Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /** |
2 | * @file nmi_int.c | |
3 | * | |
4 | * @remark Copyright 2002 OProfile authors | |
5 | * @remark Read the file COPYING | |
6 | * | |
7 | * @author John Levon <levon@movementarian.org> | |
8 | */ | |
9 | ||
10 | #include <linux/init.h> | |
11 | #include <linux/notifier.h> | |
12 | #include <linux/smp.h> | |
13 | #include <linux/oprofile.h> | |
14 | #include <linux/sysdev.h> | |
15 | #include <linux/slab.h> | |
1cfcea1b | 16 | #include <linux/moduleparam.h> |
1eeb66a1 | 17 | #include <linux/kdebug.h> |
1da177e4 LT |
18 | #include <asm/nmi.h> |
19 | #include <asm/msr.h> | |
20 | #include <asm/apic.h> | |
b75f53db | 21 | |
1da177e4 LT |
22 | #include "op_counter.h" |
23 | #include "op_x86_model.h" | |
2fbe7b25 | 24 | |
b75f53db | 25 | static struct op_x86_model_spec const *model; |
1da177e4 LT |
26 | static struct op_msrs cpu_msrs[NR_CPUS]; |
27 | static unsigned long saved_lvtpc[NR_CPUS]; | |
2fbe7b25 | 28 | |
1da177e4 LT |
29 | static int nmi_start(void); |
30 | static void nmi_stop(void); | |
31 | ||
32 | /* 0 == registered but off, 1 == registered and on */ | |
33 | static int nmi_enabled = 0; | |
34 | ||
35 | #ifdef CONFIG_PM | |
36 | ||
438510f6 | 37 | static int nmi_suspend(struct sys_device *dev, pm_message_t state) |
1da177e4 LT |
38 | { |
39 | if (nmi_enabled == 1) | |
40 | nmi_stop(); | |
41 | return 0; | |
42 | } | |
43 | ||
1da177e4 LT |
44 | static int nmi_resume(struct sys_device *dev) |
45 | { | |
46 | if (nmi_enabled == 1) | |
47 | nmi_start(); | |
48 | return 0; | |
49 | } | |
50 | ||
1da177e4 | 51 | static struct sysdev_class oprofile_sysclass = { |
af5ca3f4 | 52 | .name = "oprofile", |
1da177e4 LT |
53 | .resume = nmi_resume, |
54 | .suspend = nmi_suspend, | |
55 | }; | |
56 | ||
1da177e4 LT |
57 | static struct sys_device device_oprofile = { |
58 | .id = 0, | |
59 | .cls = &oprofile_sysclass, | |
60 | }; | |
61 | ||
405ae7d3 | 62 | static int __init init_sysfs(void) |
1da177e4 LT |
63 | { |
64 | int error; | |
b75f53db CM |
65 | |
66 | error = sysdev_class_register(&oprofile_sysclass); | |
67 | if (!error) | |
1da177e4 LT |
68 | error = sysdev_register(&device_oprofile); |
69 | return error; | |
70 | } | |
71 | ||
405ae7d3 | 72 | static void exit_sysfs(void) |
1da177e4 LT |
73 | { |
74 | sysdev_unregister(&device_oprofile); | |
75 | sysdev_class_unregister(&oprofile_sysclass); | |
76 | } | |
77 | ||
78 | #else | |
405ae7d3 RD |
79 | #define init_sysfs() do { } while (0) |
80 | #define exit_sysfs() do { } while (0) | |
1da177e4 LT |
81 | #endif /* CONFIG_PM */ |
82 | ||
c7c19f8e AB |
83 | static int profile_exceptions_notify(struct notifier_block *self, |
84 | unsigned long val, void *data) | |
1da177e4 | 85 | { |
2fbe7b25 DZ |
86 | struct die_args *args = (struct die_args *)data; |
87 | int ret = NOTIFY_DONE; | |
88 | int cpu = smp_processor_id(); | |
89 | ||
b75f53db | 90 | switch (val) { |
2fbe7b25 DZ |
91 | case DIE_NMI: |
92 | if (model->check_ctrs(args->regs, &cpu_msrs[cpu])) | |
93 | ret = NOTIFY_STOP; | |
94 | break; | |
95 | default: | |
96 | break; | |
97 | } | |
