src/core/bpf-firewall.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #include <arpa/inet.h>
   4 #include <assert.h>
   5 #include <errno.h>
   6 #include <fcntl.h>
   7 #include <linux/bpf_insn.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 "memory-util.h"
  24 #include "missing_syscall.h"
  25 #include "unit.h"
  26 #include "strv.h"
  27 #include "virt.h"
  28
  29 enum {
  30         MAP_KEY_PACKETS,
  31         MAP_KEY_BYTES,
  32 };
  33
  34 enum {
  35         ACCESS_ALLOWED = 1,
  36         ACCESS_DENIED  = 2,
  37 };
  38
  39 /* Compile instructions for one list of addresses, one direction and one specific verdict on matches. */
  40
  41 static int add_lookup_instructions(
  42                 BPFProgram *p,
  43                 int map_fd,
  44                 int protocol,
  45                 bool is_ingress,
  46                 int verdict) {
  47
  48         int r, addr_offset, addr_size;
  49
  50         assert(p);
  51         assert(map_fd >= 0);
  52
  53         switch (protocol) {
  54
  55         case ETH_P_IP:
  56                 addr_size = sizeof(uint32_t);
  57                 addr_offset = is_ingress ?
  58                         offsetof(struct iphdr, saddr) :
  59                         offsetof(struct iphdr, daddr);
  60                 break;
  61
  62         case ETH_P_IPV6:
  63                 addr_size = 4 * sizeof(uint32_t);
  64                 addr_offset = is_ingress ?
  65                         offsetof(struct ip6_hdr, ip6_src.s6_addr) :
  66                         offsetof(struct ip6_hdr, ip6_dst.s6_addr);
  67                 break;
  68
  69         default:
  70                 return -EAFNOSUPPORT;
  71         }
  72
  73         do {
  74                 /* Compare IPv4 with one word instruction (32bit) */
  75                 struct bpf_insn insn[] = {
  76                         /* If skb->protocol != ETH_P_IP, skip this whole block. The offset will be set later. */
  77                         BPF_JMP_IMM(BPF_JNE, BPF_REG_7, htobe16(protocol), 0),
  78
  79                         /*
  80                          * Call into BPF_FUNC_skb_load_bytes to load the dst/src IP address
  81                          *
  82                          * R1: Pointer to the skb
  83                          * R2: Data offset
  84                          * R3: Destination buffer on the stack (r10 - 4)
  85                          * R4: Number of bytes to read (4)
  86                          */
  87
  88                         BPF_MOV64_REG(BPF_REG_1, BPF_REG_6),
  89                         BPF_MOV32_IMM(BPF_REG_2, addr_offset),
  90
  91                         BPF_MOV64_REG(BPF_REG_3, BPF_REG_10),
  92                         BPF_ALU64_IMM(BPF_ADD, BPF_REG_3, -addr_size),
  93
  94                         BPF_MOV32_IMM(BPF_REG_4, addr_size),
  95                         BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_skb_load_bytes),
  96
  97                         /*
  98                          * Call into BPF_FUNC_map_lookup_elem to see if the address matches any entry in the
  99                          * LPM trie map. For this to work, the prefixlen field of 'struct bpf_lpm_trie_key'
 100                          * has to be set to the maximum possible value.
 101                          *
 102                          * On success, the looked up value is stored in R0. For this application, the actual
 103                          * value doesn't matter, however; we just set the bit in @verdict in R8 if we found any
 104                          * matching value.
 105                          */
 106
 107                         BPF_LD_MAP_FD(BPF_REG_1, map_fd),
 108                         BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),
 109                         BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -addr_size - sizeof(uint32_t)),
 110                         BPF_ST_MEM(BPF_W, BPF_REG_2, 0, addr_size * 8),
 111
 112                         BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
 113                         BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 1),
 114                         BPF_ALU32_IMM(BPF_OR, BPF_REG_8, verdict),
 115                 };
 116
 117                 /* Jump label fixup */
 118                 insn[0].off = ELEMENTSOF(insn) - 1;
 119
 120                 r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn));
 121                 if (r < 0)
 122                         return r;
 123
 124         } while (false);
 125
 126         return 0;
 127 }
 128
 129 static int add_instructions_for_ip_any(
 130                 BPFProgram *p,
 131                 int verdict) {
 132         int r;
 133
 134         assert(p);
 135
 136         struct bpf_insn insn[] = {
 137                 BPF_ALU32_IMM(BPF_OR, BPF_REG_8, verdict),
 138         };
 139
 140         r = bpf_program_add_instructions(p, insn, 1);
 141         if (r < 0)
 142                 return r;
 143
 144         return 0;
 145 }
 146
 147 static int bpf_firewall_compile_bpf(
 148                 Unit *u,
 149                 bool is_ingress,
 150                 BPFProgram **ret,
 151                 bool ip_allow_any,
 152                 bool ip_deny_any) {
 153
 154         struct bpf_insn pre_insn[] = {
 155                 /*
 156                  * When the eBPF program is entered, R1 contains the address of the skb.
