]> asedeno.scripts.mit.edu Git - linux.git/blob - samples/bpf/bpf_load.c
Merge branch 'patchwork' into v4l_for_linus
[linux.git] / samples / bpf / bpf_load.c
1 #include <stdio.h>
2 #include <sys/types.h>
3 #include <sys/stat.h>
4 #include <fcntl.h>
5 #include <libelf.h>
6 #include <gelf.h>
7 #include <errno.h>
8 #include <unistd.h>
9 #include <string.h>
10 #include <stdbool.h>
11 #include <stdlib.h>
12 #include <linux/bpf.h>
13 #include <linux/filter.h>
14 #include <linux/perf_event.h>
15 #include <linux/netlink.h>
16 #include <linux/rtnetlink.h>
17 #include <sys/types.h>
18 #include <sys/socket.h>
19 #include <sys/syscall.h>
20 #include <sys/ioctl.h>
21 #include <sys/mman.h>
22 #include <poll.h>
23 #include <ctype.h>
24 #include "libbpf.h"
25 #include "bpf_helpers.h"
26 #include "bpf_load.h"
27
28 #define DEBUGFS "/sys/kernel/debug/tracing/"
29
30 static char license[128];
31 static int kern_version;
32 static bool processed_sec[128];
33 int map_fd[MAX_MAPS];
34 int prog_fd[MAX_PROGS];
35 int event_fd[MAX_PROGS];
36 int prog_cnt;
37 int prog_array_fd = -1;
38
39 static int populate_prog_array(const char *event, int prog_fd)
40 {
41         int ind = atoi(event), err;
42
43         err = bpf_update_elem(prog_array_fd, &ind, &prog_fd, BPF_ANY);
44         if (err < 0) {
45                 printf("failed to store prog_fd in prog_array\n");
46                 return -1;
47         }
48         return 0;
49 }
50
51 static int load_and_attach(const char *event, struct bpf_insn *prog, int size)
52 {
53         bool is_socket = strncmp(event, "socket", 6) == 0;
54         bool is_kprobe = strncmp(event, "kprobe/", 7) == 0;
55         bool is_kretprobe = strncmp(event, "kretprobe/", 10) == 0;
56         bool is_tracepoint = strncmp(event, "tracepoint/", 11) == 0;
57         bool is_xdp = strncmp(event, "xdp", 3) == 0;
58         bool is_perf_event = strncmp(event, "perf_event", 10) == 0;
59         bool is_cgroup_skb = strncmp(event, "cgroup/skb", 10) == 0;
60         bool is_cgroup_sk = strncmp(event, "cgroup/sock", 11) == 0;
61         enum bpf_prog_type prog_type;
62         char buf[256];
63         int fd, efd, err, id;
64         struct perf_event_attr attr = {};
65
66         attr.type = PERF_TYPE_TRACEPOINT;
67         attr.sample_type = PERF_SAMPLE_RAW;
68         attr.sample_period = 1;
69         attr.wakeup_events = 1;
70
71         if (is_socket) {
72                 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
73         } else if (is_kprobe || is_kretprobe) {
74                 prog_type = BPF_PROG_TYPE_KPROBE;
75         } else if (is_tracepoint) {
76                 prog_type = BPF_PROG_TYPE_TRACEPOINT;
77         } else if (is_xdp) {
78                 prog_type = BPF_PROG_TYPE_XDP;
79         } else if (is_perf_event) {
80                 prog_type = BPF_PROG_TYPE_PERF_EVENT;
81         } else if (is_cgroup_skb) {
82                 prog_type = BPF_PROG_TYPE_CGROUP_SKB;
83         } else if (is_cgroup_sk) {
84                 prog_type = BPF_PROG_TYPE_CGROUP_SOCK;
85         } else {
86                 printf("Unknown event '%s'\n", event);
87                 return -1;
88         }
89
90         fd = bpf_prog_load(prog_type, prog, size, license, kern_version);
91         if (fd < 0) {
92                 printf("bpf_prog_load() err=%d\n%s", errno, bpf_log_buf);
93                 return -1;
94         }
95
96         prog_fd[prog_cnt++] = fd;
97
98         if (is_xdp || is_perf_event || is_cgroup_skb || is_cgroup_sk)
99                 return 0;
100
101         if (is_socket) {
102                 event += 6;
103                 if (*event != '/')
104                         return 0;
105                 event++;
106                 if (!isdigit(*event)) {
107                         printf("invalid prog number\n");
108                         return -1;
109                 }
110                 return populate_prog_array(event, fd);
111         }
112
113         if (is_kprobe || is_kretprobe) {
114                 if (is_kprobe)
115                         event += 7;
116                 else
117                         event += 10;
118
119                 if (*event == 0) {
120                         printf("event name cannot be empty\n");
121                         return -1;
122                 }
123
124                 if (isdigit(*event))
125                         return populate_prog_array(event, fd);
126
127                 snprintf(buf, sizeof(buf),
128                          "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
129                          is_kprobe ? 'p' : 'r', event, event);
130                 err = system(buf);
131                 if (err < 0) {
132                         printf("failed to create kprobe '%s' error '%s'\n",
133                                event, strerror(errno));
134                         return -1;
135                 }
136
137                 strcpy(buf, DEBUGFS);