98 | return ret; | |
1da177e4 | 99 | } |
2fbe7b25 | 100 | |
b75f53db | 101 | static void nmi_cpu_save_registers(struct op_msrs *msrs) |
1da177e4 LT |
102 | { |
103 | unsigned int const nr_ctrs = model->num_counters; | |
b75f53db CM |
104 | unsigned int const nr_ctrls = model->num_controls; |
105 | struct op_msr *counters = msrs->counters; | |
106 | struct op_msr *controls = msrs->controls; | |
1da177e4 LT |
107 | unsigned int i; |
108 | ||
109 | for (i = 0; i < nr_ctrs; ++i) { | |
b75f53db | 110 | if (counters[i].addr) { |
cb9c448c DZ |
111 | rdmsr(counters[i].addr, |
112 | counters[i].saved.low, | |
113 | counters[i].saved.high); | |
114 | } | |
1da177e4 | 115 | } |
b75f53db | 116 | |
1da177e4 | 117 | for (i = 0; i < nr_ctrls; ++i) { |
b75f53db | 118 | if (controls[i].addr) { |
cb9c448c DZ |
119 | rdmsr(controls[i].addr, |
120 | controls[i].saved.low, | |
121 | controls[i].saved.high); | |
122 | } | |
1da177e4 LT |
123 | } |
124 | } | |
125 | ||
b75f53db | 126 | static void nmi_save_registers(void *dummy) |
1da177e4 LT |
127 | { |
128 | int cpu = smp_processor_id(); | |
b75f53db | 129 | struct op_msrs *msrs = &cpu_msrs[cpu]; |
1da177e4 LT |
130 | nmi_cpu_save_registers(msrs); |
131 | } | |
132 | ||
1da177e4 LT |
133 | static void free_msrs(void) |
134 | { | |
135 | int i; | |
c8912599 | 136 | for_each_possible_cpu(i) { |
1da177e4 LT |
137 | kfree(cpu_msrs[i].counters); |
138 | cpu_msrs[i].counters = NULL; | |
139 | kfree(cpu_msrs[i].controls); | |
140 | cpu_msrs[i].controls = NULL; | |
141 | } | |
142 | } | |
143 | ||
1da177e4 LT |
144 | static int allocate_msrs(void) |
145 | { | |
146 | int success = 1; | |
147 | size_t controls_size = sizeof(struct op_msr) * model->num_controls; | |
148 | size_t counters_size = sizeof(struct op_msr) * model->num_counters; | |
149 | ||
150 | int i; | |
0939c17c | 151 | for_each_possible_cpu(i) { |
1da177e4 LT |
152 | cpu_msrs[i].counters = kmalloc(counters_size, GFP_KERNEL); |
153 | if (!cpu_msrs[i].counters) { | |
154 | success = 0; | |
155 | break; | |
156 | } | |
157 | cpu_msrs[i].controls = kmalloc(controls_size, GFP_KERNEL); | |
158 | if (!cpu_msrs[i].controls) { | |
159 | success = 0; | |
160 | break; | |
161 | } | |
162 | } | |
163 | ||
164 | if (!success) | |
165 | free_msrs(); | |
166 | ||
167 | return success; | |
168 | } | |
169 | ||
b75f53db | 170 | static void nmi_cpu_setup(void *dummy) |
1da177e4 LT |
171 | { |
172 | int cpu = smp_processor_id(); | |
b75f53db | 173 | struct op_msrs *msrs = &cpu_msrs[cpu]; |
1da177e4 LT |
174 | spin_lock(&oprofilefs_lock); |
175 | model->setup_ctrs(msrs); | |
176 | spin_unlock(&oprofilefs_lock); | |
177 | saved_lvtpc[cpu] = apic_read(APIC_LVTPC); | |
178 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
179 | } | |
180 | ||
2fbe7b25 DZ |
181 | static struct notifier_block profile_exceptions_nb = { |
182 | .notifier_call = profile_exceptions_notify, | |
183 | .next = NULL, | |
184 | .priority = 0 | |
185 | }; | |
1da177e4 LT |
186 | |
187 | static int nmi_setup(void) | |
188 | { | |
b75f53db | 189 | int err = 0; |
6c977aad | 190 | int cpu; |
2fbe7b25 | 191 | |
1da177e4 LT |
