4 * @remark Copyright 2002-2009 OProfile authors
5 * @remark Read the file COPYING
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>
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>
27 #include "op_counter.h"
28 #include "op_x86_model.h"
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);
34 /* 0 == registered but off, 1 == registered and on */
35 static int nmi_enabled = 0;
38 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
39 extern atomic_t multiplex_counter;
42 struct op_counter_config counter_config[OP_MAX_COUNTER];
44 /* common functions */
46 u64 op_x86_get_ctrl(struct op_x86_model_spec const *model,
47 struct op_counter_config *counter_config)
50 u16 event = (u16)counter_config->event;
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;
58 val |= (event & 0x0F00) << 24;
64 static int profile_exceptions_notify(struct notifier_block *self,
65 unsigned long val, void *data)
67 struct die_args *args = (struct die_args *)data;
68 int ret = NOTIFY_DONE;
69 int cpu = smp_processor_id();
74 model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu));
83 static void nmi_cpu_save_registers(struct op_msrs *msrs)
85 struct op_msr *counters = msrs->counters;
86 struct op_msr *controls = msrs->controls;
89 for (i = 0; i < model->num_counters; ++i) {
91 rdmsrl(counters[i].addr, counters[i].saved);
94 for (i = 0; i < model->num_controls; ++i) {
96 rdmsrl(controls[i].addr, controls[i].saved);
100 static void nmi_cpu_start(void *dummy)
102 struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
106 static int nmi_start(void)
108 on_each_cpu(nmi_cpu_start, NULL, 1);
112 static void nmi_cpu_stop(void *dummy)
114 struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs);
118 static void nmi_stop(void)
120 on_each_cpu(nmi_cpu_stop, NULL, 1);
123 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
125 static DEFINE_PER_CPU(int, switch_index);
127 inline int op_x86_phys_to_virt(int phys)
129 return __get_cpu_var(switch_index) + phys;
132 static void nmi_shutdown_mux(void)
135 for_each_possible_cpu(i) {
136 kfree(per_cpu(cpu_msrs, i).multiplex);
137 per_cpu(cpu_msrs, i).multiplex = NULL;
138 per_cpu(switch_index, i) = 0;
142 static int nmi_setup_mux(void)
144 size_t multiplex_size =
145 sizeof(struct op_msr) * model->num_virt_counters;
147 for_each_possible_cpu(i) {
148 per_cpu(cpu_msrs, i).multiplex =
149 kmalloc(multiplex_size, GFP_KERNEL);
150 if (!per_cpu(cpu_msrs, i).multiplex)
156 static void nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs)
159 struct op_msr *multiplex = msrs->multiplex;
161 for (i = 0; i < model->num_virt_counters; ++i) {
162 if (counter_config[i].enabled) {
163 multiplex[i].saved = -(u64)counter_config[i].count;
165 multiplex[i].addr = 0;
166 multiplex[i].saved = 0;
170 per_cpu(switch_index, cpu) = 0;
173 static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs)
175 struct op_msr *multiplex = msrs->multiplex;
178 for (i = 0; i < model->num_counters; ++i) {
179 int virt = op_x86_phys_to_virt(i);
180 if (multiplex[virt].addr)
181 rdmsrl(multiplex[virt].addr, multiplex[virt].saved);
185 static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs)
187 struct op_msr *multiplex = msrs->multiplex;
190 for (i = 0; i < model->num_counters; ++i) {
191 int virt = op_x86_phys_to_virt(i);
192 if (multiplex[virt].addr)
193 wrmsrl(multiplex[virt].addr, multiplex[virt].saved);
197 static void nmi_cpu_switch(void *dummy)
199 int cpu = smp_processor_id();
200 int si = per_cpu(switch_index, cpu);
201 struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
204 nmi_cpu_save_mpx_registers(msrs);
206 /* move to next set */
207 si += model->num_counters;
208 if ((si > model->num_virt_counters) || (counter_config[si].count == 0))
209 per_cpu(switch_index, cpu) = 0;
211 per_cpu(switch_index, cpu) = si;
213 model->switch_ctrl(model, msrs);
214 nmi_cpu_restore_mpx_registers(msrs);
221 * Quick check to see if multiplexing is necessary.
