1 /* $NetBSD: rumpuser_sp.c,v 1.77 2020/05/06 12:44:36 christos Exp $ */ 2 3 /* 4 * Copyright (c) 2010, 2011 Antti Kantee. All Rights Reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS 16 * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 17 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 18 * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 21 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * Sysproxy routines. This provides system RPC support over host sockets. 30 * The most notable limitation is that the client and server must share 31 * the same ABI. This does not mean that they have to be the same 32 * machine or that they need to run the same version of the host OS, 33 * just that they must agree on the data structures. This even *might* 34 * work correctly from one hardware architecture to another. 35 */ 36 37 #include "rumpuser_port.h" 38 39 #if !defined(lint) 40 __RCSID("$NetBSD: rumpuser_sp.c,v 1.77 2020/05/06 12:44:36 christos Exp $"); 41 #endif /* !lint */ 42 43 #include <sys/types.h> 44 #include <sys/mman.h> 45 #include <sys/socket.h> 46 47 #include <arpa/inet.h> 48 #include <netinet/in.h> 49 #include <netinet/tcp.h> 50 51 #include <assert.h> 52 #include <errno.h> 53 #include <fcntl.h> 54 #include <poll.h> 55 #include <pthread.h> 56 #include <stdarg.h> 57 #include <stdio.h> 58 #include <stdlib.h> 59 #include <string.h> 60 #include <unistd.h> 61 62 #include <rump/rump.h> /* XXX: for rfork flags */ 63 #include <rump/rumpuser.h> 64 65 #include "rumpuser_int.h" 66 67 #include "sp_common.c" 68 69 #ifndef MAXCLI 70 #define MAXCLI 256 71 #endif 72 #ifndef MAXWORKER 73 #define MAXWORKER 128 74 #endif 75 #ifndef IDLEWORKER 76 #define IDLEWORKER 16 77 #endif 78 int rumpsp_maxworker = MAXWORKER; 79 int rumpsp_idleworker = IDLEWORKER; 80 81 static struct pollfd pfdlist[MAXCLI]; 82 static struct spclient spclist[MAXCLI]; 83 static unsigned int disco; 84 static volatile int spfini; 85 86 static char banner[MAXBANNER]; 87 88 #define PROTOMAJOR 0 89 #define PROTOMINOR 4 90 91 92 /* either no atomic ops, or we haven't figured out how to use them */ 93 #if defined(__linux__) || defined(__APPLE__) || defined(__CYGWIN__) || defined(__OpenBSD__) || defined(__GNU__) || defined(__GLIBC__) 94 static pthread_mutex_t discomtx = PTHREAD_MUTEX_INITIALIZER; 95 96 static void 97 signaldisco(void) 98 { 99 100 pthread_mutex_lock(&discomtx); 101 disco++; 102 pthread_mutex_unlock(&discomtx); 103 } 104 105 static unsigned int 106 getdisco(void) 107 { 108 unsigned int discocnt; 109 110 pthread_mutex_lock(&discomtx); 111 discocnt = disco; 112 disco = 0; 113 pthread_mutex_unlock(&discomtx); 114 115 return discocnt; 116 } 117 118 #elif defined(__FreeBSD__) || defined(__DragonFly__) 119 120 #include <machine/atomic.h> 121 #define signaldisco() atomic_add_int(&disco, 1) 122 #define getdisco() atomic_readandclear_int(&disco) 123 124 #else /* NetBSD */ 125 126 #include <sys/atomic.h> 127 #define signaldisco() atomic_inc_uint(&disco) 128 #define getdisco() atomic_swap_uint(&disco, 0) 129 130 #endif 131 132 133 struct prefork { 134 uint32_t pf_auth[AUTHLEN]; 135 struct lwp *pf_lwp; 136 137 LIST_ENTRY(prefork) pf_entries; /* global list */ 138 LIST_ENTRY(prefork) pf_spcentries; /* linked from forking spc */ 139 }; 140 static LIST_HEAD(, prefork) preforks = LIST_HEAD_INITIALIZER(preforks); 141 static pthread_mutex_t pfmtx; 142 143 /* 144 * This version is for the server. It's optimized for multiple threads 145 * and is *NOT* reentrant wrt to signals. 146 */ 147 static int 148 waitresp(struct spclient *spc, struct respwait *rw) 149 { 150 int spcstate; 151 int rv = 0; 152 153 pthread_mutex_lock(&spc->spc_mtx); 154 sendunlockl(spc); 155 while (!rw->rw_done && spc->spc_state != SPCSTATE_DYING) { 156 pthread_cond_wait(&rw->rw_cv, &spc->spc_mtx); 157 } 158 TAILQ_REMOVE(&spc->spc_respwait, rw, rw_entries); 159 spcstate = spc->spc_state; 160 pthread_mutex_unlock(&spc->spc_mtx); 161 162 pthread_cond_destroy(&rw->rw_cv); 163 164 if (rv) 165 return rv; 166 if (spcstate == SPCSTATE_DYING) 167 return ENOTCONN; 168 return rw->rw_error; 169 } 170 171 /* 172 * Manual wrappers, since librump does not have access to the 173 * user namespace wrapped interfaces. 