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 <sys/syscall.h>
  16 #include <sys/ioctl.h>
  17 #include <sys/mman.h>
  18 #include <poll.h>
  19 #include <ctype.h>
  20 #include "libbpf.h"
  21 #include "bpf_helpers.h"
  22 #include "bpf_load.h"
  23
  24 #define DEBUGFS "/sys/kernel/debug/tracing/"
  25
  26 static char license[128];
  27 static int kern_version;
  28 static bool processed_sec[128];
  29 int map_fd[MAX_MAPS];
  30 int prog_fd[MAX_PROGS];
  31 int event_fd[MAX_PROGS];
  32 int prog_cnt;
  33 int prog_array_fd = -1;
  34
  35 static int populate_prog_array(const char *event, int prog_fd)
  36 {
  37         int ind = atoi(event), err;
  38
  39         err = bpf_update_elem(prog_array_fd, &ind, &prog_fd, BPF_ANY);
  40         if (err < 0) {
  41                 printf("failed to store prog_fd in prog_array\n");
  42                 return -1;
  43         }
  44         return 0;
  45 }
  46
  47 static int load_and_attach(const char *event, struct bpf_insn *prog, int size)
  48 {
  49         bool is_socket = strncmp(event, "socket", 6) == 0;
  50         bool is_kprobe = strncmp(event, "kprobe/", 7) == 0;
  51         bool is_kretprobe = strncmp(event, "kretprobe/", 10) == 0;
  52         bool is_tracepoint = strncmp(event, "tracepoint/", 11) == 0;
  53         bool is_xdp = strncmp(event, "xdp", 3) == 0;
  54         enum bpf_prog_type prog_type;
  55         char buf[256];
  56         int fd, efd, err, id;
  57         struct perf_event_attr attr = {};
  58
  59         attr.type = PERF_TYPE_TRACEPOINT;
  60         attr.sample_type = PERF_SAMPLE_RAW;
  61         attr.sample_period = 1;
  62         attr.wakeup_events = 1;
  63
  64         if (is_socket) {
  65                 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
  66         } else if (is_kprobe || is_kretprobe) {
  67                 prog_type = BPF_PROG_TYPE_KPROBE;
  68         } else if (is_tracepoint) {
  69                 prog_type = BPF_PROG_TYPE_TRACEPOINT;
  70         } else if (is_xdp) {
  71                 prog_type = BPF_PROG_TYPE_XDP;
  72         } else {
  73                 printf("Unknown event '%s'\n", event);
  74                 return -1;
  75         }
  76
  77         fd = bpf_prog_load(prog_type, prog, size, license, kern_version);
  78         if (fd < 0) {
  79                 printf("bpf_prog_load() err=%d\n%s", errno, bpf_log_buf);
  80                 return -1;
  81         }
  82
  83         prog_fd[prog_cnt++] = fd;
  84
  85         if (is_xdp)
  86                 return 0;
  87
  88         if (is_socket) {
  89                 event += 6;
  90                 if (*event != '/')
  91                         return 0;
  92                 event++;
  93                 if (!isdigit(*event)) {
  94                         printf("invalid prog number\n");
  95                         return -1;
  96                 }
  97                 return populate_prog_array(event, fd);
  98         }
  99
 100         if (is_kprobe || is_kretprobe) {
 101                 if (is_kprobe)
 102                         event += 7;
 103                 else
 104                         event += 10;
 105
 106                 if (*event == 0) {
 107                         printf("event name cannot be empty\n");
 108                         return -1;
 109                 }
 110
 111                 if (isdigit(*event))
 112                         return populate_prog_array(event, fd);
 113
 114                 snprintf(buf, sizeof(buf),
 115                          "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
 116                          is_kprobe ? 'p' : 'r', event, event);
 117                 err = system(buf);
 118                 if (err < 0) {
 119                         printf("failed to create kprobe '%s' error '%s'\n",
 120                                event, strerror(errno));
 121                         return -1;
 122                 }
 123
 124                 strcpy(buf, DEBUGFS);
 125                 strcat(buf, "events/kprobes/");
 126                 strcat(buf, event);
 127                 strcat(buf, "/id");
 128         } else if (is_tracepoint) {
 129                 event += 11;
 130
 131                 if (*event == 0) {
 132                         printf("event name cannot be empty\n");
 133                         return -1;
 134                 }
 135                 strcpy(buf, DEBUGFS);
 136                 strcat(buf, "events/");
 137                 strcat(buf, event);
 138                 strcat(buf, "/id");
 139         }
 140
 141         efd = open(buf, O_RDONLY, 0);
 142         if (efd < 0) {
 143                 printf("failed to open event %s\n", event);
 144                 return -1;
 145         }
 146
 147         err = read(efd, buf, sizeof(buf));
 148         if (err < 0 || err >= sizeof(buf)) {
 149                 printf("read from '%s' failed '%s'\n", event, strerror(errno));
 150                 return -1;
 151         }
 152
 153         close(efd);
 154
 155         buf[err] = 0;
 156         id = atoi(buf);
