rump.c revision 1.161
1 /* $NetBSD: rump.c,v 1.161 2010/04/17 16:34:29 pooka Exp $ */ 2 3 /* 4 * Copyright (c) 2007 Antti Kantee. All Rights Reserved. 5 * 6 * Development of this software was supported by Google Summer of Code. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS 18 * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 19 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 20 * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 23 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __KERNEL_RCSID(0, "$NetBSD: rump.c,v 1.161 2010/04/17 16:34:29 pooka Exp $"); 32 33 #include <sys/param.h> 34 #include <sys/atomic.h> 35 #include <sys/buf.h> 36 #include <sys/callout.h> 37 #include <sys/conf.h> 38 #include <sys/cpu.h> 39 #include <sys/device.h> 40 #include <sys/evcnt.h> 41 #include <sys/event.h> 42 #include <sys/exec_elf.h> 43 #include <sys/filedesc.h> 44 #include <sys/iostat.h> 45 #include <sys/kauth.h> 46 #include <sys/kernel.h> 47 #include <sys/kmem.h> 48 #include <sys/kprintf.h> 49 #include <sys/ksyms.h> 50 #include <sys/msgbuf.h> 51 #include <sys/module.h> 52 #include <sys/once.h> 53 #include <sys/percpu.h> 54 #include <sys/pipe.h> 55 #include <sys/queue.h> 56 #include <sys/reboot.h> 57 #include <sys/resourcevar.h> 58 #include <sys/select.h> 59 #include <sys/sysctl.h> 60 #include <sys/syscall.h> 61 #include <sys/timetc.h> 62 #include <sys/tty.h> 63 #include <sys/uidinfo.h> 64 #include <sys/vmem.h> 65 #include <sys/xcall.h> 66 67 #include <rump/rumpuser.h> 68 69 #include <secmodel/suser/suser.h> 70 71 #include <prop/proplib.h> 72 73 #include <uvm/uvm_readahead.h> 74 75 #include "rump_private.h" 76 #include "rump_net_private.h" 77 #include "rump_vfs_private.h" 78 #include "rump_dev_private.h" 79 80 struct proc proc0; 81 struct session rump_session = { 82 .s_count = 1, 83 .s_flags = 0, 84 .s_leader = &proc0, 85 .s_login = "rumphobo", 86 .s_sid = 0, 87 }; 88 struct pgrp rump_pgrp = { 89 .pg_members = LIST_HEAD_INITIALIZER(pg_members), 90 .pg_session = &rump_session, 91 .pg_jobc = 1, 92 }; 93 struct pstats rump_stats; 94 struct plimit rump_limits; 95 struct filedesc rump_filedesc0; 96 struct proclist allproc; 97 char machine[] = "rump"; 98 static kauth_cred_t rump_susercred; 99 100 /* pretend the master rump proc is init */ 101 struct proc *initproc = &proc0; 102 103 struct rumpuser_mtx *rump_giantlock; 104 105 sigset_t sigcantmask; 106 107 struct device rump_rootdev = { 108 .dv_class = DV_VIRTUAL 109 }; 110 111 #ifdef RUMP_WITHOUT_THREADS 112 int rump_threads = 0; 113 #else 114 int rump_threads = 1; 115 #endif 116 117 static char rump_msgbuf[16*1024]; /* 16k should be enough for std rump needs */ 118 119 static void 120 rump_aiodone_worker(struct work *wk, void *dummy) 121 { 122 struct buf *bp = (struct buf *)wk; 123 124 KASSERT(&bp->b_work == wk); 125 bp->b_iodone(bp); 126 } 127 128 static int rump_inited; 129 130 int rump__unavailable(void); 