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 static inline int has_mux(void)
129 return !!model
->switch_ctrl
;
132 inline int op_x86_phys_to_virt(int phys
)
134 return __get_cpu_var(switch_index
) + phys
;
137 static void nmi_shutdown_mux(void)
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;
151 static int nmi_setup_mux(void)
153 size_t multiplex_size
=
154 sizeof(struct op_msr
) * model
->num_virt_counters
;
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
)
170 static void nmi_cpu_setup_mux(int cpu
, struct op_msrs
const * const msrs
)
173 struct op_msr
*multiplex
= msrs
->multiplex
;
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
;
182 multiplex
[i
].addr
= 0;
183 multiplex
[i
].saved
= 0;
187 per_cpu(switch_index
, cpu
) = 0;
190 static void nmi_cpu_save_mpx_registers(struct op_msrs
*msrs
)
192 struct op_msr
*multiplex
= msrs
->multiplex
;
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
);
202 static void nmi_cpu_restore_mpx_registers(struct op_msrs
*msrs
)
204 struct op_msr
*multiplex
= msrs
->multiplex
;
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
);
214 static void nmi_cpu_switch(void *dummy
)
216 int cpu
= smp_processor_id();
217 int si
= per_cpu(switch_index
, cpu
);
218 struct op_msrs
*msrs
= &per_cpu(cpu_msrs
, cpu
);
221 nmi_cpu_save_mpx_registers(msrs
);
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;
228 per_cpu(switch_index
, cpu
) = si
;
230 model
->switch_ctrl(model
, msrs
);
231 nmi_cpu_restore_mpx_registers(msrs
);
238 * Quick check to see if multiplexing is necessary.
239 * The check should be sufficient since counters are used
242 static int nmi_multiplex_on(void)
244 return counter_config
[model
->num_counters
].count
? 0 : -EINVAL
;
247 static int nmi_switch_event(void)
250 return -ENOSYS
; /* not implemented */
251 if (nmi_multiplex_on() < 0)
252 return -EINVAL
; /* not necessary */
254 on_each_cpu(nmi_cpu_switch
, NULL
, 1);
256 atomic_inc(&multiplex_counter
);
261 static inline void mux_init(struct oprofile_operations
*ops
)
264 ops
->switch_events
= nmi_switch_event
;
267 static void mux_clone(int cpu
)
272 memcpy(per_cpu(cpu_msrs
, cpu
).multiplex
,
273 per_cpu(cpu_msrs
, 0).multiplex
,
274 sizeof(struct op_msr
) * model
->num_virt_counters
);
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; }
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
) { }
289 static void free_msrs(void)
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
;
300 static int allocate_msrs(void)
302 size_t controls_size
= sizeof(struct op_msr
) * model
->num_controls
;
303 size_t counters_size
= sizeof(struct op_msr
) * model
->num_counters
;
306 for_each_possible_cpu(i
) {
307 per_cpu(cpu_msrs
, i
).counters
= kmalloc(counters_size
,
309 if (!per_cpu(cpu_msrs
, i
).counters
)
311 per_cpu(cpu_msrs
, i
).controls
= kmalloc(controls_size
,
313 if (!per_cpu(cpu_msrs
, i
).controls
)
320 static void nmi_cpu_setup(void *dummy
)
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
);
333 static struct notifier_block profile_exceptions_nb
= {
334 .notifier_call
= profile_exceptions_notify
,
339 static int nmi_setup(void)
344 if (!allocate_msrs())
346 else if (!nmi_setup_mux())
349 err
= register_die_notifier(&profile_exceptions_nb
);
357 /* We need to serialize save and setup for HT because the subset
358 * of msrs are distinct for save and setup operations
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
) {
367 memcpy(per_cpu(cpu_msrs
, cpu
).counters
,
368 per_cpu(cpu_msrs
, 0).counters
,
369 sizeof(struct op_msr
) * model
->num_counters
);
371 memcpy(per_cpu(cpu_msrs
, cpu
).controls
,
372 per_cpu(cpu_msrs
, 0).controls
,
373 sizeof(struct op_msr
) * model
->num_controls
);
377 on_each_cpu(nmi_cpu_setup
, NULL
, 1);
382 static void nmi_cpu_restore_registers(struct op_msrs
*msrs
)
384 struct op_msr
*counters
= msrs
->counters
;
385 struct op_msr
*controls
= msrs
->controls
;
388 for (i
= 0; i
< model
->num_controls
; ++i
) {
389 if (controls
[i
].addr
)
390 wrmsrl(controls
[i
].addr
, controls
[i
].saved
);
393 for (i
= 0; i
< model
->num_counters
; ++i
) {
394 if (counters
[i
].addr
)
395 wrmsrl(counters
[i
].addr
, counters
[i
].saved
);
399 static void nmi_cpu_shutdown(void *dummy
)
402 int cpu
= smp_processor_id();
403 struct op_msrs
*msrs
= &per_cpu(cpu_msrs
, cpu
);
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
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
);
417 static void nmi_shutdown(void)
419 struct op_msrs
*msrs
;
422 on_each_cpu(nmi_cpu_shutdown
, NULL
, 1);
423 unregister_die_notifier(&profile_exceptions_nb
);
425 msrs
= &get_cpu_var(cpu_msrs
);
426 model
->shutdown(msrs
);
428 put_cpu_var(cpu_msrs
);
431 static int nmi_create_files(struct super_block
*sb
, struct dentry
*root
)
435 for (i
= 0; i
< model
->num_virt_counters
; ++i
) {
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).
