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  *
   9  * Each CPU has a local buffer that stores PC value/event
  10  * pairs. We also log context switches when we notice them.
  11  * Eventually each CPU's buffer is processed into the global
  12  * event buffer by sync_buffer().
  13  *
  14  * We use a local buffer for two reasons: an NMI or similar
  15  * interrupt cannot synchronise, and high sampling rates
  16  * would lead to catastrophic global synchronisation if
  17  * a global buffer was used.
  18  */
  19
  20 #include <linux/sched.h>
  21 #include <linux/oprofile.h>
  22 #include <linux/vmalloc.h>
  23 #include <linux/errno.h>
  24
  25 #include "event_buffer.h"
  26 #include "cpu_buffer.h"
  27 #include "buffer_sync.h"
  28 #include "oprof.h"
  29
  30 struct oprofile_cpu_buffer cpu_buffer[NR_CPUS] __cacheline_aligned;
  31
  32 static void wq_sync_buffer(void *);
  33
  34 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
  35 static int work_enabled;
  36
  37 void free_cpu_buffers(void)
  38 {
  39         int i;
  40
  41         for_each_online_cpu(i) {
  42                 vfree(cpu_buffer[i].buffer);
  43         }
  44 }
  45
  46
  47 int alloc_cpu_buffers(void)
  48 {
  49         int i;
  50
  51         unsigned long buffer_size = fs_cpu_buffer_size;
  52
  53         for_each_online_cpu(i) {
  54                 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
  55
  56                 b->buffer = vmalloc(sizeof(struct op_sample) * buffer_size);
  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->cpu = i;
  69                 INIT_WORK(&b->work, wq_sync_buffer, b);
  70         }
  71         return 0;
  72
  73 fail:
  74         free_cpu_buffers();
  75         return -ENOMEM;
  76 }
  77
  78
  79 void start_cpu_work(void)
  80 {
  81         int i;
  82
  83         work_enabled = 1;
  84
  85         for_each_online_cpu(i) {
  86                 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
  87
  88                 /*
  89                  * Spread the work by 1 jiffy per cpu so they dont all
  90                  * fire at once.
  91                  */
  92                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
  93         }
  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 = &cpu_buffer[i];
 105
 106                 cancel_delayed_work(&b->work);
 107         }
 108
 109         flush_scheduled_work();
 110 }
 111
 112
 113 /* Resets the cpu buffer to a sane state. */
 114 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
 115 {
 116         /* reset these to invalid values; the next sample
 117          * collected will populate the buffer with proper
 118          * values to initialize the buffer
 119          */
 120         cpu_buf->last_is_kernel = -1;
 121         cpu_buf->last_task = NULL;
 122 }
 123
 124
 125 /* compute number of available slots in cpu_buffer queue */
 126 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
 127 {
 128         unsigned long head = b->head_pos;
 129         unsigned long tail = b->tail_pos;
 130
 131         if (tail > head)
 132                 return (tail - head) - 1;
 133
 134         return tail + (b->buffer_size - head) - 1;
 135 }
 136
 137
 138 static void increment_head(struct oprofile_cpu_buffer * b)
 139 {
 140         unsigned long new_head = b->head_pos + 1;
 141
 142         /* Ensure anything written to the slot before we
 143          * increment is visible */
 144         wmb();
 145
 146         if (new_head < b->buffer_size)
 147                 b->head_pos = new_head;
 148         else
 149                 b->head_pos = 0;
 150 }
 151
 152
 153
 154
 155 inline static void
 156 add_sample(struct oprofile_cpu_buffer * cpu_buf,
 157            unsigned long pc, unsigned long event)
 158 {
 159         struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
 160         entry->eip = pc;
 161         entry->event = event;
 162         increment_head(cpu_buf);
 163 }
 164
 165
 166 inline static void
 167 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
 168 {
 169         add_sample(buffer, ESCAPE_CODE, value);
 170 }
 171
 172
 173 /* This must be safe from any context. It's safe writing here
 174  * because of the head/tail separation of the writer and reader
 175  * of the CPU buffer.
