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         bool is_perf_event = strncmp(event, "perf_event", 10) == 0;
  55         enum bpf_prog_type prog_type;
  56         char buf[256];
  57         int fd, efd, err, id;
  58         struct perf_event_attr attr = {};
  59
  60         attr.type = PERF_TYPE_TRACEPOINT;
  61         attr.sample_type = PERF_SAMPLE_RAW;
  62         attr.sample_period = 1;
  63         attr.wakeup_events = 1;
  64
  65         if (is_socket) {
  66                 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
  67         } else if (is_kprobe || is_kretprobe) {
  68                 prog_type = BPF_PROG_TYPE_KPROBE;
  69         } else if (is_tracepoint) {
  70                 prog_type = BPF_PROG_TYPE_TRACEPOINT;
  71         } else if (is_xdp) {
  72                 prog_type = BPF_PROG_TYPE_XDP;
  73         } else if (is_perf_event) {
  74                 prog_type = BPF_PROG_TYPE_PERF_EVENT;
  75         } else {
  76                 printf("Unknown event '%s'\n", event);
  77                 return -1;
  78         }
  79
  80         fd = bpf_prog_load(prog_type, prog, size, license, kern_version);
  81         if (fd < 0) {
  82                 printf("bpf_prog_load() err=%d\n%s", errno, bpf_log_buf);
  83                 return -1;
  84         }
  85
  86         prog_fd[prog_cnt++] = fd;
  87
  88         if (is_xdp || is_perf_event)
  89                 return 0;
  90
  91         if (is_socket) {
  92                 event += 6;
  93                 if (*event != '/')
  94                         return 0;
  95                 event++;
  96                 if (!isdigit(*event)) {
  97                         printf("invalid prog number\n");
  98                         return -1;
  99                 }
 100                 return populate_prog_array(event, fd);
 101         }
 102
 103         if (is_kprobe || is_kretprobe) {
 104                 if (is_kprobe)
 105                         event += 7;
 106                 else
 107                         event += 10;
 108
 109                 if (*event == 0) {
 110                         printf("event name cannot be empty\n");
 111                         return -1;
 112                 }
 113
 114                 if (isdigit(*event))
 115                         return populate_prog_array(event, fd);
 116
 117                 snprintf(buf, sizeof(buf),
 118                          "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
 119                          is_kprobe ? 'p' : 'r', event, event);
 120                 err = system(buf);
 121                 if (err < 0) {
 122                         printf("failed to create kprobe '%s' error '%s'\n",
 123                                event, strerror(errno));
 124                         return -1;
 125                 }
 126
 127                 strcpy(buf, DEBUGFS);
 128                 strcat(buf, "events/kprobes/");
 129                 strcat(buf, event);
 130                 strcat(buf, "/id");
 131         } else if (is_tracepoint) {
 132                 event += 11;
 133
 134                 if (*event == 0) {
 135                         printf("event name cannot be empty\n");
 136                         return -1;
 137                 }
 138                 strcpy(buf, DEBUGFS);
 139                 strcat(buf, "events/");
 140                 strcat(buf, event);
 141                 strcat(buf, "/id");
 142         }
 143
 144         efd = open(buf, O_RDONLY, 0);
 145         if (efd < 0) {
 146                 printf("failed to open event %s\n", event);
 147                 return -1;
 148         }
 149
 150         err = read(efd, buf, sizeof(buf));
 151         if (err < 0 || err >= sizeof(buf)) {
 152                 printf("read from '%s' failed '%s'\n", event, strerror(errno));
 153                 return -1;
 154         }
 155
 156         close(efd);
 157
 158         buf[err] = 0;
 159         id = atoi(buf);
 160         attr.config = id;
 161
 162         efd = perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
 163         if (efd < 0) {
 164                 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
 165                 return -1;
 166         }
 167         event_fd[prog_cnt - 1] = efd;
 168         ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
 169         ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
 170
 171         return 0;
 172 }
 173
 174 static int load_maps(struct bpf_map_def *maps, int len)
 175 {
 176         int i;
 177
 178         for (i = 0; i < len / sizeof(struct bpf_map_def); i++) {
 179
 180                 map_fd[i] = bpf_create_map(maps[i].type,
 181                                            maps[i].key_size,
