]>
Commit | Line | Data |
---|---|---|
53e1b683 | 1 | /* SPDX-License-Identifier: LGPL-2.1+ */ |
1988a9d1 DM |
2 | /*** |
3 | This file is part of systemd. | |
4 | ||
5 | Copyright 2016 Daniel Mack | |
6 | ||
7 | systemd is free software; you can redistribute it and/or modify it | |
8 | under the terms of the GNU Lesser General Public License as published by | |
9 | the Free Software Foundation; either version 2.1 of the License, or | |
10 | (at your option) any later version. | |
11 | ||
12 | systemd is distributed in the hope that it will be useful, but | |
13 | WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 | Lesser General Public License for more details. | |
16 | ||
17 | You should have received a copy of the GNU Lesser General Public License | |
18 | along with systemd; If not, see <http://www.gnu.org/licenses/>. | |
19 | ***/ | |
20 | ||
21 | #include <arpa/inet.h> | |
22 | #include <assert.h> | |
23 | #include <errno.h> | |
24 | #include <fcntl.h> | |
25 | #include <linux/libbpf.h> | |
26 | #include <net/ethernet.h> | |
27 | #include <net/if.h> | |
28 | #include <netinet/ip.h> | |
29 | #include <netinet/ip6.h> | |
30 | #include <stddef.h> | |
31 | #include <stdio.h> | |
32 | #include <stdlib.h> | |
33 | #include <string.h> | |
34 | #include <unistd.h> | |
35 | ||
36 | #include "alloc-util.h" | |
37 | #include "bpf-firewall.h" | |
38 | #include "bpf-program.h" | |
39 | #include "fd-util.h" | |
40 | #include "ip-address-access.h" | |
41 | #include "unit.h" | |
42 | ||
43 | enum { | |
44 | MAP_KEY_PACKETS, | |
45 | MAP_KEY_BYTES, | |
46 | }; | |
47 | ||
48 | enum { | |
49 | ACCESS_ALLOWED = 1, | |
50 | ACCESS_DENIED = 2, | |
51 | }; | |
52 | ||
53 | /* Compile instructions for one list of addresses, one direction and one specific verdict on matches. */ | |
54 | ||
55 | static int add_lookup_instructions( | |
56 | BPFProgram *p, | |
57 | int map_fd, | |
58 | int protocol, | |
59 | bool is_ingress, | |
60 | int verdict) { | |
61 | ||
62 | int r, addr_offset, addr_size; | |
63 | ||
64 | assert(p); | |
65 | assert(map_fd >= 0); | |
66 | ||
67 | switch (protocol) { | |
68 | ||
69 | case ETH_P_IP: | |
70 | addr_size = sizeof(uint32_t); | |
71 | addr_offset = is_ingress ? | |
72 | offsetof(struct iphdr, saddr) : | |
73 | offsetof(struct iphdr, daddr); | |
74 | break; | |
75 | ||
76 | case ETH_P_IPV6: | |
77 | addr_size = 4 * sizeof(uint32_t); | |
78 | addr_offset = is_ingress ? | |
79 | offsetof(struct ip6_hdr, ip6_src.s6_addr) : | |
80 | offsetof(struct ip6_hdr, ip6_dst.s6_addr); | |
81 | break; | |
82 | ||
83 | default: | |
84 | return -EAFNOSUPPORT; | |
85 | } | |
86 | ||
87 | do { | |
88 | /* Compare IPv4 with one word instruction (32bit) */ | |
89 | struct bpf_insn insn[] = { | |
90 | /* If skb->protocol != ETH_P_IP, skip this whole block. The offset will be set later. */ | |
91 | BPF_JMP_IMM(BPF_JNE, BPF_REG_7, htobe16(protocol), 0), | |
92 | ||
93 | /* | |
94 | * Call into BPF_FUNC_skb_load_bytes to load the dst/src IP address | |
95 | * | |
96 | * R1: Pointer to the skb | |
97 | * R2: Data offset | |
98 | * R3: Destination buffer on the stack (r10 - 4) | |
99 | * R4: Number of bytes to read (4) | |
100 | */ | |
101 | ||
102 | BPF_MOV64_REG(BPF_REG_1, BPF_REG_6), | |
103 | BPF_MOV32_IMM(BPF_REG_2, addr_offset), | |
104 | ||
105 | BPF_MOV64_REG(BPF_REG_3, BPF_REG_10), | |
106 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_3, -addr_size), | |
107 | ||
108 | BPF_MOV32_IMM(BPF_REG_4, addr_size), | |
109 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_skb_load_bytes), | |
110 | ||
111 | /* | |
112 | * Call into BPF_FUNC_map_lookup_elem to see if the address matches any entry in the | |
113 | * LPM trie map. For this to work, the prefixlen field of 'struct bpf_lpm_trie_key' | |
