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_load.h"
  26
  27 #define DEBUGFS "/sys/kernel/debug/tracing/"
  28
  29 static char license[128];
  30 static int kern_version;
  31 static bool processed_sec[128];
  32 char bpf_log_buf[BPF_LOG_BUF_SIZE];
  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 struct bpf_map_def {
  40         unsigned int type;
  41         unsigned int key_size;
  42         unsigned int value_size;
  43         unsigned int max_entries;
  44         unsigned int map_flags;
  45 };
  46
  47 static int populate_prog_array(const char *event, int prog_fd)
  48 {
  49         int ind = atoi(event), err;
  50
  51         err = bpf_map_update_elem(prog_array_fd, &ind, &prog_fd, BPF_ANY);
  52         if (err < 0) {
  53                 printf("failed to store prog_fd in prog_array\n");
  54                 return -1;
  55         }
  56         return 0;
  57 }
  58
  59 static int load_and_attach(const char *event, struct bpf_insn *prog, int size)
  60 {
  61         bool is_socket = strncmp(event, "socket", 6) == 0;
  62         bool is_kprobe = strncmp(event, "kprobe/", 7) == 0;
  63         bool is_kretprobe = strncmp(event, "kretprobe/", 10) == 0;
  64         bool is_tracepoint = strncmp(event, "tracepoint/", 11) == 0;
  65         bool is_xdp = strncmp(event, "xdp", 3) == 0;
  66         bool is_perf_event = strncmp(event, "perf_event", 10) == 0;
  67         bool is_cgroup_skb = strncmp(event, "cgroup/skb", 10) == 0;
  68         bool is_cgroup_sk = strncmp(event, "cgroup/sock", 11) == 0;
  69         enum bpf_prog_type prog_type;
  70         char buf[256];
  71         int fd, efd, err, id;
  72         struct perf_event_attr attr = {};
  73
  74         attr.type = PERF_TYPE_TRACEPOINT;
  75         attr.sample_type = PERF_SAMPLE_RAW;
  76         attr.sample_period = 1;
  77         attr.wakeup_events = 1;
  78
  79         if (is_socket) {
  80                 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
  81         } else if (is_kprobe || is_kretprobe) {
  82                 prog_type = BPF_PROG_TYPE_KPROBE;
  83         } else if (is_tracepoint) {
  84                 prog_type = BPF_PROG_TYPE_TRACEPOINT;
  85         } else if (is_xdp) {
  86                 prog_type = BPF_PROG_TYPE_XDP;
  87         } else if (is_perf_event) {
  88                 prog_type = BPF_PROG_TYPE_PERF_EVENT;
  89         } else if (is_cgroup_skb) {
  90                 prog_type = BPF_PROG_TYPE_CGROUP_SKB;
  91         } else if (is_cgroup_sk) {
  92                 prog_type = BPF_PROG_TYPE_CGROUP_SOCK;
  93         } else {
  94                 printf("Unknown event '%s'\n", event);
  95                 return -1;
  96         }
  97
  98         fd = bpf_load_program(prog_type, prog, size, license, kern_version,
  99                               bpf_log_buf, BPF_LOG_BUF_SIZE);
 100         if (fd < 0) {
 101                 printf("bpf_load_program() err=%d\n%s", errno, bpf_log_buf);
 102                 return -1;
 103         }
 104
 105         prog_fd[prog_cnt++] = fd;
 106
 107         if (is_xdp || is_perf_event || is_cgroup_skb || is_cgroup_sk)
 108                 return 0;
 109
 110         if (is_socket) {
 111                 event += 6;
 112                 if (*event != '/')
 113                         return 0;
 114                 event++;
 115                 if (!isdigit(*event)) {
 116                         printf("invalid prog number\n");
 117                         return -1;
 118                 }
 119                 return populate_prog_array(event, fd);
 120         }
 121
 122         if (is_kprobe || is_kretprobe) {
 123                 if (is_kprobe)
 124                         event += 7;
 125                 else
 126                         event += 10;
 127
 128                 if (*event == 0) {
 129                         printf("event name cannot be empty\n");
 130                         return -1;
 131                 }
 132
 133                 if (isdigit(*event))
 134                         return populate_prog_array(event, fd);
 135
 136                 snprintf(buf, sizeof(buf),
 137                          "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
 138                          is_kprobe ? 'p' : 'r', event, event);
 139                 err = system(buf);
 140                 if (err < 0) {
 141                         printf("failed to create kprobe '%s' error '%s'\n",
 142                                event, strerror(errno));
 143                         return -1;
 144                 }
 145
 146                 strcpy(buf, DEBUGFS);
 147                 strcat(buf, "events/kprobes/");
 148                 strcat(buf, event);
 149                 strcat(buf, "/id");
 150         } else if (is_tracepoint) {
 151                 event += 11;
 152
 153                 if (*event == 0) {
 154                         printf("event name cannot be empty\n");
