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  * @author Barry Kasindorf <barry.kasindorf@amd.com>
   9  *
  10  * Each CPU has a local buffer that stores PC value/event
  11  * pairs. We also log context switches when we notice them.
  12  * Eventually each CPU's buffer is processed into the global
  13  * event buffer by sync_buffer().
  14  *
  15  * We use a local buffer for two reasons: an NMI or similar
  16  * interrupt cannot synchronise, and high sampling rates
  17  * would lead to catastrophic global synchronisation if
  18  * a global buffer was used.
  19  */
  20
  21 #include <linux/sched.h>
  22 #include <linux/oprofile.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/errno.h>
  25
  26 #include "event_buffer.h"
  27 #include "cpu_buffer.h"
  28 #include "buffer_sync.h"
  29 #include "oprof.h"
  30
  31 DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
  32
  33 static void wq_sync_buffer(struct work_struct *work);
  34
  35 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
  36 static int work_enabled;
  37
  38 void free_cpu_buffers(void)
  39 {
  40         int i;
  41
  42         for_each_online_cpu(i) {
  43                 vfree(per_cpu(cpu_buffer, i).buffer);
  44                 per_cpu(cpu_buffer, i).buffer = NULL;
  45         }
  46 }
  47
  48 int alloc_cpu_buffers(void)
  49 {
  50         int i;
  51
  52         unsigned long buffer_size = fs_cpu_buffer_size;
  53
  54         for_each_online_cpu(i) {
  55                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
  56
  57                 b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
  58                         cpu_to_node(i));
  59                 if (!b->buffer)
  60                         goto fail;
  61
  62                 b->last_task = NULL;
  63                 b->last_is_kernel = -1;
  64                 b->tracing = 0;
  65                 b->buffer_size = buffer_size;
  66                 b->tail_pos = 0;
  67                 b->head_pos = 0;
  68                 b->sample_received = 0;
  69                 b->sample_lost_overflow = 0;
  70                 b->backtrace_aborted = 0;
  71                 b->sample_invalid_eip = 0;
  72                 b->cpu = i;
  73                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
  74         }
  75         return 0;
  76
  77 fail:
  78         free_cpu_buffers();
  79         return -ENOMEM;
  80 }
  81
  82 void start_cpu_work(void)
  83 {
  84         int i;
  85
  86         work_enabled = 1;
  87
  88         for_each_online_cpu(i) {
  89                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
  90
  91                 /*
  92                  * Spread the work by 1 jiffy per cpu so they dont all
  93                  * fire at once.
  94                  */
  95                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
  96         }
  97 }
  98
  99 void end_cpu_work(void)
 100 {
 101         int i;
 102
 103         work_enabled = 0;
 104
 105         for_each_online_cpu(i) {
 106                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
 107
 108                 cancel_delayed_work(&b->work);
 109         }
 110
 111         flush_scheduled_work();
 112 }
 113
 114 /* Resets the cpu buffer to a sane state. */
 115 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
 116 {
 117         /* reset these to invalid values; the next sample
 118          * collected will populate the buffer with proper
 119          * values to initialize the buffer
 120          */
 121         cpu_buf->last_is_kernel = -1;
 122         cpu_buf->last_task = NULL;
 123 }
 124
 125 /* compute number of available slots in cpu_buffer queue */
 126 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
 127 {
 128         unsigned long head = b->head_pos;
 129         unsigned long tail = b->tail_pos;
 130
 131         if (tail > head)
 132                 return (tail - head) - 1;
 133
 134         return tail + (b->buffer_size - head) - 1;
 135 }
 136
 137 static void increment_head(struct oprofile_cpu_buffer * b)
 138 {
 139         unsigned long new_head = b->head_pos + 1;
 140
 141         /* Ensure anything written to the slot before we
 142          * increment is visible */
 143         wmb();
 144
 145         if (new_head < b->buffer_size)
 146                 b->head_pos = new_head;
 147         else
 148                 b->head_pos = 0;
 149 }
 150
 151 static inline void
 152 add_sample(struct oprofile_cpu_buffer * cpu_buf,
 153            unsigned long pc, unsigned long event)
 154 {
 155         struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
 156         entry->eip = pc;
 157         entry->event = event;
 158         increment_head(cpu_buf);
 159 }
 160
 161 static inline void
 162 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
 163 {
 164         add_sample(buffer, ESCAPE_CODE, value);
 165 }
 166
 167 /* This must be safe from any context. It's safe writing here
 168  * because of the head/tail separation of the writer and reader
 169  * of the CPU buffer.
