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 <linux/types.h>
18 #include <sys/types.h>
19 #include <sys/socket.h>
20 #include <sys/syscall.h>
21 #include <sys/ioctl.h>
30 #define DEBUGFS "/sys/kernel/debug/tracing/"
32 static char license[128];
33 static int kern_version;
34 static bool processed_sec[128];
35 char bpf_log_buf[BPF_LOG_BUF_SIZE];
37 int prog_fd[MAX_PROGS];
38 int event_fd[MAX_PROGS];
40 int prog_array_fd = -1;
42 struct bpf_map_data map_data[MAX_MAPS];
43 int map_data_count = 0;
45 static int populate_prog_array(const char *event, int prog_fd)
47 int ind = atoi(event), err;
49 err = bpf_map_update_elem(prog_array_fd, &ind, &prog_fd, BPF_ANY);
51 printf("failed to store prog_fd in prog_array\n");
57 static int load_and_attach(const char *event, struct bpf_insn *prog, int size)
59 bool is_socket = strncmp(event, "socket", 6) == 0;
60 bool is_kprobe = strncmp(event, "kprobe/", 7) == 0;
61 bool is_kretprobe = strncmp(event, "kretprobe/", 10) == 0;
62 bool is_tracepoint = strncmp(event, "tracepoint/", 11) == 0;
63 bool is_xdp = strncmp(event, "xdp", 3) == 0;
64 bool is_perf_event = strncmp(event, "perf_event", 10) == 0;
65 bool is_cgroup_skb = strncmp(event, "cgroup/skb", 10) == 0;
66 bool is_cgroup_sk = strncmp(event, "cgroup/sock", 11) == 0;
67 bool is_sockops = strncmp(event, "sockops", 7) == 0;
68 size_t insns_cnt = size / sizeof(struct bpf_insn);
69 enum bpf_prog_type prog_type;
72 struct perf_event_attr attr = {};
74 attr.type = PERF_TYPE_TRACEPOINT;
75 attr.sample_type = PERF_SAMPLE_RAW;
76 attr.sample_period = 1;
77 attr.wakeup_events = 1;
80 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
81 } else if (is_kprobe || is_kretprobe) {
82 prog_type = BPF_PROG_TYPE_KPROBE;
83 } else if (is_tracepoint) {
84 prog_type = BPF_PROG_TYPE_TRACEPOINT;
86 prog_type = BPF_PROG_TYPE_XDP;
87 } else if (is_perf_event) {
88 prog_type = BPF_PROG_TYPE_PERF_EVENT;
89 } else if (is_cgroup_skb) {
90 prog_type = BPF_PROG_TYPE_CGROUP_SKB;
91 } else if (is_cgroup_sk) {
92 prog_type = BPF_PROG_TYPE_CGROUP_SOCK;
93 } else if (is_sockops) {
94 prog_type = BPF_PROG_TYPE_SOCK_OPS;
96 printf("Unknown event '%s'\n", event);
100 fd = bpf_load_program(prog_type, prog, insns_cnt, license, kern_version,
101 bpf_log_buf, BPF_LOG_BUF_SIZE);
103 printf("bpf_load_program() err=%d\n%s", errno, bpf_log_buf);
107 prog_fd[prog_cnt++] = fd;
109 if (is_xdp || is_perf_event || is_cgroup_skb || is_cgroup_sk)
112 if (is_socket || is_sockops) {
120 if (!isdigit(*event)) {
121 printf("invalid prog number\n");
124 return populate_prog_array(event, fd);
127 if (is_kprobe || is_kretprobe) {
134 printf("event name cannot be empty\n");
139 return populate_prog_array(event, fd);
141 snprintf(buf, sizeof(buf),
142 "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
143 is_kprobe ? 'p' : 'r', event, event);
146 printf("failed to create kprobe '%s' error '%s'\n",
147 event, strerror(errno));
151 strcpy(buf, DEBUGFS);
152 strcat(buf, "events/kprobes/");
155 } else if (is_tracepoint) {
159 printf("event name cannot be empty\n");
162 strcpy(buf, DEBUGFS);
163 strcat(buf, "events/");
168 efd = open(buf, O_RDONLY, 0);
170 printf("failed to open event %s\n", event);
174 err = read(efd, buf, sizeof(buf));
175 if (err < 0 || err >= sizeof(buf)) {
176 printf("read from '%s' failed '%s'\n", event, strerror(errno));
186 efd = sys_perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
188 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
191 event_fd[prog_cnt - 1] = efd;
192 ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
193 ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
198 static int load_maps(struct bpf_map_data *maps, int nr_maps,
199 fixup_map_cb fixup_map)
203 for (i = 0; i < nr_maps; i++) {
205 fixup_map(&maps[i], i);