 155                         return -1;
 156                 }
 157                 strcpy(buf, DEBUGFS);
 158                 strcat(buf, "events/");
 159                 strcat(buf, event);
 160                 strcat(buf, "/id");
 161         }
 162
 163         efd = open(buf, O_RDONLY, 0);
 164         if (efd < 0) {
 165                 printf("failed to open event %s\n", event);
 166                 return -1;
 167         }
 168
 169         err = read(efd, buf, sizeof(buf));
 170         if (err < 0 || err >= sizeof(buf)) {
 171                 printf("read from '%s' failed '%s'\n", event, strerror(errno));
 172                 return -1;
 173         }
 174
 175         close(efd);
 176
 177         buf[err] = 0;
 178         id = atoi(buf);
 179         attr.config = id;
 180
 181         efd = perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
 182         if (efd < 0) {
 183                 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
 184                 return -1;
 185         }
 186         event_fd[prog_cnt - 1] = efd;
 187         ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
 188         ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
 189
 190         return 0;
 191 }
 192
 193 static int load_maps(struct bpf_map_def *maps, int len)
 194 {
 195         int i;
 196
 197         for (i = 0; i < len / sizeof(struct bpf_map_def); i++) {
 198
 199                 map_fd[i] = bpf_create_map(maps[i].type,
 200                                            maps[i].key_size,
 201                                            maps[i].value_size,
 202                                            maps[i].max_entries,
 203                                            maps[i].map_flags);
 204                 if (map_fd[i] < 0) {
 205                         printf("failed to create a map: %d %s\n",
 206                                errno, strerror(errno));
 207                         return 1;
 208                 }
 209
 210                 if (maps[i].type == BPF_MAP_TYPE_PROG_ARRAY)
 211                         prog_array_fd = map_fd[i];
 212         }
 213         return 0;
 214 }
 215
 216 static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
 217                    GElf_Shdr *shdr, Elf_Data **data)
 218 {
 219         Elf_Scn *scn;
 220
 221         scn = elf_getscn(elf, i);
 222         if (!scn)
 223                 return 1;
 224
 225         if (gelf_getshdr(scn, shdr) != shdr)
 226                 return 2;
 227
 228         *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
 229         if (!*shname || !shdr->sh_size)
 230                 return 3;
 231
 232         *data = elf_getdata(scn, 0);
 233         if (!*data || elf_getdata(scn, *data) != NULL)
 234                 return 4;
 235
 236         return 0;
 237 }
 238
 239 static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
 240                                 GElf_Shdr *shdr, struct bpf_insn *insn)
 241 {
 242         int i, nrels;
 243
 244         nrels = shdr->sh_size / shdr->sh_entsize;
 245
 246         for (i = 0; i < nrels; i++) {
 247                 GElf_Sym sym;
 248                 GElf_Rel rel;
 249                 unsigned int insn_idx;
 250
 251                 gelf_getrel(data, i, &rel);
 252
 253                 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
 254
 255                 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
 256
 257                 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
 258                         printf("invalid relo for insn[%d].code 0x%x\n",
 259                                insn_idx, insn[insn_idx].code);
 260                         return 1;
 261                 }
 262                 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
 263                 insn[insn_idx].imm = map_fd[sym.st_value / sizeof(struct bpf_map_def)];
 264         }
 265
 266         return 0;
 267 }
 268
 269 int load_bpf_file(char *path)
 270 {
 271         int fd, i;
 272         Elf *elf;
 273         GElf_Ehdr ehdr;
 274         GElf_Shdr shdr, shdr_prog;
 275         Elf_Data *data, *data_prog, *symbols = NULL;
 276         char *shname, *shname_prog;
 277
 278         if (elf_version(EV_CURRENT) == EV_NONE)
 279                 return 1;
 280
 281         fd = open(path, O_RDONLY, 0);
 282         if (fd < 0)
 283                 return 1;
 284
 285         elf = elf_begin(fd, ELF_C_READ, NULL);
 286
 287         if (!elf)
 288                 return 1;
 289
 290         if (gelf_getehdr(elf, &ehdr) != &ehdr)
 291                 return 1;
 292
 293         /* clear all kprobes */
 294         i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
 295
 296         /* scan over all elf sections to get license and map info */
 297         for (i = 1; i < ehdr.e_shnum; i++) {
 298
 299                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 300                         continue;
 301
