libbpf-tools: Use raw_tp rather than tp_btf for biolatency.

[eiffel-fl]

Hi.

In this PR, I converted biolatency to use raw_tp rather than tp_btf.
As a result, the tool can now run on kernel older than 5.5.

This PR is one step toward closing #4231.

Best regards and thank you in advance for any feedback.

[chenhengqi]

Sorry, I forgot your commets in #4231 , I've already worked on biolatency. Would you mind taking biosnoop/biostacks/bitesize instead ? Thanks.

[chenhengqi]

From 4e5c7c3ef039576e05fec6641f248f115a87c545 Mon Sep 17 00:00:00 2001
From: Hengqi Chen <chenhengqi@outlook.com>
Date: Mon, 10 Oct 2022 18:04:25 +0800
Subject: [PATCH] libbpf-tools: Enable biolatency on kernels without BPF
 trampoline

On kernels without BPF trampoline, let's switch to raw tracepoint instead.

Signed-off-by: Hengqi Chen <chenhengqi@outlook.com>
---
 libbpf-tools/biolatency.bpf.c | 74 +++++++++++++++++++++++++----------
 libbpf-tools/biolatency.c     | 38 +++++++++---------
 2 files changed, 71 insertions(+), 41 deletions(-)

diff --git a/libbpf-tools/biolatency.bpf.c b/libbpf-tools/biolatency.bpf.c
index bab62b1d..429412db 100644
--- a/libbpf-tools/biolatency.bpf.c
+++ b/libbpf-tools/biolatency.bpf.c
@@ -4,6 +4,7 @@
 #include <bpf/bpf_helpers.h>
 #include <bpf/bpf_core_read.h>
 #include <bpf/bpf_tracing.h>
+
 #include "biolatency.h"
 #include "bits.bpf.h"
 #include "core_fixes.bpf.h"
@@ -43,13 +44,17 @@ struct {
 	__type(value, struct hist);
 } hists SEC(".maps");
 
-static __always_inline
-int trace_rq_start(struct request *rq, int issue)
+static int __always_inline trace_rq_start(struct request *rq, int issue)
 {
-	if (issue && targ_queued && BPF_CORE_READ(rq->q, elevator))
+	u64 ts;
+
+	if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
 		return 0;
 
-	u64 ts = bpf_ktime_get_ns();
+	if (issue && targ_queued && BPF_CORE_READ(rq, q, elevator))
+		return 0;
+
+	ts = bpf_ktime_get_ns();
 
 	if (filter_dev) {
 		struct gendisk *disk = get_disk(rq);
@@ -64,12 +69,8 @@ int trace_rq_start(struct request *rq, int issue)
 	return 0;
 }
 
-SEC("tp_btf/block_rq_insert")
-int block_rq_insert(u64 *ctx)
+static int handle_block_rq_insert(__u64 *ctx)
 {
-	if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
-		return 0;
-
 	/**
 	 * commit a54895fa (v5.11-rc1) changed tracepoint argument list
 	 * from TP_PROTO(struct request_queue *q, struct request *rq)
@@ -81,12 +82,8 @@ int block_rq_insert(u64 *ctx)
 		return trace_rq_start((void *)ctx[0], false);
 }
 
-SEC("tp_btf/block_rq_issue")
-int block_rq_issue(u64 *ctx)
+static int handle_block_rq_issue(__u64 *ctx)
 {
-	if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
-		return 0;
-
 	/**
 	 * commit a54895fa (v5.11-rc1) changed tracepoint argument list
 	 * from TP_PROTO(struct request_queue *q, struct request *rq)
@@ -98,21 +95,20 @@ int block_rq_issue(u64 *ctx)
 		return trace_rq_start((void *)ctx[0], true);
 }
 
-SEC("tp_btf/block_rq_complete")
-int BPF_PROG(block_rq_complete, struct request *rq, int error,
-	unsigned int nr_bytes)
+static int handle_block_rq_complete(struct request *rq, int error, unsigned int nr_bytes)
 {
-	if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
-		return 0;
-
 	u64 slot, *tsp, ts = bpf_ktime_get_ns();
 	struct hist_key hkey = {};
 	struct hist *histp;
 	s64 delta;
 
+	if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
+		return 0;
+
 	tsp = bpf_map_lookup_elem(&start, &rq);
 	if (!tsp)
 		return 0;
+
 	delta = (s64)(ts - *tsp);
 	if (delta < 0)
 		goto cleanup;
@@ -124,7 +120,7 @@ int BPF_PROG(block_rq_complete, struct request *rq, int error,
 					BPF_CORE_READ(disk, first_minor)) : 0;
 	}
 	if (targ_per_flag)
-		hkey.cmd_flags = rq->cmd_flags;
+		hkey.cmd_flags = BPF_CORE_READ(rq, cmd_flags);
 
 	histp = bpf_map_lookup_elem(&hists, &hkey);
 	if (!histp) {
@@ -148,4 +144,40 @@ cleanup:
 	return 0;
 }
 