 170  *
 171  * is_kernel is needed because on some architectures you cannot
 172  * tell if you are in kernel or user space simply by looking at
 173  * pc. We tag this in the buffer by generating kernel enter/exit
 174  * events whenever is_kernel changes
 175  */
 176 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
 177                       int is_kernel, unsigned long event)
 178 {
 179         struct task_struct * task;
 180
 181         cpu_buf->sample_received++;
 182
 183         if (pc == ESCAPE_CODE) {
 184                 cpu_buf->sample_invalid_eip++;
 185                 return 0;
 186         }
 187
 188         if (nr_available_slots(cpu_buf) < 3) {
 189                 cpu_buf->sample_lost_overflow++;
 190                 return 0;
 191         }
 192
 193         is_kernel = !!is_kernel;
 194
 195         task = current;
 196
 197         /* notice a switch from user->kernel or vice versa */
 198         if (cpu_buf->last_is_kernel != is_kernel) {
 199                 cpu_buf->last_is_kernel = is_kernel;
 200                 add_code(cpu_buf, is_kernel);
 201         }
 202
 203         /* notice a task switch */
 204         if (cpu_buf->last_task != task) {
 205                 cpu_buf->last_task = task;
 206                 add_code(cpu_buf, (unsigned long)task);
 207         }
 208
 209         add_sample(cpu_buf, pc, event);
 210         return 1;
 211 }
 212
 213 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 214 {
 215         if (nr_available_slots(cpu_buf) < 4) {
 216                 cpu_buf->sample_lost_overflow++;
 217                 return 0;
 218         }
 219
 220         add_code(cpu_buf, CPU_TRACE_BEGIN);
 221         cpu_buf->tracing = 1;
 222         return 1;
 223 }
 224
 225 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
 226 {
 227         cpu_buf->tracing = 0;
 228 }
 229
 230 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 231                                 unsigned long event, int is_kernel)
 232 {
 233         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 234
 235         if (!backtrace_depth) {
 236                 log_sample(cpu_buf, pc, is_kernel, event);
 237                 return;
 238         }
 239
 240         if (!oprofile_begin_trace(cpu_buf))
 241                 return;
 242
 243         /* if log_sample() fail we can't backtrace since we lost the source
 244          * of this event */
 245         if (log_sample(cpu_buf, pc, is_kernel, event))
 246                 oprofile_ops.backtrace(regs, backtrace_depth);
 247         oprofile_end_trace(cpu_buf);
 248 }
 249
 250 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 251 {
 252         int is_kernel = !user_mode(regs);
 253         unsigned long pc = profile_pc(regs);
 254
 255         oprofile_add_ext_sample(pc, regs, event, is_kernel);
 256 }
 257
 258 #ifdef CONFIG_OPROFILE_IBS
 259
 260 #define MAX_IBS_SAMPLE_SIZE     14
 261 static int log_ibs_sample(struct oprofile_cpu_buffer *cpu_buf,
 262         unsigned long pc, int is_kernel, unsigned  int *ibs, int ibs_code)
 263 {
 264         struct task_struct *task;
 265
 266         cpu_buf->sample_received++;
 267
 268         if (nr_available_slots(cpu_buf) < MAX_IBS_SAMPLE_SIZE) {
 269                 cpu_buf->sample_lost_overflow++;
 270                 return 0;
 271         }
 272
 273         is_kernel = !!is_kernel;
 274
 275         /* notice a switch from user->kernel or vice versa */
 276         if (cpu_buf->last_is_kernel != is_kernel) {
