selftests/bpf: support stat filtering in comparison mode in veristat

Finally add support for filtering stats values, similar to
non-comparison mode filtering. For comparison mode 4 variants of stats
are important for filtering, as they allow to filter either A or B side,
but even more importantly they allow to filter based on value
difference, and for verdict stat value difference is MATCH/MISMATCH
classification. So with these changes it's finally possible to easily
check if there were any mismatches between failure/success outcomes on
two separate data sets. Like in an example below:

  $ ./veristat -e file,prog,verdict,insns -C ~/baseline-results.csv ~/shortest-results.csv -f verdict_diff=mismatch
  File                                   Program                Verdict (A)  Verdict (B)  Verdict (DIFF)  Insns (A)  Insns (B)  Insns        (DIFF)
  -------------------------------------  ---------------------  -----------  -----------  --------------  ---------  ---------  -------------------
  dynptr_success.bpf.linked1.o           test_data_slice        success      failure      MISMATCH               85          0       -85 (-100.00%)
  dynptr_success.bpf.linked1.o           test_read_write        success      failure      MISMATCH             1992          0     -1992 (-100.00%)
  dynptr_success.bpf.linked1.o           test_ringbuf           success      failure      MISMATCH               74          0       -74 (-100.00%)
  kprobe_multi.bpf.linked1.o             test_kprobe            failure      success      MISMATCH                0        246      +246 (+100.00%)
  kprobe_multi.bpf.linked1.o             test_kprobe_manual     failure      success      MISMATCH                0        246      +246 (+100.00%)
  kprobe_multi.bpf.linked1.o             test_kretprobe         failure      success      MISMATCH                0        248      +248 (+100.00%)
  kprobe_multi.bpf.linked1.o             test_kretprobe_manual  failure      success      MISMATCH                0        248      +248 (+100.00%)
  kprobe_multi.bpf.linked1.o             trigger                failure      success      MISMATCH                0          2        +2 (+100.00%)
  netcnt_prog.bpf.linked1.o              bpf_nextcnt            failure      success      MISMATCH                0         56       +56 (+100.00%)
  pyperf600_nounroll.bpf.linked1.o       on_event               success      failure      MISMATCH           568128    1000001    +431873 (+76.02%)
  ringbuf_bench.bpf.linked1.o            bench_ringbuf          success      failure      MISMATCH                8          0        -8 (-100.00%)
  strobemeta.bpf.linked1.o               on_event               success      failure      MISMATCH           557149    1000001    +442852 (+79.49%)
  strobemeta_nounroll1.bpf.linked1.o     on_event               success      failure      MISMATCH            57240    1000001  +942761 (+1647.03%)
  strobemeta_nounroll2.bpf.linked1.o     on_event               success      failure      MISMATCH           501725    1000001    +498276 (+99.31%)
  strobemeta_subprogs.bpf.linked1.o      on_event               success      failure      MISMATCH            65420    1000001  +934581 (+1428.59%)
  test_map_in_map_invalid.bpf.linked1.o  xdp_noop0              success      failure      MISMATCH                2          0        -2 (-100.00%)
  test_mmap.bpf.linked1.o                test_mmap              success      failure      MISMATCH               46          0       -46 (-100.00%)
  test_verif_scale3.bpf.linked1.o        balancer_ingress       success      failure      MISMATCH           845499    1000001    +154502 (+18.27%)
  -------------------------------------  ---------------------  -----------  -----------  --------------  ---------  ---------  -------------------

Note that by filtering on verdict_diff=mismatch, it's now extremely easy and
fast to see any changes in verdict. Example above showcases both failure ->
success transitions (which are generally surprising) and success -> failure
transitions (which are expected if bugs are present).

