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 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
77 if (!xen_feature(XENFEAT_supervisor_mode_kernel
)) {
78 xen_enable_sysenter();
81 cpu
= smp_processor_id();
82 smp_store_cpu_info(cpu
);
83 cpu_data(cpu
).x86_max_cores
= 1;
84 set_cpu_sibling_map(cpu
);
86 xen_setup_cpu_clockevents();
88 notify_cpu_starting(cpu
);
90 set_cpu_online(cpu
, true);
92 this_cpu_write(cpu_state
, CPU_ONLINE
);
96 /* We can take interrupts now: we're officially "up". */
99 wmb(); /* make sure everything is out */
102 /* Note: cpu parameter is only relevant for PVH */
103 static void cpu_bringup_and_idle(int cpu
)
106 if (xen_feature(XENFEAT_auto_translated_physmap
) &&
107 xen_feature(XENFEAT_supervisor_mode_kernel
))
108 xen_pvh_secondary_vcpu_init(cpu
);
111 cpu_startup_entry(CPUHP_ONLINE
);
114 static void xen_smp_intr_free(unsigned int cpu
)
116 if (per_cpu(xen_resched_irq
, cpu
).irq
>= 0) {
117 unbind_from_irqhandler(per_cpu(xen_resched_irq
, cpu
).irq
, NULL
);
118 per_cpu(xen_resched_irq
, cpu
).irq
= -1;
119 kfree(per_cpu(xen_resched_irq
, cpu
).name
);
120 per_cpu(xen_resched_irq
, cpu
).name
= NULL
;
122 if (per_cpu(xen_callfunc_irq
, cpu
).irq
>= 0) {
123 unbind_from_irqhandler(per_cpu(xen_callfunc_irq
, cpu
).irq
, NULL
);
124 per_cpu(xen_callfunc_irq
, cpu
).irq
= -1;
125 kfree(per_cpu(xen_callfunc_irq
, cpu
).name
);
126 per_cpu(xen_callfunc_irq
, cpu
).name
= NULL
;
128 if (per_cpu(xen_debug_irq
, cpu
).irq
>= 0) {
129 unbind_from_irqhandler(per_cpu(xen_debug_irq
, cpu
).irq
, NULL
);
130 per_cpu(xen_debug_irq
, cpu
).irq
= -1;
131 kfree(per_cpu(xen_debug_irq
, cpu
).name
);
132 per_cpu(xen_debug_irq
, cpu
).name
= NULL
;
134 if (per_cpu(xen_callfuncsingle_irq
, cpu
).irq
>= 0) {
135 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq
, cpu
).irq
,
137 per_cpu(xen_callfuncsingle_irq
, cpu
).irq
= -1;
138 kfree(per_cpu(xen_callfuncsingle_irq
, cpu
).name
);
139 per_cpu(xen_callfuncsingle_irq
, cpu
).name
= NULL
;
141 if (xen_hvm_domain())
144 if (per_cpu(xen_irq_work
, cpu
).irq
>= 0) {
145 unbind_from_irqhandler(per_cpu(xen_irq_work
, cpu
).irq
, NULL
);
146 per_cpu(xen_irq_work
, cpu
).irq
= -1;
147 kfree(per_cpu(xen_irq_work
, cpu
).name
);
148 per_cpu(xen_irq_work
, cpu
).name
= NULL
;
151 static int xen_smp_intr_init(unsigned int cpu
)
154 char *resched_name
, *callfunc_name
, *debug_name
;
156 resched_name
= kasprintf(GFP_KERNEL
, "resched%d", cpu
);
157 rc
= bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR
,
159 xen_reschedule_interrupt
,
160 IRQF_PERCPU
|IRQF_NOBALANCING
,
165 per_cpu(xen_resched_irq
, cpu
).irq
= rc
;
166 per_cpu(xen_resched_irq
, cpu
).name
= resched_name
;
168 callfunc_name
= kasprintf(GFP_KERNEL
, "callfunc%d", cpu
);
169 rc
= bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR
,
171 xen_call_function_interrupt
,
172 IRQF_PERCPU
|IRQF_NOBALANCING
,
177 per_cpu(xen_callfunc_irq
, cpu
).irq
= rc
;
178 per_cpu(xen_callfunc_irq
, cpu
).name
= callfunc_name
;
180 debug_name
