]>
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> |
1da177e4 LT |
17 | #include <asm/nmi.h> |
18 | #include <asm/msr.h> | |
19 | #include <asm/apic.h> | |
20 | ||
21 | #include "op_counter.h" | |
22 | #include "op_x86_model.h" | |
23 | ||
24 | static struct op_x86_model_spec const * model; | |
25 | static struct op_msrs cpu_msrs[NR_CPUS]; | |
26 | static unsigned long saved_lvtpc[NR_CPUS]; | |
27 | ||
28 | static int nmi_start(void); | |
29 | static void nmi_stop(void); | |
30 | ||
31 | /* 0 == registered but off, 1 == registered and on */ | |
32 | static int nmi_enabled = 0; | |
33 | ||
34 | #ifdef CONFIG_PM | |
35 | ||
438510f6 | 36 | static int nmi_suspend(struct sys_device *dev, pm_message_t state) |
1da177e4 LT |
37 | { |
38 | if (nmi_enabled == 1) | |
39 | nmi_stop(); | |
40 | return 0; | |
41 | } | |
42 | ||
43 | ||
44 | static int nmi_resume(struct sys_device *dev) | |
45 | { | |
46 | if (nmi_enabled == 1) | |
47 | nmi_start(); | |
48 | return 0; | |
49 | } | |
50 | ||
51 | ||
52 | static struct sysdev_class oprofile_sysclass = { | |
53 | set_kset_name("oprofile"), | |
54 | .resume = nmi_resume, | |
55 | .suspend = nmi_suspend, | |
56 | }; | |
57 | ||
58 | ||
59 | static struct sys_device device_oprofile = { | |
60 | .id = 0, | |
61 | .cls = &oprofile_sysclass, | |
62 | }; | |
63 | ||
64 | ||
65 | static int __init init_driverfs(void) | |
66 | { | |
67 | int error; | |
68 | if (!(error = sysdev_class_register(&oprofile_sysclass))) | |
69 | error = sysdev_register(&device_oprofile); | |
70 | return error; | |
71 | } | |
72 | ||
73 | ||
74 | static void exit_driverfs(void) | |
75 | { | |
76 | sysdev_unregister(&device_oprofile); | |
77 | sysdev_class_unregister(&oprofile_sysclass); | |
78 | } | |
79 | ||
80 | #else | |
81 | #define init_driverfs() do { } while (0) | |
82 | #define exit_driverfs() do { } while (0) | |
83 | #endif /* CONFIG_PM */ | |
84 | ||
85 | ||
86 | static int nmi_callback(struct pt_regs * regs, int cpu) | |
87 | { | |
88 | return model->check_ctrs(regs, &cpu_msrs[cpu]); | |
89 | } | |
90 | ||
91 | ||
92 | static void nmi_cpu_save_registers(struct op_msrs * msrs) | |
93 | { | |
94 | unsigned int const nr_ctrs = model->num_counters; | |
95 | unsigned int const nr_ctrls = model->num_controls; | |
96 | struct op_msr * counters = msrs->counters; | |
97 | struct op_msr * controls = msrs->controls; | |
98 | unsigned int i; | |
99 | ||
100 | for (i = 0; i < nr_ctrs; ++i) { | |
cb9c448c DZ |
101 | if (counters[i].addr){ |
102 | rdmsr(counters[i].addr, | |
103 | counters[i].saved.low, | |
104 | counters[i].saved.high); | |
105 | } | |
1da177e4 LT |
106 | } |
107 | ||
108 | for (i = 0; i < nr_ctrls; ++i) { | |
cb9c448c DZ |
109 | if (controls[i].addr){ |
110 | rdmsr(controls[i].addr, | |
111 | controls[i].saved.low, | |
112 | controls[i].saved.high); | |
113 | } | |
1da177e4 LT |
114 | } |
115 | } | |
116 | ||
117 | ||
118 | static void nmi_save_registers(void * dummy) | |
119 | { | |
120 | int cpu = smp_processor_id(); | |
121 | struct op_msrs * msrs = &cpu_msrs[cpu]; | |
122 | model->fill_in_addresses(msrs); | |
123 | nmi_cpu_save_registers(msrs); | |
124 | } | |
125 | ||
126 | ||
127 | static void free_msrs(void) | |
128 | { | |
129 | int i; | |
c8912599 | 130 | for_each_possible_cpu(i) { |
