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) == 0x3e4, 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) == 0x288, 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 = NULL; 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 == NULL, ("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 cpu_thread_new_kstack(td); 807 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td); 808 return (td); 809 } 810 811 int 812 thread_recycle(struct thread *td, int pages) 813 { 814 if (td->td_kstack == NULL || td->td_kstack_pages != pages) { 815 if (td->td_kstack != NULL) 816 vm_thread_dispose(td); 817 if (!vm_thread_new(td, pages)) 818 return (ENOMEM); 819 cpu_thread_new_kstack(td); 820 } 821 kasan_thread_alloc(td); 822 kmsan_thread_alloc(td); 823 return (0); 824 } 825 826 /* 827 * Deallocate a thread. 828 */ 829 static void 830 thread_free_batched(struct thread *td) 831 { 832 833 lock_profile_thread_exit(td); 834 if (td->td_cpuset) 835 cpuset_rel(td->td_cpuset); 836 td->td_cpuset = NULL; 837 cpu_thread_free(td); 838 if (td->td_kstack != NULL) 839 vm_thread_dispose(td); 840 callout_drain(&td->td_slpcallout); 841 /* 842 * Freeing handled by the caller. 843 */ 844 td->td_tid = -1; 845 kmsan_thread_free(td); 846 uma_zfree(thread_zone, td); 847 } 848 849 void 850 thread_free(struct thread *td) 851 { 852 lwpid_t tid; 853 854 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td); 855 tid = td->td_tid; 856 thread_free_batched(td); 857 tid_free(tid); 858 thread_count_dec(); 859 } 860 861 void 862 thread_cow_get_proc(struct thread *newtd, struct proc *p) 863 { 864 865 PROC_LOCK_ASSERT(p, MA_OWNED); 866 newtd->td_realucred = crcowget(p->p_ucred); 867 newtd->td_ucred = newtd->td_realucred; 868 newtd->td_limit = lim_hold(p->p_limit); 869 newtd->td_cowgen = p->p_cowgen; 870 } 871 872 void 873 thread_cow_get(struct thread *newtd, struct thread *td) 874 { 875 876 MPASS(td->td_realucred == td->td_ucred); 877 newtd->td_realucred = crcowget(td->td_realucred); 878 newtd->td_ucred = newtd->td_realucred; 879 newtd->td_limit = lim_hold(td->td_limit); 880 newtd->td_cowgen = td->td_cowgen; 881 } 882 883 void 884 thread_cow_free(struct thread *td) 885 { 886 887 if (td->td_realucred != NULL) 888 crcowfree(td); 889 if (td->td_limit != NULL) 890 lim_free(td->td_limit); 891 } 892 893 void 894 thread_cow_update(struct thread *td) 895 { 896 struct proc *p; 897 struct ucred *oldcred; 898 struct plimit *oldlimit; 899 900 p = td->td_proc; 901 PROC_LOCK(p); 902 oldcred = crcowsync(); 903 oldlimit = lim_cowsync(); 904 td->td_cowgen = p->p_cowgen; 905 PROC_UNLOCK(p); 906 if (oldcred != NULL) 907 crfree(oldcred); 908 if (oldlimit != NULL) 909 lim_free(oldlimit); 910 } 911 912 void 913 thread_cow_synced(struct thread *td) 914 { 915 struct proc *p; 916 917 p = td->td_proc; 918 PROC_LOCK_ASSERT(p, MA_OWNED); 919 MPASS(td->td_cowgen != p->p_cowgen); 920 MPASS(td->td_ucred == p->p_ucred); 921 MPASS(td->td_limit == p->p_limit); 922 td->td_cowgen = p->p_cowgen; 923 } 924 925 /* 926 * Discard the current thread and exit from its context. 927 * Always called with scheduler locked. 928 * 929 * Because we can't free a thread while we're operating under its context, 930 * push the current thread into our CPU's deadthread holder. This means 931 * we needn't worry about someone else grabbing our context before we 932 * do a cpu_throw(). 