kprobes: Optimize hot path by using percpu counter to collect 'nhit' statistics
When doing ebpf+kprobe on some hot TCP functions (e.g.
tcp_rcv_established), the kprobe_dispatcher() function
shows up in 'perf report'.
In kprobe_dispatcher(), there is a lot of cache bouncing
on 'tk->nhit++'. 'tk->nhit' and 'tk->tp.flags' also share
the same cacheline.
perf report (cycles:pp):
8.30% ipv4_dst_check
4.74% copy_user_enhanced_fast_string
3.93% dst_release
2.80% tcp_v4_rcv
2.31% queued_spin_lock_slowpath
2.30% _raw_spin_lock
1.88% mlx4_en_process_rx_cq
1.84% eth_get_headlen
1.81% ip_rcv_finish
~~~~
1.71% kprobe_dispatcher
~~~~
1.55% mlx4_en_xmit
1.09% __probe_kernel_read
perf report after patch:
9.15% ipv4_dst_check
5.00% copy_user_enhanced_fast_string
4.12% dst_release
2.96% tcp_v4_rcv
2.50% _raw_spin_lock
2.39% queued_spin_lock_slowpath
2.11% eth_get_headlen
2.03% mlx4_en_process_rx_cq
1.69% mlx4_en_xmit
1.19% ip_rcv_finish
1.12% __probe_kernel_read
1.02% ehci_hcd_cleanup
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Kernel Team <kernel-team@fb.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1454531308-2441898-1-git-send-email-kafai@fb.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
diff --git a/kernel/trace/trace_kprobe.c b/kernel/trace/trace_kprobe.c
index c995644..21b81a4 100644
--- a/kernel/trace/trace_kprobe.c
+++ b/kernel/trace/trace_kprobe.c
@@ -30,7 +30,7 @@
struct trace_kprobe {
struct list_head list;
struct kretprobe rp; /* Use rp.kp for kprobe use */
- unsigned long nhit;
+ unsigned long __percpu *nhit;
const char *symbol; /* symbol name */
struct trace_probe tp;
};
@@ -274,6 +274,10 @@
if (!tk)
return ERR_PTR(ret);
+ tk->nhit = alloc_percpu(unsigned long);
+ if (!tk->nhit)
+ goto error;
+
if (symbol) {
tk->symbol = kstrdup(symbol, GFP_KERNEL);
if (!tk->symbol)
@@ -313,6 +317,7 @@
error:
kfree(tk->tp.call.name);
kfree(tk->symbol);
+ free_percpu(tk->nhit);
kfree(tk);
return ERR_PTR(ret);
}
@@ -327,6 +332,7 @@
kfree(tk->tp.call.class->system);
kfree(tk->tp.call.name);
kfree(tk->symbol);
+ free_percpu(tk->nhit);
kfree(tk);
}
@@ -874,9 +880,14 @@
static int probes_profile_seq_show(struct seq_file *m, void *v)
{
struct trace_kprobe *tk = v;
+ unsigned long nhit = 0;
+ int cpu;
+
+ for_each_possible_cpu(cpu)
+ nhit += *per_cpu_ptr(tk->nhit, cpu);
seq_printf(m, " %-44s %15lu %15lu\n",
- trace_event_name(&tk->tp.call), tk->nhit,
+ trace_event_name(&tk->tp.call), nhit,
tk->rp.kp.nmissed);
return 0;
@@ -1225,7 +1236,7 @@
{
struct trace_kprobe *tk = container_of(kp, struct trace_kprobe, rp.kp);
- tk->nhit++;
+ raw_cpu_inc(*tk->nhit);
if (tk->tp.flags & TP_FLAG_TRACE)
kprobe_trace_func(tk, regs);
@@ -1242,7 +1253,7 @@
{
struct trace_kprobe *tk = container_of(ri->rp, struct trace_kprobe, rp);
- tk->nhit++;
+ raw_cpu_inc(*tk->nhit);
if (tk->tp.flags & TP_FLAG_TRACE)
kretprobe_trace_func(tk, ri, regs);