 157         attr.config = id;
 158
 159         efd = perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
 160         if (efd < 0) {
 161                 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
 162                 return -1;
 163         }
 164         event_fd[prog_cnt - 1] = efd;
 165         ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
 166         ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
 167
 168         return 0;
 169 }
 170
 171 static int load_maps(struct bpf_map_def *maps, int len)
 172 {
 173         int i;
 174
 175         for (i = 0; i < len / sizeof(struct bpf_map_def); i++) {
 176
 177                 map_fd[i] = bpf_create_map(maps[i].type,
 178                                            maps[i].key_size,
 179                                            maps[i].value_size,
 180                                            maps[i].max_entries,
 181                                            maps[i].map_flags);
 182                 if (map_fd[i] < 0) {
 183                         printf("failed to create a map: %d %s\n",
 184                                errno, strerror(errno));
 185                         return 1;
 186                 }
 187
 188                 if (maps[i].type == BPF_MAP_TYPE_PROG_ARRAY)
 189                         prog_array_fd = map_fd[i];
 190         }
 191         return 0;
 192 }
 193
 194 static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
 195                    GElf_Shdr *shdr, Elf_Data **data)
 196 {
 197         Elf_Scn *scn;
 198
 199         scn = elf_getscn(elf, i);
 200         if (!scn)
 201                 return 1;
 202
 203         if (gelf_getshdr(scn, shdr) != shdr)
 204                 return 2;
 205
 206         *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
 207         if (!*shname || !shdr->sh_size)
 208                 return 3;
 209
 210         *data = elf_getdata(scn, 0);
 211         if (!*data || elf_getdata(scn, *data) != NULL)
 212                 return 4;
 213
 214         return 0;
 215 }
 216
 217 static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
 218                                 GElf_Shdr *shdr, struct bpf_insn *insn)
 219 {
 220         int i, nrels;
 221
 222         nrels = shdr->sh_size / shdr->sh_entsize;
 223
 224         for (i = 0; i < nrels; i++) {
 225                 GElf_Sym sym;
 226                 GElf_Rel rel;
 227                 unsigned int insn_idx;
 228
 229                 gelf_getrel(data, i, &rel);
 230
 231                 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
 232
 233                 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
 234
 235                 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
 236                         printf("invalid relo for insn[%d].code 0x%x\n",
 237                                insn_idx, insn[insn_idx].code);
 238                         return 1;
 239                 }
 240                 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
 241                 insn[insn_idx].imm = map_fd[sym.st_value / sizeof(struct bpf_map_def)];
 242         }
 243
 244         return 0;
 245 }
 246
 247 int load_bpf_file(char *path)
 248 {
 249         int fd, i;
 250         Elf *elf;
 251         GElf_Ehdr ehdr;
 252         GElf_Shdr shdr, shdr_prog;
 253         Elf_Data *data, *data_prog, *symbols = NULL;
 254         char *shname, *shname_prog;
 255
 256         if (elf_version(EV_CURRENT) == EV_NONE)
 257                 return 1;
 258
 259         fd = open(path, O_RDONLY, 0);
 260         if (fd < 0)
 261                 return 1;
 262
 263         elf = elf_begin(fd, ELF_C_READ, NULL);
 264
 265         if (!elf)
 266                 return 1;
 267
 268         if (gelf_getehdr(elf, &ehdr) != &ehdr)
 269                 return 1;
 270
 271         /* clear all kprobes */
 272         i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
 273
 274         /* scan over all elf sections to get license and map info */
 275         for (i = 1; i < ehdr.e_shnum; i++) {
 276
 277                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 278                         continue;
 279
 280                 if (0) /* helpful for llvm debugging */
 281                         printf("section %d:%s data %p size %zd link %d flags %d\n",
 282                                i, shname, data->d_buf, data->d_size,
 283                                shdr.sh_link, (int) shdr.sh_flags);
 284
 285                 if (strcmp(shname, "license") == 0) {
 286                         processed_sec[i] = true;
 287                         memcpy(license, data->d_buf, data->d_size);
 288                 } else if (strcmp(shname, "version") == 0) {
 289                         processed_sec[i] = true;
 290                         if (data->d_size != sizeof(int)) {
 291                                 printf("invalid size of version section %zd\n",
 292                                        data->d_size);
 293                                 return 1;
 294                         }
