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 #include "perf-sys.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 char bpf_log_buf[BPF_LOG_BUF_SIZE];
  34 int map_fd[MAX_MAPS];
  35 int prog_fd[MAX_PROGS];
  36 int event_fd[MAX_PROGS];
  37 int prog_cnt;
  38 int prog_array_fd = -1;
  39
  40 struct bpf_map_def {
  41         unsigned int type;
  42         unsigned int key_size;
  43         unsigned int value_size;
  44         unsigned int max_entries;
  45         unsigned int map_flags;
  46 };
  47
  48 static int populate_prog_array(const char *event, int prog_fd)
  49 {
  50         int ind = atoi(event), err;
  51
  52         err = bpf_map_update_elem(prog_array_fd, &ind, &prog_fd, BPF_ANY);
  53         if (err < 0) {
  54                 printf("failed to store prog_fd in prog_array\n");
  55                 return -1;
  56         }
  57         return 0;
  58 }
  59
  60 static int load_and_attach(const char *event, struct bpf_insn *prog, int size)
  61 {
  62         bool is_socket = strncmp(event, "socket", 6) == 0;
  63         bool is_kprobe = strncmp(event, "kprobe/", 7) == 0;
  64         bool is_kretprobe = strncmp(event, "kretprobe/", 10) == 0;
  65         bool is_tracepoint = strncmp(event, "tracepoint/", 11) == 0;
  66         bool is_xdp = strncmp(event, "xdp", 3) == 0;
  67         bool is_perf_event = strncmp(event, "perf_event", 10) == 0;
  68         bool is_cgroup_skb = strncmp(event, "cgroup/skb", 10) == 0;
  69         bool is_cgroup_sk = strncmp(event, "cgroup/sock", 11) == 0;
  70         size_t insns_cnt = size / sizeof(struct bpf_insn);
  71         enum bpf_prog_type prog_type;
  72         char buf[256];
  73         int fd, efd, err, id;
  74         struct perf_event_attr attr = {};
  75
  76         attr.type = PERF_TYPE_TRACEPOINT;
  77         attr.sample_type = PERF_SAMPLE_RAW;
  78         attr.sample_period = 1;
  79         attr.wakeup_events = 1;
  80
  81         if (is_socket) {
  82                 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
  83         } else if (is_kprobe || is_kretprobe) {
  84                 prog_type = BPF_PROG_TYPE_KPROBE;
  85         } else if (is_tracepoint) {
  86                 prog_type = BPF_PROG_TYPE_TRACEPOINT;
  87         } else if (is_xdp) {
  88                 prog_type = BPF_PROG_TYPE_XDP;
  89         } else if (is_perf_event) {
  90                 prog_type = BPF_PROG_TYPE_PERF_EVENT;
  91         } else if (is_cgroup_skb) {
  92                 prog_type = BPF_PROG_TYPE_CGROUP_SKB;
  93         } else if (is_cgroup_sk) {
  94                 prog_type = BPF_PROG_TYPE_CGROUP_SOCK;
  95         } else {
  96                 printf("Unknown event '%s'\n", event);
  97                 return -1;
  98         }
  99
 100         fd = bpf_load_program(prog_type, prog, insns_cnt, license, kern_version,
 101                               bpf_log_buf, BPF_LOG_BUF_SIZE);
 102         if (fd < 0) {
 103                 printf("bpf_load_program() err=%d\n%s", errno, bpf_log_buf);
 104                 return -1;
 105         }
 106
 107         prog_fd[prog_cnt++] = fd;
 108
 109         if (is_xdp || is_perf_event || is_cgroup_skb || is_cgroup_sk)
 110                 return 0;
 111
 112         if (is_socket) {
 113                 event += 6;
 114                 if (*event != '/')
 115                         return 0;
 116                 event++;
 117                 if (!isdigit(*event)) {
 118                         printf("invalid prog number\n");
 119                         return -1;
 120                 }
 121                 return populate_prog_array(event, fd);
 122         }
 123
 124         if (is_kprobe || is_kretprobe) {
 125                 if (is_kprobe)
 126                         event += 7;
 127                 else
 128                         event += 10;
 129
 130                 if (*event == 0) {
 131                         printf("event name cannot be empty\n");
 132                         return -1;
 133                 }
 134
 135                 if (isdigit(*event))
 136                         return populate_prog_array(event, fd);
 137
 138                 snprintf(buf, sizeof(buf),
 139                          "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
 140                          is_kprobe ? 'p' : 'r', event, event);
 141                 err = system(buf);
 142                 if (err < 0) {
 143                         printf("failed to create kprobe '%s' error '%s'\n",
 144                                event, strerror(errno));
 145                         return -1;
 146                 }
 147
 148                 strcpy(buf, DEBUGFS);
 149                 strcat(buf, "events/kprobes/");
 150                 strcat(buf, event);
 151                 strcat(buf, "/id");
 152         } else if (is_tracepoint) {
 153                 event += 11;
 154
 155                 if (*event == 0) {
 156                         printf("event name cannot be empty\n");
