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 size_t insns_cnt = size / sizeof(struct bpf_insn);
68 enum bpf_prog_type prog_type;
71 struct perf_event_attr attr = {};
73 attr.type = PERF_TYPE_TRACEPOINT;
74 attr.sample_type = PERF_SAMPLE_RAW;
75 attr.sample_period = 1;
76 attr.wakeup_events = 1;
79 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
80 } else if (is_kprobe || is_kretprobe) {
81 prog_type = BPF_PROG_TYPE_KPROBE;
82 } else if (is_tracepoint) {
83 prog_type = BPF_PROG_TYPE_TRACEPOINT;
85 prog_type = BPF_PROG_TYPE_XDP;
86 } else if (is_perf_event) {
87 prog_type = BPF_PROG_TYPE_PERF_EVENT;
88 } else if (is_cgroup_skb) {
89 prog_type = BPF_PROG_TYPE_CGROUP_SKB;
90 } else if (is_cgroup_sk) {
91 prog_type = BPF_PROG_TYPE_CGROUP_SOCK;
93 printf("Unknown event '%s'\n", event);
97 fd = bpf_load_program(prog_type, prog, insns_cnt, license, kern_version,
98 bpf_log_buf, BPF_LOG_BUF_SIZE);
100 printf("bpf_load_program() err=%d\n%s", errno, bpf_log_buf);
104 prog_fd[prog_cnt++] = fd;
106 if (is_xdp || is_perf_event || is_cgroup_skb || is_cgroup_sk)
114 if (!isdigit(*event)) {
115 printf("invalid prog number\n");
118 return populate_prog_array(event, fd);
121 if (is_kprobe || is_kretprobe) {
128 printf("event name cannot be empty\n");
133 return populate_prog_array(event, fd);
135 snprintf(buf, sizeof(buf),
136 "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
137 is_kprobe ? 'p' : 'r', event, event);
140 printf("failed to create kprobe '%s' error '%s'\n",
141 event, strerror(errno));
145 strcpy(buf, DEBUGFS);
146 strcat(buf, "events/kprobes/");
149 } else if (is_tracepoint) {
153 printf("event name cannot be empty\n");
156 strcpy(buf, DEBUGFS);
157 strcat(buf, "events/");
162 efd = open(buf, O_RDONLY, 0);
164 printf("failed to open event %s\n", event);
168 err = read(efd, buf, sizeof(buf));
169 if (err < 0 || err >= sizeof(buf)) {
170 printf("read from '%s' failed '%s'\n", event, strerror(errno));
180 efd = sys_perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
182 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
185 event_fd[prog_cnt - 1] = efd;
186 ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
187 ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
192 static int load_maps(struct bpf_map_data *maps, int nr_maps,
193 fixup_map_cb fixup_map)
197 for (i = 0; i < nr_maps; i++) {
199 fixup_map(&maps[i], i);
200 /* Allow userspace to assign map FD prior to creation */
201 if (maps[i].fd != -1) {
202 map_fd[i] = maps[i].fd;
207 if (maps[i].def.type == BPF_MAP_TYPE_ARRAY_OF_MAPS ||
208 maps[i].def.type == BPF_MAP_TYPE_HASH_OF_MAPS) {
209 int inner_map_fd = map_fd[maps[i].def.inner_map_idx];
211 map_fd[i] = bpf_create_map_in_map(maps[i].def.type,
212 maps[i].def.key_size,
214 maps[i].def.max_entries,
215 maps[i].def.map_flags);
217 map_fd[i] = bpf_create_map(maps[i].def.type,
218 maps[i].def.key_size,
219 maps[i].def.value_size,
220 maps[i].def.max_entries,
221 maps[i].def.map_flags);
224 printf("failed to create a map: %d %s\n",
225 errno, strerror(errno));
228 maps[i].fd = map_fd[i];
230 if (maps[i].def.type == BPF_MAP_TYPE_PROG_ARRAY)
231 prog_array_fd = map_fd[i];
236 static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
237 GElf_Shdr *shdr, Elf_Data **data)
241 scn = elf_getscn(elf, i);
245 if (gelf_getshdr(scn, shdr) != shdr)
248 *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
249 if (!*shname || !shdr->sh_size)
252 *data = elf_getdata(scn, 0);
253 if (!*data || elf_getdata(scn, *data) != NULL)
