Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /** |
2 | * @file nmi_int.c | |
3 | * | |
adf5ec0b | 4 | * @remark Copyright 2002-2008 OProfile authors |
1da177e4 LT |
5 | * @remark Read the file COPYING |
6 | * | |
7 | * @author John Levon <levon@movementarian.org> | |
adf5ec0b | 8 | * @author Robert Richter <robert.richter@amd.com> |
1da177e4 LT |
9 | */ |
10 | ||
11 | #include <linux/init.h> | |
12 | #include <linux/notifier.h> | |
13 | #include <linux/smp.h> | |
14 | #include <linux/oprofile.h> | |
15 | #include <linux/sysdev.h> | |
16 | #include <linux/slab.h> | |
1cfcea1b | 17 | #include <linux/moduleparam.h> |
1eeb66a1 | 18 | #include <linux/kdebug.h> |
1da177e4 LT |
19 | #include <asm/nmi.h> |
20 | #include <asm/msr.h> | |
21 | #include <asm/apic.h> | |
b75f53db | 22 | |
1da177e4 LT |
23 | #include "op_counter.h" |
24 | #include "op_x86_model.h" | |
2fbe7b25 | 25 | |
b75f53db | 26 | static struct op_x86_model_spec const *model; |
d18d00f5 MT |
27 | static DEFINE_PER_CPU(struct op_msrs, cpu_msrs); |
28 | static DEFINE_PER_CPU(unsigned long, saved_lvtpc); | |
2fbe7b25 | 29 | |
1da177e4 LT |
30 | static int nmi_start(void); |
31 | static void nmi_stop(void); | |
32 | ||
33 | /* 0 == registered but off, 1 == registered and on */ | |
34 | static int nmi_enabled = 0; | |
35 | ||
36 | #ifdef CONFIG_PM | |
37 | ||
438510f6 | 38 | static int nmi_suspend(struct sys_device *dev, pm_message_t state) |
1da177e4 LT |
39 | { |
40 | if (nmi_enabled == 1) | |
41 | nmi_stop(); | |
42 | return 0; | |
43 | } | |
44 | ||
1da177e4 LT |
45 | static int nmi_resume(struct sys_device *dev) |
46 | { | |
47 | if (nmi_enabled == 1) | |
48 | nmi_start(); | |
49 | return 0; | |
50 | } | |
51 | ||
1da177e4 | 52 | static struct sysdev_class oprofile_sysclass = { |
af5ca3f4 | 53 | .name = "oprofile", |
1da177e4 LT |
54 | .resume = nmi_resume, |
55 | .suspend = nmi_suspend, | |
56 | }; | |
57 | ||
1da177e4 LT |
58 | static struct sys_device device_oprofile = { |
59 | .id = 0, | |
60 | .cls = &oprofile_sysclass, | |
61 | }; | |
62 | ||
405ae7d3 | 63 | static int __init init_sysfs(void) |
1da177e4 LT |
64 | { |
65 | int error; | |
b75f53db CM |
66 | |
67 | error = sysdev_class_register(&oprofile_sysclass); | |
68 | if (!error) | |
1da177e4 LT |
69 | error = sysdev_register(&device_oprofile); |
70 | return error; | |
71 | } | |
72 | ||
405ae7d3 | 73 | static void exit_sysfs(void) |
1da177e4 LT |
74 | { |
75 | sysdev_unregister(&device_oprofile); | |
76 | sysdev_class_unregister(&oprofile_sysclass); | |
77 | } | |
78 | ||
79 | #else | |
405ae7d3 RD |
80 | #define init_sysfs() do { } while (0) |
81 | #define exit_sysfs() do { } while (0) | |
1da177e4 LT |
82 | #endif /* CONFIG_PM */ |
83 | ||
c7c19f8e AB |
84 | static int profile_exceptions_notify(struct notifier_block *self, |
85 | unsigned long val, void *data) | |
1da177e4 | 86 | { |
2fbe7b25 DZ |
87 | struct die_args *args = (struct die_args *)data; |
88 | int ret = NOTIFY_DONE; | |
89 | int cpu = smp_processor_id(); | |
90 | ||
b75f53db | 91 | switch (val) { |
2fbe7b25 | 92 | case DIE_NMI: |
d18d00f5 | 93 | if (model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu))) |
2fbe7b25 DZ |
94 | ret = NOTIFY_STOP; |
95 | break; | |
96 | default: | |
97 | break; | |
98 | } | |
99 | return ret; | |
1da177e4 | 100 | } |
2fbe7b25 | 101 | |
b75f53db | 102 | static void nmi_cpu_save_registers(struct op_msrs *msrs) |