= kasprintf(GFP_KERNEL
, "debug%d", cpu
);
181 rc
= bind_virq_to_irqhandler(VIRQ_DEBUG
, cpu
, xen_debug_interrupt
,
182 IRQF_PERCPU
| IRQF_NOBALANCING
,
186 per_cpu(xen_debug_irq
, cpu
).irq
= rc
;
187 per_cpu(xen_debug_irq
, cpu
).name
= debug_name
;
189 callfunc_name
= kasprintf(GFP_KERNEL
, "callfuncsingle%d", cpu
);
190 rc
= bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR
,
192 xen_call_function_single_interrupt
,
193 IRQF_PERCPU
|IRQF_NOBALANCING
,
198 per_cpu(xen_callfuncsingle_irq
, cpu
).irq
= rc
;
199 per_cpu(xen_callfuncsingle_irq
, cpu
).name
= callfunc_name
;
202 * The IRQ worker on PVHVM goes through the native path and uses the
205 if (xen_hvm_domain())
208 callfunc_name
= kasprintf(GFP_KERNEL
, "irqwork%d", cpu
);
209 rc
= bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR
,
211 xen_irq_work_interrupt
,
212 IRQF_PERCPU
|IRQF_NOBALANCING
,
217 per_cpu(xen_irq_work
, cpu
).irq
= rc
;
218 per_cpu(xen_irq_work
, cpu
).name
= callfunc_name
;
223 xen_smp_intr_free(cpu
);
227 static void __init
xen_fill_possible_map(void)
231 if (xen_initial_domain())
234 for (i
= 0; i
< nr_cpu_ids
; i
++) {
235 rc
= HYPERVISOR_vcpu_op(VCPUOP_is_up
, i
, NULL
);
238 set_cpu_possible(i
, true);
243 static void __init
xen_filter_cpu_maps(void)
246 unsigned int subtract
= 0;
248 if (!xen_initial_domain())
253 for (i
= 0; i
< nr_cpu_ids
; i
++) {
254 rc
= HYPERVISOR_vcpu_op(VCPUOP_is_up
, i
, NULL
);
257 set_cpu_possible(i
, true);
259 set_cpu_possible(i
, false);
260 set_cpu_present(i
, false);
264 #ifdef CONFIG_HOTPLUG_CPU
265 /* This is akin to using 'nr_cpus' on the Linux command line.
266 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
267 * have up to X, while nr_cpu_ids is greater than X. This
268 * normally is not a problem, except when CPU hotplugging
269 * is involved and then there might be more than X CPUs
270 * in the guest - which will not work as there is no
271 * hypercall to expand the max number of VCPUs an already
272 * running guest has. So cap it up to X. */
274 nr_cpu_ids
= nr_cpu_ids
- subtract
;
279 static void __init
xen_smp_prepare_boot_cpu(void)
281 BUG_ON(smp_processor_id() != 0);
282 native_smp_prepare_boot_cpu();
284 if (xen_pv_domain()) {
285 if (!xen_feature(XENFEAT_writable_page_tables
))
286 /* We've switched to the "real" per-cpu gdt, so make
287 * sure the old memory can be recycled. */
288 make_lowmem_page_readwrite(xen_initial_gdt
);
292 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
295 loadsegment(ds
, __USER_DS
);
296 loadsegment(es
, __USER_DS
);
299 xen_filter_cpu_maps();
300 xen_setup_vcpu_info_placement();
303 * The alternative logic (which patches the unlock/lock) runs before
304 * the smp bootup up code is activated. Hence we need to set this up
305 * the core kernel is being patched. Otherwise we will have only
306 * modules patched but not core code.
308 xen_init_spinlocks();
311 static void __init
xen_smp_prepare_cpus(unsigned int max_cpus
)
316 if (skip_ioapic_setup
) {
317 char *m
= (max_cpus
== 0) ?
