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 <[email protected]>
   8  * @author Barry Kasindorf <[email protected]>
   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_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
  32
  33 static void wq_sync_buffer(struct work_struct *work);
  34
  35 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
  36 static int work_enabled;
  37
  38 void free_cpu_buffers(void)
  39 {
  40         int i;
  41
  42         for_each_online_cpu(i)
  43                 vfree(per_cpu(cpu_buffer, i).buffer);
  44 }
  45
  46 int alloc_cpu_buffers(void)
  47 {
  48         int i;
  49
  50         unsigned long buffer_size = fs_cpu_buffer_size;
  51
  52         for_each_online_cpu(i) {
  53                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
  54
  55                 b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
  56                         cpu_to_node(i));
  57                 if (!b->buffer)
  58                         goto fail;
  59
  60                 b->last_task = NULL;
  61                 b->last_is_kernel = -1;
  62                 b->tracing = 0;
  63                 b->buffer_size = buffer_size;
  64                 b->tail_pos = 0;
  65                 b->head_pos = 0;
  66                 b->sample_received = 0;
  67                 b->sample_lost_overflow = 0;
  68                 b->backtrace_aborted = 0;
  69                 b->sample_invalid_eip = 0;
  70                 b->cpu = i;
  71                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
  72         }
  73         return 0;
  74
  75 fail:
  76         free_cpu_buffers();
  77         return -ENOMEM;
  78 }
  79
  80 void start_cpu_work(void)
  81 {
  82         int i;
  83
  84         work_enabled = 1;
  85
  86         for_each_online_cpu(i) {
  87                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
  88
  89                 /*
  90                  * Spread the work by 1 jiffy per cpu so they dont all
  91                  * fire at once.
  92                  */
  93                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
  94         }
  95 }
  96
  97 void end_cpu_work(void)
  98 {
  99         int i;
 100
 101         work_enabled = 0;
 102
 103         for_each_online_cpu(i) {
 104                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
 105
 106                 cancel_delayed_work(&b->work);
 107         }
 108
 109         flush_scheduled_work();
 110 }
 111
 112 /* Resets the cpu buffer to a sane state. */
 113 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
 114 {
 115         /* reset these to invalid values; the next sample
 116          * collected will populate the buffer with proper
 117          * values to initialize the buffer
 118          */
 119         cpu_buf->last_is_kernel = -1;
 120         cpu_buf->last_task = NULL;
 121 }
 122
 123 /* compute number of available slots in cpu_buffer queue */
 124 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
 125 {
 126         unsigned long head = b->head_pos;
 127         unsigned long tail = b->tail_pos;
 128
 129         if (tail > head)
 130                 return (tail - head) - 1;
 131
 132         return tail + (b->buffer_size - head) - 1;
 133 }
 134
 135 static void increment_head(struct oprofile_cpu_buffer * b)
 136 {
 137         unsigned long new_head = b->head_pos + 1;
 138
 139         /* Ensure anything written to the slot before we
 140          * increment is visible */
 141         wmb();
 142
 143         if (new_head < b->buffer_size)
 144                 b->head_pos = new_head;
 145         else
 146                 b->head_pos = 0;
 147 }
 148
 149 static inline void
 150 add_sample(struct oprofile_cpu_buffer * cpu_buf,
 151            unsigned long pc, unsigned long event)
 152 {
 153         struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
 154         entry->eip = pc;
 155         entry->event = event;
 156         increment_head(cpu_buf);
 157 }
 158
 159 static inline void
 160 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
 161 {
 162         add_sample(buffer, ESCAPE_CODE, value);
 163 }
 164
 165 /* This must be safe from any context. It's safe writing here
 166  * because of the head/tail separation of the writer and reader
 167  * of the CPU buffer.
 168  *
 169  * is_kernel is needed because on some architectures you cannot
 170  * tell if you are in kernel or user space simply by looking at
 171  * pc. We tag this in the buffer by generating kernel enter/exit
 172  * events whenever is_kernel changes
 173  */
 174 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
 175                       int is_kernel, unsigned long event)
 176 {
 177         struct task_struct * task;
 178
 179         cpu_buf->sample_received++;
 180
 181         if (pc == ESCAPE_CODE) {
 182                 cpu_buf->sample_invalid_eip++;
 183                 return 0;
 184         }
 185
 186         if (nr_available_slots(cpu_buf) < 3) {
 187                 cpu_buf->sample_lost_overflow++;
 188                 return 0;
 189         }
 190
 191         is_kernel = !!is_kernel;
 192
 193         task = current;
 194
 195         /* notice a switch from user->kernel or vice versa */
 196         if (cpu_buf->last_is_kernel != is_kernel) {
 197                 cpu_buf->last_is_kernel = is_kernel;
 198                 add_code(cpu_buf, is_kernel);
 199         }
 200
 201         /* notice a task switch */
 202         if (cpu_buf->last_task != task) {
 203                 cpu_buf->last_task = task;
 204                 add_code(cpu_buf, (unsigned long)task);
 205         }
 206
 207         add_sample(cpu_buf, pc, event);
