perf/core: Implement the 'perf_kprobe' PMU

A new PMU type, perf_kprobe is added. Based on attr from perf_event_open(),
perf_kprobe creates a kprobe (or kretprobe) for the perf_event. This
kprobe is private to this perf_event, and thus not added to global
lists, and not available in tracefs.

Two functions, create_local_trace_kprobe() and
destroy_local_trace_kprobe()  are added to created and destroy these
local trace_kprobe.

Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Yonghong Song <yhs@fb.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Cc: <daniel@iogearbox.net>
Cc: <davem@davemloft.net>
Cc: <kernel-team@fb.com>
Cc: <rostedt@goodmis.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20171206224518.3598254-6-songliubraving@fb.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>

include/linux/trace_events.h

@@ -533,6 +533,10 @@ extern int  perf_trace_init(struct perf_event *event);
 extern void perf_trace_destroy(struct perf_event *event);
 extern int  perf_trace_add(struct perf_event *event, int flags);
 extern void perf_trace_del(struct perf_event *event, int flags);
+#ifdef CONFIG_KPROBE_EVENTS
+extern int  perf_kprobe_init(struct perf_event *event, bool is_retprobe);
+extern void perf_kprobe_destroy(struct perf_event *event);
+#endif
 extern int  ftrace_profile_set_filter(struct perf_event *event, int event_id,
 				     char *filter_str);
 extern void ftrace_profile_free_filter(struct perf_event *event);

kernel/events/core.c

@@ -7992,9 +7992,77 @@ static struct pmu perf_tracepoint = {
 	.read		= perf_swevent_read,
 };
 
+#ifdef CONFIG_KPROBE_EVENTS
+/*
+ * Flags in config, used by dynamic PMU kprobe and uprobe
+ * The flags should match following PMU_FORMAT_ATTR().
+ *
+ * PERF_PROBE_CONFIG_IS_RETPROBE if set, create kretprobe/uretprobe
+ *                               if not set, create kprobe/uprobe
+ */
+enum perf_probe_config {
+	PERF_PROBE_CONFIG_IS_RETPROBE = 1U << 0,  /* [k,u]retprobe */
+};
+
+PMU_FORMAT_ATTR(retprobe, "config:0");
+
+static struct attribute *probe_attrs[] = {
+	&format_attr_retprobe.attr,
+	NULL,
+};
+
+static struct attribute_group probe_format_group = {
+	.name = "format",
+	.attrs = probe_attrs,
+};
+
+static const struct attribute_group *probe_attr_groups[] = {
+	&probe_format_group,
+	NULL,
+};
+
+static int perf_kprobe_event_init(struct perf_event *event);
+static struct pmu perf_kprobe = {
+	.task_ctx_nr	= perf_sw_context,
+	.event_init	= perf_kprobe_event_init,
+	.add		= perf_trace_add,
+	.del		= perf_trace_del,
+	.start		= perf_swevent_start,
+	.stop		= perf_swevent_stop,
+	.read		= perf_swevent_read,
+	.attr_groups	= probe_attr_groups,
+};
+
+static int perf_kprobe_event_init(struct perf_event *event)
+{
+	int err;
+	bool is_retprobe;
+
+	if (event->attr.type != perf_kprobe.type)
+		return -ENOENT;
+	/*
+	 * no branch sampling for probe events
+	 */
+	if (has_branch_stack(event))
+		return -EOPNOTSUPP;
+
+	is_retprobe = event->attr.config & PERF_PROBE_CONFIG_IS_RETPROBE;
+	err = perf_kprobe_init(event, is_retprobe);
+	if (err)
+		return err;
+
+	event->destroy = perf_kprobe_destroy;
+
+	return 0;
+}
+#endif /* CONFIG_KPROBE_EVENTS */
+
 static inline void perf_tp_register(void)
 {
 	perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
+#ifdef CONFIG_KPROBE_EVENTS
+	perf_pmu_register(&perf_kprobe, "kprobe", -1);
+#endif
 }
 
 static void perf_event_free_filter(struct perf_event *event)
@@ -8071,13 +8139,28 @@ static void perf_event_free_bpf_handler(struct perf_event *event)
 }
 #endif
 
+/*
+ * returns true if the event is a tracepoint, or a kprobe/upprobe created
+ * with perf_event_open()
+ */
+static inline bool perf_event_is_tracing(struct perf_event *event)
+{
+	if (event->pmu == &perf_tracepoint)
+		return true;
+#ifdef CONFIG_KPROBE_EVENTS
+	if (event->pmu == &perf_kprobe)
+		return true;
+#endif
+	return false;
+}
+
 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
 {
 	bool is_kprobe, is_tracepoint, is_syscall_tp;
 	struct bpf_prog *prog;
 	int ret;
 
