src/libsystemd/sd-bus/bus-socket.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #include <endian.h>
   4 #include <poll.h>
   5 #include <stdlib.h>
   6 #include <unistd.h>
   7
   8 #include "sd-bus.h"
   9 #include "sd-daemon.h"
  10
  11 #include "alloc-util.h"
  12 #include "bus-internal.h"
  13 #include "bus-message.h"
  14 #include "bus-socket.h"
  15 #include "fd-util.h"
  16 #include "format-util.h"
  17 #include "fs-util.h"
  18 #include "hexdecoct.h"
  19 #include "io-util.h"
  20 #include "macro.h"
  21 #include "missing.h"
  22 #include "path-util.h"
  23 #include "process-util.h"
  24 #include "selinux-util.h"
  25 #include "signal-util.h"
  26 #include "stdio-util.h"
  27 #include "string-util.h"
  28 #include "user-util.h"
  29 #include "utf8.h"
  30 #include "util.h"
  31
  32 #define SNDBUF_SIZE (8*1024*1024)
  33
  34 static void iovec_advance(struct iovec iov[], unsigned *idx, size_t size) {
  35
  36         while (size > 0) {
  37                 struct iovec *i = iov + *idx;
  38
  39                 if (i->iov_len > size) {
  40                         i->iov_base = (uint8_t*) i->iov_base + size;
  41                         i->iov_len -= size;
  42                         return;
  43                 }
  44
  45                 size -= i->iov_len;
  46
  47                 i->iov_base = NULL;
  48                 i->iov_len = 0;
  49
  50                 (*idx)++;
  51         }
  52 }
  53
  54 static int append_iovec(sd_bus_message *m, const void *p, size_t sz) {
  55         assert(m);
  56         assert(p);
  57         assert(sz > 0);
  58
  59         m->iovec[m->n_iovec].iov_base = (void*) p;
  60         m->iovec[m->n_iovec].iov_len = sz;
  61         m->n_iovec++;
  62
  63         return 0;
  64 }
  65
  66 static int bus_message_setup_iovec(sd_bus_message *m) {
  67         struct bus_body_part *part;
  68         unsigned n, i;
  69         int r;
  70
  71         assert(m);
  72         assert(m->sealed);
  73
  74         if (m->n_iovec > 0)
  75                 return 0;
  76
  77         assert(!m->iovec);
  78
  79         n = 1 + m->n_body_parts;
  80         if (n < ELEMENTSOF(m->iovec_fixed))
  81                 m->iovec = m->iovec_fixed;
  82         else {
  83                 m->iovec = new(struct iovec, n);
  84                 if (!m->iovec) {
  85                         r = -ENOMEM;
  86                         goto fail;
  87                 }
  88         }
  89
  90         r = append_iovec(m, m->header, BUS_MESSAGE_BODY_BEGIN(m));
  91         if (r < 0)
  92                 goto fail;
  93
  94         MESSAGE_FOREACH_PART(part, i, m)  {
  95                 r = bus_body_part_map(part);
  96                 if (r < 0)
  97                         goto fail;
  98
  99                 r = append_iovec(m, part->data, part->size);
 100                 if (r < 0)
 101                         goto fail;
 102         }
 103
 104         assert(n == m->n_iovec);
 105
 106         return 0;
 107
 108 fail:
 109         m->poisoned = true;
 110         return r;
 111 }
 112
 113 bool bus_socket_auth_needs_write(sd_bus *b) {
 114
 115         unsigned i;
 116
 117         if (b->auth_index >= ELEMENTSOF(b->auth_iovec))
 118                 return false;
 119
 120         for (i = b->auth_index; i < ELEMENTSOF(b->auth_iovec); i++) {
 121                 struct iovec *j = b->auth_iovec + i;
 122
 123                 if (j->iov_len > 0)
 124                         return true;
 125         }
 126
 127         return false;
 128 }
 129
 130 static int bus_socket_write_auth(sd_bus *b) {
 131         ssize_t k;
 132
 133         assert(b);
 134         assert(b->state == BUS_AUTHENTICATING);
 135
 136         if (!bus_socket_auth_needs_write(b))
 137                 return 0;
 138
 139         if (b->prefer_writev)
 140                 k = writev(b->output_fd, b->auth_iovec + b->auth_index, ELEMENTSOF(b->auth_iovec) - b->auth_index);
 141         else {
 142                 struct msghdr mh;
 143                 zero(mh);
 144
 145                 mh.msg_iov = b->auth_iovec + b->auth_index;
 146                 mh.msg_iovlen = ELEMENTSOF(b->auth_iovec) - b->auth_index;
 147
 148                 k = sendmsg(b->output_fd, &mh, MSG_DONTWAIT|MSG_NOSIGNAL);
 149                 if (k < 0 && errno == ENOTSOCK) {
 150                         b->prefer_writev = true;
 151                         k = writev(b->output_fd, b->auth_iovec + b->auth_index, ELEMENTSOF(b->auth_iovec) - b->auth_index);
 152                 }
 153         }
 154
 155         if (k < 0)
 156                 return errno == EAGAIN ? 0 : -errno;
 157
 158         iovec_advance(b->auth_iovec, &b->auth_index, (size_t) k);
 159         return 1;
 160 }
 161
 162 static int bus_socket_auth_verify_client(sd_bus *b) {
 163         char *e, *f, *start;
 164         sd_id128_t peer;
 165         unsigned i;
 166         int r;
 167
 168         assert(b);
 169
 170         /* We expect two response lines: "OK" and possibly
 171          * "AGREE_UNIX_FD" */
 172
 173         e = memmem_safe(b->rbuffer, b->rbuffer_size, "\r\n", 2);
 174         if (!e)
 175                 return 0;
 176
 177         if (b->accept_fd) {
 178                 f = memmem(e + 2, b->rbuffer_size - (e - (char*) b->rbuffer) - 2, "\r\n", 2);
 179                 if (!f)
 180                         return 0;
 181
 182                 start = f + 2;
 183         } else {
 184                 f = NULL;
 185                 start = e + 2;
 186         }
 187
 188         /* Nice! We got all the lines we need. First check the OK
 189          * line */
 190
 191         if (e - (char*) b->rbuffer != 3 + 32)
 192                 return -EPERM;
 193
 194         if (memcmp(b->rbuffer, "OK ", 3))
 195                 return -EPERM;
 196
 197         b->auth = b->anonymous_auth ? BUS_AUTH_ANONYMOUS : BUS_AUTH_EXTERNAL;