131 int rump__unavailable() {return EOPNOTSUPP;} 132 __weak_alias(rump_net_init,rump__unavailable); 133 __weak_alias(rump_vfs_init,rump__unavailable); 134 __weak_alias(rump_dev_init,rump__unavailable); 135 136 __weak_alias(rump_vfs_fini,rump__unavailable); 137 138 __weak_alias(biodone,rump__unavailable); 139 __weak_alias(sopoll,rump__unavailable); 140 141 void rump__unavailable_vfs_panic(void); 142 void rump__unavailable_vfs_panic() {panic("vfs component not available");} 143 __weak_alias(usermount_common_policy,rump__unavailable_vfs_panic); 144 145 rump_proc_vfs_init_fn rump_proc_vfs_init; 146 rump_proc_vfs_release_fn rump_proc_vfs_release; 147 148 static void add_linkedin_modules(const struct modinfo *const *, size_t); 149 150 static void __noinline 151 messthestack(void) 152 { 153 volatile uint32_t mess[64]; 154 uint64_t d1, d2; 155 int i, error; 156 157 for (i = 0; i < 64; i++) { 158 rumpuser_gettime(&d1, &d2, &error); 159 mess[i] = d2; 160 } 161 } 162 163 /* 164 * Create kern.hostname. why only this you ask. well, init_sysctl 165 * is a kitchen sink in need of some gardening. but i want to use 166 * kern.hostname today. 167 */ 168 static void 169 mksysctls(void) 170 { 171 172 sysctl_createv(NULL, 0, NULL, NULL, 173 CTLFLAG_PERMANENT, CTLTYPE_NODE, "kern", NULL, 174 NULL, 0, NULL, 0, CTL_KERN, CTL_EOL); 175 176 /* XXX: setting hostnamelen is missing */ 177 sysctl_createv(NULL, 0, NULL, NULL, 178 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, CTLTYPE_STRING, "hostname", 179 SYSCTL_DESCR("System hostname"), NULL, 0, 180 &hostname, MAXHOSTNAMELEN, CTL_KERN, KERN_HOSTNAME, CTL_EOL); 181 } 182 183 int 184 rump__init(int rump_version) 185 { 186 char buf[256]; 187 struct timespec ts; 188 uint64_t sec, nsec; 189 struct proc *p; 190 struct lwp *l; 191 int i; 192 int error; 193 194 /* not reentrant */ 195 if (rump_inited) 196 return 0; 197 else if (rump_inited == -1) 198 panic("rump_init: host process restart required"); 199 else 200 rump_inited = 1; 201 202 if (rumpuser_getenv("RUMP_VERBOSE", buf, sizeof(buf), &error) == 0) { 203 if (*buf != '0') 204 boothowto = AB_VERBOSE; 205 } 206 207 rumpuser_gettime(&sec, &nsec, &error); 208 boottime.tv_sec = sec; 209 boottime.tv_nsec = nsec; 210 211 initmsgbuf(rump_msgbuf, sizeof(rump_msgbuf)); 212 aprint_verbose("%s%s", copyright, version); 213 214 /* 215 * Seed arc4random() with a "reasonable" amount of randomness. 216 * Yes, this is a quick kludge which depends on the arc4random 217 * implementation. 