]>
Commit | Line | Data |
---|---|---|
53e1b683 | 1 | /* SPDX-License-Identifier: LGPL-2.1+ */ |
1988a9d1 DM |
2 | |
3 | #include <arpa/inet.h> | |
4 | #include <assert.h> | |
5 | #include <errno.h> | |
6 | #include <fcntl.h> | |
7 | #include <linux/libbpf.h> | |
8 | #include <net/ethernet.h> | |
9 | #include <net/if.h> | |
10 | #include <netinet/ip.h> | |
11 | #include <netinet/ip6.h> | |
12 | #include <stddef.h> | |
13 | #include <stdio.h> | |
14 | #include <stdlib.h> | |
15 | #include <string.h> | |
16 | #include <unistd.h> | |
17 | ||
18 | #include "alloc-util.h" | |
19 | #include "bpf-firewall.h" | |
20 | #include "bpf-program.h" | |
21 | #include "fd-util.h" | |
22 | #include "ip-address-access.h" | |
23 | #include "unit.h" | |
24 | ||
25 | enum { | |
26 | MAP_KEY_PACKETS, | |
27 | MAP_KEY_BYTES, | |
28 | }; | |
29 | ||
30 | enum { | |
31 | ACCESS_ALLOWED = 1, | |
32 | ACCESS_DENIED = 2, | |
33 | }; | |
34 | ||
35 | /* Compile instructions for one list of addresses, one direction and one specific verdict on matches. */ | |
36 | ||
37 | static int add_lookup_instructions( | |
38 | BPFProgram *p, | |
39 | int map_fd, | |
40 | int protocol, | |
41 | bool is_ingress, | |
42 | int verdict) { | |
43 | ||
44 | int r, addr_offset, addr_size; | |
45 | ||
46 | assert(p); | |
47 | assert(map_fd >= 0); | |
48 | ||
49 | switch (protocol) { | |
50 | ||
51 | case ETH_P_IP: | |
52 | addr_size = sizeof(uint32_t); | |
53 | addr_offset = is_ingress ? | |
54 | offsetof(struct iphdr, saddr) : | |
55 | offsetof(struct iphdr, daddr); | |
56 | break; | |
57 | ||
58 | case ETH_P_IPV6: | |
59 | addr_size = 4 * sizeof(uint32_t); | |
60 | addr_offset = is_ingress ? | |
61 | offsetof(struct ip6_hdr, ip6_src.s6_addr) : | |
62 | offsetof(struct ip6_hdr, ip6_dst.s6_addr); | |
63 | break; | |
64 | ||
65 | default: | |
66 | return -EAFNOSUPPORT; | |
67 | } | |
68 | ||
69 | do { | |
70 | /* Compare IPv4 with one word instruction (32bit) */ | |
71 | struct bpf_insn insn[] = { | |
72 | /* If skb->protocol != ETH_P_IP, skip this whole block. The offset will be set later. */ | |
73 | BPF_JMP_IMM(BPF_JNE, BPF_REG_7, htobe16(protocol), 0), | |
74 | ||
75 | /* | |
76 | * Call into BPF_FUNC_skb_load_bytes to load the dst/src IP address | |
77 | * | |
78 | * R1: Pointer to the skb | |
79 | * R2: Data offset | |
80 | * R3: Destination buffer on the stack (r10 - 4) | |
81 | * R4: Number of bytes to read (4) | |
82 | */ | |
83 | ||
84 | BPF_MOV64_REG(BPF_REG_1, BPF_REG_6), | |
85 | BPF_MOV32_IMM(BPF_REG_2, addr_offset), | |
86 | ||
87 | BPF_MOV64_REG(BPF_REG_3, BPF_REG_10), | |
88 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_3, -addr_size), | |
89 | ||
90 | BPF_MOV32_IMM(BPF_REG_4, addr_size), | |
91 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_skb_load_bytes), | |
92 | ||
93 | /* | |
94 | * Call into BPF_FUNC_map_lookup_elem to see if the address matches any entry in the | |
95 | * LPM trie map. For this to work, the prefixlen field of 'struct bpf_lpm_trie_key' | |
96 | * has to be set to the maximum possible value. | |
97 | * | |
98 | * On success, the looked up value is stored in R0. For this application, the actual | |
99 | * value doesn't matter, however; we just set the bit in @verdict in R8 if we found any | |
100 | * matching value. | |
101 | */ | |
102 | ||
103 | BPF_LD_MAP_FD(BPF_REG_1, map_fd), | |
104 | BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), | |
105 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -addr_size - sizeof(uint32_t)), | |
106 | BPF_ST_MEM(BPF_W, BPF_REG_2, 0, addr_size * 8), | |
107 | ||
108 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), | |
109 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 1), | |
