Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
tjh.dev
kernel
os
linux
1// SPDX-License-Identifier: GPL-2.0
2#include <stdio.h>
3#include <sys/types.h>
4#include <sys/stat.h>
5#include <fcntl.h>
6#include <libelf.h>
7#include <gelf.h>
8#include <errno.h>
9#include <unistd.h>
10#include <string.h>
11#include <stdbool.h>
12#include <stdlib.h>
13#include <linux/bpf.h>
14#include <linux/filter.h>
15#include <linux/perf_event.h>
16#include <linux/netlink.h>
17#include <linux/rtnetlink.h>
18#include <linux/types.h>
19#include <sys/types.h>
20#include <sys/socket.h>
21#include <sys/syscall.h>
22#include <sys/ioctl.h>
23#include <sys/mman.h>
24#include <poll.h>
25#include <ctype.h>
26#include <assert.h>
27#include "libbpf.h"
28#include "bpf_load.h"
29#include "perf-sys.h"
30
31#define DEBUGFS "/sys/kernel/debug/tracing/"
32
33static char license[128];
34static int kern_version;
35static bool processed_sec[128];
36char bpf_log_buf[BPF_LOG_BUF_SIZE];
37int map_fd[MAX_MAPS];
38int prog_fd[MAX_PROGS];
39int event_fd[MAX_PROGS];
40int prog_cnt;
41int prog_array_fd = -1;
42
43struct bpf_map_data map_data[MAX_MAPS];
44int map_data_count = 0;
45
46static int populate_prog_array(const char *event, int prog_fd)
47{
48 int ind = atoi(event), err;
49
50 err = bpf_map_update_elem(prog_array_fd, &ind, &prog_fd, BPF_ANY);
51 if (err < 0) {
52 printf("failed to store prog_fd in prog_array\n");
53 return -1;
54 }
55 return 0;
56}
57
58static int load_and_attach(const char *event, struct bpf_insn *prog, int size)
59{
60 bool is_socket = strncmp(event, "socket", 6) == 0;
61 bool is_kprobe = strncmp(event, "kprobe/", 7) == 0;
62 bool is_kretprobe = strncmp(event, "kretprobe/", 10) == 0;
63 bool is_tracepoint = strncmp(event, "tracepoint/", 11) == 0;
64 bool is_xdp = strncmp(event, "xdp", 3) == 0;
65 bool is_perf_event = strncmp(event, "perf_event", 10) == 0;
66 bool is_cgroup_skb = strncmp(event, "cgroup/skb", 10) == 0;
67 bool is_cgroup_sk = strncmp(event, "cgroup/sock", 11) == 0;
68 bool is_sockops = strncmp(event, "sockops", 7) == 0;
69 bool is_sk_skb = strncmp(event, "sk_skb", 6) == 0;
70 size_t insns_cnt = size / sizeof(struct bpf_insn);
71 enum bpf_prog_type prog_type;
72 char buf[256];
73 int fd, efd, err, id;
74 struct perf_event_attr attr = {};
75
76 attr.type = PERF_TYPE_TRACEPOINT;
77 attr.sample_type = PERF_SAMPLE_RAW;
78 attr.sample_period = 1;
79 attr.wakeup_events = 1;
80
81 if (is_socket) {
82 prog_type = BPF_PROG_TYPE_SOCKET_FILTER;
83 } else if (is_kprobe || is_kretprobe) {
84 prog_type = BPF_PROG_TYPE_KPROBE;
85 } else if (is_tracepoint) {
86 prog_type = BPF_PROG_TYPE_TRACEPOINT;
87 } else if (is_xdp) {
88 prog_type = BPF_PROG_TYPE_XDP;
89 } else if (is_perf_event) {
90 prog_type = BPF_PROG_TYPE_PERF_EVENT;
91 } else if (is_cgroup_skb) {
92 prog_type = BPF_PROG_TYPE_CGROUP_SKB;
93 } else if (is_cgroup_sk) {
94 prog_type = BPF_PROG_TYPE_CGROUP_SOCK;
95 } else if (is_sockops) {
96 prog_type = BPF_PROG_TYPE_SOCK_OPS;
97 } else if (is_sk_skb) {
98 prog_type = BPF_PROG_TYPE_SK_SKB;
