linux/drivers/oprofile/cpu_buffer.c
Robert Richter 9966718dae oprofile: remove ring buffer inline functions in cpu_buffer.h
This patch moves ring buffer inline functions to cpu_buffer.c.

Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Robert Richter <robert.richter@amd.com>
2008-12-29 15:19:19 +01:00

423 lines
9.9 KiB
C

/**
* @file cpu_buffer.c
*
* @remark Copyright 2002 OProfile authors
* @remark Read the file COPYING
*
* @author John Levon <levon@movementarian.org>
* @author Barry Kasindorf <barry.kasindorf@amd.com>
*
* Each CPU has a local buffer that stores PC value/event
* pairs. We also log context switches when we notice them.
* Eventually each CPU's buffer is processed into the global
* event buffer by sync_buffer().
*
* We use a local buffer for two reasons: an NMI or similar
* interrupt cannot synchronise, and high sampling rates
* would lead to catastrophic global synchronisation if
* a global buffer was used.
*/
#include <linux/sched.h>
#include <linux/oprofile.h>
#include <linux/vmalloc.h>
#include <linux/errno.h>
#include "event_buffer.h"
#include "cpu_buffer.h"
#include "buffer_sync.h"
#include "oprof.h"
#define OP_BUFFER_FLAGS 0
/*
* Read and write access is using spin locking. Thus, writing to the
* buffer by NMI handler (x86) could occur also during critical
* sections when reading the buffer. To avoid this, there are 2
* buffers for independent read and write access. Read access is in
* process context only, write access only in the NMI handler. If the
* read buffer runs empty, both buffers are swapped atomically. There
* is potentially a small window during swapping where the buffers are
* disabled and samples could be lost.
*
* Using 2 buffers is a little bit overhead, but the solution is clear
* and does not require changes in the ring buffer implementation. It
* can be changed to a single buffer solution when the ring buffer
* access is implemented as non-locking atomic code.
*/
static struct ring_buffer *op_ring_buffer_read;
static struct ring_buffer *op_ring_buffer_write;
DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
static void wq_sync_buffer(struct work_struct *work);
#define DEFAULT_TIMER_EXPIRE (HZ / 10)
static int work_enabled;
void free_cpu_buffers(void)
{
if (op_ring_buffer_read)
ring_buffer_free(op_ring_buffer_read);
op_ring_buffer_read = NULL;
if (op_ring_buffer_write)
ring_buffer_free(op_ring_buffer_write);
op_ring_buffer_write = NULL;
}
unsigned long oprofile_get_cpu_buffer_size(void)
{
return oprofile_cpu_buffer_size;
}
void oprofile_cpu_buffer_inc_smpl_lost(void)
{
struct oprofile_cpu_buffer *cpu_buf
= &__get_cpu_var(cpu_buffer);
cpu_buf->sample_lost_overflow++;
}
int alloc_cpu_buffers(void)
{
int i;
unsigned long buffer_size = oprofile_cpu_buffer_size;
op_ring_buffer_read = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
if (!op_ring_buffer_read)
goto fail;
op_ring_buffer_write = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
if (!op_ring_buffer_write)
goto fail;
for_each_possible_cpu(i) {
struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
b->last_task = NULL;
b->last_is_kernel = -1;
b->tracing = 0;
b->buffer_size = buffer_size;
b->tail_pos = 0;
b->head_pos = 0;
b->sample_received = 0;
b->sample_lost_overflow = 0;
b->backtrace_aborted = 0;
b->sample_invalid_eip = 0;
b->cpu = i;
INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
}
return 0;
fail:
free_cpu_buffers();
return -ENOMEM;
}
void start_cpu_work(void)
{
int i;
work_enabled = 1;
for_each_online_cpu(i) {
struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
/*
* Spread the work by 1 jiffy per cpu so they dont all
* fire at once.
*/
schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
}
}
void end_cpu_work(void)
{
int i;
work_enabled = 0;
for_each_online_cpu(i) {
struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
cancel_delayed_work(&b->work);
}
flush_scheduled_work();
}
int op_cpu_buffer_write_entry(struct op_entry *entry)
{
entry->event = ring_buffer_lock_reserve(op_ring_buffer_write,
sizeof(struct op_sample),
&entry->irq_flags);
if (entry->event)
entry->sample = ring_buffer_event_data(entry->event);
else
entry->sample = NULL;
if (!entry->sample)
return -ENOMEM;
return 0;
}
int op_cpu_buffer_write_commit(struct op_entry *entry)
{
return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event,
entry->irq_flags);
}
struct op_sample *op_cpu_buffer_read_entry(int cpu)
{
struct ring_buffer_event *e;
e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
if (e)
return ring_buffer_event_data(e);
if (ring_buffer_swap_cpu(op_ring_buffer_read,
op_ring_buffer_write,
cpu))
return NULL;
e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
if (e)
return ring_buffer_event_data(e);
return NULL;
}
unsigned long op_cpu_buffer_entries(int cpu)
{
return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
+ ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
}
static inline int
add_sample(struct oprofile_cpu_buffer *cpu_buf,
unsigned long pc, unsigned long event)