 157                  * However, R1-R5 are scratch registers that are not preserved when calling
 158                  * into kernel functions, so we need to save anything that's supposed to
 159                  * stay around to R6-R9. Save the skb to R6.
 160                  */
 161                 BPF_MOV64_REG(BPF_REG_6, BPF_REG_1),
 162
 163                 /*
 164                  * Although we cannot access the skb data directly from eBPF programs used in this
 165                  * scenario, the kernel has prepared some fields for us to access through struct __sk_buff.
 166                  * Load the protocol (IPv4, IPv6) used by the packet in flight once and cache it in R7
 167                  * for later use.
 168                  */
 169                 BPF_LDX_MEM(BPF_W, BPF_REG_7, BPF_REG_6, offsetof(struct __sk_buff, protocol)),
 170
 171                 /*
 172                  * R8 is used to keep track of whether any address check has explicitly allowed or denied the packet
 173                  * through ACCESS_DENIED or ACCESS_ALLOWED bits. Reset them both to 0 in the beginning.
 174                  */
 175                 BPF_MOV32_IMM(BPF_REG_8, 0),
 176         };
 177
 178         /*
 179          * The access checkers compiled for the configured allowance and denial lists
 180          * write to R8 at runtime. The following code prepares for an early exit that
 181          * skip the accounting if the packet is denied.
 182          *
 183          * R0 = 1
 184          * if (R8 == ACCESS_DENIED)
 185          *     R0 = 0
 186          *
 187          * This means that if both ACCESS_DENIED and ACCESS_ALLOWED are set, the packet
 188          * is allowed to pass.
 189          */
 190         struct bpf_insn post_insn[] = {
 191                 BPF_MOV64_IMM(BPF_REG_0, 1),
 192                 BPF_JMP_IMM(BPF_JNE, BPF_REG_8, ACCESS_DENIED, 1),
 193                 BPF_MOV64_IMM(BPF_REG_0, 0),
 194         };
 195
 196         _cleanup_(bpf_program_unrefp) BPFProgram *p = NULL;
 197         int accounting_map_fd, r;
 198         bool access_enabled;
 199
 200         assert(u);
 201         assert(ret);
 202
 203         accounting_map_fd = is_ingress ?
 204                 u->ip_accounting_ingress_map_fd :
 205                 u->ip_accounting_egress_map_fd;
 206
 207         access_enabled =
 208                 u->ipv4_allow_map_fd >= 0 ||
 209                 u->ipv6_allow_map_fd >= 0 ||
 210                 u->ipv4_deny_map_fd >= 0 ||
 211                 u->ipv6_deny_map_fd >= 0 ||
 212                 ip_allow_any ||
 213                 ip_deny_any;
 214
 215         if (accounting_map_fd < 0 && !access_enabled) {
 216                 *ret = NULL;
 217                 return 0;
 218         }
 219
 220         r = bpf_program_new(BPF_PROG_TYPE_CGROUP_SKB, &p);
 221         if (r < 0)
 222                 return r;
 223
 224         r = bpf_program_add_instructions(p, pre_insn, ELEMENTSOF(pre_insn));
 225         if (r < 0)
 226                 return r;
 227
 228         if (access_enabled) {
 229                 /*
 230                  * The simple rule this function translates into eBPF instructions is:
 231                  *
 232                  * - Access will be granted when an address matches an entry in @list_allow
 233                  * - Otherwise, access will be denied when an address matches an entry in @list_deny
 234                  * - Otherwise, access will be granted
 235                  */
 236
 237                 if (u->ipv4_deny_map_fd >= 0) {
 238                         r = add_lookup_instructions(p, u->ipv4_deny_map_fd, ETH_P_IP, is_ingress, ACCESS_DENIED);
 239                         if (r < 0)
 240                                 return r;
 241                 }
 242
 243                 if (u->ipv6_deny_map_fd >= 0) {
 244                         r = add_lookup_instructions(p, u->ipv6_deny_map_fd, ETH_P_IPV6, is_ingress, ACCESS_DENIED);
 245                         if (r < 0)
 246                                 return r;
 247                 }
 248
 249                 if (u->ipv4_allow_map_fd >= 0) {
 250                         r = add_lookup_instructions(p, u->ipv4_allow_map_fd, ETH_P_IP, is_ingress, ACCESS_ALLOWED);
 251                         if (r < 0)
 252                                 return r;
 253                 }
 254
 255                 if (u->ipv6_allow_map_fd >= 0) {
 256                         r = add_lookup_instructions(p, u->ipv6_allow_map_fd, ETH_P_IPV6, is_ingress, ACCESS_ALLOWED);
 257                         if (r < 0)
 258                                 return r;
 259                 }
 260
 261                 if (ip_allow_any) {
 262                         r = add_instructions_for_ip_any(p, ACCESS_ALLOWED);
 263                         if (r < 0)
 264                                 return r;
 265                 }
 266
 267                 if (ip_deny_any) {
 268                         r = add_instructions_for_ip_any(p, ACCESS_DENIED);
 269                         if (r < 0)
 270                                 return r;
 271                 }
 272         }
 273
 274         r = bpf_program_add_instructions(p, post_insn, ELEMENTSOF(post_insn));
 275         if (r < 0)
 276                 return r;
 277
 278         if (accounting_map_fd >= 0) {
 279                 struct bpf_insn insn[] = {
 280                         /*
 281                          * If R0 == 0, the packet will be denied; skip the accounting instructions in this case.
