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->sample_received = 0;
 101                 b->sample_lost_overflow = 0;
 102                 b->backtrace_aborted = 0;
 103                 b->sample_invalid_eip = 0;
 104                 b->cpu = i;
 105                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
 106         }
 107         return 0;
 108
 109 fail:
 110         free_cpu_buffers();
 111         return -ENOMEM;
 112 }
 113
 114 void start_cpu_work(void)
 115 {
 116         int i;
 117
 118         work_enabled = 1;
 119
 120         for_each_online_cpu(i) {
 121                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
 122
 123                 /*
 124                  * Spread the work by 1 jiffy per cpu so they dont all
 125                  * fire at once.
 126                  */
 127                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
 128         }
 129 }
 130
 131 void end_cpu_work(void)
 132 {
 133         int i;
 134
 135         work_enabled = 0;
 136
 137         for_each_online_cpu(i) {
 138                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
 139
 140                 cancel_delayed_work(&b->work);
 141         }
 142
 143         flush_scheduled_work();
 144 }
 145
 146 int op_cpu_buffer_write_entry(struct op_entry *entry)
 147 {
 148         entry->event = ring_buffer_lock_reserve(op_ring_buffer_write,
 149                                                 sizeof(struct op_sample),
 150                                                 &entry->irq_flags);
 151         if (entry->event)
 152                 entry->sample = ring_buffer_event_data(entry->event);
 153         else
 154                 entry->sample = NULL;
 155
 156         if (!entry->sample)
 157                 return -ENOMEM;
 158
 159         return 0;
 160 }
 161
 162 int op_cpu_buffer_write_commit(struct op_entry *entry)
 163 {
 164         return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event,
 165                                          entry->irq_flags);
 166 }
 167
 168 struct op_sample *op_cpu_buffer_read_entry(int cpu)
 169 {
 170         struct ring_buffer_event *e;
 171         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
 172         if (e)
 173                 return ring_buffer_event_data(e);
 174         if (ring_buffer_swap_cpu(op_ring_buffer_read,
 175                                  op_ring_buffer_write,
 176                                  cpu))
 177                 return NULL;
 178         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
 179         if (e)
 180                 return ring_buffer_event_data(e);
 181         return NULL;
 182 }
 183
 184 unsigned long op_cpu_buffer_entries(int cpu)
 185 {
 186         return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
 187                 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
 188 }
 189
 190 static inline int
 191 add_sample(struct oprofile_cpu_buffer *cpu_buf,
 192            unsigned long pc, unsigned long event)
 193 {
 194         struct op_entry entry;
 195         int ret;
 196
 197         ret = op_cpu_buffer_write_entry(&entry);
 198         if (ret)
 199                 return ret;
 200
 201         entry.sample->eip = pc;
 202         entry.sample->event = event;
 203
 204         return op_cpu_buffer_write_commit(&entry);
 205 }
 206
 207 static inline int
 208 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
 209 {
 210         return add_sample(buffer, ESCAPE_CODE, value);
 211 }
 212
 213 /* This must be safe from any context. It's safe writing here
 214  * because of the head/tail separation of the writer and reader
 215  * of the CPU buffer.
 216  *
 217  * is_kernel is needed because on some architectures you cannot
 218  * tell if you are in kernel or user space simply by looking at
 219  * pc. We tag this in the buffer by generating kernel enter/exit
 220  * events whenever is_kernel changes
 221  */
 222 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 223                       int is_kernel, unsigned long event)
 224 {
 225         struct task_struct *task;
 226
 227         cpu_buf->sample_received++;
 228
 229         if (pc == ESCAPE_CODE) {
 230                 cpu_buf->sample_invalid_eip++;
 231                 return 0;
 232         }
 233
 234         is_kernel = !!is_kernel;
 235
 236         task = current;
 237
 238         /* notice a switch from user->kernel or vice versa */
 239         if (cpu_buf->last_is_kernel != is_kernel) {
 240                 cpu_buf->last_is_kernel = is_kernel;
 241                 if (add_code(cpu_buf, is_kernel))
 242                         goto fail;
 243         }
 244
 245         /* notice a task switch */
 246         if (cpu_buf->last_task != task) {
 247                 cpu_buf->last_task = task;
 248                 if (add_code(cpu_buf, (unsigned long)task))
 249                         goto fail;
 250         }
 251
 252         if (add_sample(cpu_buf, pc, event))
 253                 goto fail;
 254
 255         return 1;
 256
 257 fail:
 258         cpu_buf->sample_lost_overflow++;
 259         return 0;
 260 }
 261
 262 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 263 {
 264         add_code(cpu_buf, CPU_TRACE_BEGIN);
 265         cpu_buf->tracing = 1;
 266 }
 267
 268 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 269 {
 270         cpu_buf->tracing = 0;
 271 }
 272
 273 static inline void
 274 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 275                           unsigned long event, int is_kernel)
 276 {
 277         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 278
 279         if (!oprofile_backtrace_depth) {
 280                 log_sample(cpu_buf, pc, is_kernel, event);
 281                 return;
 282         }
 283
 284         oprofile_begin_trace(cpu_buf);
 285
 286         /*
 287          * if log_sample() fail we can't backtrace since we lost the
 288          * source of this event
 289          */
 290         if (log_sample(cpu_buf, pc, is_kernel, event))
 291                 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
 292
 293         oprofile_end_trace(cpu_buf);
 294 }
 295
 296 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 297                              unsigned long event, int is_kernel)
 298 {
 299         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
 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 }