lttng-context-perf-counters.c

   1 /*
   2  * (C) Copyright        2009-2011 -
   3  *              Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   4  *
   5  * LTTng performance monitoring counters (perf-counters) integration module.
   6  *
   7  * Dual LGPL v2.1/GPL v2 license.
   8  */
   9
  10 #include <linux/module.h>
  11 #include <linux/slab.h>
  12 #include <linux/perf_event.h>
  13 #include <linux/list.h>
  14 #include <linux/string.h>
  15 #include "ltt-events.h"
  16 #include "wrapper/ringbuffer/frontend_types.h"
  17 #include "wrapper/vmalloc.h"
  18 #include "ltt-tracer.h"
  19
  20 static
  21 size_t perf_counter_get_size(size_t offset)
  22 {
  23         size_t size = 0;
  24
  25         size += lib_ring_buffer_align(offset, ltt_alignof(uint64_t));
  26         size += sizeof(uint64_t);
  27         return size;
  28 }
  29
  30 static
  31 void perf_counter_record(struct lttng_ctx_field *field,
  32                          struct lib_ring_buffer_ctx *ctx,
  33                          struct ltt_channel *chan)
  34 {
  35         struct perf_event *event;
  36         uint64_t value;
  37
  38         event = field->u.perf_counter->e[ctx->cpu];
  39         if (likely(event)) {
  40                 if (unlikely(event->state == PERF_EVENT_STATE_ERROR)) {
  41                         value = 0;
  42                 } else {
  43                         event->pmu->read(event);
  44                         value = local64_read(&event->count);
  45                 }
  46         } else {
  47                 /*
  48                  * Perf chooses not to be clever and not to support enabling a
  49                  * perf counter before the cpu is brought up. Therefore, we need
  50                  * to support having events coming (e.g. scheduler events)
  51                  * before the counter is setup. Write an arbitrary 0 in this
  52                  * case.
  53                  */
  54                 value = 0;
  55         }
  56         lib_ring_buffer_align_ctx(ctx, ltt_alignof(value));
  57         chan->ops->event_write(ctx, &value, sizeof(value));
  58 }
  59
  60 static
  61 void overflow_callback(struct perf_event *event, int nmi,
  62                        struct perf_sample_data *data,
  63                        struct pt_regs *regs)
  64 {
  65 }
  66
  67 static
  68 void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field)
  69 {
  70         struct perf_event **events = field->u.perf_counter->e;
  71         int cpu;
  72
  73         get_online_cpus();
  74         for_each_online_cpu(cpu)
  75                 perf_event_release_kernel(events[cpu]);
  76         put_online_cpus();
  77 #ifdef CONFIG_HOTPLUG_CPU
  78         unregister_cpu_notifier(&field->u.perf_counter->nb);
  79 #endif
  80         kfree(field->event_field.name);
  81         kfree(field->u.perf_counter->attr);
  82         kfree(events);
  83         kfree(field->u.perf_counter);
  84 }
  85
  86 #ifdef CONFIG_HOTPLUG_CPU
  87
  88 /**
  89  *      lttng_perf_counter_hp_callback - CPU hotplug callback
  90  *      @nb: notifier block
  91  *      @action: hotplug action to take
  92  *      @hcpu: CPU number
  93  *
  94  *      Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
  95  *
  96  * We can setup perf counters when the cpu is online (up prepare seems to be too
  97  * soon).
  98  */
  99 static
 100 int __cpuinit lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb,
 101                                                  unsigned long action,
 102                                                  void *hcpu)
 103 {
 104         unsigned int cpu = (unsigned long) hcpu;
 105         struct lttng_perf_counter_field *perf_field =
 106                 container_of(nb, struct lttng_perf_counter_field, nb);
 107         struct perf_event **events = perf_field->e;
 108         struct perf_event_attr *attr = perf_field->attr;
 109         struct perf_event *pevent;
 110
 111         if (!perf_field->hp_enable)
 112                 return NOTIFY_OK;
 113
 114         switch (action) {
 115         case CPU_ONLINE:
 116         case CPU_ONLINE_FROZEN:
 117                 pevent = perf_event_create_kernel_counter(attr,
 118                                 cpu, NULL, overflow_callback);
 119                 if (!pevent || IS_ERR(pevent))
 120                         return NOTIFY_BAD;
 121                 if (pevent->state == PERF_EVENT_STATE_ERROR) {
 122                         perf_event_release_kernel(pevent);
 123                         return NOTIFY_BAD;
 124                 }
 125                 barrier();      /* Create perf counter before setting event */
 126                 events[cpu] = pevent;
 127                 break;
 128         case CPU_UP_CANCELED:
 129         case CPU_UP_CANCELED_FROZEN:
 130         case CPU_DEAD:
 131         case CPU_DEAD_FROZEN:
 132                 pevent = events[cpu];
 133                 events[cpu] = NULL;
 134                 barrier();      /* NULLify event before perf counter teardown */
 135                 perf_event_release_kernel(pevent);
 136                 break;
 137         }
 138         return NOTIFY_OK;
 139 }
 140
 141 #endif
 142
 143 int lttng_add_perf_counter_to_ctx(uint32_t type,
 144                                   uint64_t config,
 145                                   const char *name,
 146                                   struct lttng_ctx **ctx)
 147 {
 148         struct lttng_ctx_field *field;
 149         struct lttng_perf_counter_field *perf_field;
 150         struct perf_event **events;
 151         struct perf_event_attr *attr;
 152         int ret;
 153         int cpu;
 154         char *name_alloc;
 155
 156         events = kzalloc(num_possible_cpus() * sizeof(*events), GFP_KERNEL);
 157         if (!events)
 158                 return -ENOMEM;
 159
 160         attr = kzalloc(sizeof(struct perf_event_attr), GFP_KERNEL);
 161         if (!attr) {
 162                 ret = -ENOMEM;
 163                 goto error_attr;
 164         }
 165
 166         attr->type = type;
 167         attr->config = config;
 168         attr->size = sizeof(struct perf_event_attr);
 169         attr->pinned = 1;
 170         attr->disabled = 0;
 171
 172         perf_field = kzalloc(sizeof(struct lttng_perf_counter_field), GFP_KERNEL);
 173         if (!perf_field) {
 174                 ret = -ENOMEM;
 175                 goto error_alloc_perf_field;
 176         }
 177         perf_field->e = events;
 178         perf_field->attr = attr;
 179
 180         name_alloc = kstrdup(name, GFP_KERNEL);
 181         if (!name_alloc) {
 182                 ret = -ENOMEM;
 183                 goto name_alloc_error;
 184         }
 185
 186         field = lttng_append_context(ctx);
 187         if (!field) {
 188                 ret = -ENOMEM;
 189                 goto append_context_error;
 190         }
 191         if (lttng_find_context(*ctx, name_alloc)) {
 192                 ret = -EEXIST;
 193                 goto find_error;
 194         }
 195
 196 #ifdef CONFIG_HOTPLUG_CPU
 197         perf_field->nb.notifier_call =
 198                 lttng_perf_counter_cpu_hp_callback;
 199         perf_field->nb.priority = 0;
 200         register_cpu_notifier(&perf_field->nb);
 201 #endif
 202
 203         get_online_cpus();
 204         for_each_online_cpu(cpu) {
 205                 events[cpu] = perf_event_create_kernel_counter(attr,
 206                                         cpu, NULL, overflow_callback);
 207                 if (!events[cpu] || IS_ERR(events[cpu])) {
 208                         ret = -EINVAL;
 209                         goto counter_error;
 210                 }
 211                 if (events[cpu]->state == PERF_EVENT_STATE_ERROR) {
 212                         ret = -EBUSY;
 213                         goto counter_busy;
 214                 }
 215         }
 216         put_online_cpus();
 217
 218         field->destroy = lttng_destroy_perf_counter_field;
 219
 220         field->event_field.name = name_alloc;
 221         field->event_field.type.atype = atype_integer;
 222         field->event_field.type.u.basic.integer.size = sizeof(unsigned long) * CHAR_BIT;
 223         field->event_field.type.u.basic.integer.alignment = ltt_alignof(unsigned long) * CHAR_BIT;
 224         field->event_field.type.u.basic.integer.signedness = is_signed_type(unsigned long);
 225         field->event_field.type.u.basic.integer.reverse_byte_order = 0;
 226         field->event_field.type.u.basic.integer.base = 10;
 227         field->event_field.type.u.basic.integer.encoding = lttng_encode_none;
 228         field->get_size = perf_counter_get_size;
 229         field->record = perf_counter_record;
 230         field->u.perf_counter = perf_field;
 231         perf_field->hp_enable = 1;
 232
 233         wrapper_vmalloc_sync_all();
 234         return 0;
 235
 236 counter_busy:
 237 counter_error:
 238         for_each_online_cpu(cpu) {
 239                 if (events[cpu] && !IS_ERR(events[cpu]))
 240                         perf_event_release_kernel(events[cpu]);
 241         }
 242         put_online_cpus();
 243 #ifdef CONFIG_HOTPLUG_CPU
 244         unregister_cpu_notifier(&perf_field->nb);
 245 #endif
 246 find_error:
 247         lttng_remove_context_field(ctx, field);
 248 append_context_error:
 249         kfree(name_alloc);
 250 name_alloc_error:
 251         kfree(perf_field);
 252 error_alloc_perf_field:
 253         kfree(attr);
 254 error_attr:
 255         kfree(events);
 256         return ret;
 257 }
 258
 259 MODULE_LICENSE("GPL and additional rights");
 260 MODULE_AUTHOR("Mathieu Desnoyers");
 261 MODULE_DESCRIPTION("Linux Trace Toolkit Perf Support");