Skip to contentSkip to conversation

uprobes: make trace_uprobe->nhit counter a per-CPU one

PATCH In review Andrii Nakryiko 4 participants 7 messages Aug 9, 2024 – Aug 26, 2024

Subsystems (2)
lore
Share & tools
lore

Andrii Nakryiko

author maintainer
#0 lore raw

PATCH v2 1/1 uprobes: make trace_uprobe->nhit counter a per-CPU one

trace_uprobe->nhit counter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.

Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.

Solve both problems by making this a per-CPU counter.

---
 kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++---
 1 file changed, 21 insertions(+), 3 deletions(-)
kernel/trace/trace_uprobe.c +21 -3
index 52e76a73fa7c..002f801a7ab4 100644
@@ -17,6 +17,7 @@
 #include <linux/string.h>
 #include <linux/rculist.h>
 #include <linux/filter.h>
+#include <linux/percpu.h>
 
 #include "trace_dynevent.h"
 #include "trace_probe.h"
@@ -62,7 +63,7 @@ struct trace_uprobe {
 	struct uprobe			*uprobe;
 	unsigned long			offset;
 	unsigned long			ref_ctr_offset;
-	unsigned long			nhit;
+	unsigned long __percpu		*nhits;
 	struct trace_probe		tp;
 };
 
@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)
 	if (!tu)
 		return ERR_PTR(-ENOMEM);
 
+	tu->nhits = alloc_percpu(unsigned long);
+	if (!tu->nhits) {
+		ret = -ENOMEM;
+		goto error;
+	}
+
 	ret = trace_probe_init(&tu->tp, event, group, true, nargs);
 	if (ret < 0)
 		goto error;
@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)
 	return tu;
 
 error:
+	free_percpu(tu->nhits);
 	kfree(tu);
 
 	return ERR_PTR(ret);
@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)
 	path_put(&tu->path);
 	trace_probe_cleanup(&tu->tp);
 	kfree(tu->filename);
+	free_percpu(tu->nhits);
 	kfree(tu);
 }
 
@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v)
 {
 	struct dyn_event *ev = v;
 	struct trace_uprobe *tu;
+	unsigned long nhits;
+	int cpu;
 
 	if (!is_trace_uprobe(ev))
 		return 0;
 
 	tu = to_trace_uprobe(ev);
+
+	nhits = 0;
+	for_each_possible_cpu(cpu) {
+		nhits += READ_ONCE(*per_cpu_ptr(tu->nhits, cpu));
+	}
+
 	seq_printf(m, "  %s %-44s %15lu\n", tu->filename,
-			trace_probe_name(&tu->tp), tu->nhit);
+		   trace_probe_name(&tu->tp), nhits);
 	return 0;
 }
 
@@ -1507,7 +1524,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)
 	int ret = 0;
 
 	tu = container_of(con, struct trace_uprobe, consumer);
-	tu->nhit++;
+
+	this_cpu_inc(*tu->nhits);
 
 	udd.tu = tu;
 	udd.bp_addr = instruction_pointer(regs);
Signature · 1 line
2.43.5
#1 lore raw
Quoted 9 lines · On Fri, 9 Aug 2024 12:23:57 -0700 Andrii Nakryiko <[email protected]> wrote:

trace_uprobe->nhit counter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.

Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.

Solve both problems by making this a per-CPU counter.

This looks good to me. I would like to pick this to linux-trace/probes/for-next.

kernel/trace/trace_uprobe.c struct trace_uprobe { · line 63

Quoted 2 lines
kernel/trace/trace_uprobe.c +0 -0
@@ -62,7 +63,7 @@ struct trace_uprobe {
 	struct uprobe			*uprobe;

BTW, what is this change? I couldn't cleanly apply this to the v6.11-rc3. Which tree would you working on? (I missed something?)

Thanks,

Quoted 72 lines
kernel/trace/trace_uprobe.c +20 -3
 	unsigned long			offset;
 	unsigned long			ref_ctr_offset;
-	unsigned long			nhit;
+	unsigned long __percpu		*nhits;
 	struct trace_probe		tp;
 };
 
@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)
 	if (!tu)
 		return ERR_PTR(-ENOMEM);
 
+	tu->nhits = alloc_percpu(unsigned long);
+	if (!tu->nhits) {
+		ret = -ENOMEM;
+		goto error;
+	}
+
 	ret = trace_probe_init(&tu->tp, event, group, true, nargs);
 	if (ret < 0)
 		goto error;
@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)
 	return tu;
 
 error:
+	free_percpu(tu->nhits);
 	kfree(tu);
 
 	return ERR_PTR(ret);
@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)
 	path_put(&tu->path);
 	trace_probe_cleanup(&tu->tp);
 	kfree(tu->filename);
+	free_percpu(tu->nhits);
 	kfree(tu);
 }
 
@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v)
 {
 	struct dyn_event *ev = v;
 	struct trace_uprobe *tu;
+	unsigned long nhits;
+	int cpu;
 
 	if (!is_trace_uprobe(ev))
 		return 0;
 
 	tu = to_trace_uprobe(ev);
+
+	nhits = 0;
+	for_each_possible_cpu(cpu) {
+		nhits += READ_ONCE(*per_cpu_ptr(tu->nhits, cpu));
+	}
+
 	seq_printf(m, "  %s %-44s %15lu\n", tu->filename,
-			trace_probe_name(&tu->tp), tu->nhit);
+		   trace_probe_name(&tu->tp), nhits);
 	return 0;
 }
 
