Skip to contentSkip to conversation

tracing/filters: filtering event fields with a cpumask

RFC PATCH New Valentin Schneider 1 participant 6 messages Jun 9, 2023

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover tracing/filters: filtering event fields with a cpumask

Hi folks,

In the context of CPU isolation / NOHZ_FULL interference investigation, we now have the ipi_send_cpu and ipi_send_cpumask events. However, enabling these events without any filtering can yield pretty massive traces with a lot of uninteresting or irrelevant data (e.g. everything targeting housekeeping CPUs).

This series is about adding event filtering via a user-provided cpumask. This enables filtering using cpumask fields (e.g. ipi_send_cpumask) and extends this to scalar and the local CPU common fields.

With this, it becomes fairly easy to trace events both happening on and targeting CPUs of interest, e.g.:

  trace-cmd record -e 'sched_switch' -f "CPU & MASK{$ISOLATED_CPUS}"         \
		   -e 'sched_wakeup' -f "target_cpu & MASK{$ISOLATED_CPUS}"  \
		   -e 'ipi_send_cpu' -f "cpu & MASK{$ISOLATED_CPUS}"         \
		   -e 'ipi_send_cpumask' -f "cpumask & MASK{$ISOLATED_CPUS}" \
		   hackbench

The MASK{} thing is a bit crude but seems to work well enough without break^C overhauling the predicate parsing logic.

Cheers,
Valentin

Valentin Schneider (5):
tracing/filters: Dynamically allocate filter_pred.regex tracing/filters: Enable filtering a cpumask field by another cpumask tracing/filters: Enable filtering a scalar field by a cpumask tracing/filters: Enable filtering the CPU common field by a cpumask tracing/filters: Document cpumask filtering

 Documentation/trace/events.rst     |  14 ++
 include/linux/trace_events.h       |   1 +
 kernel/trace/trace_events_filter.c | 228 +++++++++++++++++++++++++----
 3 files changed, 216 insertions(+), 27 deletions(-)

--
2.31.1

#1 lore raw

RFC PATCH 1/5 tracing/filters: Dynamically allocate filter_pred.regex

Every predicate allocation includes a MAX_FILTER_STR_VAL (256) char array in the regex field, even if the predicate function does not use the field.

A later commit will introduce a dynamically allocated cpumask to struct filter_pred, which will require a dedicated freeing function. Bite the bullet and make filter_pred.regex dynamically allocated.

While at it, reorder the fields of filter_pred to fill in the byte holes. The struct now fits on a single cacheline.

No change in behaviour intended.

The kfree()'s were patched via Coccinelle:
@@
struct filter_pred *pred;
@@

  -kfree(pred);
  +free_predicate(pred);
---
 kernel/trace/trace_events_filter.c | 62 ++++++++++++++++++------------
 1 file changed, 37 insertions(+), 25 deletions(-)
kernel/trace/trace_events_filter.c +37 -25
index 1dad64267878c..d999a218fe833 100644
@@ -70,15 +70,15 @@ enum filter_pred_fn {
 };
 
 struct filter_pred {
-	enum filter_pred_fn 	fn_num;
-	u64 			val;
-	u64 			val2;
-	struct regex		regex;
+	struct regex		*regex;
 	unsigned short		*ops;
 	struct ftrace_event_field *field;
-	int 			offset;
+	u64			val;
+	u64			val2;
+	enum filter_pred_fn	fn_num;
+	int			offset;
 	int			not;
-	int 			op;
+	int			op;
 };
 
 /*
@@ -186,6 +186,12 @@ enum {
 	PROCESS_OR	= 4,
 };
 