445 if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i
)))
447 #endif /* CONFIG_OPROFILE_EVENT_MULTIPLEX */
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
);
463 static int oprofile_cpu_notifier(struct notifier_block
*b
, unsigned long action
,
466 int cpu
= (unsigned long)data
;
468 case CPU_DOWN_FAILED
:
470 smp_call_function_single(cpu
, nmi_cpu_start
, NULL
, 0);
472 case CPU_DOWN_PREPARE
:
473 smp_call_function_single(cpu
, nmi_cpu_stop
, NULL
, 1);
479 static struct notifier_block oprofile_cpu_nb
= {
480 .notifier_call
= oprofile_cpu_notifier
486 static int nmi_suspend(struct sys_device
*dev
, pm_message_t state
)
488 /* Only one CPU left, just stop that one */
489 if (nmi_enabled
== 1)
494 static int nmi_resume(struct sys_device
*dev
)
496 if (nmi_enabled
== 1)
501 static struct sysdev_class oprofile_sysclass
= {
503 .resume
= nmi_resume
,
504 .suspend
= nmi_suspend
,
507 static struct sys_device device_oprofile
= {
509 .cls
= &oprofile_sysclass
,
512 static int __init
init_sysfs(void)
516 error
= sysdev_class_register(&oprofile_sysclass
);
518 error
= sysdev_register(&device_oprofile
);
522 static void exit_sysfs(void)
524 sysdev_unregister(&device_oprofile
);
525 sysdev_class_unregister(&oprofile_sysclass
);
529 #define init_sysfs() do { } while (0)
530 #define exit_sysfs() do { } while (0)
531 #endif /* CONFIG_PM */
533 static int __init
p4_init(char **cpu_type
)
535 __u8 cpu_model
= boot_cpu_data
.x86_model
;
537 if (cpu_model
> 6 || cpu_model
== 5)
541 *cpu_type
= "i386/p4";
545 switch (smp_num_siblings
) {
547 *cpu_type
= "i386/p4";
552 *cpu_type
= "i386/p4-ht";
553 model
= &op_p4_ht2_spec
;
558 printk(KERN_INFO
"oprofile: P4 HyperThreading detected with > 2 threads\n");
559 printk(KERN_INFO
"oprofile: Reverting to timer mode.\n");
563 static int force_arch_perfmon
;
564 static int force_cpu_type(const char *str
, struct kernel_param
*kp
)
566 if (!strcmp(str
, "arch_perfmon")) {
567 force_arch_perfmon
= 1;
568 printk(KERN_INFO
"oprofile: forcing architectural perfmon\n");
573 module_param_call(cpu_type
, force_cpu_type
, NULL
, NULL
, 0);
575 static int __init
ppro_init(char **cpu_type
)
577 __u8 cpu_model
= boot_cpu_data
.x86_model
;
578 struct op_x86_model_spec
*spec
= &op_ppro_spec
; /* default */
580 if (force_arch_perfmon
&& cpu_has_arch_perfmon
)
585 *cpu_type
= "i386/ppro";
588 *cpu_type
= "i386/pii";
592 *cpu_type
= "i386/piii";
596 *cpu_type
= "i386/p6_mobile";
599 *cpu_type
= "i386/core";
602 *cpu_type
= "i386/core_2";
605 spec
= &op_arch_perfmon_spec
;
606 *cpu_type
= "i386/core_i7";
609 *cpu_type
= "i386/atom";
620 /* in order to get sysfs right */
621 static int using_nmi
;
623 int __init
op_nmi_init(struct oprofile_operations
*ops
)
625 __u8 vendor
= boot_cpu_data
.x86_vendor
;
626 __u8 family
= boot_cpu_data
.x86
;
627 char *cpu_type
= NULL
;
635 /* Needs to be at least an Athlon (or hammer in 32bit mode) */
639 cpu_type
= "i386/athlon";
643 * Actually it could be i386/hammer too, but
644 * give user space an consistent name.
646 cpu_type
= "x86-64/hammer";
649 cpu_type
= "x86-64/family10";
652 cpu_type
= "x86-64/family11h";
657 model
= &op_amd_spec
;
660 case X86_VENDOR_INTEL
:
667 /* A P6-class processor */
669 ppro_init(&cpu_type
);
679 if (!cpu_has_arch_perfmon
)
682 /* use arch perfmon as fallback */
683 cpu_type
= "i386/arch_perfmon";
684 model
= &op_arch_perfmon_spec
;
692 register_cpu_notifier(&oprofile_cpu_nb
);
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
;
703 ret
= model
->init(ops
);
707 if (!model
->num_virt_counters
)
708 model
->num_virt_counters
= model
->num_counters
;
714 printk(KERN_INFO
"oprofile: using NMI interrupt.\n");
718 void op_nmi_exit(void)
723 unregister_cpu_notifier(&oprofile_cpu_nb
);