linux/kernel/trace/trace_event_perf.c

194 lines
4.0 KiB
C
Raw Normal View History

/*
* trace event based perf event profiling/tracing
*
* Copyright (C) 2009 Red Hat Inc, Peter Zijlstra <pzijlstr@redhat.com>
perf: Take a hot regs snapshot for trace events We are taking a wrong regs snapshot when a trace event triggers. Either we use get_irq_regs(), which gives us the interrupted registers if we are in an interrupt, or we use task_pt_regs() which gives us the state before we entered the kernel, assuming we are lucky enough to be no kernel thread, in which case task_pt_regs() returns the initial set of regs when the kernel thread was started. What we want is different. We need a hot snapshot of the regs, so that we can get the instruction pointer to record in the sample, the frame pointer for the callchain, and some other things. Let's use the new perf_fetch_caller_regs() for that. Comparison with perf record -e lock: -R -a -f -g Before: perf [kernel] [k] __do_softirq | --- __do_softirq | |--55.16%-- __open | --44.84%-- __write_nocancel After: perf [kernel] [k] perf_tp_event | --- perf_tp_event | |--41.07%-- lock_acquire | | | |--39.36%-- _raw_spin_lock | | | | | |--7.81%-- hrtimer_interrupt | | | smp_apic_timer_interrupt | | | apic_timer_interrupt The old case was producing unreliable callchains. Now having right frame and instruction pointers, we have the trace we want. Also syscalls and kprobe events already have the right regs, let's use them instead of wasting a retrieval. v2: Follow the rename perf_save_regs() -> perf_fetch_caller_regs() Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Masami Hiramatsu <mhiramat@redhat.com> Cc: Jason Baron <jbaron@redhat.com> Cc: Archs <linux-arch@vger.kernel.org>
2010-03-03 14:16:16 +08:00
* Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com>
*/
#include <linux/module.h>
#include <linux/kprobes.h>
#include "trace.h"
static char *perf_trace_buf[4];
/*
* Force it to be aligned to unsigned long to avoid misaligned accesses
* suprises
*/
typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
perf_trace_t;
/* Count the events in use (per event id, not per instance) */
static int total_ref_count;
static int perf_trace_event_init(struct ftrace_event_call *tp_event,
struct perf_event *p_event)
{
struct hlist_head *list;
int ret = -ENOMEM;
int cpu;
p_event->tp_event = tp_event;
if (tp_event->perf_refcount++ > 0)
return 0;
list = alloc_percpu(struct hlist_head);
if (!list)
goto fail;
for_each_possible_cpu(cpu)
INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
tp_event->perf_events = list;
if (!total_ref_count) {
char *buf;
int i;
for (i = 0; i < 4; i++) {
buf = (char *)alloc_percpu(perf_trace_t);
if (!buf)
goto fail;
perf_trace_buf[i] = buf;
}
}
if (tp_event->class->reg)
ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER);
tracing: Remove per event trace registering This patch removes the register functions of TRACE_EVENT() to enable and disable tracepoints. The registering of a event is now down directly in the trace_events.c file. The tracepoint_probe_register() is now called directly. The prototypes are no longer type checked, but this should not be an issue since the tracepoints are created automatically by the macros. If a prototype is incorrect in the TRACE_EVENT() macro, then other macros will catch it. The trace_event_class structure now holds the probes to be called by the callbacks. This removes needing to have each event have a separate pointer for the probe. To handle kprobes and syscalls, since they register probes in a different manner, a "reg" field is added to the ftrace_event_class structure. If the "reg" field is assigned, then it will be called for enabling and disabling of the probe for either ftrace or perf. To let the reg function know what is happening, a new enum (trace_reg) is created that has the type of control that is needed. With this new rework, the 82 kernel events and 618 syscall events has their footprint dramatically lowered: text data bss dec hex filename 4913961 1088356 861512 6863829 68bbd5 vmlinux.orig 4914025 1088868 861512 6864405 68be15 vmlinux.class 4918492 1084612 861512 6864616 68bee8 vmlinux.tracepoint 4900252 1057412 861512 6819176 680d68 vmlinux.regs The size went from 6863829 to 6819176, that's a total of 44K in savings. With tracepoints being continuously added, this is critical that the footprint becomes minimal. v5: Added #ifdef CONFIG_PERF_EVENTS around a reference to perf specific structure in trace_events.c. v4: Fixed trace self tests to check probe because regfunc no longer exists. v3: Updated to handle void *data in beginning of probe parameters. Also added the tracepoint: check_trace_callback_type_##call(). v2: Changed the callback probes to pass void * and typecast the value within the function. Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Acked-by: Masami Hiramatsu <mhiramat@redhat.com> Acked-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2010-04-22 00:27:06 +08:00
else
ret = tracepoint_probe_register(tp_event->name,
tp_event->class->perf_probe,
tp_event);
if (ret)
goto fail;
total_ref_count++;
return 0;
fail:
if (!total_ref_count) {
int i;
for (i = 0; i < 4; i++) {
free_percpu(perf_trace_buf[i]);
perf_trace_buf[i] = NULL;
}
}
if (!--tp_event->perf_refcount) {
free_percpu(tp_event->perf_events);
tp_event->perf_events = NULL;
}
return ret;
}
int perf_trace_init(struct perf_event *p_event)
{
struct ftrace_event_call *tp_event;
int event_id = p_event->attr.config;
int ret = -EINVAL;
mutex_lock(&event_mutex);
list_for_each_entry(tp_event, &ftrace_events, list) {
if (tp_event->event.type == event_id &&
tp_event->class &&
(tp_event->class->perf_probe ||
tp_event->class->reg) &&
try_module_get(tp_event->mod)) {
ret = perf_trace_event_init(tp_event, p_event);
break;
}
}
mutex_unlock(&event_mutex);
return ret;
}
int perf_trace_enable(struct perf_event *p_event)
{
struct ftrace_event_call *tp_event = p_event->tp_event;
struct hlist_head *list;
list = tp_event->perf_events;
if (WARN_ON_ONCE(!list))
return -EINVAL;
list = this_cpu_ptr(list);
hlist_add_head_rcu(&p_event->hlist_entry, list);
return 0;
}
void perf_trace_disable(struct perf_event *p_event)
{
hlist_del_rcu(&p_event->hlist_entry);
}
void perf_trace_destroy(struct perf_event *p_event)
{
struct ftrace_event_call *tp_event = p_event->tp_event;
int i;
mutex_lock(&event_mutex);
if (--tp_event->perf_refcount > 0)
goto out;
if (tp_event->class->reg)
tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER);
else
tracepoint_probe_unregister(tp_event->name,
tp_event->class->perf_probe,
tp_event);
/*
* Ensure our callback won't be called anymore. See
* tracepoint_probe_unregister() and __DO_TRACE().
*/
synchronize_sched();
free_percpu(tp_event->perf_events);
tp_event->perf_events = NULL;
if (!--total_ref_count) {
for (i = 0; i < 4; i++) {
free_percpu(perf_trace_buf[i]);
perf_trace_buf[i] = NULL;
}
}
out:
mutex_unlock(&event_mutex);
}
__kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
struct pt_regs *regs, int *rctxp)
{
struct trace_entry *entry;
unsigned long flags;
char *raw_data;
int pc;
BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
pc = preempt_count();
*rctxp = perf_swevent_get_recursion_context();
if (*rctxp < 0)
return NULL;
raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
/* zero the dead bytes from align to not leak stack to user */
memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
entry = (struct trace_entry *)raw_data;
local_save_flags(flags);
tracing_generic_entry_update(entry, flags, pc);
entry->type = type;
return raw_data;
}
EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);