933 */ 934 void 935 thread_exit(void) 936 { 937 uint64_t runtime, new_switchtime; 938 struct thread *td; 939 struct thread *td2; 940 struct proc *p; 941 942 td = curthread; 943 p = td->td_proc; 944 945 PROC_SLOCK_ASSERT(p, MA_OWNED); 946 mtx_assert(&Giant, MA_NOTOWNED); 947 948 PROC_LOCK_ASSERT(p, MA_OWNED); 949 KASSERT(p != NULL, ("thread exiting without a process")); 950 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td, 951 (long)p->p_pid, td->td_name); 952 SDT_PROBE0(proc, , , lwp__exit); 953 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending")); 954 MPASS(td->td_realucred == td->td_ucred); 955 956 /* 957 * drop FPU & debug register state storage, or any other 958 * architecture specific resources that 959 * would not be on a new untouched process. 960 */ 961 cpu_thread_exit(td); 962 963 /* 964 * The last thread is left attached to the process 965 * So that the whole bundle gets recycled. Skip 966 * all this stuff if we never had threads. 967 * EXIT clears all sign of other threads when 968 * it goes to single threading, so the last thread always 969 * takes the short path. 970 */ 971 if (p->p_flag & P_HADTHREADS) { 972 if (p->p_numthreads > 1) { 973 atomic_add_int(&td->td_proc->p_exitthreads, 1); 974 thread_unlink(td); 975 td2 = FIRST_THREAD_IN_PROC(p); 976 sched_exit_thread(td2, td); 977 978 /* 979 * The test below is NOT true if we are the 980 * sole exiting thread. P_STOPPED_SINGLE is unset 981 * in exit1() after it is the only survivor. 982 */ 983 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 984 if (p->p_numthreads == p->p_suspcount) { 985 thread_lock(p->p_singlethread); 986 thread_unsuspend_one(p->p_singlethread, 987 p, false); 988 } 989 } 990 991 PCPU_SET(deadthread, td); 992 } else { 993 /* 994 * The last thread is exiting.. but not through exit() 995 */ 996 panic ("thread_exit: Last thread exiting on its own"); 997 } 998 } 999 #ifdef HWPMC_HOOKS 1000 /* 1001 * If this thread is part of a process that is being tracked by hwpmc(4), 1002 * inform the module of the thread's impending exit. 1003 */ 1004 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) { 1005 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1006 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL); 1007 } else if (PMC_SYSTEM_SAMPLING_ACTIVE()) 1008 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL); 1009 #endif 1010 1011 #ifdef HWT_HOOKS 1012 HWT_CALL_HOOK(td, HWT_THREAD_EXIT, NULL); 1013 #endif 1014 1015 PROC_UNLOCK(p); 1016 PROC_STATLOCK(p); 1017 thread_lock(td); 1018 PROC_SUNLOCK(p); 1019 1020 /* Do the same timestamp bookkeeping that mi_switch() would do. */ 1021 new_switchtime = cpu_ticks(); 1022 runtime = new_switchtime - PCPU_GET(switchtime); 1023 td->td_runtime += runtime; 1024 td->td_incruntime += runtime; 1025 PCPU_SET(switchtime, new_switchtime); 1026 PCPU_SET(switchticks, ticks); 1027 VM_CNT_INC(v_swtch); 1028 1029 /* Save our resource usage in our process. */ 1030 td->td_ru.ru_nvcsw++; 1031 ruxagg_locked(p, td); 1032 rucollect(&p->p_ru, &td->td_ru); 1033 PROC_STATUNLOCK(p); 1034 1035 TD_SET_STATE(td, TDS_INACTIVE); 1036 #ifdef WITNESS 1037 witness_thread_exit(td); 1038 #endif 1039 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td); 1040 sched_throw(td); 1041 panic("I'm a teapot!"); 1042 /* NOTREACHED */ 1043 } 1044 1045 /* 1046 * Do any thread specific cleanups that may be needed in wait() 1047 * called with Giant, proc and schedlock not held. 1048 */ 1049 void 1050 thread_wait(struct proc *p) 1051 { 1052 struct thread *td; 1053 1054 mtx_assert(&Giant, MA_NOTOWNED); 1055 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()")); 1056 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking")); 1057 td = FIRST_THREAD_IN_PROC(p); 1058 /* Lock the last thread so we spin until it exits cpu_throw(). */ 1059 thread_lock(td); 1060 thread_unlock(td); 1061 lock_profile_thread_exit(td); 1062 cpuset_rel(td->td_cpuset); 1063 td->td_cpuset = NULL; 1064 cpu_thread_clean(td); 1065 thread_cow_free(td); 1066 callout_drain(&td->td_slpcallout); 1067 thread_reap(); /* check for zombie threads etc. */ 1068 } 1069 1070 /* 1071 * Link a thread to a process. 