114 | * has to be set to the maximum possible value. | |
115 | * | |
116 | * On success, the looked up value is stored in R0. For this application, the actual | |
117 | * value doesn't matter, however; we just set the bit in @verdict in R8 if we found any | |
118 | * matching value. | |
119 | */ | |
120 | ||
121 | BPF_LD_MAP_FD(BPF_REG_1, map_fd), | |
122 | BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), | |
123 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -addr_size - sizeof(uint32_t)), | |
124 | BPF_ST_MEM(BPF_W, BPF_REG_2, 0, addr_size * 8), | |
125 | ||
126 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), | |
127 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 1), | |
128 | BPF_ALU32_IMM(BPF_OR, BPF_REG_8, verdict), | |
129 | }; | |
130 | ||
131 | /* Jump label fixup */ | |
132 | insn[0].off = ELEMENTSOF(insn) - 1; | |
133 | ||
134 | r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn)); | |
135 | if (r < 0) | |
136 | return r; | |
137 | ||
138 | } while (false); | |
139 | ||
140 | return 0; | |
141 | } | |
142 | ||
143 | static int bpf_firewall_compile_bpf( | |
144 | Unit *u, | |
145 | bool is_ingress, | |
146 | BPFProgram **ret) { | |
147 | ||
148 | struct bpf_insn pre_insn[] = { | |
149 | /* | |
150 | * When the eBPF program is entered, R1 contains the address of the skb. | |
151 | * However, R1-R5 are scratch registers that are not preserved when calling | |
152 | * into kernel functions, so we need to save anything that's supposed to | |
153 | * stay around to R6-R9. Save the skb to R6. | |
154 | */ | |
155 | BPF_MOV64_REG(BPF_REG_6, BPF_REG_1), | |
156 | ||
157 | /* | |
158 | * Although we cannot access the skb data directly from eBPF programs used in this | |
159 | * scenario, the kernel has prepared some fields for us to access through struct __sk_buff. | |
160 | * Load the protocol (IPv4, IPv6) used by the packet in flight once and cache it in R7 | |
161 | * for later use. | |
162 | */ | |
163 | BPF_LDX_MEM(BPF_W, BPF_REG_7, BPF_REG_6, offsetof(struct __sk_buff, protocol)), | |
164 | ||
165 | /* | |
166 | * R8 is used to keep track of whether any address check has explicitly allowed or denied the packet | |
167 | * through ACCESS_DENIED or ACCESS_ALLOWED bits. Reset them both to 0 in the beginning. | |
168 | */ | |
169 | BPF_MOV32_IMM(BPF_REG_8, 0), | |
170 | }; | |
171 | ||
172 | /* | |
173 | * The access checkers compiled for the configured allowance and denial lists | |
174 | * write to R8 at runtime. The following code prepares for an early exit that | |
175 | * skip the accounting if the packet is denied. | |
176 | * | |
177 | * R0 = 1 | |
178 | * if (R8 == ACCESS_DENIED) | |
179 | * R0 = 0 | |
180 | * | |
181 | * This means that if both ACCESS_DENIED and ACCESS_ALLOWED are set, the packet | |
182 | * is allowed to pass. | |
183 | */ | |
184 | struct bpf_insn post_insn[] = { | |
185 | BPF_MOV64_IMM(BPF_REG_0, 1), | |
186 | BPF_JMP_IMM(BPF_JNE, BPF_REG_8, ACCESS_DENIED, 1), | |
187 | BPF_MOV64_IMM(BPF_REG_0, 0), | |
188 | }; | |
189 | ||
190 | _cleanup_(bpf_program_unrefp) BPFProgram *p = NULL; | |
191 | int accounting_map_fd, r; | |
192 | bool access_enabled; | |
193 | ||
194 | assert(u); | |
195 | assert(ret); | |
196 | ||
197 | accounting_map_fd = is_ingress ? | |
198 | u->ip_accounting_ingress_map_fd : | |
199 | u->ip_accounting_egress_map_fd; | |
200 | ||
201 | access_enabled = | |
202 | u->ipv4_allow_map_fd >= 0 || | |
203 | u->ipv6_allow_map_fd >= 0 || | |
204 | u->ipv4_deny_map_fd >= 0 || | |
205 | u->ipv6_deny_map_fd >= 0; | |
206 | ||
207 | if (accounting_map_fd < 0 && !access_enabled) { | |
208 | *ret = NULL; | |
209 | return 0; | |
210 | } | |
211 | ||
212 | r = bpf_program_new(BPF_PROG_TYPE_CGROUP_SKB, &p); | |
213 | if (r < 0) | |
214 | return r; | |
215 | ||