110 | BPF_ALU32_IMM(BPF_OR, BPF_REG_8, verdict), | |
111 | }; | |
112 | ||
113 | /* Jump label fixup */ | |
114 | insn[0].off = ELEMENTSOF(insn) - 1; | |
115 | ||
116 | r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn)); | |
117 | if (r < 0) | |
118 | return r; | |
119 | ||
120 | } while (false); | |
121 | ||
122 | return 0; | |
123 | } | |
124 | ||
125 | static int bpf_firewall_compile_bpf( | |
126 | Unit *u, | |
127 | bool is_ingress, | |
128 | BPFProgram **ret) { | |
129 | ||
130 | struct bpf_insn pre_insn[] = { | |
131 | /* | |
132 | * When the eBPF program is entered, R1 contains the address of the skb. | |
133 | * However, R1-R5 are scratch registers that are not preserved when calling | |
134 | * into kernel functions, so we need to save anything that's supposed to | |
135 | * stay around to R6-R9. Save the skb to R6. | |
136 | */ | |
137 | BPF_MOV64_REG(BPF_REG_6, BPF_REG_1), | |
138 | ||
139 | /* | |
140 | * Although we cannot access the skb data directly from eBPF programs used in this | |
141 | * scenario, the kernel has prepared some fields for us to access through struct __sk_buff. | |
142 | * Load the protocol (IPv4, IPv6) used by the packet in flight once and cache it in R7 | |
143 | * for later use. | |
144 | */ | |
145 | BPF_LDX_MEM(BPF_W, BPF_REG_7, BPF_REG_6, offsetof(struct __sk_buff, protocol)), | |
146 | ||
147 | /* | |
148 | * R8 is used to keep track of whether any address check has explicitly allowed or denied the packet | |
149 | * through ACCESS_DENIED or ACCESS_ALLOWED bits. Reset them both to 0 in the beginning. | |
150 | */ | |
151 | BPF_MOV32_IMM(BPF_REG_8, 0), | |
152 | }; | |
153 | ||
154 | /* | |
155 | * The access checkers compiled for the configured allowance and denial lists | |
156 | * write to R8 at runtime. The following code prepares for an early exit that | |
157 | * skip the accounting if the packet is denied. | |
158 | * | |
159 | * R0 = 1 | |
160 | * if (R8 == ACCESS_DENIED) | |
161 | * R0 = 0 | |
162 | * | |
163 | * This means that if both ACCESS_DENIED and ACCESS_ALLOWED are set, the packet | |
164 | * is allowed to pass. | |
165 | */ | |
166 | struct bpf_insn post_insn[] = { | |
167 | BPF_MOV64_IMM(BPF_REG_0, 1), | |
168 | BPF_JMP_IMM(BPF_JNE, BPF_REG_8, ACCESS_DENIED, 1), | |
169 | BPF_MOV64_IMM(BPF_REG_0, 0), | |
170 | }; | |
171 | ||
172 | _cleanup_(bpf_program_unrefp) BPFProgram *p = NULL; | |
173 | int accounting_map_fd, r; | |
174 | bool access_enabled; | |
175 | ||
176 | assert(u); | |
177 | assert(ret); | |
178 | ||
179 | accounting_map_fd = is_ingress ? | |
180 | u->ip_accounting_ingress_map_fd : | |
181 | u->ip_accounting_egress_map_fd; | |
182 | ||
183 | access_enabled = | |
184 | u->ipv4_allow_map_fd >= 0 || | |
185 | u->ipv6_allow_map_fd >= 0 || | |
186 | u->ipv4_deny_map_fd >= 0 || | |
187 | u->ipv6_deny_map_fd >= 0; | |
188 | ||
189 | if (accounting_map_fd < 0 && !access_enabled) { | |
190 | *ret = NULL; | |
191 | return 0; | |
192 | } | |
193 | ||
194 | r = bpf_program_new(BPF_PROG_TYPE_CGROUP_SKB, &p); | |
195 | if (r < 0) | |
196 | return r; | |
197 | ||
198 | r = bpf_program_add_instructions(p, pre_insn, ELEMENTSOF(pre_insn)); | |
199 | if (r < 0) | |
200 | return r; | |
201 | ||
202 | if (access_enabled) { | |
203 | /* | |
204 | * The simple rule this function translates into eBPF instructions is: | |
205 | * | |
206 | * - Access will be granted when an address matches an entry in @list_allow | |
207 | * - Otherwise, access will be denied when an address matches an entry in @list_deny | |
208 | * - Otherwise, access will be granted | |
209 | */ | |
210 | ||
211 | if (u->ipv4_deny_map_fd >= 0) { | |
212 | r = add_lookup_instructions(p, u->ipv4_deny_map_fd, ETH_P_IP, is_ingress, ACCESS_DENIED); | |
