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_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_possible_cpu(i) {
  43                 vfree(per_cpu(cpu_buffer, i).buffer);
  44                 per_cpu(cpu_buffer, i).buffer = NULL;
  45         }
  46 }
  47
  48 unsigned long oprofile_get_cpu_buffer_size(void)
  49 {
  50         return fs_cpu_buffer_size;
  51 }
  52
  53 void oprofile_cpu_buffer_inc_smpl_lost(void)
  54 {
  55         struct oprofile_cpu_buffer *cpu_buf
  56                 = &__get_cpu_var(cpu_buffer);
  57
  58         cpu_buf->sample_lost_overflow++;
  59 }
  60
  61 int alloc_cpu_buffers(void)
  62 {
  63         int i;
  64
  65         unsigned long buffer_size = fs_cpu_buffer_size;
  66
  67         for_each_possible_cpu(i) {
  68                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
  69
  70                 b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
  71                         cpu_to_node(i));
  72                 if (!b->buffer)
  73                         goto fail;
  74
  75                 b->last_task = NULL;
  76                 b->last_is_kernel = -1;
  77                 b->tracing = 0;
  78                 b->buffer_size = buffer_size;
  79                 b->tail_pos = 0;
  80                 b->head_pos = 0;
  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(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(cpu_buffer, i);
 120
 121                 cancel_delayed_work(&b->work);
 122         }
 123
 124         flush_scheduled_work();
 125 }
 126
 127 /* Resets the cpu buffer to a sane state. */
 128 void cpu_buffer_reset(struct oprofile_cpu_buffer *cpu_buf)
 129 {
 130         /*
 131          * reset these to invalid values; the next sample collected
 132          * will populate the buffer with proper values to initialize
 133          * the buffer
 134          */
 135         cpu_buf->last_is_kernel = -1;
 136         cpu_buf->last_task = NULL;
 137 }
 138
 139 /* compute number of available slots in cpu_buffer queue */
 140 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const *b)
 141 {
 142         unsigned long head = b->head_pos;
 143         unsigned long tail = b->tail_pos;
 144
 145         if (tail > head)
 146                 return (tail - head) - 1;
 147
 148         return tail + (b->buffer_size - head) - 1;
 149 }
 150
 151 static void increment_head(struct oprofile_cpu_buffer *b)
 152 {
 153         unsigned long new_head = b->head_pos + 1;
 154
 155         /*
 156          * Ensure anything written to the slot before we increment is
 157          * visible
 158          */
 159         wmb();
 160
 161         if (new_head < b->buffer_size)
 162                 b->head_pos = new_head;
 163         else
 164                 b->head_pos = 0;
 165 }
 166
 167 static inline void
 168 add_sample(struct oprofile_cpu_buffer *cpu_buf,
 169            unsigned long pc, unsigned long event)
 170 {
 171         struct op_sample *entry = &cpu_buf->buffer[cpu_buf->head_pos];
 172         entry->eip = pc;
 173         entry->event = event;
 174         increment_head(cpu_buf);
 175 }
 176
 177 static inline void
 178 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
 179 {
 180         add_sample(buffer, ESCAPE_CODE, value);
 181 }
 182
 183 /* This must be safe from any context. It's safe writing here
 184  * because of the head/tail separation of the writer and reader
 185  * of the CPU buffer.
 186  *
 187  * is_kernel is needed because on some architectures you cannot
 188  * tell if you are in kernel or user space simply by looking at
 189  * pc. We tag this in the buffer by generating kernel enter/exit
 190  * events whenever is_kernel changes
 191  */
 192 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 193                       int is_kernel, unsigned long event)
 194 {
 195         struct task_struct *task;
 196
 197         cpu_buf->sample_received++;
 198
 199         if (pc == ESCAPE_CODE) {
 200                 cpu_buf->sample_invalid_eip++;
 201                 return 0;
 202         }
 203
 204         if (nr_available_slots(cpu_buf) < 3) {
 205                 cpu_buf->sample_lost_overflow++;
 206                 return 0;
 207         }
 208
 209         is_kernel = !!is_kernel;
 210
 211         task = current;
 212
 213         /* notice a switch from user->kernel or vice versa */
 214         if (cpu_buf->last_is_kernel != is_kernel) {
 215                 cpu_buf->last_is_kernel = is_kernel;
 216                 add_code(cpu_buf, is_kernel);
 217         }
 218
 219         /* notice a task switch */
 220         if (cpu_buf->last_task != task) {
 221                 cpu_buf->last_task = task;
 222                 add_code(cpu_buf, (unsigned long)task);
 223         }
 224
 225         add_sample(cpu_buf, pc, event);
 226         return 1;
 227 }
 228
 229 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 230 {
 231         if (nr_available_slots(cpu_buf) < 4) {
 232                 cpu_buf->sample_lost_overflow++;
 233                 return 0;
 234         }
 235
 236         add_code(cpu_buf, CPU_TRACE_BEGIN);
