x86/oprofile: Implement mux_clone()
[deliverable/linux.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
38 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
39 extern atomic_t multiplex_counter;
40 #endif
41
42 struct op_counter_config counter_config[OP_MAX_COUNTER];
43
44 /* common functions */
45
46 u64 op_x86_get_ctrl(struct op_x86_model_spec const *model,
47 struct op_counter_config *counter_config)
48 {
49 u64 val = 0;
50 u16 event = (u16)counter_config->event;
51
52 val |= ARCH_PERFMON_EVENTSEL_INT;
53 val |= counter_config->user ? ARCH_PERFMON_EVENTSEL_USR : 0;
54 val |= counter_config->kernel ? ARCH_PERFMON_EVENTSEL_OS : 0;
55 val |= (counter_config->unit_mask & 0xFF) << 8;
56 event &= model->event_mask ? model->event_mask : 0xFF;
57 val |= event & 0xFF;
58 val |= (event & 0x0F00) << 24;
59
60 return val;
61 }
62
63
64 static int profile_exceptions_notify(struct notifier_block *self,
65 unsigned long val, void *data)
66 {
67 struct die_args *args = (struct die_args *)data;
68 int ret = NOTIFY_DONE;
69 int cpu = smp_processor_id();
70
71 switch (val) {
72 case DIE_NMI:
73 case DIE_NMI_IPI:
74 model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu));
75 ret = NOTIFY_STOP;
76 break;
77 default:
78 break;
79 }
80 return ret;
81 }
82
83 static void nmi_cpu_save_registers(struct op_msrs *msrs)
84 {
85 struct op_msr *counters = msrs->counters;
86 struct op_msr *controls = msrs->controls;
87 unsigned int i;
88
89 for (i = 0; i < model->num_counters; ++i) {
90 if (counters[i].addr)
91 rdmsrl(counters[i].addr, counters[i].saved);
92 }
93
94 for (i = 0; i < model->num_controls; ++i) {
95 if (controls[i].addr)
96 rdmsrl(controls[i].addr, controls[i].saved);
97 }
98 }
99
100 static void nmi_cpu_start(void *dummy)
101 {
102 struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
103 model->start(msrs);
104 }
105
106 static int nmi_start(void)
107 {
108 on_each_cpu(nmi_cpu_start, NULL, 1);
109 return 0;
110 }
111
112 static void nmi_cpu_stop(void *dummy)
113 {
114 struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
115 model->stop(msrs);
116 }
117
118 static void nmi_stop(void)
119 {
120 on_each_cpu(nmi_cpu_stop, NULL, 1);
121 }
122
123 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
124
125 static DEFINE_PER_CPU(int, switch_index);
126
127 static inline int has_mux(void)
128 {
129 return !!model->switch_ctrl;
130 }
131
132 inline int op_x86_phys_to_virt(int phys)
133 {
134 return __get_cpu_var(switch_index) + phys;
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 kmalloc(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].addr = 0;
183 multiplex[i].saved = 0;
184 }
185 }
186
187 per_cpu(switch_index, cpu) = 0;
188 }
189
190 static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs)
191 {
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 (multiplex[virt].addr)
198 rdmsrl(multiplex[virt].addr, multiplex[virt].saved);
199 }
200 }
201
202 static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs)
203 {
204 struct op_msr *multiplex = msrs->multiplex;
205 int i;
206
207 for (i = 0; i < model->num_counters; ++i) {
208 int virt = op_x86_phys_to_virt(i);
209 if (multiplex[virt].addr)
210 wrmsrl(multiplex[virt].addr, multiplex[virt].saved);
211 }
212 }
213
214 static void nmi_cpu_switch(void *dummy)
215 {
216 int cpu = smp_processor_id();
217 int si = per_cpu(switch_index, cpu);
218 struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
219
220 nmi_cpu_stop(NULL);
221 nmi_cpu_save_mpx_registers(msrs);
222
223 /* move to next set */
224 si += model->num_counters;
225 if ((si > model->num_virt_counters) || (counter_config[si].count == 0))
226 per_cpu(switch_index, cpu) = 0;
227 else
228 per_cpu(switch_index, cpu) = si;
229
230 model->switch_ctrl(model, msrs);
231 nmi_cpu_restore_mpx_registers(msrs);
232
233 nmi_cpu_start(NULL);
234 }
235
236
237 /*
238 * Quick check to see if multiplexing is necessary.
239 * The check should be sufficient since counters are used
240 * in ordre.
