1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice(s), this list of conditions and the following disclaimer as 12 * the first lines of this file unmodified other than the possible 13 * addition of one or more copyright notices. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice(s), this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY 19 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 20 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 21 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY 22 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 25 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH 28 * DAMAGE. 29 */ 30 31 #include "opt_witness.h" 32 #include "opt_hwpmc_hooks.h" 33 #include "opt_hwt_hooks.h" 34 35 #include <sys/systm.h> 36 #include <sys/asan.h> 37 #include <sys/kernel.h> 38 #include <sys/lock.h> 39 #include <sys/msan.h> 40 #include <sys/mutex.h> 41 #include <sys/proc.h> 42 #include <sys/bitstring.h> 43 #include <sys/epoch.h> 44 #include <sys/rangelock.h> 45 #include <sys/resourcevar.h> 46 #include <sys/sdt.h> 47 #include <sys/smp.h> 48 #include <sys/sched.h> 49 #include <sys/sleepqueue.h> 50 #include <sys/selinfo.h> 51 #include <sys/syscallsubr.h> 52 #include <sys/dtrace_bsd.h> 53 #include <sys/sysent.h> 54 #include <sys/turnstile.h> 55 #include <sys/taskqueue.h> 56 #include <sys/ktr.h> 57 #include <sys/rwlock.h> 58 #include <sys/umtxvar.h> 59 #include <sys/vmmeter.h> 60 #include <sys/cpuset.h> 61 #ifdef HWPMC_HOOKS 62 #include <sys/pmckern.h> 63 #endif 64 #ifdef HWT_HOOKS 65 #include <dev/hwt/hwt_hook.h> 66 #endif 67 #include <sys/priv.h> 68 69 #include <security/audit/audit.h> 70 71 #include <vm/pmap.h> 72 #include <vm/vm.h> 73 #include <vm/vm_extern.h> 74 #include <vm/uma.h> 75 #include <vm/vm_phys.h> 76 #include <sys/eventhandler.h> 77 78 /* 79 * Asserts below verify the stability of struct thread and struct proc 80 * layout, as exposed by KBI to modules. On head, the KBI is allowed 81 * to drift, change to the structures must be accompanied by the 82 * assert update. 83 * 84 * On the stable branches after KBI freeze, conditions must not be 85 * violated. Typically new fields are moved to the end of the 86 * structures. 87 */ 88 #ifdef __amd64__ 89 _Static_assert(offsetof(struct thread, td_flags) == 0x108, 90 "struct thread KBI td_flags"); 91 _Static_assert(offsetof(struct thread, td_pflags) == 0x114, 92 "struct thread KBI td_pflags"); 93 _Static_assert(offsetof(struct thread, td_frame) == 0x4e8, 94 "struct thread KBI td_frame"); 95 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6f0, 96 "struct thread KBI td_emuldata"); 97 _Static_assert(offsetof(struct proc, p_flag) == 0xb8, 98 "struct proc KBI p_flag"); 99 _Static_assert(offsetof(struct proc, p_pid) == 0xc4, 100 "struct proc KBI p_pid"); 101 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8, 102 "struct proc KBI p_filemon"); 103 _Static_assert(offsetof(struct proc, p_comm) == 0x3e0, 104 "struct proc KBI p_comm"); 105 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4d0, 106 "struct proc KBI p_emuldata"); 107 #endif 108 #ifdef __i386__ 109 _Static_assert(offsetof(struct thread, td_flags) == 0x9c, 110 "struct thread KBI td_flags"); 111 _Static_assert(offsetof(struct thread, td_pflags) == 0xa8, 112 "struct thread KBI td_pflags"); 113 _Static_assert(offsetof(struct thread, td_frame) == 0x33c, 114 "struct thread KBI td_frame"); 115 _Static_assert(offsetof(struct thread, td_emuldata) == 0x380, 116 "struct thread KBI td_emuldata"); 117 _Static_assert(offsetof(struct proc, p_flag) == 0x6c, 118 "struct proc KBI p_flag"); 119 _Static_assert(offsetof(struct proc, p_pid) == 0x78, 120 "struct proc KBI p_pid"); 121 _Static_assert(offsetof(struct proc, p_filemon) == 0x270, 122 "struct proc KBI p_filemon"); 123 _Static_assert(offsetof(struct proc, p_comm) == 0x284, 124 "struct proc KBI p_comm"); 125 _Static_assert(offsetof(struct proc, p_emuldata) == 0x318, 126 "struct proc KBI p_emuldata"); 127 #endif 128 129 SDT_PROVIDER_DECLARE(proc); 130 SDT_PROBE_DEFINE(proc, , , lwp__exit); 131 132 /* 133 * thread related storage. 134 */ 135 static uma_zone_t thread_zone; 136 137 struct thread_domain_data { 138 struct thread *tdd_zombies; 139 int tdd_reapticks; 140 } __aligned(CACHE_LINE_SIZE); 141 142 static struct thread_domain_data thread_domain_data[MAXMEMDOM]; 143 144 static struct task thread_reap_task; 145 static struct callout thread_reap_callout; 146 147 static void thread_zombie(struct thread *); 148 static void thread_reap(void); 149 static void thread_reap_all(void); 150 static void thread_reap_task_cb(void *, int); 151 static void thread_reap_callout_cb(void *); 152 static void thread_unsuspend_one(struct thread *td, struct proc *p, 153 bool boundary); 154 static void thread_free_batched(struct thread *td); 155 156 static __exclusive_cache_line struct mtx tid_lock; 157 static bitstr_t *tid_bitmap; 158 159 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash"); 160 161 static int maxthread; 162 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN, 163 &maxthread, 0, "Maximum number of threads"); 164 165 static __exclusive_cache_line int nthreads; 166 167 static LIST_HEAD(tidhashhead, thread) *tidhashtbl; 168 static u_long tidhash; 169 static u_long tidhashlock; 170 static struct rwlock *tidhashtbl_lock; 171 #define TIDHASH(tid) (&tidhashtbl[(tid) & tidhash]) 172 #define TIDHASHLOCK(tid) (&tidhashtbl_lock[(tid) & tidhashlock]) 173 174 EVENTHANDLER_LIST_DEFINE(thread_ctor); 175 EVENTHANDLER_LIST_DEFINE(thread_dtor); 176 EVENTHANDLER_LIST_DEFINE(thread_init); 177 EVENTHANDLER_LIST_DEFINE(thread_fini); 178 179 static bool 180 thread_count_inc_try(void) 181 { 182 int nthreads_new; 183 184 nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1; 185 if (nthreads_new >= maxthread - 100) { 186 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 || 187 nthreads_new >= maxthread) { 188 atomic_subtract_int(&nthreads, 1); 189 return (false); 190 } 191 } 192 return (true); 193 } 194 195 static bool 196 thread_count_inc(void) 197 { 198 static struct timeval lastfail; 199 static int curfail; 200 201 thread_reap(); 202 if (thread_count_inc_try()) { 203 return (true); 204 } 205 206 thread_reap_all(); 207 if (thread_count_inc_try()) { 208 return (true); 209 } 210 211 if (ppsratecheck(&lastfail, &curfail, 1)) { 212 printf("maxthread limit exceeded by uid %u " 213 "(pid %d); consider increasing kern.maxthread\n", 214 curthread->td_ucred->cr_ruid, curproc->p_pid); 215 } 216 return (false); 217 } 218 219 static void 220 thread_count_sub(int n) 221 { 222 223 atomic_subtract_int(&nthreads, n); 224 } 225 226 static void 227 thread_count_dec(void) 228 { 229 230 thread_count_sub(1); 231 } 232 233 static lwpid_t 234 tid_alloc(void) 235 { 236 static lwpid_t trytid; 237 lwpid_t tid; 238 239 mtx_lock(&tid_lock); 240 /* 241 * It is an invariant that the bitmap is big enough to hold maxthread 242 * IDs. If we got to this point there has to be at least one free. 243 */ 244 if (trytid >= maxthread) 245 trytid = 0; 246 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid); 247 if (tid == -1) { 248 KASSERT(trytid != 0, ("unexpectedly ran out of IDs")); 249 trytid = 0; 250 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid); 251 KASSERT(tid != -1, ("unexpectedly ran out of IDs")); 252 } 253 bit_set(tid_bitmap, tid); 254 trytid = tid + 1; 255 mtx_unlock(&tid_lock); 256 return (tid + NO_PID); 257 } 258 259 static void 260 tid_free_locked(lwpid_t rtid) 261 { 262 lwpid_t tid; 263 264 mtx_assert(&tid_lock, MA_OWNED); 265 KASSERT(rtid >= NO_PID, 266 ("%s: invalid tid %d\n", __func__, rtid)); 267 tid = rtid - NO_PID; 268 KASSERT(bit_test(tid_bitmap, tid) != 0, 269 ("thread ID %d not allocated\n", rtid)); 270 bit_clear(tid_bitmap, tid); 271 } 272 273 static void 274 tid_free(lwpid_t rtid) 275 { 276 277 mtx_lock(&tid_lock); 278 tid_free_locked(rtid); 279 mtx_unlock(&tid_lock); 280 } 281 282 static void 283 tid_free_batch(lwpid_t *batch, int n) 284 { 285 int i; 286 287 mtx_lock(&tid_lock); 288 for (i = 0; i < n; i++) { 289 tid_free_locked(batch[i]); 290 } 291 mtx_unlock(&tid_lock); 292 } 293 294 /* 295 * Batching for thread reapping. 296 */ 297 struct tidbatch { 298 lwpid_t tab[16]; 299 int n; 300 }; 301 302 static void 303 tidbatch_prep(struct tidbatch *tb) 304 { 305 306 tb->n = 0; 307 } 308 309 static void 310 tidbatch_add(struct tidbatch *tb, struct thread *td) 311 { 312 313 KASSERT(tb->n < nitems(tb->tab), 314 ("%s: count too high %d", __func__, tb->n)); 315 tb->tab[tb->n] = td->td_tid; 316 tb->n++; 317 } 318 319 static void 320 tidbatch_process(struct tidbatch *tb) 321 { 322 323 KASSERT(tb->n <= nitems(tb->tab), 324 ("%s: count too high %d", __func__, tb->n)); 325 if (tb->n == nitems(tb->tab)) { 326 tid_free_batch(tb->tab, tb->n); 327 tb->n = 0; 328 } 329 } 330 331 static void 332 tidbatch_final(struct tidbatch *tb) 333 { 334 335 KASSERT(tb->n <= nitems(tb->tab), 336 ("%s: count too high %d", __func__, tb->n)); 337 if (tb->n != 0) { 338 tid_free_batch(tb->tab, tb->n); 339 } 340 } 341 342 /* 343 * Batching thread count free, for consistency 344 */ 345 struct tdcountbatch { 346 int n; 347 }; 348 349 static void 350 tdcountbatch_prep(struct tdcountbatch *tb) 351 { 352 353 tb->n = 0; 354 } 355 356 static void 357 tdcountbatch_add(struct tdcountbatch *tb, struct thread *td __unused) 358 { 359 360 tb->n++; 361 } 362 363 static void 364 tdcountbatch_process(struct tdcountbatch *tb) 365 { 366 367 if (tb->n == 32) { 368 thread_count_sub(tb->n); 369 tb->n = 0; 370 } 371 } 372 373 static void 374 tdcountbatch_final(struct tdcountbatch *tb) 375 { 376 377 if (tb->n != 0) { 378 thread_count_sub(tb->n); 379 } 380 } 381 382 /* 383 * Prepare a thread for use. 384 */ 385 static int 386 thread_ctor(void *mem, int size, void *arg, int flags) 387 { 388 struct thread *td; 389 390 td = (struct thread *)mem; 391 TD_SET_STATE(td, TDS_INACTIVE); 392 td->td_lastcpu = td->td_oncpu = NOCPU; 393 394 /* 395 * Note that td_critnest begins life as 1 because the thread is not 396 * running and is thereby implicitly waiting to be on the receiving 397 * end of a context switch. 398 */ 399 td->td_critnest = 1; 400 td->td_lend_user_pri = PRI_MAX; 401 #ifdef AUDIT 402 audit_thread_alloc(td); 403 #endif 404 #ifdef KDTRACE_HOOKS 405 kdtrace_thread_ctor(td); 406 #endif 407 umtx_thread_alloc(td); 408 MPASS(td->td_sel == NULL); 409 return (0); 410 } 411 412 /* 413 * Reclaim a thread after use. 414 */ 415 static void 416 thread_dtor(void *mem, int size, void *arg) 417 { 418 struct thread *td; 419 420 td = (struct thread *)mem; 421 422 #ifdef INVARIANTS 423 /* Verify that this thread is in a safe state to free. */ 424 switch (TD_GET_STATE(td)) { 425 case TDS_INHIBITED: 426 case TDS_RUNNING: 427 case TDS_CAN_RUN: 428 case TDS_RUNQ: 429 /* 430 * We must never unlink a thread that is in one of 431 * these states, because it is currently active. 432 */ 433 panic("bad state for thread unlinking"); 434 /* NOTREACHED */ 435 case TDS_INACTIVE: 436 break; 437 default: 438 panic("bad thread state"); 439 /* NOTREACHED */ 440 } 441 #endif 442 #ifdef AUDIT 443 audit_thread_free(td); 444 #endif 445 #ifdef KDTRACE_HOOKS 446 kdtrace_thread_dtor(td); 447 #endif 448 /* Free all OSD associated to this thread. */ 449 osd_thread_exit(td); 450 ast_kclear(td); 451 seltdfini(td); 452 } 453 454 /* 455 * Initialize type-stable parts of a thread (when newly created). 456 */ 457 static int 458 thread_init(void *mem, int size, int flags) 459 { 460 struct thread *td; 461 462 td = (struct thread *)mem; 463 464 td->td_allocdomain = vm_phys_domain(vtophys(td)); 465 td->td_sleepqueue = sleepq_alloc(); 466 td->td_turnstile = turnstile_alloc(); 467 EVENTHANDLER_DIRECT_INVOKE(thread_init, td); 468 umtx_thread_init(td); 469 td->td_kstack = 0; 470 td->td_sel = NULL; 471 return (0); 472 } 473 474 /* 475 * Tear down type-stable parts of a thread (just before being discarded). 