222 * The check should be sufficient since counters are used
225 static int nmi_multiplex_on(void)
227 return counter_config[model->num_counters].count ? 0 : -EINVAL;
230 static int nmi_switch_event(void)
232 if (!model->switch_ctrl)
233 return -ENOSYS; /* not implemented */
234 if (nmi_multiplex_on() < 0)
235 return -EINVAL; /* not necessary */
237 on_each_cpu(nmi_cpu_switch, NULL, 1);
239 atomic_inc(&multiplex_counter);
246 inline int op_x86_phys_to_virt(int phys) { return phys; }
247 static inline void nmi_shutdown_mux(void) { }
248 static inline int nmi_setup_mux(void) { return 1; }
250 nmi_cpu_setup_mux(int cpu, struct op_msrs const * const msrs) { }
254 static void free_msrs(void)
257 for_each_possible_cpu(i) {
258 kfree(per_cpu(cpu_msrs, i).counters);
259 per_cpu(cpu_msrs, i).counters = NULL;
260 kfree(per_cpu(cpu_msrs, i).controls);
261 per_cpu(cpu_msrs, i).controls = NULL;
265 static int allocate_msrs(void)
267 size_t controls_size = sizeof(struct op_msr) * model->num_controls;
268 size_t counters_size = sizeof(struct op_msr) * model->num_counters;
271 for_each_possible_cpu(i) {
272 per_cpu(cpu_msrs, i).counters = kmalloc(counters_size,
274 if (!per_cpu(cpu_msrs, i).counters)
276 per_cpu(cpu_msrs, i).controls = kmalloc(controls_size,
278 if (!per_cpu(cpu_msrs, i).controls)
285 static void nmi_cpu_setup(void *dummy)
287 int cpu = smp_processor_id();
288 struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
289 nmi_cpu_save_registers(msrs);
290 spin_lock(&oprofilefs_lock);
291 model->setup_ctrs(model, msrs);
292 nmi_cpu_setup_mux(cpu, msrs);
293 spin_unlock(&oprofilefs_lock);
294 per_cpu(saved_lvtpc, cpu) = apic_read(APIC_LVTPC);
295 apic_write(APIC_LVTPC, APIC_DM_NMI);
298 static struct notifier_block profile_exceptions_nb = {
299 .notifier_call = profile_exceptions_notify,
304 static int nmi_setup(void)
309 if (!allocate_msrs())
311 else if (!nmi_setup_mux())
314 err = register_die_notifier(&profile_exceptions_nb);
322 /* We need to serialize save and setup for HT because the subset
323 * of msrs are distinct for save and setup operations
326 /* Assume saved/restored counters are the same on all CPUs */
327 model->fill_in_addresses(&per_cpu(cpu_msrs, 0));
328 for_each_possible_cpu(cpu) {
330 memcpy(per_cpu(cpu_msrs, cpu).counters,
331 per_cpu(cpu_msrs, 0).counters,
332 sizeof(struct op_msr) * model->num_counters);
334 memcpy(per_cpu(cpu_msrs, cpu).controls,
335 per_cpu(cpu_msrs, 0).controls,
336 sizeof(struct op_msr) * model->num_controls);
337 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
338 memcpy(per_cpu(cpu_msrs, cpu).multiplex,
339 per_cpu(cpu_msrs, 0).multiplex,
340 sizeof(struct op_msr) * model->num_virt_counters);
344 on_each_cpu(nmi_cpu_setup, NULL, 1);
349 static void nmi_cpu_restore_registers(struct op_msrs *msrs)
351 struct op_msr *counters = msrs->counters;
352 struct op_msr *controls = msrs->controls;
355 for (i = 0; i < model->num_controls; ++i) {
356 if (controls[i].addr)
357 wrmsrl(controls[i].addr, controls[i].saved);
360 for (i = 0; i < model->num_counters; ++i) {
361 if (counters[i].addr)
362 wrmsrl(counters[i].addr, counters[i].saved);
366 static void nmi_cpu_shutdown(void *dummy)
369 int cpu = smp_processor_id();
370 struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
372 /* restoring APIC_LVTPC can trigger an apic error because the delivery
373 * mode and vector nr combination can be illegal. That's by design: on
374 * power on apic lvt contain a zero vector nr which are legal only for
375 * NMI delivery mode. So inhibit apic err before restoring lvtpc
377 v = apic_read(APIC_LVTERR);
378 apic_write(APIC_LVTERR, v | APIC_LVT_MASKED);
379 apic_write(APIC_LVTPC, per_cpu(saved_lvtpc, cpu));
380 apic_write(APIC_LVTERR, v);
381 nmi_cpu_restore_registers(msrs);
384 static void nmi_shutdown(void)
386 struct op_msrs *msrs;
389 on_each_cpu(nmi_cpu_shutdown, NULL, 1);
390 unregister_die_notifier(&profile_exceptions_nb);
392 msrs = &get_cpu_var(cpu_msrs);
393 model->shutdown(msrs);
395 put_cpu_var(cpu_msrs);
398 static int nmi_create_files(struct super_block *sb, struct dentry *root)
402 for (i = 0; i < model->num_virt_counters; ++i) {
406 #ifndef CONFIG_OPROFILE_EVENT_MULTIPLEX
407 /* quick little hack to _not_ expose a counter if it is not
408 * available for use. This should protect userspace app.
409 * NOTE: assumes 1:1 mapping here (that counters are organized
410 * sequentially in their struct assignment).