1da177e4 LT |
131 | kfree(cpu_msrs[i].counters); |
132 | cpu_msrs[i].counters = NULL; | |
133 | kfree(cpu_msrs[i].controls); | |
134 | cpu_msrs[i].controls = NULL; | |
135 | } | |
136 | } | |
137 | ||
138 | ||
139 | static int allocate_msrs(void) | |
140 | { | |
141 | int success = 1; | |
142 | size_t controls_size = sizeof(struct op_msr) * model->num_controls; | |
143 | size_t counters_size = sizeof(struct op_msr) * model->num_counters; | |
144 | ||
145 | int i; | |
394e3902 | 146 | for_each_online_cpu(i) { |
1da177e4 LT |
147 | cpu_msrs[i].counters = kmalloc(counters_size, GFP_KERNEL); |
148 | if (!cpu_msrs[i].counters) { | |
149 | success = 0; | |
150 | break; | |
151 | } | |
152 | cpu_msrs[i].controls = kmalloc(controls_size, GFP_KERNEL); | |
153 | if (!cpu_msrs[i].controls) { | |
154 | success = 0; | |
155 | break; | |
156 | } | |
157 | } | |
158 | ||
159 | if (!success) | |
160 | free_msrs(); | |
161 | ||
162 | return success; | |
163 | } | |
164 | ||
165 | ||
166 | static void nmi_cpu_setup(void * dummy) | |
167 | { | |
168 | int cpu = smp_processor_id(); | |
169 | struct op_msrs * msrs = &cpu_msrs[cpu]; | |
170 | spin_lock(&oprofilefs_lock); | |
171 | model->setup_ctrs(msrs); | |
172 | spin_unlock(&oprofilefs_lock); | |
173 | saved_lvtpc[cpu] = apic_read(APIC_LVTPC); | |
174 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
175 | } | |
176 | ||
177 | ||
178 | static int nmi_setup(void) | |
179 | { | |
180 | if (!allocate_msrs()) | |
181 | return -ENOMEM; | |
182 | ||
183 | /* We walk a thin line between law and rape here. | |
184 | * We need to be careful to install our NMI handler | |
185 | * without actually triggering any NMIs as this will | |
186 | * break the core code horrifically. | |
187 | */ | |
188 | if (reserve_lapic_nmi() < 0) { | |
189 | free_msrs(); | |
190 | return -EBUSY; | |
191 | } | |
192 | /* We need to serialize save and setup for HT because the subset | |
193 | * of msrs are distinct for save and setup operations | |
194 | */ | |
195 | on_each_cpu(nmi_save_registers, NULL, 0, 1); | |
196 | on_each_cpu(nmi_cpu_setup, NULL, 0, 1); | |
197 | set_nmi_callback(nmi_callback); | |
198 | nmi_enabled = 1; | |
199 | return 0; | |
200 | } | |
201 | ||
202 | ||
203 | static void nmi_restore_registers(struct op_msrs * msrs) | |
204 | { | |
205 | unsigned int const nr_ctrs = model->num_counters; | |
206 | unsigned int const nr_ctrls = model->num_controls; | |
207 | struct op_msr * counters = msrs->counters; | |
208 | struct op_msr * controls = msrs->controls; | |
209 | unsigned int i; | |
210 | ||
211 | for (i = 0; i < nr_ctrls; ++i) { | |
cb9c448c DZ |
212 | if (controls[i].addr){ |
213 | wrmsr(controls[i].addr, | |
214 | controls[i].saved.low, | |
215 | controls[i].saved.high); | |
216 | } | |
1da177e4 LT |
217 | } |
218 | ||
219 | for (i = 0; i < nr_ctrs; ++i) { | |
cb9c448c DZ |
220 | if (counters[i].addr){ |
221 | wrmsr(counters[i].addr, | |
222 | counters[i].saved.low, | |
223 | counters[i].saved.high); | |
224 | } | |
1da177e4 LT |
225 | } |
226 | } | |
227 | ||
228 | ||
229 | static void nmi_cpu_shutdown(void * dummy) | |
230 | { | |
231 | unsigned int v; | |
232 | int cpu = smp_processor_id(); | |
233 | struct op_msrs * msrs = &cpu_msrs[cpu]; | |
234 | ||