 282                          * The jump label will be fixed up later.
 283                          */
 284                         BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 0),
 285
 286                         /* Count packets */
 287                         BPF_MOV64_IMM(BPF_REG_0, MAP_KEY_PACKETS), /* r0 = 0 */
 288                         BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_0, -4), /* *(u32 *)(fp - 4) = r0 */
 289                         BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),
 290                         BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), /* r2 = fp - 4 */
 291                         BPF_LD_MAP_FD(BPF_REG_1, accounting_map_fd), /* load map fd to r1 */
 292                         BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
 293                         BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2),
 294                         BPF_MOV64_IMM(BPF_REG_1, 1), /* r1 = 1 */
 295                         BPF_RAW_INSN(BPF_STX | BPF_XADD | BPF_DW, BPF_REG_0, BPF_REG_1, 0, 0), /* xadd r0 += r1 */
 296
 297                         /* Count bytes */
 298                         BPF_MOV64_IMM(BPF_REG_0, MAP_KEY_BYTES), /* r0 = 1 */
 299                         BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_0, -4), /* *(u32 *)(fp - 4) = r0 */
 300                         BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),
 301                         BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), /* r2 = fp - 4 */
 302                         BPF_LD_MAP_FD(BPF_REG_1, accounting_map_fd),
 303                         BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
 304                         BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2),
 305                         BPF_LDX_MEM(BPF_W, BPF_REG_1, BPF_REG_6, offsetof(struct __sk_buff, len)), /* r1 = skb->len */
 306                         BPF_RAW_INSN(BPF_STX | BPF_XADD | BPF_DW, BPF_REG_0, BPF_REG_1, 0, 0), /* xadd r0 += r1 */
 307
 308                         /* Allow the packet to pass */
 309                         BPF_MOV64_IMM(BPF_REG_0, 1),
 310                 };
 311
 312                 /* Jump label fixup */
 313                 insn[0].off = ELEMENTSOF(insn) - 1;
 314
 315                 r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn));
 316                 if (r < 0)
 317                         return r;
 318         }
 319
 320         do {
 321                 /*
 322                  * Exit from the eBPF program, R0 contains the verdict.
 323                  * 0 means the packet is denied, 1 means the packet may pass.
 324                  */
 325                 struct bpf_insn insn[] = {
 326                         BPF_EXIT_INSN()
 327                 };
 328
 329                 r = bpf_program_add_instructions(p, insn, ELEMENTSOF(insn));
 330                 if (r < 0)
 331                         return r;
 332         } while (false);
 333
 334         *ret = TAKE_PTR(p);
 335
 336         return 0;
 337 }
 338
 339 static int bpf_firewall_count_access_items(IPAddressAccessItem *list, size_t *n_ipv4, size_t *n_ipv6) {
 340         IPAddressAccessItem *a;
 341
 342         assert(n_ipv4);
 343         assert(n_ipv6);
 344
 345         LIST_FOREACH(items, a, list) {
 346                 switch (a->family) {
 347
 348                 case AF_INET:
 349                         (*n_ipv4)++;
 350                         break;
 351
 352                 case AF_INET6:
 353                         (*n_ipv6)++;
 354                         break;
 355
 356                 default:
 357                         return -EAFNOSUPPORT;
 358                 }
 359         }
 360
 361         return 0;
 362 }
 363
 364 static int bpf_firewall_add_access_items(
 365                 IPAddressAccessItem *list,
 366                 int ipv4_map_fd,
 367                 int ipv6_map_fd,
 368                 int verdict) {
 369
 370         struct bpf_lpm_trie_key *key_ipv4, *key_ipv6;
 371         uint64_t value = verdict;
 372         IPAddressAccessItem *a;
 373         int r;
 374
 375         key_ipv4 = alloca0(offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t));
 376         key_ipv6 = alloca0(offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t) * 4);
 377
 378         LIST_FOREACH(items, a, list) {
 379                 switch (a->family) {
 380
 381                 case AF_INET:
 382                         key_ipv4->prefixlen = a->prefixlen;
 383                         memcpy(key_ipv4->data, &a->address, sizeof(uint32_t));
 384
 385                         r = bpf_map_update_element(ipv4_map_fd, key_ipv4, &value);
 386                         if (r < 0)
 387                                 return r;
 388
 389                         break;
 390
 391                 case AF_INET6:
 392                         key_ipv6->prefixlen = a->prefixlen;
 393                         memcpy(key_ipv6->data, &a->address, 4 * sizeof(uint32_t));
 394
 395                         r = bpf_map_update_element(ipv6_map_fd, key_ipv6, &value);