138                 strcat(buf, "events/kprobes/");
139                 strcat(buf, event);
140                 strcat(buf, "/id");
141         } else if (is_tracepoint) {
142                 event += 11;
143
144                 if (*event == 0) {
145                         printf("event name cannot be empty\n");
146                         return -1;
147                 }
148                 strcpy(buf, DEBUGFS);
149                 strcat(buf, "events/");
150                 strcat(buf, event);
151                 strcat(buf, "/id");
152         }
153
154         efd = open(buf, O_RDONLY, 0);
155         if (efd < 0) {
156                 printf("failed to open event %s\n", event);
157                 return -1;
158         }
159
160         err = read(efd, buf, sizeof(buf));
161         if (err < 0 || err >= sizeof(buf)) {
162                 printf("read from '%s' failed '%s'\n", event, strerror(errno));
163                 return -1;
164         }
165
166         close(efd);
167
168         buf[err] = 0;
169         id = atoi(buf);
170         attr.config = id;
171
172         efd = perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
173         if (efd < 0) {
174                 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
175                 return -1;
176         }
177         event_fd[prog_cnt - 1] = efd;
178         ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
179         ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
180
181         return 0;
182 }
183
184 static int load_maps(struct bpf_map_def *maps, int len)
185 {
186         int i;
187
188         for (i = 0; i < len / sizeof(struct bpf_map_def); i++) {
189
190                 map_fd[i] = bpf_create_map(maps[i].type,
191                                            maps[i].key_size,
192                                            maps[i].value_size,
193                                            maps[i].max_entries,
194                                            maps[i].map_flags);
195                 if (map_fd[i] < 0) {
196                         printf("failed to create a map: %d %s\n",
197                                errno, strerror(errno));
198                         return 1;
199                 }
200
201                 if (maps[i].type == BPF_MAP_TYPE_PROG_ARRAY)
202                         prog_array_fd = map_fd[i];
203         }
204         return 0;
205 }
206
207 static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
208                    GElf_Shdr *shdr, Elf_Data **data)
209 {
210         Elf_Scn *scn;
211
212         scn = elf_getscn(elf, i);
213         if (!scn)
214                 return 1;
215
216         if (gelf_getshdr(scn, shdr) != shdr)
217                 return 2;
218
219         *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
220         if (!*shname || !shdr->sh_size)
221                 return 3;
222
223         *data = elf_getdata(scn, 0);
224         if (!*data || elf_getdata(scn, *data) != NULL)
225                 return 4;
226
227         return 0;
228 }
229
230 static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
231                                 GElf_Shdr *shdr, struct bpf_insn *insn)
232 {
233         int i, nrels;
234
235         nrels = shdr->sh_size / shdr->sh_entsize;
236
237         for (i = 0; i < nrels; i++) {
238                 GElf_Sym sym;
239                 GElf_Rel rel;
240                 unsigned int insn_idx;
241
242                 gelf_getrel(data, i, &rel);
243
244                 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
245
246                 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
247
248                 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
249                         printf("invalid relo for insn[%d].code 0x%x\n",
250                                insn_idx, insn[insn_idx].code);
251                         return 1;
252                 }
253                 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
254                 insn[insn_idx].imm = map_fd[sym.st_value / sizeof(struct bpf_map_def)];
255         }
256
257         return 0;
258 }
259
260 int load_bpf_file(char *path)
261 {
262         int fd, i;
263         Elf *elf;
264         GElf_Ehdr ehdr;
265         GElf_Shdr shdr, shdr_prog;
266         Elf_Data *data, *data_prog, *symbols = NULL;
267         char *shname, *shname_prog;
268
269         if (elf_version(EV_CURRENT) == EV_NONE)
270                 return 1;
271
272         fd = open(path, O_RDONLY, 0);
273         if (fd < 0)
274                 return 1;
275
276         elf = elf_begin(fd, ELF_C_READ, NULL);
277
278         if (!elf)
279                 return 1;
280
281         if (gelf_getehdr(elf, &ehdr) != &ehdr)
282                 return 1;
283
284         /* clear all kprobes */
285         i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
286
287         /* scan over all elf sections to get license and map info */
