drivers/oprofile/cpu_buffer.c

   1 /**
   2  * @file cpu_buffer.c
   3  *
   4  * @remark Copyright 2002-2009 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  * @author Robert Richter <robert.richter@amd.com>
  10  *
  11  * Each CPU has a local buffer that stores PC value/event
  12  * pairs. We also log context switches when we notice them.
  13  * Eventually each CPU's buffer is processed into the global
  14  * event buffer by sync_buffer().
  15  *
  16  * We use a local buffer for two reasons: an NMI or similar
  17  * interrupt cannot synchronise, and high sampling rates
  18  * would lead to catastrophic global synchronisation if
  19  * a global buffer was used.
  20  */
  21
  22 #include <linux/sched.h>
  23 #include <linux/oprofile.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 static struct ring_buffer *op_ring_buffer;
  34 DEFINE_PER_CPU(struct oprofile_cpu_buffer, op_cpu_buffer);
  35
  36 static void wq_sync_buffer(struct work_struct *work);
  37
  38 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
  39 static int work_enabled;
  40
  41 unsigned long oprofile_get_cpu_buffer_size(void)
  42 {
  43         return oprofile_cpu_buffer_size;
  44 }
  45
  46 void oprofile_cpu_buffer_inc_smpl_lost(void)
  47 {
  48         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
  49
  50         cpu_buf->sample_lost_overflow++;
  51 }
  52
  53 void free_cpu_buffers(void)
  54 {
  55         if (op_ring_buffer)
  56                 ring_buffer_free(op_ring_buffer);
  57         op_ring_buffer = NULL;
  58 }
  59
  60 #define RB_EVENT_HDR_SIZE 4
  61
  62 int alloc_cpu_buffers(void)
  63 {
  64         int i;
  65
  66         unsigned long buffer_size = oprofile_cpu_buffer_size;
  67         unsigned long byte_size = buffer_size * (sizeof(struct op_sample) +
  68                                                  RB_EVENT_HDR_SIZE);
  69
  70         op_ring_buffer = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS);
  71         if (!op_ring_buffer)
  72                 goto fail;
  73
  74         for_each_possible_cpu(i) {
  75                 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
  76
  77                 b->last_task = NULL;
  78                 b->last_is_kernel = -1;
  79                 b->tracing = 0;
  80                 b->buffer_size = buffer_size;
  81                 b->sample_received = 0;
  82                 b->sample_lost_overflow = 0;
  83                 b->backtrace_aborted = 0;
  84                 b->sample_invalid_eip = 0;
  85                 b->cpu = i;
  86                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
  87         }
  88         return 0;
  89
  90 fail:
  91         free_cpu_buffers();
  92         return -ENOMEM;
  93 }
  94
  95 void start_cpu_work(void)
  96 {
  97         int i;
  98
  99         work_enabled = 1;
 100
 101         for_each_online_cpu(i) {
 102                 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
 103
 104                 /*
 105                  * Spread the work by 1 jiffy per cpu so they dont all
 106                  * fire at once.
 107                  */
 108                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
 109         }
 110 }
 111
 112 void end_cpu_work(void)
 113 {
 114         int i;
 115
 116         work_enabled = 0;
 117
 118         for_each_online_cpu(i) {
 119                 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
 120
 121                 cancel_delayed_work(&b->work);
 122         }
 123
 124         flush_scheduled_work();
 125 }
 126
 127 /*
 128  * This function prepares the cpu buffer to write a sample.
 129  *
 130  * Struct op_entry is used during operations on the ring buffer while
 131  * struct op_sample contains the data that is stored in the ring
 132  * buffer. Struct entry can be uninitialized. The function reserves a
 133  * data array that is specified by size. Use
 134  * op_cpu_buffer_write_commit() after preparing the sample. In case of
 135  * errors a null pointer is returned, otherwise the pointer to the
 136  * sample.