259 static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
260 GElf_Shdr *shdr, struct bpf_insn *insn,
261 struct bpf_map_data *maps, int nr_maps)
265 nrels = shdr->sh_size / shdr->sh_entsize;
267 for (i = 0; i < nrels; i++) {
270 unsigned int insn_idx;
274 gelf_getrel(data, i, &rel);
276 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
278 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
280 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
281 printf("invalid relo for insn[%d].code 0x%x\n",
282 insn_idx, insn[insn_idx].code);
285 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
287 /* Match FD relocation against recorded map_data[] offset */
288 for (map_idx = 0; map_idx < nr_maps; map_idx++) {
289 if (maps[map_idx].elf_offset == sym.st_value) {
295 insn[insn_idx].imm = maps[map_idx].fd;
297 printf("invalid relo for insn[%d] no map_data match\n",
306 static int cmp_symbols(const void *l, const void *r)
308 const GElf_Sym *lsym = (const GElf_Sym *)l;
309 const GElf_Sym *rsym = (const GElf_Sym *)r;
311 if (lsym->st_value < rsym->st_value)
313 else if (lsym->st_value > rsym->st_value)
319 static int load_elf_maps_section(struct bpf_map_data *maps, int maps_shndx,
320 Elf *elf, Elf_Data *symbols, int strtabidx)
322 int map_sz_elf, map_sz_copy;
323 bool validate_zero = false;
335 /* Get data for maps section via elf index */
336 scn = elf_getscn(elf, maps_shndx);
338 data_maps = elf_getdata(scn, NULL);
339 if (!scn || !data_maps) {
340 printf("Failed to get Elf_Data from maps section %d\n",
345 /* For each map get corrosponding symbol table entry */
346 sym = calloc(MAX_MAPS+1, sizeof(GElf_Sym));
347 for (i = 0, nr_maps = 0; i < symbols->d_size / sizeof(GElf_Sym); i++) {
348 assert(nr_maps < MAX_MAPS+1);
349 if (!gelf_getsym(symbols, i, &sym[nr_maps]))
351 if (sym[nr_maps].st_shndx != maps_shndx)
353 /* Only increment iif maps section */
357 /* Align to map_fd[] order, via sort on offset in sym.st_value */
358 qsort(sym, nr_maps, sizeof(GElf_Sym), cmp_symbols);
360 /* Keeping compatible with ELF maps section changes
361 * ------------------------------------------------
362 * The program size of struct bpf_map_def is known by loader
363 * code, but struct stored in ELF file can be different.
365 * Unfortunately sym[i].st_size is zero. To calculate the
366 * struct size stored in the ELF file, assume all struct have
367 * the same size, and simply divide with number of map
370 map_sz_elf = data_maps->d_size / nr_maps;
371 map_sz_copy = sizeof(struct bpf_map_def);
372 if (map_sz_elf < map_sz_copy) {
374 * Backward compat, loading older ELF file with
375 * smaller struct, keeping remaining bytes zero.
377 map_sz_copy = map_sz_elf;
378 } else if (map_sz_elf > map_sz_copy) {
380 * Forward compat, loading newer ELF file with larger
381 * struct with unknown features. Assume zero means
382 * feature not used. Thus, validate rest of struct
385 validate_zero = true;
388 /* Memcpy relevant part of ELF maps data to loader maps */
389 for (i = 0; i < nr_maps; i++) {
390 unsigned char *addr, *end;
391 struct bpf_map_def *def;
392 const char *map_name;
395 map_name = elf_strptr(elf, strtabidx, sym[i].st_name);
396 maps[i].name = strdup(map_name);
398 printf("strdup(%s): %s(%d)\n", map_name,
399 strerror(errno), errno);
404 /* Symbol value is offset into ELF maps section data area */
405 offset = sym[i].st_value;
406 def = (struct bpf_map_def *)(data_maps->d_buf + offset);
407 maps[i].elf_offset = offset;
408 memset(&maps[i].def, 0, sizeof(struct bpf_map_def));
