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 OP_BUFFER_FLAGS 0
  32
  33 /*
  34  * Read and write access is using spin locking. Thus, writing to the
  35  * buffer by NMI handler (x86) could occur also during critical
  36  * sections when reading the buffer. To avoid this, there are 2
  37  * buffers for independent read and write access. Read access is in
  38  * process context only, write access only in the NMI handler. If the
  39  * read buffer runs empty, both buffers are swapped atomically. There
  40  * is potentially a small window during swapping where the buffers are
  41  * disabled and samples could be lost.
  42  *
  43  * Using 2 buffers is a little bit overhead, but the solution is clear
  44  * and does not require changes in the ring buffer implementation. It
  45  * can be changed to a single buffer solution when the ring buffer
  46  * access is implemented as non-locking atomic code.
  47  */
  48 static struct ring_buffer *op_ring_buffer_read;
  49 static struct ring_buffer *op_ring_buffer_write;
  50 DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
  51
  52 static void wq_sync_buffer(struct work_struct *work);
  53
  54 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
  55 static int work_enabled;
  56
  57 unsigned long oprofile_get_cpu_buffer_size(void)
  58 {
  59         return oprofile_cpu_buffer_size;
  60 }
  61
  62 void oprofile_cpu_buffer_inc_smpl_lost(void)
  63 {
  64         struct oprofile_cpu_buffer *cpu_buf
  65                 = &__get_cpu_var(cpu_buffer);
  66
  67         cpu_buf->sample_lost_overflow++;
  68 }
  69
  70 void free_cpu_buffers(void)
  71 {
  72         if (op_ring_buffer_read)
  73                 ring_buffer_free(op_ring_buffer_read);
  74         op_ring_buffer_read = NULL;
  75         if (op_ring_buffer_write)
  76                 ring_buffer_free(op_ring_buffer_write);
  77         op_ring_buffer_write = NULL;
  78 }
  79
  80 int alloc_cpu_buffers(void)
  81 {
  82         int i;
  83
  84         unsigned long buffer_size = oprofile_cpu_buffer_size;
  85
  86         op_ring_buffer_read = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
  87         if (!op_ring_buffer_read)
  88                 goto fail;
  89         op_ring_buffer_write = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
  90         if (!op_ring_buffer_write)
  91                 goto fail;
  92
  93         for_each_possible_cpu(i) {
  94                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
  95
  96                 b->last_task = NULL;
  97                 b->last_is_kernel = -1;
  98                 b->tracing = 0;
  99                 b->buffer_size = buffer_size;
 100                 b->tail_pos = 0;
 101                 b->head_pos = 0;
 102                 b->sample_received = 0;
 103                 b->sample_lost_overflow = 0;
 104                 b->backtrace_aborted = 0;
 105                 b->sample_invalid_eip = 0;
 106                 b->cpu = i;
 107                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
 108         }
 109         return 0;
 110
 111 fail:
 112         free_cpu_buffers();
 113         return -ENOMEM;
 114 }
 115
 116 void start_cpu_work(void)
 117 {
 118         int i;
 119
 120         work_enabled = 1;
 121
 122         for_each_online_cpu(i) {
 123                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
 124
 125                 /*
 126                  * Spread the work by 1 jiffy per cpu so they dont all
 127                  * fire at once.
