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 /*
  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, op_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 = &__get_cpu_var(op_cpu_buffer);
  65
  66         cpu_buf->sample_lost_overflow++;
  67 }
  68
  69 void free_cpu_buffers(void)
  70 {
  71         if (op_ring_buffer_read)
  72                 ring_buffer_free(op_ring_buffer_read);
  73         op_ring_buffer_read = NULL;
  74         if (op_ring_buffer_write)
  75                 ring_buffer_free(op_ring_buffer_write);
  76         op_ring_buffer_write = NULL;
  77 }
  78
  79 #define RB_EVENT_HDR_SIZE 4
  80
  81 int alloc_cpu_buffers(void)
  82 {
  83         int i;
  84
  85         unsigned long buffer_size = oprofile_cpu_buffer_size;
  86         unsigned long byte_size = buffer_size * (sizeof(struct op_sample) +
  87                                                  RB_EVENT_HDR_SIZE);
  88
  89         op_ring_buffer_read = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS);
  90         if (!op_ring_buffer_read)
  91                 goto fail;
  92         op_ring_buffer_write = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS);
  93         if (!op_ring_buffer_write)
  94                 goto fail;
  95
  96         for_each_possible_cpu(i) {
  97                 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
  98
  99                 b->last_task = NULL;
 100                 b->last_is_kernel = -1;
 101                 b->tracing = 0;
 102                 b->buffer_size = buffer_size;
 103                 b->sample_received = 0;
 104                 b->sample_lost_overflow = 0;
 105                 b->backtrace_aborted = 0;
 106                 b->sample_invalid_eip = 0;
 107                 b->cpu = i;
 108                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
 109         }
 110         return 0;
 111
 112 fail:
 113         free_cpu_buffers();
 114         return -ENOMEM;
 115 }
 116
 117 void start_cpu_work(void)
 118 {
 119         int i;
 120
 121         work_enabled = 1;
 122
 123         for_each_online_cpu(i) {
 124                 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
 125
 126                 /*
 127                  * Spread the work by 1 jiffy per cpu so they dont all
 128                  * fire at once.
 129                  */
 130                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
 131         }
 132 }
 133
 134 void end_cpu_work(void)
 135 {
 136         int i;
 137
 138         work_enabled = 0;
 139
 140         for_each_online_cpu(i) {
 141                 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
 142
 143                 cancel_delayed_work(&b->work);
 144         }
 145
 146         flush_scheduled_work();
 147 }
 148
 149 /*
 150  * This function prepares the cpu buffer to write a sample.
 151  *
 152  * Struct op_entry is used during operations on the ring buffer while
 153  * struct op_sample contains the data that is stored in the ring
 154  * buffer. Struct entry can be uninitialized. The function reserves a
 155  * data array that is specified by size. Use
 156  * op_cpu_buffer_write_commit() after preparing the sample. In case of
 157  * errors a null pointer is returned, otherwise the pointer to the
 158  * sample.