218 */ 219 messthestack(); 220 arc4random(); 221 222 if (rump_version != RUMP_VERSION) { 223 printf("rump version mismatch, %d vs. %d\n", 224 rump_version, RUMP_VERSION); 225 return EPROGMISMATCH; 226 } 227 228 if (rumpuser_getenv("RUMP_THREADS", buf, sizeof(buf), &error) == 0) { 229 rump_threads = *buf != '0'; 230 } 231 rumpuser_thrinit(rump_user_schedule, rump_user_unschedule, 232 rump_threads); 233 rump_intr_init(); 234 235 /* init minimal lwp/cpu context */ 236 l = &lwp0; 237 l->l_lid = 1; 238 l->l_cpu = rump_cpu; 239 rumpuser_set_curlwp(l); 240 241 mutex_init(&tty_lock, MUTEX_DEFAULT, IPL_NONE); 242 rumpuser_mutex_recursive_init(&rump_giantlock); 243 ksyms_init(); 244 rumpvm_init(); 245 evcnt_init(); 246 247 once_init(); 248 prop_kern_init(); 249 250 pool_subsystem_init(); 251 kmem_init(); 252 253 uvm_ra_init(); 254 255 mutex_obj_init(); 256 callout_startup(); 257 258 kprintf_init(); 259 loginit(); 260 261 kauth_init(); 262 rump_susercred = rump_cred_create(0, 0, 0, NULL); 263 264 /* init proc0 and rest of lwp0 now that we can allocate memory */ 265 p = &proc0; 266 p->p_stats = &rump_stats; 267 p->p_limit = &rump_limits; 268 p->p_pgrp = &rump_pgrp; 269 p->p_pid = 0; 270 p->p_fd = &rump_filedesc0; 271 p->p_vmspace = &rump_vmspace; 272 p->p_emul = &emul_netbsd; 273 p->p_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE); 274 l->l_cred = rump_cred_suserget(); 275 l->l_proc = p; 276 LIST_INIT(&allproc); 277 LIST_INSERT_HEAD(&allproc, &proc0, p_list); 278 proc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE); 279 lwpinit_specificdata(); 280 281 rump_limits.pl_rlimit[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY; 282 rump_limits.pl_rlimit[RLIMIT_NOFILE].rlim_cur = RLIM_INFINITY; 283 rump_limits.pl_rlimit[RLIMIT_SBSIZE].rlim_cur = RLIM_INFINITY; 284 285 rump_scheduler_init(); 286 /* revert temporary context and schedule a real context */ 287 l->l_cpu = NULL; 288 rumpuser_set_curlwp(NULL); 289 rump_schedule(); 290 291 percpu_init(); 292 inittimecounter(); 293 ntp_init(); 294 295 rumpuser_gettime(&sec, &nsec, &error); 296 ts.tv_sec = sec; 297 ts.tv_nsec = nsec; 298 tc_setclock(&ts); 299 300 /* we are mostly go. do per-cpu subsystem init */ 301 for (i = 0; i < ncpu; i++) { 302 struct cpu_info *ci = cpu_lookup(i); 303 304 callout_init_cpu(ci); 305 softint_init(ci); 306 xc_init_cpu(ci); 307 pool_cache_cpu_init(ci); 308 selsysinit(ci); 309 percpu_init_cpu(ci); 310 } 311 312 sysctl_init(); 313 kqueue_init(); 314 iostat_init(); 315 uid_init(); 316 fd_sys_init(); 317 module_init(); 318 devsw_init(); 319 pipe_init(); 320 321 rumpuser_dl_bootstrap(add_linkedin_modules, rump_kernelfsym_load); 322 323 /* these do nothing if not present */ 324 rump_vfs_init(); 325 rump_net_init(); 326 rump_dev_init(); 327 cold = 0; 328 329 /* aieeeedondest */ 330 if (rump_threads) { 331 if (workqueue_create(&uvm.aiodone_queue, "aiodoned", 332 rump_aiodone_worker, NULL, 0, 0, WQ_MPSAFE)) 333 panic("aiodoned"); 334 } 335 336 mksysctls(); 337 sysctl_finalize(); 338 339 module_init_class(MODULE_CLASS_ANY); 340 341 rumpuser_gethostname(hostname, MAXHOSTNAMELEN, &error); 342 hostnamelen = strlen(hostname); 343 344 sigemptyset(&sigcantmask); 345 346 lwp0.l_fd = proc0.p_fd = fd_init(&rump_filedesc0); 347 348 if (rump_threads) 349 vmem_rehash_start(); 350 351 rump_unschedule(); 352 353 return 0; 354 } 355 356 /* maybe support sys_reboot some day for remote shutdown */ 357 void 358 rump_reboot(int howto) 359 { 360 361 /* dump means we really take the dive here */ 362 if ((howto & RB_DUMP) || panicstr) { 363 rumpuser_exit(RUMPUSER_PANIC); 364 /*NOTREACHED*/ 365 } 366 367 /* try to sync */ 368 if (!