235 | /* restoring APIC_LVTPC can trigger an apic error because the delivery | |
236 | * mode and vector nr combination can be illegal. That's by design: on | |
237 | * power on apic lvt contain a zero vector nr which are legal only for | |
238 | * NMI delivery mode. So inhibit apic err before restoring lvtpc | |
239 | */ | |
240 | v = apic_read(APIC_LVTERR); | |
241 | apic_write(APIC_LVTERR, v | APIC_LVT_MASKED); | |
242 | apic_write(APIC_LVTPC, saved_lvtpc[cpu]); | |
243 | apic_write(APIC_LVTERR, v); | |
244 | nmi_restore_registers(msrs); | |
cb9c448c | 245 | model->shutdown(msrs); |
1da177e4 LT |
246 | } |
247 | ||
248 | ||
249 | static void nmi_shutdown(void) | |
250 | { | |
251 | nmi_enabled = 0; | |
252 | on_each_cpu(nmi_cpu_shutdown, NULL, 0, 1); | |
253 | unset_nmi_callback(); | |
254 | release_lapic_nmi(); | |
255 | free_msrs(); | |
256 | } | |
257 | ||
258 | ||
259 | static void nmi_cpu_start(void * dummy) | |
260 | { | |
261 | struct op_msrs const * msrs = &cpu_msrs[smp_processor_id()]; | |
262 | model->start(msrs); | |
263 | } | |
264 | ||
265 | ||
266 | static int nmi_start(void) | |
267 | { | |
268 | on_each_cpu(nmi_cpu_start, NULL, 0, 1); | |
269 | return 0; | |
270 | } | |
271 | ||
272 | ||
273 | static void nmi_cpu_stop(void * dummy) | |
274 | { | |
275 | struct op_msrs const * msrs = &cpu_msrs[smp_processor_id()]; | |
276 | model->stop(msrs); | |
277 | } | |
278 | ||
279 | ||
280 | static void nmi_stop(void) | |
281 | { | |
282 | on_each_cpu(nmi_cpu_stop, NULL, 0, 1); | |
283 | } | |
284 | ||
285 | ||
286 | struct op_counter_config counter_config[OP_MAX_COUNTER]; | |
287 | ||
288 | static int nmi_create_files(struct super_block * sb, struct dentry * root) | |
289 | { | |
290 | unsigned int i; | |
291 | ||
292 | for (i = 0; i < model->num_counters; ++i) { | |
293 | struct dentry * dir; | |
0c6856f7 | 294 | char buf[4]; |
1da177e4 | 295 | |
cb9c448c DZ |
296 | /* quick little hack to _not_ expose a counter if it is not |
297 | * available for use. This should protect userspace app. | |
298 | * NOTE: assumes 1:1 mapping here (that counters are organized | |
299 | * sequentially in their struct assignment). | |
300 | */ | |
301 | if (unlikely(!avail_to_resrv_perfctr_nmi_bit(i))) | |
302 | continue; | |
303 | ||
0c6856f7 | 304 | snprintf(buf, sizeof(buf), "%d", i); |
1da177e4 LT |
305 | dir = oprofilefs_mkdir(sb, root, buf); |
306 | oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled); | |
307 | oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event); | |
308 | oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count); | |
309 | oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask); | |
310 | oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel); | |
311 | oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user); | |
312 | } | |
313 | ||
314 | return 0; | |
315 | } | |
316 | ||
1cfcea1b AK |
317 | static int p4force; |
318 | module_param(p4force, int, 0); | |
1da177e4 LT |
319 | |
320 | static int __init p4_init(char ** cpu_type) | |
321 | { | |
322 | __u8 cpu_model = boot_cpu_data.x86_model; | |
323 | ||
1cfcea1b | 324 | if (!p4force && (cpu_model > 6 || cpu_model == 5)) |
1da177e4 LT |
325 | return 0; |
326 | ||
327 | #ifndef CONFIG_SMP | |
328 | *cpu_type = "i386/p4"; | |
