[RFC PATCH 2/7] core/metricfs: add support for percpu metricfs files

From: Jonathan Adams
Date: Fri Aug 07 2020 - 17:30:50 EST


Add a simple mechanism for exporting percpu data through metricfs.
The API follows the existing metricfs pattern. A percpu file is
defined with:

METRIC_EXPORT_PERCPU_INT(name, desc, fn)
METRIC_EXPORT_PERCPU_COUNTER(name, desc, fn)

The first defines a file for exposing a percpu int. The second is
similar, but is for a counter that accumulates since boot. The
'name' is used as the metricfs file. The 'desc' is a description
of the metric. The 'fn' is a callback function to emit a single
percpu value:

void (*fn)(struct metric_emitter *e, int cpu);

The callback must call METRIC_EMIT_PERCPU_INT with the value for
the specified CPU.

Signed-off-by: Jonathan Adams <jwadams@xxxxxxxxxx>

---

jwadams@xxxxxxxxxx: rebased to 5.6-pre6, renamed funcs to start with
metric_. This is work originally done by another engineer at
google, who would rather not have their name associated with this
patchset. They're okay with me sending it under my name.
---
include/linux/metricfs.h | 28 +++++++++++++++++++
kernel/metricfs.c | 58 ++++++++++++++++++++++++++++++++++++----
2 files changed, 81 insertions(+), 5 deletions(-)

diff --git a/include/linux/metricfs.h b/include/linux/metricfs.h
index 65a1baa8e8c1..f103dc8c44ec 100644
--- a/include/linux/metricfs.h
+++ b/include/linux/metricfs.h
@@ -22,6 +22,19 @@ void metric_exit_##name(void) \
metric_unregister(metric_##name); \
}

+#define METRIC_EXPORT_PERCPU(name, desc, fn, cumulative) \
+static struct metric *metric_##name; \
+void metric_init_##name(struct metricfs_subsys *parent) \
+{ \
+ metric_##name = metric_register_percpu(__stringify(name), (parent), \
+ (desc), (fn), \
+ (cumulative), THIS_MODULE); \
+} \
+void metric_exit_##name(void) \
+{ \
+ metric_unregister(metric_##name); \
+}
+
/*
* Metricfs only deals with two types: int64_t and const char*.
*
@@ -47,6 +60,11 @@ void metric_exit_##name(void) \
METRIC_EXPORT_GENERIC(name, (desc), (fname0), (fname1), (fn), \
true, false)

+#define METRIC_EXPORT_PERCPU_INT(name, desc, fn) \
+ METRIC_EXPORT_PERCPU(name, (desc), (fn), false)
+#define METRIC_EXPORT_PERCPU_COUNTER(name, desc, fn) \
+ METRIC_EXPORT_PERCPU(name, (desc), (fn), true)
+
/* Subsystem support. */
/* Pass NULL as 'parent' to create a new top-level subsystem. */
struct metricfs_subsys *metricfs_create_subsys(const char *name,
@@ -69,6 +87,8 @@ struct metric_emitter;
metric_emit_int_value((e), (v), (f0), (f1))
#define METRIC_EMIT_STR(e, v, f0, f1) \
metric_emit_str_value((e), (v), (f0), (f1))
+#define METRIC_EMIT_PERCPU_INT(e, cpu, v) \
+ metric_emit_percpu_int_value((e), (cpu), (v))

/* Users don't have to call any functions below;
* use the macro definitions above instead.
@@ -77,6 +97,7 @@ void metric_emit_int_value(struct metric_emitter *e,
int64_t v, const char *f0, const char *f1);
void metric_emit_str_value(struct metric_emitter *e,
const char *v, const char *f0, const char *f1);
+void metric_emit_percpu_int_value(struct metric_emitter *e, int cpu, int64_t v);

struct metric *metric_register(const char *name,
struct metricfs_subsys *parent,
@@ -98,6 +119,13 @@ struct metric *metric_register_parm(const char *name,
bool is_cumulative,
struct module *owner);

+struct metric *metric_register_percpu(const char *name,
+ struct metricfs_subsys *parent,
+ const char *description,
+ void (*fn)(struct metric_emitter *e, int cpu),
+ bool is_cumulative,
+ struct module *owner);
+
void metric_unregister(struct metric *m);

#endif /* _METRICFS_H_ */
diff --git a/kernel/metricfs.c b/kernel/metricfs.c
index 676b7b04aa2b..992fdd9a4d0a 100644
--- a/kernel/metricfs.c
+++ b/kernel/metricfs.c
@@ -76,6 +76,8 @@ struct metric {
bool is_string;
bool is_cumulative;
bool has_parm;
+ bool is_percpu;
+ void (*percpu_fn)(struct metric_emitter *e, int cpu);

/* dentry for the directory that contains the metric */
struct dentry *dentry;
@@ -285,6 +287,19 @@ void metric_emit_str_value(struct metric_emitter *e, const char *v,
}
EXPORT_SYMBOL(metric_emit_str_value);

+void metric_emit_percpu_int_value(struct metric_emitter *e, int cpu, int64_t v)
+{
+ char *ckpt = e->buf;
+ bool ok = true;
+
+ ok &= emit_int(e, cpu);
+ ok &= emit_string(e, " ");
+ ok &= emit_int(e, v);
+ ok &= emit_string(e, "\n");
+ if (!ok)
+ e->buf = ckpt;
+}
+
/* Contains file data generated at open() */
struct metricfs_file_private {
size_t bytes_written;
@@ -400,11 +415,15 @@ static int metricfs_fields_open(struct inode *inode, struct file *filp)
}
ok &= emit_string(&e, "value\n");

- if (m->fname0)
- ok &= emit_string(&e, "str ");
- if (m->fname1)
- ok &= emit_string(&e, "str ");
- ok &= emit_string(&e, (m->is_string) ? "str\n" : "int\n");
+ if (m->is_percpu) {
+ ok &= emit_string(&e, "int int\n");
+ } else {
+ if (m->fname0)
+ ok &= emit_string(&e, "str ");
+ if (m->fname1)
+ ok &= emit_string(&e, "str ");
+ ok &= emit_string(&e, (m->is_string) ? "str\n" : "int\n");
+ }

/* Emit all or nothing. */
if (ok) {
@@ -640,6 +659,35 @@ struct metric *metric_register(const char *name,
}
EXPORT_SYMBOL(metric_register);

+static void metric_emit_percpu(struct metric_emitter *e)
+{
+ int cpu;
+
+ for_each_possible_cpu(cpu)
+ e->metric->percpu_fn(e, cpu);
+}
+
+struct metric *metric_register_percpu(const char *name,
+ struct metricfs_subsys *parent,
+ const char *description,
+ void (*fn)(struct metric_emitter *e, int cpu),
+ bool is_cumulative,
+ struct module *owner)
+{
+ struct metric *metric =
+ metric_register(name, parent, description,
+ "cpu", NULL,
+ metric_emit_percpu,
+ false,
+ is_cumulative, owner);
+ if (metric) {
+ metric->is_percpu = true;
+ metric->percpu_fn = fn;
+ }
+ return metric;
+}
+EXPORT_SYMBOL(metric_register_percpu);
+
struct metric *metric_register_parm(const char *name,
struct metricfs_subsys *parent,
const char *description,
--
2.28.0.236.gb10cc79966-goog