+static void free_predicate(struct filter_pred *pred)
+{
+	kfree(pred->regex);
+	kfree(pred);
+}
+
 /*
  * Without going into a formal proof, this explains the method that is used in
  * parsing the logical expressions.
@@ -623,7 +629,7 @@ predicate_parse(const char *str, int nr_parens, int nr_preds,
 	kfree(inverts);
 	if (prog_stack) {
 		for (i = 0; prog_stack[i].pred; i++)
-			kfree(prog_stack[i].pred);
+			free_predicate(prog_stack[i].pred);
 		kfree(prog_stack);
 	}
 	return ERR_PTR(ret);
@@ -750,7 +756,7 @@ static int filter_pred_string(struct filter_pred *pred, void *event)
 	char *addr = (char *)(event + pred->offset);
 	int cmp, match;
 
-	cmp = pred->regex.match(addr, &pred->regex, pred->regex.field_len);
+	cmp = pred->regex->match(addr, pred->regex, pred->regex->field_len);
 
 	match = cmp ^ pred->not;
 
@@ -763,7 +769,7 @@ static __always_inline int filter_pchar(struct filter_pred *pred, char *str)
 	int len;
 
 	len = strlen(str) + 1;	/* including tailing '\0' */
-	cmp = pred->regex.match(str, &pred->regex, len);
+	cmp = pred->regex->match(str, pred->regex, len);
 
 	match = cmp ^ pred->not;
 
@@ -813,7 +819,7 @@ static int filter_pred_strloc(struct filter_pred *pred, void *event)
 	char *addr = (char *)(event + str_loc);
 	int cmp, match;
 
-	cmp = pred->regex.match(addr, &pred->regex, str_len);
+	cmp = pred->regex->match(addr, pred->regex, str_len);
 
 	match = cmp ^ pred->not;
 
@@ -836,7 +842,7 @@ static int filter_pred_strrelloc(struct filter_pred *pred, void *event)
 	char *addr = (char *)(&item[1]) + str_loc;
 	int cmp, match;
 
-	cmp = pred->regex.match(addr, &pred->regex, str_len);
+	cmp = pred->regex->match(addr, pred->regex, str_len);
 
 	match = cmp ^ pred->not;
 
@@ -874,7 +880,7 @@ static int filter_pred_comm(struct filter_pred *pred, void *event)
 {
 	int cmp;
 
-	cmp = pred->regex.match(current->comm, &pred->regex,
+	cmp = pred->regex->match(current->comm, pred->regex,
 				TASK_COMM_LEN);
 	return cmp ^ pred->not;
 }
@@ -1004,7 +1010,7 @@ enum regex_type filter_parse_regex(char *buff, int len, char **search, int *not)
 
 static void filter_build_regex(struct filter_pred *pred)
 {
-	struct regex *r = &pred->regex;
+	struct regex *r = pred->regex;
 	char *search;
 	enum regex_type type = MATCH_FULL;
 
@@ -1169,7 +1175,7 @@ static void free_prog(struct event_filter *filter)
 		return;
 
 	for (i = 0; prog[i].pred; i++)
-		kfree(prog[i].pred);
+		free_predicate(prog[i].pred);
 	kfree(prog);
 }
 
@@ -1553,9 +1559,12 @@ static int parse_pred(const char *str, void *data,
 			goto err_free;
 		}
 
-		pred->regex.len = len;
-		strncpy(pred->regex.pattern, str + s, len);
-		pred->regex.pattern[len] = 0;
+		pred->regex = kzalloc(sizeof(*pred->regex), GFP_KERNEL);
+		if (!pred->regex)
+			goto err_mem;
+		pred->regex->len = len;
+		strncpy(pred->regex->pattern, str + s, len);
+		pred->regex->pattern[len] = 0;
 
 	/* This is either a string, or an integer */
 	} else if (str[i] == '\'' || str[i] == '"') {
@@ -1597,9 +1606,12 @@ static int parse_pred(const char *str, void *data,
 			goto err_free;
 		}
 
-		pred->regex.len = len;
-		strncpy(pred->regex.pattern, str + s, len);
-		pred->regex.pattern[len] = 0;
+		pred->regex = kzalloc(sizeof(*pred->regex), GFP_KERNEL);
+		if (!pred->regex)
+			goto err_mem;
+		pred->regex->len = len;
+		strncpy(pred->regex->pattern, str + s, len);
+		pred->regex->pattern[len] = 0;
 
 		filter_build_regex(pred);
 
