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> | |
4a6ea683 | 18 | #include <wrapper/cpu.h> |
241ae9a8 MD |
19 | #include <wrapper/vmalloc.h> |
20 | #include <wrapper/perf.h> | |
21 | #include <lttng-tracer.h> | |
833ad6a0 | 22 | |
f1676205 MD |
23 | static |
24 | size_t perf_counter_get_size(size_t offset) | |
25 | { | |
26 | size_t size = 0; | |
27 | ||
a90917c3 | 28 | size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t)); |
f1676205 MD |
29 | size += sizeof(uint64_t); |
30 | return size; | |
31 | } | |
32 | ||
833ad6a0 MD |
33 | static |
34 | void perf_counter_record(struct lttng_ctx_field *field, | |
35 | struct lib_ring_buffer_ctx *ctx, | |
a90917c3 | 36 | struct lttng_channel *chan) |
833ad6a0 MD |
37 | { |
38 | struct perf_event *event; | |
39 | uint64_t value; | |
40 | ||
2001023e | 41 | event = field->u.perf_counter->e[ctx->cpu]; |
0478c519 | 42 | if (likely(event)) { |
7b745a96 MD |
43 | if (unlikely(event->state == PERF_EVENT_STATE_ERROR)) { |
44 | value = 0; | |
45 | } else { | |
46 | event->pmu->read(event); | |
47 | value = local64_read(&event->count); | |
48 | } | |
f91fd73b MD |
49 | } else { |
50 | /* | |
51 | * Perf chooses not to be clever and not to support enabling a | |
52 | * perf counter before the cpu is brought up. Therefore, we need | |
53 | * to support having events coming (e.g. scheduler events) | |
54 | * before the counter is setup. Write an arbitrary 0 in this | |
55 | * case. | |
56 | */ | |
57 | value = 0; | |
58 | } | |
a90917c3 | 59 | lib_ring_buffer_align_ctx(ctx, lttng_alignof(value)); |
833ad6a0 MD |
60 | chan->ops->event_write(ctx, &value, sizeof(value)); |
61 | } | |
62 | ||
2d042821 | 63 | #if defined(CONFIG_PERF_EVENTS) && (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(3,1,0)) |
90f5546c MD |
64 | static |
65 | void overflow_callback(struct perf_event *event, | |
66 | struct perf_sample_data *data, | |
67 | struct pt_regs *regs) | |
68 | { | |
69 | } | |
70 | #else | |
833ad6a0 MD |
71 | static |
72 | void overflow_callback(struct perf_event *event, int nmi, | |
73 | struct perf_sample_data *data, | |
74 | struct pt_regs *regs) | |
75 | { | |
76 | } | |
90f5546c | 77 | #endif |
833ad6a0 | 78 | |
2dccf128 MD |
79 | static |
80 | void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field) | |
81 | { | |
2001023e | 82 | struct perf_event **events = field->u.perf_counter->e; |
2dccf128 | 83 | |
2d042821 | 84 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
85 | { |
86 | int ret; | |
87 | ||
88 | ret = cpuhp_state_remove_instance(lttng_hp_online, | |
89 | &field->u.perf_counter->cpuhp_online.node); | |
90 | WARN_ON(ret); | |
91 | ret = cpuhp_state_remove_instance(lttng_hp_prepare, | |
92 | &field->u.perf_counter->cpuhp_prepare.node); | |
93 | WARN_ON(ret); | |
94 | } | |
2d042821 | 95 | #else /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 MD |
96 | { |
97 | int cpu; | |
98 | ||
4a6ea683 | 99 | lttng_cpus_read_lock(); |
1e367326 MD |
100 | for_each_online_cpu(cpu) |
101 | perf_event_release_kernel(events[cpu]); | |
4a6ea683 | 102 | lttng_cpus_read_unlock(); |
8289661d | 103 | #ifdef CONFIG_HOTPLUG_CPU |
1e367326 | 104 | unregister_cpu_notifier(&field->u.perf_counter->nb); |
8289661d | 105 | #endif |
1e367326 | 106 | } |
2d042821 | 107 | #endif /* #else #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
