Merge branch 'core-rcu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[pandora-kernel.git] / arch / x86 / oprofile / nmi_int.c
1 /**
2  * @file nmi_int.c
3  *
4  * @remark Copyright 2002-2009 OProfile authors
5  * @remark Read the file COPYING
6  *
7  * @author John Levon <levon@movementarian.org>
8  * @author Robert Richter <robert.richter@amd.com>
9  * @author Barry Kasindorf <barry.kasindorf@amd.com>
10  * @author Jason Yeh <jason.yeh@amd.com>
11  * @author Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
12  */
13
14 #include <linux/init.h>
15 #include <linux/notifier.h>
16 #include <linux/smp.h>
17 #include <linux/oprofile.h>
18 #include <linux/sysdev.h>
19 #include <linux/slab.h>
20 #include <linux/moduleparam.h>
21 #include <linux/kdebug.h>
22 #include <linux/cpu.h>
23 #include <asm/nmi.h>
24 #include <asm/msr.h>
25 #include <asm/apic.h>
26
27 #include "op_counter.h"
28 #include "op_x86_model.h"
29
30 static struct op_x86_model_spec *model;
31 static DEFINE_PER_CPU(struct op_msrs, cpu_msrs);
32 static DEFINE_PER_CPU(unsigned long, saved_lvtpc);
33
34 /* 0 == registered but off, 1 == registered and on */
35 static int nmi_enabled = 0;
36
37 struct op_counter_config counter_config[OP_MAX_COUNTER];
38
39 /* common functions */
40
41 u64 op_x86_get_ctrl(struct op_x86_model_spec const *model,
42                     struct op_counter_config *counter_config)
43 {
44         u64 val = 0;
45         u16 event = (u16)counter_config->event;
46
47         val |= ARCH_PERFMON_EVENTSEL_INT;
48         val |= counter_config->user ? ARCH_PERFMON_EVENTSEL_USR : 0;
49         val |= counter_config->kernel ? ARCH_PERFMON_EVENTSEL_OS : 0;
50         val |= (counter_config->unit_mask & 0xFF) << 8;
51         event &= model->event_mask ? model->event_mask : 0xFF;
52         val |= event & 0xFF;
53         val |= (event & 0x0F00) << 24;
54
55         return val;
56 }
57
58
59 static int profile_exceptions_notify(struct notifier_block *self,
60                                      unsigned long val, void *data)
61 {
62         struct die_args *args = (struct die_args *)data;
63         int ret = NOTIFY_DONE;
64         int cpu = smp_processor_id();
65
66         switch (val) {
67         case DIE_NMI:
68         case DIE_NMI_IPI:
69                 model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu));
70                 ret = NOTIFY_STOP;
71                 break;
72         default:
73                 break;
74         }
75         return ret;
76 }
77
78 static void nmi_cpu_save_registers(struct op_msrs *msrs)
79 {
80         struct op_msr *counters = msrs->counters;
81         struct op_msr *controls = msrs->controls;
82         unsigned int i;
83
84         for (i = 0; i < model->num_counters; ++i) {
85                 if (counters[i].addr)
86                         rdmsrl(counters[i].addr, counters[i].saved);
87         }
88
89         for (i = 0; i < model->num_controls; ++i) {
90                 if (controls[i].addr)
91                         rdmsrl(controls[i].addr, controls[i].saved);
92         }
93 }
94
95 static void nmi_cpu_start(void *dummy)
96 {
97         struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
98         model->start(msrs);
99 }
100
101 static int nmi_start(void)
102 {
103         on_each_cpu(nmi_cpu_start, NULL, 1);
104         return 0;
105 }
106
107 static void nmi_cpu_stop(void *dummy)
108 {
109         struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
110         model->stop(msrs);
111 }
112
113 static void nmi_stop(void)
114 {
115         on_each_cpu(nmi_cpu_stop, NULL, 1);
116 }
117
118 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
119
120 static DEFINE_PER_CPU(int, switch_index);
121
122 static inline int has_mux(void)
123 {
124         return !!model->switch_ctrl;
125 }
126
127 inline int op_x86_phys_to_virt(int phys)
128 {
129         return __get_cpu_var(switch_index) + phys;
130 }
131
132 inline int op_x86_virt_to_phys(int virt)
133 {
134         return virt % model->num_counters;
135 }
136
137 static void nmi_shutdown_mux(void)
138 {
139         int i;
140
141         if (!has_mux())
142                 return;
143
144         for_each_possible_cpu(i) {
145                 kfree(per_cpu(cpu_msrs, i).multiplex);
146                 per_cpu(cpu_msrs, i).multiplex = NULL;
147                 per_cpu(switch_index, i) = 0;
148         }
149 }
150
151 static int nmi_setup_mux(void)
152 {
153         size_t multiplex_size =
154                 sizeof(struct op_msr) * model->num_virt_counters;
155         int i;
156
157         if (!has_mux())
158                 return 1;
159
160         for_each_possible_cpu(i) {
161                 per_cpu(cpu_msrs, i).multiplex =