 198
 199         for (i = 0; i < 32; i += 2) {
 200                 int x, y;
 201
 202                 x = unhexchar(((char*) b->rbuffer)[3 + i]);
 203                 y = unhexchar(((char*) b->rbuffer)[3 + i + 1]);
 204
 205                 if (x < 0 || y < 0)
 206                         return -EINVAL;
 207
 208                 peer.bytes[i/2] = ((uint8_t) x << 4 | (uint8_t) y);
 209         }
 210
 211         if (!sd_id128_is_null(b->server_id) &&
 212             !sd_id128_equal(b->server_id, peer))
 213                 return -EPERM;
 214
 215         b->server_id = peer;
 216
 217         /* And possibly check the second line, too */
 218
 219         if (f)
 220                 b->can_fds =
 221                         (f - e == STRLEN("\r\nAGREE_UNIX_FD")) &&
 222                         memcmp(e + 2, "AGREE_UNIX_FD",
 223                                STRLEN("AGREE_UNIX_FD")) == 0;
 224
 225         b->rbuffer_size -= (start - (char*) b->rbuffer);
 226         memmove(b->rbuffer, start, b->rbuffer_size);
 227
 228         r = bus_start_running(b);
 229         if (r < 0)
 230                 return r;
 231
 232         return 1;
 233 }
 234
 235 static bool line_equals(const char *s, size_t m, const char *line) {
 236         size_t l;
 237
 238         l = strlen(line);
 239         if (l != m)
 240                 return false;
 241
 242         return memcmp(s, line, l) == 0;
 243 }
 244
 245 static bool line_begins(const char *s, size_t m, const char *word) {
 246         const char *p;
 247
 248         p = memory_startswith(s, m, word);
 249         return p && (p == (s + m) || *p == ' ');
 250 }
 251
 252 static int verify_anonymous_token(sd_bus *b, const char *p, size_t l) {
 253         _cleanup_free_ char *token = NULL;
 254         size_t len;
 255         int r;
 256
 257         if (!b->anonymous_auth)
 258                 return 0;
 259
 260         if (l <= 0)
 261                 return 1;
 262
 263         assert(p[0] == ' ');
 264         p++; l--;
 265
 266         if (l % 2 != 0)
 267                 return 0;
 268
 269         r = unhexmem(p, l, (void **) &token, &len);
 270         if (r < 0)
 271                 return 0;
 272
 273         if (memchr(token, 0, len))
 274                 return 0;
 275
 276         return !!utf8_is_valid(token);
 277 }
 278
 279 static int verify_external_token(sd_bus *b, const char *p, size_t l) {
 280         _cleanup_free_ char *token = NULL;
 281         size_t len;
 282         uid_t u;
 283         int r;
 284
 285         /* We don't do any real authentication here. Instead, we if
 286          * the owner of this bus wanted authentication he should have
 287          * checked SO_PEERCRED before even creating the bus object. */
 288
 289         if (!b->anonymous_auth && !b->ucred_valid)
 290                 return 0;
 291
 292         if (l <= 0)
 293                 return 1;
 294
 295         assert(p[0] == ' ');
 296         p++; l--;
 297
 298         if (l % 2 != 0)
 299                 return 0;
 300
 301         r = unhexmem(p, l, (void**) &token, &len);
 302         if (r < 0)
 303                 return 0;
 304
 305         if (memchr(token, 0, len))
 306                 return 0;
 307
 308         r = parse_uid(token, &u);
 309         if (r < 0)
 310                 return 0;
 311
 312         /* We ignore the passed value if anonymous authentication is
 313          * on anyway. */
 314         if (!b->anonymous_auth && u != b->ucred.uid)
 315                 return 0;
 316
 317         return 1;
 318 }
 319
 320 static int bus_socket_auth_write(sd_bus *b, const char *t) {
 321         char *p;
 322         size_t l;
 323
 324         assert(b);
 325         assert(t);
 326
 327         /* We only make use of the first iovec */
 328         assert(IN_SET(b->auth_index, 0, 1));
 329
 330         l = strlen(t);
 331         p = malloc(b->auth_iovec[0].iov_len + l);
 332         if (!p)
 333                 return -ENOMEM;
 334
 335         memcpy_safe(p, b->auth_iovec[0].iov_base, b->auth_iovec[0].iov_len);
 336         memcpy(p + b->auth_iovec[0].iov_len, t, l);
 337
 338         b->auth_iovec[0].iov_base = p;
 339         b->auth_iovec[0].iov_len += l;
 340
 341         free(b->auth_buffer);
 342         b->auth_buffer = p;
 343         b->auth_index = 0;
 344         return 0;
 345 }
 346
 347 static int bus_socket_auth_write_ok(sd_bus *b) {
 348         char t[3 + 32 + 2 + 1];
 349
 350         assert(b);
 351
 352         xsprintf(t, "OK " SD_ID128_FORMAT_STR "\r\n", SD_ID128_FORMAT_VAL(b->server_id));
 353
 354         return bus_socket_auth_write(b, t);
 355 }
 356
 357 static int bus_socket_auth_verify_server(sd_bus *b) {
 358         char *e;
 359         const char *line;
 360         size_t l;
 361         bool processed = false;
 362         int r;
 363
 364         assert(b);
 365
 366         if (b->rbuffer_size < 1)
 367                 return 0;
 368
 369         /* First char must be a NUL byte */
 370         if (*(char*) b->rbuffer != 0)
 371                 return -EIO;
 372
 373         if (b->rbuffer_size < 3)
 374                 return 0;
 375
 376         /* Begin with the first line */
 377         if (b->auth_rbegin <= 0)
 378                 b->auth_rbegin = 1;
 379
 380         for (;;) {
 381                 /* Check if line is complete */
 382                 line = (char*) b->rbuffer + b->auth_rbegin;
 383                 e = memmem(line, b->rbuffer_size - b->auth_rbegin, "\r\n", 2);
 384                 if (!e)
 385                         return processed;
 386
 387                 l = e - line;
 388
 389                 if (line_begins(line, l, "AUTH ANONYMOUS")) {
 390
 391                         r = verify_anonymous_token(b, line + 14, l - 14);
 392                         if (r < 0)
 393                                 return r;
 394                         if (r == 0)