241 */
242 static int nmi_multiplex_on(void)
243 {
244 return counter_config[model->num_counters].count ? 0 : -EINVAL;
245 }
246
247 static int nmi_switch_event(void)
248 {
249 if (!has_mux())
250 return -ENOSYS; /* not implemented */
251 if (nmi_multiplex_on() < 0)
252 return -EINVAL; /* not necessary */
253
254 on_each_cpu(nmi_cpu_switch, NULL, 1);
255
256 atomic_inc(&multiplex_counter);
257
258 return 0;
259 }
260
261 static inline void mux_init(struct oprofile_operations *ops)
262 {
263 if (has_mux())
264 ops->switch_events = nmi_switch_event;
265 }
266
267 static void mux_clone(int cpu)
268 {
269 if (!has_mux())
270 return;
271
272 memcpy(per_cpu(cpu_msrs, cpu).multiplex,
273 per_cpu(cpu_msrs, 0).multiplex,
274 sizeof(struct op_msr) * model->num_virt_counters);
275 }
276
277 #else
278
279 inline int op_x86_phys_to_virt(int phys) { return phys; }
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 = kmalloc(counters_size,
308 GFP_KERNEL);
309 if (!per_cpu(cpu_msrs, i).counters)
310 return 0;
311 per_cpu(cpu_msrs, i).controls = kmalloc(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 #ifndef CONFIG_OPROFILE_EVENT_MULTIPLEX
440 /* quick little hack to _not_ expose a counter if it is not
441 * available for use. This should protect userspace app.
442 * NOTE: assumes 1:1 mapping here (that counters are organized
443 * sequentially in their struct assignment).
444 */
445 if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i)))
446 continue;
447 #endif /* CONFIG_OPROFILE_EVENT_MULTIPLEX */
448
449 snprintf(buf, sizeof(buf), "%d", i);
450 dir = oprofilefs_mkdir(sb, root, buf);
451 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
452 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
453 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
454 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
455 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
456 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
457 }
458
459 return 0;
460 }
461
462 #ifdef CONFIG_SMP
463 static int oprofile_cpu_notifier(struct notifier_block *b, unsigned long action,
464 void *data)
465 {
466 int cpu = (unsigned long)data;
467 switch (action) {
468 case CPU_DOWN_FAILED:
469 case CPU_ONLINE:
470 smp_call_function_single(cpu, nmi_cpu_start, NULL, 0);
471 break;
472 case CPU_DOWN_PREPARE:
473 smp_call_function_single(cpu, nmi_cpu_stop, NULL, 1);
474 break;
475 }
476 return NOTIFY_DONE;
477 }
478
479 static struct notifier_block oprofile_cpu_nb = {
480 .notifier_call = oprofile_cpu_notifier
481 };
482 #endif
483
484 #ifdef CONFIG_PM
485
486 static int nmi_suspend(struct sys_device *dev, pm_message_t state)
487 {
488 /* Only one CPU left, just stop that one */
489 if (nmi_enabled == 1)
490 nmi_cpu_stop(NULL);
491 return 0;
492 }
493
494 static int nmi_resume(struct sys_device *dev)
495 {
496 if (nmi_enabled == 1)
497 nmi_cpu_start(NULL);
498 return 0;
499 }
500
501 static struct sysdev_class oprofile_sysclass = {
502 .name = "oprofile",
503 .resume = nmi_resume,
504 .suspend = nmi_suspend,
505 };
506
507 static struct sys_device device_oprofile = {
508 .id = 0,
509 .cls = &oprofile_sysclass,
510 };
511
512 static int __init init_sysfs(void)
513 {
514 int error;
515
516 error = sysdev_class_register(&oprofile_sysclass);
517 if (!error)
518 error = sysdev_register(&device_oprofile);
519 return error;
520 }
521
522 static void exit_sysfs(void)
523 {
524 sysdev_unregister(&device_oprofile);
525 sysdev_class_unregister(&oprofile_sysclass);
526 }
527
528 #else
529 #define init_sysfs() do { } while (0)
530 #define exit_sysfs() do { } while (0)
531 #endif /* CONFIG_PM */
532
533 static int __init p4_init(char **cpu_type)
534 {
535 __u8 cpu_model = boot_cpu_data.x86_model;
536
537 if (cpu_model > 6 || cpu_model == 5)
538 return 0;
539
540 #ifndef CONFIG_SMP
541 *cpu_type = "i386/p4";
542 model = &op_p4_spec;
543 return 1;