213 | if (r < 0) | |
214 | return r; | |
215 | } | |
216 | ||
217 | if (u->ipv6_deny_map_fd >= 0) { | |
218 | r = add_lookup_instructions(p, u->ipv6_deny_map_fd, ETH_P_IPV6, is_ingress, ACCESS_DENIED); | |
219 | if (r < 0) | |
220 | return r; | |
221 | } | |
222 | ||
223 | if (u->ipv4_allow_map_fd >= 0) { | |
224 | r = add_lookup_instructions(p, u->ipv4_allow_map_fd, ETH_P_IP, is_ingress, ACCESS_ALLOWED); | |
225 | if (r < 0) | |
226 | return r; | |
227 | } | |
228 | ||
229 | if (u->ipv6_allow_map_fd >= 0) { | |
230 | r = add_lookup_instructions(p, u->ipv6_allow_map_fd, ETH_P_IPV6, is_ingress, ACCESS_ALLOWED); | |
231 | if (r < 0) | |
232 | return r; | |
233 | } | |
234 | } | |
235 | ||
236 | r = bpf_program_add_instructions(p, post_insn, ELEMENTSOF(post_insn)); | |
237 | if (r < 0) | |
238 | return r; | |
239 | ||
240 | if (accounting_map_fd >= 0) { | |
241 | struct bpf_insn insn[] = { | |
242 | /* | |
243 | * If R0 == 0, the packet will be denied; skip the accounting instructions in this case. | |
244 | * The jump label will be fixed up later. | |
245 | */ | |
246 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 0), | |
247 | ||
248 | /* Count packets */ | |
249 | BPF_MOV64_IMM(BPF_REG_0, MAP_KEY_PACKETS), /* r0 = 0 */ | |
250 | BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_0, -4), /* *(u32 *)(fp - 4) = r0 */ | |
251 | BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), | |
252 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), /* r2 = fp - 4 */ | |
253 | BPF_LD_MAP_FD(BPF_REG_1, accounting_map_fd), /* load map fd to r1 */ | |
254 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), | |
255 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2), | |
256 | BPF_MOV64_IMM(BPF_REG_1, 1), /* r1 = 1 */ | |
257 | BPF_RAW_INSN(BPF_STX | BPF_XADD | BPF_DW, BPF_REG_0, BPF_REG_1, 0, 0), /* xadd r0 += r1 */ | |
258 | ||
259 | /* Count bytes */ | |
260 | BPF_MOV64_IMM(BPF_REG_0, MAP_KEY_BYTES), /* r0 = 1 */ | |
261 | BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_0, -4), /* *(u32 *)(fp - 4) = r0 */ | |
262 | BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), | |
263 | BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), /* r2 = fp - 4 */ | |
264 | BPF_LD_MAP_FD(BPF_REG_1, accounting_map_fd), | |
265 | BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), | |
266 | BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2), | |
267 | BPF_LDX_MEM(BPF_W, BPF_REG_1, BPF_REG_6, offsetof(struct __sk_buff, len)), /* r1 = skb->len */ | |
268 | BPF_RAW_INSN(BPF_STX | BPF_XADD | BPF_DW, BPF_REG_0, BPF_REG_1, 0, 0), /* xadd r0 += r1 */ | |
269 | ||
270 | /* Allow the packet to pass */ | |
271 | BPF_MOV64_IMM(BPF_REG_0, 1), | |
272 | }; | |
273 | ||
274 | /* Jump label fixup */ | |
275 | insn[0].off = ELEMENTSOF(insn) - 1; | |
276 | ||
277 | r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn)); | |
278 | if (r < 0) | |
279 | return r; | |
280 | } | |
281 | ||
282 | do { | |
283 | /* | |
284 | * Exit from the eBPF program, R0 contains the verdict. | |
285 | * 0 means the packet is denied, 1 means the packet may pass. | |
286 | */ | |
287 | struct bpf_insn insn[] = { | |
288 | BPF_EXIT_INSN() | |
289 | }; | |
290 | ||
291 | r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn)); | |
292 | if (r < 0) | |
293 | return r; | |
294 | } while (false); | |
295 | ||
1cc6c93a | 296 | *ret = TAKE_PTR(p); |
1988a9d1 DM |
297 | |
298 | return 0; | |
299 | } | |
300 | ||
301 | static int bpf_firewall_count_access_items(IPAddressAccessItem *list, size_t *n_ipv4, size_t *n_ipv6) { | |
302 | IPAddressAccessItem *a; | |
303 | ||
304 | assert(n_ipv4); | |
305 | assert(n_ipv6); | |