((howto & RB_NOSYNC) || panicstr)) { 369 rump_vfs_fini(); 370 } 371 372 /* your wish is my command */ 373 if (howto & RB_HALT) { 374 for (;;) { 375 uint64_t sec = 5, nsec = 0; 376 int error; 377 378 rumpuser_nanosleep(&sec, &nsec, &error); 379 } 380 } 381 rump_inited = -1; 382 } 383 384 struct uio * 385 rump_uio_setup(void *buf, size_t bufsize, off_t offset, enum rump_uiorw rw) 386 { 387 struct uio *uio; 388 enum uio_rw uiorw; 389 390 switch (rw) { 391 case RUMPUIO_READ: 392 uiorw = UIO_READ; 393 break; 394 case RUMPUIO_WRITE: 395 uiorw = UIO_WRITE; 396 break; 397 default: 398 panic("%s: invalid rw %d", __func__, rw); 399 } 400 401 uio = kmem_alloc(sizeof(struct uio), KM_SLEEP); 402 uio->uio_iov = kmem_alloc(sizeof(struct iovec), KM_SLEEP); 403 404 uio->uio_iov->iov_base = buf; 405 uio->uio_iov->iov_len = bufsize; 406 407 uio->uio_iovcnt = 1; 408 uio->uio_offset = offset; 409 uio->uio_resid = bufsize; 410 uio->uio_rw = uiorw; 411 uio->uio_vmspace = UIO_VMSPACE_SYS; 412 413 return uio; 414 } 415 416 size_t 417 rump_uio_getresid(struct uio *uio) 418 { 419 420 return uio->uio_resid; 421 } 422 423 off_t 424 rump_uio_getoff(struct uio *uio) 425 { 426 427 return uio->uio_offset; 428 } 429 430 size_t 431 rump_uio_free(struct uio *uio) 432 { 433 size_t resid; 434 435 resid = uio->uio_resid; 436 kmem_free(uio->uio_iov, sizeof(*uio->uio_iov)); 437 kmem_free(uio, sizeof(*uio)); 438 439 return resid; 440 } 441 442 static pid_t nextpid = 1; 443 struct lwp * 444 rump_newproc_switch() 445 { 446 struct lwp *l; 447 pid_t mypid; 448 449 mypid = atomic_inc_uint_nv(&nextpid); 450 if (__predict_false(mypid == 0)) 451 mypid = atomic_inc_uint_nv(&nextpid); 452 453 l = rump_lwp_alloc(mypid, 0); 454 rump_lwp_switch(l); 455 456 return l; 457 } 458 459 struct lwp * 460 rump_lwp_alloc_and_switch(pid_t pid, lwpid_t lid) 461 { 462 struct lwp *l; 463 464 l = rump_lwp_alloc(pid, lid); 465 rump_lwp_switch(l); 466 467 return l; 468 } 469 470 struct lwp * 471 rump_lwp_alloc(pid_t pid, lwpid_t lid) 472 { 473 struct lwp *l; 474 struct proc *p; 475 476 l = kmem_zalloc(sizeof(*l), KM_SLEEP); 477 if (pid != 0) { 478 p = kmem_zalloc(sizeof(*p), KM_SLEEP); 479 if (rump_proc_vfs_init) 480 rump_proc_vfs_init(p); 481 p->p_stats = &rump_stats; 482 p->p_limit = &rump_limits; 483 p->p_pid = pid; 484 p->p_vmspace = &rump_vmspace; 485 p->p_emul = &emul_netbsd; 486 p->p_fd = fd_init(NULL); 487 p->p_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE); 488 l->l_cred = rump_cred_suserget(); 489 } else { 490 p = &proc0; 491 l->l_cred = rump_susercred; 492 } 493 494 l->l_proc = p; 