1072 * set up anything that needs to be initialized for it to 1073 * be used by the process. 1074 */ 1075 void 1076 thread_link(struct thread *td, struct proc *p) 1077 { 1078 1079 /* 1080 * XXX This can't be enabled because it's called for proc0 before 1081 * its lock has been created. 1082 * PROC_LOCK_ASSERT(p, MA_OWNED); 1083 */ 1084 TD_SET_STATE(td, TDS_INACTIVE); 1085 td->td_proc = p; 1086 td->td_flags = TDF_INMEM; 1087 1088 LIST_INIT(&td->td_contested); 1089 LIST_INIT(&td->td_lprof[0]); 1090 LIST_INIT(&td->td_lprof[1]); 1091 #ifdef EPOCH_TRACE 1092 SLIST_INIT(&td->td_epochs); 1093 #endif 1094 sigqueue_init(&td->td_sigqueue, p); 1095 callout_init(&td->td_slpcallout, 1); 1096 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist); 1097 p->p_numthreads++; 1098 } 1099 1100 /* 1101 * Called from: 1102 * thread_exit() 1103 */ 1104 void 1105 thread_unlink(struct thread *td) 1106 { 1107 struct proc *p = td->td_proc; 1108 1109 PROC_LOCK_ASSERT(p, MA_OWNED); 1110 #ifdef EPOCH_TRACE 1111 MPASS(SLIST_EMPTY(&td->td_epochs)); 1112 #endif 1113 1114 TAILQ_REMOVE(&p->p_threads, td, td_plist); 1115 p->p_numthreads--; 1116 /* could clear a few other things here */ 1117 /* Must NOT clear links to proc! */ 1118 } 1119 1120 static int 1121 calc_remaining(struct proc *p, int mode) 1122 { 1123 int remaining; 1124 1125 PROC_LOCK_ASSERT(p, MA_OWNED); 1126 PROC_SLOCK_ASSERT(p, MA_OWNED); 1127 if (mode == SINGLE_EXIT) 1128 remaining = p->p_numthreads; 1129 else if (mode == SINGLE_BOUNDARY) 1130 remaining = p->p_numthreads - p->p_boundary_count; 1131 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC) 1132 remaining = p->p_numthreads - p->p_suspcount; 1133 else 1134 panic("calc_remaining: wrong mode %d", mode); 1135 return (remaining); 1136 } 1137 1138 static int 1139 remain_for_mode(int mode) 1140 { 1141 1142 return (mode == SINGLE_ALLPROC ? 0 : 1); 1143 } 1144 1145 static void 1146 weed_inhib(int mode, struct thread *td2, struct proc *p) 1147 { 1148 PROC_LOCK_ASSERT(p, MA_OWNED); 1149 PROC_SLOCK_ASSERT(p, MA_OWNED); 1150 THREAD_LOCK_ASSERT(td2, MA_OWNED); 1151 1152 /* 1153 * Since the thread lock is dropped by the scheduler we have 1154 * to retry to check for races. 1155 */ 1156 restart: 1157 switch (mode) { 1158 case SINGLE_EXIT: 1159 if (TD_IS_SUSPENDED(td2)) { 1160 thread_unsuspend_one(td2, p, true); 1161 thread_lock(td2); 1162 goto restart; 1163 } 1164 if (TD_CAN_ABORT(td2)) { 1165 sleepq_abort(td2, EINTR); 1166 return; 1167 } 1168 break; 1169 case SINGLE_BOUNDARY: 1170 case SINGLE_NO_EXIT: 1171 if (TD_IS_SUSPENDED(td2) && 1172 (td2->td_flags & TDF_BOUNDARY) == 0) { 1173 thread_unsuspend_one(td2, p, false); 1174 thread_lock(td2); 1175 goto restart; 1176 } 1177 if (TD_CAN_ABORT(td2)) { 1178 sleepq_abort(td2, ERESTART); 1179 return; 1180 } 1181 break; 1182 case SINGLE_ALLPROC: 1183 /* 1184 * ALLPROC suspend tries to avoid spurious EINTR for 1185 * threads sleeping interruptable, by suspending the 1186 * thread directly, similarly to sig_suspend_threads(). 1187 * Since such sleep is not neccessary performed at the user 1188 * boundary, TDF_ALLPROCSUSP is used to avoid immediate 1189 * un-suspend. 1190 */ 1191 if (TD_IS_SUSPENDED(td2) && 1192 (td2->td_flags & TDF_ALLPROCSUSP) == 0) { 1193 thread_unsuspend_one(td2, p, false); 1194 thread_lock(td2); 1195 goto restart; 1196 } 1197 if (TD_CAN_ABORT(td2)) { 1198 td2->td_flags |= TDF_ALLPROCSUSP; 1199 sleepq_abort(td2, ERESTART); 1200 return; 1201 } 1202 break; 1203 default: 1204 break; 1205 } 1206 thread_unlock(td2); 1207 } 1208 1209 /* 1210 * Enforce single-threading. 