| 1 | /* |
| 2 | * lttng-context-perf-counters.c |
| 3 | * |
| 4 | * LTTng performance monitoring counters (perf-counters) integration module. |
| 5 | * |
| 6 | * Copyright (C) 2009-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 7 | * |
| 8 | * This library is free software; you can redistribute it and/or |
| 9 | * modify it under the terms of the GNU Lesser General Public |
| 10 | * License as published by the Free Software Foundation; only |
| 11 | * version 2.1 of the License. |
| 12 | * |
| 13 | * This library is distributed in the hope that it will be useful, |
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 16 | * Lesser General Public License for more details. |
| 17 | * |
| 18 | * You should have received a copy of the GNU Lesser General Public |
| 19 | * License along with this library; if not, write to the Free Software |
| 20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 21 | */ |
| 22 | |
| 23 | #include <linux/module.h> |
| 24 | #include <linux/slab.h> |
| 25 | #include <linux/perf_event.h> |
| 26 | #include <linux/list.h> |
| 27 | #include <linux/string.h> |
| 28 | #include <linux/cpu.h> |
| 29 | #include "lttng-events.h" |
| 30 | #include "wrapper/ringbuffer/frontend_types.h" |
| 31 | #include "wrapper/vmalloc.h" |
| 32 | #include "wrapper/perf.h" |
| 33 | #include "lttng-tracer.h" |
| 34 | |
| 35 | static |
| 36 | size_t perf_counter_get_size(size_t offset) |
| 37 | { |
| 38 | size_t size = 0; |
| 39 | |
| 40 | size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t)); |
| 41 | size += sizeof(uint64_t); |
| 42 | return size; |
| 43 | } |
| 44 | |
| 45 | static |
| 46 | void perf_counter_record(struct lttng_ctx_field *field, |
| 47 | struct lib_ring_buffer_ctx *ctx, |
| 48 | struct lttng_channel *chan) |
| 49 | { |
| 50 | struct perf_event *event; |
| 51 | uint64_t value; |
| 52 | |
| 53 | event = field->u.perf_counter->e[ctx->cpu]; |
| 54 | if (likely(event)) { |
| 55 | if (unlikely(event->state == PERF_EVENT_STATE_ERROR)) { |
| 56 | value = 0; |
| 57 | } else { |
| 58 | event->pmu->read(event); |
| 59 | value = local64_read(&event->count); |
| 60 | } |
| 61 | } else { |
| 62 | /* |
| 63 | * Perf chooses not to be clever and not to support enabling a |
| 64 | * perf counter before the cpu is brought up. Therefore, we need |
| 65 | * to support having events coming (e.g. scheduler events) |
| 66 | * before the counter is setup. Write an arbitrary 0 in this |
| 67 | * case. |
| 68 | */ |
| 69 | value = 0; |
| 70 | } |
| 71 | lib_ring_buffer_align_ctx(ctx, lttng_alignof(value)); |
| 72 | chan->ops->event_write(ctx, &value, sizeof(value)); |
| 73 | } |
| 74 | |
| 75 | #if defined(CONFIG_PERF_EVENTS) && (LINUX_VERSION_CODE >= KERNEL_VERSION(3,0,99)) |
| 76 | static |
| 77 | void overflow_callback(struct perf_event *event, |
| 78 | struct perf_sample_data *data, |
| 79 | struct pt_regs *regs) |
| 80 | { |
| 81 | } |
| 82 | #else |
| 83 | static |
| 84 | void overflow_callback(struct perf_event *event, int nmi, |
| 85 | struct perf_sample_data *data, |
| 86 | struct pt_regs *regs) |
| 87 | { |
| 88 | } |
| 89 | #endif |
| 90 | |
| 91 | static |
| 92 | void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field) |
| 93 | { |
| 94 | struct perf_event **events = field->u.perf_counter->e; |
| 95 | int cpu; |
| 96 | |
| 97 | get_online_cpus(); |