 157                         return -1;
 158                 }
 159                 strcpy(buf, DEBUGFS);
 160                 strcat(buf, "events/");
 161                 strcat(buf, event);
 162                 strcat(buf, "/id");
 163         }
 164
 165         efd = open(buf, O_RDONLY, 0);
 166         if (efd < 0) {
 167                 printf("failed to open event %s\n", event);
 168                 return -1;
 169         }
 170
 171         err = read(efd, buf, sizeof(buf));
 172         if (err < 0 || err >= sizeof(buf)) {
 173                 printf("read from '%s' failed '%s'\n", event, strerror(errno));
 174                 return -1;
 175         }
 176
 177         close(efd);
 178
 179         buf[err] = 0;
 180         id = atoi(buf);
 181         attr.config = id;
 182
 183         efd = sys_perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
 184         if (efd < 0) {
 185                 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
 186                 return -1;
 187         }
 188         event_fd[prog_cnt - 1] = efd;
 189         ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
 190         ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
 191
 192         return 0;
 193 }
 194
 195 static int load_maps(struct bpf_map_def *maps, int len)
 196 {
 197         int i;
 198
 199         for (i = 0; i < len / sizeof(struct bpf_map_def); i++) {
 200
 201                 map_fd[i] = bpf_create_map(maps[i].type,
 202                                            maps[i].key_size,
 203                                            maps[i].value_size,
 204                                            maps[i].max_entries,
 205                                            maps[i].map_flags);
 206                 if (map_fd[i] < 0) {
 207                         printf("failed to create a map: %d %s\n",
 208                                errno, strerror(errno));
 209                         return 1;
 210                 }
 211
 212                 if (maps[i].type == BPF_MAP_TYPE_PROG_ARRAY)
 213                         prog_array_fd = map_fd[i];
 214         }
 215         return 0;
 216 }
 217
 218 static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
 219                    GElf_Shdr *shdr, Elf_Data **data)
 220 {
 221         Elf_Scn *scn;
 222
 223         scn = elf_getscn(elf, i);
 224         if (!scn)
 225                 return 1;
 226
 227         if (gelf_getshdr(scn, shdr) != shdr)
 228                 return 2;
 229
 230         *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
 231         if (!*shname || !shdr->sh_size)
 232                 return 3;
 233
 234         *data = elf_getdata(scn, 0);
 235         if (!*data || elf_getdata(scn, *data) != NULL)
 236                 return 4;
 237
 238         return 0;
 239 }
 240
 241 static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
 242                                 GElf_Shdr *shdr, struct bpf_insn *insn)
 243 {
 244         int i, nrels;
 245
 246         nrels = shdr->sh_size / shdr->sh_entsize;
 247
 248         for (i = 0; i < nrels; i++) {
 249                 GElf_Sym sym;
 250                 GElf_Rel rel;
 251                 unsigned int insn_idx;
 252
 253                 gelf_getrel(data, i, &rel);
 254
 255                 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
 256
 257                 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
 258
 259                 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
 260                         printf("invalid relo for insn[%d].code 0x%x\n",
 261                                insn_idx, insn[insn_idx].code);
 262                         return 1;
 263                 }
 264                 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
 265                 insn[insn_idx].imm = map_fd[sym.st_value / sizeof(struct bpf_map_def)];
 266         }
 267
 268         return 0;
 269 }
 270
 271 int load_bpf_file(char *path)
 272 {
 273         int fd, i;
 274         Elf *elf;
 275         GElf_Ehdr ehdr;
 276         GElf_Shdr shdr, shdr_prog;
 277         Elf_Data *data, *data_prog, *symbols = NULL;
 278         char *shname, *shname_prog;
 279
 280         if (elf_version(EV_CURRENT) == EV_NONE)
 281                 return 1;
 282
 283         fd = open(path, O_RDONLY, 0);
 284         if (fd < 0)
 285                 return 1;
 286
 287         elf = elf_begin(fd, ELF_C_READ, NULL);
 288
 289         if (!elf)
 290                 return 1;
 291
 292         if (gelf_getehdr(elf, &ehdr) != &ehdr)
 293                 return 1;
 294
 295         /* clear all kprobes */
 296         i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
 297
 298         /* scan over all elf sections to get license and map info */
 299         for (i = 1; i < ehdr.e_shnum; i++) {
 300
 301                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 302                         continue;
 303
