP_KERNEL
);
361 gdt
= get_cpu_gdt_table(cpu
);
363 ctxt
->flags
= VGCF_IN_KERNEL
;
364 ctxt
->user_regs
.ss
= __KERNEL_DS
;
366 ctxt
->user_regs
.fs
= __KERNEL_PERCPU
;
367 ctxt
->user_regs
.gs
= __KERNEL_STACK_CANARY
;
369 ctxt
->gs_base_kernel
= per_cpu_offset(cpu
);
371 ctxt
->user_regs
.eip
= (unsigned long)cpu_bringup_and_idle
;
373 memset(&ctxt
->fpu_ctxt
, 0, sizeof(ctxt
->fpu_ctxt
));
376 ctxt
->user_regs
.eflags
= 0x1000; /* IOPL_RING1 */
377 ctxt
->user_regs
.ds
= __USER_DS
;
378 ctxt
->user_regs
.es
= __USER_DS
;
380 xen_copy_trap_info(ctxt
->trap_ctxt
);
384 BUG_ON((unsigned long)gdt
& ~PAGE_MASK
);
386 gdt_mfn
= arbitrary_virt_to_mfn(gdt
);
387 make_lowmem_page_readonly(gdt
);
388 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn
));
390 ctxt
->gdt_frames
[0] = gdt_mfn
;
391 ctxt
->gdt_ents
= GDT_ENTRIES
;
393 ctxt
->kernel_ss
= __KERNEL_DS
;
394 ctxt
->kernel_sp
= idle
->thread
.sp0
;
397 ctxt
->event_callback_cs
= __KERNEL_CS
;
398 ctxt
->failsafe_callback_cs
= __KERNEL_CS
;
400 ctxt
->event_callback_eip
=
401 (unsigned long)xen_hypervisor_callback
;
402 ctxt
->failsafe_callback_eip
=
403 (unsigned long)xen_failsafe_callback
;
405 ctxt
->user_regs
.cs
= __KERNEL_CS
;
406 ctxt
->user_regs
.esp
= idle
->thread
.sp0
- sizeof(struct pt_regs
);
408 per_cpu(xen_cr3
, cpu
) = __pa(swapper_pg_dir
);
409 ctxt
->ctrlreg
[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir
));
411 if (HYPERVISOR_vcpu_op(VCPUOP_initialise
, cpu
, ctxt
))
418 static int xen_cpu_up(unsigned int cpu
, struct task_struct
*idle
)
422 per_cpu(current_task
, cpu
) = idle
;
426 clear_tsk_thread_flag(idle
, TIF_FORK
);
427 per_cpu(kernel_stack
, cpu
) =
428 (unsigned long)task_stack_page(idle
) -
429 KERNEL_STACK_OFFSET
+ THREAD_SIZE
;
431 xen_setup_runstate_info(cpu
);
432 xen_setup_timer(cpu
);
433 xen_init_lock_cpu(cpu
);
435 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
437 /* make sure interrupts start blocked */
438 per_cpu(xen_vcpu
, cpu
)->evtchn_upcall_mask
= 1;
440 rc
= cpu_initialize_context(cpu
, idle
);
444 if (num_online_cpus() == 1)
445 /* Just in case we booted with a single CPU. */
446 alternatives_enable_smp();
448 rc
= xen_smp_intr_init(cpu
);
452 rc
= HYPERVISOR_vcpu_op(VCPUOP_up
, cpu
, NULL
);
455 while(per_cpu(cpu_state
, cpu
) != CPU_ONLINE
) {
456 HYPERVISOR_sched_op(SCHEDOP_yield
, NULL
);
463 static void xen_smp_cpus_done(unsigned int max_cpus
)
467 #ifdef CONFIG_HOTPLUG_CPU
468 static int xen_cpu_disable(void)
470 unsigned int cpu
= smp_processor_id();
474 cpu_disable_common();
476 load_cr3(swapper_pg_dir
);
480 static void xen_cpu_die(unsigned int cpu
)
482 while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up
, cpu
, NULL
)) {
483 current
->state
= TASK_UNINTERRUPTIBLE
;
484 schedule_timeout(HZ
/10);
486 xen_smp_intr_free(cpu
);
487 xen_uninit_lock_cpu(cpu
);
488 xen_teardown_timer(cpu
);
491 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
494 HYPERVISOR_vcpu_op(VCPUOP_down
, smp_processor_id(), NULL
);
497 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
498 * clears certain data that the cpu_idle loop (which called us
499 * and that we return from) expects. The only way to get that
500 * data back is to call:
502 tick_nohz_idle_enter();
505 #else /* !CONFIG_HOTPLUG_CPU */
506 static int xen_cpu_disable(void)
511 static void xen_cpu_die(unsigned int cpu
)
516 static void xen_play_dead(void)
522 static void stop_self(void *v
)
524 int cpu
= smp_processor_id();
526 /* make sure we're not pinning something down */
527 load_cr3(swapper_pg_dir
);
528 /* should set up a minimal gdt */
530 set_cpu_online(cpu
, false);
532 HYPERVISOR_vcpu_op(VCPUOP_down
, cpu
, NULL
);
536 static void xen_stop_other_cpus(int wait
)
538 smp_call_function(stop_self
, NULL
, wait
);
541 static void xen_smp_send_reschedule(int cpu
)
543 xen_send_IPI_one(cpu
, XEN_RESCHEDULE_VECTOR
);
546 static void __xen_send_IPI_mask(const struct cpumask
*mask
,
551 for_each_cpu_and(cpu
, mask
, cpu_online_mask
)
552 xen_send_IPI_one(cpu
, vector
);
555 static void xen_smp_send_call_function_ipi(const struct cpumask
*mask
)
559 __xen_send_IPI_mask(mask
, XEN_CALL_FUNCTION_VECTOR
