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