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 DEFINE_PER_CPU_SHARED_ALIGNED(struct oprofile_cpu_buffer, cpu_buffer);
  31
  32 static void wq_sync_buffer(struct work_struct *work);
  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(per_cpu(cpu_buffer, i).buffer);
  43 }
  44
  45 int alloc_cpu_buffers(void)
  46 {
  47         int i;
  48
  49         unsigned long buffer_size = fs_cpu_buffer_size;
  50
  51         for_each_online_cpu(i) {
  52                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
  53
  54                 b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
  55                         cpu_to_node(i));
  56                 if (!b->buffer)
  57                         goto fail;
  58
  59                 b->last_task = NULL;
  60                 b->last_is_kernel = -1;
  61                 b->tracing = 0;
  62                 b->buffer_size = buffer_size;
  63                 b->tail_pos = 0;
  64                 b->head_pos = 0;
  65                 b->sample_received = 0;
  66                 b->sample_lost_overflow = 0;
  67                 b->backtrace_aborted = 0;
  68                 b->sample_invalid_eip = 0;
  69                 b->cpu = i;
  70                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
  71         }
  72         return 0;
  73
  74 fail:
  75         free_cpu_buffers();
  76         return -ENOMEM;
  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 = &per_cpu(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 void end_cpu_work(void)
  97 {
  98         int i;
  99
 100         work_enabled = 0;
 101
 102         for_each_online_cpu(i) {
 103                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
 104
 105                 cancel_delayed_work(&b->work);
 106         }
 107
 108         flush_scheduled_work();
 109 }
 110
 111 /* Resets the cpu buffer to a sane state. */
 112 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
 113 {
 114         /* reset these to invalid values; the next sample
 115          * collected will populate the buffer with proper
 116          * values to initialize the buffer
 117          */
 118         cpu_buf->last_is_kernel = -1;
 119         cpu_buf->last_task = NULL;
 120 }
 121
 122 /* compute number of available slots in cpu_buffer queue */
 123 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
 124 {
 125         unsigned long head = b->head_pos;
 126         unsigned long tail = b->tail_pos;
 127
 128         if (tail > head)
 129                 return (tail - head) - 1;
 130
 131         return tail + (b->buffer_size - head) - 1;
 132 }
 133
 134 static void increment_head(struct oprofile_cpu_buffer * b)
 135 {
 136         unsigned long new_head = b->head_pos + 1;
 137
 138         /* Ensure anything written to the slot before we
 139          * increment is visible */
 140         wmb();
 141
 142         if (new_head < b->buffer_size)
 143                 b->head_pos = new_head;
 144         else
 145                 b->head_pos = 0;
 146 }
 147
 148 static inline void
 149 add_sample(struct oprofile_cpu_buffer * cpu_buf,
 150            unsigned long pc, unsigned long event)
 151 {
 152         struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
 153         entry->eip = pc;
 154         entry->event = event;
 155         increment_head(cpu_buf);
 156 }
 157
 158 static inline void
 159 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
 160 {
 161         add_sample(buffer, ESCAPE_CODE, value);
 162 }
 163
 164 /* This must be safe from any context. It's safe writing here
 165  * because of the head/tail separation of the writer and reader
 166  * of the CPU buffer.
 167  *
 168  * is_kernel is needed because on some architectures you cannot
 169  * tell if you are in kernel or user space simply by looking at
 170  * pc. We tag this in the buffer by generating kernel enter/exit
 171  * events whenever is_kernel changes
 172  */
 173 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
 174                       int is_kernel, unsigned long event)
 175 {
 176         struct task_struct * task;
 177
 178         cpu_buf->sample_received++;
 179
 180         if (pc == ESCAPE_CODE) {
 181                 cpu_buf->sample_invalid_eip++;
 182                 return 0;
 183         }
 184
 185         if (nr_available_slots(cpu_buf) < 3) {
 186                 cpu_buf->sample_lost_overflow++;
 187                 return 0;
 188         }
 189
 190         is_kernel = !!is_kernel;
 191
 192         task = current;
 193
 194         /* notice a switch from user->kernel or vice versa */
 195         if (cpu_buf->last_is_kernel != is_kernel) {
 196                 cpu_buf->last_is_kernel = is_kernel;
 197                 add_code(cpu_buf, is_kernel);
 198         }
 199
 200         /* notice a task switch */
 201         if (cpu_buf->last_task != task) {
 202                 cpu_buf->last_task = task;
 203                 add_code(cpu_buf, (unsigned long)task);
 204         }
 205
 206         add_sample(cpu_buf, pc, event);
 207         return 1;
 208 }
 209
 210 static int oprofile_begin_trace(struct oprofile_cpu_buffer * cpu_buf)
 211 {
 212         if (nr_available_slots(cpu_buf) < 4) {
 213                 cpu_buf->sample_lost_overflow++;
 214                 return 0;
 215         }
 216
 217         add_code(cpu_buf, CPU_TRACE_BEGIN);
 218         cpu_buf->tracing = 1;
 219         return 1;
 220 }
 221
 222 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
 223 {
 224         cpu_buf->tracing = 0;
 225 }
 226
 227 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 228                                 unsigned long event, int is_kernel)
 229 {
 230         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 231
 232         if (!backtrace_depth) {
 233                 log_sample(cpu_buf, pc, is_kernel, event);
 234                 return;
 235         }
 236
 237         if (!oprofile_begin_trace(cpu_buf))
 238                 return;
 239
 240         /* if log_sample() fail we can't backtrace since we lost the source
 241          * of this event */
 242         if (log_sample(cpu_buf, pc, is_kernel, event))
 243                 oprofile_ops.backtrace(regs, backtrace_depth);
 244         oprofile_end_trace(cpu_buf);
 245 }
 246
 247 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 248 {
 249         int is_kernel = !user_mode(regs);
 250         unsigned long pc = profile_pc(regs);
 251
 252         oprofile_add_ext_sample(pc, regs, event, is_kernel);
 253 }
 254
 255 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 256 {
 257         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 258         log_sample(cpu_buf, pc, is_kernel, event);
 259 }
 260
 261 void oprofile_add_trace(unsigned long pc)
 262 {
 263         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 264
 265         if (!cpu_buf->tracing)
 266                 return;
 267
 268         if (nr_available_slots(cpu_buf) < 1) {
 269                 cpu_buf->tracing = 0;
 270                 cpu_buf->sample_lost_overflow++;
 271                 return;
 272         }
 273
 274         /* broken frame can give an eip with the same value as an escape code,
 275          * abort the trace if we get it */
 276         if (pc == ESCAPE_CODE) {
 277                 cpu_buf->tracing = 0;
 278                 cpu_buf->backtrace_aborted++;
 279                 return;
 280         }
 281
 282         add_sample(cpu_buf, pc, 0);
 283 }
 284
 285 /*
 286  * This serves to avoid cpu buffer overflow, and makes sure
 287  * the task mortuary progresses
 288  *
 289  * By using schedule_delayed_work_on and then schedule_delayed_work
 290  * we guarantee this will stay on the correct cpu
 291  */
 292 static void wq_sync_buffer(struct work_struct *work)
 293 {
 294         struct oprofile_cpu_buffer * b =
 295                 container_of(work, struct oprofile_cpu_buffer, work.work);
 296         if (b->cpu != smp_processor_id()) {
 297                 printk("WQ on CPU%d, prefer CPU%d\n",
 298                        smp_processor_id(), b->cpu);
 299         }
 300         sync_buffer(b->cpu);
 301
 302         /* don't re-add the work if we're shutting down */
 303         if (work_enabled)
 304                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 305 }