192 | if (!allocate_msrs()) |
193 | return -ENOMEM; | |
194 | ||
b75f53db CM |
195 | err = register_die_notifier(&profile_exceptions_nb); |
196 | if (err) { | |
1da177e4 | 197 | free_msrs(); |
2fbe7b25 | 198 | return err; |
1da177e4 | 199 | } |
2fbe7b25 | 200 | |
1da177e4 LT |
201 | /* We need to serialize save and setup for HT because the subset |
202 | * of msrs are distinct for save and setup operations | |
203 | */ | |
6c977aad AK |
204 | |
205 | /* Assume saved/restored counters are the same on all CPUs */ | |
206 | model->fill_in_addresses(&cpu_msrs[0]); | |
b75f53db | 207 | for_each_possible_cpu(cpu) { |
0939c17c CW |
208 | if (cpu != 0) { |
209 | memcpy(cpu_msrs[cpu].counters, cpu_msrs[0].counters, | |
210 | sizeof(struct op_msr) * model->num_counters); | |
211 | ||
212 | memcpy(cpu_msrs[cpu].controls, cpu_msrs[0].controls, | |
213 | sizeof(struct op_msr) * model->num_controls); | |
214 | } | |
215 | ||
6c977aad | 216 | } |
1da177e4 LT |
217 | on_each_cpu(nmi_save_registers, NULL, 0, 1); |
218 | on_each_cpu(nmi_cpu_setup, NULL, 0, 1); | |
1da177e4 LT |
219 | nmi_enabled = 1; |
220 | return 0; | |
221 | } | |
222 | ||
b75f53db | 223 | static void nmi_restore_registers(struct op_msrs *msrs) |
1da177e4 LT |
224 | { |
225 | unsigned int const nr_ctrs = model->num_counters; | |
b75f53db CM |
226 | unsigned int const nr_ctrls = model->num_controls; |
227 | struct op_msr *counters = msrs->counters; | |
228 | struct op_msr *controls = msrs->controls; | |
1da177e4 LT |
229 | unsigned int i; |
230 | ||
231 | for (i = 0; i < nr_ctrls; ++i) { | |
b75f53db | 232 | if (controls[i].addr) { |
cb9c448c DZ |
233 | wrmsr(controls[i].addr, |
234 | controls[i].saved.low, | |
235 | controls[i].saved.high); | |
236 | } | |
1da177e4 | 237 | } |
b75f53db | 238 | |
1da177e4 | 239 | for (i = 0; i < nr_ctrs; ++i) { |
b75f53db | 240 | if (counters[i].addr) { |
cb9c448c DZ |
241 | wrmsr(counters[i].addr, |
242 | counters[i].saved.low, | |
243 | counters[i].saved.high); | |
244 | } | |
1da177e4 LT |
245 | } |
246 | } | |
1da177e4 | 247 | |
b75f53db | 248 | static void nmi_cpu_shutdown(void *dummy) |
1da177e4 LT |
249 | { |
250 | unsigned int v; | |
251 | int cpu = smp_processor_id(); | |
b75f53db CM |
252 | struct op_msrs *msrs = &cpu_msrs[cpu]; |
253 | ||
1da177e4 LT |
254 | /* restoring APIC_LVTPC can trigger an apic error because the delivery |
255 | * mode and vector nr combination can be illegal. That's by design: on | |
256 | * power on apic lvt contain a zero vector nr which are legal only for | |
257 | * NMI delivery mode. So inhibit apic err before restoring lvtpc | |
258 | */ | |
259 | v = apic_read(APIC_LVTERR); | |
260 | apic_write(APIC_LVTERR, v | APIC_LVT_MASKED); | |
261 | apic_write(APIC_LVTPC, saved_lvtpc[cpu]); | |
262 | apic_write(APIC_LVTERR, v); | |
263 | nmi_restore_registers(msrs); | |
264 | } | |
265 | ||
1da177e4 LT |
266 | static void nmi_shutdown(void) |
267 | { | |
268 | nmi_enabled = 0; | |
269 | on_each_cpu(nmi_cpu_shutdown, NULL, 0, 1); | |
2fbe7b25 | 270 | unregister_die_notifier(&profile_exceptions_nb); |
0f8e45a2 | 271 | model->shutdown(cpu_msrs); |
