/root/src/xen/xen/arch/x86/oprofile/nmi_int.c
Line | Count | Source (jump to first uncovered line) |
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 | | * Modified for Xen: by Aravind Menon & Jose Renato Santos |
10 | | * These modifications are: |
11 | | * Copyright (C) 2005 Hewlett-Packard Co. |
12 | | */ |
13 | | |
14 | | #include <xen/event.h> |
15 | | #include <xen/types.h> |
16 | | #include <xen/errno.h> |
17 | | #include <xen/init.h> |
18 | | #include <xen/string.h> |
19 | | #include <xen/delay.h> |
20 | | #include <xen/xenoprof.h> |
21 | | #include <public/xen.h> |
22 | | #include <asm/msr.h> |
23 | | #include <asm/apic.h> |
24 | | #include <asm/regs.h> |
25 | | #include <asm/current.h> |
26 | | #include <asm/nmi.h> |
27 | | |
28 | | #include "op_counter.h" |
29 | | #include "op_x86_model.h" |
30 | | |
31 | | struct op_counter_config counter_config[OP_MAX_COUNTER]; |
32 | | struct op_ibs_config ibs_config; |
33 | | |
34 | | struct op_x86_model_spec const *__read_mostly model; |
35 | | static struct op_msrs cpu_msrs[NR_CPUS]; |
36 | | static unsigned long saved_lvtpc[NR_CPUS]; |
37 | | |
38 | | static char *cpu_type; |
39 | | |
40 | | static int passive_domain_msr_op_checks(unsigned int msr, int *typep, int *indexp) |
41 | 50 | { |
42 | 50 | struct vpmu_struct *vpmu = vcpu_vpmu(current); |
43 | 50 | if ( model == NULL ) |
44 | 0 | return 0; |
45 | 50 | if ( model->is_arch_pmu_msr == NULL ) |
46 | 0 | return 0; |
47 | 50 | if ( !model->is_arch_pmu_msr(msr, typep, indexp) ) |
48 | 50 | return 0; |
49 | 50 | |
50 | 0 | if ( !vpmu_is_set(vpmu, VPMU_PASSIVE_DOMAIN_ALLOCATED) ) |
51 | 0 | if ( ! model->allocated_msr(current) ) |
52 | 0 | return 0; |
53 | 0 | return 1; |
54 | 0 | } |
55 | | |
56 | | int passive_domain_do_rdmsr(unsigned int msr, uint64_t *msr_content) |
57 | 0 | { |
58 | 0 | int type, index; |
59 | 0 |
|
60 | 0 | if ( !passive_domain_msr_op_checks(msr, &type, &index)) |
61 | 0 | return 0; |
62 | 0 |
|
63 | 0 | model->load_msr(current, type, index, msr_content); |
64 | 0 | return 1; |
65 | 0 | } |
66 | | |
67 | | int passive_domain_do_wrmsr(unsigned int msr, uint64_t msr_content) |
68 | 50 | { |
69 | 50 | int type, index; |
70 | 50 | |
71 | 50 | if ( !passive_domain_msr_op_checks(msr, &type, &index)) |
72 | 50 | return 0; |
73 | 50 | |
74 | 0 | model->save_msr(current, type, index, msr_content); |
75 | 0 | return 1; |
76 | 50 | } |
77 | | |
78 | | void passive_domain_destroy(struct vcpu *v) |
79 | 0 | { |
80 | 0 | struct vpmu_struct *vpmu = vcpu_vpmu(v); |
81 | 0 | if ( vpmu_is_set(vpmu, VPMU_PASSIVE_DOMAIN_ALLOCATED) ) |
82 | 0 | model->free_msr(v); |
83 | 0 | } |
84 | | |
85 | | static int nmi_callback(const struct cpu_user_regs *regs, int cpu) |
86 | 0 | { |
87 | 0 | int xen_mode, ovf; |
88 | 0 |
|
89 | 0 | ovf = model->check_ctrs(cpu, &cpu_msrs[cpu], regs); |
90 | 0 | xen_mode = ring_0(regs); |
