tracing/filters: Support specifying filter hook to a TRACE_EVENT field

Add __field_ext(), so a field can be assigned to a specific filter_type, which matches a corresponding filter function.

For example, a later patch will allow this:

	__field_ext(const char *, str, FILTER_PTR_STR);

Signed-off-by Li Zefan <[email protected]>

---
 include/linux/ftrace_event.h       |   11 +++++++++--
 include/trace/ftrace.h             |   28 ++++++++++++++++++++++------
 kernel/trace/trace_events.c        |    9 +++++++--
 kernel/trace/trace_events_filter.c |    6 ------
 kernel/trace/trace_export.c        |    7 ++++---
 5 files changed, 42 insertions(+), 19 deletions(-)

include/linux/ftrace_event.h +0 -0

index 26d3673..14c388e 100644

include/linux/ftrace_event.h +4 -2

+
 extern int trace_define_field(struct ftrace_event_call *call, char *type,
-			      char *name, int offset, int size, int is_signed);
+			      char *name, int offset, int size,
+			      int is_signed, int filter_type);
 
 #define is_signed_type(type)	(((type)(-1)) < 0)
 

@@ -167,7 +174,7 @@ do {									\
 #define __common_field(type, item, is_signed)				\
 	ret = trace_define_field(event_call, #type, "common_" #item,	\
 				 offsetof(typeof(field.ent), item),	\
-				 sizeof(field.ent.item), is_signed);	\
+				 sizeof(field.ent.item), is_signed, -1);\
 	if (ret)							\
 		return ret;
 

include/trace/ftrace.h +3 -0

index 380b603..8d8518f 100644
@@ -21,6 +21,9 @@
 #undef __field
 #define __field(type, item)		type	item;
 
+#undef __field_ext
+#define __field_ext(type, item, filter_type)	type	item;
+
 #undef __array
 #define __array(type, item, len)	type	item[len];
 

include/trace/ftrace.h +6 -2

+
 #undef __array
 #define __array(type, item, len)
 

@@ -110,6 +116,9 @@
 	if (!ret)							\
 		return 0;
 
+#undef __field_ext
+#define __field_ext(type, item, filter_type)	__field(type, item)
+
 #undef __array
 #define __array(type, item, len)						\
 	ret = trace_seq_printf(s, "\tfield:" #type " " #item "[" #len "];\t"	\

@@ -264,28 +273,32 @@ ftrace_raw_output_##call(struct trace_iterator *iter, int flags)	\
 
 #include TRACE_INCLUDE(TRACE_INCLUDE_FILE)
 
-#undef __field
-#define __field(type, item)						\
+#undef __field_ext
+#define __field_ext(type, item, filter_type)				\

include/trace/ftrace.h +2 -1

+
 #undef __array
 #define __array(type, item, len)					\
 	BUILD_BUG_ON(len > MAX_FILTER_STR_VAL);				\
 	ret = trace_define_field(event_call, #type "[" #len "]", #item,	\
 				 offsetof(typeof(field), item),		\
-				 sizeof(field.item), 0);		\
+				 sizeof(field.item), 0, -1);		\

include/trace/ftrace.h +3 -0

 #undef __string
 #define __string(item, src) __dynamic_array(char, item, -1)

@@ -322,6 +335,9 @@ ftrace_define_fields_##call(void)					\
 #undef __field
 #define __field(type, item)
 
+#undef __field_ext
+#define __field_ext(type, item, filter_type)
+
 #undef __array
 #define __array(type, item, len)
 

kernel/trace/trace_events.c +7 -2

index 71b2085..0cbad89 100644
@@ -28,7 +28,8 @@ DEFINE_MUTEX(event_mutex);
 LIST_HEAD(ftrace_events);
 
 int trace_define_field(struct ftrace_event_call *call, char *type,
-		       char *name, int offset, int size, int is_signed)
+		       char *name, int offset, int size,
+		       int is_signed, int filter_type)
 {
 	struct ftrace_event_field *field;
 

@@ -44,7 +45,11 @@ int trace_define_field(struct ftrace_event_call *call, char *type,
 	if (!field->type)
 		goto err;
 