99 } else {
100 printf("Unknown event '%s'\n", event);
101 return -1;
102 }
103
104 fd = bpf_load_program(prog_type, prog, insns_cnt, license, kern_version,
105 bpf_log_buf, BPF_LOG_BUF_SIZE);
106 if (fd < 0) {
107 printf("bpf_load_program() err=%d\n%s", errno, bpf_log_buf);
108 return -1;
109 }
110
111 prog_fd[prog_cnt++] = fd;
112
113 if (is_xdp || is_perf_event || is_cgroup_skb || is_cgroup_sk)
114 return 0;
115
116 if (is_socket || is_sockops || is_sk_skb) {
117 if (is_socket)
118 event += 6;
119 else
120 event += 7;
121 if (*event != '/')
122 return 0;
123 event++;
124 if (!isdigit(*event)) {
125 printf("invalid prog number\n");
126 return -1;
127 }
128 return populate_prog_array(event, fd);
129 }
130
131 if (is_kprobe || is_kretprobe) {
132 if (is_kprobe)
133 event += 7;
134 else
135 event += 10;
136
137 if (*event == 0) {
138 printf("event name cannot be empty\n");
139 return -1;
140 }
141
142 if (isdigit(*event))
143 return populate_prog_array(event, fd);
144
145 snprintf(buf, sizeof(buf),
146 "echo '%c:%s %s' >> /sys/kernel/debug/tracing/kprobe_events",
147 is_kprobe ? 'p' : 'r', event, event);
148 err = system(buf);
149 if (err < 0) {
150 printf("failed to create kprobe '%s' error '%s'\n",
151 event, strerror(errno));
152 return -1;
153 }
154
155 strcpy(buf, DEBUGFS);
156 strcat(buf, "events/kprobes/");
157 strcat(buf, event);
158 strcat(buf, "/id");
159 } else if (is_tracepoint) {
160 event += 11;
161
162 if (*event == 0) {
163 printf("event name cannot be empty\n");
164 return -1;
165 }
166 strcpy(buf, DEBUGFS);
167 strcat(buf, "events/");
168 strcat(buf, event);
169 strcat(buf, "/id");
170 }
171
172 efd = open(buf, O_RDONLY, 0);
173 if (efd < 0) {
174 printf("failed to open event %s\n", event);
175 return -1;
176 }
177
178 err = read(efd, buf, sizeof(buf));
179 if (err < 0 || err >= sizeof(buf)) {
180 printf("read from '%s' failed '%s'\n", event, strerror(errno));
181 return -1;
182 }
183
184 close(efd);
185
186 buf[err] = 0;
187 id = atoi(buf);
188 attr.config = id;
189
190 efd = sys_perf_event_open(&attr, -1/*pid*/, 0/*cpu*/, -1/*group_fd*/, 0);
191 if (efd < 0) {
192 printf("event %d fd %d err %s\n", id, efd, strerror(errno));
193 return -1;
194 }
195 event_fd[prog_cnt - 1] = efd;
196 err = ioctl(efd, PERF_EVENT_IOC_ENABLE, 0);
197 if (err < 0) {
198 printf("ioctl PERF_EVENT_IOC_ENABLE failed err %s\n",
199 strerror(errno));
200 return -1;
201 }
202 err = ioctl(efd, PERF_EVENT_IOC_SET_BPF, fd);
203 if (err < 0) {
204 printf("ioctl PERF_EVENT_IOC_SET_BPF failed err %s\n",
205 strerror(errno));
206 return -1;
207 }
208
209 return 0;
210}
211
212static int load_maps(struct bpf_map_data *maps, int nr_maps,
213 fixup_map_cb fixup_map)
214{
215 int i, numa_node;
216
217 for (i = 0; i < nr_maps; i++) {
218 if (fixup_map) {
219 fixup_map(&maps[i], i);
220 /* Allow userspace to assign map FD prior to creation */
221 if (maps[i].fd != -1) {
222 map_fd[i] = maps[i].fd;
223 continue;
224 }
225 }
226
227 numa_node = maps[i].def.map_flags & BPF_F_NUMA_NODE ?