216 | r = bpf_program_add_instructions(p, pre_insn, ELEMENTSOF(pre_insn)); | |
217 | if (r < 0) | |
218 | return r; | |
219 | ||
220 | if (access_enabled) { | |
221 | /* | |
222 | * The simple rule this function translates into eBPF instructions is: | |
223 | * | |
224 | * - Access will be granted when an address matches an entry in @list_allow | |
225 | * - Otherwise, access will be denied when an address matches an entry in @list_deny | |
226 | * - Otherwise, access will be granted | |
227 | */ | |
228 | ||
229 | if (u->ipv4_deny_map_fd >= 0) { | |
230 | r = add_lookup_instructions(p, u->ipv4_deny_map_fd, ETH_P_IP, is_ingress, ACCESS_DENIED); | |
231 | if (r < 0) | |
232 | return r; | |
233 | } | |
234 | ||
235 | if (u->ipv6_deny_map_fd >= 0) { | |
236 | r = add_lookup_instructions(p, u->ipv6_deny_map_fd, ETH_P_IPV6, is_ingress, ACCESS_DENIED); | |
237 | if (r < 0) | |
238 | return r; | |
239 | } | |
240 | ||
241 | if (u->ipv4_allow_map_fd >= 0) { | |
242 | r = add_lookup_instructions(p, u->ipv4_allow_map_fd, ETH_P_IP, is_ingress, ACCESS_ALLOWED); | |
243 | if (r < 0) | |
244 | return r; | |
245 | } | |
246 | ||
247 | if (u->ipv6_allow_map_fd >= 0) { | |
248 | r = add_lookup_instructions(p, u->ipv6_allow_map_fd, ETH_P_IPV6, is_ingress, ACCESS_ALLOWED); | |
249 | if (r < 0) | |
250 | return r; | |
251 | } | |
252 | } | |
253 | ||
254 | r = bpf_program_add_instructions(p, post_insn, ELEMENTSOF(post_insn)); | |
255 | if (r < 0) | |
256 | return r; | |
257 | ||
258 | if (accounting_map_fd >= 0) { | |
259 | struct bpf_insn insn[] = { | |
260 | /* | |
261 | * If R0 == 0, the packet will be denied; skip the accounting instructions in this case. | |
262 | * The jump label will be fixed up later. | |
263 | */ | |
264 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 0), | |
265 | ||
266 | /* Count packets */ | |
267 | BPF_MOV64_IMM(BPF_REG_0, MAP_KEY_PACKETS), /* r0 = 0 */ | |
268 | BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_0, -4), /* *(u32 *)(fp - 4) = r0 */ | |
269 | BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), | |
270 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), /* r2 = fp - 4 */ | |
271 | BPF_LD_MAP_FD(BPF_REG_1, accounting_map_fd), /* load map fd to r1 */ | |
272 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), | |
273 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2), | |
274 | BPF_MOV64_IMM(BPF_REG_1, 1), /* r1 = 1 */ | |
275 | BPF_RAW_INSN(BPF_STX | BPF_XADD | BPF_DW, BPF_REG_0, BPF_REG_1, 0, 0), /* xadd r0 += r1 */ | |
276 | ||
277 | /* Count bytes */ | |
278 | BPF_MOV64_IMM(BPF_REG_0, MAP_KEY_BYTES), /* r0 = 1 */ | |
279 | BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_0, -4), /* *(u32 *)(fp - 4) = r0 */ | |
280 | BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), | |
281 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), /* r2 = fp - 4 */ | |
282 | BPF_LD_MAP_FD(BPF_REG_1, accounting_map_fd), | |
283 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), | |
284 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2), | |
285 | BPF_LDX_MEM(BPF_W, BPF_REG_1, BPF_REG_6, offsetof(struct __sk_buff, len)), /* r1 = skb->len */ | |
286 | BPF_RAW_INSN(BPF_STX | BPF_XADD | BPF_DW, BPF_REG_0, BPF_REG_1, 0, 0), /* xadd r0 += r1 */ | |
287 | ||
288 | /* Allow the packet to pass */ | |
289 | BPF_MOV64_IMM(BPF_REG_0, 1), | |
290 | }; | |
291 | ||
292 | /* Jump label fixup */ | |
293 | insn[0].off = ELEMENTSOF(insn) - 1; | |
294 | ||
295 | r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn)); | |
296 | if (r < 0) | |
297 | return r; | |
298 | } | |
299 | ||
300 | do { | |
301 | /* | |
302 | * Exit from the eBPF program, R0 contains the verdict. | |
303 | * 0 means the packet is denied, 1 means the packet may pass. | |
304 | */ | |
305 | struct bpf_insn insn[] = { | |