@@ -1608,7 +1620,7 @@ static int parse_pred(const char *str, void *data,
 
 		} else if (field->filter_type == FILTER_STATIC_STRING) {
 			pred->fn_num = FILTER_PRED_FN_STRING;
-			pred->regex.field_len = field->size;
+			pred->regex->field_len = field->size;
 
 		} else if (field->filter_type == FILTER_DYN_STRING) {
 			pred->fn_num = FILTER_PRED_FN_STRLOC;
@@ -1691,10 +1703,10 @@ static int parse_pred(const char *str, void *data,
 	return i;
 
 err_free:
-	kfree(pred);
+	free_predicate(pred);
 	return -EINVAL;
 err_mem:
-	kfree(pred);
+	free_predicate(pred);
 	return -ENOMEM;
 }
 
@@ -2287,8 +2299,8 @@ static int ftrace_function_set_filter_pred(struct filter_pred *pred,
 		return ret;
 
 	return __ftrace_function_set_filter(pred->op == OP_EQ,
-					    pred->regex.pattern,
-					    pred->regex.len,
+					    pred->regex->pattern,
+					    pred->regex->len,
 					    data);
 }
 
Signature · 1 line
2.31.1
#2 lore raw

RFC PATCH 2/5 tracing/filters: Enable filtering a cpumask field by another cpumask

In reply to Valentin Schneider

The recently introduced ipi_send_cpumask trace event contains a cpumask field, but it currently cannot be used in filter expressions.

Make event filtering aware of cpumask fields, and allow these to be filtered by a user-provided cpumask.

The user-provided cpumask is to be given in cpulist format and wrapped as:
"MASK{$cpulist}". The use of curly braces instead of parentheses is to prevent predicate_parse() from parsing the contents of MASK{...} as a full-fledged predicate subexpression.

This enables e.g.:

$ trace-cmd record -e 'ipi_send_cpumask' -f 'cpu & MASK{2,4,6,8-32}'
---
 include/linux/trace_events.h       |  1 +
 kernel/trace/trace_events_filter.c | 85 +++++++++++++++++++++++++++++-
 2 files changed, 84 insertions(+), 2 deletions(-)
include/linux/trace_events.h +1 -0
index 7c4a0b72334eb..974ef37a06c83 100644
@@ -804,6 +804,7 @@ enum {
 	FILTER_RDYN_STRING,
 	FILTER_PTR_STRING,
 	FILTER_TRACE_FN,
+	FILTER_CPUMASK,
 	FILTER_COMM,
 	FILTER_CPU,
 	FILTER_STACKTRACE,
kernel/trace/trace_events_filter.c +83 -2

139 lines not on this page · raw message

Signature · 1 line
2.31.1
#3 lore raw

RFC PATCH 3/5 tracing/filters: Enable filtering a scalar field by a cpumask

In reply to Valentin Schneider

Several events use a scalar field to denote a CPU:
o sched_wakeup.target_cpu
o sched_migrate_task.orig_cpu,dest_cpu
o sched_move_numa.src_cpu,dst_cpu
o ipi_send_cpu.cpu
o ...

Filtering these currently requires using arithmetic comparison functions, which can be tedious when dealing with interleaved SMT or NUMA CPU ids.

