lttng-context-perf-counters.c

   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 "lttng-events.h"
  29 #include "wrapper/ringbuffer/frontend_types.h"
  30 #include "wrapper/vmalloc.h"
  31 #include "wrapper/perf.h"
  32 #include "lttng-tracer.h"
  33
  34 static
  35 size_t perf_counter_get_size(size_t offset)
  36 {
  37         size_t size = 0;
  38
  39         size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t));
  40         size += sizeof(uint64_t);
  41         return size;
  42 }
  43
  44 static
  45 void perf_counter_record(struct lttng_ctx_field *field,
  46                          struct lib_ring_buffer_ctx *ctx,
  47                          struct lttng_channel *chan)
  48 {
  49         struct perf_event *event;
  50         uint64_t value;
  51
  52         event = field->u.perf_counter->e[ctx->cpu];
  53         if (likely(event)) {
  54                 if (unlikely(event->state == PERF_EVENT_STATE_ERROR)) {
  55                         value = 0;
  56                 } else {
  57                         event->pmu->read(event);
  58                         value = local64_read(&event->count);
  59                 }
  60         } else {
  61                 /*
  62                  * Perf chooses not to be clever and not to support enabling a
  63                  * perf counter before the cpu is brought up. Therefore, we need
  64                  * to support having events coming (e.g. scheduler events)
  65                  * before the counter is setup. Write an arbitrary 0 in this
  66                  * case.
  67                  */
  68                 value = 0;
  69         }
  70         lib_ring_buffer_align_ctx(ctx, lttng_alignof(value));
  71         chan->ops->event_write(ctx, &value, sizeof(value));
  72 }
  73
  74 #if defined(CONFIG_PERF_EVENTS) && (LINUX_VERSION_CODE >= KERNEL_VERSION(3,0,99))
  75 static
  76 void overflow_callback(struct perf_event *event,
  77                        struct perf_sample_data *data,
  78                        struct pt_regs *regs)
  79 {
  80 }
  81 #else
  82 static
  83 void overflow_callback(struct perf_event *event, int nmi,
  84                        struct perf_sample_data *data,
  85                        struct pt_regs *regs)
  86 {
  87 }
  88 #endif
  89
  90 static
  91 void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field)
  92 {
  93         struct perf_event **events = field->u.perf_counter->e;
  94         int cpu;
  95
  96         get_online_cpus();
  97         for_each_online_cpu(cpu)
  98                 perf_event_release_kernel(events[cpu]);
  99         put_online_cpus();
 100 #ifdef CONFIG_HOTPLUG_CPU
 101         unregister_cpu_notifier(&field->u.perf_counter->nb);
 102 #endif
 103         kfree(field->event_field.name);
 104         kfree(field->u.perf_counter->attr);
 105         kfree(events);
 106         kfree(field->u.perf_counter);
 107 }
 108
 109 #ifdef CONFIG_HOTPLUG_CPU
 110
 111 /**
 112  *      lttng_perf_counter_hp_callback - CPU hotplug callback
 113  *      @nb: notifier block
 114  *      @action: hotplug action to take
 115  *      @hcpu: CPU number
 116  *
 117  *      Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
 118  *
 119  * We can setup perf counters when the cpu is online (up prepare seems to be too
 120  * soon).
 121  */
 122 static
 123 int __cpuinit lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb,
 124                                                  unsigned long action,
 125                                                  void *hcpu)
 126 {
 127         unsigned int cpu = (unsigned long) hcpu;
 128         struct lttng_perf_counter_field *perf_field =
 129                 container_of(nb, struct lttng_perf_counter_field, nb);
 130         struct perf_event **events = perf_field->e;
 131         struct perf_event_attr *attr = perf_field->attr;
 132         struct perf_event *pevent;
 133
 134         if (!perf_field->hp_enable)
 135                 return NOTIFY_OK;
 136
 137         switch (action) {
 138         case CPU_ONLINE:
 139         case CPU_ONLINE_FROZEN:
 140                 pevent = wrapper_perf_event_create_kernel_counter(attr,
 141                                 cpu, NULL, overflow_callback);
 142                 if (!pevent || IS_ERR(pevent))
 143                         return NOTIFY_BAD;
 144                 if (pevent->state == PERF_EVENT_STATE_ERROR) {
 145                         perf_event_release_kernel(pevent);
 146                         return NOTIFY_BAD;
 147                 }
 148                 barrier();      /* Create perf counter before setting event */
 149                 events[cpu] = pevent;
 150                 break;
 151         case CPU_UP_CANCELED:
 152         case CPU_UP_CANCELED_FROZEN:
 153         case CPU_DEAD:
 154         case CPU_DEAD_FROZEN:
 155                 pevent = events[cpu];
 156                 events[cpu] = NULL;
 157                 barrier();      /* NULLify event before perf counter teardown */
 158                 perf_event_release_kernel(pevent);
 159                 break;
 160         }
 161         return NOTIFY_OK;
 162 }
 163
 164 #endif
 165
 166 int lttng_add_perf_counter_to_ctx(uint32_t type,
 167                                   uint64_t config,
 168                                   const char *name,
 169                                   struct lttng_ctx **ctx)
 170 {
 171         struct lttng_ctx_field *field;
 172         struct lttng_perf_counter_field *perf_field;
 173         struct perf_event **events;
 174         struct perf_event_attr *attr;
 175         int ret;
 176         int cpu;
 177         char *name_alloc;
 178
 179         events = kzalloc(num_possible_cpus() * sizeof(*events), GFP_KERNEL);
 180         if (!events)
 181                 return -ENOMEM;
 182
 183         attr = kzalloc(sizeof(struct perf_event_attr), GFP_KERNEL);
 184         if (!attr) {
 185                 ret = -ENOMEM;
 186                 goto error_attr;
 187         }
 188
 189         attr->type = type;
 190         attr->config = config;
 191         attr->size = sizeof(struct perf_event_attr);
 192         attr->pinned = 1;
 193         attr->disabled = 0;
 194
 195         perf_field = kzalloc(sizeof(struct lttng_perf_counter_field), GFP_KERNEL);
 196         if (!perf_field) {
 197                 ret = -ENOMEM;
 198                 goto error_alloc_perf_field;
 199         }
 200         perf_field->e = events;
 201         perf_field->attr = attr;
 202
 203         name_alloc = kstrdup(name, GFP_KERNEL);
 204         if (!name_alloc) {
 205                 ret = -ENOMEM;
 206                 goto name_alloc_error;
 207         }
 208
 209         field = lttng_append_context(ctx);
 210         if (!field) {
 211                 ret = -ENOMEM;
 212                 goto append_context_error;
 213         }
 214         if (lttng_find_context(*ctx, name_alloc)) {
 215                 ret = -EEXIST;
 216                 goto find_error;
 217         }
 218
 219 #ifdef CONFIG_HOTPLUG_CPU
 220         perf_field->nb.notifier_call =
 221                 lttng_perf_counter_cpu_hp_callback;
 222         perf_field->nb.priority = 0;
 223         register_cpu_notifier(&perf_field->nb);
 224 #endif
 225
 226         get_online_cpus();
 227         for_each_online_cpu(cpu) {
 228                 events[cpu] = wrapper_perf_event_create_kernel_counter(attr,
 229                                         cpu, NULL, overflow_callback);
 230                 if (!events[cpu] || IS_ERR(events[cpu])) {
 231                         ret = -EINVAL;
 232                         goto counter_error;
 233                 }
 234                 if (events[cpu]->state == PERF_EVENT_STATE_ERROR) {
 235                         ret = -EBUSY;
 236                         goto counter_busy;
 237                 }
 238         }
 239         put_online_cpus();
 240
 241         field->destroy = lttng_destroy_perf_counter_field;
 242
 243         field->event_field.name = name_alloc;
 244         field->event_field.type.atype = atype_integer;
 245         field->event_field.type.u.basic.integer.size = sizeof(uint64_t) * CHAR_BIT;
 246         field->event_field.type.u.basic.integer.alignment = lttng_alignof(uint64_t) * CHAR_BIT;
 247         field->event_field.type.u.basic.integer.signedness = lttng_is_signed_type(uint64_t);
 248         field->event_field.type.u.basic.integer.reverse_byte_order = 0;
 249         field->event_field.type.u.basic.integer.base = 10;
 250         field->event_field.type.u.basic.integer.encoding = lttng_encode_none;
 251         field->get_size = perf_counter_get_size;
 252         field->record = perf_counter_record;
 253         field->u.perf_counter = perf_field;
 254         perf_field->hp_enable = 1;
 255
 256         wrapper_vmalloc_sync_all();
 257         return 0;
 258
 259 counter_busy:
 260 counter_error:
 261         for_each_online_cpu(cpu) {
 262                 if (events[cpu] && !IS_ERR(events[cpu]))
 263                         perf_event_release_kernel(events[cpu]);
 264         }
 265         put_online_cpus();
 266 #ifdef CONFIG_HOTPLUG_CPU
 267         unregister_cpu_notifier(&perf_field->nb);
 268 #endif
 269 find_error:
 270         lttng_remove_context_field(ctx, field);
 271 append_context_error:
 272         kfree(name_alloc);
 273 name_alloc_error:
 274         kfree(perf_field);
 275 error_alloc_perf_field:
 276         kfree(attr);
 277 error_attr:
 278         kfree(events);
 279         return ret;
 280 }
 281
 282 MODULE_LICENSE("GPL and additional rights");
 283 MODULE_AUTHOR("Mathieu Desnoyers");
 284 MODULE_DESCRIPTION("Linux Trace Toolkit Perf Support");