306 | BPF_EXIT_INSN() | |
307 | }; | |
308 | ||
309 | r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn)); | |
310 | if (r < 0) | |
311 | return r; | |
312 | } while (false); | |
313 | ||
314 | *ret = p; | |
315 | p = NULL; | |
316 | ||
317 | return 0; | |
318 | } | |
319 | ||
320 | static int bpf_firewall_count_access_items(IPAddressAccessItem *list, size_t *n_ipv4, size_t *n_ipv6) { | |
321 | IPAddressAccessItem *a; | |
322 | ||
323 | assert(n_ipv4); | |
324 | assert(n_ipv6); | |
325 | ||
326 | LIST_FOREACH(items, a, list) { | |
327 | switch (a->family) { | |
328 | ||
329 | case AF_INET: | |
330 | (*n_ipv4)++; | |
331 | break; | |
332 | ||
333 | case AF_INET6: | |
334 | (*n_ipv6)++; | |
335 | break; | |
336 | ||
337 | default: | |
338 | return -EAFNOSUPPORT; | |
339 | } | |
340 | } | |
341 | ||
342 | return 0; | |
343 | } | |
344 | ||
345 | static int bpf_firewall_add_access_items( | |
346 | IPAddressAccessItem *list, | |
347 | int ipv4_map_fd, | |
348 | int ipv6_map_fd, | |
349 | int verdict) { | |
350 | ||
351 | struct bpf_lpm_trie_key *key_ipv4, *key_ipv6; | |
352 | uint64_t value = verdict; | |
353 | IPAddressAccessItem *a; | |
354 | int r; | |
355 | ||
356 | key_ipv4 = alloca0(offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t)); | |
357 | key_ipv6 = alloca0(offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t) * 4); | |
358 | ||
359 | LIST_FOREACH(items, a, list) { | |
360 | switch (a->family) { | |
361 | ||
362 | case AF_INET: | |
363 | key_ipv4->prefixlen = a->prefixlen; | |
364 | memcpy(key_ipv4->data, &a->address, sizeof(uint32_t)); | |
365 | ||
366 | r = bpf_map_update_element(ipv4_map_fd, key_ipv4, &value); | |
367 | if (r < 0) | |
368 | return r; | |
369 | ||
370 | break; | |
371 | ||
372 | case AF_INET6: | |
373 | key_ipv6->prefixlen = a->prefixlen; | |
374 | memcpy(key_ipv6->data, &a->address, 4 * sizeof(uint32_t)); | |
375 | ||
376 | r = bpf_map_update_element(ipv6_map_fd, key_ipv6, &value); | |
377 | if (r < 0) | |
378 | return r; | |
379 | ||
380 | break; | |
381 | ||
382 | default: | |
383 | return -EAFNOSUPPORT; | |
384 | } | |
385 | } | |
386 | ||
387 | return 0; | |
388 | } | |
389 | ||
390 | static int bpf_firewall_prepare_access_maps( | |
391 | Unit *u, | |
392 | int verdict, | |
393 | int *ret_ipv4_map_fd, | |
394 | int *ret_ipv6_map_fd) { | |
395 | ||
396 | _cleanup_close_ int ipv4_map_fd = -1, ipv6_map_fd = -1; | |
397 | size_t n_ipv4 = 0, n_ipv6 = 0; | |
398 | Unit *p; | |
399 | int r; | |
400 | ||
401 | assert(ret_ipv4_map_fd); | |
402 | assert(ret_ipv6_map_fd); | |
403 | ||
404 | for (p = u; p; p = UNIT_DEREF(p->slice)) { | |
405 | CGroupContext *cc; | |
406 | ||
407 | cc = unit_get_cgroup_context(p); | |
408 | if (!cc) | |
409 | continue; | |
410 | ||
411 | bpf_firewall_count_access_items(verdict == ACCESS_ALLOWED ? cc->ip_address_allow : cc->ip_address_deny, &n_ipv4, &n_ipv6); | |
412 | } | |
413 | ||
414 | if (n_ipv4 > 0) { | |
415 | ipv4_map_fd = bpf_map_new( | |
416 | BPF_MAP_TYPE_LPM_TRIE, | |
417 | offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t), | |
418 | sizeof(uint64_t), | |
419 | n_ipv4, | |
420 | BPF_F_NO_PREALLOC); | |
421 | if (ipv4_map_fd < 0) | |
422 | return ipv4_map_fd; | |
423 | } | |
424 | ||
425 | if (n_ipv6 > 0) { | |
426 | ipv6_map_fd = bpf_map_new( | |
427 | BPF_MAP_TYPE_LPM_TRIE, | |
428 | offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t)*4, | |
429 | sizeof(uint64_t), | |
430 | n_ipv6, | |
431 | BPF_F_NO_PREALLOC); | |
432 | if (ipv6_map_fd < 0) | |
433 | return ipv6_map_fd; | |
434 | } | |
435 | ||
436 | for (p = u; p; p = UNIT_DEREF(p->slice)) { | |
437 | CGroupContext *cc; | |
438 | ||
439 | cc = unit_get_cgroup_context(p); | |
440 | if (!cc) | |
441 | continue; | |
442 | ||
443 | r = bpf_firewall_add_access_items(verdict == ACCESS_ALLOWED ? cc->ip_address_allow : cc->ip_address_deny, | |