 396                         if (r < 0)
 397                                 return r;
 398
 399                         break;
 400
 401                 default:
 402                         return -EAFNOSUPPORT;
 403                 }
 404         }
 405
 406         return 0;
 407 }
 408
 409 static int bpf_firewall_prepare_access_maps(
 410                 Unit *u,
 411                 int verdict,
 412                 int *ret_ipv4_map_fd,
 413                 int *ret_ipv6_map_fd,
 414                 bool *ret_has_any) {
 415
 416         _cleanup_close_ int ipv4_map_fd = -1, ipv6_map_fd = -1;
 417         size_t n_ipv4 = 0, n_ipv6 = 0;
 418         IPAddressAccessItem *list;
 419         Unit *p;
 420         int r;
 421
 422         assert(ret_ipv4_map_fd);
 423         assert(ret_ipv6_map_fd);
 424         assert(ret_has_any);
 425
 426         for (p = u; p; p = UNIT_DEREF(p->slice)) {
 427                 CGroupContext *cc;
 428
 429                 cc = unit_get_cgroup_context(p);
 430                 if (!cc)
 431                         continue;
 432
 433                 list = verdict == ACCESS_ALLOWED ? cc->ip_address_allow : cc->ip_address_deny;
 434
 435                 bpf_firewall_count_access_items(list, &n_ipv4, &n_ipv6);
 436
 437                 /* Skip making the LPM trie map in cases where we are using "any" in order to hack around
 438                  * needing CAP_SYS_ADMIN for allocating LPM trie map. */
 439                 if (ip_address_access_item_is_any(list)) {
 440                         *ret_has_any = true;
 441                         return 0;
 442                 }
 443         }
 444
 445         if (n_ipv4 > 0) {
 446                 ipv4_map_fd = bpf_map_new(
 447                                 BPF_MAP_TYPE_LPM_TRIE,
 448                                 offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t),
 449                                 sizeof(uint64_t),
 450                                 n_ipv4,
 451                                 BPF_F_NO_PREALLOC);
 452                 if (ipv4_map_fd < 0)
 453                         return ipv4_map_fd;
 454         }
 455
 456         if (n_ipv6 > 0) {
 457                 ipv6_map_fd = bpf_map_new(
 458                                 BPF_MAP_TYPE_LPM_TRIE,
 459                                 offsetof(struct bpf_lpm_trie_key, data) + sizeof(uint32_t)*4,
 460                                 sizeof(uint64_t),
 461                                 n_ipv6,
 462                                 BPF_F_NO_PREALLOC);
 463                 if (ipv6_map_fd < 0)
 464                         return ipv6_map_fd;
 465         }
 466
 467         for (p = u; p; p = UNIT_DEREF(p->slice)) {
 468                 CGroupContext *cc;
 469
 470                 cc = unit_get_cgroup_context(p);
 471                 if (!cc)
 472                         continue;
 473
 474                 r = bpf_firewall_add_access_items(verdict == ACCESS_ALLOWED ? cc->ip_address_allow : cc->ip_address_deny,
 475                                                   ipv4_map_fd, ipv6_map_fd, verdict);
 476                 if (r < 0)
 477                         return r;
 478         }
 479
 480         *ret_ipv4_map_fd = TAKE_FD(ipv4_map_fd);
 481         *ret_ipv6_map_fd = TAKE_FD(ipv6_map_fd);
 482         *ret_has_any = false;
 483         return 0;
 484 }
 485
 486 static int bpf_firewall_prepare_accounting_maps(Unit *u, bool enabled, int *fd_ingress, int *fd_egress) {
 487         int r;
 488
 489         assert(u);
 490         assert(fd_ingress);
 491         assert(fd_egress);
 492
 493         if (enabled) {
 494                 if (*fd_ingress < 0) {
 495                         r = bpf_map_new(BPF_MAP_TYPE_ARRAY, sizeof(int), sizeof(uint64_t), 2, 0);
 496                         if (r < 0)
 497                                 return r;
 498
 499                         *fd_ingress = r;
 500                 }
 501
 502                 if (*fd_egress < 0) {
 503
 504                         r = bpf_map_new(BPF_MAP_TYPE_ARRAY, sizeof(int), sizeof(uint64_t), 2, 0);
 505                         if (r < 0)
 506                                 return r;
 507
 508                         *fd_egress = r;
 509                 }
 510
 511         } else {
 512                 *fd_ingress = safe_close(*fd_ingress);
 513                 *fd_egress = safe_close(*fd_egress);
 514
 515                 zero(u->ip_accounting_extra);
 516         }
 517
 518         return 0;
 519 }
 520
 521 int bpf_firewall_compile(Unit *u) {
 522         CGroupContext *cc;
 523         int r, supported;
 524         bool ip_allow_any = false, ip_deny_any = false;
 525
 526         assert(u);
 527
 528         cc = unit_get_cgroup_context(u);
 529         if (!cc)
 530                 return -EINVAL;
 531
 532         supported = bpf_firewall_supported();
 533         if (supported < 0)
 534                 return supported;
 535         if (supported == BPF_FIREWALL_UNSUPPORTED)
 536                 return log_unit_debug_errno(u, SYNTHETIC_ERRNO(EOPNOTSUPP),