 182                                            maps[i].value_size,
 183                                            maps[i].max_entries,
 184                                            maps[i].map_flags);
 185                 if (map_fd[i] < 0) {
 186                         printf("failed to create a map: %d %s\n",
 187                                errno, strerror(errno));
 188                         return 1;
 189                 }
 190
 191                 if (maps[i].type == BPF_MAP_TYPE_PROG_ARRAY)
 192                         prog_array_fd = map_fd[i];
 193         }
 194         return 0;
 195 }
 196
 197 static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
 198                    GElf_Shdr *shdr, Elf_Data **data)
 199 {
 200         Elf_Scn *scn;
 201
 202         scn = elf_getscn(elf, i);
 203         if (!scn)
 204                 return 1;
 205
 206         if (gelf_getshdr(scn, shdr) != shdr)
 207                 return 2;
 208
 209         *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
 210         if (!*shname || !shdr->sh_size)
 211                 return 3;
 212
 213         *data = elf_getdata(scn, 0);
 214         if (!*data || elf_getdata(scn, *data) != NULL)
 215                 return 4;
 216
 217         return 0;
 218 }
 219
 220 static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
 221                                 GElf_Shdr *shdr, struct bpf_insn *insn)
 222 {
 223         int i, nrels;
 224
 225         nrels = shdr->sh_size / shdr->sh_entsize;
 226
 227         for (i = 0; i < nrels; i++) {
 228                 GElf_Sym sym;
 229                 GElf_Rel rel;
 230                 unsigned int insn_idx;
 231
 232                 gelf_getrel(data, i, &rel);
 233
 234                 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
 235
 236                 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
 237
 238                 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
 239                         printf("invalid relo for insn[%d].code 0x%x\n",
 240                                insn_idx, insn[insn_idx].code);
 241                         return 1;
 242                 }
 243                 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
 244                 insn[insn_idx].imm = map_fd[sym.st_value / sizeof(struct bpf_map_def)];
 245         }
 246
 247         return 0;
 248 }
 249
 250 int load_bpf_file(char *path)
 251 {
 252         int fd, i;
 253         Elf *elf;
 254         GElf_Ehdr ehdr;
 255         GElf_Shdr shdr, shdr_prog;
 256         Elf_Data *data, *data_prog, *symbols = NULL;
 257         char *shname, *shname_prog;
 258
 259         if (elf_version(EV_CURRENT) == EV_NONE)
 260                 return 1;
 261
 262         fd = open(path, O_RDONLY, 0);
 263         if (fd < 0)
 264                 return 1;
 265
 266         elf = elf_begin(fd, ELF_C_READ, NULL);
 267
 268         if (!elf)
 269                 return 1;
 270
 271         if (gelf_getehdr(elf, &ehdr) != &ehdr)
 272                 return 1;
 273
 274         /* clear all kprobes */
 275         i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
 276
 277         /* scan over all elf sections to get license and map info */
 278         for (i = 1; i < ehdr.e_shnum; i++) {
 279
 280                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 281                         continue;
 282
 283                 if (0) /* helpful for llvm debugging */
 284                         printf("section %d:%s data %p size %zd link %d flags %d\n",
 285                                i, shname, data->d_buf, data->d_size,
 286                                shdr.sh_link, (int) shdr.sh_flags);
 287
 288                 if (strcmp(shname, "license") == 0) {
 289                         processed_sec[i] = true;
 290                         memcpy(license, data->d_buf, data->d_size);
 291                 } else if (strcmp(shname, "version") == 0) {
 292                         processed_sec[i] = true;
 293                         if (data->d_size != sizeof(int)) {
 294                                 printf("invalid size of version section %zd\n",
 295                                        data->d_size);
 296                                 return 1;
 297                         }
 298                         memcpy(&kern_version, data->d_buf, sizeof(int));
 299                 } else if (strcmp(shname, "maps") == 0) {
 300                         processed_sec[i] = true;
 301                         if (load_maps(data->d_buf, data->d_size))
 302                                 return 1;
 303                 } else if (shdr.sh_type == SHT_SYMTAB) {
 304                         symbols = data;
 305                 }
 306         }
 307
