1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 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 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/kernel.h> 40 #include <sys/lock.h> 41 #include <sys/msan.h> 42 #include <sys/mutex.h> 43 #include <sys/proc.h> 44 #include <sys/bitstring.h> 45 #include <sys/epoch.h> 46 #include <sys/rangelock.h> 47 #include <sys/resourcevar.h> 48 #include <sys/sdt.h> 49 #include <sys/smp.h> 50 #include <sys/sched.h> 51 #include <sys/sleepqueue.h> 52 #include <sys/selinfo.h> 53 #include <sys/syscallsubr.h> 54 #include <sys/dtrace_bsd.h> 55 #include <sys/sysent.h> 56 #include <sys/turnstile.h> 57 #include <sys/taskqueue.h> 58 #include <sys/ktr.h> 59 #include <sys/rwlock.h> 60 #include <sys/umtxvar.h> 61 #include <sys/vmmeter.h> 62 #include <sys/cpuset.h> 63 #ifdef HWPMC_HOOKS 64 #include <sys/pmckern.h> 65 #endif 66 #include <sys/priv.h> 67 68 #include <security/audit/audit.h> 69 70 #include <vm/pmap.h> 71 #include <vm/vm.h> 72 #include <vm/vm_extern.h> 73 #include <vm/uma.h> 74 #include <vm/vm_phys.h> 75 #include <sys/eventhandler.h> 76 77 /* 78 * Asserts below verify the stability of struct thread and struct proc 79 * layout, as exposed by KBI to modules. On head, the KBI is allowed 80 * to drift, change to the structures must be accompanied by the 81 * assert update. 82 * 83 * On the stable branches after KBI freeze, conditions must not be 84 * violated. Typically new fields are moved to the end of the 85 * structures. 86 */ 87 #ifdef __amd64__ 88 _Static_assert(offsetof(struct thread, td_flags) == 0x108, 89 "struct thread KBI td_flags"); 90 _Static_assert(offsetof(struct thread, td_pflags) == 0x110, 91 "struct thread KBI td_pflags"); 92 _Static_assert(offsetof(struct thread, td_frame) == 0x4a8, 93 "struct thread KBI td_frame"); 94 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0, 95 "struct thread KBI td_emuldata"); 96 _Static_assert(offsetof(struct proc, p_flag) == 0xb8, 97 "struct proc KBI p_flag"); 98 _Static_assert(offsetof(struct proc, p_pid) == 0xc4, 99 "struct proc KBI p_pid"); 100 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8, 101 "struct proc KBI p_filemon"); 102 _Static_assert(offsetof(struct proc, p_comm) == 0x3e4, 103 "struct proc KBI p_comm"); 104 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4c8, 105 "struct proc KBI p_emuldata"); 106 #endif 107 #ifdef __i386__ 108 _Static_assert(offsetof(struct thread, td_flags) == 0x9c, 109 "struct thread KBI td_flags"); 110 _Static_assert(offsetof(struct thread, td_pflags) == 0xa4, 111 "struct thread KBI td_pflags"); 112 _Static_assert(offsetof(struct thread, td_frame) == 0x308, 113 "struct thread KBI td_frame"); 114 _Static_assert(offsetof(struct thread, td_emuldata) == 0x34c, 115 "struct thread KBI td_emuldata"); 116 _Static_assert(offsetof(struct proc, p_flag) == 0x6c, 117 "struct proc KBI p_flag"); 118 _Static_assert(offsetof(struct proc, p_pid) == 0x78, 119 "struct proc KBI p_pid"); 120 _Static_assert(offsetof(struct proc, p_filemon) == 0x270, 121 "struct proc KBI p_filemon"); 122 _Static_assert(offsetof(struct proc, p_comm) == 0x288, 123 "struct proc KBI p_comm"); 124 _Static_assert(offsetof(struct proc, p_emuldata) == 0x314, 125 "struct proc KBI p_emuldata"); 126 #endif 127 128 SDT_PROVIDER_DECLARE(proc); 129 SDT_PROBE_DEFINE(proc, , , lwp__exit); 130 131 /* 132 * thread related storage. 