174 */ 175 176 static void 177 lwproc_switch(struct lwp *l) 178 { 179 180 rumpuser__hyp.hyp_schedule(); 181 rumpuser__hyp.hyp_lwproc_switch(l); 182 rumpuser__hyp.hyp_unschedule(); 183 } 184 185 static void 186 lwproc_release(void) 187 { 188 189 rumpuser__hyp.hyp_schedule(); 190 rumpuser__hyp.hyp_lwproc_release(); 191 rumpuser__hyp.hyp_unschedule(); 192 } 193 194 static int 195 lwproc_rfork(struct spclient *spc, int flags, const char *comm) 196 { 197 int rv; 198 199 rumpuser__hyp.hyp_schedule(); 200 rv = rumpuser__hyp.hyp_lwproc_rfork(spc, flags, comm); 201 rumpuser__hyp.hyp_unschedule(); 202 203 return rv; 204 } 205 206 static int 207 lwproc_newlwp(pid_t pid) 208 { 209 int rv; 210 211 rumpuser__hyp.hyp_schedule(); 212 rv = rumpuser__hyp.hyp_lwproc_newlwp(pid); 213 rumpuser__hyp.hyp_unschedule(); 214 215 return rv; 216 } 217 218 static struct lwp * 219 lwproc_curlwp(void) 220 { 221 struct lwp *l; 222 223 rumpuser__hyp.hyp_schedule(); 224 l = rumpuser__hyp.hyp_lwproc_curlwp(); 225 rumpuser__hyp.hyp_unschedule(); 226 227 return l; 228 } 229 230 static pid_t 231 lwproc_getpid(void) 232 { 233 pid_t p; 234 235 rumpuser__hyp.hyp_schedule(); 236 p = rumpuser__hyp.hyp_getpid(); 237 rumpuser__hyp.hyp_unschedule(); 238 239 return p; 240 } 241 242 static void 243 lwproc_execnotify(const char *comm) 244 { 245 246 rumpuser__hyp.hyp_schedule(); 247 rumpuser__hyp.hyp_execnotify(comm); 248 rumpuser__hyp.hyp_unschedule(); 249 } 250 251 static void 252 lwproc_lwpexit(void) 253 { 254 255 rumpuser__hyp.hyp_schedule(); 256 rumpuser__hyp.hyp_lwpexit(); 257 rumpuser__hyp.hyp_unschedule(); 258 } 259 260 static int 261 rumpsyscall(int sysnum, void *data, register_t *regrv) 262 { 263 long retval[2] = {0, 0}; 264 int rv; 265 266 rumpuser__hyp.hyp_schedule(); 267 rv = rumpuser__hyp.hyp_syscall(sysnum, data, retval); 268 rumpuser__hyp.hyp_unschedule(); 269 270 regrv[0] = retval[0]; 271 regrv[1] = retval[1]; 272 return rv; 273 } 274 275 static uint64_t 276 nextreq(struct spclient *spc) 277 { 278 uint64_t nw; 279 280 pthread_mutex_lock(&spc->spc_mtx); 281 nw = spc->spc_nextreq++; 282 pthread_mutex_unlock(&spc->spc_mtx); 283 284 return nw; 285 } 286 287 /* 288 * XXX: we send responses with "blocking" I/O. This is not 289 * ok for the main thread. XXXFIXME 290 */ 291 292 static void 293 send_error_resp(struct spclient *spc, uint64_t reqno, enum rumpsp_err error) 294 { 295 struct rsp_hdr rhdr; 296 struct iovec iov[1]; 297 298 rhdr.rsp_len = sizeof(rhdr); 299 rhdr.rsp_reqno = reqno; 300 rhdr.rsp_class = RUMPSP_ERROR; 301 rhdr.rsp_type = 0; 302 rhdr.rsp_error = error; 303 304 IOVPUT(iov[0], rhdr); 305 306 sendlock(spc); 307 (void)SENDIOV(spc, iov); 308 sendunlock(spc); 309 } 310 311 static int 312 send_handshake_resp(struct spclient *spc, uint64_t reqno, int error) 313 { 314 struct rsp_hdr rhdr; 315 struct iovec iov[2]; 316 int rv; 317 318 rhdr.rsp_len = sizeof(rhdr) + sizeof(error); 319 rhdr.rsp_reqno = reqno; 320 rhdr.rsp_class = RUMPSP_RESP; 321 rhdr.rsp_type = RUMPSP_HANDSHAKE; 322 rhdr.rsp_error = 0; 323 324 IOVPUT(iov[0], rhdr); 325 IOVPUT(iov[1], error); 326 327 sendlock(spc); 328 rv = SENDIOV(spc, iov); 329 sendunlock(spc); 330 331 return rv; 332 } 333 334 static int 335 send_syscall_resp(struct spclient *spc, uint64_t reqno, int error, 336 register_t *retval) 337 { 338 struct rsp_hdr rhdr; 339 struct rsp_sysresp sysresp; 340 struct iovec iov[2]; 341 int rv; 342 343 rhdr.rsp_len = sizeof(rhdr) + sizeof(sysresp); 344 rhdr.rsp_reqno = reqno; 345 rhdr.rsp_class = RUMPSP_RESP; 346 rhdr.rsp_type = RUMPSP_SYSCALL; 347 rhdr.rsp_sysnum = 0; 348 349 sysresp.rsys_error = error; 350 memcpy(sysresp.rsys_retval, retval, sizeof(sysresp.rsys_retval)); 351 352 IOVPUT(iov[0], rhdr); 353 IOVPUT(iov[1], sysresp); 354 355 sendlock(spc); 356 rv = SENDIOV(spc, iov); 357 sendunlock(spc); 358 359 return rv; 360 } 361 362 static int 363 send_prefork_resp(struct spclient *spc, uint64_t reqno, uint32_t *auth) 364 { 365 struct rsp_hdr rhdr; 366 struct iovec