 304                 if (0) /* helpful for llvm debugging */
 305                         printf("section %d:%s data %p size %zd link %d flags %d\n",
 306                                i, shname, data->d_buf, data->d_size,
 307                                shdr.sh_link, (int) shdr.sh_flags);
 308
 309                 if (strcmp(shname, "license") == 0) {
 310                         processed_sec[i] = true;
 311                         memcpy(license, data->d_buf, data->d_size);
 312                 } else if (strcmp(shname, "version") == 0) {
 313                         processed_sec[i] = true;
 314                         if (data->d_size != sizeof(int)) {
 315                                 printf("invalid size of version section %zd\n",
 316                                        data->d_size);
 317                                 return 1;
 318                         }
 319                         memcpy(&kern_version, data->d_buf, sizeof(int));
 320                 } else if (strcmp(shname, "maps") == 0) {
 321                         processed_sec[i] = true;
 322                         if (load_maps(data->d_buf, data->d_size))
 323                                 return 1;
 324                 } else if (shdr.sh_type == SHT_SYMTAB) {
 325                         symbols = data;
 326                 }
 327         }
 328
 329         /* load programs that need map fixup (relocations) */
 330         for (i = 1; i < ehdr.e_shnum; i++) {
 331
 332                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 333                         continue;
 334                 if (shdr.sh_type == SHT_REL) {
 335                         struct bpf_insn *insns;
 336
 337                         if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
 338                                     &shdr_prog, &data_prog))
 339                                 continue;
 340
 341                         if (shdr_prog.sh_type != SHT_PROGBITS ||
 342                             !(shdr_prog.sh_flags & SHF_EXECINSTR))
 343                                 continue;
 344
 345                         insns = (struct bpf_insn *) data_prog->d_buf;
 346
 347                         processed_sec[shdr.sh_info] = true;
 348                         processed_sec[i] = true;
 349
 350                         if (parse_relo_and_apply(data, symbols, &shdr, insns))
 351                                 continue;
 352
 353                         if (memcmp(shname_prog, "kprobe/", 7) == 0 ||
 354                             memcmp(shname_prog, "kretprobe/", 10) == 0 ||
 355                             memcmp(shname_prog, "tracepoint/", 11) == 0 ||
 356                             memcmp(shname_prog, "xdp", 3) == 0 ||
 357                             memcmp(shname_prog, "perf_event", 10) == 0 ||
 358                             memcmp(shname_prog, "socket", 6) == 0 ||
 359                             memcmp(shname_prog, "cgroup/", 7) == 0)
 360                                 load_and_attach(shname_prog, insns, data_prog->d_size);
 361                 }
 362         }
 363
 364         /* load programs that don't use maps */
 365         for (i = 1; i < ehdr.e_shnum; i++) {
 366
 367                 if (processed_sec[i])
 368                         continue;
 369
 370                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 371                         continue;
 372
 373                 if (memcmp(shname, "kprobe/", 7) == 0 ||
 374                     memcmp(shname, "kretprobe/", 10) == 0 ||
 375                     memcmp(shname, "tracepoint/", 11) == 0 ||
 376                     memcmp(shname, "xdp", 3) == 0 ||
 377                     memcmp(shname, "perf_event", 10) == 0 ||
 378                     memcmp(shname, "socket", 6) == 0 ||
 379                     memcmp(shname, "cgroup/", 7) == 0)
 380                         load_and_attach(shname, data->d_buf, data->d_size);
 381         }
 382
 383         close(fd);
 384         return 0;
 385 }
 386
 387 void read_trace_pipe(void)
 388 {
 389         int trace_fd;
 390
 391         trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
 392         if (trace_fd < 0)
 393                 return;
 394
 395         while (1) {
 396                 static char buf[4096];
 397                 ssize_t sz;
 398
 399                 sz = read(trace_fd, buf, sizeof(buf));
 400                 if (sz > 0) {
 401                         buf[sz] = 0;
 402                         puts(buf);
 403                 }
 404         }
 405 }
 406
 407 #define MAX_SYMS 300000
 408 static struct ksym syms[MAX_SYMS];
 409 static int sym_cnt;
 410
 411 static int ksym_cmp(const void *p1, const void *p2)
 412 {
 413         return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
 414 }
 415
 416 int load_kallsyms(void)
 417 {
 418         FILE *f = fopen("/proc/kallsyms", "r");
 419         char func[256], buf[256];
 420         char symbol;
 421         void *addr;