1da177e4 LT |
103 | { |
104 | unsigned int const nr_ctrs = model->num_counters; | |
b75f53db CM |
105 | unsigned int const nr_ctrls = model->num_controls; |
106 | struct op_msr *counters = msrs->counters; | |
107 | struct op_msr *controls = msrs->controls; | |
1da177e4 LT |
108 | unsigned int i; |
109 | ||
110 | for (i = 0; i < nr_ctrs; ++i) { | |
b75f53db | 111 | if (counters[i].addr) { |
cb9c448c DZ |
112 | rdmsr(counters[i].addr, |
113 | counters[i].saved.low, | |
114 | counters[i].saved.high); | |
115 | } | |
1da177e4 | 116 | } |
b75f53db | 117 | |
1da177e4 | 118 | for (i = 0; i < nr_ctrls; ++i) { |
b75f53db | 119 | if (controls[i].addr) { |
cb9c448c DZ |
120 | rdmsr(controls[i].addr, |
121 | controls[i].saved.low, | |
122 | controls[i].saved.high); | |
123 | } | |
1da177e4 LT |
124 | } |
125 | } | |
126 | ||
b75f53db | 127 | static void nmi_save_registers(void *dummy) |
1da177e4 LT |
128 | { |
129 | int cpu = smp_processor_id(); | |
d18d00f5 | 130 | struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu); |
1da177e4 LT |
131 | nmi_cpu_save_registers(msrs); |
132 | } | |
133 | ||
1da177e4 LT |
134 | static void free_msrs(void) |
135 | { | |
136 | int i; | |
c8912599 | 137 | for_each_possible_cpu(i) { |
d18d00f5 MT |
138 | kfree(per_cpu(cpu_msrs, i).counters); |
139 | per_cpu(cpu_msrs, i).counters = NULL; | |
140 | kfree(per_cpu(cpu_msrs, i).controls); | |
141 | per_cpu(cpu_msrs, i).controls = NULL; | |
1da177e4 LT |
142 | } |
143 | } | |
144 | ||
1da177e4 LT |
145 | static int allocate_msrs(void) |
146 | { | |
147 | int success = 1; | |
148 | size_t controls_size = sizeof(struct op_msr) * model->num_controls; | |
149 | size_t counters_size = sizeof(struct op_msr) * model->num_counters; | |
150 | ||
151 | int i; | |
0939c17c | 152 | for_each_possible_cpu(i) { |
d18d00f5 MT |
153 | per_cpu(cpu_msrs, i).counters = kmalloc(counters_size, |
154 | GFP_KERNEL); | |
155 | if (!per_cpu(cpu_msrs, i).counters) { | |
1da177e4 LT |
156 | success = 0; |
157 | break; | |
158 | } | |
d18d00f5 MT |
159 | per_cpu(cpu_msrs, i).controls = kmalloc(controls_size, |
160 | GFP_KERNEL); | |
161 | if (!per_cpu(cpu_msrs, i).controls) { | |
1da177e4 LT |
162 | success = 0; |
163 | break; | |
164 | } | |
165 | } | |
166 | ||
167 | if (!success) | |
168 | free_msrs(); | |
169 | ||
170 | return success; | |
171 | } | |
172 | ||
b75f53db | 173 | static void nmi_cpu_setup(void *dummy) |
1da177e4 LT |
174 | { |
175 | int cpu = smp_processor_id(); | |
d18d00f5 | 176 | struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu); |
1da177e4 LT |
177 | spin_lock(&oprofilefs_lock); |
178 | model->setup_ctrs(msrs); | |
179 | spin_unlock(&oprofilefs_lock); | |
d18d00f5 | 180 | per_cpu(saved_lvtpc, cpu) = apic_read(APIC_LVTPC); |
1da177e4 LT |
181 | apic_write(APIC_LVTPC, APIC_DM_NMI); |
182 | } | |
183 | ||
2fbe7b25 DZ |
184 | static struct notifier_block profile_exceptions_nb = { |
185 | .notifier_call = profile_exceptions_notify, | |
186 | .next = NULL, | |
187 | .priority = 0 | |
188 | }; | |
1da177e4 LT |
189 | |
190 | static int nmi_setup(void) | |
191 | { | |
b75f53db | 192 | int err = 0; |
6c977aad | 193 | int cpu; |
2fbe7b25 | 194 | |
1da177e4 LT |
195 | if (!allocate_msrs()) |
196 | return -ENOMEM; | |
197 | ||
b75f53db CM |
198 | err = register_die_notifier(&profile_exceptions_nb); |