444 | ipv4_map_fd, ipv6_map_fd, verdict); | |
445 | if (r < 0) | |
446 | return r; | |
447 | } | |
448 | ||
449 | *ret_ipv4_map_fd = ipv4_map_fd; | |
450 | *ret_ipv6_map_fd = ipv6_map_fd; | |
451 | ||
452 | ipv4_map_fd = ipv6_map_fd = -1; | |
453 | return 0; | |
454 | } | |
455 | ||
51283461 | 456 | static int bpf_firewall_prepare_accounting_maps(Unit *u, bool enabled, int *fd_ingress, int *fd_egress) { |
1988a9d1 DM |
457 | int r; |
458 | ||
51283461 | 459 | assert(u); |
1988a9d1 DM |
460 | assert(fd_ingress); |
461 | assert(fd_egress); | |
462 | ||
463 | if (enabled) { | |
464 | if (*fd_ingress < 0) { | |
465 | r = bpf_map_new(BPF_MAP_TYPE_ARRAY, sizeof(int), sizeof(uint64_t), 2, 0); | |
466 | if (r < 0) | |
467 | return r; | |
468 | ||
469 | *fd_ingress = r; | |
470 | } | |
471 | ||
472 | if (*fd_egress < 0) { | |
473 | ||
474 | r = bpf_map_new(BPF_MAP_TYPE_ARRAY, sizeof(int), sizeof(uint64_t), 2, 0); | |
475 | if (r < 0) | |
476 | return r; | |
477 | ||
478 | *fd_egress = r; | |
479 | } | |
51283461 | 480 | |
1988a9d1 DM |
481 | } else { |
482 | *fd_ingress = safe_close(*fd_ingress); | |
483 | *fd_egress = safe_close(*fd_egress); | |
51283461 LP |
484 | |
485 | zero(u->ip_accounting_extra); | |
1988a9d1 DM |
486 | } |
487 | ||
488 | return 0; | |
489 | } | |
490 | ||
491 | int bpf_firewall_compile(Unit *u) { | |
492 | CGroupContext *cc; | |
acf7f253 | 493 | int r, supported; |
1988a9d1 DM |
494 | |
495 | assert(u); | |
496 | ||
51283461 LP |
497 | cc = unit_get_cgroup_context(u); |
498 | if (!cc) | |
499 | return -EINVAL; | |
500 | ||
acf7f253 LP |
501 | supported = bpf_firewall_supported(); |
502 | if (supported < 0) | |
503 | return supported; | |
504 | if (supported == BPF_FIREWALL_UNSUPPORTED) { | |
418cdd69 | 505 | log_debug("BPF firewalling not supported on this manager, proceeding without."); |
1988a9d1 DM |
506 | return -EOPNOTSUPP; |
507 | } | |
acf7f253 LP |
508 | if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI && u->type == UNIT_SLICE) { |
509 | /* If BPF_F_ALLOW_MULTI is not supported we don't support any BPF magic on inner nodes (i.e. on slice | |
510 | * units), since that would mean leaf nodes couldn't do any BPF anymore at all. Under the assumption | |
511 | * that BPF is more interesting on leaf nodes we hence avoid it on inner nodes in that case. This is | |
512 | * consistent with old systemd behaviour from before v238, where BPF wasn't supported in inner nodes at | |
513 | * all, either. */ | |
514 | log_debug("BPF_F_ALLOW_MULTI is not supported on this manager, not doing BPF firewall on slice units."); | |
515 | return -EOPNOTSUPP; | |
516 | } | |
1988a9d1 DM |
517 | |
518 | /* Note that when we compile a new firewall we first flush out the access maps and the BPF programs themselves, | |
519 | * but we reuse the the accounting maps. That way the firewall in effect always maps to the actual | |
520 | * configuration, but we don't flush out the accounting unnecessarily */ | |
521 | ||
522 | u->ip_bpf_ingress = bpf_program_unref(u->ip_bpf_ingress); | |
523 | u->ip_bpf_egress = bpf_program_unref(u->ip_bpf_egress); | |
524 | ||
525 | u->ipv4_allow_map_fd = safe_close(u->ipv4_allow_map_fd); | |
526 | u->ipv4_deny_map_fd = safe_close(u->ipv4_deny_map_fd); | |
527 | ||
528 | u->ipv6_allow_map_fd = safe_close(u->ipv6_allow_map_fd); | |
529 | u->ipv6_deny_map_fd = safe_close(u->ipv6_deny_map_fd); | |
530 | ||
acf7f253 LP |
531 | if (u->type != UNIT_SLICE) { |
532 | /* In inner nodes we only do accounting, we do not actually bother with access control. However, leaf | |
533 | * nodes will incorporate all IP access rules set on all their parent nodes. This has the benefit that | |
534 | * they can optionally cancel out system-wide rules. Since inner nodes can't contain processes this | |