iov[2]; 367 int rv; 368 369 rhdr.rsp_len = sizeof(rhdr) + AUTHLEN*sizeof(*auth); 370 rhdr.rsp_reqno = reqno; 371 rhdr.rsp_class = RUMPSP_RESP; 372 rhdr.rsp_type = RUMPSP_PREFORK; 373 rhdr.rsp_sysnum = 0; 374 375 IOVPUT(iov[0], rhdr); 376 IOVPUT_WITHSIZE(iov[1], auth, AUTHLEN*sizeof(*auth)); 377 378 sendlock(spc); 379 rv = SENDIOV(spc, iov); 380 sendunlock(spc); 381 382 return rv; 383 } 384 385 static int 386 copyin_req(struct spclient *spc, const void *remaddr, size_t *dlen, 387 int wantstr, void **resp) 388 { 389 struct rsp_hdr rhdr; 390 struct rsp_copydata copydata; 391 struct respwait rw; 392 struct iovec iov[2]; 393 int rv; 394 395 DPRINTF(("copyin_req: %zu bytes from %p\n", *dlen, remaddr)); 396 397 rhdr.rsp_len = sizeof(rhdr) + sizeof(copydata); 398 rhdr.rsp_class = RUMPSP_REQ; 399 if (wantstr) 400 rhdr.rsp_type = RUMPSP_COPYINSTR; 401 else 402 rhdr.rsp_type = RUMPSP_COPYIN; 403 rhdr.rsp_sysnum = 0; 404 405 copydata.rcp_addr = __UNCONST(remaddr); 406 copydata.rcp_len = *dlen; 407 408 IOVPUT(iov[0], rhdr); 409 IOVPUT(iov[1], copydata); 410 411 putwait(spc, &rw, &rhdr); 412 rv = SENDIOV(spc, iov); 413 if (rv) { 414 unputwait(spc, &rw); 415 return rv; 416 } 417 418 rv = waitresp(spc, &rw); 419 420 DPRINTF(("copyin: response %d\n", rv)); 421 422 *resp = rw.rw_data; 423 if (wantstr) 424 *dlen = rw.rw_dlen; 425 426 return rv; 427 428 } 429 430 static int 431 send_copyout_req(struct spclient *spc, const void *remaddr, 432 const void *data, size_t dlen) 433 { 434 struct rsp_hdr rhdr; 435 struct rsp_copydata copydata; 436 struct iovec iov[3]; 437 int rv; 438 439 DPRINTF(("copyout_req (async): %zu bytes to %p\n", dlen, remaddr)); 440 441 rhdr.rsp_len = sizeof(rhdr) + sizeof(copydata) + dlen; 442 rhdr.rsp_reqno = nextreq(spc); 443 rhdr.rsp_class = RUMPSP_REQ; 444 rhdr.rsp_type = RUMPSP_COPYOUT; 445 rhdr.rsp_sysnum = 0; 446 447 copydata.rcp_addr = __UNCONST(remaddr); 448 copydata.rcp_len = dlen; 449 450 IOVPUT(iov[0], rhdr); 451 IOVPUT(iov[1], copydata); 452 IOVPUT_WITHSIZE(iov[2], __UNCONST(data), dlen); 453 454 sendlock(spc); 455 rv = SENDIOV(spc, iov); 456 sendunlock(spc); 457 458 return rv; 459 } 460 461 static int 462 anonmmap_req(struct spclient *spc, size_t howmuch, void **resp) 463 { 464 struct rsp_hdr rhdr; 465 struct respwait rw; 466 struct iovec iov[2]; 467 int rv; 468 469 DPRINTF(("anonmmap_req: %zu bytes\n", howmuch)); 470 471 rhdr.rsp_len = sizeof(rhdr) + sizeof(howmuch); 472 rhdr.rsp_class = RUMPSP_REQ; 473 rhdr.rsp_type = RUMPSP_ANONMMAP; 474 rhdr.rsp_sysnum = 0; 475 476 IOVPUT(iov[0], rhdr); 477 IOVPUT(iov[1], howmuch); 478 479 putwait(spc, &rw, &rhdr); 480 rv = SENDIOV(spc, iov); 481 if (rv) { 482 unputwait(spc, &rw); 483 return rv; 484 } 485 486 rv = waitresp(spc, &rw); 487 488 *resp = rw.rw_data; 489 490 DPRINTF(("anonmmap: mapped at %p\n", **(void ***)resp)); 491 492 return rv; 493 } 494 495 static int 496 send_raise_req(struct spclient *spc, int signo) 497 { 498 struct rsp_hdr rhdr; 499 struct iovec iov[1]; 500 int rv; 501 502 rhdr.rsp_len = sizeof(rhdr); 503 rhdr.rsp_class = RUMPSP_REQ; 504 rhdr.rsp_type = RUMPSP_RAISE; 505 rhdr.rsp_signo = signo; 506 507 IOVPUT(iov[0], rhdr); 508 509 sendlock(spc); 510 rv = SENDIOV(spc, iov); 511 sendunlock(spc); 512 513 return rv; 514 } 515 516 static void 517 spcref(struct spclient *spc) 518 { 519 520 pthread_mutex_lock(&spc->spc_mtx); 521 spc->spc_refcnt++; 522 pthread_mutex_unlock(&spc->spc_mtx); 523 } 524 525 static void 526 spcrelease(struct spclient *spc) 527 { 528 int ref; 529 530 pthread_mutex_lock(&spc->spc_mtx); 531 ref = --spc->spc_refcnt; 532 if (__predict_false(spc->spc_inexec && ref <= 2)) 533 pthread_cond_broadcast(&spc->spc_cv); 534 pthread_mutex_unlock(&spc->spc_mtx); 535 536 if (ref > 0) 537 return; 538 539 DPRINTF(("rump_sp: spcrelease: spc %p fd %d\n", spc, spc->spc_fd)); 540 541 _DIAGASSERT(TAILQ_EMPTY(&spc->spc_respwait)); 542 _DIAGASSERT(spc->spc_buf == NULL); 543 544 if (spc->spc_mainlwp) { 545 lwproc_switch(spc->spc_mainlwp); 546 lwproc_release(); 547 } 548 spc->spc_mainlwp = NULL; 549 550 close(spc->spc_fd); 551 spc->spc_fd = -1; 552 