306 | ||
307 | LIST_FOREACH(items, a, list) { | |
308 | switch (a->family) { | |
309 | ||
310 | case AF_INET: | |
311 | (*n_ipv4)++; | |
312 | break; | |
313 | ||
314 | case AF_INET6: | |
315 | (*n_ipv6)++; | |
316 | break; | |
317 | ||
318 | default: | |
319 | return -EAFNOSUPPORT; | |
320 | } | |
321 | } | |
322 | ||
323 | return 0; | |
324 | } | |
325 | ||
326 | static int bpf_firewall_add_access_items( | |
327 | IPAddressAccessItem *list, | |
328 | int ipv4_map_fd, | |
329 | int ipv6_map_fd, | |
330 | int verdict) { | |
331 | ||
332 | struct bpf_lpm_trie_key *key_ipv4, *key_ipv6; | |
333 | uint64_t value = verdict; | |
334 | IPAddressAccessItem *a; | |
335 | int r; | |
336 | ||
337 | key_ipv4 = alloca0(offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t)); | |
338 | key_ipv6 = alloca0(offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t) * 4); | |
339 | ||
340 | LIST_FOREACH(items, a, list) { | |
341 | switch (a->family) { | |
342 | ||
343 | case AF_INET: | |
344 | key_ipv4->prefixlen = a->prefixlen; | |
345 | memcpy(key_ipv4->data, &a->address, sizeof(uint32_t)); | |
346 | ||
347 | r = bpf_map_update_element(ipv4_map_fd, key_ipv4, &value); | |
348 | if (r < 0) | |
349 | return r; | |
350 | ||
351 | break; | |
352 | ||
353 | case AF_INET6: | |
354 | key_ipv6->prefixlen = a->prefixlen; | |
355 | memcpy(key_ipv6->data, &a->address, 4 * sizeof(uint32_t)); | |
356 | ||
357 | r = bpf_map_update_element(ipv6_map_fd, key_ipv6, &value); | |
358 | if (r < 0) | |
359 | return r; | |
360 | ||
361 | break; | |
362 | ||
363 | default: | |
364 | return -EAFNOSUPPORT; | |
365 | } | |
366 | } | |
367 | ||
368 | return 0; | |
369 | } | |
370 | ||
371 | static int bpf_firewall_prepare_access_maps( | |
372 | Unit *u, | |
373 | int verdict, | |
374 | int *ret_ipv4_map_fd, | |
375 | int *ret_ipv6_map_fd) { | |
376 | ||
377 | _cleanup_close_ int ipv4_map_fd = -1, ipv6_map_fd = -1; | |
378 | size_t n_ipv4 = 0, n_ipv6 = 0; | |
379 | Unit *p; | |
380 | int r; | |
381 | ||
382 | assert(ret_ipv4_map_fd); | |
383 | assert(ret_ipv6_map_fd); | |
384 | ||
385 | for (p = u; p; p = UNIT_DEREF(p->slice)) { | |
386 | CGroupContext *cc; | |
387 | ||
388 | cc = unit_get_cgroup_context(p); | |
389 | if (!cc) | |
390 | continue; | |
391 | ||
392 | bpf_firewall_count_access_items(verdict == ACCESS_ALLOWED ? cc->ip_address_allow : cc->ip_address_deny, &n_ipv4, &n_ipv6); | |
393 | } | |
394 | ||
395 | if (n_ipv4 > 0) { | |
396 | ipv4_map_fd = bpf_map_new( | |
397 | BPF_MAP_TYPE_LPM_TRIE, | |
398 | offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t), | |
399 | sizeof(uint64_t), | |
400 | n_ipv4, | |
401 | BPF_F_NO_PREALLOC); | |
402 | if (ipv4_map_fd < 0) | |
403 | return ipv4_map_fd; | |
404 | } | |
405 | ||
406 | if (n_ipv6 > 0) { | |
407 | ipv6_map_fd = bpf_map_new( | |
408 | BPF_MAP_TYPE_LPM_TRIE, | |
409 | offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t)*4, | |
410 | sizeof(uint64_t), | |
411 | n_ipv6, | |
412 | BPF_F_NO_PREALLOC); | |
413 | if (ipv6_map_fd < 0) | |
414 | return ipv6_map_fd; | |
415 | } | |
416 | ||
417 | for (p = u; p; p = UNIT_DEREF(p->slice)) { | |
418 | CGroupContext *cc; | |
419 | ||
420 | cc = unit_get_cgroup_context(p); | |
421 | if (!cc) | |
422 | continue; | |
423 | ||
424 | r = bpf_firewall_add_access_items(verdict == ACCESS_ALLOWED ? cc->ip_address_allow : cc->ip_address_deny, | |
425 | ipv4_map_fd, ipv6_map_fd, verdict); | |
426 | if (r < 0) | |
427 | return r; | |
428 | } | |
429 | ||
430 | *ret_ipv4_map_fd = ipv4_map_fd; | |
431 | *ret_ipv6_map_fd = ipv6_map_fd; | |
432 | ||
433 | ipv4_map_fd = ipv6_map_fd = -1; | |
434 | return 0; | |
435 | } | |
436 | ||