535 | * means that all configure IP access rules *will* take effect on processes, even though we never | |
536 | * compile them for inner nodes. */ | |
1988a9d1 | 537 | |
acf7f253 LP |
538 | r = bpf_firewall_prepare_access_maps(u, ACCESS_ALLOWED, &u->ipv4_allow_map_fd, &u->ipv6_allow_map_fd); |
539 | if (r < 0) | |
540 | return log_error_errno(r, "Preparation of eBPF allow maps failed: %m"); | |
541 | ||
542 | r = bpf_firewall_prepare_access_maps(u, ACCESS_DENIED, &u->ipv4_deny_map_fd, &u->ipv6_deny_map_fd); | |
543 | if (r < 0) | |
544 | return log_error_errno(r, "Preparation of eBPF deny maps failed: %m"); | |
545 | } | |
1988a9d1 | 546 | |
51283461 | 547 | r = bpf_firewall_prepare_accounting_maps(u, cc->ip_accounting, &u->ip_accounting_ingress_map_fd, &u->ip_accounting_egress_map_fd); |
1988a9d1 DM |
548 | if (r < 0) |
549 | return log_error_errno(r, "Preparation of eBPF accounting maps failed: %m"); | |
550 | ||
551 | r = bpf_firewall_compile_bpf(u, true, &u->ip_bpf_ingress); | |
552 | if (r < 0) | |
553 | return log_error_errno(r, "Compilation for ingress BPF program failed: %m"); | |
554 | ||
555 | r = bpf_firewall_compile_bpf(u, false, &u->ip_bpf_egress); | |
556 | if (r < 0) | |
557 | return log_error_errno(r, "Compilation for egress BPF program failed: %m"); | |
558 | ||
559 | return 0; | |
560 | } | |
561 | ||
562 | int bpf_firewall_install(Unit *u) { | |
563 | _cleanup_free_ char *path = NULL; | |
9f2e6892 | 564 | CGroupContext *cc; |
acf7f253 | 565 | int r, supported; |
aa2b6f1d | 566 | uint32_t flags; |
1988a9d1 DM |
567 | |
568 | assert(u); | |
569 | ||
9f2e6892 LP |
570 | cc = unit_get_cgroup_context(u); |
571 | if (!cc) | |
572 | return -EINVAL; | |
aa2b6f1d LP |
573 | if (!u->cgroup_path) |
574 | return -EINVAL; | |
575 | if (!u->cgroup_realized) | |
576 | return -EINVAL; | |
9f2e6892 | 577 | |
acf7f253 LP |
578 | supported = bpf_firewall_supported(); |
579 | if (supported < 0) | |
580 | return supported; | |
581 | if (supported == BPF_FIREWALL_UNSUPPORTED) { | |
418cdd69 | 582 | log_debug("BPF firewalling not supported on this manager, proceeding without."); |
1988a9d1 DM |
583 | return -EOPNOTSUPP; |
584 | } | |
acf7f253 LP |
585 | if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI && u->type == UNIT_SLICE) { |
586 | log_debug("BPF_F_ALLOW_MULTI is not supported on this manager, not doing BPF firewall on slice units."); | |
587 | return -EOPNOTSUPP; | |
588 | } | |
1988a9d1 DM |
589 | |
590 | r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, NULL, &path); | |
591 | if (r < 0) | |
592 | return log_error_errno(r, "Failed to determine cgroup path: %m"); | |
593 | ||
acf7f253 LP |
594 | flags = (supported == BPF_FIREWALL_SUPPORTED_WITH_MULTI && |
595 | (u->type == UNIT_SLICE || unit_cgroup_delegate(u))) ? BPF_F_ALLOW_MULTI : 0; | |
596 | ||
aa2b6f1d LP |
597 | /* Unref the old BPF program (which will implicitly detach it) right before attaching the new program, to |
598 | * minimize the time window when we don't account for IP traffic. */ | |
599 | u->ip_bpf_egress_installed = bpf_program_unref(u->ip_bpf_egress_installed); | |
600 | u->ip_bpf_ingress_installed = bpf_program_unref(u->ip_bpf_ingress_installed); | |
1988a9d1 | 601 | |
aa2b6f1d | 602 | if (u->ip_bpf_egress) { |
acf7f253 | 603 | r = bpf_program_cgroup_attach(u->ip_bpf_egress, BPF_CGROUP_INET_EGRESS, path, flags); |
1988a9d1 DM |
604 | if (r < 0) |
605 | return log_error_errno(r, "Attaching egress BPF program to cgroup %s failed: %m", path); | |
aa2b6f1d LP |
606 | |
607 | /* Remember that this BPF program is installed now. */ | |
608 | u->ip_bpf_egress_installed = bpf_program_ref(u->ip_bpf_egress); | |
1988a9d1 DM |
609 | } |
610 | ||
611 | if (u->ip_bpf_ingress) { | |
