Merge git://git.kernel.org/pub/scm/linux/kernel/git/wim/linux-2.6-watchdog
[linux-2.6] / 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/smp.h>
18
19 #include <asm/paravirt.h>
20 #include <asm/desc.h>
21 #include <asm/pgtable.h>
22 #include <asm/cpu.h>
23
24 #include <xen/interface/xen.h>
25 #include <xen/interface/vcpu.h>
26
27 #include <asm/xen/interface.h>
28 #include <asm/xen/hypercall.h>
29
30 #include <xen/page.h>
31 #include <xen/events.h>
32
33 #include "xen-ops.h"
34 #include "mmu.h"
35
36 cpumask_var_t xen_cpu_initialized_map;
37
38 static DEFINE_PER_CPU(int, resched_irq);
39 static DEFINE_PER_CPU(int, callfunc_irq);
40 static DEFINE_PER_CPU(int, callfuncsingle_irq);
41 static DEFINE_PER_CPU(int, debug_irq) = -1;
42
43 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
44 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
45
46 /*
47  * Reschedule call back. Nothing to do,
48  * all the work is done automatically when
49  * we return from the interrupt.
50  */
51 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
52 {
53 #ifdef CONFIG_X86_32
54         __get_cpu_var(irq_stat).irq_resched_count++;
55 #else
56         add_pda(irq_resched_count, 1);
57 #endif
58
59         return IRQ_HANDLED;
60 }
61
62 static __cpuinit void cpu_bringup(void)
63 {
64         int cpu = smp_processor_id();
65
66         cpu_init();
67         touch_softlockup_watchdog();
68         preempt_disable();
69
70         xen_enable_sysenter();
71         xen_enable_syscall();
72
73         cpu = smp_processor_id();
74         smp_store_cpu_info(cpu);
75         cpu_data(cpu).x86_max_cores = 1;
76         set_cpu_sibling_map(cpu);
77
78         xen_setup_cpu_clockevents();
79
80         cpu_set(cpu, cpu_online_map);
81         x86_write_percpu(cpu_state, CPU_ONLINE);
82         wmb();
83
84         /* We can take interrupts now: we're officially "up". */
85         local_irq_enable();
86
87         wmb();                  /* make sure everything is out */
88 }
89
90 static __cpuinit void cpu_bringup_and_idle(void)
91 {
92         cpu_bringup();
93         cpu_idle();
94 }
95
96 static int xen_smp_intr_init(unsigned int cpu)
97 {
98         int rc;
99         const char *resched_name, *callfunc_name, *debug_name;
100
101         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
102         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
103                                     cpu,
104                                     xen_reschedule_interrupt,
105                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
106                                     resched_name,
107                                     NULL);
108         if (rc < 0)
109                 goto fail;
110         per_cpu(resched_irq, cpu) = rc;
111
112         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
113         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
114                                     cpu,
115                                     xen_call_function_interrupt,
116                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
117                                     callfunc_name,
118                                     NULL);
119         if (rc < 0)
120                 goto fail;
121         per_cpu(callfunc_irq, cpu) = rc;
122
123         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
124         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
125                                      IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING,
126                                      debug_name, NULL);
127         if (rc < 0)
128                 goto fail;
129         per_cpu(debug_irq, cpu) = rc;
130
131         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
132         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
133                                     cpu,
134                                     xen_call_function_single_interrupt,
135                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
136                                     callfunc_name,
137                                     NULL);
138         if (rc < 0)
139                 goto fail;
140         per_cpu(callfuncsingle_irq, cpu) = rc;
141
142         return 0;
143
144  fail:
145         if (per_cpu(resched_irq, cpu) >= 0)
146                 unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL);
147         if (per_cpu(callfunc_irq, cpu) >= 0)
148                 unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL);
149         if (per_cpu(debug_irq, cpu) >= 0)
150                 unbind_from_irqhandler(per_cpu(debug_irq, cpu), NULL);
151         if (per_cpu(callfuncsingle_irq, cpu) >= 0)
152                 unbind_from_irqhandler(per_cpu(callfuncsingle_irq, cpu), NULL);
153
154         return rc;
155 }
156
157 static void __init xen_fill_possible_map(void)
158 {
159         int i, rc;
160
161         for (i = 0; i < nr_cpu_ids; i++) {
162                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
