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 /* compute number of available slots in cpu_buffer queue */
 128 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const *b)
 129 {
 130         unsigned long head = b->head_pos;
 131         unsigned long tail = b->tail_pos;
 132
 133         if (tail > head)
 134                 return (tail - head) - 1;
 135
 136         return tail + (b->buffer_size - head) - 1;
 137 }
 138
 139 static inline void
 140 add_sample(struct oprofile_cpu_buffer *cpu_buf,
 141            unsigned long pc, unsigned long event)
 142 {
 143         struct op_sample *entry = cpu_buffer_write_entry(cpu_buf);
 144         entry->eip = pc;
 145         entry->event = event;
 146         cpu_buffer_write_commit(cpu_buf);
 147 }
 148
 149 static inline void
 150 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
 151 {
 152         add_sample(buffer, ESCAPE_CODE, value);
 153 }
 154
 155 /* This must be safe from any context. It's safe writing here
 156  * because of the head/tail separation of the writer and reader
 157  * of the CPU buffer.
 158  *
 159  * is_kernel is needed because on some architectures you cannot
 160  * tell if you are in kernel or user space simply by looking at
 161  * pc. We tag this in the buffer by generating kernel enter/exit
 162  * events whenever is_kernel changes
 163  */
 164 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 165                       int is_kernel, unsigned long event)
 166 {
 167         struct task_struct *task;
 168
 169         cpu_buf->sample_received++;
 170
 171         if (pc == ESCAPE_CODE) {
 172                 cpu_buf->sample_invalid_eip++;
 173                 return 0;
 174         }
 175
 176         if (nr_available_slots(cpu_buf) < 3) {
 177                 cpu_buf->sample_lost_overflow++;
 178                 return 0;
 179         }
 180
 181         is_kernel = !!is_kernel;
 182
 183         task = current;
 184
 185         /* notice a switch from user->kernel or vice versa */
 186         if (cpu_buf->last_is_kernel != is_kernel) {
 187                 cpu_buf->last_is_kernel = is_kernel;
 188                 add_code(cpu_buf, is_kernel);
 189         }
 190
 191         /* notice a task switch */
 192         if (cpu_buf->last_task != task) {
 193                 cpu_buf->last_task = task;
 194                 add_code(cpu_buf, (unsigned long)task);
 195         }
 196
 197         add_sample(cpu_buf, pc, event);
 198         return 1;
 199 }
 200
 201 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 202 {
 203         if (nr_available_slots(cpu_buf) < 4) {
 204                 cpu_buf->sample_lost_overflow++;
 205                 return 0;
 206         }
 207
 208         add_code(cpu_buf, CPU_TRACE_BEGIN);
 209         cpu_buf->tracing = 1;
 210         return 1;
 211 }
 212
 213 static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 214 {
 215         cpu_buf->tracing = 0;
 216 }
 217
 218 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
 219                                 unsigned long event, int is_kernel)
 220 {
 221         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 222
 223         if (!backtrace_depth) {
 224                 log_sample(cpu_buf, pc, is_kernel, event);
 225                 return;
 226         }
 227
 228         if (!oprofile_begin_trace(cpu_buf))
 229                 return;
 230
 231         /*
 232          * if log_sample() fail we can't backtrace since we lost the
 233          * source of this event
 234          */
 235         if (log_sample(cpu_buf, pc, is_kernel, event))
 236                 oprofile_ops.backtrace(regs, backtrace_depth);
 237         oprofile_end_trace(cpu_buf);
 238 }
 239
 240 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 241 {
 242         int is_kernel = !user_mode(regs);
 243         unsigned long pc = profile_pc(regs);
 244
 245         oprofile_add_ext_sample(pc, regs, event, is_kernel);
 246 }
 247
 248 #ifdef CONFIG_OPROFILE_IBS
 249
 250 #define MAX_IBS_SAMPLE_SIZE 14
 251
 252 void oprofile_add_ibs_sample(struct pt_regs * const regs,
 253                              unsigned int * const ibs_sample, int ibs_code)
 254 {
 255         int is_kernel = !user_mode(regs);
 256         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 257         struct task_struct *task;
 258
 259         cpu_buf->sample_received++;
 260
 261         if (nr_available_slots(cpu_buf) < MAX_IBS_SAMPLE_SIZE) {
 262                 /* we can't backtrace since we lost the source of this event */
 263                 cpu_buf->sample_lost_overflow++;
 264                 return;
 265         }
 266
 267         /* notice a switch from user->kernel or vice versa */
 268         if (cpu_buf->last_is_kernel != is_kernel) {
 269                 cpu_buf->last_is_kernel = is_kernel;
 270                 add_code(cpu_buf, is_kernel);
 271         }
 272
 273         /* notice a task switch */
 274         if (!is_kernel) {
 275                 task = current;
 276                 if (cpu_buf->last_task != task) {
 277                         cpu_buf->last_task = task;
 278                         add_code(cpu_buf, (unsigned long)task);
 279                 }
 280         }
 281
 282         add_code(cpu_buf, ibs_code);
 283         add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
 284         add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
 285         add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
 286
 287         if (ibs_code == IBS_OP_BEGIN) {
 288                 add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
 289                 add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
 290                 add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
 291         }
 292
 293         if (backtrace_depth)
 294                 oprofile_ops.backtrace(regs, backtrace_depth);
 295 }
 296
 297 #endif
 298
 299 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 300 {
 301         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 302         log_sample(cpu_buf, pc, is_kernel, event);
 303 }
 304
 305 void oprofile_add_trace(unsigned long pc)
 306 {
 307         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
 308
 309         if (!cpu_buf->tracing)
 310                 return;
 311
 312         if (nr_available_slots(cpu_buf) < 1) {
 313                 cpu_buf->tracing = 0;
 314                 cpu_buf->sample_lost_overflow++;
 315                 return;
 316         }
 317
 318         /*
 319          * broken frame can give an eip with the same value as an
 320          * escape code, abort the trace if we get it
 321          */
 322         if (pc == ESCAPE_CODE) {
 323                 cpu_buf->tracing = 0;
 324                 cpu_buf->backtrace_aborted++;
 325                 return;
 326         }
 327
 328         add_sample(cpu_buf, pc, 0);
 329 }
 330
 331 /*
 332  * This serves to avoid cpu buffer overflow, and makes sure
 333  * the task mortuary progresses
 334  *
 335  * By using schedule_delayed_work_on and then schedule_delayed_work
 336  * we guarantee this will stay on the correct cpu
 337  */
 338 static void wq_sync_buffer(struct work_struct *work)
 339 {
 340         struct oprofile_cpu_buffer *b =
 341                 container_of(work, struct oprofile_cpu_buffer, work.work);
 342         if (b->cpu != smp_processor_id()) {
 343                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
 344                        smp_processor_id(), b->cpu);
 345
 346                 if (!cpu_online(b->cpu)) {
 347                         cancel_delayed_work(&b->work);
 348                         return;
 349                 }
 350         }
 351         sync_buffer(b->cpu);
 352
 353         /* don't re-add the work if we're shutting down */
 354         if (work_enabled)
 355                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
 356 }