1211 * 1212 * Returns 1 if the caller must abort (another thread is waiting to 1213 * exit the process or similar). Process is locked! 1214 * Returns 0 when you are successfully the only thread running. 1215 * A process has successfully single threaded in the suspend mode when 1216 * There are no threads in user mode. Threads in the kernel must be 1217 * allowed to continue until they get to the user boundary. They may even 1218 * copy out their return values and data before suspending. They may however be 1219 * accelerated in reaching the user boundary as we will wake up 1220 * any sleeping threads that are interruptable. (PCATCH). 1221 */ 1222 int 1223 thread_single(struct proc *p, int mode) 1224 { 1225 struct thread *td; 1226 struct thread *td2; 1227 int remaining; 1228 1229 td = curthread; 1230 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 1231 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 1232 ("invalid mode %d", mode)); 1233 /* 1234 * If allowing non-ALLPROC singlethreading for non-curproc 1235 * callers, calc_remaining() and remain_for_mode() should be 1236 * adjusted to also account for td->td_proc != p. For now 1237 * this is not implemented because it is not used. 1238 */ 1239 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) || 1240 (mode != SINGLE_ALLPROC && td->td_proc == p), 1241 ("mode %d proc %p curproc %p", mode, p, td->td_proc)); 1242 mtx_assert(&Giant, MA_NOTOWNED); 1243 PROC_LOCK_ASSERT(p, MA_OWNED); 1244 1245 /* 1246 * Is someone already single threading? 1247 * Or may be singlethreading is not needed at all. 1248 */ 1249 if (mode == SINGLE_ALLPROC) { 1250 while ((p->p_flag & P_STOPPED_SINGLE) != 0) { 1251 if ((p->p_flag2 & P2_WEXIT) != 0) 1252 return (1); 1253 msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0); 1254 } 1255 if ((p->p_flag & (P_STOPPED_SIG | P_TRACED)) != 0 || 1256 (p->p_flag2 & P2_WEXIT) != 0) 1257 return (1); 1258 } else if ((p->p_flag & P_HADTHREADS) == 0) 1259 return (0); 1260 if (p->p_singlethread != NULL && p->p_singlethread != td) 1261 return (1); 1262 1263 if (mode == SINGLE_EXIT) { 1264 p->p_flag |= P_SINGLE_EXIT; 1265 p->p_flag &= ~P_SINGLE_BOUNDARY; 1266 } else { 1267 p->p_flag &= ~P_SINGLE_EXIT; 1268 if (mode == SINGLE_BOUNDARY) 1269 p->p_flag |= P_SINGLE_BOUNDARY; 1270 else 1271 p->p_flag &= ~P_SINGLE_BOUNDARY; 1272 } 1273 if (mode == SINGLE_ALLPROC) 1274 p->p_flag |= P_TOTAL_STOP; 1275 p->p_flag |= P_STOPPED_SINGLE; 1276 PROC_SLOCK(p); 1277 p->p_singlethread = td; 1278 remaining = calc_remaining(p, mode); 1279 while (remaining != remain_for_mode(mode)) { 1280 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE) 1281 goto stopme; 1282 FOREACH_THREAD_IN_PROC(p, td2) { 1283 if (td2 == td) 1284 continue; 1285 thread_lock(td2); 1286 ast_sched_locked(td2, TDA_SUSPEND); 1287 if (TD_IS_INHIBITED(td2)) { 1288 weed_inhib(mode, td2, p); 1289 #ifdef SMP 1290 } else if (TD_IS_RUNNING(td2)) { 1291 forward_signal(td2); 1292 thread_unlock(td2); 1293 #endif 1294 } else 1295 thread_unlock(td2); 1296 } 1297 remaining = calc_remaining(p, mode); 1298 1299 /* 1300 * Maybe we suspended some threads.. was it enough? 1301 */ 1302 if (remaining == remain_for_mode(mode)) 1303 break; 1304 1305 stopme: 1306 /* 1307 * Wake us up when everyone else has suspended. 1308 * In the mean time we suspend as well. 1309 */ 1310 thread_suspend_switch(td, p); 1311 remaining = calc_remaining(p, mode); 1312 } 1313 if (mode == SINGLE_EXIT) { 1314 /* 1315 * Convert the process to an unthreaded process. The 1316 * SINGLE_EXIT is called by exit1() or execve(), in 1317 * both cases other threads must be retired. 