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