162                         kzalloc(multiplex_size, GFP_KERNEL);
163                 if (!per_cpu(cpu_msrs, i).multiplex)
164                         return 0;
165         }
166
167         return 1;
168 }
169
170 static void nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs)
171 {
172         int i;
173         struct op_msr *multiplex = msrs->multiplex;
174
175         if (!has_mux())
176                 return;
177
178         for (i = 0; i < model->num_virt_counters; ++i) {
179                 if (counter_config[i].enabled) {
180                         multiplex[i].saved = -(u64)counter_config[i].count;
181                 } else {
182                         multiplex[i].saved = 0;
183                 }
184         }
185
186         per_cpu(switch_index, cpu) = 0;
187 }
188
189 static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs)
190 {
191         struct op_msr *counters = msrs->counters;
192         struct op_msr *multiplex = msrs->multiplex;
193         int i;
194
195         for (i = 0; i < model->num_counters; ++i) {
196                 int virt = op_x86_phys_to_virt(i);
197                 if (counters[i].addr)
198                         rdmsrl(counters[i].addr, multiplex[virt].saved);
199         }
200 }
201
202 static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs)
203 {
204         struct op_msr *counters = msrs->counters;
205         struct op_msr *multiplex = msrs->multiplex;
206         int i;
207
208         for (i = 0; i < model->num_counters; ++i) {
209                 int virt = op_x86_phys_to_virt(i);
210                 if (counters[i].addr)
211                         wrmsrl(counters[i].addr, multiplex[virt].saved);
212         }
213 }
214
215 static void nmi_cpu_switch(void *dummy)
216 {
217         int cpu = smp_processor_id();
218         int si = per_cpu(switch_index, cpu);
219         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
220
221         nmi_cpu_stop(NULL);
222         nmi_cpu_save_mpx_registers(msrs);
223
224         /* move to next set */
225         si += model->num_counters;
226         if ((si >= model->num_virt_counters) || (counter_config[si].count == 0))
227                 per_cpu(switch_index, cpu) = 0;
228         else
229                 per_cpu(switch_index, cpu) = si;
230
231         model->switch_ctrl(model, msrs);
232         nmi_cpu_restore_mpx_registers(msrs);
233
234         nmi_cpu_start(NULL);
235 }
236
237
238 /*
239  * Quick check to see if multiplexing is necessary.
240  * The check should be sufficient since counters are used
241  * in ordre.
242  */
243 static int nmi_multiplex_on(void)
244 {
245         return counter_config[model->num_counters].count ? 0 : -EINVAL;
246 }
247
248 static int nmi_switch_event(void)
249 {
250         if (!has_mux())
251                 return -ENOSYS;         /* not implemented */
252         if (nmi_multiplex_on() < 0)
253                 return -EINVAL;         /* not necessary */
254
255         on_each_cpu(nmi_cpu_switch, NULL, 1);
256
257         return 0;
258 }
259
260 static inline void mux_init(struct oprofile_operations *ops)
261 {
262         if (has_mux())
263                 ops->switch_events = nmi_switch_event;
264 }
265
266 static void mux_clone(int cpu)
267 {
268         if (!has_mux())
269                 return;
270
271         memcpy(per_cpu(cpu_msrs, cpu).multiplex,
272                per_cpu(cpu_msrs, 0).multiplex,
273                sizeof(struct op_msr) * model->num_virt_counters);
274 }
275
276 #else
277
278 inline int op_x86_phys_to_virt(int phys) { return phys; }
279 inline int op_x86_virt_to_phys(int virt) { return virt; }
280 static inline void nmi_shutdown_mux(void) { }
281 static inline int nmi_setup_mux(void) { return 1; }
282 static inline void
283 nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs) { }
284 static inline void mux_init(struct oprofile_operations *ops) { }
285 static void mux_clone(int cpu) { }
286
287 #endif
288
289 static void free_msrs(void)
290 {
291         int i;
292         for_each_possible_cpu(i) {
293                 kfree(per_cpu(cpu_msrs, i).counters);
294                 per_cpu(cpu_msrs, i).counters = NULL;
295                 kfree(per_cpu(cpu_msrs, i).controls);
296                 per_cpu(cpu_msrs, i).controls = NULL;
297         }
298 }
299
300 static int allocate_msrs(void)
301 {
302         size_t controls_size = sizeof(struct op_msr) * model->num_controls;
303         size_t counters_size = sizeof(struct op_msr) * model->num_counters;
304
305         int i;
306         for_each_possible_cpu(i) {
307                 per_cpu(cpu_msrs, i).counters = kzalloc(counters_size,
308                                                         GFP_KERNEL);
