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 void free_cpu_buffers(void)
  58 {
  59         if (op_ring_buffer_read)
  60                 ring_buffer_free(op_ring_buffer_read);
  61         op_ring_buffer_read = NULL;
  62         if (op_ring_buffer_write)
  63                 ring_buffer_free(op_ring_buffer_write);
  64         op_ring_buffer_write = NULL;
  65 }
  66
  67 unsigned long oprofile_get_cpu_buffer_size(void)
  68 {
  69         return oprofile_cpu_buffer_size;
  70 }
  71
  72 void oprofile_cpu_buffer_inc_smpl_lost(void)
  73 {
  74         struct oprofile_cpu_buffer *cpu_buf
  75                 = &__get_cpu_var(cpu_buffer);
  76
  77         cpu_buf->sample_lost_overflow++;
  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         ret = op_cpu_buffer_write_commit(&entry);
 207         if (ret)
 208                 return ret;
 209
 210         return 0;
 211 }
 212
 213 static inline int
 214 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
 215 {
 216         return add_sample(buffer, ESCAPE_CODE, value);
 217 }
 218
 219 /* This must be safe from any context. It's safe writing here
 220  * because of the head/tail separation of the writer and reader
 221  * of the CPU buffer.
 222  *
 223  * is_kernel is needed because on some architectures you cannot
 224  * tell if you are in kernel or user space simply by looking at
 225  * pc. We tag this in the buffer by generating kernel enter/exit
 226  * events whenever is_kernel changes
 227  */
 228 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 229                       int is_kernel, unsigned long event)
 230 {
 231         struct task_struct *task;
 232
 233         cpu_buf->sample_received++;
 234
 235         if (pc == ESCAPE_CODE) {
 236                 cpu_buf->sample_invalid_eip++;
 237                 return 0;
 238         }
 239
 240         is_kernel = !!is_kernel;
 241
 242         task = current;
 243
 244         /* notice a switch from user->kernel or vice versa */
 245         if (cpu_buf->last_is_kernel != is_kernel) {
 246                 cpu_buf->last_is_kernel = is_kernel;
 247                 if (add_code(cpu_buf, is_kernel))
 248                         goto fail;
 249         }
 250
 251         /* notice a task switch */
 252         if (cpu_buf->last_task != task) {
 253                 cpu_buf->last_task = task;
 254                 if (add_code(cpu_buf, (unsigned long)task))
 255                         goto fail;
 256         }
 257
 258         if (add_sample(cpu_buf, pc, event))
 259                 goto fail;
 260
 261         return 1;
 262
 263 fail:
 264         cpu_buf->sample_lost_overflow++;
 265         return 0;
 266 }
 267
 268 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 269 {
 270         add_code(cpu_buf, CPU_TRACE_BEGIN);
 271         cpu_buf->tracing = 1;
 272         return 1;
 273 }
 274
 275 static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 276 {
 277         cpu_buf->tracing = 0;
 278 }
 279
 280 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 281                                 unsigned long event, int is_kernel)
 282 {
 283         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 284
 285         if (!oprofile_backtrace_depth) {
 286                 log_sample(cpu_buf, pc, is_kernel, event);
 287                 return;
 288         }
 289
 290         if (!oprofile_begin_trace(cpu_buf))
 291                 return;
 292
 293         /*
 294          * if log_sample() fail we can't backtrace since we lost the
 295          * source of this event
 296          */
 297         if (log_sample(cpu_buf, pc, is_kernel, event))
 298                 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
 299         oprofile_end_trace(cpu_buf);
 300 }
 301
 302 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 303 {
 304         int is_kernel = !user_mode(regs);
 305         unsigned long pc = profile_pc(regs);
 306
 307         oprofile_add_ext_sample(pc, regs, event, is_kernel);
 308 }
 309
 310 #ifdef CONFIG_OPROFILE_IBS
 311
 312 #define MAX_IBS_SAMPLE_SIZE 14
 313
 314 void oprofile_add_ibs_sample(struct pt_regs * const regs,
 315                              unsigned int * const ibs_sample, int ibs_code)
 316 {
 317         int is_kernel = !user_mode(regs);
 318         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 319         struct task_struct *task;
 320         int fail = 0;
 321
 322         cpu_buf->sample_received++;
 323
 324         /* notice a switch from user->kernel or vice versa */
 325         if (cpu_buf->last_is_kernel != is_kernel) {
 326                 if (add_code(cpu_buf, is_kernel))
 327                         goto fail;
 328                 cpu_buf->last_is_kernel = is_kernel;
 329         }
 330
 331         /* notice a task switch */
 332         if (!is_kernel) {
 333                 task = current;
 334                 if (cpu_buf->last_task != task) {
 335                         if (add_code(cpu_buf, (unsigned long)task))
 336                                 goto fail;
 337                         cpu_buf->last_task = task;
 338                 }
 339         }
 340
 341         fail = fail || add_code(cpu_buf, ibs_code);
 342         fail = fail || add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
 343         fail = fail || add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
 344         fail = fail || add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
 345
 346         if (ibs_code == IBS_OP_BEGIN) {
 347                 fail = fail || add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
 348                 fail = fail || add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
 349                 fail = fail || add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
 350         }
 351
 352         if (fail)
 353                 goto fail;
 354
 355         if (oprofile_backtrace_depth)
 356                 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
 357
 358         return;
 359
 360 fail:
 361         cpu_buf->sample_lost_overflow++;
 362         return;
 363 }
 364
 365 #endif
 366
 367 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 368 {
 369         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 370         log_sample(cpu_buf, pc, is_kernel, event);
 371 }
 372
 373 void oprofile_add_trace(unsigned long pc)
 374 {
 375         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 376
 377         if (!cpu_buf->tracing)
 378                 return;
 379
 380         /*
 381          * broken frame can give an eip with the same value as an
 382          * escape code, abort the trace if we get it
 383          */
 384         if (pc == ESCAPE_CODE)
 385                 goto fail;
 386
 387         if (add_sample(cpu_buf, pc, 0))
 388                 goto fail;
 389
 390         return;
 391 fail:
 392         cpu_buf->tracing = 0;
 393         cpu_buf->backtrace_aborted++;
 394         return;
 395 }
 396
 397 /*
 398  * This serves to avoid cpu buffer overflow, and makes sure
 399  * the task mortuary progresses
 400  *
 401  * By using schedule_delayed_work_on and then schedule_delayed_work
 402  * we guarantee this will stay on the correct cpu
 403  */
 404 static void wq_sync_buffer(struct work_struct *work)
 405 {
 406         struct oprofile_cpu_buffer *b =
 407                 container_of(work, struct oprofile_cpu_buffer, work.work);
 408         if (b->cpu != smp_processor_id()) {
 409                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 410                        smp_processor_id(), b->cpu);
 411
 412                 if (!cpu_online(b->cpu)) {
 413                         cancel_delayed_work(&b->work);
 414                         return;
 415                 }
 416         }
 417         sync_buffer(b->cpu);
 418
 419         /* don't re-add the work if we're shutting down */
 420         if (work_enabled)
 421                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 422 }