1318 */ 1319 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads")); 1320 p->p_singlethread = NULL; 1321 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS); 1322 1323 /* 1324 * Wait for any remaining threads to exit cpu_throw(). 1325 */ 1326 while (p->p_exitthreads != 0) { 1327 PROC_SUNLOCK(p); 1328 PROC_UNLOCK(p); 1329 sched_relinquish(td); 1330 PROC_LOCK(p); 1331 PROC_SLOCK(p); 1332 } 1333 } else if (mode == SINGLE_BOUNDARY) { 1334 /* 1335 * Wait until all suspended threads are removed from 1336 * the processors. The thread_suspend_check() 1337 * increments p_boundary_count while it is still 1338 * running, which makes it possible for the execve() 1339 * to destroy vmspace while our other threads are 1340 * still using the address space. 1341 * 1342 * We lock the thread, which is only allowed to 1343 * succeed after context switch code finished using 1344 * the address space. 1345 */ 1346 FOREACH_THREAD_IN_PROC(p, td2) { 1347 if (td2 == td) 1348 continue; 1349 thread_lock(td2); 1350 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0, 1351 ("td %p not on boundary", td2)); 1352 KASSERT(TD_IS_SUSPENDED(td2), 1353 ("td %p is not suspended", td2)); 1354 thread_unlock(td2); 1355 } 1356 } 1357 PROC_SUNLOCK(p); 1358 return (0); 1359 } 1360 1361 bool 1362 thread_suspend_check_needed(void) 1363 { 1364 struct proc *p; 1365 struct thread *td; 1366 1367 td = curthread; 1368 p = td->td_proc; 1369 PROC_LOCK_ASSERT(p, MA_OWNED); 1370 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 && 1371 (td->td_dbgflags & TDB_SUSPEND) != 0)); 1372 } 1373 1374 /* 1375 * Called in from locations that can safely check to see 1376 * whether we have to suspend or at least throttle for a 1377 * single-thread event (e.g. fork). 1378 * 1379 * Such locations include userret(). 1380 * If the "return_instead" argument is non zero, the thread must be able to 1381 * accept 0 (caller may continue), or 1 (caller must abort) as a result. 1382 * 1383 * The 'return_instead' argument tells the function if it may do a 1384 * thread_exit() or suspend, or whether the caller must abort and back 1385 * out instead. 1386 * 1387 * If the thread that set the single_threading request has set the 1388 * P_SINGLE_EXIT bit in the process flags then this call will never return 1389 * if 'return_instead' is false, but will exit. 1390 * 1391 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0 1392 *---------------+--------------------+--------------------- 1393 * 0 | returns 0 | returns 0 or 1 1394 * | when ST ends | immediately 1395 *---------------+--------------------+--------------------- 1396 * 1 | thread exits | returns 1 1397 * | | immediately 1398 * 0 = thread_exit() or suspension ok, 1399 * other = return error instead of stopping the thread. 1400 * 1401 * While a full suspension is under effect, even a single threading 1402 * thread would be suspended if it made this call (but it shouldn't). 1403 * This call should only be made from places where 1404 * thread_exit() would be safe as that may be the outcome unless 1405 * return_instead is set. 1406 */ 1407 int 1408 thread_suspend_check(int return_instead) 1409 { 1410 struct thread *td; 1411 struct proc *p; 1412 1413 td = curthread; 1414 p = td->td_proc; 1415 mtx_assert(&Giant, MA_NOTOWNED); 1416 PROC_LOCK_ASSERT(p, MA_OWNED); 1417 while (thread_suspend_check_needed()) { 1418 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 1419 KASSERT(p->p_singlethread != NULL, 1420 ("singlethread not set")); 1421 /* 1422 * The only suspension in action is a 1423 * single-threading. Single threader need not stop. 1424 * It is safe to access p->p_singlethread unlocked 1425 * because it can only be set to our address by us. 1426 */ 1427 if (p->p_singlethread == td) 1428 return (0); /* Exempt from stopping. */ 1429 } 1430 if ((p->p_flag & P_SINGLE_EXIT) && return_instead) 1431 return (EINTR); 1432 1433 /* Should we goto user boundary if we didn't come from there? */ 1434 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 1435 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead) 1436 return (ERESTART); 1437 1438 /* 1439 * Ignore suspend requests if they are deferred. 