309                 if (!per_cpu(cpu_msrs, i).counters)
310                         return 0;
311                 per_cpu(cpu_msrs, i).controls = kzalloc(controls_size,
312                                                         GFP_KERNEL);
313                 if (!per_cpu(cpu_msrs, i).controls)
314                         return 0;
315         }
316
317         return 1;
318 }
319
320 static void nmi_cpu_setup(void *dummy)
321 {
322         int cpu = smp_processor_id();
323         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
324         nmi_cpu_save_registers(msrs);
325         spin_lock(&oprofilefs_lock);
326         model->setup_ctrs(model, msrs);
327         nmi_cpu_setup_mux(cpu, msrs);
328         spin_unlock(&oprofilefs_lock);
329         per_cpu(saved_lvtpc, cpu) = apic_read(APIC_LVTPC);
330         apic_write(APIC_LVTPC, APIC_DM_NMI);
331 }
332
333 static struct notifier_block profile_exceptions_nb = {
334         .notifier_call = profile_exceptions_notify,
335         .next = NULL,
336         .priority = 2
337 };
338
339 static int nmi_setup(void)
340 {
341         int err = 0;
342         int cpu;
343
344         if (!allocate_msrs())
345                 err = -ENOMEM;
346         else if (!nmi_setup_mux())
347                 err = -ENOMEM;
348         else
349                 err = register_die_notifier(&profile_exceptions_nb);
350
351         if (err) {
352                 free_msrs();
353                 nmi_shutdown_mux();
354                 return err;
355         }
356
357         /* We need to serialize save and setup for HT because the subset
358          * of msrs are distinct for save and setup operations
359          */
360
361         /* Assume saved/restored counters are the same on all CPUs */
362         model->fill_in_addresses(&per_cpu(cpu_msrs, 0));
363         for_each_possible_cpu(cpu) {
364                 if (!cpu)
365                         continue;
366
367                 memcpy(per_cpu(cpu_msrs, cpu).counters,
368                        per_cpu(cpu_msrs, 0).counters,
369                        sizeof(struct op_msr) * model->num_counters);
370
371                 memcpy(per_cpu(cpu_msrs, cpu).controls,
372                        per_cpu(cpu_msrs, 0).controls,
373                        sizeof(struct op_msr) * model->num_controls);
374
375                 mux_clone(cpu);
376         }
377         on_each_cpu(nmi_cpu_setup, NULL, 1);
378         nmi_enabled = 1;
379         return 0;
380 }
381
382 static void nmi_cpu_restore_registers(struct op_msrs *msrs)
383 {
384         struct op_msr *counters = msrs->counters;
385         struct op_msr *controls = msrs->controls;
386         unsigned int i;
387
388         for (i = 0; i < model->num_controls; ++i) {
389                 if (controls[i].addr)
390                         wrmsrl(controls[i].addr, controls[i].saved);
391         }
392
393         for (i = 0; i < model->num_counters; ++i) {
394                 if (counters[i].addr)
395                         wrmsrl(counters[i].addr, counters[i].saved);
396         }
397 }
398
399 static void nmi_cpu_shutdown(void *dummy)
400 {
401         unsigned int v;
402         int cpu = smp_processor_id();
403         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
404
405         /* restoring APIC_LVTPC can trigger an apic error because the delivery
406          * mode and vector nr combination can be illegal. That's by design: on
407          * power on apic lvt contain a zero vector nr which are legal only for
408          * NMI delivery mode. So inhibit apic err before restoring lvtpc
409          */
410         v = apic_read(APIC_LVTERR);
411         apic_write(APIC_LVTERR, v | APIC_LVT_MASKED);
412         apic_write(APIC_LVTPC, per_cpu(saved_lvtpc, cpu));
413         apic_write(APIC_LVTERR, v);
414         nmi_cpu_restore_registers(msrs);
415 }
416
417 static void nmi_shutdown(void)
418 {
419         struct op_msrs *msrs;
420
421         nmi_enabled = 0;
422         on_each_cpu(nmi_cpu_shutdown, NULL, 1);
423         unregister_die_notifier(&profile_exceptions_nb);
424         nmi_shutdown_mux();
425         msrs = &get_cpu_var(cpu_msrs);
426         model->shutdown(msrs);
427         free_msrs();
428         put_cpu_var(cpu_msrs);
429 }
430
431 static int nmi_create_files(struct super_block *sb, struct dentry *root)
432 {
433         unsigned int i;
434
435         for (i = 0; i < model->num_virt_counters; ++i) {
436                 struct dentry *dir;
437                 char buf[4];
438
439                 /* quick little hack to _not_ expose a counter if it is not
440                  * available for use.  This should protect userspace app.