409 memcpy(&maps[i].def, def, map_sz_copy);
411 /* Verify no newer features were requested */
413 addr = (unsigned char*) def + map_sz_copy;
414 end = (unsigned char*) def + map_sz_elf;
415 for (; addr < end; addr++) {
428 static int do_load_bpf_file(const char *path, fixup_map_cb fixup_map)
430 int fd, i, ret, maps_shndx = -1, strtabidx = -1;
433 GElf_Shdr shdr, shdr_prog;
434 Elf_Data *data, *data_prog, *data_maps = NULL, *symbols = NULL;
435 char *shname, *shname_prog;
438 /* reset global variables */
440 memset(license, 0, sizeof(license));
441 memset(processed_sec, 0, sizeof(processed_sec));
443 if (elf_version(EV_CURRENT) == EV_NONE)
446 fd = open(path, O_RDONLY, 0);
450 elf = elf_begin(fd, ELF_C_READ, NULL);
455 if (gelf_getehdr(elf, &ehdr) != &ehdr)
458 /* clear all kprobes */
459 i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
461 /* scan over all elf sections to get license and map info */
462 for (i = 1; i < ehdr.e_shnum; i++) {
464 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
467 if (0) /* helpful for llvm debugging */
468 printf("section %d:%s data %p size %zd link %d flags %d\n",
469 i, shname, data->d_buf, data->d_size,
470 shdr.sh_link, (int) shdr.sh_flags);
472 if (strcmp(shname, "license") == 0) {
473 processed_sec[i] = true;
474 memcpy(license, data->d_buf, data->d_size);
475 } else if (strcmp(shname, "version") == 0) {
476 processed_sec[i] = true;
477 if (data->d_size != sizeof(int)) {
478 printf("invalid size of version section %zd\n",
482 memcpy(&kern_version, data->d_buf, sizeof(int));
483 } else if (strcmp(shname, "maps") == 0) {
488 for (j = 0; j < MAX_MAPS; j++)
490 } else if (shdr.sh_type == SHT_SYMTAB) {
491 strtabidx = shdr.sh_link;
499 printf("missing SHT_SYMTAB section\n");
504 nr_maps = load_elf_maps_section(map_data, maps_shndx,
505 elf, symbols, strtabidx);
507 printf("Error: Failed loading ELF maps (errno:%d):%s\n",
508 nr_maps, strerror(-nr_maps));
512 if (load_maps(map_data, nr_maps, fixup_map))
514 map_data_count = nr_maps;
516 processed_sec[maps_shndx] = true;
519 /* load programs that need map fixup (relocations) */
520 for (i = 1; i < ehdr.e_shnum; i++) {
521 if (processed_sec[i])
524 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
526 if (shdr.sh_type == SHT_REL) {
527 struct bpf_insn *insns;
529 if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
530 &shdr_prog, &data_prog))
533 if (shdr_prog.sh_type != SHT_PROGBITS ||
534 !(shdr_prog.sh_flags & SHF_EXECINSTR))
537 insns = (struct bpf_insn *) data_prog->d_buf;
539 processed_sec[shdr.sh_info] = true;
540 processed_sec[i] = true;
542 if (parse_relo_and_apply(data, symbols, &shdr, insns,
546 if (memcmp(shname_prog, "kprobe/", 7) == 0 ||
547 memcmp(shname_prog, "kretprobe/", 10) == 0 ||
548 memcmp(shname_prog, "tracepoint/", 11) == 0 ||
549 memcmp(shname_prog, "xdp", 3) == 0 ||
550 memcmp(shname_prog, "perf_event", 10) == 0 ||
551 memcmp(shname_prog, "socket", 6) == 0 ||
552 memcmp(shname_prog, "cgroup/", 7) == 0)
553 load_and_attach(shname_prog, insns, data_prog->d_size);
557 /* load programs that don't use maps */
558 for (i = 1; i < ehdr.e_shnum; i++) {
560 if (processed_sec[i])
563 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
566 if (memcmp(shname, "kprobe/", 7) == 0 ||
567 memcmp(shname, "kretprobe/", 10) == 0 ||
568 memcmp(shname, "tracepoint/", 11) == 0 ||
569 memcmp(shname, "xdp", 3) == 0 ||
570 memcmp(shname, "perf_event", 10) == 0 ||
571 memcmp(shname, "socket", 6) == 0 ||
572 memcmp(shname, "cgroup/", 7) == 0)