 176  *
 177  * is_kernel is needed because on some architectures you cannot
 178  * tell if you are in kernel or user space simply by looking at
 179  * pc. We tag this in the buffer by generating kernel enter/exit
 180  * events whenever is_kernel changes
 181  */
 182 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
 183                       int is_kernel, unsigned long event)
 184 {
 185         struct task_struct * task;
 186
 187         cpu_buf->sample_received++;
 188
 189         if (nr_available_slots(cpu_buf) < 3) {
 190                 cpu_buf->sample_lost_overflow++;
 191                 return 0;
 192         }
 193
 194         is_kernel = !!is_kernel;
 195
 196         task = current;
 197
 198         /* notice a switch from user->kernel or vice versa */
 199         if (cpu_buf->last_is_kernel != is_kernel) {
 200                 cpu_buf->last_is_kernel = is_kernel;
 201                 add_code(cpu_buf, is_kernel);
 202         }
 203
 204         /* notice a task switch */
 205         if (cpu_buf->last_task != task) {
 206                 cpu_buf->last_task = task;
 207                 add_code(cpu_buf, (unsigned long)task);
 208         }
 209
 210         add_sample(cpu_buf, pc, event);
 211         return 1;
 212 }
 213
 214 static int oprofile_begin_trace(struct oprofile_cpu_buffer * cpu_buf)
 215 {
 216         if (nr_available_slots(cpu_buf) < 4) {
 217                 cpu_buf->sample_lost_overflow++;
 218                 return 0;
 219         }
 220
 221         add_code(cpu_buf, CPU_TRACE_BEGIN);
 222         cpu_buf->tracing = 1;
 223         return 1;
 224 }
 225
 226
 227 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
 228 {
 229         cpu_buf->tracing = 0;
 230 }
 231
 232
 233 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 234 {
 235         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
 236         unsigned long pc = profile_pc(regs);
 237         int is_kernel = !user_mode(regs);
 238
 239         if (!backtrace_depth) {
 240                 log_sample(cpu_buf, pc, is_kernel, event);
 241                 return;
 242         }
 243
 244         if (!oprofile_begin_trace(cpu_buf))
 245                 return;
 246
 247         /* if log_sample() fail we can't backtrace since we lost the source
 248          * of this event */
 249         if (log_sample(cpu_buf, pc, is_kernel, event))
 250                 oprofile_ops.backtrace(regs, backtrace_depth);
 251         oprofile_end_trace(cpu_buf);
 252 }
 253
 254
 255 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 256 {
 257         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
 258         log_sample(cpu_buf, pc, is_kernel, event);
 259 }
 260
 261
 262 void oprofile_add_trace(unsigned long pc)
 263 {
 264         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
 265
 266         if (!cpu_buf->tracing)
 267                 return;
 268
 269         if (nr_available_slots(cpu_buf) < 1) {
 270                 cpu_buf->tracing = 0;
 271                 cpu_buf->sample_lost_overflow++;
 272                 return;
 273         }
 274
 275         /* broken frame can give an eip with the same value as an escape code,
 276          * abort the trace if we get it */
 277         if (pc == ESCAPE_CODE) {
 278                 cpu_buf->tracing = 0;
 279                 cpu_buf->backtrace_aborted++;
 280                 return;
 281         }
 282
 283         add_sample(cpu_buf, pc, 0);
 284 }
 285
 286
 287
 288 /*
 289  * This serves to avoid cpu buffer overflow, and makes sure
 290  * the task mortuary progresses
 291  *
 292  * By using schedule_delayed_work_on and then schedule_delayed_work
 293  * we guarantee this will stay on the correct cpu
 294  */
 295 static void wq_sync_buffer(void * data)
 296 {
 297         struct oprofile_cpu_buffer * b = data;
 298         if (b->cpu != smp_processor_id()) {
 299                 printk("WQ on CPU%d, prefer CPU%d\n",
 300                        smp_processor_id(), b->cpu);
 301         }
 302         sync_buffer(b->cpu);
 303
 304         /* don't re-add the work if we're shutting down */
 305         if (work_enabled)
 306                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 307 }