199 | if (err) { | |
1da177e4 | 200 | free_msrs(); |
2fbe7b25 | 201 | return err; |
1da177e4 | 202 | } |
2fbe7b25 | 203 | |
1da177e4 LT |
204 | /* We need to serialize save and setup for HT because the subset |
205 | * of msrs are distinct for save and setup operations | |
206 | */ | |
6c977aad AK |
207 | |
208 | /* Assume saved/restored counters are the same on all CPUs */ | |
d18d00f5 | 209 | model->fill_in_addresses(&per_cpu(cpu_msrs, 0)); |
b75f53db | 210 | for_each_possible_cpu(cpu) { |
0939c17c | 211 | if (cpu != 0) { |
d18d00f5 MT |
212 | memcpy(per_cpu(cpu_msrs, cpu).counters, |
213 | per_cpu(cpu_msrs, 0).counters, | |
0939c17c CW |
214 | sizeof(struct op_msr) * model->num_counters); |
215 | ||
d18d00f5 MT |
216 | memcpy(per_cpu(cpu_msrs, cpu).controls, |
217 | per_cpu(cpu_msrs, 0).controls, | |
0939c17c CW |
218 | sizeof(struct op_msr) * model->num_controls); |
219 | } | |
220 | ||
6c977aad | 221 | } |
15c8b6c1 JA |
222 | on_each_cpu(nmi_save_registers, NULL, 1); |
223 | on_each_cpu(nmi_cpu_setup, NULL, 1); | |
1da177e4 LT |
224 | nmi_enabled = 1; |
225 | return 0; | |
226 | } | |
227 | ||
b75f53db | 228 | static void nmi_restore_registers(struct op_msrs *msrs) |
1da177e4 LT |
229 | { |
230 | unsigned int const nr_ctrs = model->num_counters; | |
b75f53db CM |
231 | unsigned int const nr_ctrls = model->num_controls; |
232 | struct op_msr *counters = msrs->counters; | |
233 | struct op_msr *controls = msrs->controls; | |
1da177e4 LT |
234 | unsigned int i; |
235 | ||
236 | for (i = 0; i < nr_ctrls; ++i) { | |
b75f53db | 237 | if (controls[i].addr) { |
cb9c448c DZ |
238 | wrmsr(controls[i].addr, |
239 | controls[i].saved.low, | |
240 | controls[i].saved.high); | |
241 | } | |
1da177e4 | 242 | } |
b75f53db | 243 | |
1da177e4 | 244 | for (i = 0; i < nr_ctrs; ++i) { |
b75f53db | 245 | if (counters[i].addr) { |
cb9c448c DZ |
246 | wrmsr(counters[i].addr, |
247 | counters[i].saved.low, | |
248 | counters[i].saved.high); | |
249 | } | |
1da177e4 LT |
250 | } |
251 | } | |
1da177e4 | 252 | |
b75f53db | 253 | static void nmi_cpu_shutdown(void *dummy) |
1da177e4 LT |
254 | { |
255 | unsigned int v; | |
256 | int cpu = smp_processor_id(); | |
d18d00f5 | 257 | struct op_msrs *msrs = &__get_cpu_var(cpu_msrs); |
b75f53db | 258 | |
1da177e4 LT |
259 | /* restoring APIC_LVTPC can trigger an apic error because the delivery |
260 | * mode and vector nr combination can be illegal. That's by design: on | |
261 | * power on apic lvt contain a zero vector nr which are legal only for | |
262 | * NMI delivery mode. So inhibit apic err before restoring lvtpc | |
263 | */ | |
264 | v = apic_read(APIC_LVTERR); | |
265 | apic_write(APIC_LVTERR, v | APIC_LVT_MASKED); | |
d18d00f5 | 266 | apic_write(APIC_LVTPC, per_cpu(saved_lvtpc, cpu)); |
1da177e4 LT |
267 | apic_write(APIC_LVTERR, v); |
268 | nmi_restore_registers(msrs); | |
269 | } | |
270 | ||
1da177e4 LT |
271 | static void nmi_shutdown(void) |
272 | { | |
93e1ade5 | 273 | struct op_msrs *msrs = &get_cpu_var(cpu_msrs); |
1da177e4 | 274 | nmi_enabled = 0; |
15c8b6c1 | 275 | on_each_cpu(nmi_cpu_shutdown, NULL, 1); |
2fbe7b25 | 276 | unregister_die_notifier(&profile_exceptions_nb); |
d18d00f5 | 277 | model->shutdown(msrs); |
1da177e4 | 278 | free_msrs(); |
93e1ade5 | 279 | put_cpu_var(cpu_msrs); |
1da177e4 LT |
280 | } |
281 | ||
b75f53db | 282 | static void nmi_cpu_start(void *dummy) |
