[PATCH bpf-next v1 2/3] selftests/bpf: Add lock wait time stats to rqspinlock stress test

[Kumar Kartikeya Dwivedi <memxor@gmail.com>]

Add statistics per-CPU broken down by context and various timing windows
for the time taken to acquire an rqspinlock. Cases where all
acquisitions fit into the 10ms window are skipped from printing,
otherwise the full breakdown is displayed when printing the summary.
This allows capturing precisely the number of times outlier attempts
happened for a given lock in a given context.

A critical detail is that time is captured regardless of success or
failure, which is important to capture events for failed but long
waiting timeout attempts.

Output:

[   64.279459] rqspinlock acquisition latency histogram (ms):
[   64.279472]  cpu1: total 528426 (normal 526559, nmi 1867)
[   64.279477]    0-1ms: total 524697 (normal 524697, nmi 0)
[   64.279480]    2-2ms: total 3652 (normal 1811, nmi 1841)
[   64.279482]    3-3ms: total 66 (normal 47, nmi 19)
[   64.279485]    4-4ms: total 2 (normal 1, nmi 1)
[   64.279487]    5-5ms: total 1 (normal 1, nmi 0)
[   64.279489]    6-6ms: total 1 (normal 0, nmi 1)
[   64.279490]    101-150ms: total 1 (normal 0, nmi 1)
[   64.279492]    >= 251ms: total 6 (normal 2, nmi 4)
...

Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
---
 .../bpf/test_kmods/bpf_test_rqspinlock.c      | 104 ++++++++++++++++++
 1 file changed, 104 insertions(+)

diff --git a/tools/testing/selftests/bpf/test_kmods/bpf_test_rqspinlock.c b/tools/testing/selftests/bpf/test_kmods/bpf_test_rqspinlock.c
index 8096624cf9c1..4ea7ec420e4e 100644
--- a/tools/testing/selftests/bpf/test_kmods/bpf_test_rqspinlock.c
+++ b/tools/testing/selftests/bpf/test_kmods/bpf_test_rqspinlock.c
@@ -5,6 +5,7 @@
 #include <linux/delay.h>
 #include <linux/module.h>
 #include <linux/prandom.h>
+#include <linux/ktime.h>
 #include <asm/rqspinlock.h>
 #include <linux/perf_event.h>
 #include <linux/kthread.h>
@@ -24,6 +25,21 @@ static rqspinlock_t lock_a;
 static rqspinlock_t lock_b;
 static rqspinlock_t lock_c;
 
+#define RQSL_SLOW_THRESHOLD_MS 10
+static const unsigned int rqsl_hist_ms[] = {
+	1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
+	12, 14, 16, 18, 20, 25, 30, 40, 50, 75,
+	100, 150, 200, 250, 1000,
+};
+#define RQSL_NR_HIST_BUCKETS ARRAY_SIZE(rqsl_hist_ms)
+
+struct rqsl_cpu_hist {
+	atomic64_t normal[RQSL_NR_HIST_BUCKETS];
+	atomic64_t nmi[RQSL_NR_HIST_BUCKETS];
+};
+
+static DEFINE_PER_CPU(struct rqsl_cpu_hist, rqsl_cpu_hists);
+
 enum rqsl_mode {
 	RQSL_MODE_AA = 0,
 	RQSL_MODE_ABBA,
@@ -79,10 +95,33 @@ static struct rqsl_lock_pair rqsl_get_lock_pair(int cpu)
 	}
 }
 
+static u32 rqsl_hist_bucket_idx(u32 delta_ms)
+{
+	int i;
+
+	for (i = 0; i < RQSL_NR_HIST_BUCKETS; i++) {
+		if (delta_ms <= rqsl_hist_ms[i])
+			return i;
+	}
+
+	return RQSL_NR_HIST_BUCKETS - 1;
+}
+
+static void rqsl_record_lock_time(u64 delta_ns, bool is_nmi)
+{
+	struct rqsl_cpu_hist *hist = this_cpu_ptr(&rqsl_cpu_hists);
+	u32 delta_ms = DIV_ROUND_UP_ULL(delta_ns, NSEC_PER_MSEC);
+	u32 bucket = rqsl_hist_bucket_idx(delta_ms);
+	atomic64_t *buckets = is_nmi ? hist->nmi : hist->normal;
+
+	atomic64_inc(&buckets[bucket]);
+}
+
 static int rqspinlock_worker_fn(void *arg)
 {
 	int cpu = smp_processor_id();
 	unsigned long flags;
+	u64 start_ns;
 	int ret;
 