228 maps[i].def.numa_node : -1;
229
230 if (maps[i].def.type == BPF_MAP_TYPE_ARRAY_OF_MAPS ||
231 maps[i].def.type == BPF_MAP_TYPE_HASH_OF_MAPS) {
232 int inner_map_fd = map_fd[maps[i].def.inner_map_idx];
233
234 map_fd[i] = bpf_create_map_in_map_node(maps[i].def.type,
235 maps[i].name,
236 maps[i].def.key_size,
237 inner_map_fd,
238 maps[i].def.max_entries,
239 maps[i].def.map_flags,
240 numa_node);
241 } else {
242 map_fd[i] = bpf_create_map_node(maps[i].def.type,
243 maps[i].name,
244 maps[i].def.key_size,
245 maps[i].def.value_size,
246 maps[i].def.max_entries,
247 maps[i].def.map_flags,
248 numa_node);
249 }
250 if (map_fd[i] < 0) {
251 printf("failed to create a map: %d %s\n",
252 errno, strerror(errno));
253 return 1;
254 }
255 maps[i].fd = map_fd[i];
256
257 if (maps[i].def.type == BPF_MAP_TYPE_PROG_ARRAY)
258 prog_array_fd = map_fd[i];
259 }
260 return 0;
261}
262
263static int get_sec(Elf *elf, int i, GElf_Ehdr *ehdr, char **shname,
264 GElf_Shdr *shdr, Elf_Data **data)
265{
266 Elf_Scn *scn;
267
268 scn = elf_getscn(elf, i);
269 if (!scn)
270 return 1;
271
272 if (gelf_getshdr(scn, shdr) != shdr)
273 return 2;
274
275 *shname = elf_strptr(elf, ehdr->e_shstrndx, shdr->sh_name);
276 if (!*shname || !shdr->sh_size)
277 return 3;
278
279 *data = elf_getdata(scn, 0);
280 if (!*data || elf_getdata(scn, *data) != NULL)
281 return 4;
282
283 return 0;
284}
285
286static int parse_relo_and_apply(Elf_Data *data, Elf_Data *symbols,
287 GElf_Shdr *shdr, struct bpf_insn *insn,
288 struct bpf_map_data *maps, int nr_maps)
289{
290 int i, nrels;
291
292 nrels = shdr->sh_size / shdr->sh_entsize;
293
294 for (i = 0; i < nrels; i++) {
295 GElf_Sym sym;
296 GElf_Rel rel;
297 unsigned int insn_idx;
298 bool match = false;
299 int j, map_idx;
300
301 gelf_getrel(data, i, &rel);
302
303 insn_idx = rel.r_offset / sizeof(struct bpf_insn);
304
305 gelf_getsym(symbols, GELF_R_SYM(rel.r_info), &sym);
306
307 if (insn[insn_idx].code != (BPF_LD | BPF_IMM | BPF_DW)) {
308 printf("invalid relo for insn[%d].code 0x%x\n",
309 insn_idx, insn[insn_idx].code);
310 return 1;
311 }
312 insn[insn_idx].src_reg = BPF_PSEUDO_MAP_FD;
313
314 /* Match FD relocation against recorded map_data[] offset */
315 for (map_idx = 0; map_idx < nr_maps; map_idx++) {
316 if (maps[map_idx].elf_offset == sym.st_value) {
317 match = true;
318 break;
319 }
320 }
321 if (match) {
322 insn[insn_idx].imm = maps[map_idx].fd;
323 } else {
324 printf("invalid relo for insn[%d] no map_data match\n",
325 insn_idx);
326 return 1;