@@ -1507,7 +1524,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)
 	int ret = 0;
 
 	tu = container_of(con, struct trace_uprobe, consumer);
-	tu->nhit++;
+
+	this_cpu_inc(*tu->nhits);
 
 	udd.tu = tu;
 	udd.bp_addr = instruction_pointer(regs);
Signature · 1 line
2.43.5
Signature · 1 line
Masami Hiramatsu (Google) <[email protected]>
#2 lore raw
Quoted 6 lines · Masami Hiramatsu

kernel/trace/trace_uprobe.c struct trace_uprobe { · line 63

Quoted 3 lines
kernel/trace/trace_uprobe.c +0 -0
@@ -62,7 +63,7 @@ struct trace_uprobe {
 	struct uprobe			*uprobe;

BTW, what is this change? I couldn't cleanly apply this to the v6.11-rc3. Which tree would you working on? (I missed something?)

tip/perf/core

See https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/diff/kernel/trace/trace_uprobe.c?h=perf/core&id=3c83a9ad0295eb63bdeb81d821b8c3b9417fbcac

Oleg.

#3 lore raw
Quoted 11 lines · On Tue, 13 Aug 2024 17:41:04 +0200 Oleg Nesterov <[email protected]> wrote:
Quoted 6 lines · Masami Hiramatsu

kernel/trace/trace_uprobe.c struct trace_uprobe { · line 63

Quoted 3 lines
kernel/trace/trace_uprobe.c +0 -0
@@ -62,7 +63,7 @@ struct trace_uprobe {
 	struct uprobe			*uprobe;

BTW, what is this change? I couldn't cleanly apply this to the v6.11-rc3. Which tree would you working on? (I missed something?)

tip/perf/core

See https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/diff/kernel/trace/trace_uprobe.c?h=perf/core&id=3c83a9ad0295eb63bdeb81d821b8c3b9417fbcac

OK, let me consider to rebase on tip/perf/core.

Thank you,

Quoted 3 lines

Oleg.

Signature · 1 line
Masami Hiramatsu (Google) <[email protected]>
#4 lore raw
Quoted 12 lines · Masami Hiramatsu
Quoted 8 lines · On Tue, 13 Aug 2024 17:41:04 +0200 Oleg Nesterov <[email protected]> wrote:
Quoted 6 lines · Masami Hiramatsu

kernel/trace/trace_uprobe.c struct trace_uprobe { · line 63

Quoted 3 lines
kernel/trace/trace_uprobe.c +0 -0
@@ -62,7 +63,7 @@ struct trace_uprobe {
  struct uprobe                   *uprobe;

BTW, what is this change? I couldn't cleanly apply this to the v6.11-rc3. Which tree would you working on? (I missed something?)

Quoted 6 lines
Quoted 3 lines

tip/perf/core

See https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/diff/kernel/trace/trace_uprobe.c?h=perf/core&id=3c83a9ad0295eb63bdeb81d821b8c3b9417fbcac

OK, let me consider to rebase on tip/perf/core.

Hey Masami,

I've posted v3 rebased onto linux-trace/probes/for-next, so you shouldn't need to rebase anything just for this. See [0] for the latest revision.

[0] https://lore.kernel.org/linux-trace-kernel/[email protected]/

Quoted 9 lines

Thank you,

Quoted 3 lines

Oleg.

--
Masami Hiramatsu (Google) <[email protected]>

Reviewed-by on v2 1/1
#5 lore raw
In reply to Andrii Nakryiko
Quoted 6 lines · Andrii Nakryiko
kernel/trace/trace_uprobe.c +1 -0
@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v)
 {
 	struct dyn_event *ev = v;
 	struct trace_uprobe *tu;
+	unsigned long nhits;

kernel/trace/trace_uprobe.c static int probes_profile_seq_show(struct seq_file *m, void *v) · line 837

Quoted 10 lines
kernel/trace/trace_uprobe.c +5 -0
+	int cpu;
 
 	if (!is_trace_uprobe(ev))
 		return 0;
 
 	tu = to_trace_uprobe(ev);
+
+	nhits = 0;
+	for_each_possible_cpu(cpu) {
+		nhits += READ_ONCE(*per_cpu_ptr(tu->nhits, cpu));

why not

		nhits += per_cpu(*tu->nhits, cpu);

?

See for example per_cpu_sum() or nr_processes(), per_cpu() should work just fine...

Other than that

#6 lore raw
Quoted 21 lines · Oleg Nesterov
Quoted 6 lines · Andrii Nakryiko
kernel/trace/trace_uprobe.c +1 -0
@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v)
 {
      struct dyn_event *ev = v;
      struct trace_uprobe *tu;
+     unsigned long nhits;

kernel/trace/trace_uprobe.c static int probes_profile_seq_show(struct seq_file *m, void *v) · line 837

Quoted 10 lines
kernel/trace/trace_uprobe.c +5 -0
+     int cpu;
 
      if (!is_trace_uprobe(ev))
              return 0;
 
      tu = to_trace_uprobe(ev);
+
+     nhits = 0;
+     for_each_possible_cpu(cpu) {
+             nhits += READ_ONCE(*per_cpu_ptr(tu->nhits, cpu));

why not

Quoted 6 lines
                nhits += per_cpu(*tu->nhits, cpu);

?

See for example per_cpu_sum() or nr_processes(), per_cpu() should work just fine...

I just monkeyed it from some existing code somewhere in the BPF code base. I like per_cpu, will send a v3 and rebase it onto a linux-trace tree.

Quoted 4 lines

Other than that