51283461 | 437 | static int bpf_firewall_prepare_accounting_maps(Unit *u, bool enabled, int *fd_ingress, int *fd_egress) { |
1988a9d1 DM |
438 | int r; |
439 | ||
51283461 | 440 | assert(u); |
1988a9d1 DM |
441 | assert(fd_ingress); |
442 | assert(fd_egress); | |
443 | ||
444 | if (enabled) { | |
445 | if (*fd_ingress < 0) { | |
446 | r = bpf_map_new(BPF_MAP_TYPE_ARRAY, sizeof(int), sizeof(uint64_t), 2, 0); | |
447 | if (r < 0) | |
448 | return r; | |
449 | ||
450 | *fd_ingress = r; | |
451 | } | |
452 | ||
453 | if (*fd_egress < 0) { | |
454 | ||
455 | r = bpf_map_new(BPF_MAP_TYPE_ARRAY, sizeof(int), sizeof(uint64_t), 2, 0); | |
456 | if (r < 0) | |
457 | return r; | |
458 | ||
459 | *fd_egress = r; | |
460 | } | |
51283461 | 461 | |
1988a9d1 DM |
462 | } else { |
463 | *fd_ingress = safe_close(*fd_ingress); | |
464 | *fd_egress = safe_close(*fd_egress); | |
51283461 LP |
465 | |
466 | zero(u->ip_accounting_extra); | |
1988a9d1 DM |
467 | } |
468 | ||
469 | return 0; | |
470 | } | |
471 | ||
472 | int bpf_firewall_compile(Unit *u) { | |
473 | CGroupContext *cc; | |
acf7f253 | 474 | int r, supported; |
1988a9d1 DM |
475 | |
476 | assert(u); | |
477 | ||
51283461 LP |
478 | cc = unit_get_cgroup_context(u); |
479 | if (!cc) | |
480 | return -EINVAL; | |
481 | ||
acf7f253 LP |
482 | supported = bpf_firewall_supported(); |
483 | if (supported < 0) | |
484 | return supported; | |
485 | if (supported == BPF_FIREWALL_UNSUPPORTED) { | |
418cdd69 | 486 | log_debug("BPF firewalling not supported on this manager, proceeding without."); |
1988a9d1 DM |
487 | return -EOPNOTSUPP; |
488 | } | |
acf7f253 LP |
489 | if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI && u->type == UNIT_SLICE) { |
490 | /* If BPF_F_ALLOW_MULTI is not supported we don't support any BPF magic on inner nodes (i.e. on slice | |
491 | * units), since that would mean leaf nodes couldn't do any BPF anymore at all. Under the assumption | |
492 | * that BPF is more interesting on leaf nodes we hence avoid it on inner nodes in that case. This is | |
493 | * consistent with old systemd behaviour from before v238, where BPF wasn't supported in inner nodes at | |
494 | * all, either. */ | |
495 | log_debug("BPF_F_ALLOW_MULTI is not supported on this manager, not doing BPF firewall on slice units."); | |
496 | return -EOPNOTSUPP; | |
497 | } | |
1988a9d1 DM |
498 | |
499 | /* Note that when we compile a new firewall we first flush out the access maps and the BPF programs themselves, | |
500 | * but we reuse the the accounting maps. That way the firewall in effect always maps to the actual | |
501 | * configuration, but we don't flush out the accounting unnecessarily */ | |
502 | ||
503 | u->ip_bpf_ingress = bpf_program_unref(u->ip_bpf_ingress); | |
504 | u->ip_bpf_egress = bpf_program_unref(u->ip_bpf_egress); | |
505 | ||
506 | u->ipv4_allow_map_fd = safe_close(u->ipv4_allow_map_fd); | |
507 | u->ipv4_deny_map_fd = safe_close(u->ipv4_deny_map_fd); | |
508 | ||
509 | u->ipv6_allow_map_fd = safe_close(u->ipv6_allow_map_fd); | |
510 | u->ipv6_deny_map_fd = safe_close(u->ipv6_deny_map_fd); | |
511 | ||
acf7f253 LP |
512 | if (u->type != UNIT_SLICE) { |
513 | /* In inner nodes we only do accounting, we do not actually bother with access control. However, leaf | |
514 | * nodes will incorporate all IP access rules set on all their parent nodes. This has the benefit that | |
515 | * they can optionally cancel out system-wide rules. Since inner nodes can't contain processes this | |
516 | * means that all configure IP access rules *will* take effect on processes, even though we never | |
517 | * compile them for inner nodes. */ | |
1988a9d1 | 518 | |
acf7f253 LP |
519 | r = bpf_firewall_prepare_access_maps(u, ACCESS_ALLOWED, &u->ipv4_allow_map_fd, &u->ipv6_allow_map_fd); |