476 */ 477 static void 478 thread_fini(void *mem, int size) 479 { 480 struct thread *td; 481 482 td = (struct thread *)mem; 483 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td); 484 turnstile_free(td->td_turnstile); 485 sleepq_free(td->td_sleepqueue); 486 umtx_thread_fini(td); 487 MPASS(td->td_sel == NULL); 488 } 489 490 /* 491 * For a newly created process, 492 * link up all the structures and its initial threads etc. 493 * called from: 494 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc. 495 * proc_dtor() (should go away) 496 * proc_init() 497 */ 498 void 499 proc_linkup0(struct proc *p, struct thread *td) 500 { 501 TAILQ_INIT(&p->p_threads); /* all threads in proc */ 502 proc_linkup(p, td); 503 } 504 505 void 506 proc_linkup(struct proc *p, struct thread *td) 507 { 508 509 sigqueue_init(&p->p_sigqueue, p); 510 p->p_ksi = ksiginfo_alloc(M_WAITOK); 511 if (p->p_ksi != NULL) { 512 /* XXX p_ksi may be null if ksiginfo zone is not ready */ 513 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS; 514 } 515 LIST_INIT(&p->p_mqnotifier); 516 p->p_numthreads = 0; 517 thread_link(td, p); 518 } 519 520 static void 521 ast_suspend(struct thread *td, int tda __unused) 522 { 523 struct proc *p; 524 525 p = td->td_proc; 526 /* 527 * We need to check to see if we have to exit or wait due to a 528 * single threading requirement or some other STOP condition. 529 */ 530 PROC_LOCK(p); 531 thread_suspend_check(0); 532 PROC_UNLOCK(p); 533 } 534 535 extern int max_threads_per_proc; 536 537 /* 538 * Initialize global thread allocation resources. 539 */ 540 void 541 threadinit(void) 542 { 543 u_long i; 544 lwpid_t tid0; 545 546 /* 547 * Place an upper limit on threads which can be allocated. 548 * 549 * Note that other factors may make the de facto limit much lower. 550 * 551 * Platform limits are somewhat arbitrary but deemed "more than good 552 * enough" for the foreseable future. 553 */ 554 if (maxthread == 0) { 555 #ifdef _LP64 556 maxthread = MIN(maxproc * max_threads_per_proc, 1000000); 557 #else 558 maxthread = MIN(maxproc * max_threads_per_proc, 100000); 559 #endif 560 } 561 562 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF); 563 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK); 564 /* 565 * Handle thread0. 566 */ 567 thread_count_inc(); 568 tid0 = tid_alloc(); 569 if (tid0 != THREAD0_TID) 570 panic("tid0 %d != %d\n", tid0, THREAD0_TID); 571 572 /* 573 * Thread structures are specially aligned so that (at least) the 574 * 5 lower bits of a pointer to 'struct thread' must be 0. These bits 575 * are used by synchronization primitives to store flags in pointers to 576 * such structures. 577 */ 578 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(), 579 thread_ctor, thread_dtor, thread_init, thread_fini, 580 UMA_ALIGN_CACHE_AND_MASK(32 - 1), UMA_ZONE_NOFREE); 581 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash); 582 tidhashlock = (tidhash + 1) / 64; 583 if (tidhashlock > 0) 584 tidhashlock--; 585 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1), 586 M_TIDHASH, M_WAITOK | M_ZERO); 587 for (i = 0; i < tidhashlock + 1; i++) 588 rw_init(&tidhashtbl_lock[i], "tidhash"); 589 590 TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL); 591 callout_init(&thread_reap_callout, 1); 592 callout_reset(&thread_reap_callout, 5 * hz, 593 thread_reap_callout_cb, NULL); 594 ast_register(TDA_SUSPEND, ASTR_ASTF_REQUIRED, 0, ast_suspend); 595 } 596 597 /* 598 * Place an unused thread on the zombie list. 599 */ 600 void 601 thread_zombie(struct thread *td) 602 { 603 struct thread_domain_data *tdd; 604 struct thread *ztd; 605 606 tdd = &thread_domain_data[td->td_allocdomain]; 607 ztd = atomic_load_ptr(&tdd->tdd_zombies); 608 for (;;) { 609 td->td_zombie = ztd; 610 if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies, 611 (uintptr_t *)&ztd, (uintptr_t)td)) 612 break; 613 continue; 614 } 615 } 616 617 /* 618 * Release a thread that has exited after cpu_throw(). 619 */ 620 void 621 thread_stash(struct thread *td) 622 { 623 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1); 624 thread_zombie(td); 625 } 626 627 /* 628 * Reap zombies from passed domain. 629 */ 630 static void 631 thread_reap_domain(struct thread_domain_data *tdd) 632 { 633 struct thread *itd, *ntd; 634 struct tidbatch tidbatch; 635 struct credbatch credbatch; 636 struct limbatch limbatch; 637 struct tdcountbatch tdcountbatch; 638 639 /* 640 * Reading upfront is pessimal if followed by concurrent atomic_swap, 641 * but most of the time the list is empty. 642 */ 643 if (tdd->tdd_zombies == NULL) 644 return; 645 646 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies, 647 (uintptr_t)NULL); 648 if (itd == NULL) 649 return; 650 651 /* 652 * Multiple CPUs can get here, the race is fine as ticks is only 653 * advisory. 654 */ 655 tdd->tdd_reapticks = ticks; 656 657 tidbatch_prep(&tidbatch); 658 credbatch_prep(&credbatch); 659 limbatch_prep(&limbatch); 660 tdcountbatch_prep(&tdcountbatch); 661 662 while (itd != NULL) { 663 ntd = itd->td_zombie; 664 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd); 665 666 tidbatch_add(&tidbatch, itd); 667 credbatch_add(&credbatch, itd); 668 limbatch_add(&limbatch, itd); 669 tdcountbatch_add(&tdcountbatch, itd); 670 671 thread_free_batched(itd); 672 673 tidbatch_process(&tidbatch); 674 credbatch_process(&credbatch); 675 limbatch_process(&limbatch); 676 tdcountbatch_process(&tdcountbatch); 677 678 itd = ntd; 679 } 680 681 tidbatch_final(&tidbatch); 682 credbatch_final(&credbatch); 683 limbatch_final(&limbatch); 684 tdcountbatch_final(&tdcountbatch); 685 } 686 687 /* 688 * Reap zombies from all domains. 689 */ 690 static void 691 thread_reap_all(void) 692 { 693 struct thread_domain_data *tdd; 694 int i, domain; 695 696 domain = PCPU_GET(domain); 697 for (i = 0; i < vm_ndomains; i++) { 698 tdd = &thread_domain_data[(i + domain) % vm_ndomains]; 699 thread_reap_domain(tdd); 700 } 701 } 702 703 /* 704 * Reap zombies from local domain. 