 	if (cpu) {
@@ -96,7 +135,9 @@ static int rqspinlock_worker_fn(void *arg)
 				msleep(1000);
 				continue;
 			}
+			start_ns = ktime_get_mono_fast_ns();
 			ret = raw_res_spin_lock_irqsave(worker_lock, flags);
+			rqsl_record_lock_time(ktime_get_mono_fast_ns() - start_ns, false);
 			mdelay(20);
 			if (!ret)
 				raw_res_spin_unlock_irqrestore(worker_lock, flags);
@@ -130,13 +171,16 @@ static void nmi_cb(struct perf_event *event, struct perf_sample_data *data,
 	struct rqsl_lock_pair locks;
 	int cpu = smp_processor_id();
 	unsigned long flags;
+	u64 start_ns;
 	int ret;
 
 	if (!cpu || READ_ONCE(pause))
 		return;
 
 	locks = rqsl_get_lock_pair(cpu);
+	start_ns = ktime_get_mono_fast_ns();
 	ret = raw_res_spin_lock_irqsave(locks.nmi_lock, flags);
+	rqsl_record_lock_time(ktime_get_mono_fast_ns() - start_ns, true);
 
 	mdelay(10);
 
@@ -235,10 +279,70 @@ static int bpf_test_rqspinlock_init(void)
 
 module_init(bpf_test_rqspinlock_init);
 
+static void rqsl_print_histograms(void)
+{
+	int cpu, i;
+
+	pr_err("rqspinlock acquisition latency histogram (ms):\n");
+
+	for_each_online_cpu(cpu) {
+		struct rqsl_cpu_hist *hist = per_cpu_ptr(&rqsl_cpu_hists, cpu);
+		u64 norm_counts[RQSL_NR_HIST_BUCKETS];
+		u64 nmi_counts[RQSL_NR_HIST_BUCKETS];
+		u64 total_counts[RQSL_NR_HIST_BUCKETS];
+		u64 norm_total = 0, nmi_total = 0, total = 0;
+		bool has_slow = false;
+
+		for (i = 0; i < RQSL_NR_HIST_BUCKETS; i++) {
+			norm_counts[i] = atomic64_read(&hist->normal[i]);
+			nmi_counts[i] = atomic64_read(&hist->nmi[i]);
+			total_counts[i] = norm_counts[i] + nmi_counts[i];
+			norm_total += norm_counts[i];
+			nmi_total += nmi_counts[i];
+			total += total_counts[i];
+			if (rqsl_hist_ms[i] > RQSL_SLOW_THRESHOLD_MS &&
+			    total_counts[i])
+				has_slow = true;
+		}
+
+		if (!total)
+			continue;
+
+		if (!has_slow) {
+			pr_err(" cpu%d: total %llu (normal %llu, nmi %llu), all within 0-%ums\n",
+			       cpu, total, norm_total, nmi_total, RQSL_SLOW_THRESHOLD_MS);
+			continue;
+		}
+
+		pr_err(" cpu%d: total %llu (normal %llu, nmi %llu)\n",
+		       cpu, total, norm_total, nmi_total);
+		for (i = 0; i < RQSL_NR_HIST_BUCKETS; i++) {
+			unsigned int start_ms;
+
+			if (!total_counts[i])
+				continue;
+
+			start_ms = i == 0 ? 0 : rqsl_hist_ms[i - 1] + 1;
+			if (i == RQSL_NR_HIST_BUCKETS - 1) {
+				pr_err("   >= %ums: total %llu (normal %llu, nmi %llu)\n",
+				       start_ms, total_counts[i],
+				       norm_counts[i], nmi_counts[i]);
+			} else {
+				pr_err("   %u-%ums: total %llu (normal %llu, nmi %llu)\n",
+				       start_ms, rqsl_hist_ms[i],
+				       total_counts[i],
+				       norm_counts[i], nmi_counts[i]);
+			}
+		}
+	}
+}
+
 static void bpf_test_rqspinlock_exit(void)
 {
+	WRITE_ONCE(pause, 1);
 	free_rqsl_threads();
 	free_rqsl_evts();
+	rqsl_print_histograms();
 }
 
 module_exit(bpf_test_rqspinlock_exit);
-- 
2.51.0