Commit | Line | Data |
---|---|---|
9f36eaed MJ |
1 | /* SPDX-License-Identifier: (GPL-2.0 or LGPL-2.1) |
2 | * | |
886d51a3 | 3 | * lttng-context-perf-counters.c |
833ad6a0 MD |
4 | * |
5 | * LTTng performance monitoring counters (perf-counters) integration module. | |
6 | * | |
886d51a3 | 7 | * Copyright (C) 2009-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
833ad6a0 MD |
8 | */ |
9 | ||
10 | #include <linux/module.h> | |
11 | #include <linux/slab.h> | |
12 | #include <linux/perf_event.h> | |
13 | #include <linux/list.h> | |
c24a0d71 | 14 | #include <linux/string.h> |
5ca7b8a3 | 15 | #include <linux/cpu.h> |
241ae9a8 MD |
16 | #include <lttng-events.h> |
17 | #include <wrapper/ringbuffer/frontend_types.h> | |
18 | #include <wrapper/vmalloc.h> | |
19 | #include <wrapper/perf.h> | |
20 | #include <lttng-tracer.h> | |
833ad6a0 | 21 | |
f1676205 MD |
22 | static |
23 | size_t perf_counter_get_size(size_t offset) | |
24 | { | |
25 | size_t size = 0; | |
26 | ||
a90917c3 | 27 | size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t)); |
f1676205 MD |
28 | size += sizeof(uint64_t); |
29 | return size; | |
30 | } | |
31 | ||
833ad6a0 MD |
32 | static |
33 | void perf_counter_record(struct lttng_ctx_field *field, | |
34 | struct lib_ring_buffer_ctx *ctx, | |
a90917c3 | 35 | struct lttng_channel *chan) |
833ad6a0 MD |
36 | { |
37 | struct perf_event *event; | |
38 | uint64_t value; | |
39 | ||
2001023e | 40 | event = field->u.perf_counter->e[ctx->cpu]; |
0478c519 | 41 | if (likely(event)) { |
7b745a96 MD |
42 | if (unlikely(event->state == PERF_EVENT_STATE_ERROR)) { |
43 | value = 0; | |
44 | } else { | |
45 | event->pmu->read(event); | |
46 | value = local64_read(&event->count); | |
47 | } | |
f91fd73b MD |
48 | } else { |
49 | /* | |
50 | * Perf chooses not to be clever and not to support enabling a | |
51 | * perf counter before the cpu is brought up. Therefore, we need | |
52 | * to support having events coming (e.g. scheduler events) | |
53 | * before the counter is setup. Write an arbitrary 0 in this | |
54 | * case. | |
55 | */ | |
56 | value = 0; | |
57 | } | |
a90917c3 | 58 | lib_ring_buffer_align_ctx(ctx, lttng_alignof(value)); |
833ad6a0 MD |
59 | chan->ops->event_write(ctx, &value, sizeof(value)); |
60 | } | |
61 | ||
90f5546c MD |
62 | #if defined(CONFIG_PERF_EVENTS) && (LINUX_VERSION_CODE >= KERNEL_VERSION(3,0,99)) |
63 | static | |
64 | void overflow_callback(struct perf_event *event, | |
65 | struct perf_sample_data *data, | |
66 | struct pt_regs *regs) | |
67 | { | |
68 | } | |
69 | #else | |
833ad6a0 MD |
70 | static |
71 | void overflow_callback(struct perf_event *event, int nmi, | |
72 | struct perf_sample_data *data, | |
73 | struct pt_regs *regs) | |
74 | { | |
75 | } | |
90f5546c | 76 | #endif |
833ad6a0 | 77 | |
2dccf128 MD |
78 | static |
79 | void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field) | |
80 | { | |
2001023e | 81 | struct perf_event **events = field->u.perf_counter->e; |
2dccf128 | 82 | |
1e367326 MD |
83 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
84 | { | |
85 | int ret; | |
86 | ||
87 | ret = cpuhp_state_remove_instance(lttng_hp_online, | |
88 | &field->u.perf_counter->cpuhp_online.node); | |
89 | WARN_ON(ret); | |
90 | ret = cpuhp_state_remove_instance(lttng_hp_prepare, | |
91 | &field->u.perf_counter->cpuhp_prepare.node); | |
92 | WARN_ON(ret); | |
93 | } | |