705 */ 706 static void 707 thread_reap(void) 708 { 709 struct thread_domain_data *tdd; 710 int domain; 711 712 domain = PCPU_GET(domain); 713 tdd = &thread_domain_data[domain]; 714 715 thread_reap_domain(tdd); 716 } 717 718 static void 719 thread_reap_task_cb(void *arg __unused, int pending __unused) 720 { 721 722 thread_reap_all(); 723 } 724 725 static void 726 thread_reap_callout_cb(void *arg __unused) 727 { 728 struct thread_domain_data *tdd; 729 int i, cticks, lticks; 730 bool wantreap; 731 732 wantreap = false; 733 cticks = atomic_load_int(&ticks); 734 for (i = 0; i < vm_ndomains; i++) { 735 tdd = &thread_domain_data[i]; 736 lticks = tdd->tdd_reapticks; 737 if (tdd->tdd_zombies != NULL && 738 (u_int)(cticks - lticks) > 5 * hz) { 739 wantreap = true; 740 break; 741 } 742 } 743 744 if (wantreap) 745 taskqueue_enqueue(taskqueue_thread, &thread_reap_task); 746 callout_reset(&thread_reap_callout, 5 * hz, 747 thread_reap_callout_cb, NULL); 748 } 749 750 /* 751 * Calling this function guarantees that any thread that exited before 752 * the call is reaped when the function returns. By 'exited' we mean 753 * a thread removed from the process linkage with thread_unlink(). 754 * Practically this means that caller must lock/unlock corresponding 755 * process lock before the call, to synchronize with thread_exit(). 756 */ 757 void 758 thread_reap_barrier(void) 759 { 760 struct task *t; 761 762 /* 763 * First do context switches to each CPU to ensure that all 764 * PCPU pc_deadthreads are moved to zombie list. 765 */ 766 quiesce_all_cpus("", PDROP); 767 768 /* 769 * Second, fire the task in the same thread as normal 770 * thread_reap() is done, to serialize reaping. 771 */ 772 t = malloc(sizeof(*t), M_TEMP, M_WAITOK); 773 TASK_INIT(t, 0, thread_reap_task_cb, t); 774 taskqueue_enqueue(taskqueue_thread, t); 775 taskqueue_drain(taskqueue_thread, t); 776 free(t, M_TEMP); 777 } 778 779 /* 780 * Allocate a thread. 781 */ 782 struct thread * 783 thread_alloc(int pages) 784 { 785 struct thread *td; 786 lwpid_t tid; 787 788 if (!thread_count_inc()) { 789 return (NULL); 790 } 791 792 tid = tid_alloc(); 793 td = uma_zalloc(thread_zone, M_WAITOK); 794 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack")); 795 if (!vm_thread_new(td, pages)) { 796 uma_zfree(thread_zone, td); 797 tid_free(tid); 798 thread_count_dec(); 799 return (NULL); 800 } 801 td->td_tid = tid; 802 bzero(&td->td_sa.args, sizeof(td->td_sa.args)); 803 kasan_thread_alloc(td); 804 kmsan_thread_alloc(td); 805 cpu_thread_alloc(td); 806 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td); 807 return (td); 808 } 809 810 int 811 thread_recycle(struct thread *td, int pages) 812 { 813 if (td->td_kstack == 0 || td->td_kstack_pages != pages) { 814 if (td->td_kstack != 0) 815 vm_thread_dispose(td); 816 if (!vm_thread_new(td, pages)) 817 return (ENOMEM); 818 cpu_thread_alloc(td); 819 } 820 kasan_thread_alloc(td); 821 kmsan_thread_alloc(td); 822 return (0); 823 } 824 825 /* 826 * Deallocate a thread. 827 */ 828 static void 829 thread_free_batched(struct thread *td) 830 { 831 832 lock_profile_thread_exit(td); 833 if (td->td_cpuset) 834 cpuset_rel(td->td_cpuset); 835 td->td_cpuset = NULL; 836 cpu_thread_free(td); 837 if (td->td_kstack != 0) 838 vm_thread_dispose(td); 839 callout_drain(&td->td_slpcallout); 840 /* 841 * Freeing handled by the caller. 842 */ 843 td->td_tid = -1; 844 kmsan_thread_free(td); 845 uma_zfree(thread_zone, td); 846 } 847 848 void 849 thread_free(struct thread *td) 850 { 851 lwpid_t tid; 852 853 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td); 854 tid = td->td_tid; 855 thread_free_batched(td); 856 tid_free(tid); 857 thread_count_dec(); 858 } 859 860 void 861 thread_cow_get_proc(struct thread *newtd, struct proc *p) 862 { 863 864 PROC_LOCK_ASSERT(p, MA_OWNED); 865 newtd->td_realucred = crcowget(p->p_ucred); 866 newtd->td_ucred = newtd->td_realucred; 867 newtd->td_limit = lim_hold(p->p_limit); 868 newtd->td_cowgen = p->p_cowgen; 869 } 870 871 void 872 thread_cow_get(struct thread *newtd, struct thread *td) 873 { 874 875 MPASS(td->td_realucred == td->td_ucred); 876 newtd->td_realucred = crcowget(td->td_realucred); 877 newtd->td_ucred = newtd->td_realucred; 878 newtd->td_limit = lim_hold(td->td_limit); 879 newtd->td_cowgen = td->td_cowgen; 880 } 881 882 void 883 thread_cow_free(struct thread *td) 884 { 885 886 if (td->td_realucred != NULL) 887 crcowfree(td); 888 if (td->td_limit != NULL) 889 lim_free(td->td_limit); 890 } 891 892 void 893 thread_cow_update(struct thread *td) 894 { 895 struct proc *p; 896 struct ucred *oldcred; 897 struct plimit *oldlimit; 898 899 p = td->td_proc; 900 PROC_LOCK(p); 901 oldcred = crcowsync(); 902 oldlimit = lim_cowsync(); 903 td->td_cowgen = p->p_cowgen; 904 PROC_UNLOCK(p); 905 if (oldcred != NULL) 906 crfree(oldcred); 907 if (oldlimit != NULL) 908 lim_free(oldlimit); 909 } 910 911 void 912 thread_cow_synced(struct thread *td) 913 { 914 struct proc *p; 915 916 p = td->td_proc; 917 PROC_LOCK_ASSERT(p, MA_OWNED); 918 MPASS(td->td_cowgen != p->p_cowgen); 919 MPASS(td->td_ucred == p->p_ucred); 920 MPASS(td->td_limit == p->p_limit); 921 td->td_cowgen = p->p_cowgen; 922 } 923 924 /* 925 * Discard the current thread and exit from its context. 926 * Always called with scheduler locked. 927 * 928 * Because we can't free a thread while we're operating under its context, 929 * push the current thread into our CPU's deadthread holder. This means 930 * we needn't worry about someone else grabbing our context before we 931 * do a cpu_throw(). 932 */ 933 void 934 thread_exit(void) 935 { 936 uint64_t runtime, new_switchtime; 937 struct thread *td; 938 struct thread *td2; 939 struct proc *p; 940 941 td = curthread; 942 p = td->td_proc; 943 944 PROC_SLOCK_ASSERT(p, MA_OWNED); 945 mtx_assert(&Giant, MA_NOTOWNED); 946 947 PROC_LOCK_ASSERT(p, MA_OWNED); 948 KASSERT(p != NULL, ("thread exiting without a process")); 949 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td, 950 (long)p->p_pid, td->td_name); 951 SDT_PROBE0(proc, , , lwp__exit); 952 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending")); 953 MPASS(td->td_realucred == td->td_ucred); 954 955 /* 956 * drop FPU & debug register state storage, or any other 957 * architecture specific resources that 958 * would not be on a new untouched process. 959 */ 960 cpu_thread_exit(td); 961 962 /* 963 * The last thread is left attached to the process 964 * So that the whole bundle gets recycled. Skip 965 * all this stuff if we never had threads. 