 137  *
 138  */
 139 struct op_sample
 140 *op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size)
 141 {
 142         entry->event = ring_buffer_lock_reserve
 143                 (op_ring_buffer, sizeof(struct op_sample) +
 144                  size * sizeof(entry->sample->data[0]));
 145         if (!entry->event)
 146                 return NULL;
 147         entry->sample = ring_buffer_event_data(entry->event);
 148         entry->size = size;
 149         entry->data = entry->sample->data;
 150
 151         return entry->sample;
 152 }
 153
 154 int op_cpu_buffer_write_commit(struct op_entry *entry)
 155 {
 156         return ring_buffer_unlock_commit(op_ring_buffer, entry->event);
 157 }
 158
 159 struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
 160 {
 161         struct ring_buffer_event *e;
 162         e = ring_buffer_consume(op_ring_buffer, cpu, NULL, NULL);
 163         if (!e)
 164                 return NULL;
 165
 166         entry->event = e;
 167         entry->sample = ring_buffer_event_data(e);
 168         entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
 169                 / sizeof(entry->sample->data[0]);
 170         entry->data = entry->sample->data;
 171         return entry->sample;
 172 }
 173
 174 unsigned long op_cpu_buffer_entries(int cpu)
 175 {
 176         return ring_buffer_entries_cpu(op_ring_buffer, cpu);
 177 }
 178
 179 static int
 180 op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
 181             int is_kernel, struct task_struct *task)
 182 {
 183         struct op_entry entry;
 184         struct op_sample *sample;
 185         unsigned long flags;
 186         int size;
 187
 188         flags = 0;
 189
 190         if (backtrace)
 191                 flags |= TRACE_BEGIN;
 192
 193         /* notice a switch from user->kernel or vice versa */
 194         is_kernel = !!is_kernel;
 195         if (cpu_buf->last_is_kernel != is_kernel) {
 196                 cpu_buf->last_is_kernel = is_kernel;
 197                 flags |= KERNEL_CTX_SWITCH;
 198                 if (is_kernel)
 199                         flags |= IS_KERNEL;
 200         }
 201
 202         /* notice a task switch */
 203         if (cpu_buf->last_task != task) {
 204                 cpu_buf->last_task = task;
 205                 flags |= USER_CTX_SWITCH;
 206         }
 207
 208         if (!flags)
 209                 /* nothing to do */
 210                 return 0;
 211
 212         if (flags & USER_CTX_SWITCH)
 213                 size = 1;
 214         else
 215                 size = 0;
 216
 217         sample = op_cpu_buffer_write_reserve(&entry, size);
 218         if (!sample)
 219                 return -ENOMEM;
 220
 221         sample->eip = ESCAPE_CODE;
 222         sample->event = flags;
 223
 224         if (size)
 225                 op_cpu_buffer_add_data(&entry, (unsigned long)task);
 226
 227         op_cpu_buffer_write_commit(&entry);
 228
 229         return 0;
 230 }
 231
 232 static inline int
 233 op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
 234               unsigned long pc, unsigned long event)
 235 {
 236         struct op_entry entry;
 237         struct op_sample *sample;
 238
 239         sample = op_cpu_buffer_write_reserve(&entry, 0);
 240         if (!sample)
 241                 return -ENOMEM;
 242
 243         sample->eip = pc;
 244         sample->event = event;
 245
 246         return op_cpu_buffer_write_commit(&entry);
 247 }
 248
 249 /*
 250  * This must be safe from any context.
 251  *
 252  * is_kernel is needed because on some architectures you cannot
 253  * tell if you are in kernel or user space simply by looking at
 254  * pc. We tag this in the buffer by generating kernel enter/exit
 255  * events whenever is_kernel changes
 256  */
 257 static int
 258 log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 259            unsigned long backtrace, int is_kernel, unsigned long event)
 260 {
 261         cpu_buf->sample_received++;
 262
 263         if (pc == ESCAPE_CODE) {
 264                 cpu_buf->sample_invalid_eip++;
 265                 return 0;
 266         }
 267
 268         if (op_add_code(cpu_buf, backtrace, is_kernel, current))
 269                 goto fail;
 270
 271         if (op_add_sample(cpu_buf, pc, event))
 272                 goto fail;
 273
 274         return 1;
 275
 276 fail:
 277         cpu_buf->sample_lost_overflow++;
 278         return 0;
 279 }
 280
 281 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 282 {
 283         cpu_buf->tracing = 1;
 284 }
 285
 286 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 287 {
 288         cpu_buf->tracing = 0;
 289 }
 290
 291 static inline void
 292 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 293                           unsigned long event, int is_kernel)
 294 {
 295         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 296         unsigned long backtrace = oprofile_backtrace_depth;
 297
 298         /*
 299          * if log_sample() fail we can't backtrace since we lost the
 300          * source of this event
 301          */
 302         if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event))
 303                 /* failed */
 304                 return;
 305
 306         if (!backtrace)
 307                 return;
 308
 309         oprofile_begin_trace(cpu_buf);
 310         oprofile_ops.backtrace(regs, backtrace);
 311         oprofile_end_trace(cpu_buf);
 312 }
 313
 314 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 315                              unsigned long event, int is_kernel)
 316 {
 317         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
 318 }
 319
 320 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 321 {
 322         int is_kernel;
 323         unsigned long pc;
 324
 325         if (likely(regs)) {
 326                 is_kernel = !user_mode(regs);
 327                 pc = profile_pc(regs);
 328         } else {
 329                 is_kernel = 0;    /* This value will not be used */
 330                 pc = ESCAPE_CODE; /* as this causes an early return. */
 331         }
 332
 333         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
 334 }
 335
 336 /*
 337  * Add samples with data to the ring buffer.