 537                                             "BPF firewalling not supported on this manager, proceeding without.");
 538         if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI && u->type == UNIT_SLICE)
 539                 /* If BPF_F_ALLOW_MULTI is not supported we don't support any BPF magic on inner nodes (i.e. on slice
 540                  * units), since that would mean leaf nodes couldn't do any BPF anymore at all. Under the assumption
 541                  * that BPF is more interesting on leaf nodes we hence avoid it on inner nodes in that case. This is
 542                  * consistent with old systemd behaviour from before v238, where BPF wasn't supported in inner nodes at
 543                  * all, either. */
 544                 return log_unit_debug_errno(u, SYNTHETIC_ERRNO(EOPNOTSUPP),
 545                                             "BPF_F_ALLOW_MULTI is not supported on this manager, not doing BPF firewall on slice units.");
 546
 547         /* Note that when we compile a new firewall we first flush out the access maps and the BPF programs themselves,
 548          * but we reuse the the accounting maps. That way the firewall in effect always maps to the actual
 549          * configuration, but we don't flush out the accounting unnecessarily */
 550
 551         u->ip_bpf_ingress = bpf_program_unref(u->ip_bpf_ingress);
 552         u->ip_bpf_egress = bpf_program_unref(u->ip_bpf_egress);
 553
 554         u->ipv4_allow_map_fd = safe_close(u->ipv4_allow_map_fd);
 555         u->ipv4_deny_map_fd = safe_close(u->ipv4_deny_map_fd);
 556
 557         u->ipv6_allow_map_fd = safe_close(u->ipv6_allow_map_fd);
 558         u->ipv6_deny_map_fd = safe_close(u->ipv6_deny_map_fd);
 559
 560         if (u->type != UNIT_SLICE) {
 561                 /* In inner nodes we only do accounting, we do not actually bother with access control. However, leaf
 562                  * nodes will incorporate all IP access rules set on all their parent nodes. This has the benefit that
 563                  * they can optionally cancel out system-wide rules. Since inner nodes can't contain processes this
 564                  * means that all configure IP access rules *will* take effect on processes, even though we never
 565                  * compile them for inner nodes. */
 566
 567                 r = bpf_firewall_prepare_access_maps(u, ACCESS_ALLOWED, &u->ipv4_allow_map_fd, &u->ipv6_allow_map_fd, &ip_allow_any);
 568                 if (r < 0)
 569                         return log_unit_error_errno(u, r, "Preparation of eBPF allow maps failed: %m");
 570
 571                 r = bpf_firewall_prepare_access_maps(u, ACCESS_DENIED, &u->ipv4_deny_map_fd, &u->ipv6_deny_map_fd, &ip_deny_any);
 572                 if (r < 0)
 573                         return log_unit_error_errno(u, r, "Preparation of eBPF deny maps failed: %m");
 574         }
 575
 576         r = bpf_firewall_prepare_accounting_maps(u, cc->ip_accounting, &u->ip_accounting_ingress_map_fd, &u->ip_accounting_egress_map_fd);
 577         if (r < 0)
 578                 return log_unit_error_errno(u, r, "Preparation of eBPF accounting maps failed: %m");
 579
 580         r = bpf_firewall_compile_bpf(u, true, &u->ip_bpf_ingress, ip_allow_any, ip_deny_any);
 581         if (r < 0)
 582                 return log_unit_error_errno(u, r, "Compilation for ingress BPF program failed: %m");
 583
 584         r = bpf_firewall_compile_bpf(u, false, &u->ip_bpf_egress, ip_allow_any, ip_deny_any);
 585         if (r < 0)
 586                 return log_unit_error_errno(u, r, "Compilation for egress BPF program failed: %m");
 587
 588         return 0;
 589 }
 590
 591 DEFINE_PRIVATE_HASH_OPS_WITH_VALUE_DESTRUCTOR(filter_prog_hash_ops, void, trivial_hash_func, trivial_compare_func, BPFProgram, bpf_program_unref);
 592
 593 static int load_bpf_progs_from_fs_to_set(Unit *u, char **filter_paths, Set **set) {
 594         char **bpf_fs_path;
 595
 596         set_clear(*set);
 597
 598         STRV_FOREACH(bpf_fs_path, filter_paths) {
 599                 _cleanup_free_ BPFProgram *prog = NULL;
 600                 int r;
 601
 602                 r = bpf_program_new(BPF_PROG_TYPE_CGROUP_SKB, &prog);
 603                 if (r < 0)
 604                         return log_unit_error_errno(u, r, "Can't allocate CGROUP SKB BPF program: %m");
 605
 606                 r = bpf_program_load_from_bpf_fs(prog, *bpf_fs_path);
 607                 if (r < 0)
 608                         return log_unit_error_errno(u, r, "Loading of ingress BPF program %s failed: %m", *bpf_fs_path);
 609
 610                 r = set_ensure_allocated(set, &filter_prog_hash_ops);
 611                 if (r < 0)
 612                         return log_unit_error_errno(u, r, "Can't allocate BPF program set: %m");
 613
 614                 r = set_put(*set, prog);
 615                 if (r < 0)
 616                         return log_unit_error_errno(u, r, "Can't add program to BPF program set: %m");
 617                 TAKE_PTR(prog);
 618         }
 619
 620         return 0;
 621 }
 622