966 * EXIT clears all sign of other threads when 967 * it goes to single threading, so the last thread always 968 * takes the short path. 969 */ 970 if (p->p_flag & P_HADTHREADS) { 971 if (p->p_numthreads > 1) { 972 atomic_add_int(&td->td_proc->p_exitthreads, 1); 973 thread_unlink(td); 974 td2 = FIRST_THREAD_IN_PROC(p); 975 sched_exit_thread(td2, td); 976 977 /* 978 * The test below is NOT true if we are the 979 * sole exiting thread. P_STOPPED_SINGLE is unset 980 * in exit1() after it is the only survivor. 981 */ 982 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 983 if (p->p_numthreads == p->p_suspcount) { 984 thread_lock(p->p_singlethread); 985 thread_unsuspend_one(p->p_singlethread, 986 p, false); 987 } 988 } 989 990 PCPU_SET(deadthread, td); 991 } else { 992 /* 993 * The last thread is exiting.. but not through exit() 994 */ 995 panic ("thread_exit: Last thread exiting on its own"); 996 } 997 } 998 #ifdef HWPMC_HOOKS 999 /* 1000 * If this thread is part of a process that is being tracked by hwpmc(4), 1001 * inform the module of the thread's impending exit. 1002 */ 1003 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) { 1004 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1005 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL); 1006 } else if (PMC_SYSTEM_SAMPLING_ACTIVE()) 1007 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL); 1008 #endif 1009 1010 #ifdef HWT_HOOKS 1011 HWT_CALL_HOOK(td, HWT_THREAD_EXIT, NULL); 1012 #endif 1013 1014 PROC_UNLOCK(p); 1015 PROC_STATLOCK(p); 1016 thread_lock(td); 1017 PROC_SUNLOCK(p); 1018 1019 /* Do the same timestamp bookkeeping that mi_switch() would do. */ 1020 new_switchtime = cpu_ticks(); 1021 runtime = new_switchtime - PCPU_GET(switchtime); 1022 td->td_runtime += runtime; 1023 td->td_incruntime += runtime; 1024 PCPU_SET(switchtime, new_switchtime); 1025 PCPU_SET(switchticks, ticks); 1026 VM_CNT_INC(v_swtch); 1027 1028 /* Save our resource usage in our process. */ 1029 td->td_ru.ru_nvcsw++; 1030 ruxagg_locked(p, td); 1031 rucollect(&p->p_ru, &td->td_ru); 1032 PROC_STATUNLOCK(p); 1033 1034 TD_SET_STATE(td, TDS_INACTIVE); 1035 #ifdef WITNESS 1036 witness_thread_exit(td); 1037 #endif 1038 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td); 1039 sched_throw(td); 1040 panic("I'm a teapot!"); 1041 /* NOTREACHED */ 1042 } 1043 1044 /* 1045 * Do any thread specific cleanups that may be needed in wait() 1046 * called with Giant, proc and schedlock not held. 1047 */ 1048 void 1049 thread_wait(struct proc *p) 1050 { 1051 struct thread *td; 1052 1053 mtx_assert(&Giant, MA_NOTOWNED); 1054 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()")); 1055 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking")); 1056 td = FIRST_THREAD_IN_PROC(p); 1057 /* Lock the last thread so we spin until it exits cpu_throw(). */ 1058 thread_lock(td); 1059 thread_unlock(td); 1060 lock_profile_thread_exit(td); 1061 cpuset_rel(td->td_cpuset); 1062 td->td_cpuset = NULL; 1063 cpu_thread_clean(td); 1064 thread_cow_free(td); 1065 callout_drain(&td->td_slpcallout); 1066 thread_reap(); /* check for zombie threads etc. */ 1067 } 1068 1069 /* 1070 * Link a thread to a process. 1071 * set up anything that needs to be initialized for it to 1072 * be used by the process. 1073 */ 1074 void 1075 thread_link(struct thread *td, struct proc *p) 1076 { 1077 1078 /* 1079 * XXX This can't be enabled because it's called for proc0 before 1080 * its lock has been created. 1081 * PROC_LOCK_ASSERT(p, MA_OWNED); 1082 */ 1083 TD_SET_STATE(td, TDS_INACTIVE); 1084 td->td_proc = p; 1085 td->td_flags = TDF_INMEM; 1086 1087 LIST_INIT(&td->td_contested); 1088 LIST_INIT(&td->td_lprof[0]); 1089 LIST_INIT(&td->td_lprof[1]); 1090 #ifdef EPOCH_TRACE 1091 SLIST_INIT(&td->td_epochs); 1092 #endif 1093 sigqueue_init(&td->td_sigqueue, p); 1094 callout_init(&td->td_slpcallout, 1); 1095 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist); 1096 p->p_numthreads++; 1097 } 1098 1099 /* 1100 * Called from: 1101 * thread_exit() 1102 */ 1103 void 1104 thread_unlink(struct thread *td) 1105 { 1106 struct proc *p = td->td_proc; 1107 1108 PROC_LOCK_ASSERT(p, MA_OWNED); 1109 #ifdef EPOCH_TRACE 1110 MPASS(SLIST_EMPTY(&td->td_epochs)); 1111 #endif 1112 1113 TAILQ_REMOVE(&p->p_threads, td, td_plist); 1114 p->p_numthreads--; 1115 /* could clear a few other things here */ 1116 /* Must NOT clear links to proc! */ 1117 } 1118 1119 static int 1120 calc_remaining(struct proc *p, int mode) 1121 { 1122 int remaining; 1123 1124 PROC_LOCK_ASSERT(p, MA_OWNED); 1125 PROC_SLOCK_ASSERT(p, MA_OWNED); 1126 if (mode == SINGLE_EXIT) 1127 remaining = p->p_numthreads; 1128 else if (mode == SINGLE_BOUNDARY) 1129 remaining = p->p_numthreads - p->p_boundary_count; 1130 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC) 1131 remaining = p->p_numthreads - p->p_suspcount; 1132 else 1133 panic("calc_remaining: wrong mode %d", mode); 1134 return (remaining); 1135 } 1136 1137 static int 1138 remain_for_mode(int mode) 1139 { 1140 1141 return (mode == SINGLE_ALLPROC ? 0 : 1); 1142 } 1143 1144 static void 1145 weed_inhib(int mode, struct thread *td2, struct proc *p) 1146 { 1147 PROC_LOCK_ASSERT(p, MA_OWNED); 1148 PROC_SLOCK_ASSERT(p, MA_OWNED); 1149 THREAD_LOCK_ASSERT(td2, MA_OWNED); 1150 1151 /* 1152 * Since the thread lock is dropped by the scheduler we have 1153 * to retry to check for races. 1154 */ 1155 restart: 1156 switch (mode) { 1157 case SINGLE_EXIT: 1158 if (TD_IS_SUSPENDED(td2)) { 1159 thread_unsuspend_one(td2, p, true); 1160 thread_lock(td2); 1161 goto restart; 1162 } 1163 if (TD_CAN_ABORT(td2)) { 1164 sleepq_abort(td2, EINTR); 1165 return; 1166 } 1167 break; 1168 case SINGLE_BOUNDARY: 1169 case SINGLE_NO_EXIT: 1170 if (TD_IS_SUSPENDED(td2) && 1171 (td2->td_flags & TDF_BOUNDARY) == 0) { 1172 thread_unsuspend_one(td2, p, false); 1173 thread_lock(td2); 1174 goto restart; 1175 } 1176 if (TD_CAN_ABORT(td2)) { 1177 sleepq_abort(td2, ERESTART); 1178 return; 1179 } 1180 break; 1181 case SINGLE_ALLPROC: 1182 /* 1183 * ALLPROC suspend tries to avoid spurious EINTR for 1184 * threads sleeping interruptable, by suspending the 1185 * thread directly, similarly to sig_suspend_threads(). 1186 * Since such sleep is not neccessary performed at the user 1187 * boundary, TDF_ALLPROCSUSP is used to avoid immediate 1188 * un-suspend. 