-	field->filter_type = filter_assign_type(type);
+	if (filter_type == -1)
+		field->filter_type = filter_assign_type(type);
+	else
+		field->filter_type = filter_type;
+
 	field->offset = offset;
 	field->size = size;
 	field->is_signed = is_signed;

kernel/trace/trace_events_filter.c +0 -6

index 4c8c5fd..5e7f031 100644
@@ -475,12 +475,6 @@ static int filter_add_pred_fn(struct filter_parse_state *ps,
 	return 0;
 }
 
-enum {
-	FILTER_STATIC_STRING = 1,
-	FILTER_DYN_STRING,
-	FILTER_OTHER,
-};
-
 int filter_assign_type(const char *type)
 {
 	if (strstr(type, "__data_loc") && strstr(type, "char"))

kernel/trace/trace_export.c +4 -3

index d06cf89..4a9f273 100644
@@ -156,7 +156,8 @@ __attribute__((section("_ftrace_events"))) event_##call = {		\
 #define TRACE_FIELD(type, item, assign)					\
 	ret = trace_define_field(event_call, #type, #item,		\
 				 offsetof(typeof(field), item),		\
-				 sizeof(field.item), is_signed_type(type));	\
+				 sizeof(field.item),			\
+				 is_signed_type(type), -1);		\
 	if (ret)							\
 		return ret;
 

@@ -164,7 +165,7 @@ __attribute__((section("_ftrace_events"))) event_##call = {		\
 #define TRACE_FIELD_SPECIAL(type, item, len, cmd)			\
 	ret = trace_define_field(event_call, #type "[" #len "]", #item,	\
 				 offsetof(typeof(field), item),		\
-				 sizeof(field.item), 0);		\
+				 sizeof(field.item), 0, -1);		\
 	if (ret)							\
 		return ret;
 

@@ -172,7 +173,7 @@ __attribute__((section("_ftrace_events"))) event_##call = {		\
 #define TRACE_FIELD_SIGN(type, item, assign, is_signed)			\
 	ret = trace_define_field(event_call, #type, #item,		\
 				 offsetof(typeof(field), item),		\
-				 sizeof(field.item), is_signed);	\
+				 sizeof(field.item), is_signed, -1);	\
 	if (ret)							\
 		return ret;
 

Signature · 1 line

1.6.3

Usually, char * entries are dangerous in traces because the string can be released whereas a pointer to it can still wait to be read from the ring buffer.

But sometimes we can assume it's safe, like in case of RO data (eg: __file__ or __line__, used in bkl trace event). If these RO data are in a module and so is the call to the trace event, then it's safe, because the ring buffer will be flushed once this module get unloaded.

To allow char * to be treated as a string:

	TRACE_EVENT(...,

		TP_STRUCT__entry(
			__field_ext(const char *, name, FILTER_PTR_STRING)
			...
		)

		...
	);

Signed-off-by Li Zefan <[email protected]>

---
 include/linux/ftrace_event.h       |    1 +
 kernel/trace/trace_events_filter.c |   26 +++++++++++++++++++++++---
 2 files changed, 24 insertions(+), 3 deletions(-)

include/linux/ftrace_event.h +1 -0

index 14c388e..1a98a61 100644
@@ -141,6 +141,7 @@ extern int filter_current_check_discard(struct ftrace_event_call *call,
 enum {
 	FILTER_STATIC_STRING = 1,
 	FILTER_DYN_STRING,
+	FILTER_PTR_STRING,
 	FILTER_OTHER,
 };
 

kernel/trace/trace_events_filter.c +18 -1

index 5e7f031..b16923e 100644
@@ -163,6 +163,20 @@ static int filter_pred_string(struct filter_pred *pred, void *event,
 	return match;
 }
 
+/* Filter predicate for char * pointers */
+static int filter_pred_pchar(struct filter_pred *pred, void *event,
+			     int val1, int val2)
+{
+	char **addr = (char **)(event + pred->offset);
+	int cmp, match;
+
+	cmp = strncmp(*addr, pred->str_val, pred->str_len);
+
+	match = (!cmp) ^ pred->not;
+
+	return match;
+}
+
 /*
  * Filter predicate for dynamic sized arrays of characters.
  * These are implemented through a list of strings at the end