133 */ 134 static uma_zone_t thread_zone; 135 136 struct thread_domain_data { 137 struct thread *tdd_zombies; 138 int tdd_reapticks; 139 } __aligned(CACHE_LINE_SIZE); 140 141 static struct thread_domain_data thread_domain_data[MAXMEMDOM]; 142 143 static struct task thread_reap_task; 144 static struct callout thread_reap_callout; 145 146 static void thread_zombie(struct thread *); 147 static void thread_reap(void); 148 static void thread_reap_all(void); 149 static void thread_reap_task_cb(void *, int); 150 static void thread_reap_callout_cb(void *); 151 static int thread_unsuspend_one(struct thread *td, struct proc *p, 152 bool boundary); 153 static void thread_free_batched(struct thread *td); 154 155 static __exclusive_cache_line struct mtx tid_lock; 156 static bitstr_t *tid_bitmap; 157 158 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash"); 159 160 static int maxthread; 161 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN, 162 &maxthread, 0, "Maximum number of threads"); 163 164 static __exclusive_cache_line int nthreads; 165 166 static LIST_HEAD(tidhashhead, thread) *tidhashtbl; 167 static u_long tidhash; 168 static u_long tidhashlock; 169 static struct rwlock *tidhashtbl_lock; 170 #define TIDHASH(tid) (&tidhashtbl[(tid) & tidhash]) 171 #define TIDHASHLOCK(tid) (&tidhashtbl_lock[(tid) & tidhashlock]) 172 173 EVENTHANDLER_LIST_DEFINE(thread_ctor); 174 EVENTHANDLER_LIST_DEFINE(thread_dtor); 175 EVENTHANDLER_LIST_DEFINE(thread_init); 176 EVENTHANDLER_LIST_DEFINE(thread_fini); 177 178 static bool 179 thread_count_inc_try(void) 180 { 181 int nthreads_new; 182 183 nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1; 184 if (nthreads_new >= maxthread - 100) { 185 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 || 186 nthreads_new >= maxthread) { 187 atomic_subtract_int(&nthreads, 1); 188 return (false); 189 } 190 } 191 return (true); 192 } 193 194 static bool 195 thread_count_inc(void) 196 { 197 static struct timeval lastfail; 198 static int curfail; 199 200 thread_reap(); 201 if (thread_count_inc_try()) { 202 return (true); 203 } 204 205 thread_reap_all(); 206 if (thread_count_inc_try()) { 207 return (true); 208 } 209 210 if (ppsratecheck(&lastfail, &curfail, 1)) { 211 printf("maxthread limit exceeded by uid %u " 212 "(pid %d); consider increasing kern.maxthread\n", 213 curthread->td_ucred->cr_ruid, curproc->p_pid); 214 } 215 return (false); 216 } 217 218 static void 219 thread_count_sub(int n) 220 { 221 222 atomic_subtract_int(&nthreads, n); 223 } 224 225 static void 226 thread_count_dec(void) 227 { 228 229 thread_count_sub(1); 230 } 231 232 static lwpid_t 233 tid_alloc(void) 234 { 235 static lwpid_t trytid; 236 lwpid_t tid; 237 238 mtx_lock(&tid_lock); 239 /* 240 * It is an invariant that the bitmap is big enough to hold maxthread 241 * IDs. If we got to this point there has to be at least one free. 242 */ 243 if (trytid >= maxthread) 244 trytid = 0; 245 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid); 246 if (tid == -1) { 247 KASSERT(trytid != 0, ("unexpectedly ran out of IDs")); 248 trytid = 0; 249 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid); 250 KASSERT(tid != -1, ("unexpectedly ran out of IDs")); 251 } 252 bit_set(tid_bitmap, tid); 253 trytid = tid + 1; 254 mtx_unlock(&tid_lock); 255 return (tid + NO_PID); 256 } 257 258 static void 259 tid_free_locked(lwpid_t rtid) 260 { 261 lwpid_t tid; 262 263 mtx_assert(&tid_lock, MA_OWNED); 264 KASSERT(rtid >= NO_PID, 265 ("%s: invalid tid %d\n", __func__, rtid)); 266 tid = rtid - NO_PID; 267 KASSERT(bit_test(tid_bitmap, tid) != 0, 268 ("thread ID %d not allocated\n", rtid)); 269 bit_clear(tid_bitmap, tid); 270 } 271 272 static void 273 tid_free(lwpid_t rtid) 274 { 275 276 mtx_lock(&tid_lock); 277 tid_free_locked(rtid); 278 mtx_unlock(&tid_lock); 279 } 280 281 static void 282 tid_free_batch(lwpid_t *batch, int n) 283 { 284 int i; 285 286 mtx_lock(&tid_lock); 287 for (i = 0; i < n; i++) { 288 tid_free_locked(batch[i]); 289 } 290 mtx_unlock(&tid_lock); 291 } 292 293 /* 294 * Batching for thread reapping. 