 128                  */
 129                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
 130         }
 131 }
 132
 133 void end_cpu_work(void)
 134 {
 135         int i;
 136
 137         work_enabled = 0;
 138
 139         for_each_online_cpu(i) {
 140                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
 141
 142                 cancel_delayed_work(&b->work);
 143         }
 144
 145         flush_scheduled_work();
 146 }
 147
 148 int op_cpu_buffer_write_entry(struct op_entry *entry)
 149 {
 150         entry->event = ring_buffer_lock_reserve(op_ring_buffer_write,
 151                                                 sizeof(struct op_sample),
 152                                                 &entry->irq_flags);
 153         if (entry->event)
 154                 entry->sample = ring_buffer_event_data(entry->event);
 155         else
 156                 entry->sample = NULL;
 157
 158         if (!entry->sample)
 159                 return -ENOMEM;
 160
 161         return 0;
 162 }
 163
 164 int op_cpu_buffer_write_commit(struct op_entry *entry)
 165 {
 166         return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event,
 167                                          entry->irq_flags);
 168 }
 169
 170 struct op_sample *op_cpu_buffer_read_entry(int cpu)
 171 {
 172         struct ring_buffer_event *e;
 173         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
 174         if (e)
 175                 return ring_buffer_event_data(e);
 176         if (ring_buffer_swap_cpu(op_ring_buffer_read,
 177                                  op_ring_buffer_write,
 178                                  cpu))
 179                 return NULL;
 180         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
 181         if (e)
 182                 return ring_buffer_event_data(e);
 183         return NULL;
 184 }
 185
 186 unsigned long op_cpu_buffer_entries(int cpu)
 187 {
 188         return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
 189                 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
 190 }
 191
 192 static inline int
 193 add_sample(struct oprofile_cpu_buffer *cpu_buf,
 194            unsigned long pc, unsigned long event)
 195 {
 196         struct op_entry entry;
 197         int ret;
 198
 199         ret = op_cpu_buffer_write_entry(&entry);
 200         if (ret)
 201                 return ret;
 202
 203         entry.sample->eip = pc;
 204         entry.sample->event = event;
 205
 206         return op_cpu_buffer_write_commit(&entry);
 207 }
 208
 209 static inline int
 210 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
 211 {
 212         return add_sample(buffer, ESCAPE_CODE, value);
 213 }
 214
 215 /* This must be safe from any context. It's safe writing here
 216  * because of the head/tail separation of the writer and reader
 217  * of the CPU buffer.
 218  *
 219  * is_kernel is needed because on some architectures you cannot
 220  * tell if you are in kernel or user space simply by looking at
 221  * pc. We tag this in the buffer by generating kernel enter/exit
 222  * events whenever is_kernel changes
 223  */
 224 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 225                       int is_kernel, unsigned long event)
 226 {
 227         struct task_struct *task;
 228
 229         cpu_buf->sample_received++;
 230
 231         if (pc == ESCAPE_CODE) {
 232                 cpu_buf->sample_invalid_eip++;
 233                 return 0;
 234         }
 235
 236         is_kernel = !!is_kernel;
 237
 238         task = current;
 239
 240         /* notice a switch from user->kernel or vice versa */
 241         if (cpu_buf->last_is_kernel != is_kernel) {
 242                 cpu_buf->last_is_kernel = is_kernel;
 243                 if (add_code(cpu_buf, is_kernel))
 244                         goto fail;
 245         }
 246
 247         /* notice a task switch */
 248         if (cpu_buf->last_task != task) {
 249                 cpu_buf->last_task = task;
 250                 if (add_code(cpu_buf, (unsigned long)task))
 251                         goto fail;
 252         }
 253
 254         if (add_sample(cpu_buf, pc, event))
 255                 goto fail;
 256
 257         return 1;
 258
 259 fail:
 260         cpu_buf->sample_lost_overflow++;
 261         return 0;
 262 }
 263
 264 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 265 {
 266         add_code(cpu_buf, CPU_TRACE_BEGIN);
 267         cpu_buf->tracing = 1;
 268 }
 269
 270 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 271 {
 272         cpu_buf->tracing = 0;
 273 }
 274
 275 static inline void
 276 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 277                           unsigned long event, int is_kernel)
 278 {
 279         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 280
 281         if (!oprofile_backtrace_depth) {