412 if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i)))
414 #endif /* CONFIG_OPROFILE_EVENT_MULTIPLEX */
416 snprintf(buf, sizeof(buf), "%d", i);
417 dir = oprofilefs_mkdir(sb, root, buf);
418 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
419 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
420 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
421 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
422 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
423 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
430 static int oprofile_cpu_notifier(struct notifier_block *b, unsigned long action,
433 int cpu = (unsigned long)data;
435 case CPU_DOWN_FAILED:
437 smp_call_function_single(cpu, nmi_cpu_start, NULL, 0);
439 case CPU_DOWN_PREPARE:
440 smp_call_function_single(cpu, nmi_cpu_stop, NULL, 1);
446 static struct notifier_block oprofile_cpu_nb = {
447 .notifier_call = oprofile_cpu_notifier
453 static int nmi_suspend(struct sys_device *dev, pm_message_t state)
455 /* Only one CPU left, just stop that one */
456 if (nmi_enabled == 1)
461 static int nmi_resume(struct sys_device *dev)
463 if (nmi_enabled == 1)
468 static struct sysdev_class oprofile_sysclass = {
470 .resume = nmi_resume,
471 .suspend = nmi_suspend,
474 static struct sys_device device_oprofile = {
476 .cls = &oprofile_sysclass,
479 static int __init init_sysfs(void)
483 error = sysdev_class_register(&oprofile_sysclass);
485 error = sysdev_register(&device_oprofile);
489 static void exit_sysfs(void)
491 sysdev_unregister(&device_oprofile);
492 sysdev_class_unregister(&oprofile_sysclass);
496 #define init_sysfs() do { } while (0)
497 #define exit_sysfs() do { } while (0)
498 #endif /* CONFIG_PM */
500 static int __init p4_init(char **cpu_type)
502 __u8 cpu_model = boot_cpu_data.x86_model;
504 if (cpu_model > 6 || cpu_model == 5)
508 *cpu_type = "i386/p4";
512 switch (smp_num_siblings) {
514 *cpu_type = "i386/p4";
519 *cpu_type = "i386/p4-ht";
520 model = &op_p4_ht2_spec;
525 printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n");
526 printk(KERN_INFO "oprofile: Reverting to timer mode.\n");
530 static int force_arch_perfmon;
531 static int force_cpu_type(const char *str, struct kernel_param *kp)
533 if (!strcmp(str, "arch_perfmon")) {
534 force_arch_perfmon = 1;
535 printk(KERN_INFO "oprofile: forcing architectural perfmon\n");
540 module_param_call(cpu_type, force_cpu_type, NULL, NULL, 0);
542 static int __init ppro_init(char **cpu_type)
544 __u8 cpu_model = boot_cpu_data.x86_model;
545 struct op_x86_model_spec *spec = &op_ppro_spec; /* default */
547 if (force_arch_perfmon && cpu_has_arch_perfmon)
552 *cpu_type = "i386/ppro";
555 *cpu_type = "i386/pii";
559 *cpu_type = "i386/piii";
563 *cpu_type = "i386/p6_mobile";
566 *cpu_type = "i386/core";
569 *cpu_type = "i386/core_2";
572 spec = &op_arch_perfmon_spec;
573 *cpu_type = "i386/core_i7";
576 *cpu_type = "i386/atom";
587 /* in order to get sysfs right */
588 static int using_nmi;
590 int __init op_nmi_init(struct oprofile_operations *ops)
592 __u8 vendor = boot_cpu_data.x86_vendor;
593 __u8 family = boot_cpu_data.x86;
594 char *cpu_type = NULL;
602 /* Needs to be at least an Athlon (or hammer in 32bit mode) */
606 cpu_type = "i386/athlon";
610 * Actually it could be i386/hammer too, but
611 * give user space an consistent name.
613 cpu_type = "x86-64/hammer";
616 cpu_type = "x86-64/family10";
619 cpu_type = "x86-64/family11h";
624 model = &op_amd_spec;
627 case X86_VENDOR_INTEL:
634 /* A P6-class processor */
636 ppro_init(&cpu_type);
646 if (!cpu_has_arch_perfmon)
649 /* use arch perfmon as fallback */
650 cpu_type = "i386/arch_perfmon";
651 model = &op_arch_perfmon_spec;
659 register_cpu_notifier(&oprofile_cpu_nb);
661 /* default values, can be overwritten by model */
662 ops->create_files = nmi_create_files;
663 ops->setup = nmi_setup;
664 ops->shutdown = nmi_shutdown;
665 ops->start = nmi_start;
666 ops->stop = nmi_stop;
667 ops->cpu_type = cpu_type;
668 #ifdef CONFIG_OPROFILE_EVENT_MULTIPLEX
669 ops->switch_events = nmi_switch_event;
673 ret = model->init(ops);
677 if (!model->num_virt_counters)
678 model->num_virt_counters = model->num_counters;
682 printk(KERN_INFO "oprofile: using NMI interrupt.\n");
686 void op_nmi_exit(void)
691 unregister_cpu_notifier(&oprofile_cpu_nb);