94 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
95 | { | |
96 | int cpu; | |
97 | ||
98 | get_online_cpus(); | |
99 | for_each_online_cpu(cpu) | |
100 | perf_event_release_kernel(events[cpu]); | |
101 | put_online_cpus(); | |
8289661d | 102 | #ifdef CONFIG_HOTPLUG_CPU |
1e367326 | 103 | unregister_cpu_notifier(&field->u.perf_counter->nb); |
8289661d | 104 | #endif |
1e367326 MD |
105 | } |
106 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
c24a0d71 | 107 | kfree(field->event_field.name); |
2001023e | 108 | kfree(field->u.perf_counter->attr); |
48f5e0b5 | 109 | lttng_kvfree(events); |
2001023e | 110 | kfree(field->u.perf_counter); |
2dccf128 MD |
111 | } |
112 | ||
1e367326 MD |
113 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
114 | ||
115 | int lttng_cpuhp_perf_counter_online(unsigned int cpu, | |
116 | struct lttng_cpuhp_node *node) | |
117 | { | |
118 | struct lttng_perf_counter_field *perf_field = | |
119 | container_of(node, struct lttng_perf_counter_field, | |
120 | cpuhp_online); | |
121 | struct perf_event **events = perf_field->e; | |
122 | struct perf_event_attr *attr = perf_field->attr; | |
123 | struct perf_event *pevent; | |
124 | ||
125 | pevent = wrapper_perf_event_create_kernel_counter(attr, | |
126 | cpu, NULL, overflow_callback); | |
127 | if (!pevent || IS_ERR(pevent)) | |
128 | return -EINVAL; | |
129 | if (pevent->state == PERF_EVENT_STATE_ERROR) { | |
130 | perf_event_release_kernel(pevent); | |
131 | return -EINVAL; | |
132 | } | |
133 | barrier(); /* Create perf counter before setting event */ | |
134 | events[cpu] = pevent; | |
135 | return 0; | |
136 | } | |
137 | ||
138 | int lttng_cpuhp_perf_counter_dead(unsigned int cpu, | |
139 | struct lttng_cpuhp_node *node) | |
140 | { | |
141 | struct lttng_perf_counter_field *perf_field = | |
142 | container_of(node, struct lttng_perf_counter_field, | |
143 | cpuhp_prepare); | |
144 | struct perf_event **events = perf_field->e; | |
145 | struct perf_event *pevent; | |
146 | ||
147 | pevent = events[cpu]; | |
148 | events[cpu] = NULL; | |
149 | barrier(); /* NULLify event before perf counter teardown */ | |
150 | perf_event_release_kernel(pevent); | |
151 | return 0; | |
152 | } | |
153 | ||
154 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
155 | ||
8289661d MD |
156 | #ifdef CONFIG_HOTPLUG_CPU |
157 | ||
158 | /** | |
159 | * lttng_perf_counter_hp_callback - CPU hotplug callback | |
160 | * @nb: notifier block | |
161 | * @action: hotplug action to take | |
162 | * @hcpu: CPU number | |
163 | * | |
164 | * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD) | |
165 | * | |
166 | * We can setup perf counters when the cpu is online (up prepare seems to be too | |
167 | * soon). | |
168 | */ | |
169 | static | |
e8f071d5 | 170 | int lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb, |
8289661d MD |
171 | unsigned long action, |
172 | void *hcpu) | |
173 | { | |
174 | unsigned int cpu = (unsigned long) hcpu; | |
2001023e MD |
175 | struct lttng_perf_counter_field *perf_field = |
176 | container_of(nb, struct lttng_perf_counter_field, nb); | |
177 | struct perf_event **events = perf_field->e; | |
178 | struct perf_event_attr *attr = perf_field->attr; | |
f91fd73b | 179 | struct perf_event *pevent; |
8289661d | 180 | |
2001023e | 181 | if (!perf_field->hp_enable) |
8289661d MD |
182 | return NOTIFY_OK; |
183 | ||
184 | switch (action) { | |
185 | case CPU_ONLINE: | |