 395                                 r = bus_socket_auth_write(b, "REJECTED\r\n");
 396                         else {
 397                                 b->auth = BUS_AUTH_ANONYMOUS;
 398                                 r = bus_socket_auth_write_ok(b);
 399                         }
 400
 401                 } else if (line_begins(line, l, "AUTH EXTERNAL")) {
 402
 403                         r = verify_external_token(b, line + 13, l - 13);
 404                         if (r < 0)
 405                                 return r;
 406                         if (r == 0)
 407                                 r = bus_socket_auth_write(b, "REJECTED\r\n");
 408                         else {
 409                                 b->auth = BUS_AUTH_EXTERNAL;
 410                                 r = bus_socket_auth_write_ok(b);
 411                         }
 412
 413                 } else if (line_begins(line, l, "AUTH"))
 414                         r = bus_socket_auth_write(b, "REJECTED EXTERNAL ANONYMOUS\r\n");
 415                 else if (line_equals(line, l, "CANCEL") ||
 416                          line_begins(line, l, "ERROR")) {
 417
 418                         b->auth = _BUS_AUTH_INVALID;
 419                         r = bus_socket_auth_write(b, "REJECTED\r\n");
 420
 421                 } else if (line_equals(line, l, "BEGIN")) {
 422
 423                         if (b->auth == _BUS_AUTH_INVALID)
 424                                 r = bus_socket_auth_write(b, "ERROR\r\n");
 425                         else {
 426                                 /* We can't leave from the auth phase
 427                                  * before we haven't written
 428                                  * everything queued, so let's check
 429                                  * that */
 430
 431                                 if (bus_socket_auth_needs_write(b))
 432                                         return 1;
 433
 434                                 b->rbuffer_size -= (e + 2 - (char*) b->rbuffer);
 435                                 memmove(b->rbuffer, e + 2, b->rbuffer_size);
 436                                 return bus_start_running(b);
 437                         }
 438
 439                 } else if (line_begins(line, l, "DATA")) {
 440
 441                         if (b->auth == _BUS_AUTH_INVALID)
 442                                 r = bus_socket_auth_write(b, "ERROR\r\n");
 443                         else {
 444                                 if (b->auth == BUS_AUTH_ANONYMOUS)
 445                                         r = verify_anonymous_token(b, line + 4, l - 4);
 446                                 else
 447                                         r = verify_external_token(b, line + 4, l - 4);
 448
 449                                 if (r < 0)
 450                                         return r;
 451                                 if (r == 0) {
 452                                         b->auth = _BUS_AUTH_INVALID;
 453                                         r = bus_socket_auth_write(b, "REJECTED\r\n");
 454                                 } else
 455                                         r = bus_socket_auth_write_ok(b);
 456                         }
 457                 } else if (line_equals(line, l, "NEGOTIATE_UNIX_FD")) {
 458                         if (b->auth == _BUS_AUTH_INVALID || !b->accept_fd)
 459                                 r = bus_socket_auth_write(b, "ERROR\r\n");
 460                         else {
 461                                 b->can_fds = true;
 462                                 r = bus_socket_auth_write(b, "AGREE_UNIX_FD\r\n");
 463                         }
 464                 } else
 465                         r = bus_socket_auth_write(b, "ERROR\r\n");
 466
 467                 if (r < 0)
 468                         return r;
 469
 470                 b->auth_rbegin = e + 2 - (char*) b->rbuffer;
 471
 472                 processed = true;
 473         }
 474 }
 475
 476 static int bus_socket_auth_verify(sd_bus *b) {
 477         assert(b);
 478
 479         if (b->is_server)
 480                 return bus_socket_auth_verify_server(b);
 481         else
 482                 return bus_socket_auth_verify_client(b);
 483 }
 484
 485 static int bus_socket_read_auth(sd_bus *b) {
 486         struct msghdr mh;
 487         struct iovec iov = {};
 488         size_t n;
 489         ssize_t k;
 490         int r;
 491         void *p;
 492         union {
 493                 struct cmsghdr cmsghdr;
 494                 uint8_t buf[CMSG_SPACE(sizeof(int) * BUS_FDS_MAX)];
 495         } control;
 496         bool handle_cmsg = false;
 497
 498         assert(b);
 499         assert(b->state == BUS_AUTHENTICATING);
 500
 501         r = bus_socket_auth_verify(b);
 502         if (r != 0)
 503                 return r;
 504
 505         n = MAX(256u, b->rbuffer_size * 2);
 506
 507         if (n > BUS_AUTH_SIZE_MAX)
 508                 n = BUS_AUTH_SIZE_MAX;
 509
 510         if (b->rbuffer_size >= n)
 511                 return -ENOBUFS;
 512
 513         p = realloc(b->rbuffer, n);
 514         if (!p)
 515                 return -ENOMEM;
 516
 517         b->rbuffer = p;
 518
 519         iov.iov_base = (uint8_t*) b->rbuffer + b->rbuffer_size;
 520         iov.iov_len = n - b->rbuffer_size;
 521
 522         if (b->prefer_readv)
 523                 k = readv(b->input_fd, &iov, 1);
 524         else {
 525                 zero(mh);
 526                 mh.msg_iov = &iov;
 527                 mh.msg_iovlen = 1;
 528                 mh.msg_control = &control;
 529                 mh.msg_controllen = sizeof(control);
 530
 531                 k = recvmsg(b->input_fd, &mh, MSG_DONTWAIT|MSG_CMSG_CLOEXEC);
 532                 if (k < 0 && errno == ENOTSOCK) {
 533                         b->prefer_readv = true;