495 l->l_lid = lid; 496 l->l_fd = p->p_fd; 497 l->l_cpu = NULL; 498 lwp_initspecific(l); 499 LIST_INSERT_HEAD(&alllwp, l, l_list); 500 501 return l; 502 } 503 504 void 505 rump_lwp_switch(struct lwp *newlwp) 506 { 507 struct lwp *l = curlwp; 508 509 rumpuser_set_curlwp(NULL); 510 newlwp->l_cpu = l->l_cpu; 511 newlwp->l_mutex = l->l_mutex; 512 l->l_mutex = NULL; 513 l->l_cpu = NULL; 514 rumpuser_set_curlwp(newlwp); 515 if (l->l_flag & LW_WEXIT) 516 rump_lwp_free(l); 517 } 518 519 /* XXX: this has effect only on non-pid0 lwps */ 520 void 521 rump_lwp_release(struct lwp *l) 522 { 523 struct proc *p; 524 525 p = l->l_proc; 526 if (p->p_pid != 0) { 527 mutex_obj_free(p->p_lock); 528 fd_free(); 529 if (rump_proc_vfs_release) 530 rump_proc_vfs_release(p); 531 rump_cred_put(l->l_cred); 532 kmem_free(p, sizeof(*p)); 533 } 534 KASSERT((l->l_flag & LW_WEXIT) == 0); 535 l->l_flag |= LW_WEXIT; 536 } 537 538 void 539 rump_lwp_free(struct lwp *l) 540 { 541 542 KASSERT(l->l_flag & LW_WEXIT); 543 KASSERT(l->l_mutex == NULL); 544 if (l->l_name) 545 kmem_free(l->l_name, MAXCOMLEN); 546 lwp_finispecific(l); 547 LIST_REMOVE(l, l_list); 548 kmem_free(l, sizeof(*l)); 549 } 550 551 struct lwp * 552 rump_lwp_curlwp(void) 553 { 554 struct lwp *l = curlwp; 555 556 if (l->l_flag & LW_WEXIT) 557 return NULL; 558 return l; 559 } 560 561 /* rump private. NEEDS WORK! */ 562 void 563 rump_set_vmspace(struct vmspace *vm) 564 { 565 struct proc *p = curproc; 566 567 p->p_vmspace = vm; 568 } 569 570 kauth_cred_t 571 rump_cred_create(uid_t uid, gid_t gid, size_t ngroups, gid_t *groups) 572 { 573 kauth_cred_t cred; 574 int rv; 575 576 cred = kauth_cred_alloc(); 577 kauth_cred_setuid(cred, uid); 578 kauth_cred_seteuid(cred, uid); 579 kauth_cred_setsvuid(cred, uid); 580 kauth_cred_setgid(cred, gid); 581 kauth_cred_setgid(cred, gid); 582 kauth_cred_setegid(cred, gid); 583 kauth_cred_setsvgid(cred, gid); 584 rv = kauth_cred_setgroups(cred, groups, ngroups, 0, UIO_SYSSPACE); 585 /* oh this is silly. and by "this" I mean kauth_cred_setgroups() */ 586 assert(rv == 0); 587 588 return cred; 589 } 590 591 void 592 rump_cred_put(kauth_cred_t cred) 593 { 594 595 kauth_cred_free(cred); 596 } 597 598 kauth_cred_t 599 rump_cred_suserget(void) 600 { 601 602 kauth_cred_hold(rump_susercred); 603 return rump_susercred; 604 } 605 606 /* 607 * Return the next system lwpid 608 */ 609 lwpid_t 610 rump_nextlid(void) 611 { 612 lwpid_t retid; 613 614 mutex_enter(proc0.p_lock); 615 /* 616 * Take next one, don't return 0 617 * XXX: most likely we'll have collisions in case this 618 * wraps around. 