1da177e4 LT |
272 | free_msrs(); |
273 | } | |
274 | ||
b75f53db | 275 | static void nmi_cpu_start(void *dummy) |
1da177e4 | 276 | { |
b75f53db | 277 | struct op_msrs const *msrs = &cpu_msrs[smp_processor_id()]; |
1da177e4 LT |
278 | model->start(msrs); |
279 | } | |
1da177e4 LT |
280 | |
281 | static int nmi_start(void) | |
282 | { | |
283 | on_each_cpu(nmi_cpu_start, NULL, 0, 1); | |
284 | return 0; | |
285 | } | |
b75f53db CM |
286 | |
287 | static void nmi_cpu_stop(void *dummy) | |
1da177e4 | 288 | { |
b75f53db | 289 | struct op_msrs const *msrs = &cpu_msrs[smp_processor_id()]; |
1da177e4 LT |
290 | model->stop(msrs); |
291 | } | |
b75f53db | 292 | |
1da177e4 LT |
293 | static void nmi_stop(void) |
294 | { | |
295 | on_each_cpu(nmi_cpu_stop, NULL, 0, 1); | |
296 | } | |
297 | ||
1da177e4 LT |
298 | struct op_counter_config counter_config[OP_MAX_COUNTER]; |
299 | ||
b75f53db | 300 | static int nmi_create_files(struct super_block *sb, struct dentry *root) |
1da177e4 LT |
301 | { |
302 | unsigned int i; | |
303 | ||
304 | for (i = 0; i < model->num_counters; ++i) { | |
b75f53db | 305 | struct dentry *dir; |
0c6856f7 | 306 | char buf[4]; |
b75f53db CM |
307 | |
308 | /* quick little hack to _not_ expose a counter if it is not | |
cb9c448c DZ |
309 | * available for use. This should protect userspace app. |
310 | * NOTE: assumes 1:1 mapping here (that counters are organized | |
311 | * sequentially in their struct assignment). | |
312 | */ | |
313 | if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i))) | |
314 | continue; | |
315 | ||
0c6856f7 | 316 | snprintf(buf, sizeof(buf), "%d", i); |
1da177e4 | 317 | dir = oprofilefs_mkdir(sb, root, buf); |
b75f53db CM |
318 | oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled); |
319 | oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event); | |
320 | oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count); | |
321 | oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask); | |
322 | oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel); | |
323 | oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user); | |
1da177e4 LT |
324 | } |
325 | ||
326 | return 0; | |
327 | } | |
b75f53db | 328 | |
1cfcea1b AK |
329 | static int p4force; |
330 | module_param(p4force, int, 0); | |
b75f53db CM |
331 | |
332 | static int __init p4_init(char **cpu_type) | |
1da177e4 LT |
333 | { |
334 | __u8 cpu_model = boot_cpu_data.x86_model; | |
335 | ||
1cfcea1b | 336 | if (!p4force && (cpu_model > 6 || cpu_model == 5)) |
1da177e4 LT |
337 | return 0; |
338 | ||
339 | #ifndef CONFIG_SMP | |
340 | *cpu_type = "i386/p4"; | |
341 | model = &op_p4_spec; | |
342 | return 1; | |
343 | #else | |
344 | switch (smp_num_siblings) { | |
b75f53db CM |
345 | case 1: |
346 | *cpu_type = "i386/p4"; | |
347 | model = &op_p4_spec; | |
348 | return 1; | |
349 | ||
350 | case 2: | |
351 | *cpu_type = "i386/p4-ht"; | |
352 | model = &op_p4_ht2_spec; | |
353 | return 1; | |
1da177e4 LT |
354 | } |
355 | #endif | |
356 | ||
357 | printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n"); | |