186 | case CPU_ONLINE_FROZEN: | |
90f5546c | 187 | pevent = wrapper_perf_event_create_kernel_counter(attr, |
8289661d | 188 | cpu, NULL, overflow_callback); |
0478c519 | 189 | if (!pevent || IS_ERR(pevent)) |
8289661d | 190 | return NOTIFY_BAD; |
7b745a96 MD |
191 | if (pevent->state == PERF_EVENT_STATE_ERROR) { |
192 | perf_event_release_kernel(pevent); | |
193 | return NOTIFY_BAD; | |
194 | } | |
f91fd73b MD |
195 | barrier(); /* Create perf counter before setting event */ |
196 | events[cpu] = pevent; | |
8289661d MD |
197 | break; |
198 | case CPU_UP_CANCELED: | |
199 | case CPU_UP_CANCELED_FROZEN: | |
200 | case CPU_DEAD: | |
201 | case CPU_DEAD_FROZEN: | |
f91fd73b MD |
202 | pevent = events[cpu]; |
203 | events[cpu] = NULL; | |
204 | barrier(); /* NULLify event before perf counter teardown */ | |
205 | perf_event_release_kernel(pevent); | |
8289661d MD |
206 | break; |
207 | } | |
208 | return NOTIFY_OK; | |
209 | } | |
210 | ||
211 | #endif | |
212 | ||
1e367326 MD |
213 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
214 | ||
833ad6a0 MD |
215 | int lttng_add_perf_counter_to_ctx(uint32_t type, |
216 | uint64_t config, | |
c24a0d71 | 217 | const char *name, |
2dccf128 | 218 | struct lttng_ctx **ctx) |
833ad6a0 MD |
219 | { |
220 | struct lttng_ctx_field *field; | |
2001023e | 221 | struct lttng_perf_counter_field *perf_field; |
833ad6a0 MD |
222 | struct perf_event **events; |
223 | struct perf_event_attr *attr; | |
224 | int ret; | |
c24a0d71 | 225 | char *name_alloc; |
833ad6a0 | 226 | |
48f5e0b5 | 227 | events = lttng_kvzalloc(num_possible_cpus() * sizeof(*events), GFP_KERNEL); |
833ad6a0 MD |
228 | if (!events) |
229 | return -ENOMEM; | |
230 | ||
2001023e | 231 | attr = kzalloc(sizeof(struct perf_event_attr), GFP_KERNEL); |
833ad6a0 MD |
232 | if (!attr) { |
233 | ret = -ENOMEM; | |
234 | goto error_attr; | |
235 | } | |
236 | ||
237 | attr->type = type; | |
238 | attr->config = config; | |
239 | attr->size = sizeof(struct perf_event_attr); | |
240 | attr->pinned = 1; | |
241 | attr->disabled = 0; | |
242 | ||
2001023e MD |
243 | perf_field = kzalloc(sizeof(struct lttng_perf_counter_field), GFP_KERNEL); |
244 | if (!perf_field) { | |
245 | ret = -ENOMEM; | |
246 | goto error_alloc_perf_field; | |
247 | } | |
248 | perf_field->e = events; | |
249 | perf_field->attr = attr; | |
250 | ||
c24a0d71 | 251 | name_alloc = kstrdup(name, GFP_KERNEL); |
bef96e48 MD |
252 | if (!name_alloc) { |
253 | ret = -ENOMEM; | |
c24a0d71 | 254 | goto name_alloc_error; |
bef96e48 | 255 | } |
c24a0d71 | 256 | |
2dccf128 MD |
257 | field = lttng_append_context(ctx); |
258 | if (!field) { | |
259 | ret = -ENOMEM; | |
8289661d | 260 | goto append_context_error; |
833ad6a0 | 261 | } |
2001023e | 262 | if (lttng_find_context(*ctx, name_alloc)) { |
44252f0f MD |
263 | ret = -EEXIST; |
264 | goto find_error; | |
265 | } | |
8289661d | 266 | |
1e367326 MD |
267 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
268 | ||
269 | perf_field->cpuhp_prepare.component = LTTNG_CONTEXT_PERF_COUNTERS; | |
270 | ret = cpuhp_state_add_instance(lttng_hp_prepare, | |
271 | &perf_field->cpuhp_prepare.node); | |
272 | if (ret) | |
273 | goto cpuhp_prepare_error; | |
274 | ||
275 | perf_field->cpuhp_online.component = LTTNG_CONTEXT_PERF_COUNTERS; | |
276 | ret = cpuhp_state_add_instance(lttng_hp_online, | |
277 | &perf_field->cpuhp_online.node); | |
278 | if (ret) | |