206 /* Allow userspace to assign map FD prior to creation */
207 if (maps[i].fd != -1) {
208 map_fd[i] = maps[i].fd;
213 if (maps[i].def.type == BPF_MAP_TYPE_ARRAY_OF_MAPS ||
214 maps[i].def.type == BPF_MAP_TYPE_HASH_OF_MAPS) {
215 int inner_map_fd = map_fd[maps[i].def.inner_map_idx];
217 map_fd[i] = bpf_create_map_in_map(maps[i].def.type,
218 maps[i].def.key_size,
220 maps[i].def.max_entries,
221 maps[i].def.map_flags);
223 map_fd[i] = bpf_create_map(maps[i].def.type,
224 maps[i].def.key_size,
225 maps[i].def.value_size,
226 maps[i].def.max_entries,
227 maps[i].def.map_flags);
230 printf("failed to create a map: %d %s\n",
231 errno, strerror(errno));
234 maps[i].fd = map_fd[i];
236 if (maps[i].def.type == BPF_MAP_TYPE_PROG_ARRAY)
237 prog_array_fd = map_fd[i];
242 static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
243 GElf_Shdr *shdr, Elf_Data **data)
247 scn = elf_getscn(elf, i);
251 if (gelf_getshdr(scn, shdr) != shdr)
254 *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
255 if (!*shname || !shdr->sh_size)
258 *data = elf_getdata(scn, 0);
259 if (!*data || elf_getdata(scn, *data) != NULL)
265 static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
266 GElf_Shdr *shdr, struct bpf_insn *insn,
267 struct bpf_map_data *maps, int nr_maps)
271 nrels = shdr->sh_size / shdr->sh_entsize;
273 for (i = 0; i < nrels; i++) {
276 unsigned int insn_idx;
280 gelf_getrel(data, i, &rel);
282 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
284 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
286 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
287 printf("invalid relo for insn[%d].code 0x%x\n",
288 insn_idx, insn[insn_idx].code);
291 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
293 /* Match FD relocation against recorded map_data[] offset */
294 for (map_idx = 0; map_idx < nr_maps; map_idx++) {
295 if (maps[map_idx].elf_offset == sym.st_value) {
301 insn[insn_idx].imm = maps[map_idx].fd;
303 printf("invalid relo for insn[%d] no map_data match\n",
312 static int cmp_symbols(const void *l, const void *r)
314 const GElf_Sym *lsym = (const GElf_Sym *)l;
315 const GElf_Sym *rsym = (const GElf_Sym *)r;
317 if (lsym->st_value < rsym->st_value)
319 else if (lsym->st_value > rsym->st_value)
325 static int load_elf_maps_section(struct bpf_map_data *maps, int maps_shndx,
326 Elf *elf, Elf_Data *symbols, int strtabidx)
328 int map_sz_elf, map_sz_copy;
329 bool validate_zero = false;
341 /* Get data for maps section via elf index */
342 scn = elf_getscn(elf, maps_shndx);
344 data_maps = elf_getdata(scn, NULL);
345 if (!scn || !data_maps) {
346 printf("Failed to get Elf_Data from maps section %d\n",
351 /* For each map get corrosponding symbol table entry */
352 sym = calloc(MAX_MAPS+1, sizeof(GElf_Sym));
353 for (i = 0, nr_maps = 0; i < symbols->d_size / sizeof(GElf_Sym); i++) {
354 assert(nr_maps < MAX_MAPS+1);
355 if (!gelf_getsym(symbols, i, &sym[nr_maps]))
357 if (sym[nr_maps].st_shndx != maps_shndx)
359 /* Only increment iif maps section */
363 /* Align to map_fd[] order, via sort on offset in sym.st_value */
364 qsort(sym, nr_maps, sizeof(GElf_Sym), cmp_symbols);
366 /* Keeping compatible with ELF maps section changes
367 * ------------------------------------------------
368 * The program size of struct bpf_map_def is known by loader
369 * code, but struct stored in ELF file can be different.
371 * Unfortunately sym[i].st_size is zero. To calculate the
372 * struct size stored in the ELF file, assume all struct have
373 * the same size, and simply divide with number of map
376 map_sz_elf = data_maps->d_size / nr_maps;
377 map_sz_copy = sizeof(struct bpf_map_def);
378 if (map_sz_elf < map_sz_copy) {
380 * Backward compat, loading older ELF file with
381 * smaller struct, keeping remaining bytes zero.