288         for (i = 1; i < ehdr.e_shnum; i++) {
289
290                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
291                         continue;
292
293                 if (0) /* helpful for llvm debugging */
294                         printf("section %d:%s data %p size %zd link %d flags %d\n",
295                                i, shname, data->d_buf, data->d_size,
296                                shdr.sh_link, (int) shdr.sh_flags);
297
298                 if (strcmp(shname, "license") == 0) {
299                         processed_sec[i] = true;
300                         memcpy(license, data->d_buf, data->d_size);
301                 } else if (strcmp(shname, "version") == 0) {
302                         processed_sec[i] = true;
303                         if (data->d_size != sizeof(int)) {
304                                 printf("invalid size of version section %zd\n",
305                                        data->d_size);
306                                 return 1;
307                         }
308                         memcpy(&kern_version, data->d_buf, sizeof(int));
309                 } else if (strcmp(shname, "maps") == 0) {
310                         processed_sec[i] = true;
311                         if (load_maps(data->d_buf, data->d_size))
312                                 return 1;
313                 } else if (shdr.sh_type == SHT_SYMTAB) {
314                         symbols = data;
315                 }
316         }
317
318         /* load programs that need map fixup (relocations) */
319         for (i = 1; i < ehdr.e_shnum; i++) {
320
321                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
322                         continue;
323                 if (shdr.sh_type == SHT_REL) {
324                         struct bpf_insn *insns;
325
326                         if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
327                                     &shdr_prog, &data_prog))
328                                 continue;
329
330                         if (shdr_prog.sh_type != SHT_PROGBITS ||
331                             !(shdr_prog.sh_flags & SHF_EXECINSTR))
332                                 continue;
333
334                         insns = (struct bpf_insn *) data_prog->d_buf;
335
336                         processed_sec[shdr.sh_info] = true;
337                         processed_sec[i] = true;
338
339                         if (parse_relo_and_apply(data, symbols, &shdr, insns))
340                                 continue;
341
342                         if (memcmp(shname_prog, "kprobe/", 7) == 0 ||
343                             memcmp(shname_prog, "kretprobe/", 10) == 0 ||
344                             memcmp(shname_prog, "tracepoint/", 11) == 0 ||
345                             memcmp(shname_prog, "xdp", 3) == 0 ||
346                             memcmp(shname_prog, "perf_event", 10) == 0 ||
347                             memcmp(shname_prog, "socket", 6) == 0 ||
348                             memcmp(shname_prog, "cgroup/", 7) == 0)
349                                 load_and_attach(shname_prog, insns, data_prog->d_size);
350                 }
351         }
352
353         /* load programs that don't use maps */
354         for (i = 1; i < ehdr.e_shnum; i++) {
355
356                 if (processed_sec[i])
357                         continue;
358
359                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
360                         continue;
361
362                 if (memcmp(shname, "kprobe/", 7) == 0 ||
363                     memcmp(shname, "kretprobe/", 10) == 0 ||
364                     memcmp(shname, "tracepoint/", 11) == 0 ||
365                     memcmp(shname, "xdp", 3) == 0 ||
366                     memcmp(shname, "perf_event", 10) == 0 ||
367                     memcmp(shname, "socket", 6) == 0 ||
368                     memcmp(shname, "cgroup/", 7) == 0)
369                         load_and_attach(shname, data->d_buf, data->d_size);
370         }
371
372         close(fd);
373         return 0;
374 }
375
376 void read_trace_pipe(void)
377 {
378         int trace_fd;
379
380         trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
381         if (trace_fd < 0)
382                 return;
383
384         while (1) {
385                 static char buf[4096];
386                 ssize_t sz;
387
388                 sz = read(trace_fd, buf, sizeof(buf));
389                 if (sz > 0) {
390                         buf[sz] = 0;
391                         puts(buf);
392                 }
393         }
394 }
395
396 #define MAX_SYMS 300000
397 static struct ksym syms[MAX_SYMS];
398 static int sym_cnt;
399
400 static int ksym_cmp(const void *p1, const void *p2)
401 {
402         return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
403 }
404
405 int load_kallsyms(void)
406 {
407         FILE *f = fopen("/proc/kallsyms", "r");