acf7f253 | 612 | r = bpf_program_cgroup_attach(u->ip_bpf_ingress, BPF_CGROUP_INET_INGRESS, path, flags); |
1988a9d1 DM |
613 | if (r < 0) |
614 | return log_error_errno(r, "Attaching ingress BPF program to cgroup %s failed: %m", path); | |
aa2b6f1d LP |
615 | |
616 | u->ip_bpf_ingress_installed = bpf_program_ref(u->ip_bpf_ingress); | |
1988a9d1 DM |
617 | } |
618 | ||
619 | return 0; | |
620 | } | |
621 | ||
622 | int bpf_firewall_read_accounting(int map_fd, uint64_t *ret_bytes, uint64_t *ret_packets) { | |
623 | uint64_t key, packets; | |
624 | int r; | |
625 | ||
626 | if (map_fd < 0) | |
627 | return -EBADF; | |
628 | ||
629 | if (ret_packets) { | |
630 | key = MAP_KEY_PACKETS; | |
631 | r = bpf_map_lookup_element(map_fd, &key, &packets); | |
632 | if (r < 0) | |
633 | return r; | |
634 | } | |
635 | ||
636 | if (ret_bytes) { | |
637 | key = MAP_KEY_BYTES; | |
638 | r = bpf_map_lookup_element(map_fd, &key, ret_bytes); | |
639 | if (r < 0) | |
640 | return r; | |
641 | } | |
642 | ||
643 | if (ret_packets) | |
644 | *ret_packets = packets; | |
645 | ||
646 | return 0; | |
647 | } | |
648 | ||
649 | int bpf_firewall_reset_accounting(int map_fd) { | |
650 | uint64_t key, value = 0; | |
651 | int r; | |
652 | ||
653 | if (map_fd < 0) | |
654 | return -EBADF; | |
655 | ||
656 | key = MAP_KEY_PACKETS; | |
657 | r = bpf_map_update_element(map_fd, &key, &value); | |
658 | if (r < 0) | |
659 | return r; | |
660 | ||
661 | key = MAP_KEY_BYTES; | |
662 | return bpf_map_update_element(map_fd, &key, &value); | |
663 | } | |
664 | ||
1988a9d1 | 665 | int bpf_firewall_supported(void) { |
93e93da5 LP |
666 | struct bpf_insn trivial[] = { |
667 | BPF_MOV64_IMM(BPF_REG_0, 1), | |
668 | BPF_EXIT_INSN() | |
669 | }; | |
670 | ||
671 | _cleanup_(bpf_program_unrefp) BPFProgram *program = NULL; | |
1988a9d1 | 672 | static int supported = -1; |
e583759b | 673 | union bpf_attr attr; |
1988a9d1 DM |
674 | int fd, r; |
675 | ||
e583759b | 676 | /* Checks whether BPF firewalling is supported. For this, we check five things: |
1988a9d1 DM |
677 | * |
678 | * a) whether we are privileged | |
679 | * b) whether the unified hierarchy is being used | |
680 | * c) the BPF implementation in the kernel supports BPF LPM TRIE maps, which we require | |
e583759b LP |
681 | * d) the BPF implementation in the kernel supports BPF_PROG_TYPE_CGROUP_SKB programs, which we require |
682 | * e) the BPF implementation in the kernel supports the BPF_PROG_ATTACH call, which we require | |
1988a9d1 DM |
683 | * |
684 | */ | |
685 | ||
686 | if (supported >= 0) | |
687 | return supported; | |
688 | ||
93e93da5 LP |
689 | if (geteuid() != 0) { |
690 | log_debug("Not enough privileges, BPF firewalling is not supported."); | |
2ae7ee58 | 691 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 | 692 | } |
1988a9d1 DM |
693 | |
694 | r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER); | |
695 | if (r < 0) | |
696 | return log_error_errno(r, "Can't determine whether the unified hierarchy is used: %m"); | |
e583759b LP |
697 | if (r == 0) { |
698 | log_debug("Not running with unified cgroups, BPF firewalling is not supported."); | |
2ae7ee58 | 699 | return supported = BPF_FIREWALL_UNSUPPORTED; |
e583759b | 700 | } |
1988a9d1 DM |
701 | |
702 | fd = bpf_map_new(BPF_MAP_TYPE_LPM_TRIE, | |
703 | offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint64_t), | |
704 | sizeof(uint64_t), | |
705 | 1, | |
706 | BPF_F_NO_PREALLOC); | |
707 | if (fd < 0) { | |
708 | log_debug_errno(r, "Can't allocate BPF LPM TRIE map, BPF firewalling is not supported: %m"); | |
2ae7ee58 | 709 | return supported = BPF_FIREWALL_UNSUPPORTED; |
1988a9d1 DM |
710 | } |
711 | ||
712 | safe_close(fd); | |
713 | ||
93e93da5 LP |
714 | if (bpf_program_new(BPF_PROG_TYPE_CGROUP_SKB, &program) < 0) { |
