Skip to contentSkip to conversation

perf tools: Get a perf cgroup more portably in BPF

PATCH Superseded Namhyung Kim 3 participants 4 messages Sep 22, 2022

Subsystems (2)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

PATCH v2 1/1 perf tools: Get a perf cgroup more portably in BPF

The perf_event_cgrp_id can be different on other configurations. To be more portable as CO-RE, it needs to get the cgroup subsys id using the bpf_core_enum_value() helper.

---
v2 changes)
* fix off_cpu.bpf.c too
* get perf_subsys_id only once

 tools/perf/util/bpf_skel/bperf_cgroup.bpf.c |  6 +++++-
 tools/perf/util/bpf_skel/off_cpu.bpf.c      | 12 ++++++++----
 2 files changed, 13 insertions(+), 5 deletions(-)
tools/perf/util/bpf_skel/bperf_cgroup.bpf.c +5 -1
index 292c430768b5..9223e4b87fe9 100644
@@ -48,6 +48,7 @@ const volatile __u32 num_cpus = 1;
 
 int enabled = 0;
 int use_cgroup_v2 = 0;
+int perf_subsys_id = -1;
 
 static inline int get_cgroup_v1_idx(__u32 *cgrps, int size)
 {
@@ -58,7 +59,10 @@ static inline int get_cgroup_v1_idx(__u32 *cgrps, int size)
 	int level;
 	int cnt;
 
-	cgrp = BPF_CORE_READ(p, cgroups, subsys[perf_event_cgrp_id], cgroup);
+	if (perf_subsys_id == -1)
+		perf_subsys_id = bpf_core_enum_value(enum cgroup_subsys_id, perf_event_cgrp_id);
+
+	cgrp = BPF_CORE_READ(p, cgroups, subsys[perf_subsys_id], cgroup);
 	level = BPF_CORE_READ(cgrp, level);
 
 	for (cnt = 0; i < MAX_LEVELS; i++) {
tools/perf/util/bpf_skel/off_cpu.bpf.c +8 -4
index c4ba2bcf179f..e917ef7b8875 100644
@@ -94,6 +94,8 @@ const volatile bool has_prev_state = false;
 const volatile bool needs_cgroup = false;
 const volatile bool uses_cgroup_v1 = false;
 
+int perf_subsys_id = -1;
+
 /*
  * Old kernel used to call it task_struct->state and now it's '__state'.
  * Use BPF CO-RE "ignored suffix rule" to deal with it like below:
@@ -119,11 +121,13 @@ static inline __u64 get_cgroup_id(struct task_struct *t)
 {
 	struct cgroup *cgrp;
 
-	if (uses_cgroup_v1)
-		cgrp = BPF_CORE_READ(t, cgroups, subsys[perf_event_cgrp_id], cgroup);
-	else
-		cgrp = BPF_CORE_READ(t, cgroups, dfl_cgrp);
+	if (!uses_cgroup_v1)
+		return BPF_CORE_READ(t, cgroups, dfl_cgrp, kn, id);
+
+	if (perf_subsys_id == -1)
+		perf_subsys_id = bpf_core_enum_value(enum cgroup_subsys_id, perf_event_cgrp_id);
 
+	cgrp = BPF_CORE_READ(t, cgroups, subsys[perf_subsys_id], cgroup);
 	return BPF_CORE_READ(cgrp, kn, id);
 }
 
Signature · 1 line
2.37.3.968.ga6b4b080e4-goog
#1 lore raw

Em Wed, Sep 21, 2022 at 09:40:23PM -0700, Namhyung Kim escreveu:

Quoted 6 lines

The perf_event_cgrp_id can be different on other configurations. To be more portable as CO-RE, it needs to get the cgroup subsys id using the bpf_core_enum_value() helper.

Applying, Ian, can I have your Reviewed-by?