318 "The nosmp parameter is incompatible with Xen; " \
319 "use Xen dom0_max_vcpus=1 parameter" :
320 "The noapic parameter is incompatible with Xen";
325 xen_init_lock_cpu(0);
327 smp_store_boot_cpu_info();
328 cpu_data(0).x86_max_cores
= 1;
330 for_each_possible_cpu(i
) {
331 zalloc_cpumask_var(&per_cpu(cpu_sibling_map
, i
), GFP_KERNEL
);
332 zalloc_cpumask_var(&per_cpu(cpu_core_map
, i
), GFP_KERNEL
);
333 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map
, i
), GFP_KERNEL
);
335 set_cpu_sibling_map(0);
337 if (xen_smp_intr_init(0))
340 if (!alloc_cpumask_var(&xen_cpu_initialized_map
, GFP_KERNEL
))
341 panic("could not allocate xen_cpu_initialized_map\n");
343 cpumask_copy(xen_cpu_initialized_map
, cpumask_of(0));
345 /* Restrict the possible_map according to max_cpus. */
346 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus
)) {
347 for (cpu
= nr_cpu_ids
- 1; !cpu_possible(cpu
); cpu
--)
349 set_cpu_possible(cpu
, false);
352 for_each_possible_cpu(cpu
)
353 set_cpu_present(cpu
, true);
357 cpu_initialize_context(unsigned int cpu
, struct task_struct
*idle
)
359 struct vcpu_guest_context
*ctxt
;
360 struct desc_struct
*gdt
;
361 unsigned long gdt_mfn
;
363 if (cpumask_test_and_set_cpu(cpu
, xen_cpu_initialized_map
))
366 ctxt
= kzalloc(sizeof(*ctxt
), GFP_KERNEL
);
370 gdt
= get_cpu_gdt_table(cpu
);
373 /* Note: PVH is not yet supported on x86_32. */
374 ctxt
->user_regs
.fs
= __KERNEL_PERCPU
;
375 ctxt
->user_regs
.gs
= __KERNEL_STACK_CANARY
;
377 ctxt
->user_regs
.eip
= (unsigned long)cpu_bringup_and_idle
;
379 memset(&ctxt
->fpu_ctxt
, 0, sizeof(ctxt
->fpu_ctxt
));
381 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
382 ctxt
->flags
= VGCF_IN_KERNEL
;
383 ctxt
->user_regs
.eflags
= 0x1000; /* IOPL_RING1 */
384 ctxt
->user_regs
.ds
= __USER_DS
;
385 ctxt
->user_regs
.es
= __USER_DS
;
386 ctxt
->user_regs
.ss
= __KERNEL_DS
;
388 xen_copy_trap_info(ctxt
->trap_ctxt
);
392 BUG_ON((unsigned long)gdt
& ~PAGE_MASK
);
394 gdt_mfn
= arbitrary_virt_to_mfn(gdt
);
395 make_lowmem_page_readonly(gdt
);
396 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn
));
398 ctxt
->gdt_frames
[0] = gdt_mfn
;
399 ctxt
->gdt_ents
= GDT_ENTRIES
;
401 ctxt
->kernel_ss
= __KERNEL_DS
;
402 ctxt
->kernel_sp
= idle
->thread
.sp0
;
405 ctxt
->event_callback_cs
= __KERNEL_CS
;
406 ctxt
->failsafe_callback_cs
= __KERNEL_CS
;
408 ctxt
->gs_base_kernel
= per_cpu_offset(cpu
);
410 ctxt
->event_callback_eip
=
411 (unsigned long)xen_hypervisor_callback
;
412 ctxt
->failsafe_callback_eip
=
413 (unsigned long)xen_failsafe_callback
;
414 ctxt
->user_regs
.cs
= __KERNEL_CS
;
415 per_cpu(xen_cr3
, cpu
) = __pa(swapper_pg_dir
);
420 /* N.B. The user_regs.eip (cpu_bringup_and_idle) is called with
421 * %rdi having the cpu number - which means are passing in
422 * as the first parameter the cpu. Subtle!