@@ -489,7 +503,8 @@ int filter_assign_type(const char *type)
 static bool is_string_field(struct ftrace_event_field *field)
 {
 	return field->filter_type == FILTER_DYN_STRING ||
-	       field->filter_type == FILTER_STATIC_STRING;
+	       field->filter_type == FILTER_STATIC_STRING ||
+	       field->filter_type == FILTER_PTR_STRING;
 }
 
 static int is_legal_op(struct ftrace_event_field *field, int op)

@@ -579,11 +594,16 @@ static int filter_add_pred(struct filter_parse_state *ps,
 	}
 
 	if (is_string_field(field)) {
+		pred->str_len = field->size;
+

kernel/trace/trace_events_filter.c +0 -0

 	} else {
 		if (field->is_signed)
 			ret = strict_strtoll(pred->str_val, 0, &val);

Signature · 1 line

1.6.3

Currently only static strings and dynamic strings have their own filter functions, other types of field are treated as integers.

This patchset allows assigning a specific filter type to a field, so a field which is defined as:

__field_ext(const char *, str, FILTER_PTR_STR)

will be treated as a string but not a plain pointer, and then we can set the filter like this:

	# echo 'str == foo' > filter

And it's easy to add more filter functions for different types to turn these into valid operations:

	(dev is of type dev_t)
	# echo 'dev == 8:0' > filter

	(callsite is of type void * or unsigned long)
	# echo 'callsite == skb_free' > filter

[PATCH 1/3] tracing/filters: Add filter_type to struct ftrace_event_field [PATCH 2/3] tracing/filters: Add __field_ext() to TRACE_EVENT [PATCH 3/3] tracing/filters: Support filtering for char * strings

---
 include/linux/ftrace_event.h       |   12 +++++++-
 include/trace/ftrace.h             |   28 ++++++++++++++++----
 kernel/trace/trace.h               |    2 +
 kernel/trace/trace_events.c        |    9 ++++++-
 kernel/trace/trace_events_filter.c |   47 +++++++++++++++++++++++++----------
 kernel/trace/trace_export.c        |    7 +++--
 6 files changed, 79 insertions(+), 26 deletions(-)

Quoted 3 lines

[PATCH 1/3] tracing/filters: Add filter_type to struct ftrace_event_field [PATCH 2/3] tracing/filters: Add __field_ext() to TRACE_EVENT [PATCH 3/3] tracing/filters: Support filtering for char * strings

I like the first two patches, but would like to see a V2 with the updates I've mentioned. The last patch I do not fully understand its purpose.

I'll wait for V2 to pull it in.

Quoted 8 lines

kernel/trace/trace_events_filter.c +3 -1

 	if (is_string_field(field)) {
+		pred->str_len = field->size;
+
 		if (field->filter_type == FILTER_STATIC_STRING)
 			fn = filter_pred_string;
-		else
+		else if (field->filter_type == FILTER_DYN_STRING)
 			fn = filter_pred_strloc;

kernel/trace/trace_events_filter.c static int filter_add_pred(struct filter_parse_state *ps, · line 606

Quoted 5 lines

kernel/trace/trace_events_filter.c +4 -1

-		pred->str_len = field->size;
+		else {
+			fn = filter_pred_pchar;
+			pred->str_len = strlen(pred->str_val);
+		}

I'm a little dense here, where do we protect against someone making a tracepoint that points to unsafe data?

Quoted 8 lines

Quoted 8 lines

kernel/trace/trace_events_filter.c +3 -1

 	if (is_string_field(field)) {
+		pred->str_len = field->size;
+
 		if (field->filter_type == FILTER_STATIC_STRING)
 			fn = filter_pred_string;
-		else
+		else if (field->filter_type == FILTER_DYN_STRING)
 			fn = filter_pred_strloc;

Quoted 9 lines

kernel/trace/trace_events_filter.c static int filter_add_pred(struct filter_parse_state *ps, · line 606

Quoted 5 lines

kernel/trace/trace_events_filter.c +4 -1

-		pred->str_len = field->size;
+		else {
+			fn = filter_pred_pchar;
+			pred->str_len = strlen(pred->str_val);
+		}

I'm a little dense here, where do we protect against someone making a tracepoint that points to unsafe data?

