oprofile/x86: fix uninitialized counter usage during cpu hotplug
[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         if (!msrs->controls)
99                 WARN_ON_ONCE(1);
100         else
101                 model->start(msrs);
102 }
103
104 static int nmi_start(void)
105 {
106         on_each_cpu(nmi_cpu_start, NULL, 1);
107         return 0;
108 }
109
110 static void nmi_cpu_stop(void *dummy)
111 {
112         struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
113         if (!msrs->controls)
114                 WARN_ON_ONCE(1);
115         else
116                 model->stop(msrs);
117 }
118
119 static void nmi_stop(void)
120 {
121         on_each_cpu(nmi_cpu_stop, NULL, 1);
122 }
123
124 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
125
126 static DEFINE_PER_CPU(int, switch_index);
127
128 static inline int has_mux(void)
129 {
130         return !!model->switch_ctrl;
131 }
132
133 inline int op_x86_phys_to_virt(int phys)
134 {
135         return __get_cpu_var(switch_index) + phys;
136 }
137
138 inline int op_x86_virt_to_phys(int virt)
139 {
140         return virt % model->num_counters;
141 }
142
143 static void nmi_shutdown_mux(void)
144 {
145         int i;
146
147         if (!has_mux())
148                 return;
149
150         for_each_possible_cpu(i) {
151                 kfree(per_cpu(cpu_msrs, i).multiplex);
152                 per_cpu(cpu_msrs, i).multiplex = NULL;
153                 per_cpu(switch_index, i) = 0;
154         }
155 }
156
157 static int nmi_setup_mux(void)
158 {
159         size_t multiplex_size =
160                 sizeof(struct op_msr) * model->num_virt_counters;
161         int i;
162
163         if (!has_mux())
164                 return 1;
165
166         for_each_possible_cpu(i) {
167                 per_cpu(cpu_msrs, i).multiplex =
168                         kzalloc(multiplex_size, GFP_KERNEL);
169                 if (!per_cpu(cpu_msrs, i).multiplex)
170                         return 0;
171         }
172
173         return 1;
174 }
175
176 static void nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs)
177 {
178         int i;
179         struct op_msr *multiplex = msrs->multiplex;
180
181         if (!has_mux())
182                 return;
183
184         for (i = 0; i < model->num_virt_counters; ++i) {
185                 if (counter_config[i].enabled) {
186                         multiplex[i].saved = -(u64)counter_config[i].count;
187                 } else {
188                         multiplex[i].saved = 0;
189                 }
190         }
191
192         per_cpu(switch_index, cpu) = 0;
193 }
194
195 static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs)
196 {
197         struct op_msr *counters = msrs->counters;
198         struct op_msr *multiplex = msrs->multiplex;
199         int i;
200
201         for (i = 0; i < model->num_counters; ++i) {
202                 int virt = op_x86_phys_to_virt(i);
203                 if (counters[i].addr)
204                         rdmsrl(counters[i].addr, multiplex[virt].saved);
205         }
206 }
207
208 static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs)
209 {
210         struct op_msr *counters = msrs->counters;
211         struct op_msr *multiplex = msrs->multiplex;
212         int i;
213
214         for (i = 0; i < model->num_counters; ++i) {
215                 int virt = op_x86_phys_to_virt(i);
216                 if (counters[i].addr)
217                         wrmsrl(counters[i].addr, multiplex[virt].saved);
218         }
219 }
220
221 static void nmi_cpu_switch(void *dummy)
222 {
223         int cpu = smp_processor_id();
224         int si = per_cpu(switch_index, cpu);
225         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
226
227         nmi_cpu_stop(NULL);
228         nmi_cpu_save_mpx_registers(msrs);
229
230         /* move to next set */
231         si += model->num_counters;
232         if ((si >= model->num_virt_counters) || (counter_config[si].count == 0))
233                 per_cpu(switch_index, cpu) = 0;
234         else
235                 per_cpu(switch_index, cpu) = si;
236
237         model->switch_ctrl(model, msrs);
238         nmi_cpu_restore_mpx_registers(msrs);
239
240         nmi_cpu_start(NULL);
241 }
242
243
244 /*
245  * Quick check to see if multiplexing is necessary.