 302                 if (0) /* helpful for llvm debugging */
 303                         printf("section %d:%s data %p size %zd link %d flags %d\n",
 304                                i, shname, data->d_buf, data->d_size,
 305                                shdr.sh_link, (int) shdr.sh_flags);
 306
 307                 if (strcmp(shname, "license") == 0) {
 308                         processed_sec[i] = true;
 309                         memcpy(license, data->d_buf, data->d_size);
 310                 } else if (strcmp(shname, "version") == 0) {
 311                         processed_sec[i] = true;
 312                         if (data->d_size != sizeof(int)) {
 313                                 printf("invalid size of version section %zd\n",
 314                                        data->d_size);
 315                                 return 1;
 316                         }
 317                         memcpy(&kern_version, data->d_buf, sizeof(int));
 318                 } else if (strcmp(shname, "maps") == 0) {
 319                         processed_sec[i] = true;
 320                         if (load_maps(data->d_buf, data->d_size))
 321                                 return 1;
 322                 } else if (shdr.sh_type == SHT_SYMTAB) {
 323                         symbols = data;
 324                 }
 325         }
 326
 327         /* load programs that need map fixup (relocations) */
 328         for (i = 1; i < ehdr.e_shnum; i++) {
 329
 330                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 331                         continue;
 332                 if (shdr.sh_type == SHT_REL) {
 333                         struct bpf_insn *insns;
 334
 335                         if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
 336                                     &shdr_prog, &data_prog))
 337                                 continue;
 338
 339                         if (shdr_prog.sh_type != SHT_PROGBITS ||
 340                             !(shdr_prog.sh_flags & SHF_EXECINSTR))
 341                                 continue;
 342
 343                         insns = (struct bpf_insn *) data_prog->d_buf;
 344
 345                         processed_sec[shdr.sh_info] = true;
 346                         processed_sec[i] = true;
 347
 348                         if (parse_relo_and_apply(data, symbols, &shdr, insns))
 349                                 continue;
 350
 351                         if (memcmp(shname_prog, "kprobe/", 7) == 0 ||
 352                             memcmp(shname_prog, "kretprobe/", 10) == 0 ||
 353                             memcmp(shname_prog, "tracepoint/", 11) == 0 ||
 354                             memcmp(shname_prog, "xdp", 3) == 0 ||
 355                             memcmp(shname_prog, "perf_event", 10) == 0 ||
 356                             memcmp(shname_prog, "socket", 6) == 0 ||
 357                             memcmp(shname_prog, "cgroup/", 7) == 0)
 358                                 load_and_attach(shname_prog, insns, data_prog->d_size);
 359                 }
 360         }
 361
 362         /* load programs that don't use maps */
 363         for (i = 1; i < ehdr.e_shnum; i++) {
 364
 365                 if (processed_sec[i])
 366                         continue;
 367
 368                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 369                         continue;
 370
 371                 if (memcmp(shname, "kprobe/", 7) == 0 ||
 372                     memcmp(shname, "kretprobe/", 10) == 0 ||
 373                     memcmp(shname, "tracepoint/", 11) == 0 ||
 374                     memcmp(shname, "xdp", 3) == 0 ||
 375                     memcmp(shname, "perf_event", 10) == 0 ||
 376                     memcmp(shname, "socket", 6) == 0 ||
 377                     memcmp(shname, "cgroup/", 7) == 0)
 378                         load_and_attach(shname, data->d_buf, data->d_size);
 379         }
 380
 381         close(fd);
 382         return 0;
 383 }
 384
 385 void read_trace_pipe(void)
 386 {
 387         int trace_fd;
 388
 389         trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
 390         if (trace_fd < 0)
 391                 return;
 392
 393         while (1) {
 394                 static char buf[4096];
 395                 ssize_t sz;
 396
 397                 sz = read(trace_fd, buf, sizeof(buf));
 398                 if (sz > 0) {
 399                         buf[sz] = 0;
 400                         puts(buf);
 401                 }
 402         }
 403 }
 404
 405 #define MAX_SYMS 300000
 406 static struct ksym syms[MAX_SYMS];
 407 static int sym_cnt;
 408
 409 static int ksym_cmp(const void *p1, const void *p2)
 410 {
 411         return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
 412 }
 413
 414 int load_kallsyms(void)
 415 {
 416         FILE *f = fopen("/proc/kallsyms", "r");
 417         char func[256], buf[256];
 418         char symbol;
 419         void *addr;
 420         int i = 0;