163                 if (rc >= 0) {
164                         num_processors++;
165                         cpu_set(i, cpu_possible_map);
166                 }
167         }
168 }
169
170 static void __init xen_smp_prepare_boot_cpu(void)
171 {
172         BUG_ON(smp_processor_id() != 0);
173         native_smp_prepare_boot_cpu();
174
175         /* We've switched to the "real" per-cpu gdt, so make sure the
176            old memory can be recycled */
177         make_lowmem_page_readwrite(&per_cpu_var(gdt_page));
178
179         xen_setup_vcpu_info_placement();
180 }
181
182 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
183 {
184         unsigned cpu;
185
186         xen_init_lock_cpu(0);
187
188         smp_store_cpu_info(0);
189         cpu_data(0).x86_max_cores = 1;
190         set_cpu_sibling_map(0);
191
192         if (xen_smp_intr_init(0))
193                 BUG();
194
195         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
196                 panic("could not allocate xen_cpu_initialized_map\n");
197
198         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
199
200         /* Restrict the possible_map according to max_cpus. */
201         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
202                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
203                         continue;
204                 cpu_clear(cpu, cpu_possible_map);
205         }
206
207         for_each_possible_cpu (cpu) {
208                 struct task_struct *idle;
209
210                 if (cpu == 0)
211                         continue;
212
213                 idle = fork_idle(cpu);
214                 if (IS_ERR(idle))
215                         panic("failed fork for CPU %d", cpu);
216
217                 cpu_set(cpu, cpu_present_map);
218         }
219 }
220
221 static __cpuinit int
222 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
223 {
224         struct vcpu_guest_context *ctxt;
225         struct desc_struct *gdt;
226
227         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
228                 return 0;
229
230         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
231         if (ctxt == NULL)
232                 return -ENOMEM;
233
234         gdt = get_cpu_gdt_table(cpu);
235
236         ctxt->flags = VGCF_IN_KERNEL;
237         ctxt->user_regs.ds = __USER_DS;
238         ctxt->user_regs.es = __USER_DS;
239         ctxt->user_regs.ss = __KERNEL_DS;
240 #ifdef CONFIG_X86_32
241         ctxt->user_regs.fs = __KERNEL_PERCPU;
242 #endif
243         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
244         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
245
246         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
247
248         xen_copy_trap_info(ctxt->trap_ctxt);
249
250         ctxt->ldt_ents = 0;
251
252         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
253         make_lowmem_page_readonly(gdt);
254
255         ctxt->gdt_frames[0] = virt_to_mfn(gdt);
256         ctxt->gdt_ents      = GDT_ENTRIES;
257
258         ctxt->user_regs.cs = __KERNEL_CS;
259         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
260
261         ctxt->kernel_ss = __KERNEL_DS;
262         ctxt->kernel_sp = idle->thread.sp0;
263
264 #ifdef CONFIG_X86_32
265         ctxt->event_callback_cs     = __KERNEL_CS;
266         ctxt->failsafe_callback_cs  = __KERNEL_CS;
267 #endif
268         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
269         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
270
271         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
272         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
273
274         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
275                 BUG();
276
277         kfree(ctxt);
278         return 0;
279 }
280
281 static int __cpuinit xen_cpu_up(unsigned int cpu)
282 {
283         struct task_struct *idle = idle_task(cpu);
284         int rc;
285
286 #ifdef CONFIG_X86_64
287         /* Allocate node local memory for AP pdas */
288         WARN_ON(cpu == 0);
289         if (cpu > 0) {
290                 rc = get_local_pda(cpu);
291                 if (rc)
292                         return rc;
293         }
294 #endif
295
296 #ifdef CONFIG_X86_32
297         init_gdt(cpu);
298         per_cpu(current_task, cpu) = idle;
299         irq_ctx_init(cpu);
300 #else
301         cpu_pda(cpu)->pcurrent = idle;
302         clear_tsk_thread_flag(idle, TIF_FORK);
303 #endif
304         xen_setup_timer(cpu);
305         xen_init_lock_cpu(cpu);
306
307         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
308
309         /* make sure interrupts start blocked */
310         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
311
312         rc = cpu_initialize_context(cpu, idle);
313         if (rc)
314                 return rc;
315
316         if (num_online_cpus() == 1)
317                 alternatives_smp_switch(1);
318