246  * The check should be sufficient since counters are used
247  * in ordre.
248  */
249 static int nmi_multiplex_on(void)
250 {
251         return counter_config[model->num_counters].count ? 0 : -EINVAL;
252 }
253
254 static int nmi_switch_event(void)
255 {
256         if (!has_mux())
257                 return -ENOSYS;         /* not implemented */
258         if (nmi_multiplex_on() < 0)
259                 return -EINVAL;         /* not necessary */
260
261         on_each_cpu(nmi_cpu_switch, NULL, 1);
262
263         return 0;
264 }
265
266 static inline void mux_init(struct oprofile_operations *ops)
267 {
268         if (has_mux())
269                 ops->switch_events = nmi_switch_event;
270 }
271
272 static void mux_clone(int cpu)
273 {
274         if (!has_mux())
275                 return;
276
277         memcpy(per_cpu(cpu_msrs, cpu).multiplex,
278                per_cpu(cpu_msrs, 0).multiplex,
279                sizeof(struct op_msr) * model->num_virt_counters);
280 }
281
282 #else
283
284 inline int op_x86_phys_to_virt(int phys) { return phys; }
285 inline int op_x86_virt_to_phys(int virt) { return virt; }
286 static inline void nmi_shutdown_mux(void) { }
287 static inline int nmi_setup_mux(void) { return 1; }
288 static inline void
289 nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs) { }
290 static inline void mux_init(struct oprofile_operations *ops) { }
291 static void mux_clone(int cpu) { }
292
293 #endif
294
295 static void free_msrs(void)
296 {
297         int i;
298         for_each_possible_cpu(i) {
299                 kfree(per_cpu(cpu_msrs, i).counters);
300                 per_cpu(cpu_msrs, i).counters = NULL;
301                 kfree(per_cpu(cpu_msrs, i).controls);
302                 per_cpu(cpu_msrs, i).controls = NULL;
303         }
304         nmi_shutdown_mux();
305 }
306
307 static int allocate_msrs(void)
308 {
309         size_t controls_size = sizeof(struct op_msr) * model->num_controls;
310         size_t counters_size = sizeof(struct op_msr) * model->num_counters;
311
312         int i;
313         for_each_possible_cpu(i) {
314                 per_cpu(cpu_msrs, i).counters = kzalloc(counters_size,
315                                                         GFP_KERNEL);
316                 if (!per_cpu(cpu_msrs, i).counters)
317                         goto fail;
318                 per_cpu(cpu_msrs, i).controls = kzalloc(controls_size,
319                                                         GFP_KERNEL);
320                 if (!per_cpu(cpu_msrs, i).controls)
321                         goto fail;
322         }
323
324         if (!nmi_setup_mux())
325                 goto fail;
326
327         return 1;
328
329 fail:
330         free_msrs();
331         return 0;
332 }
333
334 static void nmi_cpu_setup(void *dummy)
335 {
336         int cpu = smp_processor_id();
337         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
338         nmi_cpu_save_registers(msrs);
339         spin_lock(&oprofilefs_lock);
340         model->setup_ctrs(model, msrs);
341         nmi_cpu_setup_mux(cpu, msrs);
342         spin_unlock(&oprofilefs_lock);
343         per_cpu(saved_lvtpc, cpu) = apic_read(APIC_LVTPC);
344         apic_write(APIC_LVTPC, APIC_DM_NMI);
345 }
346
347 static struct notifier_block profile_exceptions_nb = {
348         .notifier_call = profile_exceptions_notify,
349         .next = NULL,
350         .priority = 2
351 };
352
353 static int nmi_setup(void)
354 {
355         int err = 0;
356         int cpu;
357
358         if (!allocate_msrs())