 534                         k = readv(b->input_fd, &iov, 1);
 535                 } else
 536                         handle_cmsg = true;
 537         }
 538         if (k < 0)
 539                 return errno == EAGAIN ? 0 : -errno;
 540         if (k == 0)
 541                 return -ECONNRESET;
 542
 543         b->rbuffer_size += k;
 544
 545         if (handle_cmsg) {
 546                 struct cmsghdr *cmsg;
 547
 548                 CMSG_FOREACH(cmsg, &mh)
 549                         if (cmsg->cmsg_level == SOL_SOCKET &&
 550                             cmsg->cmsg_type == SCM_RIGHTS) {
 551                                 int j;
 552
 553                                 /* Whut? We received fds during the auth
 554                                  * protocol? Somebody is playing games with
 555                                  * us. Close them all, and fail */
 556                                 j = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int);
 557                                 close_many((int*) CMSG_DATA(cmsg), j);
 558                                 return -EIO;
 559                         } else
 560                                 log_debug("Got unexpected auxiliary data with level=%d and type=%d",
 561                                           cmsg->cmsg_level, cmsg->cmsg_type);
 562         }
 563
 564         r = bus_socket_auth_verify(b);
 565         if (r != 0)
 566                 return r;
 567
 568         return 1;
 569 }
 570
 571 void bus_socket_setup(sd_bus *b) {
 572         assert(b);
 573
 574         /* Increase the buffers to 8 MB */
 575         (void) fd_inc_rcvbuf(b->input_fd, SNDBUF_SIZE);
 576         (void) fd_inc_sndbuf(b->output_fd, SNDBUF_SIZE);
 577
 578         b->message_version = 1;
 579         b->message_endian = 0;
 580 }
 581
 582 static void bus_get_peercred(sd_bus *b) {
 583         int r;
 584
 585         assert(b);
 586         assert(!b->ucred_valid);
 587         assert(!b->label);
 588         assert(b->n_groups == (size_t) -1);
 589
 590         /* Get the peer for socketpair() sockets */
 591         b->ucred_valid = getpeercred(b->input_fd, &b->ucred) >= 0;
 592
 593         /* Get the SELinux context of the peer */
 594         r = getpeersec(b->input_fd, &b->label);
 595         if (r < 0 && !IN_SET(r, -EOPNOTSUPP, -ENOPROTOOPT))
 596                 log_debug_errno(r, "Failed to determine peer security context: %m");
 597
 598         /* Get the list of auxiliary groups of the peer */
 599         r = getpeergroups(b->input_fd, &b->groups);
 600         if (r >= 0)
 601                 b->n_groups = (size_t) r;
 602         else if (!IN_SET(r, -EOPNOTSUPP, -ENOPROTOOPT))
 603                 log_debug_errno(r, "Failed to determine peer's group list: %m");
 604 }
 605
 606 static int bus_socket_start_auth_client(sd_bus *b) {
 607         size_t l;
 608         const char *auth_suffix, *auth_prefix;
 609
 610         assert(b);
 611
 612         if (b->anonymous_auth) {
 613                 auth_prefix = "\0AUTH ANONYMOUS ";
 614
 615                 /* For ANONYMOUS auth we send some arbitrary "trace" string */
 616                 l = 9;
 617                 b->auth_buffer = hexmem("anonymous", l);
 618         } else {
 619                 char text[DECIMAL_STR_MAX(uid_t) + 1];
 620
 621                 auth_prefix = "\0AUTH EXTERNAL ";
 622
 623                 xsprintf(text, UID_FMT, geteuid());
 624
 625                 l = strlen(text);
 626                 b->auth_buffer = hexmem(text, l);
 627         }
 628
 629         if (!b->auth_buffer)
 630                 return -ENOMEM;
 631
 632         if (b->accept_fd)
 633                 auth_suffix = "\r\nNEGOTIATE_UNIX_FD\r\nBEGIN\r\n";
 634         else
 635                 auth_suffix = "\r\nBEGIN\r\n";
 636
 637         b->auth_iovec[0].iov_base = (void*) auth_prefix;
 638         b->auth_iovec[0].iov_len = 1 + strlen(auth_prefix + 1);
 639         b->auth_iovec[1].iov_base = (void*) b->auth_buffer;
 640         b->auth_iovec[1].iov_len = l * 2;
 641         b->auth_iovec[2].iov_base = (void*) auth_suffix;
 642         b->auth_iovec[2].iov_len = strlen(auth_suffix);
 643
 644         return bus_socket_write_auth(b);
 645 }
 646
 647 int bus_socket_start_auth(sd_bus *b) {
 648         assert(b);
 649
 650         bus_get_peercred(b);
 651
 652         bus_set_state(b, BUS_AUTHENTICATING);
 653         b->auth_timeout = now(CLOCK_MONOTONIC) + BUS_AUTH_TIMEOUT;
 654
 655         if (sd_is_socket(b->input_fd, AF_UNIX, 0, 0) <= 0)
 656                 b->accept_fd = false;
 657
 658         if (b->output_fd != b->input_fd)
 659                 if (sd_is_socket(b->output_fd, AF_UNIX, 0, 0) <= 0)
 660                         b->accept_fd = false;
 661
 662         if (b->is_server)
 663                 return bus_socket_read_auth(b);
 664         else
 665                 return bus_socket_start_auth_client(b);
 666 }
 667
 668 static int bus_socket_inotify_setup(sd_bus *b) {
 669         _cleanup_free_ int *new_watches = NULL;
 670         _cleanup_free_ char *absolute = NULL;
 671         size_t n_allocated = 0, n = 0, done = 0, i;
 672         unsigned max_follow = 32;
 673         const char *p;
 674         int wd, r;
 675
 676         assert(b);
 677         assert(b->watch_bind);
 678         assert(b->sockaddr.sa.sa_family == AF_UNIX);
 679         assert(b->sockaddr.un.sun_path[0] != 0);
 680
 681         /* Sets up an inotify fd in case watch_bind is enabled: wait until the configured AF_UNIX file system socket
 682          * appears before connecting to it. The implemented is pretty simplistic: we just subscribe to relevant changes
 683          * to all prefix components of the path, and every time we get an event for that we try to reconnect again,
 684          * without actually caring what precisely the event we got told us. If we still can't connect we re-subscribe
