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_cpu_buffers().
  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/vmalloc.h>
  22 #include <linux/errno.h>
  23
  24 #include "cpu_buffer.h"
  25 #include "oprof.h"
  26
  27 struct oprofile_cpu_buffer cpu_buffer[NR_CPUS] __cacheline_aligned;
  28
  29 static void __free_cpu_buffers(int num)
  30 {
  31         int i;
  32
  33         for (i=0; i < num; ++i) {
  34                 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
  35
  36                 if (!cpu_possible(i))
  37                         continue;
  38
  39                 vfree(b->buffer);
  40         }
  41 }
  42
  43
  44 int alloc_cpu_buffers(void)
  45 {
  46         int i;
  47
  48         unsigned long buffer_size = fs_cpu_buffer_size;
  49
  50         for (i=0; i < NR_CPUS; ++i) {
  51                 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
  52
  53                 if (!cpu_possible(i))
  54                         continue;
  55
  56                 b->buffer = vmalloc(sizeof(struct op_sample) * buffer_size);
  57                 if (!b->buffer)
  58                         goto fail;
  59
  60                 b->last_task = 0;
  61                 b->last_is_kernel = -1;
  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->sample_lost_task_exit = 0;
  68         }
  69         return 0;
  70 fail:
  71         __free_cpu_buffers(i);
  72         return -ENOMEM;
  73 }
  74
  75
  76 void free_cpu_buffers(void)
  77 {
  78         __free_cpu_buffers(NR_CPUS);
  79 }
  80
  81
  82 /* compute number of available slots in cpu_buffer queue */
  83 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
  84 {
  85         unsigned long head = b->head_pos;
  86         unsigned long tail = b->tail_pos;
  87
  88         if (tail > head)
  89                 return tail - head;
  90
  91         return tail + (b->buffer_size - head);
  92 }
  93
  94
  95 static void increment_head(struct oprofile_cpu_buffer * b)
  96 {
  97         unsigned long new_head = b->head_pos + 1;
  98
  99         /* Ensure anything written to the slot before we
 100          * increment is visible */
 101         wmb();
 102
 103         if (new_head < (b->buffer_size))
 104                 b->head_pos = new_head;
 105         else
 106                 b->head_pos = 0;
 107 }
 108
 109
 110 /* This must be safe from any context. It's safe writing here
 111  * because of the head/tail separation of the writer and reader
 112  * of the CPU buffer.
 113  *
 114  * is_kernel is needed because on some architectures you cannot
 115  * tell if you are in kernel or user space simply by looking at
 116  * eip. We tag this in the buffer by generating kernel enter/exit
 117  * events whenever is_kernel changes
 118  */
 119 void oprofile_add_sample(unsigned long eip, unsigned int is_kernel,
 120         unsigned long event, int cpu)
 121 {
 122         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[cpu];
 123         struct task_struct * task;
 124
 125         is_kernel = !!is_kernel;
 126
 127         cpu_buf->sample_received++;
 128
 129
 130         if (nr_available_slots(cpu_buf) < 3) {
 131                 cpu_buf->sample_lost_overflow++;
 132                 return;
 133         }
 134
 135         task = current;
 136
 137         /* notice a switch from user->kernel or vice versa */
 138         if (cpu_buf->last_is_kernel != is_kernel) {
 139                 cpu_buf->last_is_kernel = is_kernel;
 140                 cpu_buf->buffer[cpu_buf->head_pos].eip = ~0UL;
 141                 cpu_buf->buffer[cpu_buf->head_pos].event = is_kernel;
 142                 increment_head(cpu_buf);
 143         }
 144
 145         /* notice a task switch */
 146         if (cpu_buf->last_task != task) {
 147                 cpu_buf->last_task = task;
 148                 if (!(task->flags & PF_EXITING)) {
 149                         cpu_buf->buffer[cpu_buf->head_pos].eip = ~0UL;
 150                         cpu_buf->buffer[cpu_buf->head_pos].event = (unsigned long)task;
 151                         increment_head(cpu_buf);
 152                 }
 153         }
 154
 155         /* If the task is exiting it's not safe to take a sample
 156          * as the task_struct is about to be freed. We can't just
 157          * notify at release_task() time because of CLONE_DETACHED
 158          * tasks that release_task() themselves.
 159          */
 160         if (task->flags & PF_EXITING) {
 161                 cpu_buf->sample_lost_task_exit++;
 162                 return;
 163         }
 164
 165         cpu_buf->buffer[cpu_buf->head_pos].eip = eip;
 166         cpu_buf->buffer[cpu_buf->head_pos].event = event;
 167         increment_head(cpu_buf);
 168 }
 169
 170
 171 /* Resets the cpu buffer to a sane state. */
 172 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
 173 {
 174         /* reset these to invalid values; the next sample
 175          * collected will populate the buffer with proper
 176          * values to initialize the buffer
 177          */
 178         cpu_buf->last_is_kernel = -1;
 179         cpu_buf->last_task = 0;
 180 }