{
struct op_entry entry;
int ret;
ret = op_cpu_buffer_write_entry(&entry);
if (ret)
return ret;
entry.sample->eip = pc;
entry.sample->event = event;
ret = op_cpu_buffer_write_commit(&entry);
if (ret)
return ret;
return 0;
}
static inline int
add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
{
return add_sample(buffer, ESCAPE_CODE, value);
}
/* This must be safe from any context. It's safe writing here
* because of the head/tail separation of the writer and reader
* of the CPU buffer.
*
* is_kernel is needed because on some architectures you cannot
* tell if you are in kernel or user space simply by looking at
* pc. We tag this in the buffer by generating kernel enter/exit
* events whenever is_kernel changes
*/
static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
int is_kernel, unsigned long event)
{
struct task_struct *task;
cpu_buf->sample_received++;
if (pc == ESCAPE_CODE) {
cpu_buf->sample_invalid_eip++;
return 0;
}
is_kernel = !!is_kernel;
task = current;
/* notice a switch from user->kernel or vice versa */
if (cpu_buf->last_is_kernel != is_kernel) {
cpu_buf->last_is_kernel = is_kernel;
if (add_code(cpu_buf, is_kernel))
goto fail;
}
/* notice a task switch */
if (cpu_buf->last_task != task) {
cpu_buf->last_task = task;
if (add_code(cpu_buf, (unsigned long)task))
goto fail;
}
if (add_sample(cpu_buf, pc, event))
goto fail;
return 1;
fail:
cpu_buf->sample_lost_overflow++;
return 0;
}
static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
{
add_code(cpu_buf, CPU_TRACE_BEGIN);
cpu_buf->tracing = 1;
return 1;
}
static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
{
cpu_buf->tracing = 0;
}
void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
unsigned long event, int is_kernel)
{
struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
if (!oprofile_backtrace_depth) {
log_sample(cpu_buf, pc, is_kernel, event);
return;
}
if (!oprofile_begin_trace(cpu_buf))
return;
/*
* if log_sample() fail we can't backtrace since we lost the
* source of this event
*/
if (log_sample(cpu_buf, pc, is_kernel, event))
oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
oprofile_end_trace(cpu_buf);
}
void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
{
int is_kernel = !user_mode(regs);
unsigned long pc = profile_pc(regs);
oprofile_add_ext_sample(pc, regs, event, is_kernel);
}
#ifdef CONFIG_OPROFILE_IBS
#define MAX_IBS_SAMPLE_SIZE 14
void oprofile_add_ibs_sample(struct pt_regs * const regs,
unsigned int * const ibs_sample, int ibs_code)
{
int is_kernel = !user_mode(regs);
struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
struct task_struct *task;
int fail = 0;
cpu_buf->sample_received++;
/* notice a switch from user->kernel or vice versa */
if (cpu_buf->last_is_kernel != is_kernel) {
if (add_code(cpu_buf, is_kernel))
goto fail;
cpu_buf->last_is_kernel = is_kernel;
}
/* notice a task switch */
if (!is_kernel) {
task = current;
if (cpu_buf->last_task != task) {
if (add_code(cpu_buf, (unsigned long)task))
goto fail;
cpu_buf->last_task = task;
}
}
fail = fail || add_code(cpu_buf, ibs_code);
fail = fail || add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
fail = fail || add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
fail = fail || add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
if (ibs_code == IBS_OP_BEGIN) {
fail = fail || add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
fail = fail || add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
fail = fail || add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
}
if (fail)
goto fail;
if (oprofile_backtrace_depth)
oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
return;
fail:
cpu_buf->sample_lost_overflow++;
return;
}
#endif
void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
{
struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
log_sample(cpu_buf, pc, is_kernel, event);
}
void oprofile_add_trace(unsigned long pc)
{
struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
if (!cpu_buf->tracing)
return;
/*
* broken frame can give an eip with the same value as an
* escape code, abort the trace if we get it
*/
if (pc == ESCAPE_CODE)
goto fail;
if (add_sample(cpu_buf, pc, 0))
goto fail;
return;
fail:
cpu_buf->tracing = 0;
cpu_buf->backtrace_aborted++;
return;
}
/*
* This serves to avoid cpu buffer overflow, and makes sure
* the task mortuary progresses
*
* By using schedule_delayed_work_on and then schedule_delayed_work
* we guarantee this will stay on the correct cpu
*/
static void wq_sync_buffer(struct work_struct *work)
{
struct oprofile_cpu_buffer *b =
container_of(work, struct oprofile_cpu_buffer, work.work);
if (b->cpu != smp_processor_id()) {
printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
smp_processor_id(), b->cpu);
if (!cpu_online(b->cpu)) {
cancel_delayed_work(&b->work);
return;
}
}
sync_buffer(b->cpu);
/* don't re-add the work if we're shutting down */
if (work_enabled)
schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
}