1189 */ 1190 if (TD_IS_SUSPENDED(td2) && 1191 (td2->td_flags & TDF_ALLPROCSUSP) == 0) { 1192 thread_unsuspend_one(td2, p, false); 1193 thread_lock(td2); 1194 goto restart; 1195 } 1196 if (TD_CAN_ABORT(td2)) { 1197 td2->td_flags |= TDF_ALLPROCSUSP; 1198 sleepq_abort(td2, ERESTART); 1199 return; 1200 } 1201 break; 1202 default: 1203 break; 1204 } 1205 thread_unlock(td2); 1206 } 1207 1208 /* 1209 * Enforce single-threading. 1210 * 1211 * Returns 1 if the caller must abort (another thread is waiting to 1212 * exit the process or similar). Process is locked! 1213 * Returns 0 when you are successfully the only thread running. 1214 * A process has successfully single threaded in the suspend mode when 1215 * There are no threads in user mode. Threads in the kernel must be 1216 * allowed to continue until they get to the user boundary. They may even 1217 * copy out their return values and data before suspending. They may however be 1218 * accelerated in reaching the user boundary as we will wake up 1219 * any sleeping threads that are interruptable. (PCATCH). 1220 */ 1221 int 1222 thread_single(struct proc *p, int mode) 1223 { 1224 struct thread *td; 1225 struct thread *td2; 1226 int remaining; 1227 1228 td = curthread; 1229 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 1230 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 1231 ("invalid mode %d", mode)); 1232 /* 1233 * If allowing non-ALLPROC singlethreading for non-curproc 1234 * callers, calc_remaining() and remain_for_mode() should be 1235 * adjusted to also account for td->td_proc != p. For now 1236 * this is not implemented because it is not used. 1237 */ 1238 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) || 1239 (mode != SINGLE_ALLPROC && td->td_proc == p), 1240 ("mode %d proc %p curproc %p", mode, p, td->td_proc)); 1241 mtx_assert(&Giant, MA_NOTOWNED); 1242 PROC_LOCK_ASSERT(p, MA_OWNED); 1243 1244 /* 1245 * Is someone already single threading? 1246 * Or may be singlethreading is not needed at all. 1247 */ 1248 if (mode == SINGLE_ALLPROC) { 1249 while ((p->p_flag & P_STOPPED_SINGLE) != 0) { 1250 if ((p->p_flag2 & P2_WEXIT) != 0) 1251 return (1); 1252 msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0); 1253 } 1254 if ((p->p_flag & (P_STOPPED_SIG | P_TRACED)) != 0 || 1255 (p->p_flag2 & P2_WEXIT) != 0) 1256 return (1); 1257 } else if ((p->p_flag & P_HADTHREADS) == 0) 1258 return (0); 1259 if (p->p_singlethread != NULL && p->p_singlethread != td) 1260 return (1); 1261 1262 if (mode == SINGLE_EXIT) { 1263 p->p_flag |= P_SINGLE_EXIT; 1264 p->p_flag &= ~P_SINGLE_BOUNDARY; 1265 } else { 1266 p->p_flag &= ~P_SINGLE_EXIT; 1267 if (mode == SINGLE_BOUNDARY) 1268 p->p_flag |= P_SINGLE_BOUNDARY; 1269 else 1270 p->p_flag &= ~P_SINGLE_BOUNDARY; 1271 } 1272 if (mode == SINGLE_ALLPROC) 1273 p->p_flag |= P_TOTAL_STOP; 1274 p->p_flag |= P_STOPPED_SINGLE; 1275 PROC_SLOCK(p); 1276 p->p_singlethread = td; 1277 remaining = calc_remaining(p, mode); 1278 while (remaining != remain_for_mode(mode)) { 1279 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE) 1280 goto stopme; 1281 FOREACH_THREAD_IN_PROC(p, td2) { 1282 if (td2 == td) 1283 continue; 1284 thread_lock(td2); 1285 ast_sched_locked(td2, TDA_SUSPEND); 1286 if (TD_IS_INHIBITED(td2)) { 1287 weed_inhib(mode, td2, p); 1288 #ifdef SMP 1289 } else if (TD_IS_RUNNING(td2)) { 1290 forward_signal(td2); 1291 thread_unlock(td2); 1292 #endif 1293 } else 1294 thread_unlock(td2); 1295 } 1296 remaining = calc_remaining(p, mode); 1297 1298 /* 1299 * Maybe we suspended some threads.. was it enough? 1300 */ 1301 if (remaining == remain_for_mode(mode)) 1302 break; 1303 1304 stopme: 1305 /* 1306 * Wake us up when everyone else has suspended. 1307 * In the mean time we suspend as well. 1308 */ 1309 thread_suspend_switch(td, p); 1310 remaining = calc_remaining(p, mode); 1311 } 1312 if (mode == SINGLE_EXIT) { 1313 /* 1314 * Convert the process to an unthreaded process. The 1315 * SINGLE_EXIT is called by exit1() or execve(), in 1316 * both cases other threads must be retired. 1317 */ 1318 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads")); 1319 p->p_singlethread = NULL; 1320 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS); 1321 1322 /* 1323 * Wait for any remaining threads to exit cpu_throw(). 1324 */ 1325 while (p->p_exitthreads != 0) { 1326 PROC_SUNLOCK(p); 1327 PROC_UNLOCK(p); 1328 sched_relinquish(td); 1329 PROC_LOCK(p); 1330 PROC_SLOCK(p); 1331 } 1332 } else if (mode == SINGLE_BOUNDARY) { 1333 /* 1334 * Wait until all suspended threads are removed from 1335 * the processors. The thread_suspend_check() 1336 * increments p_boundary_count while it is still 1337 * running, which makes it possible for the execve() 1338 * to destroy vmspace while our other threads are 1339 * still using the address space. 1340 * 1341 * We lock the thread, which is only allowed to 1342 * succeed after context switch code finished using 1343 * the address space. 1344 */ 1345 FOREACH_THREAD_IN_PROC(p, td2) { 1346 if (td2 == td) 1347 continue; 1348 thread_lock(td2); 1349 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0, 1350 ("td %p not on boundary", td2)); 1351 KASSERT(TD_IS_SUSPENDED(td2), 1352 ("td %p is not suspended", td2)); 1353 thread_unlock(td2); 1354 } 1355 } 1356 PROC_SUNLOCK(p); 1357 return (0); 1358 } 1359 1360 bool 1361 thread_suspend_check_needed(void) 1362 { 1363 struct proc *p; 1364 struct thread *td; 1365 1366 td = curthread; 1367 p = td->td_proc; 1368 PROC_LOCK_ASSERT(p, MA_OWNED); 1369 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 && 1370 (td->td_dbgflags & TDB_SUSPEND) != 0)); 1371 } 1372 1373 /* 1374 * Called in from locations that can safely check to see 1375 * whether we have to suspend or at least throttle for a 1376 * single-thread event (e.g. fork). 1377 * 1378 * Such locations include userret(). 1379 * If the "return_instead" argument is non zero, the thread must be able to 1380 * accept 0 (caller may continue), or 1 (caller must abort) as a result. 1381 * 1382 * The 'return_instead' argument tells the function if it may do a 1383 * thread_exit() or suspend, or whether the caller must abort and back 1384 * out instead. 