 282                 log_sample(cpu_buf, pc, is_kernel, event);
 283                 return;
 284         }
 285
 286         oprofile_begin_trace(cpu_buf);
 287
 288         /*
 289          * if log_sample() fail we can't backtrace since we lost the
 290          * source of this event
 291          */
 292         if (log_sample(cpu_buf, pc, is_kernel, event))
 293                 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
 294
 295         oprofile_end_trace(cpu_buf);
 296 }
 297
 298 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 299                              unsigned long event, int is_kernel)
 300 {
 301         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
 302 }
 303
 304 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 305 {
 306         int is_kernel = !user_mode(regs);
 307         unsigned long pc = profile_pc(regs);
 308
 309         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
 310 }
 311
 312 #ifdef CONFIG_OPROFILE_IBS
 313
 314 #define MAX_IBS_SAMPLE_SIZE 14
 315
 316 void oprofile_add_ibs_sample(struct pt_regs * const regs,
 317                              unsigned int * const ibs_sample, int ibs_code)
 318 {
 319         int is_kernel = !user_mode(regs);
 320         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 321         struct task_struct *task;
 322         int fail = 0;
 323
 324         cpu_buf->sample_received++;
 325
 326         /* notice a switch from user->kernel or vice versa */
 327         if (cpu_buf->last_is_kernel != is_kernel) {
 328                 if (add_code(cpu_buf, is_kernel))
 329                         goto fail;
 330                 cpu_buf->last_is_kernel = is_kernel;
 331         }
 332
 333         /* notice a task switch */
 334         if (!is_kernel) {
 335                 task = current;
 336                 if (cpu_buf->last_task != task) {
 337                         if (add_code(cpu_buf, (unsigned long)task))
 338                                 goto fail;
 339                         cpu_buf->last_task = task;
 340                 }
 341         }
 342
 343         fail = fail || add_code(cpu_buf, ibs_code);
 344         fail = fail || add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
 345         fail = fail || add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
 346         fail = fail || add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
 347
 348         if (ibs_code == IBS_OP_BEGIN) {
 349                 fail = fail || add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
 350                 fail = fail || add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
 351                 fail = fail || add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
 352         }
 353
 354         if (fail)
 355                 goto fail;
 356
 357         if (oprofile_backtrace_depth)
 358                 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
 359
 360         return;
 361
 362 fail:
 363         cpu_buf->sample_lost_overflow++;
 364         return;
 365 }
 366
 367 #endif
 368
 369 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 370 {
 371         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 372         log_sample(cpu_buf, pc, is_kernel, event);
 373 }
 374
 375 void oprofile_add_trace(unsigned long pc)
 376 {
 377         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 378
 379         if (!cpu_buf->tracing)
 380                 return;
 381
 382         /*
 383          * broken frame can give an eip with the same value as an
 384          * escape code, abort the trace if we get it
 385          */
 386         if (pc == ESCAPE_CODE)
 387                 goto fail;
 388
 389         if (add_sample(cpu_buf, pc, 0))
 390                 goto fail;
 391
 392         return;
 393 fail:
 394         cpu_buf->tracing = 0;
 395         cpu_buf->backtrace_aborted++;
 396         return;
 397 }
 398
 399 /*
 400  * This serves to avoid cpu buffer overflow, and makes sure
 401  * the task mortuary progresses
 402  *
 403  * By using schedule_delayed_work_on and then schedule_delayed_work
 404  * we guarantee this will stay on the correct cpu
 405  */
 406 static void wq_sync_buffer(struct work_struct *work)
 407 {
 408         struct oprofile_cpu_buffer *b =
 409                 container_of(work, struct oprofile_cpu_buffer, work.work);
 410         if (b->cpu != smp_processor_id()) {
 411                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 412                        smp_processor_id(), b->cpu);
 413
 414                 if (!cpu_online(b->cpu)) {
 415                         cancel_delayed_work(&b->work);
 416                         return;
 417                 }
 418         }
 419         sync_buffer(b->cpu);
 420
 421         /* don't re-add the work if we're shutting down */
 422         if (work_enabled)
 423                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 424 }