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}'
Signed-off-by: Valentin Schneider <[email protected]>
---
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(-)
diff --git a/kernel/trace/trace_events_filter.c b/kernel/trace/trace_events_filter.c
index 8af00caa363f7..99e111c237a93 100644
--- a/kernel/trace/trace_events_filter.c
+++ b/kernel/trace/trace_events_filter.c
@@ -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] == '"') {
--
2.31.1
On Wed, 5 Jul 2023 19:12:45 +0100
Valentin Schneider <[email protected]> wrote:
> +/* 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));
The above two can be optimized much more if the cpu is a scalar. If mask is
anything but a single CPU, then EQ will always be false, and NE will always
be true. If the mask contains a single CPU, then we should convert the mask
into the scalar CPU and just do:
case OP_EQ:
return mask_cpu == cpu;
case op_NE:
return maks_cpu != cpu;
-- Steve
> + case OP_BAND:
> + return cpumask_test_cpu(cpu, mask);
> + default:
> + return 0;
> + }
> +}