-	if (event->attr.type != PERF_TYPE_TRACEPOINT)
+	if (!perf_event_is_tracing(event))
 		return perf_event_set_bpf_handler(event, prog_fd);
 
 	is_kprobe = event->tp_event->flags & TRACE_EVENT_FL_UKPROBE;
@@ -8116,7 +8199,7 @@ static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
 
 static void perf_event_free_bpf_prog(struct perf_event *event)
 {
-	if (event->attr.type != PERF_TYPE_TRACEPOINT) {
+	if (!perf_event_is_tracing(event)) {
 		perf_event_free_bpf_handler(event);
 		return;
 	}
@@ -8535,47 +8618,36 @@ fail_clear_files:
 	return ret;
 }
 
-static int
-perf_tracepoint_set_filter(struct perf_event *event, char *filter_str)
-{
-	struct perf_event_context *ctx = event->ctx;
-	int ret;
-
-	/*
-	 * Beware, here be dragons!!
-	 *
-	 * the tracepoint muck will deadlock against ctx->mutex, but the tracepoint
-	 * stuff does not actually need it. So temporarily drop ctx->mutex. As per
-	 * perf_event_ctx_lock() we already have a reference on ctx.
-	 *
-	 * This can result in event getting moved to a different ctx, but that
-	 * does not affect the tracepoint state.
-	 */
-	mutex_unlock(&ctx->mutex);
-	ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
-	mutex_lock(&ctx->mutex);
-
-	return ret;
-}
-
 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
 {
-	char *filter_str;
 	int ret = -EINVAL;
-
+	char *filter_str;
-	if ((event->attr.type != PERF_TYPE_TRACEPOINT ||
-	    !IS_ENABLED(CONFIG_EVENT_TRACING)) &&
-	    !has_addr_filter(event))
-		return -EINVAL;
 
 	filter_str = strndup_user(arg, PAGE_SIZE);
 	if (IS_ERR(filter_str))
 		return PTR_ERR(filter_str);
 
-	if (IS_ENABLED(CONFIG_EVENT_TRACING) &&
+#ifdef CONFIG_EVENT_TRACING
-	    event->attr.type == PERF_TYPE_TRACEPOINT)
+	if (perf_event_is_tracing(event)) {
-		ret = perf_tracepoint_set_filter(event, filter_str);
+		struct perf_event_context *ctx = event->ctx;
-	else if (has_addr_filter(event))
+
+		/*
+		 * Beware, here be dragons!!
+		 *
+		 * the tracepoint muck will deadlock against ctx->mutex, but
+		 * the tracepoint stuff does not actually need it. So
+		 * temporarily drop ctx->mutex. As per perf_event_ctx_lock() we
+		 * already have a reference on ctx.
+		 *
+		 * This can result in event getting moved to a different ctx,
+		 * but that does not affect the tracepoint state.
+		 */
+		mutex_unlock(&ctx->mutex);
+		ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
+		mutex_lock(&ctx->mutex);
+	} else
+#endif
+	if (has_addr_filter(event))
 		ret = perf_event_set_addr_filter(event, filter_str);
 
 	kfree(filter_str);

kernel/trace/trace_event_perf.c

@@ -8,6 +8,7 @@
 #include <linux/module.h>
 #include <linux/kprobes.h>
 #include "trace.h"
+#include "trace_probe.h"
 
 static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
 
@@ -237,6 +238,54 @@ void perf_trace_destroy(struct perf_event *p_event)
 	mutex_unlock(&event_mutex);
 }
 
+#ifdef CONFIG_KPROBE_EVENTS
+int perf_kprobe_init(struct perf_event *p_event, bool is_retprobe)
+{
+	int ret;
+	char *func = NULL;
+	struct trace_event_call *tp_event;
+
+	if (p_event->attr.kprobe_func) {
+		func = kzalloc(KSYM_NAME_LEN, GFP_KERNEL);
+		if (!func)
+			return -ENOMEM;
+		ret = strncpy_from_user(
+			func, u64_to_user_ptr(p_event->attr.kprobe_func),
+			KSYM_NAME_LEN);
+		if (ret < 0)
+			goto out;
+
+		if (func[0] == '\0') {
+			kfree(func);
+			func = NULL;
+		}
+	}
+
+	tp_event = create_local_trace_kprobe(
+		func, (void *)(unsigned long)(p_event->attr.kprobe_addr),
+		p_event->attr.probe_offset, is_retprobe);
+	if (IS_ERR(tp_event)) {
+		ret = PTR_ERR(tp_event);
+		goto out;
+	}
+
+	ret = perf_trace_event_init(tp_event, p_event);
+	if (ret)
+		destroy_local_trace_kprobe(tp_event);
+out:
+	kfree(func);
+	return ret;
+}
+
+void perf_kprobe_destroy(struct perf_event *p_event)
+{
+	perf_trace_event_close(p_event);
+	perf_trace_event_unreg(p_event);
+
+	destroy_local_trace_kprobe(p_event->tp_event);
+}
+#endif /* CONFIG_KPROBE_EVENTS */
+
 int perf_trace_add(struct perf_event *p_event, int flags)
 {
 	struct trace_event_call *tp_event = p_event->tp_event;

kernel/trace/trace_kprobe.c

@@ -438,6 +438,14 @@ disable_trace_kprobe(struct trace_kprobe *tk, struct trace_event_file *file)
 			disable_kprobe(&tk->rp.kp);
 		wait = 1;
 	}
+
+	/*
+	 * if tk is not added to any list, it must be a local trace_kprobe
+	 * created with perf_event_open. We don't need to wait for these
+	 * trace_kprobes
+	 */
+	if (list_empty(&tk->list))
+		wait = 0;
  out:
 	if (wait) {
 		/*
@@ -1313,12 +1321,9 @@ static struct trace_event_functions kprobe_funcs = {
 	.trace		= print_kprobe_event
 };
 