327 }
328 }
329
330 return 0;
331}
332
333static int cmp_symbols(const void *l, const void *r)
334{
335 const GElf_Sym *lsym = (const GElf_Sym *)l;
336 const GElf_Sym *rsym = (const GElf_Sym *)r;
337
338 if (lsym->st_value < rsym->st_value)
339 return -1;
340 else if (lsym->st_value > rsym->st_value)
341 return 1;
342 else
343 return 0;
344}
345
346static int load_elf_maps_section(struct bpf_map_data *maps, int maps_shndx,
347 Elf *elf, Elf_Data *symbols, int strtabidx)
348{
349 int map_sz_elf, map_sz_copy;
350 bool validate_zero = false;
351 Elf_Data *data_maps;
352 int i, nr_maps;
353 GElf_Sym *sym;
354 Elf_Scn *scn;
355 int copy_sz;
356
357 if (maps_shndx < 0)
358 return -EINVAL;
359 if (!symbols)
360 return -EINVAL;
361
362 /* Get data for maps section via elf index */
363 scn = elf_getscn(elf, maps_shndx);
364 if (scn)
365 data_maps = elf_getdata(scn, NULL);
366 if (!scn || !data_maps) {
367 printf("Failed to get Elf_Data from maps section %d\n",
368 maps_shndx);
369 return -EINVAL;
370 }
371
372 /* For each map get corrosponding symbol table entry */
373 sym = calloc(MAX_MAPS+1, sizeof(GElf_Sym));
374 for (i = 0, nr_maps = 0; i < symbols->d_size / sizeof(GElf_Sym); i++) {
375 assert(nr_maps < MAX_MAPS+1);
376 if (!gelf_getsym(symbols, i, &sym[nr_maps]))
377 continue;
378 if (sym[nr_maps].st_shndx != maps_shndx)
379 continue;
380 /* Only increment iif maps section */
381 nr_maps++;
382 }
383
384 /* Align to map_fd[] order, via sort on offset in sym.st_value */
385 qsort(sym, nr_maps, sizeof(GElf_Sym), cmp_symbols);
386
387 /* Keeping compatible with ELF maps section changes
388 * ------------------------------------------------
389 * The program size of struct bpf_map_def is known by loader
390 * code, but struct stored in ELF file can be different.
391 *
392 * Unfortunately sym[i].st_size is zero. To calculate the
393 * struct size stored in the ELF file, assume all struct have
394 * the same size, and simply divide with number of map
395 * symbols.
396 */
397 map_sz_elf = data_maps->d_size / nr_maps;
398 map_sz_copy = sizeof(struct bpf_map_def);
399 if (map_sz_elf < map_sz_copy) {
400 /*
401 * Backward compat, loading older ELF file with
402 * smaller struct, keeping remaining bytes zero.
403 */
404 map_sz_copy = map_sz_elf;
405 } else if (map_sz_elf > map_sz_copy) {
406 /*
407 * Forward compat, loading newer ELF file with larger
408 * struct with unknown features. Assume zero means
409 * feature not used. Thus, validate rest of struct
410 * data is zero.
