drivers/oprofile/cpu_buffer.c

   1 /**
   2  * @file cpu_buffer.c
   3  *
   4  * @remark Copyright 2002 OProfile authors
   5  * @remark Read the file COPYING
   6  *
   7  * @author John Levon <levon@movementarian.org>
   8  *
   9  * Each CPU has a local buffer that stores PC value/event
  10  * pairs. We also log context switches when we notice them.
  11  * Eventually each CPU's buffer is processed into the global
  12  * event buffer by sync_buffer().
  13  *
  14  * We use a local buffer for two reasons: an NMI or similar
  15  * interrupt cannot synchronise, and high sampling rates
  16  * would lead to catastrophic global synchronisation if
  17  * a global buffer was used.
  18  */
  19
  20 #include <linux/sched.h>
  21 #include <linux/oprofile.h>
  22 #include <linux/vmalloc.h>
  23 #include <linux/errno.h>
  24
  25 #include "event_buffer.h"
  26 #include "cpu_buffer.h"
  27 #include "buffer_sync.h"
  28 #include "oprof.h"
  29
  30 struct oprofile_cpu_buffer cpu_buffer[NR_CPUS] __cacheline_aligned;
  31
  32 static void wq_sync_buffer(void *);
  33
  34 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
  35 static int work_enabled;
  36
  37 void free_cpu_buffers(void)
  38 {
  39         int i;
  40
  41         for_each_online_cpu(i) {
  42                 vfree(cpu_buffer[i].buffer);
  43         }
  44 }
  45
  46 int alloc_cpu_buffers(void)
  47 {
  48         int i;
  49
  50         unsigned long buffer_size = fs_cpu_buffer_size;
  51
  52         for_each_online_cpu(i) {
  53                 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
  54
  55                 b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
  56                         cpu_to_node(i));
  57                 if (!b->buffer)
  58                         goto fail;
  59
  60                 b->last_task = NULL;
  61                 b->last_is_kernel = -1;
  62                 b->tracing = 0;
  63                 b->buffer_size = buffer_size;
  64                 b->tail_pos = 0;
  65                 b->head_pos = 0;
  66                 b->sample_received = 0;
  67                 b->sample_lost_overflow = 0;
  68                 b->cpu = i;
  69                 INIT_WORK(&b->work, wq_sync_buffer, b);
  70         }
  71         return 0;
  72
  73 fail:
  74         free_cpu_buffers();
  75         return -ENOMEM;
  76 }
  77
  78 void start_cpu_work(void)
  79 {
  80         int i;
  81
  82         work_enabled = 1;
  83
  84         for_each_online_cpu(i) {
  85                 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
  86
  87                 /*
  88                  * Spread the work by 1 jiffy per cpu so they dont all
  89                  * fire at once.
  90                  */
  91                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
  92         }
  93 }
  94
  95 void end_cpu_work(void)
  96 {
  97         int i;
  98
  99         work_enabled = 0;
 100
 101         for_each_online_cpu(i) {
 102                 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
 103
 104                 cancel_delayed_work(&b->work);
 105         }
 106
 107         flush_scheduled_work();
 108 }
 109
 110 /* Resets the cpu buffer to a sane state. */
 111 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
 112 {
 113         /* reset these to invalid values; the next sample
 114          * collected will populate the buffer with proper
 115          * values to initialize the buffer
 116          */
 117         cpu_buf->last_is_kernel = -1;
 118         cpu_buf->last_task = NULL;
 119 }
 120
 121 /* compute number of available slots in cpu_buffer queue */
 122 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
 123 {
 124         unsigned long head = b->head_pos;
 125         unsigned long tail = b->tail_pos;
 126
 127         if (tail > head)
 128                 return (tail - head) - 1;
 129
 130         return tail + (b->buffer_size - head) - 1;
 131 }
 132
 133 static void increment_head(struct oprofile_cpu_buffer * b)
 134 {
 135         unsigned long new_head = b->head_pos + 1;
 136
 137         /* Ensure anything written to the slot before we
 138          * increment is visible */
 139         wmb();
 140
 141         if (new_head < b->buffer_size)
 142                 b->head_pos = new_head;
 143         else
 144                 b->head_pos = 0;
 145 }
 146
 147 static inline void
 148 add_sample(struct oprofile_cpu_buffer * cpu_buf,
 149            unsigned long pc, unsigned long event)
 150 {
 151         struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
 152         entry->eip = pc;
 153         entry->event = event;
 154         increment_head(cpu_buf);
 155 }
 156
 157 static inline void
 158 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
 159 {
 160         add_sample(buffer, ESCAPE_CODE, value);
 161 }
 162
 163 /* This must be safe from any context. It's safe writing here
 164  * because of the head/tail separation of the writer and reader
 165  * of the CPU buffer.