358 | printk(KERN_INFO "oprofile: Reverting to timer mode.\n"); | |
359 | return 0; | |
360 | } | |
361 | ||
b75f53db | 362 | static int __init ppro_init(char **cpu_type) |
1da177e4 LT |
363 | { |
364 | __u8 cpu_model = boot_cpu_data.x86_model; | |
365 | ||
64471ebe BL |
366 | if (cpu_model == 14) |
367 | *cpu_type = "i386/core"; | |
e107ebe0 | 368 | else if (cpu_model == 15 || cpu_model == 23) |
f04b92e9 | 369 | *cpu_type = "i386/core_2"; |
64471ebe | 370 | else if (cpu_model > 0xd) |
1da177e4 | 371 | return 0; |
64471ebe | 372 | else if (cpu_model == 9) { |
1da177e4 LT |
373 | *cpu_type = "i386/p6_mobile"; |
374 | } else if (cpu_model > 5) { | |
375 | *cpu_type = "i386/piii"; | |
376 | } else if (cpu_model > 2) { | |
377 | *cpu_type = "i386/pii"; | |
378 | } else { | |
379 | *cpu_type = "i386/ppro"; | |
380 | } | |
381 | ||
382 | model = &op_ppro_spec; | |
383 | return 1; | |
384 | } | |
385 | ||
405ae7d3 | 386 | /* in order to get sysfs right */ |
1da177e4 LT |
387 | static int using_nmi; |
388 | ||
96d0821c | 389 | int __init op_nmi_init(struct oprofile_operations *ops) |
1da177e4 LT |
390 | { |
391 | __u8 vendor = boot_cpu_data.x86_vendor; | |
392 | __u8 family = boot_cpu_data.x86; | |
393 | char *cpu_type; | |
394 | ||
395 | if (!cpu_has_apic) | |
396 | return -ENODEV; | |
b75f53db | 397 | |
1da177e4 | 398 | switch (vendor) { |
b75f53db CM |
399 | case X86_VENDOR_AMD: |
400 | /* Needs to be at least an Athlon (or hammer in 32bit mode) */ | |
1da177e4 | 401 | |
b75f53db CM |
402 | switch (family) { |
403 | default: | |
404 | return -ENODEV; | |
405 | case 6: | |
406 | model = &op_athlon_spec; | |
407 | cpu_type = "i386/athlon"; | |
408 | break; | |
409 | case 0xf: | |
410 | model = &op_athlon_spec; | |
411 | /* Actually it could be i386/hammer too, but give | |
412 | user space an consistent name. */ | |
413 | cpu_type = "x86-64/hammer"; | |
414 | break; | |
415 | case 0x10: | |
416 | model = &op_athlon_spec; | |
417 | cpu_type = "x86-64/family10"; | |
418 | break; | |
419 | } | |
420 | break; | |
421 | ||
422 | case X86_VENDOR_INTEL: | |
423 | switch (family) { | |
424 | /* Pentium IV */ | |
425 | case 0xf: | |
426 | if (!p4_init(&cpu_type)) | |
1da177e4 | 427 | return -ENODEV; |
1da177e4 | 428 | break; |
b75f53db CM |
429 | |
430 | /* A P6-class processor */ | |
431 | case 6: | |
432 | if (!ppro_init(&cpu_type)) | |
433 | return -ENODEV; | |
1da177e4 LT |
434 | break; |
435 | ||
436 | default: | |
437 | return -ENODEV; | |
b75f53db CM |
438 | } |
439 | break; | |
440 | ||
441 | default: | |
442 | return -ENODEV; | |
1da177e4 LT |
443 | } |
444 | ||
405ae7d3 | 445 | init_sysfs(); |
1da177e4 LT |
446 | using_nmi = 1; |
447 | ops->create_files = nmi_create_files; | |
448 | ops->setup = nmi_setup; | |
449 | ops->shutdown = nmi_shutdown; | |
450 | ops->start = nmi_start; | |
451 | ops->stop = nmi_stop; | |
452 | ops->cpu_type = cpu_type; | |
453 | printk(KERN_INFO "oprofile: using NMI interrupt.\n"); | |
454 | return 0; | |
455 | } | |
456 | ||
96d0821c | 457 | void op_nmi_exit(void) |
1da177e4 LT |
458 | { |
459 | if (using_nmi) | |
405ae7d3 | 460 | exit_sysfs(); |
1da177e4 | 461 | } |