383 map_sz_copy = map_sz_elf;
384 } else if (map_sz_elf > map_sz_copy) {
386 * Forward compat, loading newer ELF file with larger
387 * struct with unknown features. Assume zero means
388 * feature not used. Thus, validate rest of struct
391 validate_zero = true;
394 /* Memcpy relevant part of ELF maps data to loader maps */
395 for (i = 0; i < nr_maps; i++) {
396 unsigned char *addr, *end;
397 struct bpf_map_def *def;
398 const char *map_name;
401 map_name = elf_strptr(elf, strtabidx, sym[i].st_name);
402 maps[i].name = strdup(map_name);
404 printf("strdup(%s): %s(%d)\n", map_name,
405 strerror(errno), errno);
410 /* Symbol value is offset into ELF maps section data area */
411 offset = sym[i].st_value;
412 def = (struct bpf_map_def *)(data_maps->d_buf + offset);
413 maps[i].elf_offset = offset;
414 memset(&maps[i].def, 0, sizeof(struct bpf_map_def));
415 memcpy(&maps[i].def, def, map_sz_copy);
417 /* Verify no newer features were requested */
419 addr = (unsigned char*) def + map_sz_copy;
420 end = (unsigned char*) def + map_sz_elf;
421 for (; addr < end; addr++) {
434 static int do_load_bpf_file(const char *path, fixup_map_cb fixup_map)
436 int fd, i, ret, maps_shndx = -1, strtabidx = -1;
439 GElf_Shdr shdr, shdr_prog;
440 Elf_Data *data, *data_prog, *data_maps = NULL, *symbols = NULL;
441 char *shname, *shname_prog;
444 /* reset global variables */
446 memset(license, 0, sizeof(license));
447 memset(processed_sec, 0, sizeof(processed_sec));
449 if (elf_version(EV_CURRENT) == EV_NONE)
452 fd = open(path, O_RDONLY, 0);
456 elf = elf_begin(fd, ELF_C_READ, NULL);
461 if (gelf_getehdr(elf, &ehdr) != &ehdr)
464 /* clear all kprobes */
465 i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
467 /* scan over all elf sections to get license and map info */
468 for (i = 1; i < ehdr.e_shnum; i++) {
470 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
473 if (0) /* helpful for llvm debugging */
474 printf("section %d:%s data %p size %zd link %d flags %d\n",
475 i, shname, data->d_buf, data->d_size,
476 shdr.sh_link, (int) shdr.sh_flags);
478 if (strcmp(shname, "license") == 0) {
479 processed_sec[i] = true;
480 memcpy(license, data->d_buf, data->d_size);
481 } else if (strcmp(shname, "version") == 0) {
482 processed_sec[i] = true;
483 if (data->d_size != sizeof(int)) {
484 printf("invalid size of version section %zd\n",
488 memcpy(&kern_version, data->d_buf, sizeof(int));
489 } else if (strcmp(shname, "maps") == 0) {
494 for (j = 0; j < MAX_MAPS; j++)
496 } else if (shdr.sh_type == SHT_SYMTAB) {
497 strtabidx = shdr.sh_link;
505 printf("missing SHT_SYMTAB section\n");
510 nr_maps = load_elf_maps_section(map_data, maps_shndx,
511 elf, symbols, strtabidx);
513 printf("Error: Failed loading ELF maps (errno:%d):%s\n",
514 nr_maps, strerror(-nr_maps));
518 if (load_maps(map_data, nr_maps, fixup_map))
520 map_data_count = nr_maps;
522 processed_sec[maps_shndx] = true;
525 /* process all relo sections, and rewrite bpf insns for maps */
526 for (i = 1; i < ehdr.e_shnum; i++) {
527 if (processed_sec[i])
530 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
533 if (shdr.sh_type == SHT_REL) {
534 struct bpf_insn *insns;
536 /* locate prog sec that need map fixup (relocations) */
537 if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
538 &shdr_prog, &data_prog))
541 if (shdr_prog.sh_type != SHT_PROGBITS ||
542 !(shdr_prog.sh_flags & SHF_EXECINSTR))
545 insns = (struct bpf_insn *) data_prog->d_buf;
546 processed_sec[i] = true; /* relo section */
548 if (parse_relo_and_apply(data, symbols, &shdr, insns,
555 for (i = 1; i < ehdr.e_shnum; i++) {
557 if (processed_sec[i])
560 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
563 if (memcmp(shname, "kprobe/", 7) == 0 ||
564 memcmp(shname, "kretprobe/", 10) == 0 ||