 685          * to all relevant changes of anything in the path, so that our watches include any possibly newly created path
 686          * components. */
 687
 688         if (b->inotify_fd < 0) {
 689                 b->inotify_fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
 690                 if (b->inotify_fd < 0)
 691                         return -errno;
 692
 693                 b->inotify_fd = fd_move_above_stdio(b->inotify_fd);
 694         }
 695
 696         /* Make sure the path is NUL terminated */
 697         p = strndupa(b->sockaddr.un.sun_path, sizeof(b->sockaddr.un.sun_path));
 698
 699         /* Make sure the path is absolute */
 700         r = path_make_absolute_cwd(p, &absolute);
 701         if (r < 0)
 702                 goto fail;
 703
 704         /* Watch all parent directories, and don't mind any prefix that doesn't exist yet. For the innermost directory
 705          * that exists we want to know when files are created or moved into it. For all parents of it we just care if
 706          * they are removed or renamed. */
 707
 708         if (!GREEDY_REALLOC(new_watches, n_allocated, n + 1)) {
 709                 r = -ENOMEM;
 710                 goto fail;
 711         }
 712
 713         /* Start with the top-level directory, which is a bit simpler than the rest, since it can't be a symlink, and
 714          * always exists */
 715         wd = inotify_add_watch(b->inotify_fd, "/", IN_CREATE|IN_MOVED_TO);
 716         if (wd < 0) {
 717                 r = log_debug_errno(errno, "Failed to add inotify watch on /: %m");
 718                 goto fail;
 719         } else
 720                 new_watches[n++] = wd;
 721
 722         for (;;) {
 723                 _cleanup_free_ char *component = NULL, *prefix = NULL, *destination = NULL;
 724                 size_t n_slashes, n_component;
 725                 char *c = NULL;
 726
 727                 n_slashes = strspn(absolute + done, "/");
 728                 n_component = n_slashes + strcspn(absolute + done + n_slashes, "/");
 729
 730                 if (n_component == 0) /* The end */
 731                         break;
 732
 733                 component = strndup(absolute + done, n_component);
 734                 if (!component) {
 735                         r = -ENOMEM;
 736                         goto fail;
 737                 }
 738
 739                 /* A trailing slash? That's a directory, and not a socket then */
 740                 if (path_equal(component, "/")) {
 741                         r = -EISDIR;
 742                         goto fail;
 743                 }
 744
 745                 /* A single dot? Let's eat this up */
 746                 if (path_equal(component, "/.")) {
 747                         done += n_component;
 748                         continue;
 749                 }
 750
 751                 prefix = strndup(absolute, done + n_component);
 752                 if (!prefix) {
 753                         r = -ENOMEM;
 754                         goto fail;
 755                 }
 756
 757                 if (!GREEDY_REALLOC(new_watches, n_allocated, n + 1)) {
 758                         r = -ENOMEM;
 759                         goto fail;
 760                 }
 761
 762                 wd = inotify_add_watch(b->inotify_fd, prefix, IN_DELETE_SELF|IN_MOVE_SELF|IN_ATTRIB|IN_CREATE|IN_MOVED_TO|IN_DONT_FOLLOW);
 763                 log_debug("Added inotify watch for %s on bus %s: %i", prefix, strna(b->description), wd);
 764
 765                 if (wd < 0) {
 766                         if (IN_SET(errno, ENOENT, ELOOP))
 767                                 break; /* This component doesn't exist yet, or the path contains a cyclic symlink right now */
 768
 769                         r = log_debug_errno(errno, "Failed to add inotify watch on %s: %m", empty_to_root(prefix));
 770                         goto fail;
 771                 } else
 772                         new_watches[n++] = wd;
 773
 774                 /* Check if this is possibly a symlink. If so, let's follow it and watch it too. */
 775                 r = readlink_malloc(prefix, &destination);
 776                 if (r == -EINVAL) { /* not a symlink */
 777                         done += n_component;
 778                         continue;
 779                 }
 780                 if (r < 0)
 781                         goto fail;
 782
 783                 if (isempty(destination)) { /* Empty symlink target? Yuck! */
 784                         r = -EINVAL;
 785                         goto fail;
 786                 }
 787
 788                 if (max_follow <= 0) { /* Let's make sure we don't follow symlinks forever */
 789                         r = -ELOOP;
 790                         goto fail;
 791                 }
 792
 793                 if (path_is_absolute(destination)) {
 794                         /* For absolute symlinks we build the new path and start anew */
 795                         c = strjoin(destination, absolute + done + n_component);
 796                         done = 0;
 797                 } else {
 798                         _cleanup_free_ char *t = NULL;
 799
 800                         /* For relative symlinks we replace the last component, and try again */
 801                         t = strndup(absolute, done);
 802                         if (!t)
 803                                 return -ENOMEM;
 804
 805                         c = strjoin(t, "/", destination, absolute + done + n_component);
 806                 }
 807                 if (!c) {
 808                         r = -ENOMEM;
 809                         goto fail;
 810                 }
 811
 812                 free(absolute);
 813                 absolute = c;
 814
 815                 max_follow--;
 816         }
 817