| 98 | for_each_online_cpu(cpu) |
| 99 | perf_event_release_kernel(events[cpu]); |
| 100 | put_online_cpus(); |
| 101 | #ifdef CONFIG_HOTPLUG_CPU |
| 102 | unregister_cpu_notifier(&field->u.perf_counter->nb); |
| 103 | #endif |
| 104 | kfree(field->event_field.name); |
| 105 | kfree(field->u.perf_counter->attr); |
| 106 | kfree(events); |
| 107 | kfree(field->u.perf_counter); |
| 108 | } |
| 109 | |
| 110 | #ifdef CONFIG_HOTPLUG_CPU |
| 111 | |
| 112 | /** |
| 113 | * lttng_perf_counter_hp_callback - CPU hotplug callback |
| 114 | * @nb: notifier block |
| 115 | * @action: hotplug action to take |
| 116 | * @hcpu: CPU number |
| 117 | * |
| 118 | * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD) |
| 119 | * |
| 120 | * We can setup perf counters when the cpu is online (up prepare seems to be too |
| 121 | * soon). |
| 122 | */ |
| 123 | static |
| 124 | int lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb, |
| 125 | unsigned long action, |
| 126 | void *hcpu) |
| 127 | { |
| 128 | unsigned int cpu = (unsigned long) hcpu; |
| 129 | struct lttng_perf_counter_field *perf_field = |
| 130 | container_of(nb, struct lttng_perf_counter_field, nb); |
| 131 | struct perf_event **events = perf_field->e; |
| 132 | struct perf_event_attr *attr = perf_field->attr; |
| 133 | struct perf_event *pevent; |
| 134 | |
| 135 | if (!perf_field->hp_enable) |
| 136 | return NOTIFY_OK; |
| 137 | |
| 138 | switch (action) { |
| 139 | case CPU_ONLINE: |
| 140 | case CPU_ONLINE_FROZEN: |
| 141 | pevent = wrapper_perf_event_create_kernel_counter(attr, |
| 142 | cpu, NULL, overflow_callback); |
| 143 | if (!pevent || IS_ERR(pevent)) |
| 144 | return NOTIFY_BAD; |
| 145 | if (pevent->state == PERF_EVENT_STATE_ERROR) { |
| 146 | perf_event_release_kernel(pevent); |
| 147 | return NOTIFY_BAD; |
| 148 | } |
| 149 | barrier(); /* Create perf counter before setting event */ |
| 150 | events[cpu] = pevent; |
| 151 | break; |
| 152 | case CPU_UP_CANCELED: |
| 153 | case CPU_UP_CANCELED_FROZEN: |
| 154 | case CPU_DEAD: |
| 155 | case CPU_DEAD_FROZEN: |
| 156 | pevent = events[cpu]; |
| 157 | events[cpu] = NULL; |
| 158 | barrier(); /* NULLify event before perf counter teardown */ |
| 159 | perf_event_release_kernel(pevent); |
| 160 | break; |
| 161 | } |
| 162 | return NOTIFY_OK; |
| 163 | } |
| 164 | |
| 165 | #endif |
| 166 | |
| 167 | int lttng_add_perf_counter_to_ctx(uint32_t type, |
| 168 | uint64_t config, |
| 169 | const char *name, |
| 170 | struct lttng_ctx **ctx) |
| 171 | { |
| 172 | struct lttng_ctx_field *field; |
| 173 | struct lttng_perf_counter_field *perf_field; |
| 174 | struct perf_event **events; |
| 175 | struct perf_event_attr *attr; |
| 176 | int ret; |
| 177 | int cpu; |
| 178 | char *name_alloc; |
| 179 | |
| 180 | events = kzalloc(num_possible_cpus() * sizeof(*events), GFP_KERNEL); |
| 181 | if (!events) |
| 182 | return -ENOMEM; |
| 183 | |
| 184 | attr = kzalloc(sizeof(struct perf_event_attr), GFP_KERNEL); |
| 185 | if (!attr) { |
| 186 | ret = -ENOMEM; |
| 187 | goto error_attr; |
| 188 | } |
| 189 | |
| 190 | attr->type = type; |
| 191 | attr->config = config; |
| 192 | attr->size = sizeof(struct perf_event_attr); |
| 193 | attr->pinned = 1; |
| 194 | attr->disabled = 0; |
| 195 | |
| 196 | perf_field = kzalloc(sizeof(struct lttng_perf_counter_field), GFP_KERNEL); |
| 197 | if (!perf_field) { |