spc->spc_state = SPCSTATE_NEW; 553 554 signaldisco(); 555 } 556 557 static void 558 serv_handledisco(unsigned int idx) 559 { 560 struct spclient *spc = &spclist[idx]; 561 int dolwpexit; 562 563 DPRINTF(("rump_sp: disconnecting [%u]\n", idx)); 564 565 pfdlist[idx].fd = -1; 566 pfdlist[idx].revents = 0; 567 pthread_mutex_lock(&spc->spc_mtx); 568 spc->spc_state = SPCSTATE_DYING; 569 kickall(spc); 570 sendunlockl(spc); 571 /* exec uses mainlwp in another thread, but also nuked all lwps */ 572 dolwpexit = !spc->spc_inexec; 573 pthread_mutex_unlock(&spc->spc_mtx); 574 575 if (dolwpexit && spc->spc_mainlwp) { 576 lwproc_switch(spc->spc_mainlwp); 577 lwproc_lwpexit(); 578 lwproc_switch(NULL); 579 } 580 581 /* 582 * Nobody's going to attempt to send/receive anymore, 583 * so reinit info relevant to that. 584 */ 585 /*LINTED:pointer casts may be ok*/ 586 memset((char *)spc + SPC_ZEROFF, 0, sizeof(*spc) - SPC_ZEROFF); 587 588 spcrelease(spc); 589 } 590 591 static void 592 serv_shutdown(void) 593 { 594 struct spclient *spc; 595 unsigned int i; 596 597 for (i = 1; i < MAXCLI; i++) { 598 spc = &spclist[i]; 599 if (spc->spc_fd == -1) 600 continue; 601 602 shutdown(spc->spc_fd, SHUT_RDWR); 603 serv_handledisco(i); 604 605 spcrelease(spc); 606 } 607 } 608 609 static unsigned 610 serv_handleconn(int fd, connecthook_fn connhook, int busy) 611 { 612 struct sockaddr_storage ss; 613 socklen_t sl = sizeof(ss); 614 int newfd, flags; 615 unsigned i; 616 617 /*LINTED: cast ok */ 618 newfd = accept(fd, (struct sockaddr *)&ss, &sl); 619 if (newfd == -1) 620 return 0; 621 622 if (busy) { 623 close(newfd); /* EBUSY */ 624 return 0; 625 } 626 627 flags = fcntl(newfd, F_GETFL, 0); 628 if (fcntl(newfd, F_SETFL, flags | O_NONBLOCK) == -1) { 629 close(newfd); 630 return 0; 631 } 632 633 if (connhook(newfd) != 0) { 634 close(newfd); 635 return 0; 636 } 637 638 /* write out a banner for the client */ 639 if (send(newfd, banner, strlen(banner), MSG_NOSIGNAL) 640 != (ssize_t)strlen(banner)) { 641 close(newfd); 642 return 0; 643 } 644 645 /* find empty slot the simple way */ 646 for (i = 0; i < MAXCLI; i++) { 647 if (pfdlist[i].fd == -1 && spclist[i].spc_state == SPCSTATE_NEW) 648 break; 649 } 650 651 /* 652 * Although not finding a slot is impossible (cf. how this routine 653 * is called), the compiler can still think that i == MAXCLI 654 * if this code is either compiled with NDEBUG or the platform 655 * does not use __dead for assert(). Therefore, add an explicit 656 * check to avoid an array-bounds error. 657 */ 658 /* assert(i < MAXCLI); */ 659 if (i == MAXCLI) 660 abort(); 661 662 pfdlist[i].fd = newfd; 663 spclist[i].spc_fd = newfd; 664 spclist[i].spc_istatus = SPCSTATUS_BUSY; /* dedicated receiver */ 665 spclist[i].spc_refcnt = 1; 666 667 TAILQ_INIT(&spclist[i].spc_respwait); 668 669 DPRINTF(("rump_sp: added new connection fd %d at idx %u\n", newfd, i)); 670 671 return i; 672 } 673 674 static void 675 serv_handlesyscall(struct spclient *spc, struct rsp_hdr *rhdr, uint8_t *data) 676 { 677 register_t retval[2] = {0, 0}; 678 int rv, sysnum; 679 680 sysnum = (int)rhdr->rsp_sysnum; 681 DPRINTF(("rump_sp: handling syscall %d from client %d\n", 682 sysnum, spc->spc_pid)); 683 684 if (__predict_false((rv = lwproc_newlwp(spc->spc_pid)) != 0)) { 685 retval[0] = -1; 686 send_syscall_resp(spc, rhdr->rsp_reqno, rv, retval); 687 return; 688 } 689 spc->spc_syscallreq = rhdr->rsp_reqno; 690 rv = rumpsyscall(sysnum, data, retval); 691 spc->spc_syscallreq = 0; 692 lwproc_release(); 693 694 DPRINTF(("rump_sp: got return value %d & %"PRIxREGISTER 695 "/%"PRIxREGISTER"\n", 696 rv, retval[0], retval[1])); 697 698 send_syscall_resp(spc, rhdr->rsp_reqno, rv, retval); 699 } 700 701 static void 702 serv_handleexec(struct spclient *spc, struct rsp_hdr *rhdr, const char *comm) 703 { 704 pthread_mutex_lock(&spc->spc_mtx); 705 /* one for the connection and one for us */ 706 while (spc->spc_refcnt > 2) 707 pthread_cond_wait(&spc->spc_cv, &spc->spc_mtx); 708 pthread_mutex_unlock(&spc->spc_mtx); 709 710 /* 711 * ok, all the threads are dead (or one is still alive and 712 * the connection