91 | 0 | if ( ovf && is_active(current->domain) && !xen_mode ) |
92 | 0 | send_guest_vcpu_virq(current, VIRQ_XENOPROF); |
93 | 0 |
|
94 | 0 | if ( ovf == 2 ) |
95 | 0 | current->nmi_pending = 1; |
96 | 0 | return 1; |
97 | 0 | } |
98 | | |
99 | | |
100 | | static void nmi_cpu_save_registers(struct op_msrs *msrs) |
101 | 0 | { |
102 | 0 | unsigned int const nr_ctrs = model->num_counters; |
103 | 0 | unsigned int const nr_ctrls = model->num_controls; |
104 | 0 | struct op_msr *counters = msrs->counters; |
105 | 0 | struct op_msr *controls = msrs->controls; |
106 | 0 | unsigned int i; |
107 | 0 |
|
108 | 0 | for (i = 0; i < nr_ctrs; ++i) { |
109 | 0 | rdmsrl(counters[i].addr, counters[i].value); |
110 | 0 | } |
111 | 0 |
|
112 | 0 | for (i = 0; i < nr_ctrls; ++i) { |
113 | 0 | rdmsrl(controls[i].addr, controls[i].value); |
114 | 0 | } |
115 | 0 | } |
116 | | |
117 | | |
118 | | static void nmi_save_registers(void * dummy) |
119 | 0 | { |
120 | 0 | int cpu = smp_processor_id(); |
121 | 0 | struct op_msrs * msrs = &cpu_msrs[cpu]; |
122 | 0 | model->fill_in_addresses(msrs); |
123 | 0 | nmi_cpu_save_registers(msrs); |
124 | 0 | } |
125 | | |
126 | | |
127 | | static void free_msrs(void) |
128 | 0 | { |
129 | 0 | int i; |
130 | 0 | for (i = 0; i < nr_cpu_ids; ++i) { |
131 | 0 | xfree(cpu_msrs[i].counters); |
132 | 0 | cpu_msrs[i].counters = NULL; |
133 | 0 | xfree(cpu_msrs[i].controls); |
134 | 0 | cpu_msrs[i].controls = NULL; |
135 | 0 | } |
136 | 0 | } |
137 | | |
138 | | |
139 | | static int allocate_msrs(void) |
140 | 0 | { |
141 | 0 | int success = 1; |
142 | 0 | size_t controls_size = sizeof(struct op_msr) * model->num_controls; |
143 | 0 | size_t counters_size = sizeof(struct op_msr) * model->num_counters; |
144 | 0 |
|
145 | 0 | int i; |
146 | 0 | for_each_online_cpu (i) { |
147 | 0 | cpu_msrs[i].counters = xmalloc_bytes(counters_size); |
148 | 0 | if (!cpu_msrs[i].counters) { |
149 | 0 | success = 0; |
150 | 0 | break; |
151 | 0 | } |
152 | 0 | cpu_msrs[i].controls = xmalloc_bytes(controls_size); |
153 | 0 | if (!cpu_msrs[i].controls) { |
154 | 0 | success = 0; |
155 | 0 | break; |
156 | 0 | } |
157 | 0 | } |
158 | 0 |
|
159 | 0 | if (!success) |
160 | 0 | free_msrs(); |
161 | 0 |
|
162 | 0 | return success; |
163 | 0 | } |
164 | | |
165 | | |
166 | | static void nmi_cpu_setup(void * dummy) |
167 | 0 | { |
168 | 0 | int cpu = smp_processor_id(); |
169 | 0 | struct op_msrs * msrs = &cpu_msrs[cpu]; |
170 | 0 | model->setup_ctrs(msrs); |
171 | 0 | } |
172 | | |
173 | | |
174 | | int nmi_setup_events(void) |
175 | 0 | { |
176 | 0 | on_each_cpu(nmi_cpu_setup, NULL, 1); |
177 | 0 | return 0; |
178 | 0 | } |
179 | | |
180 | | int nmi_reserve_counters(void) |
181 | 0 | { |
182 | 0 | if (!allocate_msrs()) |
183 | 0 | return -ENOMEM; |
184 | 0 |
|
185 | 0 | /* We walk a thin line between law and rape here. |
186 | 0 | * We need to be careful to install our NMI handler |
