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