 237         cpu_buf->tracing = 1;
 238         return 1;
 239 }
 240
 241 static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 242 {
 243         cpu_buf->tracing = 0;
 244 }
 245
 246 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 247                                 unsigned long event, int is_kernel)
 248 {
 249         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 250
 251         if (!backtrace_depth) {
 252                 log_sample(cpu_buf, pc, is_kernel, event);
 253                 return;
 254         }
 255
 256         if (!oprofile_begin_trace(cpu_buf))
 257                 return;
 258
 259         /*
 260          * if log_sample() fail we can't backtrace since we lost the
 261          * source of this event
 262          */
 263         if (log_sample(cpu_buf, pc, is_kernel, event))
 264                 oprofile_ops.backtrace(regs, backtrace_depth);
 265         oprofile_end_trace(cpu_buf);
 266 }
 267
 268 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 269 {
 270         int is_kernel = !user_mode(regs);
 271         unsigned long pc = profile_pc(regs);
 272
 273         oprofile_add_ext_sample(pc, regs, event, is_kernel);
 274 }
 275
 276 #ifdef CONFIG_OPROFILE_IBS
 277
 278 #define MAX_IBS_SAMPLE_SIZE 14
 279
 280 void oprofile_add_ibs_sample(struct pt_regs * const regs,
 281                              unsigned int * const ibs_sample, int ibs_code)
 282 {
 283         int is_kernel = !user_mode(regs);
 284         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 285         struct task_struct *task;
 286
 287         cpu_buf->sample_received++;
 288
 289         if (nr_available_slots(cpu_buf) < MAX_IBS_SAMPLE_SIZE) {
 290                 /* we can't backtrace since we lost the source of this event */
 291                 cpu_buf->sample_lost_overflow++;
 292                 return;
 293         }
 294
 295         /* notice a switch from user->kernel or vice versa */
 296         if (cpu_buf->last_is_kernel != is_kernel) {
 297                 cpu_buf->last_is_kernel = is_kernel;
 298                 add_code(cpu_buf, is_kernel);
 299         }
 300
 301         /* notice a task switch */
 302         if (!is_kernel) {
 303                 task = current;
 304                 if (cpu_buf->last_task != task) {
 305                         cpu_buf->last_task = task;
 306                         add_code(cpu_buf, (unsigned long)task);
 307                 }
 308         }
 309
 310         add_code(cpu_buf, ibs_code);
 311         add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
 312         add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
 313         add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
 314
 315         if (ibs_code == IBS_OP_BEGIN) {
 316                 add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
 317                 add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
 318                 add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
 319         }
 320
 321         if (backtrace_depth)
 322                 oprofile_ops.backtrace(regs, backtrace_depth);
 323 }
 324
 325 #endif
 326
 327 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 328 {
 329         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 330         log_sample(cpu_buf, pc, is_kernel, event);
 331 }
 332
 333 void oprofile_add_trace(unsigned long pc)
 334 {
 335         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 336
 337         if (!cpu_buf->tracing)
 338                 return;
 339
 340         if (nr_available_slots(cpu_buf) < 1) {
 341                 cpu_buf->tracing = 0;
 342                 cpu_buf->sample_lost_overflow++;
 343                 return;
 344         }
 345
 346         /*
 347          * broken frame can give an eip with the same value as an
 348          * escape code, abort the trace if we get it
 349          */
 350         if (pc == ESCAPE_CODE) {
 351                 cpu_buf->tracing = 0;
 352                 cpu_buf->backtrace_aborted++;
 353                 return;
 354         }
 355
 356         add_sample(cpu_buf, pc, 0);
 357 }
 358
 359 /*
 360  * This serves to avoid cpu buffer overflow, and makes sure
 361  * the task mortuary progresses
 362  *
 363  * By using schedule_delayed_work_on and then schedule_delayed_work
 364  * we guarantee this will stay on the correct cpu
 365  */
 366 static void wq_sync_buffer(struct work_struct *work)
 367 {
 368         struct oprofile_cpu_buffer *b =
 369                 container_of(work, struct oprofile_cpu_buffer, work.work);
 370         if (b->cpu != smp_processor_id()) {
 371                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 372                        smp_processor_id(), b->cpu);
 373
 374                 if (!cpu_online(b->cpu)) {
 375                         cancel_delayed_work(&b->work);
 376                         return;
 377                 }
 378         }
 379         sync_buffer(b->cpu);
 380
 381         /* don't re-add the work if we're shutting down */
 382         if (work_enabled)
 383                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 384 }