c24a0d71 | 108 | kfree(field->event_field.name); |
2001023e | 109 | kfree(field->u.perf_counter->attr); |
48f5e0b5 | 110 | lttng_kvfree(events); |
2001023e | 111 | kfree(field->u.perf_counter); |
2dccf128 MD |
112 | } |
113 | ||
2d042821 | 114 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
115 | |
116 | int lttng_cpuhp_perf_counter_online(unsigned int cpu, | |
117 | struct lttng_cpuhp_node *node) | |
118 | { | |
119 | struct lttng_perf_counter_field *perf_field = | |
120 | container_of(node, struct lttng_perf_counter_field, | |
121 | cpuhp_online); | |
122 | struct perf_event **events = perf_field->e; | |
123 | struct perf_event_attr *attr = perf_field->attr; | |
124 | struct perf_event *pevent; | |
125 | ||
126 | pevent = wrapper_perf_event_create_kernel_counter(attr, | |
127 | cpu, NULL, overflow_callback); | |
128 | if (!pevent || IS_ERR(pevent)) | |
129 | return -EINVAL; | |
130 | if (pevent->state == PERF_EVENT_STATE_ERROR) { | |
131 | perf_event_release_kernel(pevent); | |
132 | return -EINVAL; | |
133 | } | |
134 | barrier(); /* Create perf counter before setting event */ | |
135 | events[cpu] = pevent; | |
136 | return 0; | |
137 | } | |
138 | ||
139 | int lttng_cpuhp_perf_counter_dead(unsigned int cpu, | |
140 | struct lttng_cpuhp_node *node) | |
141 | { | |
142 | struct lttng_perf_counter_field *perf_field = | |
143 | container_of(node, struct lttng_perf_counter_field, | |
144 | cpuhp_prepare); | |
145 | struct perf_event **events = perf_field->e; | |
146 | struct perf_event *pevent; | |
147 | ||
148 | pevent = events[cpu]; | |
149 | events[cpu] = NULL; | |
150 | barrier(); /* NULLify event before perf counter teardown */ | |
151 | perf_event_release_kernel(pevent); | |
152 | return 0; | |
153 | } | |
154 | ||
2d042821 | 155 | #else /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 | 156 | |
8289661d MD |
157 | #ifdef CONFIG_HOTPLUG_CPU |
158 | ||
159 | /** | |
160 | * lttng_perf_counter_hp_callback - CPU hotplug callback | |
161 | * @nb: notifier block | |
162 | * @action: hotplug action to take | |
163 | * @hcpu: CPU number | |
164 | * | |
165 | * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD) | |
166 | * | |
167 | * We can setup perf counters when the cpu is online (up prepare seems to be too | |
168 | * soon). | |
169 | */ | |
170 | static | |
e8f071d5 | 171 | int lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb, |
8289661d MD |
172 | unsigned long action, |
173 | void *hcpu) | |
174 | { | |
175 | unsigned int cpu = (unsigned long) hcpu; | |
2001023e MD |
176 | struct lttng_perf_counter_field *perf_field = |
177 | container_of(nb, struct lttng_perf_counter_field, nb); | |
178 | struct perf_event **events = perf_field->e; | |
179 | struct perf_event_attr *attr = perf_field->attr; | |
f91fd73b | 180 | struct perf_event *pevent; |
8289661d | 181 | |
2001023e | 182 | if (!perf_field->hp_enable) |
8289661d MD |
183 | return NOTIFY_OK; |
184 | ||
185 | switch (action) { | |
186 | case CPU_ONLINE: | |
187 | case CPU_ONLINE_FROZEN: | |
90f5546c | 188 | pevent = wrapper_perf_event_create_kernel_counter(attr, |
8289661d | 189 | cpu, NULL, overflow_callback); |
0478c519 | 190 | if (!pevent || IS_ERR(pevent)) |
8289661d | 191 | return NOTIFY_BAD; |
7b745a96 MD |
192 | if (pevent->state == PERF_EVENT_STATE_ERROR) { |