411 */
412 validate_zero = true;
413 }
414
415 /* Memcpy relevant part of ELF maps data to loader maps */
416 for (i = 0; i < nr_maps; i++) {
417 unsigned char *addr, *end;
418 struct bpf_map_def *def;
419 const char *map_name;
420 size_t offset;
421
422 map_name = elf_strptr(elf, strtabidx, sym[i].st_name);
423 maps[i].name = strdup(map_name);
424 if (!maps[i].name) {
425 printf("strdup(%s): %s(%d)\n", map_name,
426 strerror(errno), errno);
427 free(sym);
428 return -errno;
429 }
430
431 /* Symbol value is offset into ELF maps section data area */
432 offset = sym[i].st_value;
433 def = (struct bpf_map_def *)(data_maps->d_buf + offset);
434 maps[i].elf_offset = offset;
435 memset(&maps[i].def, 0, sizeof(struct bpf_map_def));
436 memcpy(&maps[i].def, def, map_sz_copy);
437
438 /* Verify no newer features were requested */
439 if (validate_zero) {
440 addr = (unsigned char*) def + map_sz_copy;
441 end = (unsigned char*) def + map_sz_elf;
442 for (; addr < end; addr++) {
443 if (*addr != 0) {
444 free(sym);
445 return -EFBIG;
446 }
447 }
448 }
449 }
450
451 free(sym);
452 return nr_maps;
453}
454
455static int do_load_bpf_file(const char *path, fixup_map_cb fixup_map)
456{
457 int fd, i, ret, maps_shndx = -1, strtabidx = -1;
458 Elf *elf;
459 GElf_Ehdr ehdr;
460 GElf_Shdr shdr, shdr_prog;
461 Elf_Data *data, *data_prog, *data_maps = NULL, *symbols = NULL;
462 char *shname, *shname_prog;
463 int nr_maps = 0;
464
465 /* reset global variables */
466 kern_version = 0;
467 memset(license, 0, sizeof(license));
468 memset(processed_sec, 0, sizeof(processed_sec));
469
470 if (elf_version(EV_CURRENT) == EV_NONE)
471 return 1;
472
473 fd = open(path, O_RDONLY, 0);
474 if (fd < 0)
475 return 1;
476
477 elf = elf_begin(fd, ELF_C_READ, NULL);
478
479 if (!elf)
480 return 1;
481
482 if (gelf_getehdr(elf, &ehdr) != &ehdr)
483 return 1;
484
485 /* clear all kprobes */
486 i = system("echo \"\" > /sys/kernel/debug/tracing/kprobe_events");
487
488 /* scan over all elf sections to get license and map info */
489 for (i = 1; i < ehdr.e_shnum; i++) {
490
491 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
492 continue;
493
494 if (0) /* helpful for llvm debugging */
495 printf("section %d:%s data %p size %zd link %d flags %d\n",
496 i, shname, data->d_buf, data->d_size,
497 shdr.sh_link, (int) shdr.sh_flags);
498
499 if (strcmp(shname, "license") == 0) {
500 processed_sec[i] = true;
501 memcpy(license, data->d_buf, data->d_size);
502 } else if (strcmp(shname, "version") == 0) {
503 processed_sec[i] = true;
504 if (data->d_size != sizeof(int)) {
505 printf("invalid size of version section %zd\n",
506 data->d_size);
507 return 1;
508 }
509 memcpy(&kern_version, data->d_buf, sizeof(int));
510 } else if (strcmp(shname, "maps") == 0) {
511 int j;
512
513 maps_shndx = i;
514 data_maps = data;
515 for (j = 0; j < MAX_MAPS; j++)
516 map_data[j].fd = -1;
517 } else if (shdr.sh_type == SHT_SYMTAB) {
518 strtabidx = shdr.sh_link;
519 symbols = data;
520 }
521 }
522
523 ret = 1;
524
525 if (!symbols) {
526 printf("missing SHT_SYMTAB section\n");
527 goto done;
528 }
529
530 if (data_maps) {
531 nr_maps = load_elf_maps_section(map_data, maps_shndx,
532 elf, symbols, strtabidx);
533 if (nr_maps < 0) {
534 printf("Error: Failed loading ELF maps (errno:%d):%s\n",
535 nr_maps, strerror(-nr_maps));
536 ret = 1;
537 goto done;
538 }
539 if (load_maps(map_data, nr_maps, fixup_map))
540 goto done;
541 map_data_count = nr_maps;