424 ctxt
->user_regs
.rdi
= cpu
;
426 ctxt
->user_regs
.esp
= idle
->thread
.sp0
- sizeof(struct pt_regs
);
427 ctxt
->ctrlreg
[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir
));
428 if (HYPERVISOR_vcpu_op(VCPUOP_initialise
, cpu
, ctxt
))
435 static int xen_cpu_up(unsigned int cpu
, struct task_struct
*idle
)
439 per_cpu(current_task
, cpu
) = idle
;
443 clear_tsk_thread_flag(idle
, TIF_FORK
);
444 per_cpu(kernel_stack
, cpu
) =
445 (unsigned long)task_stack_page(idle
) -
446 KERNEL_STACK_OFFSET
+ THREAD_SIZE
;
448 xen_setup_runstate_info(cpu
);
449 xen_setup_timer(cpu
);
450 xen_init_lock_cpu(cpu
);
452 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
454 /* make sure interrupts start blocked */
455 per_cpu(xen_vcpu
, cpu
)->evtchn_upcall_mask
= 1;
457 rc
= cpu_initialize_context(cpu
, idle
);
461 if (num_online_cpus() == 1)
462 /* Just in case we booted with a single CPU. */
463 alternatives_enable_smp();
465 rc
= xen_smp_intr_init(cpu
);
469 rc
= HYPERVISOR_vcpu_op(VCPUOP_up
, cpu
, NULL
);
472 while(per_cpu(cpu_state
, cpu
) != CPU_ONLINE
) {
473 HYPERVISOR_sched_op(SCHEDOP_yield
, NULL
);
480 static void xen_smp_cpus_done(unsigned int max_cpus
)
484 #ifdef CONFIG_HOTPLUG_CPU
485 static int xen_cpu_disable(void)
487 unsigned int cpu
= smp_processor_id();
491 cpu_disable_common();
493 load_cr3(swapper_pg_dir
);
497 static void xen_cpu_die(unsigned int cpu
)
499 while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up
, cpu
, NULL
)) {
500 current
->state
= TASK_UNINTERRUPTIBLE
;
501 schedule_timeout(HZ
/10);
503 xen_smp_intr_free(cpu
);
504 xen_uninit_lock_cpu(cpu
);
505 xen_teardown_timer(cpu
);
508 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
511 HYPERVISOR_vcpu_op(VCPUOP_down
, smp_processor_id(), NULL
);
514 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
515 * clears certain data that the cpu_idle loop (which called us
516 * and that we return from) expects. The only way to get that
517 * data back is to call:
519 tick_nohz_idle_enter();
522 #else /* !CONFIG_HOTPLUG_CPU */
523 static int xen_cpu_disable(void)
528 static void xen_cpu_die(unsigned int cpu
)
533 static void xen_play_dead(void)
539 static void stop_self(void *v
)
541 int cpu
= smp_processor_id();
543 /* make sure we're not pinning something down */
544 load_cr3(swapper_pg_dir
);
545 /* should set up a minimal gdt */
547 set_cpu_online(cpu
, false);
549 HYPERVISOR_vcpu_op(VCPUOP_down
, cpu
, NULL
);
553 static void xen_stop_other_cpus(int wait
)
555 smp_call_function(stop_self
, NULL
, wait
);
558 static void xen_smp_send_reschedule(int cpu
)
560 xen_send_IPI_one(cpu
, XEN_RESCHEDULE_VECTOR
);
563 static void __xen_send_IPI_mask(const struct cpumask
*mask
,
568 for_each_cpu_and(cpu
, mask
, cpu_online_mask
)
569 xen_send_IPI_one(cpu
, vector
);
572 static void xen_smp_send_call_function_ipi(const struct cpumask
*mask
)
576 __xen_send_IPI_mask(mask
, XEN_CALL_FUNCTION_VECTOR
);
578 /* Make sure other vcpus get a chance to run if they need to. */
579 for_each_cpu(cpu
, mask
) {
580 if (xen_vcpu_stolen(cpu
)) {
581 HYPERVISOR_sched_op(SCHEDOP_yield
, NULL
);
587 static void xen_smp_send_call_function_single_ipi(int cpu
)
589 __xen_send_IPI_mask(cpumask_of(cpu
),
590 XEN_CALL_FUNCTION_SINGLE_VECTOR
);
593 static inline int xen_map_vector(int vector
)
598 case RESCHEDULE_VECTOR
:
599 xen_vector
= XEN_RESCHEDULE_VECTOR
;
601 case CALL_FUNCTION_VECTOR
:
602 xen_vector
= XEN_CALL_FUNCTION_VECTOR
;
604 case CALL_FUNCTION_SINGLE_VECTOR
:
