Merge tag 'stable/for-linus-3.4-tag-two' of git://git.kernel.org/pub/scm/linux/kernel...
[~shefty/rdma-dev.git] / arch / x86 / xen / smp.c
1 /*
2  * Xen SMP support
3  *
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.
7  *
8  * IPIs are handled through the Xen event mechanism.
9  *
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
13  * single-threaded.
14  */
15 #include <linux/sched.h>
16 #include <linux/err.h>
17 #include <linux/slab.h>
18 #include <linux/smp.h>
19
20 #include <asm/paravirt.h>
21 #include <asm/desc.h>
22 #include <asm/pgtable.h>
23 #include <asm/cpu.h>
24
25 #include <xen/interface/xen.h>
26 #include <xen/interface/vcpu.h>
27
28 #include <asm/xen/interface.h>
29 #include <asm/xen/hypercall.h>
30
31 #include <xen/xen.h>
32 #include <xen/page.h>
33 #include <xen/events.h>
34
35 #include <xen/hvc-console.h>
36 #include "xen-ops.h"
37 #include "mmu.h"
38
39 cpumask_var_t xen_cpu_initialized_map;
40
41 static DEFINE_PER_CPU(int, xen_resched_irq);
42 static DEFINE_PER_CPU(int, xen_callfunc_irq);
43 static DEFINE_PER_CPU(int, xen_callfuncsingle_irq);
44 static DEFINE_PER_CPU(int, xen_debug_irq) = -1;
45
46 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
47 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
48
49 /*
50  * Reschedule call back.
51  */
52 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
53 {
54         inc_irq_stat(irq_resched_count);
55         scheduler_ipi();
56
57         return IRQ_HANDLED;
58 }
59
60 static void __cpuinit cpu_bringup(void)
61 {
62         int cpu = smp_processor_id();
63
64         cpu_init();
65         touch_softlockup_watchdog();
66         preempt_disable();
67
68         xen_enable_sysenter();
69         xen_enable_syscall();
70
71         cpu = smp_processor_id();
72         smp_store_cpu_info(cpu);
73         cpu_data(cpu).x86_max_cores = 1;
74         set_cpu_sibling_map(cpu);
75
76         xen_setup_cpu_clockevents();
77
78         notify_cpu_starting(cpu);
79
80         ipi_call_lock();
81         set_cpu_online(cpu, true);
82         ipi_call_unlock();
83
84         this_cpu_write(cpu_state, CPU_ONLINE);
85
86         wmb();
87
88         /* We can take interrupts now: we're officially "up". */
89         local_irq_enable();
90
91         wmb();                  /* make sure everything is out */
92 }
93
94 static void __cpuinit cpu_bringup_and_idle(void)
95 {
96         cpu_bringup();
97         cpu_idle();
98 }
99
100 static int xen_smp_intr_init(unsigned int cpu)
101 {
102         int rc;
103         const char *resched_name, *callfunc_name, *debug_name;
104
105         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
106         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
107                                     cpu,
108                                     xen_reschedule_interrupt,
109                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
110                                     resched_name,
111                                     NULL);
112         if (rc < 0)
113                 goto fail;
114         per_cpu(xen_resched_irq, cpu) = rc;
115
116         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
117         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
118                                     cpu,
119                                     xen_call_function_interrupt,
120                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
121                                     callfunc_name,
122                                     NULL);
123         if (rc < 0)
124                 goto fail;
125         per_cpu(xen_callfunc_irq, cpu) = rc;
126
127         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
128         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
129                                      IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING,
130                                      debug_name, NULL);
131         if (rc < 0)
132                 goto fail;
133         per_cpu(xen_debug_irq, cpu) = rc;
134
135         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
136         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
137                                     cpu,
138                                     xen_call_function_single_interrupt,
139                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
140                                     callfunc_name,
141                                     NULL);
142         if (rc < 0)
143                 goto fail;
144         per_cpu(xen_callfuncsingle_irq, cpu) = rc;
145
146         return 0;
147
148  fail:
149         if (per_cpu(xen_resched_irq, cpu) >= 0)