 159  *
 160  */
 161 struct op_sample
 162 *op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size)
 163 {
 164         entry->event = ring_buffer_lock_reserve
 165                 (op_ring_buffer_write, sizeof(struct op_sample) +
 166                  size * sizeof(entry->sample->data[0]));
 167         if (entry->event)
 168                 entry->sample = ring_buffer_event_data(entry->event);
 169         else
 170                 entry->sample = NULL;
 171
 172         if (!entry->sample)
 173                 return NULL;
 174
 175         entry->size = size;
 176         entry->data = entry->sample->data;
 177
 178         return entry->sample;
 179 }
 180
 181 int op_cpu_buffer_write_commit(struct op_entry *entry)
 182 {
 183         return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event);
 184 }
 185
 186 struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
 187 {
 188         struct ring_buffer_event *e;
 189         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
 190         if (e)
 191                 goto event;
 192         if (ring_buffer_swap_cpu(op_ring_buffer_read,
 193                                  op_ring_buffer_write,
 194                                  cpu))
 195                 return NULL;
 196         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
 197         if (e)
 198                 goto event;
 199         return NULL;
 200
 201 event:
 202         entry->event = e;
 203         entry->sample = ring_buffer_event_data(e);
 204         entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
 205                 / sizeof(entry->sample->data[0]);
 206         entry->data = entry->sample->data;
 207         return entry->sample;
 208 }
 209
 210 unsigned long op_cpu_buffer_entries(int cpu)
 211 {
 212         return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
 213                 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
 214 }
 215
 216 static int
 217 op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
 218             int is_kernel, struct task_struct *task)
 219 {
 220         struct op_entry entry;
 221         struct op_sample *sample;
 222         unsigned long flags;
 223         int size;
 224
 225         flags = 0;
 226
 227         if (backtrace)
 228                 flags |= TRACE_BEGIN;
 229
 230         /* notice a switch from user->kernel or vice versa */
 231         is_kernel = !!is_kernel;
 232         if (cpu_buf->last_is_kernel != is_kernel) {
 233                 cpu_buf->last_is_kernel = is_kernel;
 234                 flags |= KERNEL_CTX_SWITCH;
 235                 if (is_kernel)
 236                         flags |= IS_KERNEL;
 237         }
 238
 239         /* notice a task switch */
 240         if (cpu_buf->last_task != task) {
 241                 cpu_buf->last_task = task;
 242                 flags |= USER_CTX_SWITCH;
 243         }
 244
 245         if (!flags)
 246                 /* nothing to do */
 247                 return 0;
 248
 249         if (flags & USER_CTX_SWITCH)
 250                 size = 1;
 251         else
 252                 size = 0;
 253
 254         sample = op_cpu_buffer_write_reserve(&entry, size);
 255         if (!sample)
 256                 return -ENOMEM;
 257
 258         sample->eip = ESCAPE_CODE;
 259         sample->event = flags;
 260
 261         if (size)
 262                 op_cpu_buffer_add_data(&entry, (unsigned long)task);
 263
 264         op_cpu_buffer_write_commit(&entry);
 265
 266         return 0;
 267 }
 268
 269 static inline int
 270 op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
 271               unsigned long pc, unsigned long event)
 272 {
 273         struct op_entry entry;
 274         struct op_sample *sample;
 275
 276         sample = op_cpu_buffer_write_reserve(&entry, 0);
 277         if (!sample)
 278                 return -ENOMEM;
 279
 280         sample->eip = pc;
 281         sample->event = event;
 282
 283         return op_cpu_buffer_write_commit(&entry);
 284 }
 285
 286 /*
 287  * This must be safe from any context.
 288  *
 289  * is_kernel is needed because on some architectures you cannot
 290  * tell if you are in kernel or user space simply by looking at
 291  * pc. We tag this in the buffer by generating kernel enter/exit
 292  * events whenever is_kernel changes
 293  */
 294 static int
 295 log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 296            unsigned long backtrace, int is_kernel, unsigned long event)
 297 {
 298         cpu_buf->sample_received++;
 299
 300         if (pc == ESCAPE_CODE) {
 301                 cpu_buf->sample_invalid_eip++;
 302                 return 0;
 303         }
 304
 305         if (op_add_code(cpu_buf, backtrace, is_kernel, current))
 306                 goto fail;
 307
 308         if (op_add_sample(cpu_buf, pc, event))
 309                 goto fail;
 310
 311         return 1;
 312
 313 fail:
 314         cpu_buf->sample_lost_overflow++;
 315         return 0;
 316 }
 317
 318 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 319 {
 320         cpu_buf->tracing = 1;
 321 }
 322
 323 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 324 {
 325         cpu_buf->tracing = 0;
 326 }
 327
 328 static inline void
 329 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 330                           unsigned long event, int is_kernel)
 331 {
 332         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 333         unsigned long backtrace = oprofile_backtrace_depth;
 334
 335         /*
 336          * if log_sample() fail we can't backtrace since we lost the
 337          * source of this event
 338          */
 339         if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event))
 340                 /* failed */
 341                 return;
 342
 343         if (!backtrace)
 344                 return;
 345
 346         oprofile_begin_trace(cpu_buf);
 347         oprofile_ops.backtrace(regs, backtrace);
 348         oprofile_end_trace(cpu_buf);
 349 }
 350
 351 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 352                              unsigned long event, int is_kernel)
 353 {
 354         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
 355 }
 356
 357 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 358 {
 359         int is_kernel = !user_mode(regs);
 360         unsigned long pc = profile_pc(regs);
 361
 362         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
 363 }
 364
 365 /*
 366  * Add samples with data to the ring buffer.