187 | 0 | * without actually triggering any NMIs as this will |
188 | 0 | * break the core code horrifically. |
189 | 0 | */ |
190 | 0 | if (reserve_lapic_nmi() < 0) { |
191 | 0 | free_msrs(); |
192 | 0 | return -EBUSY; |
193 | 0 | } |
194 | 0 | /* We need to serialize save and setup for HT because the subset |
195 | 0 | * of msrs are distinct for save and setup operations |
196 | 0 | */ |
197 | 0 | on_each_cpu(nmi_save_registers, NULL, 1); |
198 | 0 | return 0; |
199 | 0 | } |
200 | | |
201 | | int nmi_enable_virq(void) |
202 | 0 | { |
203 | 0 | set_nmi_callback(nmi_callback); |
204 | 0 | return 0; |
205 | 0 | } |
206 | | |
207 | | |
208 | | void nmi_disable_virq(void) |
209 | 0 | { |
210 | 0 | unset_nmi_callback(); |
211 | 0 | } |
212 | | |
213 | | |
214 | | static void nmi_restore_registers(struct op_msrs * msrs) |
215 | 0 | { |
216 | 0 | unsigned int const nr_ctrs = model->num_counters; |
217 | 0 | unsigned int const nr_ctrls = model->num_controls; |
218 | 0 | struct op_msr * counters = msrs->counters; |
219 | 0 | struct op_msr * controls = msrs->controls; |
220 | 0 | unsigned int i; |
221 | 0 |
|
222 | 0 | for (i = 0; i < nr_ctrls; ++i) { |
223 | 0 | wrmsrl(controls[i].addr, controls[i].value); |
224 | 0 | } |
225 | 0 |
|
226 | 0 | for (i = 0; i < nr_ctrs; ++i) { |
227 | 0 | wrmsrl(counters[i].addr, counters[i].value); |
228 | 0 | } |
229 | 0 | } |
230 | | |
231 | | |
232 | | static void nmi_cpu_shutdown(void * dummy) |
233 | 0 | { |
234 | 0 | int cpu = smp_processor_id(); |
235 | 0 | struct op_msrs * msrs = &cpu_msrs[cpu]; |
236 | 0 | nmi_restore_registers(msrs); |
237 | 0 | } |
238 | | |
239 | | |
240 | | void nmi_release_counters(void) |
241 | 0 | { |
242 | 0 | on_each_cpu(nmi_cpu_shutdown, NULL, 1); |
243 | 0 | release_lapic_nmi(); |
244 | 0 | free_msrs(); |
245 | 0 | } |
246 | | |
247 | | |
248 | | static void nmi_cpu_start(void * dummy) |
249 | 0 | { |
250 | 0 | int cpu = smp_processor_id(); |
251 | 0 | struct op_msrs const * msrs = &cpu_msrs[cpu]; |
252 | 0 | saved_lvtpc[cpu] = apic_read(APIC_LVTPC); |
253 | 0 | apic_write(APIC_LVTPC, APIC_DM_NMI); |
254 | 0 | model->start(msrs); |
255 | 0 | } |
256 | | |
257 | | |
258 | | int nmi_start(void) |
259 | 0 | { |
260 | 0 | on_each_cpu(nmi_cpu_start, NULL, 1); |
261 | 0 | return 0; |
262 | 0 | } |
263 | | |
264 | | |
265 | | static void nmi_cpu_stop(void * dummy) |
266 | 0 | { |
267 | 0 | unsigned int v; |
268 | 0 | int cpu = smp_processor_id(); |
269 | 0 | struct op_msrs const * msrs = &cpu_msrs[cpu]; |
270 | 0 | model->stop(msrs); |
271 | 0 |
|
272 | 0 | /* restoring APIC_LVTPC can trigger an apic error because the delivery |
273 | 0 | * mode and vector nr combination can be illegal. That's by design: on |
274 | 0 | * power on apic lvt contain a zero vector nr which are legal only for |
275 | 0 | * NMI delivery mode. So inhibit apic err before restoring lvtpc |
276 | 0 | */ |