150                 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
151         if (per_cpu(xen_callfunc_irq, cpu) >= 0)
152                 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
153         if (per_cpu(xen_debug_irq, cpu) >= 0)
154                 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
155         if (per_cpu(xen_callfuncsingle_irq, cpu) >= 0)
156                 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu),
157                                        NULL);
158
159         return rc;
160 }
161
162 static void __init xen_fill_possible_map(void)
163 {
164         int i, rc;
165
166         if (xen_initial_domain())
167                 return;
168
169         for (i = 0; i < nr_cpu_ids; i++) {
170                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
171                 if (rc >= 0) {
172                         num_processors++;
173                         set_cpu_possible(i, true);
174                 }
175         }
176 }
177
178 static void __init xen_filter_cpu_maps(void)
179 {
180         int i, rc;
181
182         if (!xen_initial_domain())
183                 return;
184
185         num_processors = 0;
186         disabled_cpus = 0;
187         for (i = 0; i < nr_cpu_ids; i++) {
188                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
189                 if (rc >= 0) {
190                         num_processors++;
191                         set_cpu_possible(i, true);
192                 } else {
193                         set_cpu_possible(i, false);
194                         set_cpu_present(i, false);
195                 }
196         }
197 }
198
199 static void __init xen_smp_prepare_boot_cpu(void)
200 {
201         BUG_ON(smp_processor_id() != 0);
202         native_smp_prepare_boot_cpu();
203
204         /* We've switched to the "real" per-cpu gdt, so make sure the
205            old memory can be recycled */
206         make_lowmem_page_readwrite(xen_initial_gdt);
207
208         xen_filter_cpu_maps();
209         xen_setup_vcpu_info_placement();
210 }
211
212 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
213 {
214         unsigned cpu;
215         unsigned int i;
216
217         if (skip_ioapic_setup) {
218                 char *m = (max_cpus == 0) ?
219                         "The nosmp parameter is incompatible with Xen; " \
220                         "use Xen dom0_max_vcpus=1 parameter" :
221                         "The noapic parameter is incompatible with Xen";
222
223                 xen_raw_printk(m);
224                 panic(m);
225         }
226         xen_init_lock_cpu(0);
227
228         smp_store_cpu_info(0);
229         cpu_data(0).x86_max_cores = 1;
230
231         for_each_possible_cpu(i) {
232                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
233                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
234                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
235         }
236         set_cpu_sibling_map(0);
237
238         if (xen_smp_intr_init(0))
239                 BUG();
240
241         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
242                 panic("could not allocate xen_cpu_initialized_map\n");
243
244         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
245
246         /* Restrict the possible_map according to max_cpus. */
247         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
248                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
249                         continue;
250                 set_cpu_possible(cpu, false);
251         }
252
253         for_each_possible_cpu (cpu) {
254                 struct task_struct *idle;
255
256                 if (cpu == 0)
257                         continue;
258
259                 idle = fork_idle(cpu);
260                 if (IS_ERR(idle))
261                         panic("failed fork for CPU %d", cpu);
262
263                 set_cpu_present(cpu, true);
264         }
265 }
266
267 static int __cpuinit
268 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
269 {
270         struct vcpu_guest_context *ctxt;
271         struct desc_struct *gdt;
272         unsigned long gdt_mfn;
273
274         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
275                 return 0;
276
277         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
278         if (ctxt == NULL)
279                 return -ENOMEM;
280
281         gdt = get_cpu_gdt_table(cpu);
282
283         ctxt->flags = VGCF_IN_KERNEL;
284         ctxt->user_regs.ds = __USER_DS;
285         ctxt->user_regs.es = __USER_DS;