 367  *
 368  * Use oprofile_add_data(&entry, val) to add data and
 369  * oprofile_write_commit(&entry) to commit the sample.
 370  */
 371 void
 372 oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs,
 373                        unsigned long pc, int code, int size)
 374 {
 375         struct op_sample *sample;
 376         int is_kernel = !user_mode(regs);
 377         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 378
 379         cpu_buf->sample_received++;
 380
 381         /* no backtraces for samples with data */
 382         if (op_add_code(cpu_buf, 0, is_kernel, current))
 383                 goto fail;
 384
 385         sample = op_cpu_buffer_write_reserve(entry, size + 2);
 386         if (!sample)
 387                 goto fail;
 388         sample->eip = ESCAPE_CODE;
 389         sample->event = 0;              /* no flags */
 390
 391         op_cpu_buffer_add_data(entry, code);
 392         op_cpu_buffer_add_data(entry, pc);
 393
 394         return;
 395
 396 fail:
 397         entry->event = NULL;
 398         cpu_buf->sample_lost_overflow++;
 399 }
 400
 401 int oprofile_add_data(struct op_entry *entry, unsigned long val)
 402 {
 403         if (!entry->event)
 404                 return 0;
 405         return op_cpu_buffer_add_data(entry, val);
 406 }
 407
 408 int oprofile_add_data64(struct op_entry *entry, u64 val)
 409 {
 410         if (!entry->event)
 411                 return 0;
 412         if (op_cpu_buffer_get_size(entry) < 2)
 413                 /*
 414                  * the function returns 0 to indicate a too small
 415                  * buffer, even if there is some space left
 416                  */
 417                 return 0;
 418         if (!op_cpu_buffer_add_data(entry, (u32)val))
 419                 return 0;
 420         return op_cpu_buffer_add_data(entry, (u32)(val >> 32));
 421 }
 422
 423 int oprofile_write_commit(struct op_entry *entry)
 424 {
 425         if (!entry->event)
 426                 return -EINVAL;
 427         return op_cpu_buffer_write_commit(entry);
 428 }
 429
 430 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 431 {
 432         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 433         log_sample(cpu_buf, pc, 0, is_kernel, event);
 434 }
 435
 436 void oprofile_add_trace(unsigned long pc)
 437 {
 438         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
 439
 440         if (!cpu_buf->tracing)
 441                 return;
 442
 443         /*
 444          * broken frame can give an eip with the same value as an
 445          * escape code, abort the trace if we get it
 446          */
 447         if (pc == ESCAPE_CODE)
 448                 goto fail;
 449
 450         if (op_add_sample(cpu_buf, pc, 0))
 451                 goto fail;
 452
 453         return;
 454 fail:
 455         cpu_buf->tracing = 0;
 456         cpu_buf->backtrace_aborted++;
 457         return;
 458 }
 459
 460 /*
 461  * This serves to avoid cpu buffer overflow, and makes sure
 462  * the task mortuary progresses
 463  *
 464  * By using schedule_delayed_work_on and then schedule_delayed_work
 465  * we guarantee this will stay on the correct cpu
 466  */
 467 static void wq_sync_buffer(struct work_struct *work)
 468 {
 469         struct oprofile_cpu_buffer *b =
 470                 container_of(work, struct oprofile_cpu_buffer, work.work);
 471         if (b->cpu != smp_processor_id()) {
 472                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 473                        smp_processor_id(), b->cpu);
 474
 475                 if (!cpu_online(b->cpu)) {
 476                         cancel_delayed_work(&b->work);
 477                         return;
 478                 }
 479         }
 480         sync_buffer(b->cpu);
 481
 482         /* don't re-add the work if we're shutting down */
 483         if (work_enabled)
 484                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 485 }