 338  *
 339  * Use oprofile_add_data(&entry, val) to add data and
 340  * oprofile_write_commit(&entry) to commit the sample.
 341  */
 342 void
 343 oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs,
 344                        unsigned long pc, int code, int size)
 345 {
 346         struct op_sample *sample;
 347         int is_kernel = !user_mode(regs);
 348         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 349
 350         cpu_buf->sample_received++;
 351
 352         /* no backtraces for samples with data */
 353         if (op_add_code(cpu_buf, 0, is_kernel, current))
 354                 goto fail;
 355
 356         sample = op_cpu_buffer_write_reserve(entry, size + 2);
 357         if (!sample)
 358                 goto fail;
 359         sample->eip = ESCAPE_CODE;
 360         sample->event = 0;              /* no flags */
 361
 362         op_cpu_buffer_add_data(entry, code);
 363         op_cpu_buffer_add_data(entry, pc);
 364
 365         return;
 366
 367 fail:
 368         entry->event = NULL;
 369         cpu_buf->sample_lost_overflow++;
 370 }
 371
 372 int oprofile_add_data(struct op_entry *entry, unsigned long val)
 373 {
 374         if (!entry->event)
 375                 return 0;
 376         return op_cpu_buffer_add_data(entry, val);
 377 }
 378
 379 int oprofile_add_data64(struct op_entry *entry, u64 val)
 380 {
 381         if (!entry->event)
 382                 return 0;
 383         if (op_cpu_buffer_get_size(entry) < 2)
 384                 /*
 385                  * the function returns 0 to indicate a too small
 386                  * buffer, even if there is some space left
 387                  */
 388                 return 0;
 389         if (!op_cpu_buffer_add_data(entry, (u32)val))
 390                 return 0;
 391         return op_cpu_buffer_add_data(entry, (u32)(val >> 32));
 392 }
 393
 394 int oprofile_write_commit(struct op_entry *entry)
 395 {
 396         if (!entry->event)
 397                 return -EINVAL;
 398         return op_cpu_buffer_write_commit(entry);
 399 }
 400
 401 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 402 {
 403         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 404         log_sample(cpu_buf, pc, 0, is_kernel, event);
 405 }
 406
 407 void oprofile_add_trace(unsigned long pc)
 408 {
 409         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 410
 411         if (!cpu_buf->tracing)
 412                 return;
 413
 414         /*
 415          * broken frame can give an eip with the same value as an
 416          * escape code, abort the trace if we get it
 417          */
 418         if (pc == ESCAPE_CODE)
 419                 goto fail;
 420
 421         if (op_add_sample(cpu_buf, pc, 0))
 422                 goto fail;
 423
 424         return;
 425 fail:
 426         cpu_buf->tracing = 0;
 427         cpu_buf->backtrace_aborted++;
 428         return;
 429 }
 430
 431 /*
 432  * This serves to avoid cpu buffer overflow, and makes sure
 433  * the task mortuary progresses
 434  *
 435  * By using schedule_delayed_work_on and then schedule_delayed_work
 436  * we guarantee this will stay on the correct cpu
 437  */
 438 static void wq_sync_buffer(struct work_struct *work)
 439 {
 440         struct oprofile_cpu_buffer *b =
 441                 container_of(work, struct oprofile_cpu_buffer, work.work);
 442         if (b->cpu != smp_processor_id()) {
 443                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 444                        smp_processor_id(), b->cpu);
 445
 446                 if (!cpu_online(b->cpu)) {
 447                         cancel_delayed_work(&b->work);
 448                         return;
 449                 }
 450         }
 451         sync_buffer(b->cpu);
 452
 453         /* don't re-add the work if we're shutting down */
 454         if (work_enabled)
 455                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 456 }