1da177e4 | 283 | { |
d18d00f5 | 284 | struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs); |
1da177e4 LT |
285 | model->start(msrs); |
286 | } | |
1da177e4 LT |
287 | |
288 | static int nmi_start(void) | |
289 | { | |
15c8b6c1 | 290 | on_each_cpu(nmi_cpu_start, NULL, 1); |
1da177e4 LT |
291 | return 0; |
292 | } | |
b75f53db CM |
293 | |
294 | static void nmi_cpu_stop(void *dummy) | |
1da177e4 | 295 | { |
d18d00f5 | 296 | struct op_msrs const *msrs = &__get_cpu_var(cpu_msrs); |
1da177e4 LT |
297 | model->stop(msrs); |
298 | } | |
b75f53db | 299 | |
1da177e4 LT |
300 | static void nmi_stop(void) |
301 | { | |
15c8b6c1 | 302 | on_each_cpu(nmi_cpu_stop, NULL, 1); |
1da177e4 LT |
303 | } |
304 | ||
1da177e4 LT |
305 | struct op_counter_config counter_config[OP_MAX_COUNTER]; |
306 | ||
b75f53db | 307 | static int nmi_create_files(struct super_block *sb, struct dentry *root) |
1da177e4 LT |
308 | { |
309 | unsigned int i; | |
310 | ||
311 | for (i = 0; i < model->num_counters; ++i) { | |
b75f53db | 312 | struct dentry *dir; |
0c6856f7 | 313 | char buf[4]; |
b75f53db CM |
314 | |
315 | /* quick little hack to _not_ expose a counter if it is not | |
cb9c448c DZ |
316 | * available for use. This should protect userspace app. |
317 | * NOTE: assumes 1:1 mapping here (that counters are organized | |
318 | * sequentially in their struct assignment). | |
319 | */ | |
320 | if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i))) | |
321 | continue; | |
322 | ||
0c6856f7 | 323 | snprintf(buf, sizeof(buf), "%d", i); |
1da177e4 | 324 | dir = oprofilefs_mkdir(sb, root, buf); |
b75f53db CM |
325 | oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled); |
326 | oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event); | |
327 | oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count); | |
328 | oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask); | |
329 | oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel); | |
330 | oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user); | |
1da177e4 LT |
331 | } |
332 | ||
333 | return 0; | |
334 | } | |
b75f53db | 335 | |
1cfcea1b AK |
336 | static int p4force; |
337 | module_param(p4force, int, 0); | |
b75f53db CM |
338 | |
339 | static int __init p4_init(char **cpu_type) | |
1da177e4 LT |
340 | { |
341 | __u8 cpu_model = boot_cpu_data.x86_model; | |
342 | ||
1cfcea1b | 343 | if (!p4force && (cpu_model > 6 || cpu_model == 5)) |
1da177e4 LT |
344 | return 0; |
345 | ||
346 | #ifndef CONFIG_SMP | |
347 | *cpu_type = "i386/p4"; | |
348 | model = &op_p4_spec; | |
349 | return 1; | |
350 | #else | |
351 | switch (smp_num_siblings) { | |
b75f53db CM |
352 | case 1: |
353 | *cpu_type = "i386/p4"; | |
354 | model = &op_p4_spec; | |
355 | return 1; | |
356 | ||
357 | case 2: | |
358 | *cpu_type = "i386/p4-ht"; | |
359 | model = &op_p4_ht2_spec; | |
360 | return 1; | |
1da177e4 LT |
361 | } |
362 | #endif | |
363 | ||
364 | printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n"); | |
365 | printk(KERN_INFO "oprofile: Reverting to timer mode.\n"); | |
366 | return 0; | |
367 | } | |
368 | ||
b75f53db | 369 | static int __init ppro_init(char **cpu_type) |
1da177e4 LT |
370 | { |
371 | __u8 cpu_model = boot_cpu_data.x86_model; | |
372 | ||
4b9f12a3 LT |
373 | switch (cpu_model) { |
374 | case 0 ... 2: | |
375 | *cpu_type = "i386/ppro"; | |