193 | perf_event_release_kernel(pevent); | |
194 | return NOTIFY_BAD; | |
195 | } | |
f91fd73b MD |
196 | barrier(); /* Create perf counter before setting event */ |
197 | events[cpu] = pevent; | |
8289661d MD |
198 | break; |
199 | case CPU_UP_CANCELED: | |
200 | case CPU_UP_CANCELED_FROZEN: | |
201 | case CPU_DEAD: | |
202 | case CPU_DEAD_FROZEN: | |
f91fd73b MD |
203 | pevent = events[cpu]; |
204 | events[cpu] = NULL; | |
205 | barrier(); /* NULLify event before perf counter teardown */ | |
206 | perf_event_release_kernel(pevent); | |
8289661d MD |
207 | break; |
208 | } | |
209 | return NOTIFY_OK; | |
210 | } | |
211 | ||
212 | #endif | |
213 | ||
2d042821 | 214 | #endif /* #else #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 | 215 | |
833ad6a0 MD |
216 | int lttng_add_perf_counter_to_ctx(uint32_t type, |
217 | uint64_t config, | |
c24a0d71 | 218 | const char *name, |
2dccf128 | 219 | struct lttng_ctx **ctx) |
833ad6a0 MD |
220 | { |
221 | struct lttng_ctx_field *field; | |
2001023e | 222 | struct lttng_perf_counter_field *perf_field; |
833ad6a0 MD |
223 | struct perf_event **events; |
224 | struct perf_event_attr *attr; | |
225 | int ret; | |
c24a0d71 | 226 | char *name_alloc; |
833ad6a0 | 227 | |
48f5e0b5 | 228 | events = lttng_kvzalloc(num_possible_cpus() * sizeof(*events), GFP_KERNEL); |
833ad6a0 MD |
229 | if (!events) |
230 | return -ENOMEM; | |
231 | ||
2001023e | 232 | attr = kzalloc(sizeof(struct perf_event_attr), GFP_KERNEL); |
833ad6a0 MD |
233 | if (!attr) { |
234 | ret = -ENOMEM; | |
235 | goto error_attr; | |
236 | } | |
237 | ||
238 | attr->type = type; | |
239 | attr->config = config; | |
240 | attr->size = sizeof(struct perf_event_attr); | |
241 | attr->pinned = 1; | |
242 | attr->disabled = 0; | |
243 | ||
2001023e MD |
244 | perf_field = kzalloc(sizeof(struct lttng_perf_counter_field), GFP_KERNEL); |
245 | if (!perf_field) { | |
246 | ret = -ENOMEM; | |
247 | goto error_alloc_perf_field; | |
248 | } | |
249 | perf_field->e = events; | |
250 | perf_field->attr = attr; | |
251 | ||
c24a0d71 | 252 | name_alloc = kstrdup(name, GFP_KERNEL); |
bef96e48 MD |
253 | if (!name_alloc) { |
254 | ret = -ENOMEM; | |
c24a0d71 | 255 | goto name_alloc_error; |
bef96e48 | 256 | } |
c24a0d71 | 257 | |
2dccf128 MD |
258 | field = lttng_append_context(ctx); |
259 | if (!field) { | |
260 | ret = -ENOMEM; | |
8289661d | 261 | goto append_context_error; |
833ad6a0 | 262 | } |
2001023e | 263 | if (lttng_find_context(*ctx, name_alloc)) { |
44252f0f MD |
264 | ret = -EEXIST; |
265 | goto find_error; | |
266 | } | |
8289661d | 267 | |
2d042821 | 268 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
269 | |
270 | perf_field->cpuhp_prepare.component = LTTNG_CONTEXT_PERF_COUNTERS; | |
271 | ret = cpuhp_state_add_instance(lttng_hp_prepare, | |
272 | &perf_field->cpuhp_prepare.node); | |
273 | if (ret) | |
274 | goto cpuhp_prepare_error; | |
275 | ||
276 | perf_field->cpuhp_online.component = LTTNG_CONTEXT_PERF_COUNTERS; | |
277 | ret = cpuhp_state_add_instance(lttng_hp_online, | |
278 | &perf_field->cpuhp_online.node); | |
279 | if (ret) | |
280 | goto cpuhp_online_error; | |
281 | ||
2d042821 | 282 | #else /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 MD |
283 | { |
284 | int cpu; | |
285 | ||
8289661d | 286 | #ifdef CONFIG_HOTPLUG_CPU |
1e367326 MD |