286         ctxt->user_regs.ss = __KERNEL_DS;
287 #ifdef CONFIG_X86_32
288         ctxt->user_regs.fs = __KERNEL_PERCPU;
289         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
290 #else
291         ctxt->gs_base_kernel = per_cpu_offset(cpu);
292 #endif
293         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
294         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
295
296         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
297
298         xen_copy_trap_info(ctxt->trap_ctxt);
299
300         ctxt->ldt_ents = 0;
301
302         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
303
304         gdt_mfn = arbitrary_virt_to_mfn(gdt);
305         make_lowmem_page_readonly(gdt);
306         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
307
308         ctxt->gdt_frames[0] = gdt_mfn;
309         ctxt->gdt_ents      = GDT_ENTRIES;
310
311         ctxt->user_regs.cs = __KERNEL_CS;
312         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
313
314         ctxt->kernel_ss = __KERNEL_DS;
315         ctxt->kernel_sp = idle->thread.sp0;
316
317 #ifdef CONFIG_X86_32
318         ctxt->event_callback_cs     = __KERNEL_CS;
319         ctxt->failsafe_callback_cs  = __KERNEL_CS;
320 #endif
321         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
322         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
323
324         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
325         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
326
327         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
328                 BUG();
329
330         kfree(ctxt);
331         return 0;
332 }
333
334 static int __cpuinit xen_cpu_up(unsigned int cpu)
335 {
336         struct task_struct *idle = idle_task(cpu);
337         int rc;
338
339         per_cpu(current_task, cpu) = idle;
340 #ifdef CONFIG_X86_32
341         irq_ctx_init(cpu);
342 #else
343         clear_tsk_thread_flag(idle, TIF_FORK);
344         per_cpu(kernel_stack, cpu) =
345                 (unsigned long)task_stack_page(idle) -
346                 KERNEL_STACK_OFFSET + THREAD_SIZE;
347 #endif
348         xen_setup_runstate_info(cpu);
349         xen_setup_timer(cpu);
350         xen_init_lock_cpu(cpu);
351
352         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
353
354         /* make sure interrupts start blocked */
355         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
356
357         rc = cpu_initialize_context(cpu, idle);
358         if (rc)
359                 return rc;
360
361         if (num_online_cpus() == 1)
362                 alternatives_smp_switch(1);
363
364         rc = xen_smp_intr_init(cpu);
365         if (rc)
366                 return rc;
367
368         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
369         BUG_ON(rc);
370
371         while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
372                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
373                 barrier();
374         }
375
376         return 0;
377 }
378
379 static void xen_smp_cpus_done(unsigned int max_cpus)
380 {
381 }
382
383 #ifdef CONFIG_HOTPLUG_CPU
384 static int xen_cpu_disable(void)
385 {
386         unsigned int cpu = smp_processor_id();
387         if (cpu == 0)
388                 return -EBUSY;
389
390         cpu_disable_common();
391
392         load_cr3(swapper_pg_dir);
393         return 0;
394 }
395
396 static void xen_cpu_die(unsigned int cpu)
397 {
398         while (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
399                 current->state = TASK_UNINTERRUPTIBLE;
400                 schedule_timeout(HZ/10);
401         }
402         unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
403         unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
404         unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
405         unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL);
406         xen_uninit_lock_cpu(cpu);
407         xen_teardown_timer(cpu);
408
409         if (num_online_cpus() == 1)
410                 alternatives_smp_switch(0);
411 }
412
413 static void __cpuinit xen_play_dead(void) /* used only with HOTPLUG_CPU */
414 {
415         play_dead_common();
416         HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
417         cpu_bringup();
418         /*
419          * Balance out the preempt calls - as we are running in cpu_idle
420          * loop which has been called at bootup from cpu_bringup_and_idle.
421          * The cpucpu_bringup_and_idle called cpu_bringup which made a
422          * preempt_disable() So this preempt_enable will balance it out.