295 */ 296 struct tidbatch { 297 lwpid_t tab[16]; 298 int n; 299 }; 300 301 static void 302 tidbatch_prep(struct tidbatch *tb) 303 { 304 305 tb->n = 0; 306 } 307 308 static void 309 tidbatch_add(struct tidbatch *tb, struct thread *td) 310 { 311 312 KASSERT(tb->n < nitems(tb->tab), 313 ("%s: count too high %d", __func__, tb->n)); 314 tb->tab[tb->n] = td->td_tid; 315 tb->n++; 316 } 317 318 static void 319 tidbatch_process(struct tidbatch *tb) 320 { 321 322 KASSERT(tb->n <= nitems(tb->tab), 323 ("%s: count too high %d", __func__, tb->n)); 324 if (tb->n == nitems(tb->tab)) { 325 tid_free_batch(tb->tab, tb->n); 326 tb->n = 0; 327 } 328 } 329 330 static void 331 tidbatch_final(struct tidbatch *tb) 332 { 333 334 KASSERT(tb->n <= nitems(tb->tab), 335 ("%s: count too high %d", __func__, tb->n)); 336 if (tb->n != 0) { 337 tid_free_batch(tb->tab, tb->n); 338 } 339 } 340 341 /* 342 * Prepare a thread for use. 343 */ 344 static int 345 thread_ctor(void *mem, int size, void *arg, int flags) 346 { 347 struct thread *td; 348 349 td = (struct thread *)mem; 350 TD_SET_STATE(td, TDS_INACTIVE); 351 td->td_lastcpu = td->td_oncpu = NOCPU; 352 353 /* 354 * Note that td_critnest begins life as 1 because the thread is not 355 * running and is thereby implicitly waiting to be on the receiving 356 * end of a context switch. 357 */ 358 td->td_critnest = 1; 359 td->td_lend_user_pri = PRI_MAX; 360 #ifdef AUDIT 361 audit_thread_alloc(td); 362 #endif 363 #ifdef KDTRACE_HOOKS 364 kdtrace_thread_ctor(td); 365 #endif 366 umtx_thread_alloc(td); 367 MPASS(td->td_sel == NULL); 368 return (0); 369 } 370 371 /* 372 * Reclaim a thread after use. 373 */ 374 static void 375 thread_dtor(void *mem, int size, void *arg) 376 { 377 struct thread *td; 378 379 td = (struct thread *)mem; 380 381 #ifdef INVARIANTS 382 /* Verify that this thread is in a safe state to free. */ 383 switch (TD_GET_STATE(td)) { 384 case TDS_INHIBITED: 385 case TDS_RUNNING: 386 case TDS_CAN_RUN: 387 case TDS_RUNQ: 388 /* 389 * We must never unlink a thread that is in one of 390 * these states, because it is currently active. 391 */ 392 panic("bad state for thread unlinking"); 393 /* NOTREACHED */ 394 case TDS_INACTIVE: 395 break; 396 default: 397 panic("bad thread state"); 398 /* NOTREACHED */ 399 } 400 #endif 401 #ifdef AUDIT 402 audit_thread_free(td); 403 #endif 404 #ifdef KDTRACE_HOOKS 405 kdtrace_thread_dtor(td); 406 #endif 407 /* Free all OSD associated to this thread. */ 408 osd_thread_exit(td); 409 td_softdep_cleanup(td); 410 MPASS(td->td_su == NULL); 411 seltdfini(td); 412 } 413 414 /* 415 * Initialize type-stable parts of a thread (when newly created). 416 */ 417 static int 418 thread_init(void *mem, int size, int flags) 419 { 420 struct thread *td; 421 422 td = (struct thread *)mem; 423 424 td->td_allocdomain = vm_phys_domain(vtophys(td)); 425 td->td_sleepqueue = sleepq_alloc(); 426 td->td_turnstile = turnstile_alloc(); 427 td->td_rlqe = NULL; 428 EVENTHANDLER_DIRECT_INVOKE(thread_init, td); 429 umtx_thread_init(td); 430 td->td_kstack = 0; 431 td->td_sel = NULL; 432 return (0); 433 } 434 435 /* 436 * Tear down type-stable parts of a thread (just before being discarded). 437 */ 438 static void 439 thread_fini(void *mem, int size) 440 { 441 struct thread *td; 442 443 td = (struct thread *)mem; 444 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td); 445 rlqentry_free(td->td_rlqe); 446 turnstile_free(td->td_turnstile); 447 sleepq_free(td->td_sleepqueue); 448 umtx_thread_fini(td); 449 MPASS(td->td_sel == NULL); 450 } 451 452 /* 453 * For a newly created process, 454 * link up all the structures and its initial threads etc. 455 * called from: 456 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc. 457 * proc_dtor() (should go away) 458 * proc_init() 459 */ 460 void 461 proc_linkup0(struct proc *p, struct thread *td) 462 { 463 TAILQ_INIT(&p->p_threads); /* all threads in proc */ 464 proc_linkup(p, td); 465 } 466 467 void 468 proc_linkup(struct proc *p, struct thread *td) 469 { 470 471 sigqueue_init(&p->p_sigqueue, p); 472 p->p_ksi = ksiginfo_alloc(1); 473 if (p->p_ksi != NULL) { 474 /* XXX p_ksi may be null if ksiginfo zone is not ready */ 475 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS; 476 } 477 LIST_INIT(&p->p_mqnotifier); 478 p->p_numthreads = 0; 479 thread_link(td, p); 480 } 481 482 extern int max_threads_per_proc; 483 484 /* 485 * Initialize global thread allocation resources. 