 422         int i = 0;
 423
 424         if (!f)
 425                 return -ENOENT;
 426
 427         while (!feof(f)) {
 428                 if (!fgets(buf, sizeof(buf), f))
 429                         break;
 430                 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
 431                         break;
 432                 if (!addr)
 433                         continue;
 434                 syms[i].addr = (long) addr;
 435                 syms[i].name = strdup(func);
 436                 i++;
 437         }
 438         sym_cnt = i;
 439         qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
 440         return 0;
 441 }
 442
 443 struct ksym *ksym_search(long key)
 444 {
 445         int start = 0, end = sym_cnt;
 446         int result;
 447
 448         while (start < end) {
 449                 size_t mid = start + (end - start) / 2;
 450
 451                 result = key - syms[mid].addr;
 452                 if (result < 0)
 453                         end = mid;
 454                 else if (result > 0)
 455                         start = mid + 1;
 456                 else
 457                         return &syms[mid];
 458         }
 459
 460         if (start >= 1 && syms[start - 1].addr < key &&
 461             key < syms[start].addr)
 462                 /* valid ksym */
 463                 return &syms[start - 1];
 464
 465         /* out of range. return _stext */
 466         return &syms[0];
 467 }
 468
 469 int set_link_xdp_fd(int ifindex, int fd)
 470 {
 471         struct sockaddr_nl sa;
 472         int sock, seq = 0, len, ret = -1;
 473         char buf[4096];
 474         struct nlattr *nla, *nla_xdp;
 475         struct {
 476                 struct nlmsghdr  nh;
 477                 struct ifinfomsg ifinfo;
 478                 char             attrbuf[64];
 479         } req;
 480         struct nlmsghdr *nh;
 481         struct nlmsgerr *err;
 482
 483         memset(&sa, 0, sizeof(sa));
 484         sa.nl_family = AF_NETLINK;
 485
 486         sock = socket(AF_NETLINK, SOCK_RAW, NETLINK_ROUTE);
 487         if (sock < 0) {
 488                 printf("open netlink socket: %s\n", strerror(errno));
 489                 return -1;
 490         }
 491
 492         if (bind(sock, (struct sockaddr *)&sa, sizeof(sa)) < 0) {
 493                 printf("bind to netlink: %s\n", strerror(errno));
 494                 goto cleanup;
 495         }
 496
 497         memset(&req, 0, sizeof(req));
 498         req.nh.nlmsg_len = NLMSG_LENGTH(sizeof(struct ifinfomsg));
 499         req.nh.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK;
 500         req.nh.nlmsg_type = RTM_SETLINK;
 501         req.nh.nlmsg_pid = 0;
 502         req.nh.nlmsg_seq = ++seq;
 503         req.ifinfo.ifi_family = AF_UNSPEC;
 504         req.ifinfo.ifi_index = ifindex;
 505         nla = (struct nlattr *)(((char *)&req)
 506                                 + NLMSG_ALIGN(req.nh.nlmsg_len));
 507         nla->nla_type = NLA_F_NESTED | 43/*IFLA_XDP*/;
 508
 509         nla_xdp = (struct nlattr *)((char *)nla + NLA_HDRLEN);
 510         nla_xdp->nla_type = 1/*IFLA_XDP_FD*/;
 511         nla_xdp->nla_len = NLA_HDRLEN + sizeof(int);
 512         memcpy((char *)nla_xdp + NLA_HDRLEN, &fd, sizeof(fd));
 513         nla->nla_len = NLA_HDRLEN + nla_xdp->nla_len;
 514
 515         req.nh.nlmsg_len += NLA_ALIGN(nla->nla_len);
 516
 517         if (send(sock, &req, req.nh.nlmsg_len, 0) < 0) {
 518                 printf("send to netlink: %s\n", strerror(errno));
 519                 goto cleanup;
 520         }
 521
 522         len = recv(sock, buf, sizeof(buf), 0);
 523         if (len < 0) {
 524                 printf("recv from netlink: %s\n", strerror(errno));
 525                 goto cleanup;
 526         }
 527
 528         for (nh = (struct nlmsghdr *)buf; NLMSG_OK(nh, len);
 529              nh = NLMSG_NEXT(nh, len)) {
 530                 if (nh->nlmsg_pid != getpid()) {
 531                         printf("Wrong pid %d, expected %d\n",
 532                                nh->nlmsg_pid, getpid());
 533                         goto cleanup;
 534                 }
 535                 if (nh->nlmsg_seq != seq) {
 536                         printf("Wrong seq %d, expected %d\n",
 537                                nh->nlmsg_seq, seq);
 538                         goto cleanup;
 539                 }
 540                 switch (nh->nlmsg_type) {
 541                 case NLMSG_ERROR:
 542                         err = (struct nlmsgerr *)NLMSG_DATA(nh);
 543                         if (!err->error)
 544                                 continue;
 545                         printf("nlmsg error %s\n", strerror(-err->error));
 546                         goto cleanup;
 547                 case NLMSG_DONE:
 548                         break;
 549                 }
 550         }
 551
 552         ret = 0;
 553
 554 cleanup:
 555         close(sock);
 556         return ret;
 557 }