279 | goto cpuhp_online_error; | |
280 | ||
281 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
282 | { | |
283 | int cpu; | |
284 | ||
8289661d | 285 | #ifdef CONFIG_HOTPLUG_CPU |
1e367326 MD |
286 | perf_field->nb.notifier_call = |
287 | lttng_perf_counter_cpu_hp_callback; | |
288 | perf_field->nb.priority = 0; | |
289 | register_cpu_notifier(&perf_field->nb); | |
8289661d | 290 | #endif |
1e367326 MD |
291 | get_online_cpus(); |
292 | for_each_online_cpu(cpu) { | |
293 | events[cpu] = wrapper_perf_event_create_kernel_counter(attr, | |
294 | cpu, NULL, overflow_callback); | |
295 | if (!events[cpu] || IS_ERR(events[cpu])) { | |
296 | ret = -EINVAL; | |
297 | goto counter_error; | |
298 | } | |
299 | if (events[cpu]->state == PERF_EVENT_STATE_ERROR) { | |
300 | ret = -EBUSY; | |
301 | goto counter_busy; | |
302 | } | |
7b745a96 | 303 | } |
1e367326 MD |
304 | put_online_cpus(); |
305 | perf_field->hp_enable = 1; | |
8289661d | 306 | } |
1e367326 | 307 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
8289661d | 308 | |
2dccf128 | 309 | field->destroy = lttng_destroy_perf_counter_field; |
833ad6a0 | 310 | |
c24a0d71 | 311 | field->event_field.name = name_alloc; |
8070f5c0 | 312 | field->event_field.type.atype = atype_integer; |
9d7d747f | 313 | field->event_field.type.u.basic.integer.size = sizeof(uint64_t) * CHAR_BIT; |
a90917c3 | 314 | field->event_field.type.u.basic.integer.alignment = lttng_alignof(uint64_t) * CHAR_BIT; |
06254b0f | 315 | field->event_field.type.u.basic.integer.signedness = lttng_is_signed_type(uint64_t); |
8070f5c0 MD |
316 | field->event_field.type.u.basic.integer.reverse_byte_order = 0; |
317 | field->event_field.type.u.basic.integer.base = 10; | |
318 | field->event_field.type.u.basic.integer.encoding = lttng_encode_none; | |
f1676205 MD |
319 | field->get_size = perf_counter_get_size; |
320 | field->record = perf_counter_record; | |
2001023e | 321 | field->u.perf_counter = perf_field; |
a9dd15da | 322 | lttng_context_update(*ctx); |
833ad6a0 | 323 | |
ea53823c | 324 | wrapper_vmalloc_sync_mappings(); |
833ad6a0 MD |
325 | return 0; |
326 | ||
1e367326 MD |
327 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
328 | cpuhp_online_error: | |
329 | { | |
330 | int remove_ret; | |
331 | ||
332 | remove_ret = cpuhp_state_remove_instance(lttng_hp_prepare, | |
333 | &perf_field->cpuhp_prepare.node); | |
334 | WARN_ON(remove_ret); | |
335 | } | |
336 | cpuhp_prepare_error: | |
337 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
7b745a96 | 338 | counter_busy: |
8289661d | 339 | counter_error: |
ce4a2f0c MD |
340 | { |
341 | int cpu; | |
342 | ||
343 | for_each_online_cpu(cpu) { | |
344 | if (events[cpu] && !IS_ERR(events[cpu])) | |
345 | perf_event_release_kernel(events[cpu]); | |
346 | } | |
347 | put_online_cpus(); | |
8289661d | 348 | #ifdef CONFIG_HOTPLUG_CPU |
ce4a2f0c | 349 | unregister_cpu_notifier(&perf_field->nb); |
8289661d | 350 | #endif |
ce4a2f0c | 351 | } |
1e367326 | 352 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
44252f0f | 353 | find_error: |
8289661d MD |
354 | lttng_remove_context_field(ctx, field); |
355 | append_context_error: | |
356 | kfree(name_alloc); | |
357 | name_alloc_error: | |
2001023e MD |
358 | kfree(perf_field); |
359 | error_alloc_perf_field: | |
833ad6a0 MD |
360 | kfree(attr); |
361 | error_attr: | |
48f5e0b5 | 362 | lttng_kvfree(events); |
833ad6a0 MD |
363 | return ret; |
364 | } |