is dead, in which case this doesn't matter 713 * very much). proceed with exec. 714 */ 715 716 lwproc_switch(spc->spc_mainlwp); 717 lwproc_execnotify(comm); 718 lwproc_switch(NULL); 719 720 pthread_mutex_lock(&spc->spc_mtx); 721 spc->spc_inexec = 0; 722 pthread_mutex_unlock(&spc->spc_mtx); 723 send_handshake_resp(spc, rhdr->rsp_reqno, 0); 724 } 725 726 enum sbatype { SBA_SYSCALL, SBA_EXEC }; 727 728 struct servbouncearg { 729 struct spclient *sba_spc; 730 struct rsp_hdr sba_hdr; 731 enum sbatype sba_type; 732 uint8_t *sba_data; 733 734 TAILQ_ENTRY(servbouncearg) sba_entries; 735 }; 736 static pthread_mutex_t sbamtx; 737 static pthread_cond_t sbacv; 738 static int nworker, idleworker, nwork; 739 static TAILQ_HEAD(, servbouncearg) wrklist = TAILQ_HEAD_INITIALIZER(wrklist); 740 741 /*ARGSUSED*/ 742 static void * 743 serv_workbouncer(void *arg) 744 { 745 struct servbouncearg *sba; 746 747 for (;;) { 748 pthread_mutex_lock(&sbamtx); 749 if (__predict_false(idleworker - nwork >= rumpsp_idleworker)) { 750 nworker--; 751 pthread_mutex_unlock(&sbamtx); 752 break; 753 } 754 idleworker++; 755 while (TAILQ_EMPTY(&wrklist)) { 756 _DIAGASSERT(nwork == 0); 757 pthread_cond_wait(&sbacv, &sbamtx); 758 } 759 idleworker--; 760 761 sba = TAILQ_FIRST(&wrklist); 762 TAILQ_REMOVE(&wrklist, sba, sba_entries); 763 nwork--; 764 pthread_mutex_unlock(&sbamtx); 765 766 if (__predict_true(sba->sba_type == SBA_SYSCALL)) { 767 serv_handlesyscall(sba->sba_spc, 768 &sba->sba_hdr, sba->sba_data); 769 } else { 770 _DIAGASSERT(sba->sba_type == SBA_EXEC); 771 serv_handleexec(sba->sba_spc, &sba->sba_hdr, 772 (char *)sba->sba_data); 773 } 774 spcrelease(sba->sba_spc); 775 free(sba->sba_data); 776 free(sba); 777 } 778 779 return NULL; 780 } 781 782 static int 783 sp_copyin(void *arg, const void *raddr, void *laddr, size_t *len, int wantstr) 784 { 785 struct spclient *spc = arg; 786 void *rdata = NULL; /* XXXuninit */ 787 int rv, nlocks; 788 789 rumpkern_unsched(&nlocks, NULL); 790 791 rv = copyin_req(spc, raddr, len, wantstr, &rdata); 792 if (rv) 793 goto out; 794 795 memcpy(laddr, rdata, *len); 796 free(rdata); 797 798 out: 799 rumpkern_sched(nlocks, NULL); 800 if (rv) 801 rv = EFAULT; 802 ET(rv); 803 } 804 805 int 806 rumpuser_sp_copyin(void *arg, const void *raddr, void *laddr, size_t len) 807 { 808 int rv; 809 810 rv = sp_copyin(arg, raddr, laddr, &len, 0); 811 ET(rv); 812 } 813 814 int 815 rumpuser_sp_copyinstr(void *arg, const void *raddr, void *laddr, size_t *len) 816 { 817 int rv; 818 819 rv = sp_copyin(arg, raddr, laddr, len, 1); 820 ET(rv); 821 } 822 823 static int 824 sp_copyout(void *arg, const void *laddr, void *raddr, size_t dlen) 825 { 826 struct spclient *spc = arg; 827 int nlocks, rv; 828 829 rumpkern_unsched(&nlocks, NULL); 830 rv = send_copyout_req(spc, raddr, laddr, dlen); 831 rumpkern_sched(nlocks, NULL); 832 833 if (rv) 834 rv = EFAULT; 835 ET(rv); 836 } 837 838 int 839 rumpuser_sp_copyout(void *arg, const void *laddr, void *raddr, size_t dlen) 840 { 841 int rv; 842 843 rv = sp_copyout(arg, laddr, raddr, dlen); 844 ET(rv); 845 } 846 847 int 848 rumpuser_sp_copyoutstr(void *arg, const void *laddr, void *raddr, size_t *dlen) 849 { 850 int rv; 851 852 rv = sp_copyout(arg, laddr, raddr, *dlen); 853 ET(rv); 854 } 855 856 int 857 rumpuser_sp_anonmmap(void *arg, size_t howmuch, void **addr) 858 { 859 struct spclient *spc = arg; 860 void *resp, *rdata = NULL; /* XXXuninit */ 861 int nlocks, rv; 862 863 rumpkern_unsched(&nlocks, NULL); 864 865 rv = anonmmap_req(spc, howmuch, &rdata); 866 if (rv) { 867 rv = EFAULT; 868 goto out; 869 } 870 871 resp = *(void **)rdata; 872 free(rdata); 873 874 if (resp == NULL) { 875 rv = ENOMEM; 876 } 877 878 *addr = resp; 879 880 out: 881 rumpkern_sched(nlocks, NULL); 882 ET(rv); 883 } 884 885 int 886 rumpuser_sp_raise(void *arg, int signo) 887 { 888 struct spclient *spc = arg; 889 int rv, nlocks; 890 891 rumpkern_unsched(&nlocks, NULL); 892 rv = send_raise_req(spc, signo); 893 rumpkern_sched(nlocks, NULL); 894 895 return rv; 896 } 897 898 static pthread_attr_t pattr_detached; 899 static void 900 schedulework(struct