1385 * 1386 * If the thread that set the single_threading request has set the 1387 * P_SINGLE_EXIT bit in the process flags then this call will never return 1388 * if 'return_instead' is false, but will exit. 1389 * 1390 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0 1391 *---------------+--------------------+--------------------- 1392 * 0 | returns 0 | returns 0 or 1 1393 * | when ST ends | immediately 1394 *---------------+--------------------+--------------------- 1395 * 1 | thread exits | returns 1 1396 * | | immediately 1397 * 0 = thread_exit() or suspension ok, 1398 * other = return error instead of stopping the thread. 1399 * 1400 * While a full suspension is under effect, even a single threading 1401 * thread would be suspended if it made this call (but it shouldn't). 1402 * This call should only be made from places where 1403 * thread_exit() would be safe as that may be the outcome unless 1404 * return_instead is set. 1405 */ 1406 int 1407 thread_suspend_check(int return_instead) 1408 { 1409 struct thread *td; 1410 struct proc *p; 1411 1412 td = curthread; 1413 p = td->td_proc; 1414 mtx_assert(&Giant, MA_NOTOWNED); 1415 PROC_LOCK_ASSERT(p, MA_OWNED); 1416 while (thread_suspend_check_needed()) { 1417 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 1418 KASSERT(p->p_singlethread != NULL, 1419 ("singlethread not set")); 1420 /* 1421 * The only suspension in action is a 1422 * single-threading. Single threader need not stop. 1423 * It is safe to access p->p_singlethread unlocked 1424 * because it can only be set to our address by us. 1425 */ 1426 if (p->p_singlethread == td) 1427 return (0); /* Exempt from stopping. */ 1428 } 1429 if ((p->p_flag & P_SINGLE_EXIT) && return_instead) 1430 return (EINTR); 1431 1432 /* Should we goto user boundary if we didn't come from there? */ 1433 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 1434 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead) 1435 return (ERESTART); 1436 1437 /* 1438 * Ignore suspend requests if they are deferred. 1439 */ 1440 if ((td->td_flags & TDF_SBDRY) != 0) { 1441 KASSERT(return_instead, 1442 ("TDF_SBDRY set for unsafe thread_suspend_check")); 1443 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) != 1444 (TDF_SEINTR | TDF_SERESTART), 1445 ("both TDF_SEINTR and TDF_SERESTART")); 1446 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0); 1447 } 1448 1449 /* 1450 * If the process is waiting for us to exit, 1451 * this thread should just suicide. 1452 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE. 1453 */ 1454 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { 1455 PROC_UNLOCK(p); 1456 1457 /* 1458 * Allow Linux emulation layer to do some work 1459 * before thread suicide. 1460 */ 1461 if (__predict_false(p->p_sysent->sv_thread_detach != NULL)) 1462 (p->p_sysent->sv_thread_detach)(td); 1463 umtx_thread_exit(td); 1464 kern_thr_exit(td); 1465 panic("stopped thread did not exit"); 1466 } 1467 1468 PROC_SLOCK(p); 1469 thread_stopped(p); 1470 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 1471 if (p->p_numthreads == p->p_suspcount + 1) { 1472 thread_lock(p->p_singlethread); 1473 thread_unsuspend_one(p->p_singlethread, p, 1474 false); 1475 } 1476 } 1477 PROC_UNLOCK(p); 1478 thread_lock(td); 1479 /* 1480 * When a thread suspends, it just 1481 * gets taken off all queues. 1482 */ 1483 thread_suspend_one(td); 1484 if (return_instead == 0) { 1485 p->p_boundary_count++; 1486 td->td_flags |= TDF_BOUNDARY; 1487 } 1488 PROC_SUNLOCK(p); 1489 mi_switch(SW_INVOL | SWT_SUSPEND); 1490 PROC_LOCK(p); 1491 } 1492 return (0); 1493 } 1494 1495 /* 1496 * Check for possible stops and suspensions while executing a 1497 * casueword or similar transiently failing operation. 1498 * 1499 * The sleep argument controls whether the function can handle a stop 1500 * request itself or it should return ERESTART and the request is 1501 * proceed at the kernel/user boundary in ast. 1502 * 1503 * Typically, when retrying due to casueword(9) failure (rv == 1), we 1504 * should handle the stop requests there, with exception of cases when 1505 * the thread owns a kernel resource, for instance busied the umtx 1506 * key, or when functions return immediately if thread_check_susp() 1507 * returned non-zero. On the other hand, retrying the whole lock 1508 * operation, we better not stop there but delegate the handling to 1509 * ast. 1510 * 1511 * If the request is for thread termination P_SINGLE_EXIT, we cannot 1512 * handle it at all, and simply return EINTR. 1513 */ 1514 int 1515 thread_check_susp(struct thread *td, bool sleep) 1516 { 1517 struct proc *p; 1518 int error; 1519 1520 /* 1521 * The check for TDA_SUSPEND is racy, but it is enough to 1522 * eventually break the lockstep loop. 1523 */ 1524 if (!td_ast_pending(td, TDA_SUSPEND)) 1525 return (0); 1526 error = 0; 1527 p = td->td_proc; 1528 PROC_LOCK(p); 1529 if (p->p_flag & P_SINGLE_EXIT) 1530 error = EINTR; 1531 else if (P_SHOULDSTOP(p) || 1532 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) 1533 error = sleep ? thread_suspend_check(0) : ERESTART; 1534 PROC_UNLOCK(p); 1535 return (error); 1536 } 1537 1538 void 1539 thread_suspend_switch(struct thread *td, struct proc *p) 1540 { 1541 1542 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1543 PROC_LOCK_ASSERT(p, MA_OWNED); 1544 PROC_SLOCK_ASSERT(p, MA_OWNED); 1545 /* 1546 * We implement thread_suspend_one in stages here to avoid 1547 * dropping the proc lock while the thread lock is owned. 