 818         /* And now, let's remove all watches from the previous iteration we don't need anymore */
 819         for (i = 0; i < b->n_inotify_watches; i++) {
 820                 bool found = false;
 821                 size_t j;
 822
 823                 for (j = 0; j < n; j++)
 824                         if (new_watches[j] == b->inotify_watches[i]) {
 825                                 found = true;
 826                                 break;
 827                         }
 828
 829                 if (found)
 830                         continue;
 831
 832                 (void) inotify_rm_watch(b->inotify_fd, b->inotify_watches[i]);
 833         }
 834
 835         free_and_replace(b->inotify_watches, new_watches);
 836         b->n_inotify_watches = n;
 837
 838         return 0;
 839
 840 fail:
 841         bus_close_inotify_fd(b);
 842         return r;
 843 }
 844
 845 int bus_socket_connect(sd_bus *b) {
 846         bool inotify_done = false;
 847         int r;
 848
 849         assert(b);
 850
 851         for (;;) {
 852                 assert(b->input_fd < 0);
 853                 assert(b->output_fd < 0);
 854                 assert(b->sockaddr.sa.sa_family != AF_UNSPEC);
 855
 856                 b->input_fd = socket(b->sockaddr.sa.sa_family, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
 857                 if (b->input_fd < 0)
 858                         return -errno;
 859
 860                 b->input_fd = fd_move_above_stdio(b->input_fd);
 861
 862                 b->output_fd = b->input_fd;
 863                 bus_socket_setup(b);
 864
 865                 if (connect(b->input_fd, &b->sockaddr.sa, b->sockaddr_size) < 0) {
 866                         if (errno == EINPROGRESS) {
 867
 868                                 /* If we have any inotify watches open, close them now, we don't need them anymore, as
 869                                  * we have successfully initiated a connection */
 870                                 bus_close_inotify_fd(b);
 871
 872                                 /* Note that very likely we are already in BUS_OPENING state here, as we enter it when
 873                                  * we start parsing the address string. The only reason we set the state explicitly
 874                                  * here, is to undo BUS_WATCH_BIND, in case we did the inotify magic. */
 875                                 bus_set_state(b, BUS_OPENING);
 876                                 return 1;
 877                         }
 878
 879                         if (IN_SET(errno, ENOENT, ECONNREFUSED) &&  /* ENOENT → unix socket doesn't exist at all; ECONNREFUSED → unix socket stale */
 880                             b->watch_bind &&
 881                             b->sockaddr.sa.sa_family == AF_UNIX &&
 882                             b->sockaddr.un.sun_path[0] != 0) {
 883
 884                                 /* This connection attempt failed, let's release the socket for now, and start with a
 885                                  * fresh one when reconnecting. */
 886                                 bus_close_io_fds(b);
 887
 888                                 if (inotify_done) {
 889                                         /* inotify set up already, don't do it again, just return now, and remember
 890                                          * that we are waiting for inotify events now. */
 891                                         bus_set_state(b, BUS_WATCH_BIND);
 892                                         return 1;
 893                                 }
 894
 895                                 /* This is a file system socket, and the inotify logic is enabled. Let's create the necessary inotify fd. */
 896                                 r = bus_socket_inotify_setup(b);
 897                                 if (r < 0)
 898                                         return r;
 899
 900                                 /* Let's now try to connect a second time, because in theory there's otherwise a race
 901                                  * here: the socket might have been created in the time between our first connect() and
 902                                  * the time we set up the inotify logic. But let's remember that we set up inotify now,
 903                                  * so that we don't do the connect() more than twice. */
 904                                 inotify_done = true;
 905
 906                         } else
 907                                 return -errno;
 908                 } else
 909                         break;
 910         }
 911
 912         /* Yay, established, we don't need no inotify anymore! */
 913         bus_close_inotify_fd(b);
 914
 915         return bus_socket_start_auth(b);
 916 }
 917
 918 int bus_socket_exec(sd_bus *b) {
 919         int s[2], r;
 920
 921         assert(b);
 922         assert(b->input_fd < 0);
 923         assert(b->output_fd < 0);
 924         assert(b->exec_path);
 925         assert(b->busexec_pid == 0);
 926
 927         r = socketpair(AF_UNIX, SOCK_STREAM|SOCK_NONBLOCK|SOCK_CLOEXEC, 0, s);
 928         if (r < 0)
 929                 return -errno;
 930
 931         r = safe_fork_full("(sd-busexec)", s+1, 1, FORK_RESET_SIGNALS|FORK_CLOSE_ALL_FDS, &b->busexec_pid);
 932         if (r < 0) {
 933                 safe_close_pair(s);
 934                 return r;
 935         }
 936         if (r == 0) {
 937                 /* Child */
 938
 939                 if (rearrange_stdio(s[1], s[1], STDERR_FILENO) < 0)
 940                         _exit(EXIT_FAILURE);
 941
 942                 if (b->exec_argv)
 943                         execvp(b->exec_path, b->exec_argv);
 944                 else {
 945                         const char *argv[] = { b->exec_path, NULL };
 946                         execvp(b->exec_path, (char**) argv);
 947                 }
 948
 949                 _exit(EXIT_FAILURE);