spclient *spc, enum sbatype sba_type) 901 { 902 struct servbouncearg *sba; 903 pthread_t pt; 904 uint64_t reqno; 905 int retries = 0; 906 907 reqno = spc->spc_hdr.rsp_reqno; 908 while ((sba = malloc(sizeof(*sba))) == NULL) { 909 if (nworker == 0 || retries++ > 10) { 910 send_error_resp(spc, reqno, RUMPSP_ERR_TRYAGAIN); 911 spcfreebuf(spc); 912 return; 913 } 914 /* slim chance of more memory? */ 915 usleep(10000); 916 } 917 918 sba->sba_spc = spc; 919 sba->sba_type = sba_type; 920 sba->sba_hdr = spc->spc_hdr; 921 sba->sba_data = spc->spc_buf; 922 spcresetbuf(spc); 923 924 spcref(spc); 925 926 pthread_mutex_lock(&sbamtx); 927 TAILQ_INSERT_TAIL(&wrklist, sba, sba_entries); 928 nwork++; 929 if (nwork <= idleworker) { 930 /* do we have a daemon's tool (i.e. idle threads)? */ 931 pthread_cond_signal(&sbacv); 932 } else if (nworker < rumpsp_maxworker) { 933 /* 934 * Else, need to create one 935 * (if we can, otherwise just expect another 936 * worker to pick up the syscall) 937 */ 938 if (pthread_create(&pt, &pattr_detached, 939 serv_workbouncer, NULL) == 0) { 940 nworker++; 941 } 942 } 943 pthread_mutex_unlock(&sbamtx); 944 } 945 946 /* 947 * 948 * Startup routines and mainloop for server. 949 * 950 */ 951 952 struct spservarg { 953 int sps_sock; 954 connecthook_fn sps_connhook; 955 }; 956 957 static void 958 handlereq(struct spclient *spc) 959 { 960 uint64_t reqno; 961 int error; 962 963 reqno = spc->spc_hdr.rsp_reqno; 964 if (__predict_false(spc->spc_state == SPCSTATE_NEW)) { 965 if (spc->spc_hdr.rsp_type != RUMPSP_HANDSHAKE) { 966 send_error_resp(spc, reqno, RUMPSP_ERR_AUTH); 967 shutdown(spc->spc_fd, SHUT_RDWR); 968 spcfreebuf(spc); 969 return; 970 } 971 972 if (spc->spc_hdr.rsp_handshake == HANDSHAKE_GUEST) { 973 /* make sure we fork off of proc1 */ 974 _DIAGASSERT(lwproc_curlwp() == NULL); 975 976 if ((error = lwproc_rfork(spc, RUMP_RFFD_CLEAR, 977 (const char *)spc->spc_buf)) != 0) { 978 shutdown(spc->spc_fd, SHUT_RDWR); 979 } 980 981 spcfreebuf(spc); 982 if (error) 983 return; 984 985 spc->spc_mainlwp = lwproc_curlwp(); 986 987 send_handshake_resp(spc, reqno, 0); 988 } else if (spc->spc_hdr.rsp_handshake == HANDSHAKE_FORK) { 989 struct lwp *tmpmain; 990 struct prefork *pf; 991 struct handshake_fork *rfp; 992 int cancel; 993 994 if (spc->spc_off-HDRSZ != sizeof(*rfp)) { 995 send_error_resp(spc, reqno, 996 RUMPSP_ERR_MALFORMED_REQUEST); 997 shutdown(spc->spc_fd, SHUT_RDWR); 998 spcfreebuf(spc); 999 return; 1000 } 1001 1002 /*LINTED*/ 1003 rfp = (void *)spc->spc_buf; 1004 cancel = rfp->rf_cancel; 1005 1006 pthread_mutex_lock(&pfmtx); 1007 LIST_FOREACH(pf, &preforks, pf_entries) { 1008 if (memcmp(rfp->rf_auth, pf->pf_auth, 1009 sizeof(rfp->rf_auth)) == 0) { 1010 LIST_REMOVE(pf, pf_entries); 1011 LIST_REMOVE(pf, pf_spcentries); 1012 break; 1013 } 1014 } 1015 pthread_mutex_unlock(&pfmtx); 1016 spcfreebuf(spc); 1017 1018 if (!pf) { 1019 send_error_resp(spc, reqno, 1020 RUMPSP_ERR_INVALID_PREFORK); 1021 shutdown(spc->spc_fd, SHUT_RDWR); 1022 return; 1023 } 1024 1025 tmpmain = pf->pf_lwp; 1026 free(pf); 1027 lwproc_switch(tmpmain); 1028 if (cancel) { 1029 lwproc_release(); 1030 shutdown(spc->spc_fd, SHUT_RDWR); 1031 return; 1032 } 1033 1034 /* 1035 * So, we forked already during "prefork" to save 1036 * the file descriptors from a parent exit 1037 * race condition. But now we need to fork 1038 * a second time since the initial fork has 1039 * the wrong spc pointer. (yea, optimize 1040 * interfaces some day if anyone cares) 1041 */ 1042 if ((error = lwproc_rfork(spc, 1043 RUMP_RFFD_SHARE, NULL)) != 0) { 1044 send_error_resp(spc, reqno, 1045 RUMPSP_ERR_RFORK_FAILED); 1046 shutdown(spc->spc_fd, SHUT_RDWR); 1047 lwproc_release(); 1048 return; 1049 } 1050 spc->spc_mainlwp = lwproc_curlwp(); 1051 lwproc_switch(tmpmain); 1052 lwproc_release(); 1053 lwproc_switch(spc->spc_mainlwp); 1054 1055 send_handshake_resp(spc, reqno, 0); 1056 } else { 1057 send_error_resp(spc, reqno, RUMPSP_ERR_AUTH); 1058 shutdown(spc->spc_fd, SHUT_RDWR); 1059 spcfreebuf(spc); 1060 return; 1061 } 1062 1063 spc->spc_pid = lwproc_getpid(); 1064 1065 