 421
 422         if (!f)
 423                 return -ENOENT;
 424
 425         while (!feof(f)) {
 426                 if (!fgets(buf, sizeof(buf), f))
 427                         break;
 428                 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
 429                         break;
 430                 if (!addr)
 431                         continue;
 432                 syms[i].addr = (long) addr;
 433                 syms[i].name = strdup(func);
 434                 i++;
 435         }
 436         sym_cnt = i;
 437         qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
 438         return 0;
 439 }
 440
 441 struct ksym *ksym_search(long key)
 442 {
 443         int start = 0, end = sym_cnt;
 444         int result;
 445
 446         while (start < end) {
 447                 size_t mid = start + (end - start) / 2;
 448
 449                 result = key - syms[mid].addr;
 450                 if (result < 0)
 451                         end = mid;
 452                 else if (result > 0)
 453                         start = mid + 1;
 454                 else
 455                         return &syms[mid];
 456         }
 457
 458         if (start >= 1 && syms[start - 1].addr < key &&
 459             key < syms[start].addr)
 460                 /* valid ksym */
 461                 return &syms[start - 1];
 462
 463         /* out of range. return _stext */
 464         return &syms[0];
 465 }
 466
 467 int set_link_xdp_fd(int ifindex, int fd)
 468 {
 469         struct sockaddr_nl sa;
 470         int sock, seq = 0, len, ret = -1;
 471         char buf[4096];
 472         struct nlattr *nla, *nla_xdp;
 473         struct {
 474                 struct nlmsghdr  nh;
 475                 struct ifinfomsg ifinfo;
 476                 char             attrbuf[64];
 477         } req;
 478         struct nlmsghdr *nh;
 479         struct nlmsgerr *err;
 480
 481         memset(&sa, 0, sizeof(sa));
 482         sa.nl_family = AF_NETLINK;
 483
 484         sock = socket(AF_NETLINK, SOCK_RAW, NETLINK_ROUTE);
 485         if (sock < 0) {
 486                 printf("open netlink socket: %s\n", strerror(errno));
 487                 return -1;
 488         }
 489
 490         if (bind(sock, (struct sockaddr *)&sa, sizeof(sa)) < 0) {
 491                 printf("bind to netlink: %s\n", strerror(errno));
 492                 goto cleanup;
 493         }
 494
 495         memset(&req, 0, sizeof(req));
 496         req.nh.nlmsg_len = NLMSG_LENGTH(sizeof(struct ifinfomsg));
 497         req.nh.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK;
 498         req.nh.nlmsg_type = RTM_SETLINK;
 499         req.nh.nlmsg_pid = 0;
 500         req.nh.nlmsg_seq = ++seq;
 501         req.ifinfo.ifi_family = AF_UNSPEC;
 502         req.ifinfo.ifi_index = ifindex;
 503         nla = (struct nlattr *)(((char *)&req)
 504                                 + NLMSG_ALIGN(req.nh.nlmsg_len));
 505         nla->nla_type = NLA_F_NESTED | 43/*IFLA_XDP*/;
 506
 507         nla_xdp = (struct nlattr *)((char *)nla + NLA_HDRLEN);
 508         nla_xdp->nla_type = 1/*IFLA_XDP_FD*/;
 509         nla_xdp->nla_len = NLA_HDRLEN + sizeof(int);
 510         memcpy((char *)nla_xdp + NLA_HDRLEN, &fd, sizeof(fd));
 511         nla->nla_len = NLA_HDRLEN + nla_xdp->nla_len;
 512
 513         req.nh.nlmsg_len += NLA_ALIGN(nla->nla_len);
 514
 515         if (send(sock, &req, req.nh.nlmsg_len, 0) < 0) {
 516                 printf("send to netlink: %s\n", strerror(errno));
 517                 goto cleanup;
 518         }
 519
 520         len = recv(sock, buf, sizeof(buf), 0);
 521         if (len < 0) {
 522                 printf("recv from netlink: %s\n", strerror(errno));
 523                 goto cleanup;
 524         }
 525
 526         for (nh = (struct nlmsghdr *)buf; NLMSG_OK(nh, len);
 527              nh = NLMSG_NEXT(nh, len)) {
 528                 if (nh->nlmsg_pid != getpid()) {
 529                         printf("Wrong pid %d, expected %d\n",
 530                                nh->nlmsg_pid, getpid());
 531                         goto cleanup;
 532                 }
 533                 if (nh->nlmsg_seq != seq) {
 534                         printf("Wrong seq %d, expected %d\n",
 535                                nh->nlmsg_seq, seq);
 536                         goto cleanup;
 537                 }
 538                 switch (nh->nlmsg_type) {
 539                 case NLMSG_ERROR:
 540                         err = (struct nlmsgerr *)NLMSG_DATA(nh);
 541                         if (!err->error)
 542                                 continue;
 543                         printf("nlmsg error %s\n", strerror(-err->error));
 544                         goto cleanup;
 545                 case NLMSG_DONE:
 546                         break;
 547                 }
 548         }
 549
 550         ret = 0;
 551
 552 cleanup:
 553         close(sock);
 554         return ret;
 555 }