We can't prevent anyone from doing insane things deliberately, but we prevent from doing wrong things unconsciously.

Only if a TRACE_EVENT has a field defined as:

__field_ext(char *, name, FILTER_PTR_STR)

Here using FILTER_PTR_STR explicitly, he should know what he's doing.

Anyway, he can make a ptr pointing to unsafe data this way:

	TP_STRUCT__entry(
		__field(char *, name)
	)
	TP_printk("%s", name)

Quoted 23 lines · Li Zefan

Quoted 7 lines

Quoted 7 lines

kernel/trace/trace_events_filter.c +3 -1

 	if (is_string_field(field)) {
+		pred->str_len = field->size;
+
 		if (field->filter_type == FILTER_STATIC_STRING)
 			fn = filter_pred_string;
-		else
+		else if (field->filter_type == FILTER_DYN_STRING)

Quoted 9 lines

kernel/trace/trace_events_filter.c static int filter_add_pred(struct filter_parse_state *ps, · line 606

Quoted 6 lines

kernel/trace/trace_events_filter.c +4 -1

 			fn = filter_pred_strloc;
-		pred->str_len = field->size;
+		else {
+			fn = filter_pred_pchar;
+			pred->str_len = strlen(pred->str_val);
+		}

I'm a little dense here, where do we protect against someone making a tracepoint that points to unsafe data?

We can't prevent anyone from doing insane things deliberately, but we prevent from doing wrong things unconsciously.

Only if a TRACE_EVENT has a field defined as:

__field_ext(char *, name, FILTER_PTR_STR)

Quoted 8 lines

Here using FILTER_PTR_STR explicitly, he should know what he's doing.

Anyway, he can make a ptr pointing to unsafe data this way:

	TP_STRUCT__entry(
		__field(char *, name)
	)
	TP_printk("%s", name)

I guess the thing I'm missing is what's the difference of the two? Why would a developer use __field_ext instead of doing it the unsafe way of just __field?

I guess I don't see the developer doing something wrong unconsciously. Well maybe I don't see this making the developer do it right unconsciously.

What protection is this giving us?

Quoted 3 lines

What protection is this giving us?

__field(char *) suggests it should be treated as plain pointer, while __field_ext(char *, FILTER_PTR_STR) suggests he's aware it's safe to dereference the pointer, for example the case in Frederic's blk events.

In Frederic's initial version, "char *" field will always be attached to ptr_str filter function. This is unsafe, because for other fields defined as "char *" but not safe to dereference, a user still can do this:

	# echo 'name == abc' > filter

Then we'll deref a pointer that can point to unsafe data.

In this patch, this won't happen, as long as the developer is aware that his use of __field_ext(char *) is right.

Otherwise, he will just use normal __field(char *) and print the pointer itself in TP_printk().

Quoted 14 lines · Li Zefan

Quoted 2 lines

What protection is this giving us?

__field(char *) suggests it should be treated as plain pointer, while __field_ext(char *, FILTER_PTR_STR) suggests he's aware it's safe to dereference the pointer, for example the case in Frederic's blk events.

In Frederic's initial version, "char *" field will always be attached to ptr_str filter function. This is unsafe, because for other fields defined as "char *" but not safe to dereference, a user still can do this:

	# echo 'name == abc' > filter

Quoted 7 lines

Then we'll deref a pointer that can point to unsafe data.

In this patch, this won't happen, as long as the developer is aware that his use of __field_ext(char *) is right.

Otherwise, he will just use normal __field(char *) and print the pointer itself in TP_printk().

Ah, so the answer I'm looking for is:

The filtering will not dereference "char *" unless the developer explicitly sets FILTER_PTR_STR in __field_ext.

Is this above statement correct?

Quoted 7 lines · Steven Rostedt

Ah, so the answer I'm looking for is:

The filtering will not dereference "char *" unless the developer explicitly sets FILTER_PTR_STR in __field_ext.

Is this above statement correct?

Exactly. Sorry that I didn't explain it clearly in the first place. :)