441                  * NOTE:  assumes 1:1 mapping here (that counters are organized
442                  *        sequentially in their struct assignment).
443                  */
444                 if (!avail_to_resrv_perfctr_nmi_bit(op_x86_virt_to_phys(i)))
445                         continue;
446
447                 snprintf(buf,  sizeof(buf), "%d", i);
448                 dir = oprofilefs_mkdir(sb, root, buf);
449                 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
450                 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
451                 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
452                 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
453                 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
454                 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
455         }
456
457         return 0;
458 }
459
460 #ifdef CONFIG_SMP
461 static int oprofile_cpu_notifier(struct notifier_block *b, unsigned long action,
462                                  void *data)
463 {
464         int cpu = (unsigned long)data;
465         switch (action) {
466         case CPU_DOWN_FAILED:
467         case CPU_ONLINE:
468                 smp_call_function_single(cpu, nmi_cpu_start, NULL, 0);
469                 break;
470         case CPU_DOWN_PREPARE:
471                 smp_call_function_single(cpu, nmi_cpu_stop, NULL, 1);
472                 break;
473         }
474         return NOTIFY_DONE;
475 }
476
477 static struct notifier_block oprofile_cpu_nb = {
478         .notifier_call = oprofile_cpu_notifier
479 };
480 #endif
481
482 #ifdef CONFIG_PM
483
484 static int nmi_suspend(struct sys_device *dev, pm_message_t state)
485 {
486         /* Only one CPU left, just stop that one */
487         if (nmi_enabled == 1)
488                 nmi_cpu_stop(NULL);
489         return 0;
490 }
491
492 static int nmi_resume(struct sys_device *dev)
493 {
494         if (nmi_enabled == 1)
495                 nmi_cpu_start(NULL);
496         return 0;
497 }
498
499 static struct sysdev_class oprofile_sysclass = {
500         .name           = "oprofile",
501         .resume         = nmi_resume,
502         .suspend        = nmi_suspend,
503 };
504
505 static struct sys_device device_oprofile = {
506         .id     = 0,
507         .cls    = &oprofile_sysclass,
508 };
509
510 static int __init init_sysfs(void)
511 {
512         int error;
513
514         error = sysdev_class_register(&oprofile_sysclass);
515         if (!error)
516                 error = sysdev_register(&device_oprofile);
517         return error;
518 }
519
520 static void exit_sysfs(void)
521 {
522         sysdev_unregister(&device_oprofile);
523         sysdev_class_unregister(&oprofile_sysclass);
524 }
525
526 #else
527 #define init_sysfs() do { } while (0)
528 #define exit_sysfs() do { } while (0)
529 #endif /* CONFIG_PM */
530
531 static int __init p4_init(char **cpu_type)
532 {
533         __u8 cpu_model = boot_cpu_data.x86_model;
534
535         if (cpu_model > 6 || cpu_model == 5)
536                 return 0;
537
538 #ifndef CONFIG_SMP
539         *cpu_type = "i386/p4";
540         model = &op_p4_spec;
541         return 1;
542 #else
543         switch (smp_num_siblings) {
544         case 1:
545                 *cpu_type = "i386/p4";
546                 model = &op_p4_spec;
547                 return 1;
548
549         case 2:
550                 *cpu_type = "i386/p4-ht";
551                 model = &op_p4_ht2_spec;
552                 return 1;
553         }
554 #endif
555
556         printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n");
557         printk(KERN_INFO "oprofile: Reverting to timer mode.\n");
558         return 0;
559 }
560
561 static int force_arch_perfmon;
562 static int force_cpu_type(const char *str, struct kernel_param *kp)
563 {
564         if (!strcmp(str, "arch_perfmon")) {
565                 force_arch_perfmon = 1;
566                 printk(KERN_INFO "oprofile: forcing architectural perfmon\n");
567         }
568
569         return 0;
570 }
571 module_param_call(cpu_type, force_cpu_type, NULL, NULL, 0);
572
573 static int __init ppro_init(char **cpu_type)
574 {
575         __u8 cpu_model = boot_cpu_data.x86_model;
576         struct op_x86_model_spec *spec = &op_ppro_spec; /* default */
577
578         if (force_arch_perfmon && cpu_has_arch_perfmon)
579                 return 0;
580
581         switch (cpu_model) {
582         case 0 ... 2:
583                 *cpu_type = "i386/ppro";
584                 break;
585         case 3 ... 5:
586                 *cpu_type = "i386/pii";
587                 break;
588         case 6 ... 8:
589         case 10 ... 11:
590                 *cpu_type = "i386/piii";
591                 break;
592         case 9:
593         case 13:
594                 *cpu_type = "i386/p6_mobile";
595                 break;
596         case 14:
597                 *cpu_type = "i386/core";
598                 break;
599         case 15: case 23:
600                 *cpu_type = "i386/core_2";
601                 break;
602         case 0x2e:
603         case 26:
604                 spec = &op_arch_perfmon_spec;
605                 *cpu_type = "i386/core_i7";
606                 break;
607         case 28:
608                 *cpu_type = "i386/atom";
609                 break;
610         default:
611                 /* Unknown */
612                 return 0;
613         }
614
615         model = spec;
616         return 1;
617 }
618
619 /* in order to get sysfs right */
620 static int using_nmi;
621
622 int __init op_nmi_init(struct oprofile_operations *ops)
623 {
624         __u8 vendor = boot_cpu_data.x86_vendor;
625         __u8 family = boot_cpu_data.x86;
626         char *cpu_type = NULL;
627         int ret = 0;
628
629         if (!cpu_has_apic)
630                 return -ENODEV;
631
632         switch (vendor) {
633         case X86_VENDOR_AMD:
634                 /* Needs to be at least an Athlon (or hammer in 32bit mode) */
635
636                 switch (family) {
637                 case 6:
638                         cpu_type = "i386/athlon";
639                         break;
640                 case 0xf:
641                         /*
642                          * Actually it could be i386/hammer too, but
643                          * give user space an consistent name.
644                          */
645                         cpu_type = "x86-64/hammer";
646                         break;
647                 case 0x10:
648                         cpu_type = "x86-64/family10";
649                         break;
650                 case 0x11:
651                         cpu_type = "x86-64/family11h";
652                         break;
653                 default:
654                         return -ENODEV;
655                 }
656                 model = &op_amd_spec;
657                 break;
658
659         case X86_VENDOR_INTEL:
660                 switch (family) {
661                         /* Pentium IV */
662                 case 0xf:
663                         p4_init(&cpu_type);
664                         break;
665
666                         /* A P6-class processor */
667                 case 6:
668                         ppro_init(&cpu_type);
669                         break;
670
671                 default:
672                         break;
673                 }
674
675                 if (cpu_type)
676                         break;
677
678                 if (!cpu_has_arch_perfmon)
679                         return -ENODEV;
680
681                 /* use arch perfmon as fallback */
682                 cpu_type = "i386/arch_perfmon";
683                 model = &op_arch_perfmon_spec;
684                 break;
685
686         default:
687                 return -ENODEV;
688         }
689
690 #ifdef CONFIG_SMP
691         register_cpu_notifier(&oprofile_cpu_nb);
692 #endif
693         /* default values, can be overwritten by model */
694         ops->create_files       = nmi_create_files;
695         ops->setup              = nmi_setup;
696         ops->shutdown           = nmi_shutdown;
697         ops->start              = nmi_start;
698         ops->stop               = nmi_stop;
699         ops->cpu_type           = cpu_type;
700
701         if (model->init)
702                 ret = model->init(ops);
703         if (ret)
704                 return ret;
705
706         if (!model->num_virt_counters)
707                 model->num_virt_counters = model->num_counters;
708
709         mux_init(ops);
710
711         init_sysfs();
712         using_nmi = 1;
713         printk(KERN_INFO "oprofile: using NMI interrupt.\n");
714         return 0;
715 }
716
717 void op_nmi_exit(void)
718 {
719         if (using_nmi) {
720                 exit_sysfs();
721 #ifdef CONFIG_SMP
722                 unregister_cpu_notifier(&oprofile_cpu_nb);
723 #endif
724         }
725         if (model->exit)
726                 model->exit();
727 }