 166  *
 167  * is_kernel is needed because on some architectures you cannot
 168  * tell if you are in kernel or user space simply by looking at
 169  * pc. We tag this in the buffer by generating kernel enter/exit
 170  * events whenever is_kernel changes
 171  */
 172 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
 173                       int is_kernel, unsigned long event)
 174 {
 175         struct task_struct * task;
 176
 177         cpu_buf->sample_received++;
 178
 179         if (nr_available_slots(cpu_buf) < 3) {
 180                 cpu_buf->sample_lost_overflow++;
 181                 return 0;
 182         }
 183
 184         is_kernel = !!is_kernel;
 185
 186         task = current;
 187
 188         /* notice a switch from user->kernel or vice versa */
 189         if (cpu_buf->last_is_kernel != is_kernel) {
 190                 cpu_buf->last_is_kernel = is_kernel;
 191                 add_code(cpu_buf, is_kernel);
 192         }
 193
 194         /* notice a task switch */
 195         if (cpu_buf->last_task != task) {
 196                 cpu_buf->last_task = task;
 197                 add_code(cpu_buf, (unsigned long)task);
 198         }
 199
 200         add_sample(cpu_buf, pc, event);
 201         return 1;
 202 }
 203
 204 static int oprofile_begin_trace(struct oprofile_cpu_buffer * cpu_buf)
 205 {
 206         if (nr_available_slots(cpu_buf) < 4) {
 207                 cpu_buf->sample_lost_overflow++;
 208                 return 0;
 209         }
 210
 211         add_code(cpu_buf, CPU_TRACE_BEGIN);
 212         cpu_buf->tracing = 1;
 213         return 1;
 214 }
 215
 216 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
 217 {
 218         cpu_buf->tracing = 0;
 219 }
 220
 221 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 222 {
 223         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
 224         unsigned long pc = profile_pc(regs);
 225         int is_kernel = !user_mode(regs);
 226
 227         if (!backtrace_depth) {
 228                 log_sample(cpu_buf, pc, is_kernel, event);
 229                 return;
 230         }
 231
 232         if (!oprofile_begin_trace(cpu_buf))
 233                 return;
 234
 235         /* if log_sample() fail we can't backtrace since we lost the source
 236          * of this event */
 237         if (log_sample(cpu_buf, pc, is_kernel, event))
 238                 oprofile_ops.backtrace(regs, backtrace_depth);
 239         oprofile_end_trace(cpu_buf);
 240 }
 241
 242 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 243 {
 244         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
 245         log_sample(cpu_buf, pc, is_kernel, event);
 246 }
 247
 248 void oprofile_add_trace(unsigned long pc)
 249 {
 250         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
 251
 252         if (!cpu_buf->tracing)
 253                 return;
 254
 255         if (nr_available_slots(cpu_buf) < 1) {
 256                 cpu_buf->tracing = 0;
 257                 cpu_buf->sample_lost_overflow++;
 258                 return;
 259         }
 260
 261         /* broken frame can give an eip with the same value as an escape code,
 262          * abort the trace if we get it */
 263         if (pc == ESCAPE_CODE) {
 264                 cpu_buf->tracing = 0;
 265                 cpu_buf->backtrace_aborted++;
 266                 return;
 267         }
 268
 269         add_sample(cpu_buf, pc, 0);
 270 }
 271
 272 /*
 273  * This serves to avoid cpu buffer overflow, and makes sure
 274  * the task mortuary progresses
 275  *
 276  * By using schedule_delayed_work_on and then schedule_delayed_work
 277  * we guarantee this will stay on the correct cpu
 278  */
 279 static void wq_sync_buffer(void * data)
 280 {
 281         struct oprofile_cpu_buffer * b = data;
 282         if (b->cpu != smp_processor_id()) {
 283                 printk("WQ on CPU%d, prefer CPU%d\n",
 284                        smp_processor_id(), b->cpu);
 285         }
 286         sync_buffer(b->cpu);
 287
 288         /* don't re-add the work if we're shutting down */
 289         if (work_enabled)
 290                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 291 }