| 198 | ret = -ENOMEM; |
| 199 | goto error_alloc_perf_field; |
| 200 | } |
| 201 | perf_field->e = events; |
| 202 | perf_field->attr = attr; |
| 203 | |
| 204 | name_alloc = kstrdup(name, GFP_KERNEL); |
| 205 | if (!name_alloc) { |
| 206 | ret = -ENOMEM; |
| 207 | goto name_alloc_error; |
| 208 | } |
| 209 | |
| 210 | field = lttng_append_context(ctx); |
| 211 | if (!field) { |
| 212 | ret = -ENOMEM; |
| 213 | goto append_context_error; |
| 214 | } |
| 215 | if (lttng_find_context(*ctx, name_alloc)) { |
| 216 | ret = -EEXIST; |
| 217 | goto find_error; |
| 218 | } |
| 219 | |
| 220 | #ifdef CONFIG_HOTPLUG_CPU |
| 221 | perf_field->nb.notifier_call = |
| 222 | lttng_perf_counter_cpu_hp_callback; |
| 223 | perf_field->nb.priority = 0; |
| 224 | register_cpu_notifier(&perf_field->nb); |
| 225 | #endif |
| 226 | |
| 227 | get_online_cpus(); |
| 228 | for_each_online_cpu(cpu) { |
| 229 | events[cpu] = wrapper_perf_event_create_kernel_counter(attr, |
| 230 | cpu, NULL, overflow_callback); |
| 231 | if (!events[cpu] || IS_ERR(events[cpu])) { |
| 232 | ret = -EINVAL; |
| 233 | goto counter_error; |
| 234 | } |
| 235 | if (events[cpu]->state == PERF_EVENT_STATE_ERROR) { |
| 236 | ret = -EBUSY; |
| 237 | goto counter_busy; |
| 238 | } |
| 239 | } |
| 240 | put_online_cpus(); |
| 241 | |
| 242 | field->destroy = lttng_destroy_perf_counter_field; |
| 243 | |
| 244 | field->event_field.name = name_alloc; |
| 245 | field->event_field.type.atype = atype_integer; |
| 246 | field->event_field.type.u.basic.integer.size = sizeof(uint64_t) * CHAR_BIT; |
| 247 | field->event_field.type.u.basic.integer.alignment = lttng_alignof(uint64_t) * CHAR_BIT; |
| 248 | field->event_field.type.u.basic.integer.signedness = lttng_is_signed_type(uint64_t); |
| 249 | field->event_field.type.u.basic.integer.reverse_byte_order = 0; |
| 250 | field->event_field.type.u.basic.integer.base = 10; |
| 251 | field->event_field.type.u.basic.integer.encoding = lttng_encode_none; |
| 252 | field->get_size = perf_counter_get_size; |
| 253 | field->record = perf_counter_record; |
| 254 | field->u.perf_counter = perf_field; |
| 255 | perf_field->hp_enable = 1; |
| 256 | lttng_context_update(*ctx); |
| 257 | |
| 258 | wrapper_vmalloc_sync_all(); |
| 259 | return 0; |
| 260 | |
| 261 | counter_busy: |
| 262 | counter_error: |
| 263 | for_each_online_cpu(cpu) { |
| 264 | if (events[cpu] && !IS_ERR(events[cpu])) |
| 265 | perf_event_release_kernel(events[cpu]); |
| 266 | } |
| 267 | put_online_cpus(); |
| 268 | #ifdef CONFIG_HOTPLUG_CPU |
| 269 | unregister_cpu_notifier(&perf_field->nb); |
| 270 | #endif |
| 271 | find_error: |
| 272 | lttng_remove_context_field(ctx, field); |
| 273 | append_context_error: |
| 274 | kfree(name_alloc); |
| 275 | name_alloc_error: |
| 276 | kfree(perf_field); |
| 277 | error_alloc_perf_field: |
| 278 | kfree(attr); |
| 279 | error_attr: |
| 280 | kfree(events); |
| 281 | return ret; |
| 282 | } |
| 283 | |
| 284 | MODULE_LICENSE("GPL and additional rights"); |
| 285 | MODULE_AUTHOR("Mathieu Desnoyers"); |
| 286 | MODULE_DESCRIPTION("Linux Trace Toolkit Perf Support"); |
| 287 | MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "." |
| 288 | __stringify(LTTNG_MODULES_MINOR_VERSION) "." |
| 289 | __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION) |
| 290 | LTTNG_MODULES_EXTRAVERSION); |