277 | 0 | if ( (apic_read(APIC_LVTPC) & APIC_MODE_MASK) != APIC_DM_NMI |
278 | 0 | || (apic_read(APIC_LVTPC) & APIC_LVT_MASKED) ) |
279 | 0 | { |
280 | 0 | printk("nmi_stop: APIC not good %ul\n", apic_read(APIC_LVTPC)); |
281 | 0 | mdelay(5000); |
282 | 0 | } |
283 | 0 | v = apic_read(APIC_LVTERR); |
284 | 0 | apic_write(APIC_LVTERR, v | APIC_LVT_MASKED); |
285 | 0 | apic_write(APIC_LVTPC, saved_lvtpc[cpu]); |
286 | 0 | apic_write(APIC_LVTERR, v); |
287 | 0 | } |
288 | | |
289 | | |
290 | | void nmi_stop(void) |
291 | 0 | { |
292 | 0 | on_each_cpu(nmi_cpu_stop, NULL, 1); |
293 | 0 | } |
294 | | |
295 | | |
296 | | static int __init p4_init(char ** cpu_type) |
297 | 0 | { |
298 | 0 | __u8 cpu_model = current_cpu_data.x86_model; |
299 | 0 |
|
300 | 0 | if ((cpu_model > 6) || (cpu_model == 5)) { |
301 | 0 | printk("xenoprof: Initialization failed. " |
302 | 0 | "Intel processor model %d for pentium 4 family is not " |
303 | 0 | "supported\n", cpu_model); |
304 | 0 | return 0; |
305 | 0 | } |
306 | 0 |
|
307 | 0 | switch (current_cpu_data.x86_num_siblings) { |
308 | 0 | case 1: |
309 | 0 | *cpu_type = "i386/p4"; |
310 | 0 | model = &op_p4_spec; |
311 | 0 | return 1; |
312 | 0 |
|
313 | 0 | case 2: |
314 | 0 | *cpu_type = "i386/p4-ht"; |
315 | 0 | model = &op_p4_ht2_spec; |
316 | 0 | return 1; |
317 | 0 | } |
318 | 0 |
|
319 | 0 | printk("Xenoprof ERROR: P4 HyperThreading detected with > 2 threads\n"); |
320 | 0 |
|
321 | 0 | return 0; |
322 | 0 | } |
323 | | |
324 | | |
325 | | static int force_arch_perfmon; |
326 | | |
327 | | static int force_cpu_type(const char *str) |
328 | 0 | { |
329 | 0 | if (!strcmp(str, "arch_perfmon")) { |
330 | 0 | force_arch_perfmon = 1; |
331 | 0 | printk(KERN_INFO "oprofile: forcing architectural perfmon\n"); |
332 | 0 | } |
333 | 0 | else |
334 | 0 | return -EINVAL; |
335 | 0 |
|
336 | 0 | return 0; |
337 | 0 | } |
338 | | custom_param("cpu_type", force_cpu_type); |
339 | | |
340 | | static int __init ppro_init(char ** cpu_type) |
341 | 1 | { |
342 | 1 | __u8 cpu_model = current_cpu_data.x86_model; |
343 | 1 | |
344 | 1 | if (force_arch_perfmon && cpu_has_arch_perfmon) |
345 | 0 | return 0; |
346 | 1 | |
347 | 1 | switch (cpu_model) { |
348 | 0 | case 14: |
349 | 0 | *cpu_type = "i386/core"; |
350 | 0 | break; |
351 | 0 | case 15: |
352 | 0 | *cpu_type = "i386/core_2"; |
353 | 0 | ppro_has_global_ctrl = 1; |
354 | 0 | break; |
355 | 1 | default: |
356 | 1 | /* Unknown */ |
357 | 1 | return 0; |
358 | 1 | } |
359 | 1 | |
360 | 0 | model = &op_ppro_spec; |
361 | 0 | return 1; |
362 | 1 | } |
363 | | |
364 | | static int __init arch_perfmon_init(char **cpu_type) |
365 | 1 | { |
366 | 1 | if (!cpu_has_arch_perfmon) |
367 | 0 | return 0; |
368 | 1 | *cpu_type = "i386/arch_perfmon"; |
369 | 1 | model = &op_arch_perfmon_spec; |
370 | 1 | arch_perfmon_setup_counters(); |
371 | 1 | ppro_has_global_ctrl = 1; |
372 | 1 | return 1; |
373 | 1 | } |
374 | | |