544 #else
545 switch (smp_num_siblings) {
546 case 1:
547 *cpu_type = "i386/p4";
548 model = &op_p4_spec;
549 return 1;
550
551 case 2:
552 *cpu_type = "i386/p4-ht";
553 model = &op_p4_ht2_spec;
554 return 1;
555 }
556 #endif
557
558 printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n");
559 printk(KERN_INFO "oprofile: Reverting to timer mode.\n");
560 return 0;
561 }
562
563 static int force_arch_perfmon;
564 static int force_cpu_type(const char *str, struct kernel_param *kp)
565 {
566 if (!strcmp(str, "arch_perfmon")) {
567 force_arch_perfmon = 1;
568 printk(KERN_INFO "oprofile: forcing architectural perfmon\n");
569 }
570
571 return 0;
572 }
573 module_param_call(cpu_type, force_cpu_type, NULL, NULL, 0);
574
575 static int __init ppro_init(char **cpu_type)
576 {
577 __u8 cpu_model = boot_cpu_data.x86_model;
578 struct op_x86_model_spec *spec = &op_ppro_spec; /* default */
579
580 if (force_arch_perfmon && cpu_has_arch_perfmon)
581 return 0;
582
583 switch (cpu_model) {
584 case 0 ... 2:
585 *cpu_type = "i386/ppro";
586 break;
587 case 3 ... 5:
588 *cpu_type = "i386/pii";
589 break;
590 case 6 ... 8:
591 case 10 ... 11:
592 *cpu_type = "i386/piii";
593 break;
594 case 9:
595 case 13:
596 *cpu_type = "i386/p6_mobile";
597 break;
598 case 14:
599 *cpu_type = "i386/core";
600 break;
601 case 15: case 23:
602 *cpu_type = "i386/core_2";
603 break;
604 case 26:
605 spec = &op_arch_perfmon_spec;
606 *cpu_type = "i386/core_i7";
607 break;
608 case 28:
609 *cpu_type = "i386/atom";
610 break;
611 default:
612 /* Unknown */
613 return 0;
614 }
615
616 model = spec;
617 return 1;
618 }
619
620 /* in order to get sysfs right */
621 static int using_nmi;
622
623 int __init op_nmi_init(struct oprofile_operations *ops)
624 {
625 __u8 vendor = boot_cpu_data.x86_vendor;
626 __u8 family = boot_cpu_data.x86;
627 char *cpu_type = NULL;
628 int ret = 0;
629
630 if (!cpu_has_apic)
631 return -ENODEV;
632
633 switch (vendor) {
634 case X86_VENDOR_AMD:
635 /* Needs to be at least an Athlon (or hammer in 32bit mode) */
636
637 switch (family) {
638 case 6:
639 cpu_type = "i386/athlon";
640 break;
641 case 0xf:
642 /*
643 * Actually it could be i386/hammer too, but
644 * give user space an consistent name.
645 */
646 cpu_type = "x86-64/hammer";
647 break;
648 case 0x10:
649 cpu_type = "x86-64/family10";
650 break;
651 case 0x11:
652 cpu_type = "x86-64/family11h";
653 break;
654 default:
655 return -ENODEV;
656 }
657 model = &op_amd_spec;
658 break;
659
660 case X86_VENDOR_INTEL:
661 switch (family) {
662 /* Pentium IV */
663 case 0xf:
664 p4_init(&cpu_type);
665 break;
666
667 /* A P6-class processor */
668 case 6:
669 ppro_init(&cpu_type);
670 break;
671
672 default:
673 break;
674 }
675
676 if (cpu_type)
677 break;
678
679 if (!cpu_has_arch_perfmon)
680 return -ENODEV;
681
682 /* use arch perfmon as fallback */
683 cpu_type = "i386/arch_perfmon";
684 model = &op_arch_perfmon_spec;
685 break;
686
687 default:
688 return -ENODEV;
689 }
690
691 #ifdef CONFIG_SMP
692 register_cpu_notifier(&oprofile_cpu_nb);
693 #endif
694 /* default values, can be overwritten by model */
695 ops->create_files = nmi_create_files;
696 ops->setup = nmi_setup;
697 ops->shutdown = nmi_shutdown;
698 ops->start = nmi_start;
699 ops->stop = nmi_stop;
700 ops->cpu_type = cpu_type;
701
702 if (model->init)
703 ret = model->init(ops);
704 if (ret)
705 return ret;
706
707 if (!model->num_virt_counters)
708 model->num_virt_counters = model->num_counters;
709
710 mux_init(ops);
711
712 init_sysfs();
713 using_nmi = 1;
714 printk(KERN_INFO "oprofile: using NMI interrupt.\n");
715 return 0;
716 }
717
718 void op_nmi_exit(void)
719 {
720 if (using_nmi) {
721 exit_sysfs();
722 #ifdef CONFIG_SMP
723 unregister_cpu_notifier(&oprofile_cpu_nb);
724 #endif
725 }
726 if (model->exit)
727 model->exit();
728 }
This page took 0.056571 seconds and 5 git commands to generate.