359                 return -ENOMEM;
360
361         /* We need to serialize save and setup for HT because the subset
362          * of msrs are distinct for save and setup operations
363          */
364
365         /* Assume saved/restored counters are the same on all CPUs */
366         err = model->fill_in_addresses(&per_cpu(cpu_msrs, 0));
367         if (err)
368                 goto fail;
369
370         for_each_possible_cpu(cpu) {
371                 if (!cpu)
372                         continue;
373
374                 memcpy(per_cpu(cpu_msrs, cpu).counters,
375                        per_cpu(cpu_msrs, 0).counters,
376                        sizeof(struct op_msr) * model->num_counters);
377
378                 memcpy(per_cpu(cpu_msrs, cpu).controls,
379                        per_cpu(cpu_msrs, 0).controls,
380                        sizeof(struct op_msr) * model->num_controls);
381
382                 mux_clone(cpu);
383         }
384
385         err = register_die_notifier(&profile_exceptions_nb);
386         if (err)
387                 goto fail;
388
389         on_each_cpu(nmi_cpu_setup, NULL, 1);
390         nmi_enabled = 1;
391         return 0;
392 fail:
393         free_msrs();
394         return err;
395 }
396
397 static void nmi_cpu_restore_registers(struct op_msrs *msrs)
398 {
399         struct op_msr *counters = msrs->counters;
400         struct op_msr *controls = msrs->controls;
401         unsigned int i;
402
403         for (i = 0; i < model->num_controls; ++i) {
404                 if (controls[i].addr)
405                         wrmsrl(controls[i].addr, controls[i].saved);
406         }
407
408         for (i = 0; i < model->num_counters; ++i) {
409                 if (counters[i].addr)
410                         wrmsrl(counters[i].addr, counters[i].saved);
411         }
412 }
413
414 static void nmi_cpu_shutdown(void *dummy)
415 {
416         unsigned int v;
417         int cpu = smp_processor_id();
418         struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
419
420         /* restoring APIC_LVTPC can trigger an apic error because the delivery
421          * mode and vector nr combination can be illegal. That's by design: on
422          * power on apic lvt contain a zero vector nr which are legal only for
423          * NMI delivery mode. So inhibit apic err before restoring lvtpc
424          */
425         v = apic_read(APIC_LVTERR);
426         apic_write(APIC_LVTERR, v | APIC_LVT_MASKED);
427         apic_write(APIC_LVTPC, per_cpu(saved_lvtpc, cpu));
428         apic_write(APIC_LVTERR, v);
429         nmi_cpu_restore_registers(msrs);
430 }
431
432 static void nmi_shutdown(void)
433 {
434         struct op_msrs *msrs;
435
436         nmi_enabled = 0;
437         on_each_cpu(nmi_cpu_shutdown, NULL, 1);
438         unregister_die_notifier(&profile_exceptions_nb);
439         msrs = &get_cpu_var(cpu_msrs);
440         model->shutdown(msrs);
441         free_msrs();
442         put_cpu_var(cpu_msrs);
443 }
444
445 static int nmi_create_files(struct super_block *sb, struct dentry *root)
446 {
447         unsigned int i;
448
449         for (i = 0; i < model->num_virt_counters; ++i) {
450                 struct dentry *dir;
451                 char buf[4];
452
453                 /* quick little hack to _not_ expose a counter if it is not
454                  * available for use.  This should protect userspace app.
455                  * NOTE:  assumes 1:1 mapping here (that counters are organized
456                  *        sequentially in their struct assignment).