);
561 /* Make sure other vcpus get a chance to run if they need to. */
562 for_each_cpu(cpu
, mask
) {
563 if (xen_vcpu_stolen(cpu
)) {
564 HYPERVISOR_sched_op(SCHEDOP_yield
, NULL
);
570 static void xen_smp_send_call_function_single_ipi(int cpu
)
572 __xen_send_IPI_mask(cpumask_of(cpu
),
573 XEN_CALL_FUNCTION_SINGLE_VECTOR
);
576 static inline int xen_map_vector(int vector
)
581 case RESCHEDULE_VECTOR
:
582 xen_vector
= XEN_RESCHEDULE_VECTOR
;
584 case CALL_FUNCTION_VECTOR
:
585 xen_vector
= XEN_CALL_FUNCTION_VECTOR
;
587 case CALL_FUNCTION_SINGLE_VECTOR
:
588 xen_vector
= XEN_CALL_FUNCTION_SINGLE_VECTOR
;
590 case IRQ_WORK_VECTOR
:
591 xen_vector
= XEN_IRQ_WORK_VECTOR
;
595 case APIC_DM_NMI
: /* Some use that instead of NMI_VECTOR */
596 xen_vector
= XEN_NMI_VECTOR
;
601 printk(KERN_ERR
"xen: vector 0x%x is not implemented\n",
608 void xen_send_IPI_mask(const struct cpumask
*mask
,
611 int xen_vector
= xen_map_vector(vector
);
614 __xen_send_IPI_mask(mask
, xen_vector
);
617 void xen_send_IPI_all(int vector
)
619 int xen_vector
= xen_map_vector(vector
);
622 __xen_send_IPI_mask(cpu_online_mask
, xen_vector
);
625 void xen_send_IPI_self(int vector
)
627 int xen_vector
= xen_map_vector(vector
);
630 xen_send_IPI_one(smp_processor_id(), xen_vector
);
633 void xen_send_IPI_mask_allbutself(const struct cpumask
*mask
,
637 unsigned int this_cpu
= smp_processor_id();
638 int xen_vector
= xen_map_vector(vector
);
640 if (!(num_online_cpus() > 1) || (xen_vector
< 0))
643 for_each_cpu_and(cpu
, mask
, cpu_online_mask
) {
647 xen_send_IPI_one(cpu
, xen_vector
);
651 void xen_send_IPI_allbutself(int vector
)
653 xen_send_IPI_mask_allbutself(cpu_online_mask
, vector
);
656 static irqreturn_t
xen_call_function_interrupt(int irq
, void *dev_id
)
659 generic_smp_call_function_interrupt();
660 inc_irq_stat(irq_call_count
);
666 static irqreturn_t
xen_call_function_single_interrupt(int irq
, void *dev_id
)
669 generic_smp_call_function_single_interrupt();
670 inc_irq_stat(irq_call_count
);
676 static irqreturn_t
xen_irq_work_interrupt(int irq
, void *dev_id
)
680 inc_irq_stat(apic_irq_work_irqs
);
686 static const struct smp_ops xen_smp_ops __initconst
= {
687 .smp_prepare_boot_cpu
= xen_smp_prepare_boot_cpu
,
688 .smp_prepare_cpus
= xen_smp_prepare_cpus
,
689 .smp_cpus_done
= xen_smp_cpus_done
,
691 .cpu_up
= xen_cpu_up
,
692 .cpu_die
= xen_cpu_die
,
693 .cpu_disable
= xen_cpu_disable
,
694 .play_dead
= xen_play_dead
,
696 .stop_other_cpus
= xen_stop_other_cpus
,
697 .smp_send_reschedule
= xen_smp_send_reschedule
,
699 .send_call_func_ipi
= xen_smp_send_call_function_ipi
,
700 .send_call_func_single_ipi
= xen_smp_send_call_function_single_ipi
,
703 void __init
xen_smp_init(void)
705 smp_ops
= xen_smp_ops
;
706 xen_fill_possible_map();
709 static void __init
xen_hvm_smp_prepare_cpus(unsigned int max_cpus
)
711 native_smp_prepare_cpus(max_cpus
);
712 WARN_ON(xen_smp_intr_init(0));
714 xen_init_lock_cpu(0);
717 static int xen_hvm_cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
721 * xen_smp_intr_init() needs to run before native_cpu_up()
722 * so that IPI vectors are set up on the booting CPU before
723 * it is marked online in native_cpu_up().
725 rc
= xen_smp_intr_init(cpu
);
728 rc
= native_cpu_up(cpu
, tidle
);
731 * We must initialize the slowpath CPU kicker _after_ the native
732 * path has executed. If we initialized it before none of the
733 * unlocker IPI kicks would reach the booting CPU as the booting
734 * CPU had not set itself 'online' in cpu_online_mask. That mask
735 * is checked when IPIs are sent (on HVM at least).
737 xen_init_lock_cpu(cpu
);
741 static void xen_hvm_cpu_die(unsigned int cpu
)
747 void __init
xen_hvm_smp_init(void)
749 if (!xen_have_vector_callback
)
751 smp_ops
.smp_prepare_cpus
= xen_hvm_smp_prepare_cpus
;
752 smp_ops
.smp_send_reschedule
= xen_smp_send_reschedule
;
753 smp_ops
.cpu_up
= xen_hvm_cpu_up
;
754 smp_ops
.cpu_die
= xen_hvm_cpu_die
;
755 smp_ops
.send_call_func_ipi
= xen_smp_send_call_function_ipi
;
756 smp_ops
.send_call_func_single_ipi
= xen_smp_send_call_function_single_ipi
;
757 smp_ops
.smp_prepare_boot_cpu
= xen_smp_prepare_boot_cpu
;