DPRINTF(("rump_sp: handshake for client %p complete, pid %d\n", 1066 spc, spc->spc_pid)); 1067 1068 lwproc_switch(NULL); 1069 spc->spc_state = SPCSTATE_RUNNING; 1070 return; 1071 } 1072 1073 if (__predict_false(spc->spc_hdr.rsp_type == RUMPSP_PREFORK)) { 1074 struct prefork *pf; 1075 uint32_t auth[AUTHLEN]; 1076 size_t randlen; 1077 int inexec; 1078 1079 DPRINTF(("rump_sp: prefork handler executing for %p\n", spc)); 1080 spcfreebuf(spc); 1081 1082 pthread_mutex_lock(&spc->spc_mtx); 1083 inexec = spc->spc_inexec; 1084 pthread_mutex_unlock(&spc->spc_mtx); 1085 if (inexec) { 1086 send_error_resp(spc, reqno, RUMPSP_ERR_INEXEC); 1087 shutdown(spc->spc_fd, SHUT_RDWR); 1088 return; 1089 } 1090 1091 pf = malloc(sizeof(*pf)); 1092 if (pf == NULL) { 1093 send_error_resp(spc, reqno, RUMPSP_ERR_NOMEM); 1094 return; 1095 } 1096 1097 /* 1098 * Use client main lwp to fork. this is never used by 1099 * worker threads (except in exec, but we checked for that 1100 * above) so we can safely use it here. 1101 */ 1102 lwproc_switch(spc->spc_mainlwp); 1103 if ((error = lwproc_rfork(spc, RUMP_RFFD_COPY, NULL)) != 0) { 1104 DPRINTF(("rump_sp: fork failed: %d (%p)\n",error, spc)); 1105 send_error_resp(spc, reqno, RUMPSP_ERR_RFORK_FAILED); 1106 lwproc_switch(NULL); 1107 free(pf); 1108 return; 1109 } 1110 1111 /* Ok, we have a new process context and a new curlwp */ 1112 rumpuser_getrandom(auth, sizeof(auth), 0, &randlen); 1113 memcpy(pf->pf_auth, auth, sizeof(pf->pf_auth)); 1114 pf->pf_lwp = lwproc_curlwp(); 1115 lwproc_switch(NULL); 1116 1117 pthread_mutex_lock(&pfmtx); 1118 LIST_INSERT_HEAD(&preforks, pf, pf_entries); 1119 LIST_INSERT_HEAD(&spc->spc_pflist, pf, pf_spcentries); 1120 pthread_mutex_unlock(&pfmtx); 1121 1122 DPRINTF(("rump_sp: prefork handler success %p\n", spc)); 1123 1124 send_prefork_resp(spc, reqno, auth); 1125 return; 1126 } 1127 1128 if (__predict_false(spc->spc_hdr.rsp_type == RUMPSP_HANDSHAKE)) { 1129 int inexec; 1130 1131 if (spc->spc_hdr.rsp_handshake != HANDSHAKE_EXEC) { 1132 send_error_resp(spc, reqno, 1133 RUMPSP_ERR_MALFORMED_REQUEST); 1134 shutdown(spc->spc_fd, SHUT_RDWR); 1135 spcfreebuf(spc); 1136 return; 1137 } 1138 1139 pthread_mutex_lock(&spc->spc_mtx); 1140 inexec = spc->spc_inexec; 1141 pthread_mutex_unlock(&spc->spc_mtx); 1142 if (inexec) { 1143 send_error_resp(spc, reqno, RUMPSP_ERR_INEXEC); 1144 shutdown(spc->spc_fd, SHUT_RDWR); 1145 spcfreebuf(spc); 1146 return; 1147 } 1148 1149 pthread_mutex_lock(&spc->spc_mtx); 1150 spc->spc_inexec = 1; 1151 pthread_mutex_unlock(&spc->spc_mtx); 1152 1153 /* 1154 * start to drain lwps. we will wait for it to finish 1155 * in another thread 1156 */ 1157 lwproc_switch(spc->spc_mainlwp); 1158 lwproc_lwpexit(); 1159 lwproc_switch(NULL); 1160 1161 /* 1162 * exec has to wait for lwps to drain, so finish it off 1163 * in another thread 1164 */ 1165 schedulework(spc, SBA_EXEC); 1166 return; 1167 } 1168 1169 if (__predict_false(spc->spc_hdr.rsp_type != RUMPSP_SYSCALL)) { 1170 send_error_resp(spc, reqno, RUMPSP_ERR_MALFORMED_REQUEST); 1171 spcfreebuf(spc); 1172 return; 1173 } 1174 1175 schedulework(spc, SBA_SYSCALL); 1176 } 1177 1178 static void * 1179 spserver(void *arg) 1180 { 1181 struct spservarg *sarg = arg; 1182 struct spclient *spc; 1183 unsigned idx; 1184 int seen; 1185 int rv; 1186 unsigned int nfds, maxidx; 1187 1188 for (idx = 0; idx < MAXCLI; idx++) { 1189 pfdlist[idx].fd = -1; 1190 pfdlist[idx].events = POLLIN; 1191 1192 spc = &spclist[idx]; 1193 pthread_mutex_init(&spc->spc_mtx, NULL); 1194 pthread_cond_init(&spc->spc_cv, NULL); 1195 spc->spc_fd = -1; 1196 } 1197 pfdlist[0].fd = spclist[0].spc_fd = sarg->sps_sock; 1198 pfdlist[0].events = POLLIN; 1199 nfds = 1; 1200 maxidx = 0; 1201 1202 pthread_attr_init(&pattr_detached); 1203 pthread_attr_setdetachstate(&pattr_detached, PTHREAD_CREATE_DETACHED); 1204 #if NOTYET 1205 pthread_attr_setstacksize(&pattr_detached, 32*1024); 1206 #endif 1207 1208 pthread_mutex_init(&sbamtx, NULL); 1209 pthread_cond_init(&sbacv, NULL); 1210 1211 DPRINTF(("rump_sp: server mainloop\n")); 1212 1213 for (;;) { 1214 int discoed; 1215 1216 /* g/c