287 | perf_field->nb.notifier_call = |
288 | lttng_perf_counter_cpu_hp_callback; | |
289 | perf_field->nb.priority = 0; | |
290 | register_cpu_notifier(&perf_field->nb); | |
8289661d | 291 | #endif |
4a6ea683 | 292 | lttng_cpus_read_lock(); |
1e367326 MD |
293 | for_each_online_cpu(cpu) { |
294 | events[cpu] = wrapper_perf_event_create_kernel_counter(attr, | |
295 | cpu, NULL, overflow_callback); | |
296 | if (!events[cpu] || IS_ERR(events[cpu])) { | |
297 | ret = -EINVAL; | |
298 | goto counter_error; | |
299 | } | |
300 | if (events[cpu]->state == PERF_EVENT_STATE_ERROR) { | |
301 | ret = -EBUSY; | |
302 | goto counter_busy; | |
303 | } | |
7b745a96 | 304 | } |
4a6ea683 | 305 | lttng_cpus_read_unlock(); |
1e367326 | 306 | perf_field->hp_enable = 1; |
8289661d | 307 | } |
2d042821 | 308 | #endif /* #else #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
8289661d | 309 | |
2dccf128 | 310 | field->destroy = lttng_destroy_perf_counter_field; |
833ad6a0 | 311 | |
c24a0d71 | 312 | field->event_field.name = name_alloc; |
8070f5c0 | 313 | field->event_field.type.atype = atype_integer; |
9d7d747f | 314 | field->event_field.type.u.basic.integer.size = sizeof(uint64_t) * CHAR_BIT; |
a90917c3 | 315 | field->event_field.type.u.basic.integer.alignment = lttng_alignof(uint64_t) * CHAR_BIT; |
06254b0f | 316 | field->event_field.type.u.basic.integer.signedness = lttng_is_signed_type(uint64_t); |
8070f5c0 MD |
317 | field->event_field.type.u.basic.integer.reverse_byte_order = 0; |
318 | field->event_field.type.u.basic.integer.base = 10; | |
319 | field->event_field.type.u.basic.integer.encoding = lttng_encode_none; | |
f1676205 MD |
320 | field->get_size = perf_counter_get_size; |
321 | field->record = perf_counter_record; | |
2001023e | 322 | field->u.perf_counter = perf_field; |
a9dd15da | 323 | lttng_context_update(*ctx); |
833ad6a0 | 324 | |
ea53823c | 325 | wrapper_vmalloc_sync_mappings(); |
833ad6a0 MD |
326 | return 0; |
327 | ||
2d042821 | 328 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
329 | cpuhp_online_error: |
330 | { | |
331 | int remove_ret; | |
332 | ||
333 | remove_ret = cpuhp_state_remove_instance(lttng_hp_prepare, | |
334 | &perf_field->cpuhp_prepare.node); | |
335 | WARN_ON(remove_ret); | |
336 | } | |
337 | cpuhp_prepare_error: | |
2d042821 | 338 | #else /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
7b745a96 | 339 | counter_busy: |
8289661d | 340 | counter_error: |
ce4a2f0c MD |
341 | { |
342 | int cpu; | |
343 | ||
344 | for_each_online_cpu(cpu) { | |
345 | if (events[cpu] && !IS_ERR(events[cpu])) | |
346 | perf_event_release_kernel(events[cpu]); | |
347 | } | |
4a6ea683 | 348 | lttng_cpus_read_unlock(); |
8289661d | 349 | #ifdef CONFIG_HOTPLUG_CPU |
ce4a2f0c | 350 | unregister_cpu_notifier(&perf_field->nb); |
8289661d | 351 | #endif |
ce4a2f0c | 352 | } |
2d042821 | 353 | #endif /* #else #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
44252f0f | 354 | find_error: |
8289661d MD |
355 | lttng_remove_context_field(ctx, field); |
356 | append_context_error: | |
357 | kfree(name_alloc); | |
358 | name_alloc_error: | |
2001023e MD |
359 | kfree(perf_field); |
360 | error_alloc_perf_field: | |
833ad6a0 MD |
361 | kfree(attr); |
362 | error_attr: | |
48f5e0b5 | 363 | lttng_kvfree(events); |
833ad6a0 MD |
364 | return ret; |
365 | } |