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