319         rc = xen_smp_intr_init(cpu);
320         if (rc)
321                 return rc;
322
323         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
324         BUG_ON(rc);
325
326         while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
327                 HYPERVISOR_sched_op(SCHEDOP_yield, 0);
328                 barrier();
329         }
330
331         return 0;
332 }
333
334 static void xen_smp_cpus_done(unsigned int max_cpus)
335 {
336 }
337
338 #ifdef CONFIG_HOTPLUG_CPU
339 static int xen_cpu_disable(void)
340 {
341         unsigned int cpu = smp_processor_id();
342         if (cpu == 0)
343                 return -EBUSY;
344
345         cpu_disable_common();
346
347         load_cr3(swapper_pg_dir);
348         return 0;
349 }
350
351 static void xen_cpu_die(unsigned int cpu)
352 {
353         while (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
354                 current->state = TASK_UNINTERRUPTIBLE;
355                 schedule_timeout(HZ/10);
356         }
357         unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL);
358         unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL);
359         unbind_from_irqhandler(per_cpu(debug_irq, cpu), NULL);
360         unbind_from_irqhandler(per_cpu(callfuncsingle_irq, cpu), NULL);
361         xen_uninit_lock_cpu(cpu);
362         xen_teardown_timer(cpu);
363
364         if (num_online_cpus() == 1)
365                 alternatives_smp_switch(0);
366 }
367
368 static void __cpuinit xen_play_dead(void) /* used only with CPU_HOTPLUG */
369 {
370         play_dead_common();
371         HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
372         cpu_bringup();
373 }
374
375 #else /* !CONFIG_HOTPLUG_CPU */
376 static int xen_cpu_disable(void)
377 {
378         return -ENOSYS;
379 }
380
381 static void xen_cpu_die(unsigned int cpu)
382 {
383         BUG();
384 }
385
386 static void xen_play_dead(void)
387 {
388         BUG();
389 }
390
391 #endif
392 static void stop_self(void *v)
393 {
394         int cpu = smp_processor_id();
395
396         /* make sure we're not pinning something down */
397         load_cr3(swapper_pg_dir);
398         /* should set up a minimal gdt */
399
400         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
401         BUG();
402 }
403
404 static void xen_smp_send_stop(void)
405 {
406         smp_call_function(stop_self, NULL, 0);
407 }
408
409 static void xen_smp_send_reschedule(int cpu)
410 {
411         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
412 }
413
414 static void xen_send_IPI_mask(const struct cpumask *mask,
415                               enum ipi_vector vector)
416 {
417         unsigned cpu;
418
419         for_each_cpu_and(cpu, mask, cpu_online_mask)
420                 xen_send_IPI_one(cpu, vector);
421 }
422
423 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
424 {
425         int cpu;
426
427         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
428
429         /* Make sure other vcpus get a chance to run if they need to. */
430         for_each_cpu(cpu, mask) {
431                 if (xen_vcpu_stolen(cpu)) {
432                         HYPERVISOR_sched_op(SCHEDOP_yield, 0);
433                         break;
434                 }
435         }
436 }
437
438 static void xen_smp_send_call_function_single_ipi(int cpu)
439 {
440         xen_send_IPI_mask(cpumask_of(cpu),
441                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
442 }
443
444 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
445 {
446         irq_enter();
447         generic_smp_call_function_interrupt();
448 #ifdef CONFIG_X86_32
449         __get_cpu_var(irq_stat).irq_call_count++;
450 #else
451         add_pda(irq_call_count, 1);
452 #endif
453         irq_exit();
454
455         return IRQ_HANDLED;
456 }
457
458 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
459 {
460         irq_enter();
461         generic_smp_call_function_single_interrupt();
462 #ifdef CONFIG_X86_32
463         __get_cpu_var(irq_stat).irq_call_count++;
464 #else
465         add_pda(irq_call_count, 1);
466 #endif
467         irq_exit();
468
469         return IRQ_HANDLED;
470 }
471
472 static const struct smp_ops xen_smp_ops __initdata = {
473         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
474         .smp_prepare_cpus = xen_smp_prepare_cpus,
475         .smp_cpus_done = xen_smp_cpus_done,
476
477         .cpu_up = xen_cpu_up,
478         .cpu_die = xen_cpu_die,
479         .cpu_disable = xen_cpu_disable,
480         .play_dead = xen_play_dead,
481
482         .smp_send_stop = xen_smp_send_stop,
483         .smp_send_reschedule = xen_smp_send_reschedule,
484
485         .send_call_func_ipi = xen_smp_send_call_function_ipi,
486         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
487 };
488
489 void __init xen_smp_init(void)
490 {
491         smp_ops = xen_smp_ops;
492         xen_fill_possible_map();
493         xen_init_spinlocks();
494 }