573 load_and_attach(shname, data->d_buf, data->d_size);
582 int load_bpf_file(char *path)
584 return do_load_bpf_file(path, NULL);
587 int load_bpf_file_fixup_map(const char *path, fixup_map_cb fixup_map)
589 return do_load_bpf_file(path, fixup_map);
592 void read_trace_pipe(void)
596 trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
601 static char buf[4096];
604 sz = read(trace_fd, buf, sizeof(buf));
612 #define MAX_SYMS 300000
613 static struct ksym syms[MAX_SYMS];
616 static int ksym_cmp(const void *p1, const void *p2)
618 return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
621 int load_kallsyms(void)
623 FILE *f = fopen("/proc/kallsyms", "r");
624 char func[256], buf[256];
633 if (!fgets(buf, sizeof(buf), f))
635 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
639 syms[i].addr = (long) addr;
640 syms[i].name = strdup(func);
644 qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
648 struct ksym *ksym_search(long key)
650 int start = 0, end = sym_cnt;
653 while (start < end) {
654 size_t mid = start + (end - start) / 2;
656 result = key - syms[mid].addr;
665 if (start >= 1 && syms[start - 1].addr < key &&
666 key < syms[start].addr)
668 return &syms[start - 1];
670 /* out of range. return _stext */
674 int set_link_xdp_fd(int ifindex, int fd, __u32 flags)
676 struct sockaddr_nl sa;
677 int sock, seq = 0, len, ret = -1;
679 struct nlattr *nla, *nla_xdp;
682 struct ifinfomsg ifinfo;
686 struct nlmsgerr *err;
688 memset(&sa, 0, sizeof(sa));
689 sa.nl_family = AF_NETLINK;
691 sock = socket(AF_NETLINK, SOCK_RAW, NETLINK_ROUTE);
693 printf("open netlink socket: %s\n", strerror(errno));
697 if (bind(sock, (struct sockaddr *)&sa, sizeof(sa)) < 0) {
698 printf("bind to netlink: %s\n", strerror(errno));
702 memset(&req, 0, sizeof(req));
703 req.nh.nlmsg_len = NLMSG_LENGTH(sizeof(struct ifinfomsg));
704 req.nh.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK;
705 req.nh.nlmsg_type = RTM_SETLINK;
706 req.nh.nlmsg_pid = 0;
707 req.nh.nlmsg_seq = ++seq;
708 req.ifinfo.ifi_family = AF_UNSPEC;
709 req.ifinfo.ifi_index = ifindex;
711 /* started nested attribute for XDP */
712 nla = (struct nlattr *)(((char *)&req)
713 + NLMSG_ALIGN(req.nh.nlmsg_len));
714 nla->nla_type = NLA_F_NESTED | 43/*IFLA_XDP*/;
715 nla->nla_len = NLA_HDRLEN;
718 nla_xdp = (struct nlattr *)((char *)nla + nla->nla_len);
719 nla_xdp->nla_type = 1/*IFLA_XDP_FD*/;
720 nla_xdp->nla_len = NLA_HDRLEN + sizeof(int);
721 memcpy((char *)nla_xdp + NLA_HDRLEN, &fd, sizeof(fd));
722 nla->nla_len += nla_xdp->nla_len;
724 /* if user passed in any flags, add those too */
726 nla_xdp = (struct nlattr *)((char *)nla + nla->nla_len);
727 nla_xdp->nla_type = 3/*IFLA_XDP_FLAGS*/;
728 nla_xdp->nla_len = NLA_HDRLEN + sizeof(flags);
729 memcpy((char *)nla_xdp + NLA_HDRLEN, &flags, sizeof(flags));
730 nla->nla_len += nla_xdp->nla_len;
733 req.nh.nlmsg_len += NLA_ALIGN(nla->nla_len);
735 if (send(sock, &req, req.nh.nlmsg_len, 0) < 0) {
736 printf("send to netlink: %s\n", strerror(errno));
740 len = recv(sock, buf, sizeof(buf), 0);
742 printf("recv from netlink: %s\n", strerror(errno));
746 for (nh = (struct nlmsghdr *)buf; NLMSG_OK(nh, len);
747 nh = NLMSG_NEXT(nh, len)) {
748 if (nh->nlmsg_pid != getpid()) {
749 printf("Wrong pid %d, expected %d\n",
750 nh->nlmsg_pid, getpid());
753 if (nh->nlmsg_seq != seq) {
754 printf("Wrong seq %d, expected %d\n",
758 switch (nh->nlmsg_type) {
760 err = (struct nlmsgerr *)NLMSG_DATA(nh);
763 printf("nlmsg error %s\n", strerror(-err->error));