1440 */ 1441 if ((td->td_flags & TDF_SBDRY) != 0) { 1442 KASSERT(return_instead, 1443 ("TDF_SBDRY set for unsafe thread_suspend_check")); 1444 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) != 1445 (TDF_SEINTR | TDF_SERESTART), 1446 ("both TDF_SEINTR and TDF_SERESTART")); 1447 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0); 1448 } 1449 1450 /* 1451 * We might get here with return_instead == 1 if 1452 * other checks missed it. Then we must not suspend 1453 * regardless of P_SHOULDSTOP() or debugger request. 1454 */ 1455 if (return_instead) 1456 return (EINTR); 1457 1458 /* 1459 * If the process is waiting for us to exit, 1460 * this thread should just suicide. 1461 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE. 1462 */ 1463 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { 1464 PROC_UNLOCK(p); 1465 1466 /* 1467 * Allow Linux emulation layer to do some work 1468 * before thread suicide. 1469 */ 1470 if (__predict_false(p->p_sysent->sv_thread_detach != NULL)) 1471 (p->p_sysent->sv_thread_detach)(td); 1472 umtx_thread_exit(td); 1473 kern_thr_exit(td); 1474 panic("stopped thread did not exit"); 1475 } 1476 1477 PROC_SLOCK(p); 1478 thread_stopped(p); 1479 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 1480 if (p->p_numthreads == p->p_suspcount + 1) { 1481 thread_lock(p->p_singlethread); 1482 thread_unsuspend_one(p->p_singlethread, p, 1483 false); 1484 } 1485 } 1486 PROC_UNLOCK(p); 1487 thread_lock(td); 1488 /* 1489 * When a thread suspends, it just 1490 * gets taken off all queues. 1491 */ 1492 thread_suspend_one(td); 1493 MPASS(!return_instead); 1494 p->p_boundary_count++; 1495 td->td_flags |= TDF_BOUNDARY; 1496 PROC_SUNLOCK(p); 1497 mi_switch(SW_INVOL | SWT_SUSPEND); 1498 PROC_LOCK(p); 1499 } 1500 return (0); 1501 } 1502 1503 /* 1504 * Check for possible stops and suspensions while executing a 1505 * casueword or similar transiently failing operation. 1506 * 1507 * The sleep argument controls whether the function can handle a stop 1508 * request itself or it should return ERESTART and the request is 1509 * proceed at the kernel/user boundary in ast. 1510 * 1511 * Typically, when retrying due to casueword(9) failure (rv == 1), we 1512 * should handle the stop requests there, with exception of cases when 1513 * the thread owns a kernel resource, for instance busied the umtx 1514 * key, or when functions return immediately if thread_check_susp() 1515 * returned non-zero. On the other hand, retrying the whole lock 1516 * operation, we better not stop there but delegate the handling to 1517 * ast. 1518 * 1519 * If the request is for thread termination P_SINGLE_EXIT, we cannot 1520 * handle it at all, and simply return EINTR. 