375 | | static int __init nmi_init(void) |
376 | 1 | { |
377 | 1 | __u8 vendor = current_cpu_data.x86_vendor; |
378 | 1 | __u8 family = current_cpu_data.x86; |
379 | 1 | __u8 _model = current_cpu_data.x86_model; |
380 | 1 | |
381 | 1 | if (!cpu_has_apic) { |
382 | 0 | printk("xenoprof: Initialization failed. No APIC\n"); |
383 | 0 | return -ENODEV; |
384 | 0 | } |
385 | 1 | |
386 | 1 | switch (vendor) { |
387 | 0 | case X86_VENDOR_AMD: |
388 | 0 | /* Needs to be at least an Athlon (or hammer in 32bit mode) */ |
389 | 0 |
|
390 | 0 | switch (family) { |
391 | 0 | default: |
392 | 0 | printk("xenoprof: Initialization failed. " |
393 | 0 | "AMD processor family %d is not " |
394 | 0 | "supported\n", family); |
395 | 0 | return -ENODEV; |
396 | 0 | case 0xf: |
397 | 0 | model = &op_athlon_spec; |
398 | 0 | cpu_type = "x86-64/hammer"; |
399 | 0 | break; |
400 | 0 | case 0x10: |
401 | 0 | model = &op_athlon_spec; |
402 | 0 | cpu_type = "x86-64/family10"; |
403 | 0 | ibs_init(); |
404 | 0 | break; |
405 | 0 | case 0x11: |
406 | 0 | model = &op_athlon_spec; |
407 | 0 | cpu_type = "x86-64/family11h"; |
408 | 0 | break; |
409 | 0 | case 0x12: |
410 | 0 | model = &op_athlon_spec; |
411 | 0 | cpu_type = "x86-64/family12h"; |
412 | 0 | break; |
413 | 0 | case 0x14: |
414 | 0 | model = &op_athlon_spec; |
415 | 0 | cpu_type = "x86-64/family14h"; |
416 | 0 | break; |
417 | 0 | case 0x15: |
418 | 0 | model = &op_amd_fam15h_spec; |
419 | 0 | cpu_type = "x86-64/family15h"; |
420 | 0 | break; |
421 | 0 | case 0x16: |
422 | 0 | model = &op_athlon_spec; |
423 | 0 | cpu_type = "x86-64/family16h"; |
424 | 0 | break; |
425 | 0 | } |
426 | 0 | break; |
427 | 0 |
|
428 | 1 | case X86_VENDOR_INTEL: |
429 | 1 | switch (family) { |
430 | 1 | /* Pentium IV */ |
431 | 0 | case 0xf: |
432 | 0 | p4_init(&cpu_type); |
433 | 0 | break; |
434 | 0 |
|
435 | 0 | /* A P6-class processor */ |
436 | 1 | case 6: |
437 | 1 | ppro_init(&cpu_type); |
438 | 1 | break; |
439 | 0 |
|
440 | 0 | default: |
441 | 0 | break; |
442 | 1 | } |
443 | 1 | if (!cpu_type && !arch_perfmon_init(&cpu_type)) { |
444 | 0 | printk("xenoprof: Initialization failed. " |
445 | 0 | "Intel processor family %d model %d " |
446 | 0 | "is not supported\n", family, _model); |
447 | 0 | return -ENODEV; |
448 | 0 | } |
449 | 1 | break; |
450 | 1 | |
451 | 0 | default: |
452 | 0 | printk("xenoprof: Initialization failed. " |
453 | 0 | "Unsupported processor. Unknown vendor %d\n", |
454 | 0 | vendor); |
455 | 0 | return -ENODEV; |
456 | 1 | } |
457 | 1 | |
458 | 1 | return 0; |
459 | 1 | } |
460 | | |
461 | | __initcall(nmi_init); |
462 | | |
463 | | int xenoprof_arch_init(int *num_events, char *_cpu_type) |
464 | 0 | { |
465 | 0 | if (cpu_type == NULL) |
466 | 0 | return -ENODEV; |
467 | 0 | *num_events = model->num_counters; |
468 | 0 | strlcpy(_cpu_type, cpu_type, XENOPROF_CPU_TYPE_SIZE); |
469 | 0 | return 0; |
470 | 0 | } |