Allow these to be filtered by a user-provided cpumask, which enables e.g.:

$ trace-cmd record -e 'sched_wakeup' -f 'target_cpu & MASK{2,4,6,8-32}'

---
NOTE: I went with an implicit cpumask conversion of the event field, as AFAICT predicate_parse() does not support parsing the application of a function to a field (e.g. 'MASK(target_cpu) & MASK{2,4,6,8-32}')

---
 kernel/trace/trace_events_filter.c | 92 ++++++++++++++++++++++++++----
 1 file changed, 80 insertions(+), 12 deletions(-)
kernel/trace/trace_events_filter.c +80 -12
index 8af00caa363f7..99e111c237a93 100644
@@ -46,15 +46,19 @@ static const char * ops[] = { OPS };
 enum filter_pred_fn {
 	FILTER_PRED_FN_NOP,
 	FILTER_PRED_FN_64,
+	FILTER_PRED_FN_64_CPUMASK,
 	FILTER_PRED_FN_S64,
 	FILTER_PRED_FN_U64,
 	FILTER_PRED_FN_32,
+	FILTER_PRED_FN_32_CPUMASK,
 	FILTER_PRED_FN_S32,
 	FILTER_PRED_FN_U32,
 	FILTER_PRED_FN_16,
+	FILTER_PRED_FN_16_CPUMASK,
 	FILTER_PRED_FN_S16,
 	FILTER_PRED_FN_U16,
 	FILTER_PRED_FN_8,
+	FILTER_PRED_FN_8_CPUMASK,
 	FILTER_PRED_FN_S8,
 	FILTER_PRED_FN_U8,
 	FILTER_PRED_FN_COMM,
@@ -641,6 +645,37 @@ predicate_parse(const char *str, int nr_parens, int nr_preds,
 	return ERR_PTR(ret);
 }
 
+static inline int
+do_filter_cpumask(int op, const struct cpumask *mask, const struct cpumask *cmp)
+{
+	switch (op) {
+	case OP_EQ:
+		return cpumask_equal(mask, cmp);
+	case OP_NE:
+		return !cpumask_equal(mask, cmp);
+	case OP_BAND:
+		return cpumask_intersects(mask, cmp);
+	default:
+		return 0;
+	}
+}
+
+/* Optimisation of do_filter_cpumask() for scalar values */
+static inline int
+do_filter_cpumask_scalar(int op, unsigned int cpu, const struct cpumask *mask)
+{
+	switch (op) {
+	case OP_EQ:
+		return cpumask_equal(mask, cpumask_of(cpu));
+	case OP_NE:
+		return !cpumask_equal(mask, cpumask_of(cpu));
+	case OP_BAND:
+		return cpumask_test_cpu(cpu, mask);
+	default:
+		return 0;
+	}
+}
+
 enum pred_cmp_types {
 	PRED_CMP_TYPE_NOP,
 	PRED_CMP_TYPE_LT,
@@ -684,6 +719,18 @@ static int filter_pred_##type(struct filter_pred *pred, void *event)	\
 	}								\
 }
 
+#define DEFINE_CPUMASK_COMPARISON_PRED(size)					\
+static int filter_pred_##size##_cpumask(struct filter_pred *pred, void *event)	\
+{										\
+	u##size *addr = (u##size *)(event + pred->offset);			\
+	unsigned int cpu = *addr;						\
+										\
+	if (cpu >= nr_cpu_ids)							\
+		return 0;							\
+										\
+	return do_filter_cpumask_scalar(pred->op, cpu, pred->mask);		\
+}
+
 #define DEFINE_EQUALITY_PRED(size)					\
 static int filter_pred_##size(struct filter_pred *pred, void *event)	\
 {									\
@@ -705,6 +752,11 @@ DEFINE_COMPARISON_PRED(u16);
 DEFINE_COMPARISON_PRED(s8);
 DEFINE_COMPARISON_PRED(u8);
 
+DEFINE_CPUMASK_COMPARISON_PRED(64);
+DEFINE_CPUMASK_COMPARISON_PRED(32);
+DEFINE_CPUMASK_COMPARISON_PRED(16);
+DEFINE_CPUMASK_COMPARISON_PRED(8);
+
 DEFINE_EQUALITY_PRED(64);
 DEFINE_EQUALITY_PRED(32);
 DEFINE_EQUALITY_PRED(16);
@@ -889,16 +941,7 @@ static int filter_pred_cpumask(struct filter_pred *pred, void *event)
 	const struct cpumask *mask = (event + loc);
 	const struct cpumask *cmp = pred->mask;
 