520 | if (r < 0) | |
521 | return log_error_errno(r, "Preparation of eBPF allow maps failed: %m"); | |
522 | ||
523 | r = bpf_firewall_prepare_access_maps(u, ACCESS_DENIED, &u->ipv4_deny_map_fd, &u->ipv6_deny_map_fd); | |
524 | if (r < 0) | |
525 | return log_error_errno(r, "Preparation of eBPF deny maps failed: %m"); | |
526 | } | |
1988a9d1 | 527 | |
51283461 | 528 | r = bpf_firewall_prepare_accounting_maps(u, cc->ip_accounting, &u->ip_accounting_ingress_map_fd, &u->ip_accounting_egress_map_fd); |
1988a9d1 DM |
529 | if (r < 0) |
530 | return log_error_errno(r, "Preparation of eBPF accounting maps failed: %m"); | |
531 | ||
532 | r = bpf_firewall_compile_bpf(u, true, &u->ip_bpf_ingress); | |
533 | if (r < 0) | |
534 | return log_error_errno(r, "Compilation for ingress BPF program failed: %m"); | |
535 | ||
536 | r = bpf_firewall_compile_bpf(u, false, &u->ip_bpf_egress); | |
537 | if (r < 0) | |
538 | return log_error_errno(r, "Compilation for egress BPF program failed: %m"); | |
539 | ||
540 | return 0; | |
541 | } | |
542 | ||
543 | int bpf_firewall_install(Unit *u) { | |
544 | _cleanup_free_ char *path = NULL; | |
9f2e6892 | 545 | CGroupContext *cc; |
acf7f253 | 546 | int r, supported; |
aa2b6f1d | 547 | uint32_t flags; |
1988a9d1 DM |
548 | |
549 | assert(u); | |
550 | ||
9f2e6892 LP |
551 | cc = unit_get_cgroup_context(u); |
552 | if (!cc) | |
553 | return -EINVAL; | |
aa2b6f1d LP |
554 | if (!u->cgroup_path) |
555 | return -EINVAL; | |
556 | if (!u->cgroup_realized) | |
557 | return -EINVAL; | |
9f2e6892 | 558 | |
acf7f253 LP |
559 | supported = bpf_firewall_supported(); |
560 | if (supported < 0) | |
561 | return supported; | |
562 | if (supported == BPF_FIREWALL_UNSUPPORTED) { | |
418cdd69 | 563 | log_debug("BPF firewalling not supported on this manager, proceeding without."); |
1988a9d1 DM |
564 | return -EOPNOTSUPP; |
565 | } | |
acf7f253 LP |
566 | if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI && u->type == UNIT_SLICE) { |
567 | log_debug("BPF_F_ALLOW_MULTI is not supported on this manager, not doing BPF firewall on slice units."); | |
568 | return -EOPNOTSUPP; | |
569 | } | |
1988a9d1 DM |
570 | |
571 | r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, NULL, &path); | |
572 | if (r < 0) | |
573 | return log_error_errno(r, "Failed to determine cgroup path: %m"); | |
574 | ||
acf7f253 LP |
575 | flags = (supported == BPF_FIREWALL_SUPPORTED_WITH_MULTI && |
576 | (u->type == UNIT_SLICE || unit_cgroup_delegate(u))) ? BPF_F_ALLOW_MULTI : 0; | |
577 | ||
aa2b6f1d LP |
578 | /* Unref the old BPF program (which will implicitly detach it) right before attaching the new program, to |
579 | * minimize the time window when we don't account for IP traffic. */ | |
580 | u->ip_bpf_egress_installed = bpf_program_unref(u->ip_bpf_egress_installed); | |
581 | u->ip_bpf_ingress_installed = bpf_program_unref(u->ip_bpf_ingress_installed); | |
1988a9d1 | 582 | |
aa2b6f1d | 583 | if (u->ip_bpf_egress) { |
acf7f253 | 584 | r = bpf_program_cgroup_attach(u->ip_bpf_egress, BPF_CGROUP_INET_EGRESS, path, flags); |
1988a9d1 DM |
585 | if (r < 0) |
586 | return log_error_errno(r, "Attaching egress BPF program to cgroup %s failed: %m", path); | |
aa2b6f1d LP |
587 | |
588 | /* Remember that this BPF program is installed now. */ | |
589 | u->ip_bpf_egress_installed = bpf_program_ref(u->ip_bpf_egress); | |
1988a9d1 DM |
590 | } |
591 | ||
592 | if (u->ip_bpf_ingress) { | |
acf7f253 | 593 | r = bpf_program_cgroup_attach(u->ip_bpf_ingress, BPF_CGROUP_INET_INGRESS, path, flags); |
1988a9d1 DM |
594 | if (r < 0) |
595 | return log_error_errno(r, "Attaching ingress BPF program to cgroup %s failed: %m", path); | |