 295                         memcpy(&kern_version, data->d_buf, sizeof(int));
 296                 } else if (strcmp(shname, "maps") == 0) {
 297                         processed_sec[i] = true;
 298                         if (load_maps(data->d_buf, data->d_size))
 299                                 return 1;
 300                 } else if (shdr.sh_type == SHT_SYMTAB) {
 301                         symbols = data;
 302                 }
 303         }
 304
 305         /* load programs that need map fixup (relocations) */
 306         for (i = 1; i < ehdr.e_shnum; i++) {
 307
 308                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 309                         continue;
 310                 if (shdr.sh_type == SHT_REL) {
 311                         struct bpf_insn *insns;
 312
 313                         if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
 314                                     &shdr_prog, &data_prog))
 315                                 continue;
 316
 317                         insns = (struct bpf_insn *) data_prog->d_buf;
 318
 319                         processed_sec[shdr.sh_info] = true;
 320                         processed_sec[i] = true;
 321
 322                         if (parse_relo_and_apply(data, symbols, &shdr, insns))
 323                                 continue;
 324
 325                         if (memcmp(shname_prog, "kprobe/", 7) == 0 ||
 326                             memcmp(shname_prog, "kretprobe/", 10) == 0 ||
 327                             memcmp(shname_prog, "tracepoint/", 11) == 0 ||
 328                             memcmp(shname_prog, "xdp", 3) == 0 ||
 329                             memcmp(shname_prog, "socket", 6) == 0)
 330                                 load_and_attach(shname_prog, insns, data_prog->d_size);
 331                 }
 332         }
 333
 334         /* load programs that don't use maps */
 335         for (i = 1; i < ehdr.e_shnum; i++) {
 336
 337                 if (processed_sec[i])
 338                         continue;
 339
 340                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 341                         continue;
 342
 343                 if (memcmp(shname, "kprobe/", 7) == 0 ||
 344                     memcmp(shname, "kretprobe/", 10) == 0 ||
 345                     memcmp(shname, "tracepoint/", 11) == 0 ||
 346                     memcmp(shname, "xdp", 3) == 0 ||
 347                     memcmp(shname, "socket", 6) == 0)
 348                         load_and_attach(shname, data->d_buf, data->d_size);
 349         }
 350
 351         close(fd);
 352         return 0;
 353 }
 354
 355 void read_trace_pipe(void)
 356 {
 357         int trace_fd;
 358
 359         trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
 360         if (trace_fd < 0)
 361                 return;
 362
 363         while (1) {
 364                 static char buf[4096];
 365                 ssize_t sz;
 366
 367                 sz = read(trace_fd, buf, sizeof(buf));
 368                 if (sz > 0) {
 369                         buf[sz] = 0;
 370                         puts(buf);
 371                 }
 372         }
 373 }
 374
 375 #define MAX_SYMS 300000
 376 static struct ksym syms[MAX_SYMS];
 377 static int sym_cnt;
 378
 379 static int ksym_cmp(const void *p1, const void *p2)
 380 {
 381         return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
 382 }
 383
 384 int load_kallsyms(void)
 385 {
 386         FILE *f = fopen("/proc/kallsyms", "r");
 387         char func[256], buf[256];
 388         char symbol;
 389         void *addr;
 390         int i = 0;
 391
 392         if (!f)
 393                 return -ENOENT;
 394
 395         while (!feof(f)) {
 396                 if (!fgets(buf, sizeof(buf), f))
 397                         break;
 398                 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
 399                         break;
 400                 if (!addr)
 401                         continue;
 402                 syms[i].addr = (long) addr;
 403                 syms[i].name = strdup(func);
 404                 i++;
 405         }
 406         sym_cnt = i;
 407         qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
 408         return 0;
 409 }
 410
 411 struct ksym *ksym_search(long key)
 412 {
 413         int start = 0, end = sym_cnt;
 414         int result;
 415
 416         while (start < end) {
 417                 size_t mid = start + (end - start) / 2;
 418
 419                 result = key - syms[mid].addr;
 420                 if (result < 0)
 421                         end = mid;
 422                 else if (result > 0)
 423                         start = mid + 1;
 424                 else
 425                         return &syms[mid];
 426         }
 427
 428         if (start >= 1 && syms[start - 1].addr < key &&
 429             key < syms[start].addr)
 430                 /* valid ksym */
 431                 return &syms[start - 1];
 432
 433         /* out of range. return _stext */
 434         return &syms[0];
 435 }