-	switch (pred->op) {
-	case OP_EQ:
-		return cpumask_equal(mask, cmp);
-	case OP_NE:
-		return !cpumask_equal(mask, cmp);
-	case OP_BAND:
-		return cpumask_intersects(mask, cmp);
-	default:
-		return 0;
-	}
+	return do_filter_cpumask(pred->op, mask, cmp);
 }
 
 /* Filter predicate for COMM. */
@@ -1349,24 +1392,32 @@ static int filter_pred_fn_call(struct filter_pred *pred, void *event)
 	switch (pred->fn_num) {
 	case FILTER_PRED_FN_64:
 		return filter_pred_64(pred, event);
+	case FILTER_PRED_FN_64_CPUMASK:
+		return filter_pred_64_cpumask(pred, event);
 	case FILTER_PRED_FN_S64:
 		return filter_pred_s64(pred, event);
 	case FILTER_PRED_FN_U64:
 		return filter_pred_u64(pred, event);
 	case FILTER_PRED_FN_32:
 		return filter_pred_32(pred, event);
+	case FILTER_PRED_FN_32_CPUMASK:
+		return filter_pred_32_cpumask(pred, event);
 	case FILTER_PRED_FN_S32:
 		return filter_pred_s32(pred, event);
 	case FILTER_PRED_FN_U32:
 		return filter_pred_u32(pred, event);
 	case FILTER_PRED_FN_16:
 		return filter_pred_16(pred, event);
+	case FILTER_PRED_FN_16_CPUMASK:
+		return filter_pred_16_cpumask(pred, event);
 	case FILTER_PRED_FN_S16:
 		return filter_pred_s16(pred, event);
 	case FILTER_PRED_FN_U16:
 		return filter_pred_u16(pred, event);
 	case FILTER_PRED_FN_8:
 		return filter_pred_8(pred, event);
+	case FILTER_PRED_FN_8_CPUMASK:
+		return filter_pred_8_cpumask(pred, event);
 	case FILTER_PRED_FN_S8:
 		return filter_pred_s8(pred, event);
 	case FILTER_PRED_FN_U8:
@@ -1602,7 +1653,8 @@ static int parse_pred(const char *str, void *data,
 		unsigned int maskstart;
 		char *tmp;
 
-		if (field->filter_type != FILTER_CPUMASK) {
+		if (field->filter_type != FILTER_CPUMASK &&
+		    field->filter_type != FILTER_OTHER) {
 			parse_error(pe, FILT_ERR_ILLEGAL_FIELD_OP, pos + i);
 			goto err_free;
 		}
@@ -1644,8 +1696,24 @@ static int parse_pred(const char *str, void *data,
 
 		/* Move along */
 		i++;
-		if (field->filter_type == FILTER_CPUMASK)
+		if (field->filter_type == FILTER_CPUMASK) {
 			pred->fn_num = FILTER_PRED_FN_CPUMASK;
+		} else {
+			switch (field->size) {
+			case 8:
+				pred->fn_num = FILTER_PRED_FN_64_CPUMASK;
+				break;
+			case 4:
+				pred->fn_num = FILTER_PRED_FN_32_CPUMASK;
+				break;
+			case 2:
+				pred->fn_num = FILTER_PRED_FN_16_CPUMASK;
+				break;
+			case 1:
+				pred->fn_num = FILTER_PRED_FN_8_CPUMASK;
+				break;
+			}
+		}
 
 	/* This is either a string, or an integer */
 	} else if (str[i] == '\'' || str[i] == '"') {
Signature · 1 line
2.31.1
#4 lore raw

RFC PATCH 4/5 tracing/filters: Enable filtering the CPU common field by a cpumask

In reply to Valentin Schneider

The tracing_cpumask lets us specify which CPUs are traced in a buffer instance, but doesn't let us do this on a per-event basis (unless one creates an instance per event).

