1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright (c) 2020 Wenbo Zhang
3 #include <vmlinux.h>
4 #include <bpf/bpf_helpers.h>
5 #include <bpf/bpf_core_read.h>
6 #include <bpf/bpf_tracing.h>
7 #include "biolatency.h"
8 #include "bits.bpf.h"
9
10 #define MAX_ENTRIES 10240
11
12 extern int LINUX_KERNEL_VERSION __kconfig;
13
14 const volatile bool filter_cg = false;
15 const volatile bool targ_per_disk = false;
16 const volatile bool targ_per_flag = false;
17 const volatile bool targ_queued = false;
18 const volatile bool targ_ms = false;
19 const volatile bool filter_dev = false;
20 const volatile __u32 targ_dev = 0;
21
22 struct {
23 __uint(type, BPF_MAP_TYPE_CGROUP_ARRAY);
24 __type(key, u32);
25 __type(value, u32);
26 __uint(max_entries, 1);
27 } cgroup_map SEC(".maps");
28
29 struct {
30 __uint(type, BPF_MAP_TYPE_HASH);
31 __uint(max_entries, MAX_ENTRIES);
32 __type(key, struct request *);
33 __type(value, u64);
34 __uint(map_flags, BPF_F_NO_PREALLOC);
35 } start SEC(".maps");
36
37 static struct hist initial_hist;
38
39 struct {
40 __uint(type, BPF_MAP_TYPE_HASH);
41 __uint(max_entries, MAX_ENTRIES);
42 __type(key, struct hist_key);
43 __type(value, struct hist);
44 __uint(map_flags, BPF_F_NO_PREALLOC);
45 } hists SEC(".maps");
46
47 static __always_inline
trace_rq_start(struct request * rq,int issue)48 int trace_rq_start(struct request *rq, int issue)
49 {
50 if (issue && targ_queued && BPF_CORE_READ(rq->q, elevator))
51 return 0;
52
53 u64 ts = bpf_ktime_get_ns();
54
55 if (filter_dev) {
56 struct gendisk *disk = BPF_CORE_READ(rq, rq_disk);
57 u32 dev;
58
59 dev = disk ? MKDEV(BPF_CORE_READ(disk, major),
60 BPF_CORE_READ(disk, first_minor)) : 0;
61 if (targ_dev != dev)
62 return 0;
63 }
64 bpf_map_update_elem(&start, &rq, &ts, 0);
65 return 0;
66 }
67
68 SEC("tp_btf/block_rq_insert")
block_rq_insert(u64 * ctx)69 int block_rq_insert(u64 *ctx)
70 {
71 if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
72 return 0;
73
74 /**
75 * commit a54895fa (v5.11-rc1) changed tracepoint argument list
76 * from TP_PROTO(struct request_queue *q, struct request *rq)
77 * to TP_PROTO(struct request *rq)
78 */
79 if (LINUX_KERNEL_VERSION <= KERNEL_VERSION(5, 10, 0))
80 return trace_rq_start((void *)ctx[1], false);
81 else
82 return trace_rq_start((void *)ctx[0], false);
83 }
84
85 SEC("tp_btf/block_rq_issue")
block_rq_issue(u64 * ctx)86 int block_rq_issue(u64 *ctx)
87 {
88 if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
89 return 0;
90
91 /**
92 * commit a54895fa (v5.11-rc1) changed tracepoint argument list
93 * from TP_PROTO(struct request_queue *q, struct request *rq)
94 * to TP_PROTO(struct request *rq)
95 */
96 if (LINUX_KERNEL_VERSION <= KERNEL_VERSION(5, 10, 0))
97 return trace_rq_start((void *)ctx[1], true);
98 else
99 return trace_rq_start((void *)ctx[0], true);
100 }
101
102 SEC("tp_btf/block_rq_complete")
BPF_PROG(block_rq_complete,struct request * rq,int error,unsigned int nr_bytes)103 int BPF_PROG(block_rq_complete, struct request *rq, int error,
104 unsigned int nr_bytes)
105 {
106 if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
107 return 0;
108
109 u64 slot, *tsp, ts = bpf_ktime_get_ns();
110 struct hist_key hkey = {};
111 struct hist *histp;
112 s64 delta;
113
114 tsp = bpf_map_lookup_elem(&start, &rq);
115 if (!tsp)
116 return 0;
117 delta = (s64)(ts - *tsp);
118 if (delta < 0)
119 goto cleanup;
120
121 if (targ_per_disk) {
122 struct gendisk *disk = BPF_CORE_READ(rq, rq_disk);
123
124 hkey.dev = disk ? MKDEV(BPF_CORE_READ(disk, major),
125 BPF_CORE_READ(disk, first_minor)) : 0;
126 }
127 if (targ_per_flag)
128 hkey.cmd_flags = rq->cmd_flags;
129
130 histp = bpf_map_lookup_elem(&hists, &hkey);
131 if (!histp) {
132 bpf_map_update_elem(&hists, &hkey, &initial_hist, 0);
133 histp = bpf_map_lookup_elem(&hists, &hkey);
134 if (!histp)
135 goto cleanup;
136 }
137
138 if (targ_ms)
139 delta /= 1000000U;
140 else
141 delta /= 1000U;
142 slot = log2l(delta);
143 if (slot >= MAX_SLOTS)
144 slot = MAX_SLOTS - 1;
145 __sync_fetch_and_add(&histp->slots[slot], 1);
146
147 cleanup:
148 bpf_map_delete_elem(&start, &rq);
149 return 0;
150 }
151
152 char LICENSE[] SEC("license") = "GPL";
153