 623 int bpf_firewall_load_custom(Unit *u) {
 624         CGroupContext *cc;
 625         int r, supported;
 626
 627         assert(u);
 628
 629         cc = unit_get_cgroup_context(u);
 630         if (!cc)
 631                 return 0;
 632
 633         if (!(cc->ip_filters_ingress || cc->ip_filters_egress))
 634                 return 0;
 635
 636         supported = bpf_firewall_supported();
 637         if (supported < 0)
 638                 return supported;
 639
 640         if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI)
 641                 return log_unit_debug_errno(u, SYNTHETIC_ERRNO(EOPNOTSUPP), "BPF_F_ALLOW_MULTI not supported on this manager, cannot attach custom BPF programs.");
 642
 643         r = load_bpf_progs_from_fs_to_set(u, cc->ip_filters_ingress, &u->ip_bpf_custom_ingress);
 644         if (r < 0)
 645                 return r;
 646         r = load_bpf_progs_from_fs_to_set(u, cc->ip_filters_egress, &u->ip_bpf_custom_egress);
 647         if (r < 0)
 648                 return r;
 649
 650         return 0;
 651 }
 652
 653 static int attach_custom_bpf_progs(Unit *u, const char *path, int attach_type, Set **set, Set **set_installed) {
 654         BPFProgram *prog;
 655         Iterator i;
 656         int r;
 657
 658         assert(u);
 659
 660         set_clear(*set_installed);
 661
 662         SET_FOREACH(prog, *set, i) {
 663                 r = bpf_program_cgroup_attach(prog, attach_type, path, BPF_F_ALLOW_MULTI);
 664                 if (r < 0)
 665                         return log_unit_error_errno(u, r, "Attaching custom egress BPF program to cgroup %s failed: %m", path);
 666                 /* Remember that these BPF programs are installed now. */
 667                 r = set_ensure_allocated(set_installed, &filter_prog_hash_ops);
 668                 if (r < 0)
 669                         return log_unit_error_errno(u, r, "Can't allocate BPF program set: %m");
 670
 671                 r = set_put(*set_installed, prog);
 672                 if (r < 0)
 673                         return log_unit_error_errno(u, r, "Can't add program to BPF program set: %m");
 674                 bpf_program_ref(prog);
 675         }
 676
 677         return 0;
 678 }
 679
 680 int bpf_firewall_install(Unit *u) {
 681         _cleanup_free_ char *path = NULL;
 682         CGroupContext *cc;
 683         int r, supported;
 684         uint32_t flags;
 685
 686         assert(u);
 687
 688         cc = unit_get_cgroup_context(u);
 689         if (!cc)
 690                 return -EINVAL;
 691         if (!u->cgroup_path)
 692                 return -EINVAL;
 693         if (!u->cgroup_realized)
 694                 return -EINVAL;
 695
 696         supported = bpf_firewall_supported();
 697         if (supported < 0)
 698                 return supported;
 699         if (supported == BPF_FIREWALL_UNSUPPORTED) {
 700                 log_unit_debug(u, "BPF firewalling not supported on this manager, proceeding without.");
 701                 return -EOPNOTSUPP;
 702         }
 703         if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI && u->type == UNIT_SLICE) {
 704                 log_unit_debug(u, "BPF_F_ALLOW_MULTI is not supported on this manager, not doing BPF firewall on slice units.");
 705                 return -EOPNOTSUPP;
 706         }
 707         if (supported != BPF_FIREWALL_SUPPORTED_WITH_MULTI &&
 708             (!set_isempty(u->ip_bpf_custom_ingress) || !set_isempty(u->ip_bpf_custom_egress)))
 709                 return log_unit_debug_errno(u, SYNTHETIC_ERRNO(EOPNOTSUPP), "BPF_F_ALLOW_MULTI not supported on this manager, cannot attach custom BPF programs.");
 710
 711         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, NULL, &path);
 712         if (r < 0)
 713                 return log_unit_error_errno(u, r, "Failed to determine cgroup path: %m");
 714
 715         flags = (supported == BPF_FIREWALL_SUPPORTED_WITH_MULTI &&
 716                  (u->type == UNIT_SLICE || unit_cgroup_delegate(u))) ? BPF_F_ALLOW_MULTI : 0;
 717
 718         /* Unref the old BPF program (which will implicitly detach it) right before attaching the new program, to
 719          * minimize the time window when we don't account for IP traffic. */
 720         u->ip_bpf_egress_installed = bpf_program_unref(u->ip_bpf_egress_installed);
 721         u->ip_bpf_ingress_installed = bpf_program_unref(u->ip_bpf_ingress_installed);
 722
 723         if (u->ip_bpf_egress) {
 724                 r = bpf_program_cgroup_attach(u->ip_bpf_egress, BPF_CGROUP_INET_EGRESS, path,
 725                                               flags | (set_isempty(u->ip_bpf_custom_egress) ? 0 : BPF_F_ALLOW_MULTI));
 726                 if (r < 0)
 727                         return log_unit_error_errno(u, r, "Attaching egress BPF program to cgroup %s failed: %m", path);
 728
 729                 /* Remember that this BPF program is installed now. */
 730                 u->ip_bpf_egress_installed = bpf_program_ref(u->ip_bpf_egress);
 731         }
 732
 733         if (u->ip_bpf_ingress) {
