1 // SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2 // Copyright (c) 2020 Wenbo Zhang
3 //
4 // Based on drsnoop(8) from BCC by Wenbo Zhang.
5 // 28-Feb-2020 Wenbo Zhang Created this.
6 #include <argp.h>
7 #include <signal.h>
8 #include <stdio.h>
9 #include <stdlib.h>
10 #include <string.h>
11 #include <unistd.h>
12 #include <time.h>
13 #include <bpf/libbpf.h>
14 #include <bpf/bpf.h>
15 #include "drsnoop.h"
16 #include "drsnoop.skel.h"
17 #include "trace_helpers.h"
18
19 #define PERF_BUFFER_PAGES 16
20 #define PERF_POLL_TIMEOUT_MS 100
21
22 static volatile sig_atomic_t exiting = 0;
23
24 static struct env {
25 pid_t pid;
26 pid_t tid;
27 time_t duration;
28 bool extended;
29 bool verbose;
30 } env = { };
31
32 const char *argp_program_version = "drsnoop 0.1";
33 const char *argp_program_bug_address =
34 "https://github.com/iovisor/bcc/tree/master/libbpf-tools";
35 const char argp_program_doc[] =
36 "Trace direct reclaim latency.\n"
37 "\n"
38 "USAGE: drsnoop [--help] [-p PID] [-t TID] [-d DURATION] [-e]\n"
39 "\n"
40 "EXAMPLES:\n"
41 " drsnoop # trace all direct reclaim events\n"
42 " drsnoop -p 123 # trace pid 123\n"
43 " drsnoop -t 123 # trace tid 123 (use for threads only)\n"
44 " drsnoop -d 10 # trace for 10 seconds only\n"
45 " drsnoop -e # trace all direct reclaim events with extended faileds\n";
46
47 static const struct argp_option opts[] = {
48 { "duration", 'd', "DURATION", 0, "Total duration of trace in seconds" },
49 { "extended", 'e', NULL, 0, "Extended fields output" },
50 { "pid", 'p', "PID", 0, "Process PID to trace" },
51 { "tid", 't', "TID", 0, "Thread TID to trace" },
52 { "verbose", 'v', NULL, 0, "Verbose debug output" },
53 { NULL, 'h', NULL, OPTION_HIDDEN, "Show the full help" },
54 {},
55 };
56
57 static int page_size;
58
parse_arg(int key,char * arg,struct argp_state * state)59 static error_t parse_arg(int key, char *arg, struct argp_state *state)
60 {
61 time_t duration;
62 int pid;
63
64 switch (key) {
65 case 'h':
66 argp_state_help(state, stderr, ARGP_HELP_STD_HELP);
67 break;
68 case 'v':
69 env.verbose = true;
70 break;
71 case 'd':
72 errno = 0;
73 duration = strtol(arg, NULL, 10);
74 if (errno || duration <= 0) {
75 fprintf(stderr, "invalid DURATION: %s\n", arg);
76 argp_usage(state);
77 }
78 env.duration = duration;
79 break;
80 case 'e':
81 env.extended = true;
82 break;
83 case 'p':
84 errno = 0;
85 pid = strtol(arg, NULL, 10);
86 if (errno || pid <= 0) {
87 fprintf(stderr, "invalid PID: %s\n", arg);
88 argp_usage(state);
89 }
90 env.pid = pid;
91 break;
92 case 't':
93 errno = 0;
94 pid = strtol(arg, NULL, 10);
95 if (errno || pid <= 0) {
96 fprintf(stderr, "invalid TID: %s\n", arg);
97 argp_usage(state);
98 }
99 env.tid = pid;
100 break;
101 default:
102 return ARGP_ERR_UNKNOWN;
103 }
104 return 0;
105 }
106
libbpf_print_fn(enum libbpf_print_level level,const char * format,va_list args)107 static int libbpf_print_fn(enum libbpf_print_level level, const char *format, va_list args)
108 {
109 if (level == LIBBPF_DEBUG && !env.verbose)
110 return 0;
111 return vfprintf(stderr, format, args);
112 }
113
sig_int(int signo)114 static void sig_int(int signo)
115 {
116 exiting = 1;
117 }
118
handle_event(void * ctx,int cpu,void * data,__u32 data_sz)119 void handle_event(void *ctx, int cpu, void *data, __u32 data_sz)
120 {
121 const struct event *e = data;
122 struct tm *tm;
123 char ts[32];
124 time_t t;
125
126 time(&t);
127 tm = localtime(&t);
128 strftime(ts, sizeof(ts), "%H:%M:%S", tm);
129 printf("%-8s %-16s %-6d %8.3f %5lld",