605 xen_vector
= XEN_CALL_FUNCTION_SINGLE_VECTOR
;
607 case IRQ_WORK_VECTOR
:
608 xen_vector
= XEN_IRQ_WORK_VECTOR
;
612 case APIC_DM_NMI
: /* Some use that instead of NMI_VECTOR */
613 xen_vector
= XEN_NMI_VECTOR
;
618 printk(KERN_ERR
"xen: vector 0x%x is not implemented\n",
625 void xen_send_IPI_mask(const struct cpumask
*mask
,
628 int xen_vector
= xen_map_vector(vector
);
631 __xen_send_IPI_mask(mask
, xen_vector
);
634 void xen_send_IPI_all(int vector
)
636 int xen_vector
= xen_map_vector(vector
);
639 __xen_send_IPI_mask(cpu_online_mask
, xen_vector
);
642 void xen_send_IPI_self(int vector
)
644 int xen_vector
= xen_map_vector(vector
);
647 xen_send_IPI_one(smp_processor_id(), xen_vector
);
650 void xen_send_IPI_mask_allbutself(const struct cpumask
*mask
,
654 unsigned int this_cpu
= smp_processor_id();
655 int xen_vector
= xen_map_vector(vector
);
657 if (!(num_online_cpus() > 1) || (xen_vector
< 0))
660 for_each_cpu_and(cpu
, mask
, cpu_online_mask
) {
664 xen_send_IPI_one(cpu
, xen_vector
);
668 void xen_send_IPI_allbutself(int vector
)
670 xen_send_IPI_mask_allbutself(cpu_online_mask
, vector
);
673 static irqreturn_t
xen_call_function_interrupt(int irq
, void *dev_id
)
676 generic_smp_call_function_interrupt();
677 inc_irq_stat(irq_call_count
);
683 static irqreturn_t
xen_call_function_single_interrupt(int irq
, void *dev_id
)
686 generic_smp_call_function_single_interrupt();
687 inc_irq_stat(irq_call_count
);
693 static irqreturn_t
xen_irq_work_interrupt(int irq
, void *dev_id
)
697 inc_irq_stat(apic_irq_work_irqs
);
703 static const struct smp_ops xen_smp_ops __initconst
= {
704 .smp_prepare_boot_cpu
= xen_smp_prepare_boot_cpu
,
705 .smp_prepare_cpus
= xen_smp_prepare_cpus
,
706 .smp_cpus_done
= xen_smp_cpus_done
,
708 .cpu_up
= xen_cpu_up
,
709 .cpu_die
= xen_cpu_die
,
710 .cpu_disable
= xen_cpu_disable
,
711 .play_dead
= xen_play_dead
,
713 .stop_other_cpus
= xen_stop_other_cpus
,
714 .smp_send_reschedule
= xen_smp_send_reschedule
,
716 .send_call_func_ipi
= xen_smp_send_call_function_ipi
,
717 .send_call_func_single_ipi
= xen_smp_send_call_function_single_ipi
,
720 void __init
xen_smp_init(void)
722 smp_ops
= xen_smp_ops
;
723 xen_fill_possible_map();
726 static void __init
xen_hvm_smp_prepare_cpus(unsigned int max_cpus
)
728 native_smp_prepare_cpus(max_cpus
);
729 WARN_ON(xen_smp_intr_init(0));
731 xen_init_lock_cpu(0);
734 static int xen_hvm_cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
738 * xen_smp_intr_init() needs to run before native_cpu_up()
739 * so that IPI vectors are set up on the booting CPU before
740 * it is marked online in native_cpu_up().
742 rc
= xen_smp_intr_init(cpu
);
745 rc
= native_cpu_up(cpu
, tidle
);
748 * We must initialize the slowpath CPU kicker _after_ the native
749 * path has executed. If we initialized it before none of the
750 * unlocker IPI kicks would reach the booting CPU as the booting
751 * CPU had not set itself 'online' in cpu_online_mask. That mask
752 * is checked when IPIs are sent (on HVM at least).
754 xen_init_lock_cpu(cpu
);
758 static void xen_hvm_cpu_die(unsigned int cpu
)
764 void __init
xen_hvm_smp_init(void)
766 if (!xen_have_vector_callback
)
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_up
= xen_hvm_cpu_up
;
771 smp_ops
.cpu_die
= xen_hvm_cpu_die
;
772 smp_ops
.send_call_func_ipi
= xen_smp_send_call_function_ipi
;
773 smp_ops
.send_call_func_single_ipi
= xen_smp_send_call_function_single_ipi
;
774 smp_ops
.smp_prepare_boot_cpu
= xen_smp_prepare_boot_cpu
;