565 memcmp(shname, "tracepoint/", 11) == 0 ||
566 memcmp(shname, "xdp", 3) == 0 ||
567 memcmp(shname, "perf_event", 10) == 0 ||
568 memcmp(shname, "socket", 6) == 0 ||
569 memcmp(shname, "cgroup/", 7) == 0 ||
570 memcmp(shname, "sockops", 7) == 0) {
571 ret = load_and_attach(shname, data->d_buf,
584 int load_bpf_file(char *path)
586 return do_load_bpf_file(path, NULL);
589 int load_bpf_file_fixup_map(const char *path, fixup_map_cb fixup_map)
591 return do_load_bpf_file(path, fixup_map);
594 void read_trace_pipe(void)
598 trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
603 static char buf[4096];
606 sz = read(trace_fd, buf, sizeof(buf));
614 #define MAX_SYMS 300000
615 static struct ksym syms[MAX_SYMS];
618 static int ksym_cmp(const void *p1, const void *p2)
620 return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
623 int load_kallsyms(void)
625 FILE *f = fopen("/proc/kallsyms", "r");
626 char func[256], buf[256];
635 if (!fgets(buf, sizeof(buf), f))
637 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
641 syms[i].addr = (long) addr;
642 syms[i].name = strdup(func);
646 qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
650 struct ksym *ksym_search(long key)
652 int start = 0, end = sym_cnt;
655 while (start < end) {
656 size_t mid = start + (end - start) / 2;
658 result = key - syms[mid].addr;
667 if (start >= 1 && syms[start - 1].addr < key &&
668 key < syms[start].addr)
670 return &syms[start - 1];
672 /* out of range. return _stext */
676 int set_link_xdp_fd(int ifindex, int fd, __u32 flags)
678 struct sockaddr_nl sa;
679 int sock, seq = 0, len, ret = -1;
681 struct nlattr *nla, *nla_xdp;
684 struct ifinfomsg ifinfo;
688 struct nlmsgerr *err;
690 memset(&sa, 0, sizeof(sa));
691 sa.nl_family = AF_NETLINK;
693 sock = socket(AF_NETLINK, SOCK_RAW, NETLINK_ROUTE);
695 printf("open netlink socket: %s\n", strerror(errno));
699 if (bind(sock, (struct sockaddr *)&sa, sizeof(sa)) < 0) {
700 printf("bind to netlink: %s\n", strerror(errno));
704 memset(&req, 0, sizeof(req));
705 req.nh.nlmsg_len = NLMSG_LENGTH(sizeof(struct ifinfomsg));
706 req.nh.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK;
707 req.nh.nlmsg_type = RTM_SETLINK;
708 req.nh.nlmsg_pid = 0;
709 req.nh.nlmsg_seq = ++seq;
710 req.ifinfo.ifi_family = AF_UNSPEC;
711 req.ifinfo.ifi_index = ifindex;
713 /* started nested attribute for XDP */
714 nla = (struct nlattr *)(((char *)&req)
715 + NLMSG_ALIGN(req.nh.nlmsg_len));
716 nla->nla_type = NLA_F_NESTED | 43/*IFLA_XDP*/;
717 nla->nla_len = NLA_HDRLEN;
720 nla_xdp = (struct nlattr *)((char *)nla + nla->nla_len);
721 nla_xdp->nla_type = 1/*IFLA_XDP_FD*/;
722 nla_xdp->nla_len = NLA_HDRLEN + sizeof(int);
723 memcpy((char *)nla_xdp + NLA_HDRLEN, &fd, sizeof(fd));
724 nla->nla_len += nla_xdp->nla_len;
726 /* if user passed in any flags, add those too */
728 nla_xdp = (struct nlattr *)((char *)nla + nla->nla_len);
729 nla_xdp->nla_type = 3/*IFLA_XDP_FLAGS*/;
730 nla_xdp->nla_len = NLA_HDRLEN + sizeof(flags);
731 memcpy((char *)nla_xdp + NLA_HDRLEN, &flags, sizeof(flags));
732 nla->nla_len += nla_xdp->nla_len;
735 req.nh.nlmsg_len += NLA_ALIGN(nla->nla_len);
737 if (send(sock, &req, req.nh.nlmsg_len, 0) < 0) {
738 printf("send to netlink: %s\n", strerror(errno));
742 len = recv(sock, buf, sizeof(buf), 0);
744 printf("recv from netlink: %s\n", strerror(errno));
748 for (nh = (struct nlmsghdr *)buf; NLMSG_OK(nh, len);
749 nh = NLMSG_NEXT(nh, len)) {
750 if (nh->nlmsg_pid != getpid()) {
751 printf("Wrong pid %d, expected %d\n",
752 nh->nlmsg_pid, getpid());
755 if (nh->nlmsg_seq != seq) {
756 printf("Wrong seq %d, expected %d\n",
760 switch (nh->nlmsg_type) {
762 err = (struct nlmsgerr *)NLMSG_DATA(nh);
765 printf("nlmsg error %s\n", strerror(-err->error));