- Arnaldo

Quoted 66 lines

---
v2 changes)
* fix off_cpu.bpf.c too
* get perf_subsys_id only once

 tools/perf/util/bpf_skel/bperf_cgroup.bpf.c |  6 +++++-
 tools/perf/util/bpf_skel/off_cpu.bpf.c      | 12 ++++++++----
 2 files changed, 13 insertions(+), 5 deletions(-)
tools/perf/util/bpf_skel/bperf_cgroup.bpf.c +5 -1
index 292c430768b5..9223e4b87fe9 100644
@@ -48,6 +48,7 @@ const volatile __u32 num_cpus = 1;
 
 int enabled = 0;
 int use_cgroup_v2 = 0;
+int perf_subsys_id = -1;
 
 static inline int get_cgroup_v1_idx(__u32 *cgrps, int size)
 {
@@ -58,7 +59,10 @@ static inline int get_cgroup_v1_idx(__u32 *cgrps, int size)
 	int level;
 	int cnt;
 
-	cgrp = BPF_CORE_READ(p, cgroups, subsys[perf_event_cgrp_id], cgroup);
+	if (perf_subsys_id == -1)
+		perf_subsys_id = bpf_core_enum_value(enum cgroup_subsys_id, perf_event_cgrp_id);
+
+	cgrp = BPF_CORE_READ(p, cgroups, subsys[perf_subsys_id], cgroup);
 	level = BPF_CORE_READ(cgrp, level);
 
 	for (cnt = 0; i < MAX_LEVELS; i++) {
tools/perf/util/bpf_skel/off_cpu.bpf.c +8 -4
index c4ba2bcf179f..e917ef7b8875 100644
@@ -94,6 +94,8 @@ const volatile bool has_prev_state = false;
 const volatile bool needs_cgroup = false;
 const volatile bool uses_cgroup_v1 = false;
 
+int perf_subsys_id = -1;
+
 /*
  * Old kernel used to call it task_struct->state and now it's '__state'.
  * Use BPF CO-RE "ignored suffix rule" to deal with it like below:
@@ -119,11 +121,13 @@ static inline __u64 get_cgroup_id(struct task_struct *t)
 {
 	struct cgroup *cgrp;
 
-	if (uses_cgroup_v1)
-		cgrp = BPF_CORE_READ(t, cgroups, subsys[perf_event_cgrp_id], cgroup);
-	else
-		cgrp = BPF_CORE_READ(t, cgroups, dfl_cgrp);
+	if (!uses_cgroup_v1)
+		return BPF_CORE_READ(t, cgroups, dfl_cgrp, kn, id);
+
+	if (perf_subsys_id == -1)
+		perf_subsys_id = bpf_core_enum_value(enum cgroup_subsys_id, perf_event_cgrp_id);
 
+	cgrp = BPF_CORE_READ(t, cgroups, subsys[perf_subsys_id], cgroup);
 	return BPF_CORE_READ(cgrp, kn, id);
 }
 
Signature · 1 line
2.37.3.968.ga6b4b080e4-goog
Signature · 1 line
- Arnaldo
#2 lore raw
In reply to Namhyung Kim
Quoted 5 lines · Namhyung Kim

The perf_event_cgrp_id can be different on other configurations. To be more portable as CO-RE, it needs to get the cgroup subsys id using the bpf_core_enum_value() helper.

I remember using bpf_core_enum_value requires a compiler built-in. So the build will fail on old compiler such as clang-11. See [1]. Maybe we should surround it with #if __has_builtin(__builtin_preserve_enum_value) to be sure.

[1] https://www.spinics.net/lists/bpf/msg30859.html

Quoted 3 lines

---

[...]

author maintainer
#3 lore raw

Hi Hao,

Quoted 8 lines · Hao Luo
Quoted 5 lines · Namhyung Kim

The perf_event_cgrp_id can be different on other configurations. To be more portable as CO-RE, it needs to get the cgroup subsys id using the bpf_core_enum_value() helper.

Quoted 6 lines

I remember using bpf_core_enum_value requires a compiler built-in. So the build will fail on old compiler such as clang-11. See [1]. Maybe we should surround it with #if __has_builtin(__builtin_preserve_enum_value) to be sure.

[1] https://www.spinics.net/lists/bpf/msg30859.html

Thanks for pointing this out. As this is a kind of optimization I think we can fallback to using the existing value if not available. Will send v3.

Thanks,
Namhyung