329 | model = &op_p4_spec; | |
330 | return 1; | |
331 | #else | |
332 | switch (smp_num_siblings) { | |
333 | case 1: | |
334 | *cpu_type = "i386/p4"; | |
335 | model = &op_p4_spec; | |
336 | return 1; | |
337 | ||
338 | case 2: | |
339 | *cpu_type = "i386/p4-ht"; | |
340 | model = &op_p4_ht2_spec; | |
341 | return 1; | |
342 | } | |
343 | #endif | |
344 | ||
345 | printk(KERN_INFO "oprofile: P4 HyperThreading detected with > 2 threads\n"); | |
346 | printk(KERN_INFO "oprofile: Reverting to timer mode.\n"); | |
347 | return 0; | |
348 | } | |
349 | ||
350 | ||
351 | static int __init ppro_init(char ** cpu_type) | |
352 | { | |
353 | __u8 cpu_model = boot_cpu_data.x86_model; | |
354 | ||
64471ebe BL |
355 | if (cpu_model == 14) |
356 | *cpu_type = "i386/core"; | |
f04b92e9 BL |
357 | else if (cpu_model == 15) |
358 | *cpu_type = "i386/core_2"; | |
64471ebe | 359 | else if (cpu_model > 0xd) |
1da177e4 | 360 | return 0; |
64471ebe | 361 | else if (cpu_model == 9) { |
1da177e4 LT |
362 | *cpu_type = "i386/p6_mobile"; |
363 | } else if (cpu_model > 5) { | |
364 | *cpu_type = "i386/piii"; | |
365 | } else if (cpu_model > 2) { | |
366 | *cpu_type = "i386/pii"; | |
367 | } else { | |
368 | *cpu_type = "i386/ppro"; | |
369 | } | |
370 | ||
371 | model = &op_ppro_spec; | |
372 | return 1; | |
373 | } | |
374 | ||
375 | /* in order to get driverfs right */ | |
376 | static int using_nmi; | |
377 | ||
96d0821c | 378 | int __init op_nmi_init(struct oprofile_operations *ops) |
1da177e4 LT |
379 | { |
380 | __u8 vendor = boot_cpu_data.x86_vendor; | |
381 | __u8 family = boot_cpu_data.x86; | |
382 | char *cpu_type; | |
383 | ||
384 | if (!cpu_has_apic) | |
385 | return -ENODEV; | |
386 | ||
387 | switch (vendor) { | |
388 | case X86_VENDOR_AMD: | |
389 | /* Needs to be at least an Athlon (or hammer in 32bit mode) */ | |
390 | ||
391 | switch (family) { | |
392 | default: | |
393 | return -ENODEV; | |
394 | case 6: | |
395 | model = &op_athlon_spec; | |
396 | cpu_type = "i386/athlon"; | |
397 | break; | |
398 | case 0xf: | |
399 | model = &op_athlon_spec; | |
400 | /* Actually it could be i386/hammer too, but give | |
401 | user space an consistent name. */ | |
402 | cpu_type = "x86-64/hammer"; | |
403 | break; | |
404 | } | |
405 | break; | |
406 | ||
407 | case X86_VENDOR_INTEL: | |
408 | switch (family) { | |
409 | /* Pentium IV */ | |
410 | case 0xf: | |
411 | if (!p4_init(&cpu_type)) | |
412 | return -ENODEV; | |
413 | break; | |
414 | ||
415 | /* A P6-class processor */ | |
416 | case 6: | |
417 | if (!ppro_init(&cpu_type)) | |
418 | return -ENODEV; | |
419 | break; | |
420 | ||
421 | default: | |
422 | return -ENODEV; | |
423 | } | |
424 | break; | |
425 | ||
426 | default: | |
427 | return -ENODEV; | |
428 | } | |
429 | ||
430 | init_driverfs(); | |
431 | using_nmi = 1; | |
432 | ops->create_files = nmi_create_files; | |
433 | ops->setup = nmi_setup; | |
434 | ops->shutdown = nmi_shutdown; | |
435 | ops->start = nmi_start; | |
436 | ops->stop = nmi_stop; | |
437 | ops->cpu_type = cpu_type; | |
438 | printk(KERN_INFO "oprofile: using NMI interrupt.\n"); | |
439 | return 0; | |
440 | } | |
441 | ||
442 | ||
96d0821c | 443 | void op_nmi_exit(void) |
1da177e4 LT |
444 | { |
445 | if (using_nmi) | |
446 | exit_driverfs(); | |
447 | } |