457                  */
458                 if (!avail_to_resrv_perfctr_nmi_bit(op_x86_virt_to_phys(i)))
459                         continue;
460
461                 snprintf(buf,  sizeof(buf), "%d", i);
462                 dir = oprofilefs_mkdir(sb, root, buf);
463                 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
464                 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
465                 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
466                 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
467                 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
468                 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
469         }
470
471         return 0;
472 }
473
474 #ifdef CONFIG_SMP
475 static int oprofile_cpu_notifier(struct notifier_block *b, unsigned long action,
476                                  void *data)
477 {
478         int cpu = (unsigned long)data;
479         switch (action) {
480         case CPU_DOWN_FAILED:
481         case CPU_ONLINE:
482                 smp_call_function_single(cpu, nmi_cpu_start, NULL, 0);
483                 break;
484         case CPU_DOWN_PREPARE:
485                 smp_call_function_single(cpu, nmi_cpu_stop, NULL, 1);
486                 break;
487         }
488         return NOTIFY_DONE;
489 }
490
491 static struct notifier_block oprofile_cpu_nb = {
492         .notifier_call = oprofile_cpu_notifier
493 };
494 #endif
495
496 #ifdef CONFIG_PM
497
498 static int nmi_suspend(struct sys_device *dev, pm_message_t state)
499 {
500         /* Only one CPU left, just stop that one */
501         if (nmi_enabled == 1)
502                 nmi_cpu_stop(NULL);
503         return 0;
504 }
505
506 static int nmi_resume(struct sys_device *dev)
507 {
508         if (nmi_enabled == 1)
509                 nmi_cpu_start(NULL);
510         return 0;
511 }
512
513 static struct sysdev_class oprofile_sysclass = {
514         .name           = "oprofile",
515         .resume         = nmi_resume,
516         .suspend        = nmi_suspend,
517 };
518
519 static struct sys_device device_oprofile = {
520         .id     = 0,
521         .cls    = &oprofile_sysclass,
522 };
523
524 static int __init init_sysfs(void)
525 {
526         int error;
527
528         error = sysdev_class_register(&oprofile_sysclass);
529         if (!error)
530                 error = sysdev_register(&device_oprofile);
531         return error;
532 }
533
534 static void exit_sysfs(void)
535 {
536         sysdev_unregister(&device_oprofile);
537         sysdev_class_unregister(&oprofile_sysclass);
538 }
539
540 #else
541 #define init_sysfs() do { } while (0)
542 #define exit_sysfs() do { } while (0)
543 #endif /* CONFIG_PM */
544
545 static int __init p4_init(char **cpu_type)
546 {
547         __u8 cpu_model = boot_cpu_data.x86_model;
548
549         if (cpu_model > 6 || cpu_model == 5)
550                 return 0;
551
552 #ifndef CONFIG_SMP
553         *cpu_type = "i386/p4";
554         model = &op_p4_spec;
555         return 1;
556 #else
557         switch (smp_num_siblings) {
558         case 1:
559                 *cpu_type = "i386/p4";
560                 model = &op_p4_spec;
561                 return 1;
562
563         case 2:
564                 *cpu_type = "i386/p4-ht";
565                 model = &op_p4_ht2_spec;
566                 return 1;
567         }
568 #endif
569
570         printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n");
571         printk(KERN_INFO "oprofile: Reverting to timer mode.\n");
572         return 0;
573 }
574
575 static int force_arch_perfmon;
576 static int force_cpu_type(const char *str, struct kernel_param *kp)
577 {
578         if (!strcmp(str, "arch_perfmon")) {
579                 force_arch_perfmon = 1;
580                 printk(KERN_INFO "oprofile: forcing architectural perfmon\n");
581         }
582
583         return 0;
584 }
585 module_param_call(cpu_type, force_cpu_type, NULL, NULL, 0);
586
587 static int __init ppro_init(char **cpu_type)
588 {
589         __u8 cpu_model = boot_cpu_data.x86_model;
590         struct op_x86_model_spec *spec = &op_ppro_spec; /* default */
591
592         if (force_arch_perfmon && cpu_has_arch_perfmon)
593                 return 0;
594
595         switch (cpu_model) {
596         case 0 ... 2:
597                 *cpu_type = "i386/ppro";
598                 break;
599         case 3 ... 5:
600                 *cpu_type = "i386/pii";
601                 break;
602         case 6 ... 8:
603         case 10 ... 11:
604                 *cpu_type = "i386/piii";
605                 break;
606         case 9:
607         case 13:
608                 *cpu_type = "i386/p6_mobile";
609                 break;
610         case 14:
611                 *cpu_type = "i386/core";
612                 break;
613         case 15: case 23:
614                 *cpu_type = "i386/core_2";
615                 break;
616         case 0x2e:
617         case 26:
618                 spec = &op_arch_perfmon_spec;
619                 *cpu_type = "i386/core_i7";
620                 break;
621         case 28:
622                 *cpu_type = "i386/atom";
623                 break;
624         default:
625                 /* Unknown */
626                 return 0;
627         }
628
629         model = spec;
630         return 1;
631 }
632
633 /* in order to get sysfs right */
634 static int using_nmi;
635
636 int __init op_nmi_init(struct oprofile_operations *ops)
637 {
638         __u8 vendor = boot_cpu_data.x86_vendor;
639         __u8 family = boot_cpu_data.x86;
640         char *cpu_type = NULL;
641         int ret = 0;
642
643         if (!cpu_has_apic)
644                 return -ENODEV;
645
646         switch (vendor) {
647         case X86_VENDOR_AMD:
648                 /* Needs to be at least an Athlon (or hammer in 32bit mode) */
649
650                 switch (family) {
651                 case 6:
652                         cpu_type = "i386/athlon";
653                         break;
654                 case 0xf:
655                         /*
656                          * Actually it could be i386/hammer too, but
657                          * give user space an consistent name.
658                          */
659                         cpu_type = "x86-64/hammer";
660                         break;
661                 case 0x10:
662                         cpu_type = "x86-64/family10";
663                         break;
664                 case 0x11:
665                         cpu_type = "x86-64/family11h";
666                         break;
667                 default:
668                         return -ENODEV;
669                 }
670                 model = &op_amd_spec;
671                 break;
672
673         case X86_VENDOR_INTEL:
674                 switch (family) {
675                         /* Pentium IV */
676                 case 0xf:
677                         p4_init(&cpu_type);
678                         break;
679
680                         /* A P6-class processor */
681                 case 6:
682                         ppro_init(&cpu_type);
683                         break;
684
685                 default:
686                         break;
687                 }
688
689                 if (cpu_type)
690                         break;
691
692                 if (!cpu_has_arch_perfmon)
693                         return -ENODEV;
694
695                 /* use arch perfmon as fallback */
696                 cpu_type = "i386/arch_perfmon";
697                 model = &op_arch_perfmon_spec;
698                 break;
699
700         default:
701                 return -ENODEV;
702         }
703
704 #ifdef CONFIG_SMP
705         register_cpu_notifier(&oprofile_cpu_nb);
706 #endif
707         /* default values, can be overwritten by model */
708         ops->create_files       = nmi_create_files;
709         ops->setup              = nmi_setup;
710         ops->shutdown           = nmi_shutdown;
711         ops->start              = nmi_start;
712         ops->stop               = nmi_stop;
713         ops->cpu_type           = cpu_type;
714
715         if (model->init)
716                 ret = model->init(ops);
717         if (ret)
718                 return ret;
719
720         if (!model->num_virt_counters)
721                 model->num_virt_counters = model->num_counters;
722
723         mux_init(ops);
724
725         init_sysfs();
726         using_nmi = 1;
727         printk(KERN_INFO "oprofile: using NMI interrupt.\n");
728         return 0;
729 }
730
731 void op_nmi_exit(void)
732 {
733         if (using_nmi) {
734                 exit_sysfs();
735 #ifdef CONFIG_SMP
736                 unregister_cpu_notifier(&oprofile_cpu_nb);
737 #endif
738         }
739         if (model->exit)
740                 model->exit();
741 }