542
543 processed_sec[maps_shndx] = true;
544 }
545
546 /* process all relo sections, and rewrite bpf insns for maps */
547 for (i = 1; i < ehdr.e_shnum; i++) {
548 if (processed_sec[i])
549 continue;
550
551 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
552 continue;
553
554 if (shdr.sh_type == SHT_REL) {
555 struct bpf_insn *insns;
556
557 /* locate prog sec that need map fixup (relocations) */
558 if (get_sec(elf, shdr.sh_info, &ehdr, &shname_prog,
559 &shdr_prog, &data_prog))
560 continue;
561
562 if (shdr_prog.sh_type != SHT_PROGBITS ||
563 !(shdr_prog.sh_flags & SHF_EXECINSTR))
564 continue;
565
566 insns = (struct bpf_insn *) data_prog->d_buf;
567 processed_sec[i] = true; /* relo section */
568
569 if (parse_relo_and_apply(data, symbols, &shdr, insns,
570 map_data, nr_maps))
571 continue;
572 }
573 }
574
575 /* load programs */
576 for (i = 1; i < ehdr.e_shnum; i++) {
577
578 if (processed_sec[i])
579 continue;
580
581 if (get_sec(elf, i, &ehdr, &shname, &shdr, &data))
582 continue;
583
584 if (memcmp(shname, "kprobe/", 7) == 0 ||
585 memcmp(shname, "kretprobe/", 10) == 0 ||
586 memcmp(shname, "tracepoint/", 11) == 0 ||
587 memcmp(shname, "xdp", 3) == 0 ||
588 memcmp(shname, "perf_event", 10) == 0 ||
589 memcmp(shname, "socket", 6) == 0 ||
590 memcmp(shname, "cgroup/", 7) == 0 ||
591 memcmp(shname, "sockops", 7) == 0 ||
592 memcmp(shname, "sk_skb", 6) == 0) {
593 ret = load_and_attach(shname, data->d_buf,
594 data->d_size);
595 if (ret != 0)
596 goto done;
597 }
598 }
599
600 ret = 0;
601done:
602 close(fd);
603 return ret;
604}
605
606int load_bpf_file(char *path)
607{
608 return do_load_bpf_file(path, NULL);
609}
610
611int load_bpf_file_fixup_map(const char *path, fixup_map_cb fixup_map)
612{
613 return do_load_bpf_file(path, fixup_map);
614}
615
616void read_trace_pipe(void)
617{
618 int trace_fd;
619
620 trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
621 if (trace_fd < 0)
622 return;
623
624 while (1) {
625 static char buf[4096];
626 ssize_t sz;
627
628 sz = read(trace_fd, buf, sizeof(buf));
629 if (sz > 0) {
630 buf[sz] = 0;
631 puts(buf);
632 }
633 }
634}
635
636#define MAX_SYMS 300000
637static struct ksym syms[MAX_SYMS];
638static int sym_cnt;
639
640static int ksym_cmp(const void *p1, const void *p2)
641{
642 return ((struct ksym *)p1)->addr - ((struct ksym *)p2)->addr;
643}
644
645int load_kallsyms(void)
646{
647 FILE *f = fopen("/proc/kallsyms", "r");
648 char func[256], buf[256];
649 char symbol;
650 void *addr;
651 int i = 0;
652
653 if (!f)
654 return -ENOENT;
655
656 while (!feof(f)) {
657 if (!fgets(buf, sizeof(buf), f))
658 break;
659 if (sscanf(buf, "%p %c %s", &addr, &symbol, func) != 3)
660 break;
661 if (!addr)
662 continue;
663 syms[i].addr = (long) addr;
664 syms[i].name = strdup(func);
665 i++;
666 }
667 sym_cnt = i;
668 qsort(syms, sym_cnt, sizeof(struct ksym), ksym_cmp);
669 return 0;
670}
671
672struct ksym *ksym_search(long key)
673{
674 int start = 0, end = sym_cnt;
675 int result;
676
677 while (start < end) {
678 size_t mid = start + (end - start) / 2;
679
680 result = key - syms[mid].addr;
681 if (result < 0)
682 end = mid;
683 else if (result > 0)
684 start = mid + 1;
685 else
686 return &syms[mid];
687 }
688
689 if (start >= 1 && syms[start - 1].addr < key &&
690 key < syms[start].addr)
691 /* valid ksym */
692 return &syms[start - 1];
693
694 /* out of range. return _stext */
695 return &syms[0];
696}
697