376 | break; | |
377 | case 3 ... 5: | |
378 | *cpu_type = "i386/pii"; | |
379 | break; | |
380 | case 6 ... 8: | |
381 | *cpu_type = "i386/piii"; | |
382 | break; | |
383 | case 9: | |
384 | *cpu_type = "i386/p6_mobile"; | |
385 | break; | |
386 | case 10 ... 13: | |
387 | *cpu_type = "i386/p6"; | |
388 | break; | |
389 | case 14: | |
64471ebe | 390 | *cpu_type = "i386/core"; |
4b9f12a3 LT |
391 | break; |
392 | case 15: case 23: | |
393 | *cpu_type = "i386/core_2"; | |
394 | break; | |
395 | case 26: | |
f04b92e9 | 396 | *cpu_type = "i386/core_2"; |
4b9f12a3 LT |
397 | break; |
398 | default: | |
399 | /* Unknown */ | |
1da177e4 | 400 | return 0; |
1da177e4 LT |
401 | } |
402 | ||
403 | model = &op_ppro_spec; | |
404 | return 1; | |
405 | } | |
406 | ||
405ae7d3 | 407 | /* in order to get sysfs right */ |
1da177e4 LT |
408 | static int using_nmi; |
409 | ||
96d0821c | 410 | int __init op_nmi_init(struct oprofile_operations *ops) |
1da177e4 LT |
411 | { |
412 | __u8 vendor = boot_cpu_data.x86_vendor; | |
413 | __u8 family = boot_cpu_data.x86; | |
414 | char *cpu_type; | |
adf5ec0b | 415 | int ret = 0; |
1da177e4 LT |
416 | |
417 | if (!cpu_has_apic) | |
418 | return -ENODEV; | |
b75f53db | 419 | |
1da177e4 | 420 | switch (vendor) { |
b75f53db CM |
421 | case X86_VENDOR_AMD: |
422 | /* Needs to be at least an Athlon (or hammer in 32bit mode) */ | |
1da177e4 | 423 | |
b75f53db CM |
424 | switch (family) { |
425 | default: | |
426 | return -ENODEV; | |
427 | case 6: | |
428 | model = &op_athlon_spec; | |
429 | cpu_type = "i386/athlon"; | |
430 | break; | |
431 | case 0xf: | |
432 | model = &op_athlon_spec; | |
433 | /* Actually it could be i386/hammer too, but give | |
434 | user space an consistent name. */ | |
435 | cpu_type = "x86-64/hammer"; | |
436 | break; | |
437 | case 0x10: | |
438 | model = &op_athlon_spec; | |
439 | cpu_type = "x86-64/family10"; | |
440 | break; | |
12f2b261 BK |
441 | case 0x11: |
442 | model = &op_athlon_spec; | |
443 | cpu_type = "x86-64/family11h"; | |
444 | break; | |
b75f53db CM |
445 | } |
446 | break; | |
447 | ||
448 | case X86_VENDOR_INTEL: | |
449 | switch (family) { | |
450 | /* Pentium IV */ | |
451 | case 0xf: | |
452 | if (!p4_init(&cpu_type)) | |
1da177e4 | 453 | return -ENODEV; |
1da177e4 | 454 | break; |
b75f53db CM |
455 | |
456 | /* A P6-class processor */ | |
457 | case 6: | |
458 | if (!ppro_init(&cpu_type)) | |
459 | return -ENODEV; | |
1da177e4 LT |
460 | break; |
461 | ||
462 | default: | |
463 | return -ENODEV; | |
b75f53db CM |
464 | } |
465 | break; | |
466 | ||
467 | default: | |
468 | return -ENODEV; | |
1da177e4 LT |
469 | } |
470 | ||
adf5ec0b RR |
471 | if (model->init) |
472 | ret = model->init(ops); | |
473 | if (ret) | |
474 | return ret; | |
475 | ||
405ae7d3 | 476 | init_sysfs(); |
1da177e4 LT |
477 | using_nmi = 1; |
478 | ops->create_files = nmi_create_files; | |
479 | ops->setup = nmi_setup; | |
480 | ops->shutdown = nmi_shutdown; | |
481 | ops->start = nmi_start; | |
482 | ops->stop = nmi_stop; | |
483 | ops->cpu_type = cpu_type; | |
484 | printk(KERN_INFO "oprofile: using NMI interrupt.\n"); | |
485 | return 0; | |
486 | } | |
487 | ||
96d0821c | 488 | void op_nmi_exit(void) |
1da177e4 LT |
489 | { |
490 | if (using_nmi) | |
405ae7d3 | 491 | exit_sysfs(); |
adf5ec0b RR |
492 | if (model->exit) |
493 | model->exit(); | |
1da177e4 | 494 | } |