aa2b6f1d LP |
596 | |
597 | u->ip_bpf_ingress_installed = bpf_program_ref(u->ip_bpf_ingress); | |
1988a9d1 DM |
598 | } |
599 | ||
600 | return 0; | |
601 | } | |
602 | ||
603 | int bpf_firewall_read_accounting(int map_fd, uint64_t *ret_bytes, uint64_t *ret_packets) { | |
604 | uint64_t key, packets; | |
605 | int r; | |
606 | ||
607 | if (map_fd < 0) | |
608 | return -EBADF; | |
609 | ||
610 | if (ret_packets) { | |
611 | key = MAP_KEY_PACKETS; | |
612 | r = bpf_map_lookup_element(map_fd, &key, &packets); | |
613 | if (r < 0) | |
614 | return r; | |
615 | } | |
616 | ||
617 | if (ret_bytes) { | |
618 | key = MAP_KEY_BYTES; | |
619 | r = bpf_map_lookup_element(map_fd, &key, ret_bytes); | |
620 | if (r < 0) | |
621 | return r; | |
622 | } | |
623 | ||
624 | if (ret_packets) | |
625 | *ret_packets = packets; | |
626 | ||
627 | return 0; | |
628 | } | |
629 | ||
630 | int bpf_firewall_reset_accounting(int map_fd) { | |
631 | uint64_t key, value = 0; | |
632 | int r; | |
633 | ||
634 | if (map_fd < 0) | |
635 | return -EBADF; | |
636 | ||
637 | key = MAP_KEY_PACKETS; | |
638 | r = bpf_map_update_element(map_fd, &key, &value); | |
639 | if (r < 0) | |
640 | return r; | |
641 | ||
642 | key = MAP_KEY_BYTES; | |
643 | return bpf_map_update_element(map_fd, &key, &value); | |
644 | } | |
645 | ||
1988a9d1 | 646 | int bpf_firewall_supported(void) { |
93e93da5 LP |
647 | struct bpf_insn trivial[] = { |
648 | BPF_MOV64_IMM(BPF_REG_0, 1), | |
649 | BPF_EXIT_INSN() | |
650 | }; | |
651 | ||
652 | _cleanup_(bpf_program_unrefp) BPFProgram *program = NULL; | |
1988a9d1 | 653 | static int supported = -1; |
e583759b | 654 | union bpf_attr attr; |
1988a9d1 DM |
655 | int fd, r; |
656 | ||
e583759b | 657 | /* Checks whether BPF firewalling is supported. For this, we check five things: |
1988a9d1 DM |
658 | * |
659 | * a) whether we are privileged | |
660 | * b) whether the unified hierarchy is being used | |
661 | * c) the BPF implementation in the kernel supports BPF LPM TRIE maps, which we require | |
e583759b | 662 | * d) the BPF implementation in the kernel supports BPF_PROG_TYPE_CGROUP_SKB programs, which we require |
047de7e1 | 663 | * e) the BPF implementation in the kernel supports the BPF_PROG_DETACH call, which we require |
1988a9d1 DM |
664 | */ |
665 | ||
666 | if (supported >= 0) | |
667 | return supported; | |
668 | ||
93e93da5 LP |
669 | if (geteuid() != 0) { |
670 | log_debug("Not enough privileges, BPF firewalling is not supported."); | |
2ae7ee58 | 671 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 | 672 | } |
1988a9d1 DM |
673 | |
674 | r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER); | |
675 | if (r < 0) | |
676 | return log_error_errno(r, "Can't determine whether the unified hierarchy is used: %m"); | |
e583759b LP |
677 | if (r == 0) { |
678 | log_debug("Not running with unified cgroups, BPF firewalling is not supported."); | |
2ae7ee58 | 679 | return supported = BPF_FIREWALL_UNSUPPORTED; |
e583759b | 680 | } |
1988a9d1 DM |
681 | |
682 | fd = bpf_map_new(BPF_MAP_TYPE_LPM_TRIE, | |
683 | offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint64_t), | |
684 | sizeof(uint64_t), | |
685 | 1, | |
686 | BPF_F_NO_PREALLOC); | |
687 | if (fd < 0) { | |
4355f1c9 | 688 | log_debug_errno(fd, "Can't allocate BPF LPM TRIE map, BPF firewalling is not supported: %m"); |
2ae7ee58 | 689 | return supported = BPF_FIREWALL_UNSUPPORTED; |
1988a9d1 DM |
690 | } |
691 | ||
692 | safe_close(fd); | |
693 | ||
4355f1c9 ZJS |
694 | r = bpf_program_new(BPF_PROG_TYPE_CGROUP_SKB, &program); |
695 | if (r < 0) { | |
93e93da5 | 696 | log_debug_errno(r, "Can't allocate CGROUP SKB BPF program, BPF firewalling is not supported: %m"); |