619 */ 620 if (++proc0.p_nlwpid == 0) 621 ++proc0.p_nlwpid; 622 retid = proc0.p_nlwpid; 623 mutex_exit(proc0.p_lock); 624 625 return retid; 626 } 627 628 static int compcounter[RUMP_COMPONENT_MAX]; 629 630 static void 631 rump_component_init_cb(struct rump_component *rc, int type) 632 { 633 634 KASSERT(type < RUMP_COMPONENT_MAX); 635 if (rc->rc_type == type) { 636 rc->rc_init(); 637 compcounter[type]++; 638 } 639 } 640 641 int 642 rump_component_count(enum rump_component_type type) 643 { 644 645 KASSERT(type <= RUMP_COMPONENT_MAX); 646 return compcounter[type]; 647 } 648 649 void 650 rump_component_init(enum rump_component_type type) 651 { 652 653 rumpuser_dl_component_init(type, rump_component_init_cb); 654 } 655 656 /* 657 * Initialize a module which has already been loaded and linked 658 * with dlopen(). This is fundamentally the same as a builtin module. 659 */ 660 int 661 rump_module_init(const struct modinfo * const *mip, size_t nmodinfo) 662 { 663 664 return module_builtin_add(mip, nmodinfo, true); 665 } 666 667 /* 668 * Finish module (flawless victory, fatality!). 669 */ 670 int 671 rump_module_fini(const struct modinfo *mi) 672 { 673 674 return module_builtin_remove(mi, true); 675 } 676 677 /* 678 * Add loaded and linked module to the builtin list. It will 679 * later be initialized with module_init_class(). 680 */ 681 682 static void 683 add_linkedin_modules(const struct modinfo * const *mip, size_t nmodinfo) 684 { 685 686 module_builtin_add(mip, nmodinfo, false); 687 } 688 689 int 690 rump_kernelfsym_load(void *symtab, uint64_t symsize, 691 char *strtab, uint64_t strsize) 692 { 693 static int inited = 0; 694 Elf64_Ehdr ehdr; 695 696 if (inited) 697 return EBUSY; 698 inited = 1; 699 700 /* 701 * Use 64bit header since it's bigger. Shouldn't make a 702 * difference, since we're passing in all zeroes anyway. 703 */ 704 memset(&ehdr, 0, sizeof(ehdr)); 705 ksyms_addsyms_explicit(&ehdr, symtab, symsize, strtab, strsize); 706 707 return 0; 708 } 709 710 static int 711 rump_sysproxy_local(int num, void *arg, uint8_t *data, size_t dlen, 712 register_t *retval) 713 { 714 struct lwp *l; 715 struct sysent *callp; 716 int rv; 717 718 if (__predict_false(num >= SYS_NSYSENT)) 719 return ENOSYS; 720 721 callp = rump_sysent + num; 722 rump_schedule(); 723 l = curlwp; 724 rv = callp->sy_call(l, (void *)data, retval); 725 rump_unschedule(); 726 727 return rv; 728 } 729 730 int 731 rump_boot_gethowto() 732 { 733 734 return boothowto; 735 } 736 737 void 738 rump_boot_sethowto(int howto) 739 { 740 741 boothowto = howto; 742 } 743 744 rump_sysproxy_t rump_sysproxy = rump_sysproxy_local; 745 void *rump_sysproxy_arg; 746 747 /* 748 * This whole syscall-via-rpc is still taking form. For example, it 749 * may be necessary to set syscalls individually instead of lobbing 750 * them all to the same place. So don't think this interface is 751 * set in stone. 752 */ 753 int 754 rump_sysproxy_set(rump_sysproxy_t proxy, void *arg) 755 { 756 757 if (rump_sysproxy_arg) 758 return EBUSY; 759 760 rump_sysproxy_arg = arg; 761 rump_sysproxy = proxy; 762 763 return 0; 764 } 765 766 int 767 rump_getversion(void) 768 { 769 770 return __NetBSD_Version__; 771 } 772
Indexes created Tue Sep 30 06:09:59 GMT 2025