Given veristat allows to query relative percent difference values, internal
logic for comparison mode is based on floating point numbers, so requires a bit
of epsilon precision logic, deviating from typical integer simple handling
rules.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20221103055304.2904589-11-andrii@kernel.orgSigned-off-by: Alexei Starovoitov <ast@kernel.org>

tools/testing/selftests/bpf/veristat.c

@@ -456,12 +456,16 @@ static int append_filter(struct filter **filters, int *cnt, const char *str)
 		    strcasecmp(p, "t") == 0 ||
 		    strcasecmp(p, "success") == 0 ||
 		    strcasecmp(p, "succ") == 0 ||
-		    strcasecmp(p, "s") == 0) {
+		    strcasecmp(p, "s") == 0 ||
+		    strcasecmp(p, "match") == 0 ||
+		    strcasecmp(p, "m") == 0) {
 			val = 1;
 		} else if (strcasecmp(p, "false") == 0 ||
 			   strcasecmp(p, "f") == 0 ||
 			   strcasecmp(p, "failure") == 0 ||
-			   strcasecmp(p, "fail") == 0) {
+			   strcasecmp(p, "fail") == 0 ||
+			   strcasecmp(p, "mismatch") == 0 ||
+			   strcasecmp(p, "mis") == 0) {
 			val = 0;
 		} else {
 			errno = 0;
@@ -1011,6 +1015,8 @@ static void fetch_join_stat_value(const struct verif_stats_join *s,
 	case VARIANT_DIFF:
 		if (!s->stats_a || !s->stats_b)
 			*num_val = -DBL_MAX;
+		else if (id == VERDICT)
+			*num_val = v1 == v2 ? 1.0 /* MATCH */ : 0.0 /* MISMATCH */;
 		else
 			*num_val = (double)(v2 - v1);
 		return;
@@ -1532,12 +1538,61 @@ static int cmp_stats_key(const struct verif_stats *base, const struct verif_stat
 	return strcmp(base->prog_name, comp->prog_name);
 }
 
+static bool is_join_stat_filter_matched(struct filter *f, const struct verif_stats_join *stats)
+{
+	static const double eps = 1e-9;
+	const char *str = NULL;
+	double value = 0.0;
+
+	fetch_join_stat_value(stats, f->stat_id, f->stat_var, &str, &value);
+
+	switch (f->op) {
+	case OP_EQ: return value > f->value - eps && value < f->value + eps;
+	case OP_NEQ: return value < f->value - eps || value > f->value + eps;
+	case OP_LT: return value < f->value - eps;
+	case OP_LE: return value <= f->value + eps;
+	case OP_GT: return value > f->value + eps;
+	case OP_GE: return value >= f->value - eps;
+	}
+
+	fprintf(stderr, "BUG: unknown filter op %d!\n", f->op);
+	return false;
+}
+
+static bool should_output_join_stats(const struct verif_stats_join *stats)
+{
+	struct filter *f;
+	int i, allow_cnt = 0;
+
+	for (i = 0; i < env.deny_filter_cnt; i++) {
+		f = &env.deny_filters[i];
+		if (f->kind != FILTER_STAT)
+			continue;
+
+		if (is_join_stat_filter_matched(f, stats))
+			return false;
+	}
+
+	for (i = 0; i < env.allow_filter_cnt; i++) {
+		f = &env.allow_filters[i];
+		if (f->kind != FILTER_STAT)
+			continue;
+		allow_cnt++;
+
+		if (is_join_stat_filter_matched(f, stats))
+			return true;
+	}
+
+	/* if there are no stat allowed filters, pass everything through */
+	return allow_cnt == 0;
+}
+
 static int handle_comparison_mode(void)
 {
 	struct stat_specs base_specs = {}, comp_specs = {};
 	struct stat_specs tmp_sort_spec;
 	enum resfmt cur_fmt;
-	int err, i, j;
+	int err, i, j, last_idx;
 
 	if (env.filename_cnt != 2) {
 		fprintf(stderr, "Comparison mode expects exactly two input CSV files!\n\n");
@@ -1664,9 +1719,14 @@ static int handle_comparison_mode(void)
 
 	for (i = 0; i < env.join_stat_cnt; i++) {
 		const struct verif_stats_join *join = &env.join_stats[i];
-		bool last = i == env.join_stat_cnt - 1;
 
-		output_comp_stats(join, cur_fmt, last);
+		if (!should_output_join_stats(join))
+			continue;
+
+		if (cur_fmt == RESFMT_TABLE_CALCLEN)
+			last_idx = i;
+
+		output_comp_stats(join, cur_fmt, i == last_idx);
 	}
 
 	if (cur_fmt == RESFMT_TABLE_CALCLEN) {