 308         /* load programs that need map fixup (relocations) */
 309         for (i = 1; i < ehdr.e_shnum; i++) {
 310
 311                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 312                         continue;
 313                 if (shdr.sh_type == SHT_REL) {
 314                         struct bpf_insn *insns;
 315
 316                         if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
 317                                     &shdr_prog, &data_prog))
 318                                 continue;
 319
 320                         insns = (struct bpf_insn *) data_prog->d_buf;
 321
 322                         processed_sec[shdr.sh_info] = true;
 323                         processed_sec[i] = true;
 324
 325                         if (parse_relo_and_apply(data, symbols, &shdr, insns))
 326                                 continue;
 327
 328                         if (memcmp(shname_prog, "kprobe/", 7) == 0 ||
 329                             memcmp(shname_prog, "kretprobe/", 10) == 0 ||
 330                             memcmp(shname_prog, "tracepoint/", 11) == 0 ||
 331                             memcmp(shname_prog, "xdp", 3) == 0 ||
 332                             memcmp(shname_prog, "perf_event", 10) == 0 ||
 333                             memcmp(shname_prog, "socket", 6) == 0)
 334                                 load_and_attach(shname_prog, insns, data_prog->d_size);
 335                 }
 336         }
 337
 338         /* load programs that don't use maps */
 339         for (i = 1; i < ehdr.e_shnum; i++) {
 340
 341                 if (processed_sec[i])
 342                         continue;
 343
 344                 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
 345                         continue;
 346
 347                 if (memcmp(shname, "kprobe/", 7) == 0 ||
 348                     memcmp(shname, "kretprobe/", 10) == 0 ||
 349                     memcmp(shname, "tracepoint/", 11) == 0 ||
 350                     memcmp(shname, "xdp", 3) == 0 ||
 351                     memcmp(shname, "perf_event", 10) == 0 ||
 352                     memcmp(shname, "socket", 6) == 0)
 353                         load_and_attach(shname, data->d_buf, data->d_size);
 354         }
 355
 356         close(fd);
 357         return 0;
 358 }
 359
 360 void read_trace_pipe(void)
 361 {
 362         int trace_fd;
 363
 364         trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
 365         if (trace_fd < 0)
 366                 return;
 367
 368         while (1) {
 369                 static char buf[4096];
 370                 ssize_t sz;
 371
 372                 sz = read(trace_fd, buf, sizeof(buf));
 373                 if (sz > 0) {
 374                         buf[sz] = 0;
 375                         puts(buf);
 376                 }
 377         }
 378 }
 379
 380 #define MAX_SYMS 300000
 381 static struct ksym syms[MAX_SYMS];
 382 static int sym_cnt;
 383
 384 static int ksym_cmp(const void *p1, const void *p2)
 385 {
 386         return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
 387 }
 388
 389 int load_kallsyms(void)
 390 {
 391         FILE *f = fopen("/proc/kallsyms", "r");
 392         char func[256], buf[256];
 393         char symbol;
 394         void *addr;
 395         int i = 0;
 396
 397         if (!f)
 398                 return -ENOENT;
 399
 400         while (!feof(f)) {
 401                 if (!fgets(buf, sizeof(buf), f))
 402                         break;
 403                 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
 404                         break;
 405                 if (!addr)
 406                         continue;
 407                 syms[i].addr = (long) addr;
 408                 syms[i].name = strdup(func);
 409                 i++;
 410         }
 411         sym_cnt = i;
 412         qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
 413         return 0;
 414 }
 415
 416 struct ksym *ksym_search(long key)
 417 {
 418         int start = 0, end = sym_cnt;
 419         int result;
 420
 421         while (start < end) {
 422                 size_t mid = start + (end - start) / 2;
 423
 424                 result = key - syms[mid].addr;
 425                 if (result < 0)
 426                         end = mid;
 427                 else if (result > 0)
 428                         start = mid + 1;
 429                 else
 430                         return &syms[mid];
 431         }
 432
 433         if (start >= 1 && syms[start - 1].addr < key &&
 434             key < syms[start].addr)
 435                 /* valid ksym */
 436                 return &syms[start - 1];
 437
 438         /* out of range. return _stext */
 439         return &syms[0];
 440 }