2ae7ee58 | 697 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 LP |
698 | } |
699 | ||
700 | r = bpf_program_add_instructions(program, trivial, ELEMENTSOF(trivial)); | |
701 | if (r < 0) { | |
702 | log_debug_errno(r, "Can't add trivial instructions to CGROUP SKB BPF program, BPF firewalling is not supported: %m"); | |
2ae7ee58 | 703 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 LP |
704 | } |
705 | ||
706 | r = bpf_program_load_kernel(program, NULL, 0); | |
707 | if (r < 0) { | |
708 | log_debug_errno(r, "Can't load kernel CGROUP SKB BPF program, BPF firewalling is not supported: %m"); | |
2ae7ee58 | 709 | return supported = BPF_FIREWALL_UNSUPPORTED; |
93e93da5 LP |
710 | } |
711 | ||
e583759b LP |
712 | /* Unfortunately the kernel allows us to create BPF_PROG_TYPE_CGROUP_SKB programs even when CONFIG_CGROUP_BPF |
713 | * is turned off at kernel compilation time. This sucks of course: why does it allow us to create a cgroup BPF | |
714 | * program if we can't do a thing with it later? | |
715 | * | |
047de7e1 | 716 | * We detect this case by issuing the BPF_PROG_DETACH bpf() call with invalid file descriptors: if |
e583759b LP |
717 | * CONFIG_CGROUP_BPF is turned off, then the call will fail early with EINVAL. If it is turned on the |
718 | * parameters are validated however, and that'll fail with EBADF then. */ | |
719 | ||
720 | attr = (union bpf_attr) { | |
721 | .attach_type = BPF_CGROUP_INET_EGRESS, | |
722 | .target_fd = -1, | |
723 | .attach_bpf_fd = -1, | |
724 | }; | |
725 | ||
047de7e1 | 726 | if (bpf(BPF_PROG_DETACH, &attr, sizeof(attr)) < 0) { |
2ae7ee58 | 727 | if (errno != EBADF) { |
047de7e1 | 728 | log_debug_errno(errno, "Didn't get EBADF from BPF_PROG_DETACH, BPF firewalling is not supported: %m"); |
2ae7ee58 LP |
729 | return supported = BPF_FIREWALL_UNSUPPORTED; |
730 | } | |
731 | ||
732 | /* YAY! */ | |
733 | } else { | |
047de7e1 | 734 | log_debug("Wut? Kernel accepted our invalid BPF_PROG_DETACH call? Something is weird, assuming BPF firewalling is broken and hence not supported."); |
2ae7ee58 LP |
735 | return supported = BPF_FIREWALL_UNSUPPORTED; |
736 | } | |
e583759b | 737 | |
2ae7ee58 | 738 | /* So now we know that the BPF program is generally available, let's see if BPF_F_ALLOW_MULTI is also supported |
047de7e1 AF |
739 | * (which was added in kernel 4.15). We use a similar logic as before, but this time we use the BPF_PROG_ATTACH |
740 | * bpf() call and the BPF_F_ALLOW_MULTI flags value. Since the flags are checked early in the system call we'll | |
741 | * get EINVAL if it's not supported, and EBADF as before if it is available. */ | |
e583759b | 742 | |
2ae7ee58 LP |
743 | attr = (union bpf_attr) { |
744 | .attach_type = BPF_CGROUP_INET_EGRESS, | |
745 | .target_fd = -1, | |
746 | .attach_bpf_fd = -1, | |
747 | .attach_flags = BPF_F_ALLOW_MULTI, | |
748 | }; | |
749 | ||
b1c05b98 | 750 | if (bpf(BPF_PROG_ATTACH, &attr, sizeof(attr)) < 0) { |
2ae7ee58 LP |
751 | if (errno == EBADF) { |
752 | log_debug_errno(errno, "Got EBADF when using BPF_F_ALLOW_MULTI, which indicates it is supported. Yay!"); | |
753 | return supported = BPF_FIREWALL_SUPPORTED_WITH_MULTI; | |
754 | } | |
755 | ||
756 | if (errno == EINVAL) | |
757 | log_debug_errno(errno, "Got EINVAL error when using BPF_F_ALLOW_MULTI, which indicates it's not supported."); | |
758 | else | |
759 | log_debug_errno(errno, "Got unexpected error when using BPF_F_ALLOW_MULTI, assuming it's not supported: %m"); | |
760 | ||
761 | return supported = BPF_FIREWALL_SUPPORTED; | |
762 | } else { | |
763 | 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."); | |
764 | return supported = BPF_FIREWALL_UNSUPPORTED; | |
765 | } | |
1988a9d1 | 766 | } |