hangarounds (eventually) */ 1217 discoed = getdisco(); 1218 while (discoed--) { 1219 nfds--; 1220 idx = maxidx; 1221 while (idx) { 1222 if (pfdlist[idx].fd != -1) { 1223 maxidx = idx; 1224 break; 1225 } 1226 idx--; 1227 } 1228 DPRINTF(("rump_sp: set maxidx to [%u]\n", 1229 maxidx)); 1230 } 1231 1232 DPRINTF(("rump_sp: loop nfd %d\n", maxidx+1)); 1233 seen = 0; 1234 rv = poll(pfdlist, maxidx+1, INFTIM); 1235 assert(maxidx+1 <= MAXCLI); 1236 assert(rv != 0); 1237 if (rv == -1) { 1238 if (errno == EINTR) 1239 continue; 1240 fprintf(stderr, "rump_spserver: poll returned %d\n", 1241 errno); 1242 break; 1243 } 1244 1245 for (idx = 0; seen < rv && idx < MAXCLI; idx++) { 1246 if ((pfdlist[idx].revents & POLLIN) == 0) 1247 continue; 1248 1249 seen++; 1250 DPRINTF(("rump_sp: activity at [%u] %d/%d\n", 1251 idx, seen, rv)); 1252 if (idx > 0) { 1253 spc = &spclist[idx]; 1254 DPRINTF(("rump_sp: mainloop read [%u]\n", idx)); 1255 switch (readframe(spc)) { 1256 case 0: 1257 break; 1258 case -1: 1259 serv_handledisco(idx); 1260 break; 1261 default: 1262 switch (spc->spc_hdr.rsp_class) { 1263 case RUMPSP_RESP: 1264 kickwaiter(spc); 1265 break; 1266 case RUMPSP_REQ: 1267 handlereq(spc); 1268 break; 1269 default: 1270 send_error_resp(spc, 1271 spc->spc_hdr.rsp_reqno, 1272 RUMPSP_ERR_MALFORMED_REQUEST); 1273 spcfreebuf(spc); 1274 break; 1275 } 1276 break; 1277 } 1278 1279 } else { 1280 DPRINTF(("rump_sp: mainloop new connection\n")); 1281 1282 if (__predict_false(spfini)) { 1283 close(spclist[0].spc_fd); 1284 serv_shutdown(); 1285 goto out; 1286 } 1287 1288 idx = serv_handleconn(pfdlist[0].fd, 1289 sarg->sps_connhook, nfds == MAXCLI); 1290 if (idx) 1291 nfds++; 1292 if (idx > maxidx) 1293 maxidx = idx; 1294 DPRINTF(("rump_sp: maxid now %d\n", maxidx)); 1295 } 1296 } 1297 } 1298 1299 out: 1300 return NULL; 1301 } 1302 1303 static unsigned cleanupidx; 1304 static struct sockaddr *cleanupsa; 1305 int 1306 rumpuser_sp_init(const char *url, 1307 const char *ostype, const char *osrelease, const char *machine) 1308 { 1309 pthread_t pt; 1310 struct spservarg *sarg; 1311 struct sockaddr *sap; 1312 char *p; 1313 unsigned idx = 0; /* XXXgcc */ 1314 int error, s; 1315 1316 p = strdup(url); 1317 if (p == NULL) { 1318 error = ENOMEM; 1319 goto out; 1320 } 1321 error = parseurl(p, &sap, &idx, 1); 1322 free(p); 1323 if (error) 1324 goto out; 1325 1326 snprintf(banner, sizeof(banner), "RUMPSP-%d.%d-%s-%s/%s\n", 1327 PROTOMAJOR, PROTOMINOR, ostype, osrelease, machine); 1328 1329 s = socket(parsetab[idx].domain, SOCK_STREAM, 0); 1330 if (s == -1) { 1331 error = errno; 1332 goto out; 1333 } 1334 1335 sarg = malloc(sizeof(*sarg)); 1336 if (sarg == NULL) { 1337 close(s); 1338 error = ENOMEM; 1339 goto out; 1340 } 1341 1342 sarg->sps_sock = s; 1343 sarg->sps_connhook = parsetab[idx].connhook; 1344 1345 cleanupidx = idx; 1346 cleanupsa = sap; 1347 1348 /* sloppy error recovery */ 1349 1350 /*LINTED*/ 1351 if (bind(s, sap, parsetab[idx].slen) == -1) { 1352 error = errno; 1353 fprintf(stderr, "rump_sp: failed to bind to URL %s\n", url); 1354 goto out; 1355 } 1356 if (listen(s, MAXCLI) == -1) { 1357 error = errno; 1358 fprintf(stderr, "rump_sp: server listen failed\n"); 1359 goto out; 1360 } 1361 1362 if ((error = pthread_create(&pt, NULL, spserver, sarg)) != 0) { 1363 fprintf(stderr, "rump_sp: cannot create wrkr thread\n"); 1364 goto out; 1365 } 1366 pthread_detach(pt); 1367 1368 out: 1369 ET(error); 1370 } 1371 1372 void 1373 rumpuser_sp_fini(void *arg) 1374 { 1375 struct spclient *spc = arg; 1376 register_t retval[2] = {0, 0}; 1377 1378 if (spclist[0].spc_fd) { 1379 parsetab[cleanupidx].cleanup(cleanupsa); 1380 } 1381 1382 /* 1383 * stuff response into the socket, since the rump kernel container 1384 * is just about to exit 1385 */ 1386 if (spc && spc->spc_syscallreq) 1387 send_syscall_resp(spc, spc->spc_syscallreq, 0, retval); 1388 1389 if (spclist[0].spc_fd) { 1390 shutdown(spclist[0].spc_fd, SHUT_RDWR); 1391 spfini = 1; 1392 } 1393 1394 /* 1395 * could release thread, but don't bother, since the container 1396 * will be stone dead in a moment. 1397 */ 1398 } 1399