 734                 r = bpf_program_cgroup_attach(u->ip_bpf_ingress, BPF_CGROUP_INET_INGRESS, path,
 735                                               flags | (set_isempty(u->ip_bpf_custom_ingress) ? 0 : BPF_F_ALLOW_MULTI));
 736                 if (r < 0)
 737                         return log_unit_error_errno(u, r, "Attaching ingress BPF program to cgroup %s failed: %m", path);
 738
 739                 u->ip_bpf_ingress_installed = bpf_program_ref(u->ip_bpf_ingress);
 740         }
 741
 742         r = attach_custom_bpf_progs(u, path, BPF_CGROUP_INET_EGRESS, &u->ip_bpf_custom_egress, &u->ip_bpf_custom_egress_installed);
 743         if (r < 0)
 744                 return r;
 745
 746         r = attach_custom_bpf_progs(u, path, BPF_CGROUP_INET_INGRESS, &u->ip_bpf_custom_ingress, &u->ip_bpf_custom_ingress_installed);
 747         if (r < 0)
 748                 return r;
 749
 750         return 0;
 751 }
 752
 753 int bpf_firewall_read_accounting(int map_fd, uint64_t *ret_bytes, uint64_t *ret_packets) {
 754         uint64_t key, packets;
 755         int r;
 756
 757         if (map_fd < 0)
 758                 return -EBADF;
 759
 760         if (ret_packets) {
 761                 key = MAP_KEY_PACKETS;
 762                 r = bpf_map_lookup_element(map_fd, &key, &packets);
 763                 if (r < 0)
 764                         return r;
 765         }
 766
 767         if (ret_bytes) {
 768                 key = MAP_KEY_BYTES;
 769                 r = bpf_map_lookup_element(map_fd, &key, ret_bytes);
 770                 if (r < 0)
 771                         return r;
 772         }
 773
 774         if (ret_packets)
 775                 *ret_packets = packets;
 776
 777         return 0;
 778 }
 779
 780 int bpf_firewall_reset_accounting(int map_fd) {
 781         uint64_t key, value = 0;
 782         int r;
 783
 784         if (map_fd < 0)
 785                 return -EBADF;
 786
 787         key = MAP_KEY_PACKETS;
 788         r = bpf_map_update_element(map_fd, &key, &value);
 789         if (r < 0)
 790                 return r;
 791
 792         key = MAP_KEY_BYTES;
 793         return bpf_map_update_element(map_fd, &key, &value);
 794 }
 795
 796 static int bpf_firewall_unsupported_reason = 0;
 797
 798 int bpf_firewall_supported(void) {
 799         struct bpf_insn trivial[] = {
 800                 BPF_MOV64_IMM(BPF_REG_0, 1),
 801                 BPF_EXIT_INSN()
 802         };
 803
 804         _cleanup_(bpf_program_unrefp) BPFProgram *program = NULL;
 805         static int supported = -1;
 806         union bpf_attr attr;
 807         int r;
 808
 809         /* Checks whether BPF firewalling is supported. For this, we check the following things:
 810          *
 811          * - whether the unified hierarchy is being used
 812          * - the BPF implementation in the kernel supports BPF_PROG_TYPE_CGROUP_SKB programs, which we require
 813          * - the BPF implementation in the kernel supports the BPF_PROG_DETACH call, which we require
 814          */
 815         if (supported >= 0)
 816                 return supported;
 817
 818         r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
 819         if (r < 0)
 820                 return log_error_errno(r, "Can't determine whether the unified hierarchy is used: %m");
 821         if (r == 0) {
 822                 bpf_firewall_unsupported_reason =
 823                         log_debug_errno(SYNTHETIC_ERRNO(EUCLEAN),
 824                                         "Not running with unified cgroups, BPF firewalling is not supported.");
 825                 return supported = BPF_FIREWALL_UNSUPPORTED;
 826         }
 827
 828         r = bpf_program_new(BPF_PROG_TYPE_CGROUP_SKB, &program);
 829         if (r < 0) {
 830                 bpf_firewall_unsupported_reason =
 831                         log_debug_errno(r, "Can't allocate CGROUP SKB BPF program, BPF firewalling is not supported: %m");
 832                 return supported = BPF_FIREWALL_UNSUPPORTED;
 833         }
 834
 835         r = bpf_program_add_instructions(program, trivial, ELEMENTSOF(trivial));
 836         if (r < 0) {
 837                 bpf_firewall_unsupported_reason =
 838                         log_debug_errno(r, "Can't add trivial instructions to CGROUP SKB BPF program, BPF firewalling is not supported: %m");
 839                 return supported = BPF_FIREWALL_UNSUPPORTED;
 840         }
 841
 842         r = bpf_program_load_kernel(program, NULL, 0);
 843         if (r < 0) {
 844                 bpf_firewall_unsupported_reason =
 845                         log_debug_errno(r, "Can't load kernel CGROUP SKB BPF program, BPF firewalling is not supported: %m");
 846                 return supported = BPF_FIREWALL_UNSUPPORTED;
 847         }
 848
 849         /* Unfortunately the kernel allows us to create BPF_PROG_TYPE_CGROUP_SKB programs even when CONFIG_CGROUP_BPF
 850          * is turned off at kernel compilation time. This sucks of course: why does it allow us to create a cgroup BPF
 851          * program if we can't do a thing with it later?