-static int register_kprobe_event(struct trace_kprobe *tk)
+static inline void init_trace_event_call(struct trace_kprobe *tk,
+					 struct trace_event_call *call)
 {
-	struct trace_event_call *call = &tk->tp.call;
-	int ret;
-
-	/* Initialize trace_event_call */
 	INIT_LIST_HEAD(&call->class->fields);
 	if (trace_kprobe_is_return(tk)) {
 		call->event.funcs = &kretprobe_funcs;
@@ -1327,6 +1332,19 @@ static int register_kprobe_event(struct trace_kprobe *tk)
 		call->event.funcs = &kprobe_funcs;
 		call->class->define_fields = kprobe_event_define_fields;
 	}
+
+	call->flags = TRACE_EVENT_FL_KPROBE;
+	call->class->reg = kprobe_register;
+	call->data = tk;
+}
+
+static int register_kprobe_event(struct trace_kprobe *tk)
+{
+	struct trace_event_call *call = &tk->tp.call;
+	int ret = 0;
+
+	init_trace_event_call(tk, call);
+
 	if (set_print_fmt(&tk->tp, trace_kprobe_is_return(tk)) < 0)
 		return -ENOMEM;
 	ret = register_trace_event(&call->event);
@@ -1334,9 +1352,6 @@ static int register_kprobe_event(struct trace_kprobe *tk)
 		kfree(call->print_fmt);
 		return -ENODEV;
 	}
-	call->flags = TRACE_EVENT_FL_KPROBE;
-	call->class->reg = kprobe_register;
-	call->data = tk;
 	ret = trace_add_event_call(call);
 	if (ret) {
 		pr_info("Failed to register kprobe event: %s\n",
@@ -1358,6 +1373,66 @@ static int unregister_kprobe_event(struct trace_kprobe *tk)
 	return ret;
 }
 
+#ifdef CONFIG_PERF_EVENTS
+/* create a trace_kprobe, but don't add it to global lists */
+struct trace_event_call *
+create_local_trace_kprobe(char *func, void *addr, unsigned long offs,
+			  bool is_return)
+{
+	struct trace_kprobe *tk;
+	int ret;
+	char *event;
+
+	/*
+	 * local trace_kprobes are not added to probe_list, so they are never
+	 * searched in find_trace_kprobe(). Therefore, there is no concern of
+	 * duplicated name here.
+	 */
+	event = func ? func : "DUMMY_EVENT";
+
+	tk = alloc_trace_kprobe(KPROBE_EVENT_SYSTEM, event, (void *)addr, func,
+				offs, 0 /* maxactive */, 0 /* nargs */,
+				is_return);
+
+	if (IS_ERR(tk)) {
+		pr_info("Failed to allocate trace_probe.(%d)\n",
+			(int)PTR_ERR(tk));
+		return ERR_CAST(tk);
+	}
+
+	init_trace_event_call(tk, &tk->tp.call);
+
+	if (set_print_fmt(&tk->tp, trace_kprobe_is_return(tk)) < 0) {
+		ret = -ENOMEM;
+		goto error;
+	}
+
+	ret = __register_trace_kprobe(tk);
+	if (ret < 0)
+		goto error;
+
+	return &tk->tp.call;
+error:
+	free_trace_kprobe(tk);
+	return ERR_PTR(ret);
+}
+
+void destroy_local_trace_kprobe(struct trace_event_call *event_call)
+{
+	struct trace_kprobe *tk;
+
+	tk = container_of(event_call, struct trace_kprobe, tp.call);
+
+	if (trace_probe_is_enabled(&tk->tp)) {
+		WARN_ON(1);
+		return;
+	}
+
+	__unregister_trace_kprobe(tk);
+	free_trace_kprobe(tk);
+}
+#endif /* CONFIG_PERF_EVENTS */
+
 /* Make a tracefs interface for controlling probe points */
 static __init int init_kprobe_trace(void)
 {

kernel/trace/trace_probe.h

@@ -404,3 +404,10 @@ store_trace_args(int ent_size, struct trace_probe *tp, struct pt_regs *regs,
 }
 
 extern int set_print_fmt(struct trace_probe *tp, bool is_return);
+
+#ifdef CONFIG_PERF_EVENTS
+extern struct trace_event_call *
+create_local_trace_kprobe(char *func, void *addr, unsigned long offs,
+			  bool is_return);
+extern void destroy_local_trace_kprobe(struct trace_event_call *event_call);
+#endif