1521 */ 1522 int 1523 thread_check_susp(struct thread *td, bool sleep) 1524 { 1525 struct proc *p; 1526 int error; 1527 1528 /* 1529 * The check for TDA_SUSPEND is racy, but it is enough to 1530 * eventually break the lockstep loop. 1531 */ 1532 if (!td_ast_pending(td, TDA_SUSPEND)) 1533 return (0); 1534 error = 0; 1535 p = td->td_proc; 1536 PROC_LOCK(p); 1537 if (p->p_flag & P_SINGLE_EXIT) 1538 error = EINTR; 1539 else if (P_SHOULDSTOP(p) || 1540 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) 1541 error = sleep ? thread_suspend_check(0) : ERESTART; 1542 PROC_UNLOCK(p); 1543 return (error); 1544 } 1545 1546 void 1547 thread_suspend_switch(struct thread *td, struct proc *p) 1548 { 1549 1550 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1551 PROC_LOCK_ASSERT(p, MA_OWNED); 1552 PROC_SLOCK_ASSERT(p, MA_OWNED); 1553 /* 1554 * We implement thread_suspend_one in stages here to avoid 1555 * dropping the proc lock while the thread lock is owned. 1556 */ 1557 if (p == td->td_proc) { 1558 thread_stopped(p); 1559 p->p_suspcount++; 1560 } 1561 PROC_UNLOCK(p); 1562 thread_lock(td); 1563 ast_unsched_locked(td, TDA_SUSPEND); 1564 TD_SET_SUSPENDED(td); 1565 sched_sleep(td, 0); 1566 PROC_SUNLOCK(p); 1567 DROP_GIANT(); 1568 mi_switch(SW_VOL | SWT_SUSPEND); 1569 PICKUP_GIANT(); 1570 PROC_LOCK(p); 1571 PROC_SLOCK(p); 1572 } 1573 1574 void 1575 thread_suspend_one(struct thread *td) 1576 { 1577 struct proc *p; 1578 1579 p = td->td_proc; 1580 PROC_SLOCK_ASSERT(p, MA_OWNED); 1581 THREAD_LOCK_ASSERT(td, MA_OWNED); 1582 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1583 p->p_suspcount++; 1584 ast_unsched_locked(td, TDA_SUSPEND); 1585 TD_SET_SUSPENDED(td); 1586 sched_sleep(td, 0); 1587 } 1588 1589 static void 1590 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary) 1591 { 1592 1593 THREAD_LOCK_ASSERT(td, MA_OWNED); 1594 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); 1595 TD_CLR_SUSPENDED(td); 1596 td->td_flags &= ~TDF_ALLPROCSUSP; 1597 if (td->td_proc == p) { 1598 PROC_SLOCK_ASSERT(p, MA_OWNED); 1599 p->p_suspcount--; 1600 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) { 1601 td->td_flags &= ~TDF_BOUNDARY; 1602 p->p_boundary_count--; 1603 } 1604 } 1605 setrunnable(td, 0); 1606 } 1607 1608 void 1609 thread_run_flash(struct thread *td) 1610 { 1611 struct proc *p; 1612 1613 p = td->td_proc; 1614 PROC_LOCK_ASSERT(p, MA_OWNED); 1615 1616 if (TD_ON_SLEEPQ(td)) 1617 sleepq_remove_nested(td); 1618 else 1619 thread_lock(td); 1620 1621 THREAD_LOCK_ASSERT(td, MA_OWNED); 1622 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); 1623 1624 TD_CLR_SUSPENDED(td); 1625 PROC_SLOCK(p); 1626 MPASS(p->p_suspcount > 0); 1627 p->p_suspcount--; 1628 PROC_SUNLOCK(p); 1629 setrunnable(td, 0); 1630 } 1631 1632 /* 1633 * Allow all threads blocked by single threading to continue running. 1634 */ 1635 void 1636 thread_unsuspend(struct proc *p) 1637 { 1638 struct thread *td; 1639 1640 PROC_LOCK_ASSERT(p, MA_OWNED); 1641 PROC_SLOCK_ASSERT(p, MA_OWNED); 1642 if (!P_SHOULDSTOP(p)) { 1643 FOREACH_THREAD_IN_PROC(p, td) { 1644 thread_lock(td); 1645 if (TD_IS_SUSPENDED(td)) 1646 thread_unsuspend_one(td, p, true); 1647 else 1648 thread_unlock(td); 1649 } 1650 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 1651 p->p_numthreads == p->p_suspcount) { 1652 /* 1653 * Stopping everything also did the job for the single 1654 * threading request. Now we've downgraded to single-threaded, 1655 * let it continue. 1656 */ 1657 if (p->p_singlethread->td_proc == p) { 1658 thread_lock(p->p_singlethread); 1659 thread_unsuspend_one(p->p_singlethread, p, false); 1660 } 1661 } 1662 } 1663 1664 /* 1665 * End the single threading mode.. 1666 */ 1667 void 1668 thread_single_end(struct proc *p, int mode) 1669 { 1670 struct thread *td; 1671 1672 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 1673 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 1674 ("invalid mode %d", mode)); 1675 PROC_LOCK_ASSERT(p, MA_OWNED); 1676 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) || 1677 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0), 1678 ("mode %d does not match P_TOTAL_STOP", mode)); 1679 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread, 1680 ("thread_single_end from other thread %p %p", 1681 curthread, p->p_singlethread)); 1682 KASSERT(mode != SINGLE_BOUNDARY || 1683 (p->p_flag & P_SINGLE_BOUNDARY) != 0, 1684 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag)); 1685 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY | 1686 P_TOTAL_STOP); 1687 PROC_SLOCK(p); 1688 p->p_singlethread = NULL; 1689 1690 /* 1691 * If there are other threads they may now run, 1692 * unless of course there is a blanket 'stop order' 1693 * on the process. The single threader must be allowed 1694 * to continue however as this is a bad place to stop. 1695 */ 1696 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) { 1697 FOREACH_THREAD_IN_PROC(p, td) { 1698 thread_lock(td); 1699 if (TD_IS_SUSPENDED(td)) 1700 thread_unsuspend_one(td, p, true); 1701 else 1702 thread_unlock(td); 1703 } 1704 } 1705 KASSERT(mode != SINGLE_BOUNDARY || P_SHOULDSTOP(p) || 1706 p->p_boundary_count == 0, 1707 ("pid %d proc %p flags %#x inconsistent boundary count %d", 1708 p->p_pid, p, p->p_flag, p->p_boundary_count)); 1709 PROC_SUNLOCK(p); 1710 wakeup(&p->p_flag); 1711 } 1712 1713 /* 1714 * Locate a thread by number and return with proc lock held. 1715 * 1716 * thread exit establishes proc -> tidhash lock ordering, but lookup 1717 * takes tidhash first and needs to return locked proc. 1718 * 1719 * The problem is worked around by relying on type-safety of both 1720 * structures and doing the work in 2 steps: 1721 * - tidhash-locked lookup which saves both thread and proc pointers 1722 * - proc-locked verification that the found thread still matches 1723 */ 1724 static bool 1725 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp) 1726 { 1727 #define RUN_THRESH 16 1728 struct proc *p; 1729 struct thread *td; 1730 int run; 1731 bool locked; 1732 1733 run = 0; 1734 rw_rlock(TIDHASHLOCK(tid)); 1735 locked = true; 1736 LIST_FOREACH(td, TIDHASH(tid), td_hash) { 1737 if (td->td_tid != tid) { 1738 run++; 1739 continue; 1740 } 1741 p = td->td_proc; 1742 if (pid != -1 && p->p_pid != pid) { 1743 td = NULL; 1744 break; 1745 } 1746 if (run > RUN_THRESH) { 1747 if (rw_try_upgrade(TIDHASHLOCK(tid))) { 1748 LIST_REMOVE(td, td_hash); 1749 LIST_INSERT_HEAD(TIDHASH(td->td_tid), 1750 td, td_hash); 1751 rw_wunlock(TIDHASHLOCK(tid)); 1752 locked = false; 1753 break; 1754 } 1755 } 1756 break; 1757 } 1758 if (locked) 1759 rw_runlock(TIDHASHLOCK(tid)); 1760 if (td == NULL) 1761 return (false); 1762 *pp = p; 1763 *tdp = td; 1764 return (true); 1765 } 1766 1767 struct thread * 1768 tdfind(lwpid_t tid, pid_t pid) 1769 { 1770 struct proc *p; 1771 struct thread *td; 1772 1773 td = curthread; 1774 if (td->td_tid == tid) { 1775 if (pid != -1 && td->td_proc->p_pid != pid) 1776 return (NULL); 1777 PROC_LOCK(td->td_proc); 1778 return (td); 1779 } 1780 1781 for (;;) { 1782 if (!tdfind_hash(tid, pid, &p, &td)) 1783 return (NULL); 1784 PROC_LOCK(p); 1785 if (td->td_tid != tid) { 1786 PROC_UNLOCK(p); 1787 continue; 1788 } 1789 if (td->td_proc != p) { 1790 PROC_UNLOCK(p); 1791 continue; 1792 } 1793 if (p->p_state == PRS_NEW) { 1794 PROC_UNLOCK(p); 1795 return (NULL); 1796 } 1797 return (td); 1798 } 1799 } 1800 1801 void 1802 tidhash_add(struct thread *td) 1803 { 1804 rw_wlock(TIDHASHLOCK(td->td_tid)); 1805 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash); 1806 rw_wunlock(TIDHASHLOCK(td->td_tid)); 1807 } 1808 1809 void 1810 tidhash_remove(struct thread *td) 1811 { 1812 1813 rw_wlock(TIDHASHLOCK(td->td_tid)); 1814 LIST_REMOVE(td, td_hash); 1815 rw_wunlock(TIDHASHLOCK(td->td_tid)); 1816 } 1817