 852          *
 853          * We detect this case by issuing the BPF_PROG_DETACH bpf() call with invalid file descriptors: if
 854          * CONFIG_CGROUP_BPF is turned off, then the call will fail early with EINVAL. If it is turned on the
 855          * parameters are validated however, and that'll fail with EBADF then. */
 856
 857         attr = (union bpf_attr) {
 858                 .attach_type = BPF_CGROUP_INET_EGRESS,
 859                 .target_fd = -1,
 860                 .attach_bpf_fd = -1,
 861         };
 862
 863         if (bpf(BPF_PROG_DETACH, &attr, sizeof(attr)) < 0) {
 864                 if (errno != EBADF) {
 865                         bpf_firewall_unsupported_reason =
 866                                 log_debug_errno(errno, "Didn't get EBADF from BPF_PROG_DETACH, BPF firewalling is not supported: %m");
 867                         return supported = BPF_FIREWALL_UNSUPPORTED;
 868                 }
 869
 870                 /* YAY! */
 871         } else {
 872                 log_debug("Wut? Kernel accepted our invalid BPF_PROG_DETACH call? Something is weird, assuming BPF firewalling is broken and hence not supported.");
 873                 return supported = BPF_FIREWALL_UNSUPPORTED;
 874         }
 875
 876         /* So now we know that the BPF program is generally available, let's see if BPF_F_ALLOW_MULTI is also supported
 877          * (which was added in kernel 4.15). We use a similar logic as before, but this time we use the BPF_PROG_ATTACH
 878          * bpf() call and the BPF_F_ALLOW_MULTI flags value. Since the flags are checked early in the system call we'll
 879          * get EINVAL if it's not supported, and EBADF as before if it is available. */
 880
 881         attr = (union bpf_attr) {
 882                 .attach_type = BPF_CGROUP_INET_EGRESS,
 883                 .target_fd = -1,
 884                 .attach_bpf_fd = -1,
 885                 .attach_flags = BPF_F_ALLOW_MULTI,
 886         };
 887
 888         if (bpf(BPF_PROG_ATTACH, &attr, sizeof(attr)) < 0) {
 889                 if (errno == EBADF) {
 890                         log_debug_errno(errno, "Got EBADF when using BPF_F_ALLOW_MULTI, which indicates it is supported. Yay!");
 891                         return supported = BPF_FIREWALL_SUPPORTED_WITH_MULTI;
 892                 }
 893
 894                 if (errno == EINVAL)
 895                         log_debug_errno(errno, "Got EINVAL error when using BPF_F_ALLOW_MULTI, which indicates it's not supported.");
 896                 else
 897                         log_debug_errno(errno, "Got unexpected error when using BPF_F_ALLOW_MULTI, assuming it's not supported: %m");
 898
 899                 return supported = BPF_FIREWALL_SUPPORTED;
 900         } else {
 901                 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.");
 902                 return supported = BPF_FIREWALL_UNSUPPORTED;
 903         }
 904 }
 905
 906 void emit_bpf_firewall_warning(Unit *u) {
 907         static bool warned = false;
 908
 909         if (!warned) {
 910                 bool quiet = bpf_firewall_unsupported_reason == -EPERM && detect_container();
 911
 912                 log_unit_full(u, quiet ? LOG_DEBUG : LOG_WARNING, bpf_firewall_unsupported_reason,
 913                               "unit configures an IP firewall, but %s.\n"
 914                               "(This warning is only shown for the first unit using IP firewalling.)",
 915                               getuid() != 0 ? "not running as root" :
 916                                               "the local system does not support BPF/cgroup firewalling");
 917                 warned = true;
 918         }
 919 }