 277                 cpu_buf->last_is_kernel = is_kernel;
 278                 add_code(cpu_buf, is_kernel);
 279         }
 280
 281         /* notice a task switch */
 282         if (!is_kernel) {
 283                 task = current;
 284
 285                 if (cpu_buf->last_task != task) {
 286                         cpu_buf->last_task = task;
 287                         add_code(cpu_buf, (unsigned long)task);
 288                 }
 289         }
 290
 291         add_code(cpu_buf, ibs_code);
 292         add_sample(cpu_buf, ibs[0], ibs[1]);
 293         add_sample(cpu_buf, ibs[2], ibs[3]);
 294         add_sample(cpu_buf, ibs[4], ibs[5]);
 295
 296         if (ibs_code == IBS_OP_BEGIN) {
 297         add_sample(cpu_buf, ibs[6], ibs[7]);
 298         add_sample(cpu_buf, ibs[8], ibs[9]);
 299         add_sample(cpu_buf, ibs[10], ibs[11]);
 300         }
 301
 302         return 1;
 303 }
 304
 305 void oprofile_add_ibs_sample(struct pt_regs *const regs,
 306                                 unsigned int * const ibs_sample, u8 code)
 307 {
 308         int is_kernel = !user_mode(regs);
 309         unsigned long pc = profile_pc(regs);
 310
 311         struct oprofile_cpu_buffer *cpu_buf =
 312                          &per_cpu(cpu_buffer, smp_processor_id());
 313
 314         if (!backtrace_depth) {
 315                 log_ibs_sample(cpu_buf, pc, is_kernel, ibs_sample, code);
 316                 return;
 317         }
 318
 319         /* if log_sample() fails we can't backtrace since we lost the source
 320         * of this event */
 321         if (log_ibs_sample(cpu_buf, pc, is_kernel, ibs_sample, code))
 322                 oprofile_ops.backtrace(regs, backtrace_depth);
 323 }
 324
 325 #endif
 326
 327 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 328 {
 329         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 330         log_sample(cpu_buf, pc, is_kernel, event);
 331 }
 332
 333 void oprofile_add_trace(unsigned long pc)
 334 {
 335         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 336
 337         if (!cpu_buf->tracing)
 338                 return;
 339
 340         if (nr_available_slots(cpu_buf) < 1) {
 341                 cpu_buf->tracing = 0;
 342                 cpu_buf->sample_lost_overflow++;
 343                 return;
 344         }
 345
 346         /* broken frame can give an eip with the same value as an escape code,
 347          * abort the trace if we get it */
 348         if (pc == ESCAPE_CODE) {
 349                 cpu_buf->tracing = 0;
 350                 cpu_buf->backtrace_aborted++;
 351                 return;
 352         }
 353
 354         add_sample(cpu_buf, pc, 0);
 355 }
 356
 357 /*
 358  * This serves to avoid cpu buffer overflow, and makes sure
 359  * the task mortuary progresses
 360  *
 361  * By using schedule_delayed_work_on and then schedule_delayed_work
 362  * we guarantee this will stay on the correct cpu
 363  */
 364 static void wq_sync_buffer(struct work_struct *work)
 365 {
 366         struct oprofile_cpu_buffer * b =
 367                 container_of(work, struct oprofile_cpu_buffer, work.work);
 368         if (b->cpu != smp_processor_id()) {
 369                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 370                        smp_processor_id(), b->cpu);
 371         }
 372         sync_buffer(b->cpu);
 373
 374         /* don't re-add the work if we're shutting down */
 375         if (work_enabled)
 376                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 377 }