 208         return 1;
 209 }
 210
 211 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 212 {
 213         if (nr_available_slots(cpu_buf) < 4) {
 214                 cpu_buf->sample_lost_overflow++;
 215                 return 0;
 216         }
 217
 218         add_code(cpu_buf, CPU_TRACE_BEGIN);
 219         cpu_buf->tracing = 1;
 220         return 1;
 221 }
 222
 223 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
 224 {
 225         cpu_buf->tracing = 0;
 226 }
 227
 228 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 229                                 unsigned long event, int is_kernel)
 230 {
 231         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 232
 233         if (!backtrace_depth) {
 234                 log_sample(cpu_buf, pc, is_kernel, event);
 235                 return;
 236         }
 237
 238         if (!oprofile_begin_trace(cpu_buf))
 239                 return;
 240
 241         /* if log_sample() fail we can't backtrace since we lost the source
 242          * of this event */
 243         if (log_sample(cpu_buf, pc, is_kernel, event))
 244                 oprofile_ops.backtrace(regs, backtrace_depth);
 245         oprofile_end_trace(cpu_buf);
 246 }
 247
 248 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 249 {
 250         int is_kernel = !user_mode(regs);
 251         unsigned long pc = profile_pc(regs);
 252
 253         oprofile_add_ext_sample(pc, regs, event, is_kernel);
 254 }
 255
 256 #ifdef CONFIG_OPROFILE_IBS
 257
 258 #define MAX_IBS_SAMPLE_SIZE     14
 259 static int log_ibs_sample(struct oprofile_cpu_buffer *cpu_buf,
 260         unsigned long pc, int is_kernel, unsigned  int *ibs, int ibs_code)
 261 {
 262         struct task_struct *task;
 263
 264         cpu_buf->sample_received++;
 265
 266         if (nr_available_slots(cpu_buf) < MAX_IBS_SAMPLE_SIZE) {
 267                 cpu_buf->sample_lost_overflow++;
 268                 return 0;
 269         }
 270
 271         is_kernel = !!is_kernel;
 272
 273         /* notice a switch from user->kernel or vice versa */
 274         if (cpu_buf->last_is_kernel != is_kernel) {
 275                 cpu_buf->last_is_kernel = is_kernel;
 276                 add_code(cpu_buf, is_kernel);
 277         }
 278
 279         /* notice a task switch */
 280         if (!is_kernel) {
 281                 task = current;
 282
 283                 if (cpu_buf->last_task != task) {
 284                         cpu_buf->last_task = task;
 285                         add_code(cpu_buf, (unsigned long)task);
 286                 }
 287         }
 288
 289         add_code(cpu_buf, ibs_code);
 290         add_sample(cpu_buf, ibs[0], ibs[1]);
 291         add_sample(cpu_buf, ibs[2], ibs[3]);
 292         add_sample(cpu_buf, ibs[4], ibs[5]);
 293
 294         if (ibs_code == IBS_OP_BEGIN) {
 295         add_sample(cpu_buf, ibs[6], ibs[7]);
 296         add_sample(cpu_buf, ibs[8], ibs[9]);
 297         add_sample(cpu_buf, ibs[10], ibs[11]);
 298         }
 299
 300         return 1;
 301 }
 302
 303 void oprofile_add_ibs_sample(struct pt_regs *const regs,
 304                                 unsigned int * const ibs_sample, u8 code)
 305 {
 306         int is_kernel = !user_mode(regs);
 307         unsigned long pc = profile_pc(regs);
 308
 309         struct oprofile_cpu_buffer *cpu_buf =
 310                          &per_cpu(cpu_buffer, smp_processor_id());
 311
 312         if (!backtrace_depth) {
 313                 log_ibs_sample(cpu_buf, pc, is_kernel, ibs_sample, code);
 314                 return;
 315         }
 316
 317         /* if log_sample() fails we can't backtrace since we lost the source
 318         * of this event */
 319         if (log_ibs_sample(cpu_buf, pc, is_kernel, ibs_sample, code))
 320                 oprofile_ops.backtrace(regs, backtrace_depth);
 321 }
 322
 323 #endif
 324
 325 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 326 {
 327         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 328         log_sample(cpu_buf, pc, is_kernel, event);
 329 }
 330
 331 void oprofile_add_trace(unsigned long pc)
 332 {
 333         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 334
 335         if (!cpu_buf->tracing)
 336                 return;
 337
 338         if (nr_available_slots(cpu_buf) < 1) {
 339                 cpu_buf->tracing = 0;
 340                 cpu_buf->sample_lost_overflow++;
 341                 return;
 342         }
 343
 344         /* broken frame can give an eip with the same value as an escape code,
 345          * abort the trace if we get it */
 346         if (pc == ESCAPE_CODE) {
 347                 cpu_buf->tracing = 0;
 348                 cpu_buf->backtrace_aborted++;
 349                 return;
 350         }
 351
 352         add_sample(cpu_buf, pc, 0);
 353 }
 354
 355 /*
 356  * This serves to avoid cpu buffer overflow, and makes sure
 357  * the task mortuary progresses
 358  *
 359  * By using schedule_delayed_work_on and then schedule_delayed_work
 360  * we guarantee this will stay on the correct cpu
 361  */
 362 static void wq_sync_buffer(struct work_struct *work)
 363 {
 364         struct oprofile_cpu_buffer * b =
 365                 container_of(work, struct oprofile_cpu_buffer, work.work);
 366         if (b->cpu != smp_processor_id()) {
 367                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 368                        smp_processor_id(), b->cpu);
 369         }
 370         sync_buffer(b->cpu);
 371
 372         /* don't re-add the work if we're shutting down */
 373         if (work_enabled)
 374                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 375 }