130 ts, e->task, e->pid, e->delta_ns / 1000000.0,
131 e->nr_reclaimed);
132 if (env.extended)
133 printf(" %8llu", e->nr_free_pages * page_size / 1024);
134 printf("\n");
135 }
136
handle_lost_events(void * ctx,int cpu,__u64 lost_cnt)137 void handle_lost_events(void *ctx, int cpu, __u64 lost_cnt)
138 {
139 fprintf(stderr, "lost %llu events on CPU #%d\n", lost_cnt, cpu);
140 }
141
main(int argc,char ** argv)142 int main(int argc, char **argv)
143 {
144 static const struct argp argp = {
145 .options = opts,
146 .parser = parse_arg,
147 .doc = argp_program_doc,
148 };
149 struct perf_buffer *pb = NULL;
150 struct ksyms *ksyms = NULL;
151 const struct ksym *ksym;
152 struct drsnoop_bpf *obj;
153 __u64 time_end = 0;
154 int err;
155
156 err = argp_parse(&argp, argc, argv, 0, NULL, NULL);
157 if (err)
158 return err;
159
160 libbpf_set_print(libbpf_print_fn);
161
162 obj = drsnoop_bpf__open();
163 if (!obj) {
164 fprintf(stderr, "failed to open BPF object\n");
165 return 1;
166 }
167
168 /* initialize global data (filtering options) */
169 obj->rodata->targ_tgid = env.pid;
170 obj->rodata->targ_pid = env.tid;
171 if (env.extended) {
172 ksyms = ksyms__load();
173 if (!ksyms) {
174 fprintf(stderr, "failed to load kallsyms\n");
175 goto cleanup;
176 }
177 ksym = ksyms__get_symbol(ksyms, "vm_zone_stat");
178 if (!ksym) {
179 fprintf(stderr, "failed to get vm_zone_stat's addr\n");
180 goto cleanup;
181 }
182 obj->rodata->vm_zone_stat_kaddr = ksym->addr;
183 page_size = sysconf(_SC_PAGESIZE);
184 }
185
186 if (probe_tp_btf("mm_vmscan_direct_reclaim_begin")) {
187 bpf_program__set_autoload(obj->progs.direct_reclaim_begin, false);
188 bpf_program__set_autoload(obj->progs.direct_reclaim_end, false);
189 } else {
190 bpf_program__set_autoload(obj->progs.direct_reclaim_begin_btf, false);
191 bpf_program__set_autoload(obj->progs.direct_reclaim_end_btf, false);
192 }
193
194 err = drsnoop_bpf__load(obj);
195 if (err) {
196 fprintf(stderr, "failed to load BPF object: %d\n", err);
197 goto cleanup;
198 }
199
200 err = drsnoop_bpf__attach(obj);
201 if (err) {
202 fprintf(stderr, "failed to attach BPF programs\n");
203 goto cleanup;
204 }
205
206 printf("Tracing direct reclaim events");
207 if (env.duration)
208 printf(" for %ld secs.\n", env.duration);
209 else
210 printf("... Hit Ctrl-C to end.\n");
211 printf("%-8s %-16s %-6s %8s %5s",
212 "TIME", "COMM", "TID", "LAT(ms)", "PAGES");
213 if (env.extended)
214 printf(" %8s", "FREE(KB)");
215 printf("\n");
216
217 pb = perf_buffer__new(bpf_map__fd(obj->maps.events), PERF_BUFFER_PAGES,
218 handle_event, handle_lost_events, NULL, NULL);
219 if (!pb) {
220 err = -errno;
221 fprintf(stderr, "failed to open perf buffer: %d\n", err);
222 goto cleanup;
223 }
224
225 /* setup duration */
226 if (env.duration)
227 time_end = get_ktime_ns() + env.duration * NSEC_PER_SEC;
228
229 if (signal(SIGINT, sig_int) == SIG_ERR) {
230 fprintf(stderr, "can't set signal handler: %s\n", strerror(errno));
231 err = 1;
232 goto cleanup;
233 }
234
235 /* main: poll */
236 while (!exiting) {
237 err = perf_buffer__poll(pb, PERF_POLL_TIMEOUT_MS);
238 if (err < 0 && err != -EINTR) {
239 fprintf(stderr, "error polling perf buffer: %s\n", strerror(-err));
240 goto cleanup;
241 }
242 if (env.duration && get_ktime_ns() > time_end)
243 goto cleanup;
244 /* reset err to return 0 if exiting */
245 err = 0;
246 }
247
248 cleanup:
249 perf_buffer__free(pb);
250 drsnoop_bpf__destroy(obj);
251 ksyms__free(ksyms);
252
253 return err != 0;
254 }
255