715 | log_debug_errno(r, "Can't allocate CGROUP SKB BPF program, BPF firewalling is not supported: %m"); | |
2ae7ee58 | 716 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 LP |
717 | } |
718 | ||
719 | r = bpf_program_add_instructions(program, trivial, ELEMENTSOF(trivial)); | |
720 | if (r < 0) { | |
721 | log_debug_errno(r, "Can't add trivial instructions to CGROUP SKB BPF program, BPF firewalling is not supported: %m"); | |
2ae7ee58 | 722 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 LP |
723 | } |
724 | ||
725 | r = bpf_program_load_kernel(program, NULL, 0); | |
726 | if (r < 0) { | |
727 | log_debug_errno(r, "Can't load kernel CGROUP SKB BPF program, BPF firewalling is not supported: %m"); | |
2ae7ee58 | 728 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 LP |
729 | } |
730 | ||
e583759b LP |
731 | /* Unfortunately the kernel allows us to create BPF_PROG_TYPE_CGROUP_SKB programs even when CONFIG_CGROUP_BPF |
732 | * is turned off at kernel compilation time. This sucks of course: why does it allow us to create a cgroup BPF | |
733 | * program if we can't do a thing with it later? | |
734 | * | |
735 | * We detect this case by issuing the BPF_PROG_ATTACH bpf() call with invalid file descriptors: if | |
736 | * CONFIG_CGROUP_BPF is turned off, then the call will fail early with EINVAL. If it is turned on the | |
737 | * parameters are validated however, and that'll fail with EBADF then. */ | |
738 | ||
739 | attr = (union bpf_attr) { | |
740 | .attach_type = BPF_CGROUP_INET_EGRESS, | |
741 | .target_fd = -1, | |
742 | .attach_bpf_fd = -1, | |
743 | }; | |
744 | ||
745 | r = bpf(BPF_PROG_ATTACH, &attr, sizeof(attr)); | |
746 | if (r < 0) { | |
2ae7ee58 LP |
747 | if (errno != EBADF) { |
748 | log_debug_errno(errno, "Didn't get EBADF from BPF_PROG_ATTACH, BPF firewalling is not supported: %m"); | |
749 | return supported = BPF_FIREWALL_UNSUPPORTED; | |
750 | } | |
751 | ||
752 | /* YAY! */ | |
753 | } else { | |
754 | log_debug("Wut? Kernel accepted our invalid BPF_PROG_ATTACH call? Something is weird, assuming BPF firewalling is broken and hence not supported."); | |
755 | return supported = BPF_FIREWALL_UNSUPPORTED; | |
756 | } | |
e583759b | 757 | |
2ae7ee58 LP |
758 | /* So now we know that the BPF program is generally available, let's see if BPF_F_ALLOW_MULTI is also supported |
759 | * (which was added in kernel 4.15). We use a similar logic as before, but this time we use | |
760 | * BPF_F_ALLOW_MULTI. Since the flags are checked early in the system call we'll get EINVAL if it's not | |
761 | * supported, and EBADF as before if it is available. */ | |
e583759b | 762 | |
2ae7ee58 LP |
763 | attr = (union bpf_attr) { |
764 | .attach_type = BPF_CGROUP_INET_EGRESS, | |
765 | .target_fd = -1, | |
766 | .attach_bpf_fd = -1, | |
767 | .attach_flags = BPF_F_ALLOW_MULTI, | |
768 | }; | |
769 | ||
770 | r = bpf(BPF_PROG_ATTACH, &attr, sizeof(attr)); | |
771 | if (r < 0) { | |
772 | if (errno == EBADF) { | |
773 | log_debug_errno(errno, "Got EBADF when using BPF_F_ALLOW_MULTI, which indicates it is supported. Yay!"); | |
774 | return supported = BPF_FIREWALL_SUPPORTED_WITH_MULTI; | |
775 | } | |
776 | ||
777 | if (errno == EINVAL) | |
778 | log_debug_errno(errno, "Got EINVAL error when using BPF_F_ALLOW_MULTI, which indicates it's not supported."); | |
779 | else | |
780 | log_debug_errno(errno, "Got unexpected error when using BPF_F_ALLOW_MULTI, assuming it's not supported: %m"); | |
781 | ||
782 | return supported = BPF_FIREWALL_SUPPORTED; | |
783 | } else { | |
784 | log_debug("Wut? Kernel accepted our invalid BPF_PROG_ATTACH+BPF_F_ALLOW_MULTI call? Something is weird, assuming BPF firewalling is broken and hence not supported."); | |
785 | return supported = BPF_FIREWALL_UNSUPPORTED; | |
786 | } | |
1988a9d1 | 787 | } |