408         char func[256], buf[256];
409         char symbol;
410         void *addr;
411         int i = 0;
412
413         if (!f)
414                 return -ENOENT;
415
416         while (!feof(f)) {
417                 if (!fgets(buf, sizeof(buf), f))
418                         break;
419                 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
420                         break;
421                 if (!addr)
422                         continue;
423                 syms[i].addr = (long) addr;
424                 syms[i].name = strdup(func);
425                 i++;
426         }
427         sym_cnt = i;
428         qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
429         return 0;
430 }
431
432 struct ksym *ksym_search(long key)
433 {
434         int start = 0, end = sym_cnt;
435         int result;
436
437         while (start < end) {
438                 size_t mid = start + (end - start) / 2;
439
440                 result = key - syms[mid].addr;
441                 if (result < 0)
442                         end = mid;
443                 else if (result > 0)
444                         start = mid + 1;
445                 else
446                         return &syms[mid];
447         }
448
449         if (start >= 1 && syms[start - 1].addr < key &&
450             key < syms[start].addr)
451                 /* valid ksym */
452                 return &syms[start - 1];
453
454         /* out of range. return _stext */
455         return &syms[0];
456 }
457
458 int set_link_xdp_fd(int ifindex, int fd)
459 {
460         struct sockaddr_nl sa;
461         int sock, seq = 0, len, ret = -1;
462         char buf[4096];
463         struct nlattr *nla, *nla_xdp;
464         struct {
465                 struct nlmsghdr  nh;
466                 struct ifinfomsg ifinfo;
467                 char             attrbuf[64];
468         } req;
469         struct nlmsghdr *nh;
470         struct nlmsgerr *err;
471
472         memset(&sa, 0, sizeof(sa));
473         sa.nl_family = AF_NETLINK;
474
475         sock = socket(AF_NETLINK, SOCK_RAW, NETLINK_ROUTE);
476         if (sock < 0) {
477                 printf("open netlink socket: %s\n", strerror(errno));
478                 return -1;
479         }
480
481         if (bind(sock, (struct sockaddr *)&sa, sizeof(sa)) < 0) {
482                 printf("bind to netlink: %s\n", strerror(errno));
483                 goto cleanup;
484         }
485
486         memset(&req, 0, sizeof(req));
487         req.nh.nlmsg_len = NLMSG_LENGTH(sizeof(struct ifinfomsg));
488         req.nh.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK;
489         req.nh.nlmsg_type = RTM_SETLINK;
490         req.nh.nlmsg_pid = 0;
491         req.nh.nlmsg_seq = ++seq;
492         req.ifinfo.ifi_family = AF_UNSPEC;
493         req.ifinfo.ifi_index = ifindex;
494         nla = (struct nlattr *)(((char *)&req)
495                                 + NLMSG_ALIGN(req.nh.nlmsg_len));
496         nla->nla_type = NLA_F_NESTED | 43/*IFLA_XDP*/;
497
498         nla_xdp = (struct nlattr *)((char *)nla + NLA_HDRLEN);
499         nla_xdp->nla_type = 1/*IFLA_XDP_FD*/;
500         nla_xdp->nla_len = NLA_HDRLEN + sizeof(int);
501         memcpy((char *)nla_xdp + NLA_HDRLEN, &fd, sizeof(fd));
502         nla->nla_len = NLA_HDRLEN + nla_xdp->nla_len;
503
504         req.nh.nlmsg_len += NLA_ALIGN(nla->nla_len);
505
506         if (send(sock, &req, req.nh.nlmsg_len, 0) < 0) {
507                 printf("send to netlink: %s\n", strerror(errno));
508                 goto cleanup;
509         }
510
511         len = recv(sock, buf, sizeof(buf), 0);
512         if (len < 0) {
513                 printf("recv from netlink: %s\n", strerror(errno));
514                 goto cleanup;
515         }
516
517         for (nh = (struct nlmsghdr *)buf; NLMSG_OK(nh, len);
518              nh = NLMSG_NEXT(nh, len)) {
519                 if (nh->nlmsg_pid != getpid()) {
520                         printf("Wrong pid %d, expected %d\n",
521                                nh->nlmsg_pid, getpid());
522                         goto cleanup;
523                 }
524                 if (nh->nlmsg_seq != seq) {
525                         printf("Wrong seq %d, expected %d\n",
526                                nh->nlmsg_seq, seq);
527                         goto cleanup;
528                 }
529                 switch (nh->nlmsg_type) {
530                 case NLMSG_ERROR:
531                         err = (struct nlmsgerr *)NLMSG_DATA(nh);
532                         if (!err->error)
533                                 continue;
534                         printf("nlmsg error %s\n", strerror(-err->error));
535                         goto cleanup;
536                 case NLMSG_DONE:
537                         break;
538                 }
539         }
540
541         ret = 0;
542
543 cleanup:
544         close(sock);
545         return ret;
546 }