423          */
424         preempt_enable();
425 }
426
427 #else /* !CONFIG_HOTPLUG_CPU */
428 static int xen_cpu_disable(void)
429 {
430         return -ENOSYS;
431 }
432
433 static void xen_cpu_die(unsigned int cpu)
434 {
435         BUG();
436 }
437
438 static void xen_play_dead(void)
439 {
440         BUG();
441 }
442
443 #endif
444 static void stop_self(void *v)
445 {
446         int cpu = smp_processor_id();
447
448         /* make sure we're not pinning something down */
449         load_cr3(swapper_pg_dir);
450         /* should set up a minimal gdt */
451
452         set_cpu_online(cpu, false);
453
454         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
455         BUG();
456 }
457
458 static void xen_stop_other_cpus(int wait)
459 {
460         smp_call_function(stop_self, NULL, wait);
461 }
462
463 static void xen_smp_send_reschedule(int cpu)
464 {
465         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
466 }
467
468 static void xen_send_IPI_mask(const struct cpumask *mask,
469                               enum ipi_vector vector)
470 {
471         unsigned cpu;
472
473         for_each_cpu_and(cpu, mask, cpu_online_mask)
474                 xen_send_IPI_one(cpu, vector);
475 }
476
477 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
478 {
479         int cpu;
480
481         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
482
483         /* Make sure other vcpus get a chance to run if they need to. */
484         for_each_cpu(cpu, mask) {
485                 if (xen_vcpu_stolen(cpu)) {
486                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
487                         break;
488                 }
489         }
490 }
491
492 static void xen_smp_send_call_function_single_ipi(int cpu)
493 {
494         xen_send_IPI_mask(cpumask_of(cpu),
495                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
496 }
497
498 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
499 {
500         irq_enter();
501         generic_smp_call_function_interrupt();
502         inc_irq_stat(irq_call_count);
503         irq_exit();
504
505         return IRQ_HANDLED;
506 }
507
508 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
509 {
510         irq_enter();
511         generic_smp_call_function_single_interrupt();
512         inc_irq_stat(irq_call_count);
513         irq_exit();
514
515         return IRQ_HANDLED;
516 }
517
518 static const struct smp_ops xen_smp_ops __initconst = {
519         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
520         .smp_prepare_cpus = xen_smp_prepare_cpus,
521         .smp_cpus_done = xen_smp_cpus_done,
522
523         .cpu_up = xen_cpu_up,
524         .cpu_die = xen_cpu_die,
525         .cpu_disable = xen_cpu_disable,
526         .play_dead = xen_play_dead,
527
528         .stop_other_cpus = xen_stop_other_cpus,
529         .smp_send_reschedule = xen_smp_send_reschedule,
530
531         .send_call_func_ipi = xen_smp_send_call_function_ipi,
532         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
533 };
534
535 void __init xen_smp_init(void)
536 {
537         smp_ops = xen_smp_ops;
538         xen_fill_possible_map();
539         xen_init_spinlocks();
540 }
541
542 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
543 {
544         native_smp_prepare_cpus(max_cpus);
545         WARN_ON(xen_smp_intr_init(0));
546
547         xen_init_lock_cpu(0);
548 }
549
550 static int __cpuinit xen_hvm_cpu_up(unsigned int cpu)
551 {
552         int rc;
553         rc = native_cpu_up(cpu);
554         WARN_ON (xen_smp_intr_init(cpu));
555         return rc;
556 }
557
558 static void xen_hvm_cpu_die(unsigned int cpu)
559 {
560         unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
561         unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
562         unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
563         unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL);
564         native_cpu_die(cpu);
565 }
566
567 void __init xen_hvm_smp_init(void)
568 {
569         if (!xen_have_vector_callback)
570                 return;
571         smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
572         smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
573         smp_ops.cpu_up = xen_hvm_cpu_up;
574         smp_ops.cpu_die = xen_hvm_cpu_die;
575         smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
576         smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
577 }