1548 */ 1549 if (p == td->td_proc) { 1550 thread_stopped(p); 1551 p->p_suspcount++; 1552 } 1553 PROC_UNLOCK(p); 1554 thread_lock(td); 1555 ast_unsched_locked(td, TDA_SUSPEND); 1556 TD_SET_SUSPENDED(td); 1557 sched_sleep(td, 0); 1558 PROC_SUNLOCK(p); 1559 DROP_GIANT(); 1560 mi_switch(SW_VOL | SWT_SUSPEND); 1561 PICKUP_GIANT(); 1562 PROC_LOCK(p); 1563 PROC_SLOCK(p); 1564 } 1565 1566 void 1567 thread_suspend_one(struct thread *td) 1568 { 1569 struct proc *p; 1570 1571 p = td->td_proc; 1572 PROC_SLOCK_ASSERT(p, MA_OWNED); 1573 THREAD_LOCK_ASSERT(td, MA_OWNED); 1574 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1575 p->p_suspcount++; 1576 ast_unsched_locked(td, TDA_SUSPEND); 1577 TD_SET_SUSPENDED(td); 1578 sched_sleep(td, 0); 1579 } 1580 1581 static void 1582 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary) 1583 { 1584 1585 THREAD_LOCK_ASSERT(td, MA_OWNED); 1586 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); 1587 TD_CLR_SUSPENDED(td); 1588 td->td_flags &= ~TDF_ALLPROCSUSP; 1589 if (td->td_proc == p) { 1590 PROC_SLOCK_ASSERT(p, MA_OWNED); 1591 p->p_suspcount--; 1592 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) { 1593 td->td_flags &= ~TDF_BOUNDARY; 1594 p->p_boundary_count--; 1595 } 1596 } 1597 setrunnable(td, 0); 1598 } 1599 1600 void 1601 thread_run_flash(struct thread *td) 1602 { 1603 struct proc *p; 1604 1605 p = td->td_proc; 1606 PROC_LOCK_ASSERT(p, MA_OWNED); 1607 1608 if (TD_ON_SLEEPQ(td)) 1609 sleepq_remove_nested(td); 1610 else 1611 thread_lock(td); 1612 1613 THREAD_LOCK_ASSERT(td, MA_OWNED); 1614 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); 1615 1616 TD_CLR_SUSPENDED(td); 1617 PROC_SLOCK(p); 1618 MPASS(p->p_suspcount > 0); 1619 p->p_suspcount--; 1620 PROC_SUNLOCK(p); 1621 setrunnable(td, 0); 1622 } 1623 1624 /* 1625 * Allow all threads blocked by single threading to continue running. 1626 */ 1627 void 1628 thread_unsuspend(struct proc *p) 1629 { 1630 struct thread *td; 1631 1632 PROC_LOCK_ASSERT(p, MA_OWNED); 1633 PROC_SLOCK_ASSERT(p, MA_OWNED); 1634 if (!P_SHOULDSTOP(p)) { 1635 FOREACH_THREAD_IN_PROC(p, td) { 1636 thread_lock(td); 1637 if (TD_IS_SUSPENDED(td)) 1638 thread_unsuspend_one(td, p, true); 1639 else 1640 thread_unlock(td); 1641 } 1642 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 1643 p->p_numthreads == p->p_suspcount) { 1644 /* 1645 * Stopping everything also did the job for the single 1646 * threading request. Now we've downgraded to single-threaded, 1647 * let it continue. 1648 */ 1649 if (p->p_singlethread->td_proc == p) { 1650 thread_lock(p->p_singlethread); 1651 thread_unsuspend_one(p->p_singlethread, p, false); 1652 } 1653 } 1654 } 1655 1656 /* 1657 * End the single threading mode.. 1658 */ 1659 void 1660 thread_single_end(struct proc *p, int mode) 1661 { 1662 struct thread *td; 1663 1664 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 1665 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 1666 ("invalid mode %d", mode)); 1667 PROC_LOCK_ASSERT(p, MA_OWNED); 1668 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) || 1669 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0), 1670 ("mode %d does not match P_TOTAL_STOP", mode)); 1671 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread, 1672 ("thread_single_end from other thread %p %p", 1673 curthread, p->p_singlethread)); 1674 KASSERT(mode != SINGLE_BOUNDARY || 1675 (p->p_flag & P_SINGLE_BOUNDARY) != 0, 1676 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag)); 1677 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY | 1678 P_TOTAL_STOP); 1679 PROC_SLOCK(p); 1680 p->p_singlethread = NULL; 1681 1682 /* 1683 * If there are other threads they may now run, 1684 * unless of course there is a blanket 'stop order' 1685 * on the process. The single threader must be allowed 1686 * to continue however as this is a bad place to stop. 1687 */ 1688 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) { 1689 FOREACH_THREAD_IN_PROC(p, td) { 1690 thread_lock(td); 1691 if (TD_IS_SUSPENDED(td)) 1692 thread_unsuspend_one(td, p, true); 1693 else 1694 thread_unlock(td); 1695 } 1696 } 1697 KASSERT(mode != SINGLE_BOUNDARY || P_SHOULDSTOP(p) || 1698 p->p_boundary_count == 0, 1699 ("pid %d proc %p flags %#x inconsistent boundary count %d", 1700 p->p_pid, p, p->p_flag, p->p_boundary_count)); 1701 PROC_SUNLOCK(p); 1702 wakeup(&p->p_flag); 1703 } 1704 1705 /* 1706 * Locate a thread by number and return with proc lock held. 1707 * 1708 * thread exit establishes proc -> tidhash lock ordering, but lookup 1709 * takes tidhash first and needs to return locked proc. 1710 * 1711 * The problem is worked around by relying on type-safety of both 1712 * structures and doing the work in 2 steps: 1713 * - tidhash-locked lookup which saves both thread and proc pointers 1714 * - proc-locked verification that the found thread still matches 1715 */ 1716 static bool 1717 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp) 1718 { 1719 #define RUN_THRESH 16 1720 struct proc *p; 1721 struct thread *td; 1722 int run; 1723 bool locked; 1724 1725 run = 0; 1726 rw_rlock(TIDHASHLOCK(tid)); 1727 locked = true; 1728 LIST_FOREACH(td, TIDHASH(tid), td_hash) { 1729 if (td->td_tid != tid) { 1730 run++; 1731 continue; 1732 } 1733 p = td->td_proc; 1734 if (pid != -1 && p->p_pid != pid) { 1735 td = NULL; 1736 break; 1737 } 1738 if (run > RUN_THRESH) { 1739 if (rw_try_upgrade(TIDHASHLOCK(tid))) { 1740 LIST_REMOVE(td, td_hash); 1741 LIST_INSERT_HEAD(TIDHASH(td->td_tid), 1742 td, td_hash); 1743 rw_wunlock(TIDHASHLOCK(tid)); 1744 locked = false; 1745 break; 1746 } 1747 } 1748 break; 1749 } 1750 if (locked) 1751 rw_runlock(TIDHASHLOCK(tid)); 1752 if (td == NULL) 1753 return (false); 1754 *pp = p; 1755 *tdp = td; 1756 return (true); 1757 } 1758 1759 struct thread * 1760 tdfind(lwpid_t tid, pid_t pid) 1761 { 1762 struct proc *p; 1763 struct thread *td; 1764 1765 td = curthread; 1766 if (td->td_tid == tid) { 1767 if (pid != -1 && td->td_proc->p_pid != pid) 1768 return (NULL); 1769 PROC_LOCK(td->td_proc); 1770 return (td); 1771 } 1772 1773 for (;;) { 1774 if (!tdfind_hash(tid, pid, &p, &td)) 1775 return (NULL); 1776 PROC_LOCK(p); 1777 if (td->td_tid != tid) { 1778 PROC_UNLOCK(p); 1779 continue; 1780 } 1781 if (td->td_proc != p) { 1782 PROC_UNLOCK(p); 1783 continue; 1784 } 1785 if (p->p_state == PRS_NEW) { 1786 PROC_UNLOCK(p); 1787 return (NULL); 1788 } 1789 return (td); 1790 } 1791 } 1792 1793 void 1794 tidhash_add(struct thread *td) 1795 { 1796 rw_wlock(TIDHASHLOCK(td->td_tid)); 1797 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash); 1798 rw_wunlock(TIDHASHLOCK(td->td_tid)); 1799 } 1800 1801 void 1802 tidhash_remove(struct thread *td) 1803 { 1804 1805 rw_wlock(TIDHASHLOCK(td->td_tid)); 1806 LIST_REMOVE(td, td_hash); 1807 rw_wunlock(TIDHASHLOCK(td->td_tid)); 1808 } 1809