A previous commit added filtering scalar fields by a user-given cpumask, make this work with the CPU common field as well.

This enables doing things like

$ trace-cmd record -e 'sched_switch' -f 'CPU & MASK{12-52}' \
		   -e 'sched_wakeup' -f 'target_cpu & MASK{12-52}'
---
 kernel/trace/trace_events_filter.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)
kernel/trace/trace_events_filter.c +13 -0
index 99e111c237a93..b3d2612d4670a 100644
@@ -68,6 +68,7 @@ enum filter_pred_fn {
 	FILTER_PRED_FN_PCHAR_USER,
 	FILTER_PRED_FN_PCHAR,
 	FILTER_PRED_FN_CPU,
+	FILTER_PRED_FN_CPU_CPUMASK,
 	FILTER_PRED_FN_CPUMASK,
 	FILTER_PRED_FN_FUNCTION,
 	FILTER_PRED_FN_,
@@ -933,6 +934,13 @@ static int filter_pred_cpu(struct filter_pred *pred, void *event)
 	}
 }
 
+static int filter_pred_cpu_cpumask(struct filter_pred *pred, void *event)
+{
+	int cpu = raw_smp_processor_id();
+
+	return do_filter_cpumask_scalar(pred->op, cpu, pred->mask);
+}
+
 /* Filter predicate for cpumasks. */
 static int filter_pred_cpumask(struct filter_pred *pred, void *event)
 {
@@ -1436,6 +1444,8 @@ static int filter_pred_fn_call(struct filter_pred *pred, void *event)
 		return filter_pred_pchar(pred, event);
 	case FILTER_PRED_FN_CPU:
 		return filter_pred_cpu(pred, event);
+	case FILTER_PRED_FN_CPU_CPUMASK:
+		return filter_pred_cpu_cpumask(pred, event);
 	case FILTER_PRED_FN_CPUMASK:
 		return filter_pred_cpumask(pred, event);
 	case FILTER_PRED_FN_FUNCTION:
@@ -1654,6 +1664,7 @@ static int parse_pred(const char *str, void *data,
 		char *tmp;
 
 		if (field->filter_type != FILTER_CPUMASK &&
+		    field->filter_type != FILTER_CPU &&
 		    field->filter_type != FILTER_OTHER) {
 			parse_error(pe, FILT_ERR_ILLEGAL_FIELD_OP, pos + i);
 			goto err_free;
@@ -1698,6 +1709,8 @@ static int parse_pred(const char *str, void *data,
 		i++;
 		if (field->filter_type == FILTER_CPUMASK) {
 			pred->fn_num = FILTER_PRED_FN_CPUMASK;
+		} else if (field->filter_type == FILTER_CPU) {
+			pred->fn_num = FILTER_PRED_FN_CPU_CPUMASK;
 		} else {
 			switch (field->size) {
 			case 8:
Signature · 1 line
2.31.1
#5 lore raw

RFC PATCH 5/5 tracing/filters: Document cpumask filtering

In reply to Valentin Schneider

Cpumask, scalar and CPU fields can now be filtered by a user-provided cpumask, document the syntax.

---
 Documentation/trace/events.rst | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
Documentation/trace/events.rst +14 -0
index f5fcb8e1218f6..e9bc9f23891a0 100644
@@ -219,6 +219,20 @@ the function "security_prepare_creds" and less than the end of that function.
 The ".function" postfix can only be attached to values of size long, and can only
 be compared with "==" or "!=".
 
+Cpumask fields or scalar fields that encode a CPU number can be filtered using
+a user-provided cpumask in cpulist format. The format is as follows::
+
+  MASK{$cpulist}
+
+Operators available to cpumask filtering are:
+
+& (intersection), ==, !=
+
+For example, this will filter events that have their .target_cpu field present
+in the given cpumask::
+
+  target_cpu & MASK{17-42}
+
 5.2 Setting filters
 -------------------
 
Signature · 1 line
2.31.1