 950         }
 951
 952         safe_close(s[1]);
 953         b->output_fd = b->input_fd = fd_move_above_stdio(s[0]);
 954
 955         bus_socket_setup(b);
 956
 957         return bus_socket_start_auth(b);
 958 }
 959
 960 int bus_socket_take_fd(sd_bus *b) {
 961         assert(b);
 962
 963         bus_socket_setup(b);
 964
 965         return bus_socket_start_auth(b);
 966 }
 967
 968 int bus_socket_write_message(sd_bus *bus, sd_bus_message *m, size_t *idx) {
 969         struct iovec *iov;
 970         ssize_t k;
 971         size_t n;
 972         unsigned j;
 973         int r;
 974
 975         assert(bus);
 976         assert(m);
 977         assert(idx);
 978         assert(IN_SET(bus->state, BUS_RUNNING, BUS_HELLO));
 979
 980         if (*idx >= BUS_MESSAGE_SIZE(m))
 981                 return 0;
 982
 983         r = bus_message_setup_iovec(m);
 984         if (r < 0)
 985                 return r;
 986
 987         n = m->n_iovec * sizeof(struct iovec);
 988         iov = alloca(n);
 989         memcpy_safe(iov, m->iovec, n);
 990
 991         j = 0;
 992         iovec_advance(iov, &j, *idx);
 993
 994         if (bus->prefer_writev)
 995                 k = writev(bus->output_fd, iov, m->n_iovec);
 996         else {
 997                 struct msghdr mh = {
 998                         .msg_iov = iov,
 999                         .msg_iovlen = m->n_iovec,
1000                 };
1001
1002                 if (m->n_fds > 0 && *idx == 0) {
1003                         struct cmsghdr *control;
1004
1005                         mh.msg_control = control = alloca(CMSG_SPACE(sizeof(int) * m->n_fds));
1006                         mh.msg_controllen = control->cmsg_len = CMSG_LEN(sizeof(int) * m->n_fds);
1007                         control->cmsg_level = SOL_SOCKET;
1008                         control->cmsg_type = SCM_RIGHTS;
1009                         memcpy(CMSG_DATA(control), m->fds, sizeof(int) * m->n_fds);
1010                 }
1011
1012                 k = sendmsg(bus->output_fd, &mh, MSG_DONTWAIT|MSG_NOSIGNAL);
1013                 if (k < 0 && errno == ENOTSOCK) {
1014                         bus->prefer_writev = true;
1015                         k = writev(bus->output_fd, iov, m->n_iovec);
1016                 }
1017         }
1018
1019         if (k < 0)
1020                 return errno == EAGAIN ? 0 : -errno;
1021
1022         *idx += (size_t) k;
1023         return 1;
1024 }
1025
1026 static int bus_socket_read_message_need(sd_bus *bus, size_t *need) {
1027         uint32_t a, b;
1028         uint8_t e;
1029         uint64_t sum;
1030
1031         assert(bus);
1032         assert(need);
1033         assert(IN_SET(bus->state, BUS_RUNNING, BUS_HELLO));
1034
1035         if (bus->rbuffer_size < sizeof(struct bus_header)) {
1036                 *need = sizeof(struct bus_header) + 8;
1037
1038                 /* Minimum message size:
1039                  *
1040                  * Header +
1041                  *
1042                  *  Method Call: +2 string headers
1043                  *       Signal: +3 string headers
1044                  * Method Error: +1 string headers
1045                  *               +1 uint32 headers
1046                  * Method Reply: +1 uint32 headers
1047                  *
1048                  * A string header is at least 9 bytes
1049                  * A uint32 header is at least 8 bytes
1050                  *
1051                  * Hence the minimum message size of a valid message
1052                  * is header + 8 bytes */
1053
1054                 return 0;
1055         }
1056
1057         a = ((const uint32_t*) bus->rbuffer)[1];
1058         b = ((const uint32_t*) bus->rbuffer)[3];
1059
1060         e = ((const uint8_t*) bus->rbuffer)[0];
1061         if (e == BUS_LITTLE_ENDIAN) {
1062                 a = le32toh(a);
1063                 b = le32toh(b);
1064         } else if (e == BUS_BIG_ENDIAN) {
1065                 a = be32toh(a);
1066                 b = be32toh(b);
1067         } else
1068                 return -EBADMSG;
1069
1070         sum = (uint64_t) sizeof(struct bus_header) + (uint64_t) ALIGN_TO(b, 8) + (uint64_t) a;
1071         if (sum >= BUS_MESSAGE_SIZE_MAX)
1072                 return -ENOBUFS;
1073
1074         *need = (size_t) sum;
1075         return 0;
1076 }
1077
1078 static int bus_socket_make_message(sd_bus *bus, size_t size) {
1079         sd_bus_message *t;
1080         void *b;
1081         int r;
1082
1083         assert(bus);
1084         assert(bus->rbuffer_size >= size);
1085         assert(IN_SET(bus->state, BUS_RUNNING, BUS_HELLO));
1086
1087         r = bus_rqueue_make_room(bus);
1088         if (r < 0)
1089                 return r;
1090
1091         if (bus->rbuffer_size > size) {
1092                 b = memdup((const uint8_t*) bus->rbuffer + size,
1093                            bus->rbuffer_size - size);
1094                 if (!b)
1095                         return -ENOMEM;
1096         } else
1097                 b = NULL;
1098
1099         r = bus_message_from_malloc(bus,
1100                                     bus->rbuffer, size,
1101                                     bus->fds, bus->n_fds,
1102                                     NULL,
1103                                     &t);
1104         if (r < 0) {
1105                 free(b);
1106                 return r;
1107         }
1108
1109         bus->rbuffer = b;
1110         bus->rbuffer_size -= size;
1111
1112         bus->fds = NULL;
1113         bus->n_fds = 0;
1114
1115         bus->rqueue[bus->rqueue_size++] = t;
1116
1117         return 1;
1118 }
1119
1120 int bus_socket_read_message(sd_bus *bus) {
1121         struct msghdr mh;
1122         struct iovec iov = {};
1123         ssize_t k;
1124         size_t need;
1125         int r;
1126         void *b;
1127         union {
1128                 struct cmsghdr cmsghdr;