+SEC("tp_btf/block_rq_insert")
+int block_rq_insert_btf(u64 *ctx)
+{
+	return handle_block_rq_insert(ctx);
+}
+
+SEC("tp_btf/block_rq_issue")
+int block_rq_issue_btf(u64 *ctx)
+{
+	return handle_block_rq_issue(ctx);
+}
+
+SEC("tp_btf/block_rq_complete")
+int BPF_PROG(block_rq_complete_btf, struct request *rq, int error, unsigned int nr_bytes)
+{
+	return handle_block_rq_complete(rq, error, nr_bytes);
+}
+
+SEC("raw_tp/block_rq_insert")
+int BPF_PROG(block_rq_insert)
+{
+	return handle_block_rq_insert(ctx);
+}
+
+SEC("raw_tp/block_rq_issue")
+int BPF_PROG(block_rq_issue)
+{
+	return handle_block_rq_issue(ctx);
+}
+
+SEC("raw_tp/block_rq_complete")
+int BPF_PROG(block_rq_complete, struct request *rq, int error, unsigned int nr_bytes)
+{
+	return handle_block_rq_complete(rq, error, nr_bytes);
+}
+
 char LICENSE[] SEC("license") = "GPL";
diff --git a/libbpf-tools/biolatency.c b/libbpf-tools/biolatency.c
index 51afa509..4345851a 100644
--- a/libbpf-tools/biolatency.c
+++ b/libbpf-tools/biolatency.c
@@ -193,8 +193,7 @@ static void print_cmd_flags(int cmd_flags)
 		printf("Unknown");
 }
 
-static
-int print_log2_hists(struct bpf_map *hists, struct partitions *partitions)
+static int print_log2_hists(struct bpf_map *hists, struct partitions *partitions)
 {
 	struct hist_key lookup_key = { .cmd_flags = -1 }, next_key;
 	const char *units = env.milliseconds ? "msecs" : "usecs";
@@ -286,6 +285,20 @@ int main(int argc, char **argv)
 	obj->rodata->targ_queued = env.queued;
 	obj->rodata->filter_cg = env.cg;
 
+	if (probe_tp_btf("block_rq_insert")) {
+		bpf_program__set_autoload(obj->progs.block_rq_insert, false);
+		bpf_program__set_autoload(obj->progs.block_rq_issue, false);
+		bpf_program__set_autoload(obj->progs.block_rq_complete, false);
+		if (!env.queued)
+			bpf_program__set_autoload(obj->progs.block_rq_insert_btf, false);
+	} else {
+		bpf_program__set_autoload(obj->progs.block_rq_insert_btf, false);
+		bpf_program__set_autoload(obj->progs.block_rq_issue_btf, false);
+		bpf_program__set_autoload(obj->progs.block_rq_complete_btf, false);
+		if (!env.queued)
+			bpf_program__set_autoload(obj->progs.block_rq_insert, false);
+	}
+
 	err = biolatency_bpf__load(obj);
 	if (err) {
 		fprintf(stderr, "failed to load BPF object: %d\n", err);
@@ -307,24 +320,9 @@ int main(int argc, char **argv)
 		}
 	}
 
-	if (env.queued) {
-		obj->links.block_rq_insert = bpf_program__attach(obj->progs.block_rq_insert);
-		if (!obj->links.block_rq_insert) {
-			err = -errno;
-			fprintf(stderr, "failed to attach: %s\n", strerror(-err));
-			goto cleanup;
-		}
-	}
-	obj->links.block_rq_issue = bpf_program__attach(obj->progs.block_rq_issue);
-	if (!obj->links.block_rq_issue) {
-		err = -errno;
-		fprintf(stderr, "failed to attach: %s\n", strerror(-err));
-		goto cleanup;
-	}
-	obj->links.block_rq_complete = bpf_program__attach(obj->progs.block_rq_complete);
-	if (!obj->links.block_rq_complete) {
-		err = -errno;
-		fprintf(stderr, "failed to attach: %s\n", strerror(-err));
+	err = biolatency_bpf__attach(obj);
+	if (err) {
+		fprintf(stderr, "failed to attach BPF object: %d\n", err);
 		goto cleanup;
 	}
 
-- 
2.27.0

[eiffel-fl]

Sorry, I forgot your commets in #4231 , I've already worked on biolatency. Would you mind taking biosnoop/biostacks/bitesize instead ? Thanks.

No problem, I guess this kind of things happen!
I cannot ensure you I will be able to work quickly on the other tool because I am less familiar with them than biolatency.
Nonetheless, open a PR for biolatency and I will test your code and share some reviews :D.

[chenhengqi]

Close in favor of #4275 .