1129                 uint8_t buf[CMSG_SPACE(sizeof(int) * BUS_FDS_MAX)];
1130         } control;
1131         bool handle_cmsg = false;
1132
1133         assert(bus);
1134         assert(IN_SET(bus->state, BUS_RUNNING, BUS_HELLO));
1135
1136         r = bus_socket_read_message_need(bus, &need);
1137         if (r < 0)
1138                 return r;
1139
1140         if (bus->rbuffer_size >= need)
1141                 return bus_socket_make_message(bus, need);
1142
1143         b = realloc(bus->rbuffer, need);
1144         if (!b)
1145                 return -ENOMEM;
1146
1147         bus->rbuffer = b;
1148
1149         iov.iov_base = (uint8_t*) bus->rbuffer + bus->rbuffer_size;
1150         iov.iov_len = need - bus->rbuffer_size;
1151
1152         if (bus->prefer_readv)
1153                 k = readv(bus->input_fd, &iov, 1);
1154         else {
1155                 zero(mh);
1156                 mh.msg_iov = &iov;
1157                 mh.msg_iovlen = 1;
1158                 mh.msg_control = &control;
1159                 mh.msg_controllen = sizeof(control);
1160
1161                 k = recvmsg(bus->input_fd, &mh, MSG_DONTWAIT|MSG_CMSG_CLOEXEC);
1162                 if (k < 0 && errno == ENOTSOCK) {
1163                         bus->prefer_readv = true;
1164                         k = readv(bus->input_fd, &iov, 1);
1165                 } else
1166                         handle_cmsg = true;
1167         }
1168         if (k < 0)
1169                 return errno == EAGAIN ? 0 : -errno;
1170         if (k == 0)
1171                 return -ECONNRESET;
1172
1173         bus->rbuffer_size += k;
1174
1175         if (handle_cmsg) {
1176                 struct cmsghdr *cmsg;
1177
1178                 CMSG_FOREACH(cmsg, &mh)
1179                         if (cmsg->cmsg_level == SOL_SOCKET &&
1180                             cmsg->cmsg_type == SCM_RIGHTS) {
1181                                 int n, *f, i;
1182
1183                                 n = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int);
1184
1185                                 if (!bus->can_fds) {
1186                                         /* Whut? We received fds but this
1187                                          * isn't actually enabled? Close them,
1188                                          * and fail */
1189
1190                                         close_many((int*) CMSG_DATA(cmsg), n);
1191                                         return -EIO;
1192                                 }
1193
1194                                 f = reallocarray(bus->fds, bus->n_fds + n, sizeof(int));
1195                                 if (!f) {
1196                                         close_many((int*) CMSG_DATA(cmsg), n);
1197                                         return -ENOMEM;
1198                                 }
1199
1200                                 for (i = 0; i < n; i++)
1201                                         f[bus->n_fds++] = fd_move_above_stdio(((int*) CMSG_DATA(cmsg))[i]);
1202                                 bus->fds = f;
1203                         } else
1204                                 log_debug("Got unexpected auxiliary data with level=%d and type=%d",
1205                                           cmsg->cmsg_level, cmsg->cmsg_type);
1206         }
1207
1208         r = bus_socket_read_message_need(bus, &need);
1209         if (r < 0)
1210                 return r;
1211
1212         if (bus->rbuffer_size >= need)
1213                 return bus_socket_make_message(bus, need);
1214
1215         return 1;
1216 }
1217
1218 int bus_socket_process_opening(sd_bus *b) {
1219         int error = 0;
1220         socklen_t slen = sizeof(error);
1221         struct pollfd p = {
1222                 .fd = b->output_fd,
1223                 .events = POLLOUT,
1224         };
1225         int r;
1226
1227         assert(b->state == BUS_OPENING);
1228
1229         r = poll(&p, 1, 0);
1230         if (r < 0)
1231                 return -errno;
1232
1233         if (!(p.revents & (POLLOUT|POLLERR|POLLHUP)))
1234                 return 0;
1235
1236         r = getsockopt(b->output_fd, SOL_SOCKET, SO_ERROR, &error, &slen);
1237         if (r < 0)
1238                 b->last_connect_error = errno;
1239         else if (error != 0)
1240                 b->last_connect_error = error;
1241         else if (p.revents & (POLLERR|POLLHUP))
1242                 b->last_connect_error = ECONNREFUSED;
1243         else
1244                 return bus_socket_start_auth(b);
1245
1246         return bus_next_address(b);
1247 }
1248
1249 int bus_socket_process_authenticating(sd_bus *b) {
1250         int r;
1251
1252         assert(b);
1253         assert(b->state == BUS_AUTHENTICATING);
1254
1255         if (now(CLOCK_MONOTONIC) >= b->auth_timeout)
1256                 return -ETIMEDOUT;
1257
1258         r = bus_socket_write_auth(b);
1259         if (r != 0)
1260                 return r;
1261
1262         return bus_socket_read_auth(b);
1263 }
1264
1265 int bus_socket_process_watch_bind(sd_bus *b) {
1266         int r, q;
1267
1268         assert(b);
1269         assert(b->state == BUS_WATCH_BIND);
1270         assert(b->inotify_fd >= 0);
1271
1272         r = flush_fd(b->inotify_fd);
1273         if (r <= 0)
1274                 return r;
1275
1276         log_debug("Got inotify event on bus %s.", strna(b->description));
1277
1278         /* We flushed events out of the inotify fd. In that case, maybe the socket is valid now? Let's try to connect
1279          * to it again */
1280
1281         r = bus_socket_connect(b);
1282         if (r < 0)
1283                 return r;
1284
1285         q = bus_attach_io_events(b);
1286         if (q < 0)
1287                 return q;
1288
1289         q = bus_attach_inotify_event(b);
1290         if (q < 0)
1291                 return q;
1292
1293         return r;
1294 }