486 */ 487 void 488 threadinit(void) 489 { 490 u_long i; 491 lwpid_t tid0; 492 uint32_t flags; 493 494 /* 495 * Place an upper limit on threads which can be allocated. 496 * 497 * Note that other factors may make the de facto limit much lower. 498 * 499 * Platform limits are somewhat arbitrary but deemed "more than good 500 * enough" for the foreseable future. 501 */ 502 if (maxthread == 0) { 503 #ifdef _LP64 504 maxthread = MIN(maxproc * max_threads_per_proc, 1000000); 505 #else 506 maxthread = MIN(maxproc * max_threads_per_proc, 100000); 507 #endif 508 } 509 510 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF); 511 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK); 512 /* 513 * Handle thread0. 514 */ 515 thread_count_inc(); 516 tid0 = tid_alloc(); 517 if (tid0 != THREAD0_TID) 518 panic("tid0 %d != %d\n", tid0, THREAD0_TID); 519 520 flags = UMA_ZONE_NOFREE; 521 #ifdef __aarch64__ 522 /* 523 * Force thread structures to be allocated from the direct map. 524 * Otherwise, superpage promotions and demotions may temporarily 525 * invalidate thread structure mappings. For most dynamically allocated 526 * structures this is not a problem, but translation faults cannot be 527 * handled without accessing curthread. 528 */ 529 flags |= UMA_ZONE_CONTIG; 530 #endif 531 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(), 532 thread_ctor, thread_dtor, thread_init, thread_fini, 533 32 - 1, flags); 534 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash); 535 tidhashlock = (tidhash + 1) / 64; 536 if (tidhashlock > 0) 537 tidhashlock--; 538 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1), 539 M_TIDHASH, M_WAITOK | M_ZERO); 540 for (i = 0; i < tidhashlock + 1; i++) 541 rw_init(&tidhashtbl_lock[i], "tidhash"); 542 543 TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL); 544 callout_init(&thread_reap_callout, 1); 545 callout_reset(&thread_reap_callout, 5 * hz, 546 thread_reap_callout_cb, NULL); 547 } 548 549 /* 550 * Place an unused thread on the zombie list. 551 */ 552 void 553 thread_zombie(struct thread *td) 554 { 555 struct thread_domain_data *tdd; 556 struct thread *ztd; 557 558 tdd = &thread_domain_data[td->td_allocdomain]; 559 ztd = atomic_load_ptr(&tdd->tdd_zombies); 560 for (;;) { 561 td->td_zombie = ztd; 562 if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies, 563 (uintptr_t *)&ztd, (uintptr_t)td)) 564 break; 565 continue; 566 } 567 } 568 569 /* 570 * Release a thread that has exited after cpu_throw(). 571 */ 572 void 573 thread_stash(struct thread *td) 574 { 575 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1); 576 thread_zombie(td); 577 } 578 579 /* 580 * Reap zombies from passed domain. 581 */ 582 static void 583 thread_reap_domain(struct thread_domain_data *tdd) 584 { 585 struct thread *itd, *ntd; 586 struct tidbatch tidbatch; 587 struct credbatch credbatch; 588 int tdcount; 589 struct plimit *lim; 590 int limcount; 591 592 /* 593 * Reading upfront is pessimal if followed by concurrent atomic_swap, 594 * but most of the time the list is empty. 595 */ 596 if (tdd->tdd_zombies == NULL) 597 return; 598 599 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies, 600 (uintptr_t)NULL); 601 if (itd == NULL) 602 return; 603 604 /* 605 * Multiple CPUs can get here, the race is fine as ticks is only 606 * advisory. 607 */ 608 tdd->tdd_reapticks = ticks; 609 610 tidbatch_prep(&tidbatch); 611 credbatch_prep(&credbatch); 612 tdcount = 0; 613 lim = NULL; 614 limcount = 0; 615 616 while (itd != NULL) { 617 ntd = itd->td_zombie; 618 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd); 619 tidbatch_add(&tidbatch, itd); 620 credbatch_add(&credbatch, itd); 621 MPASS(itd->td_limit != NULL); 622 if (lim != itd->td_limit) { 623 if (limcount != 0) { 624 lim_freen(lim, limcount); 625 limcount = 0; 626 } 627 } 628 lim = itd->td_limit; 629 limcount++; 630 thread_free_batched(itd); 631 tidbatch_process(&tidbatch); 632 credbatch_process(&credbatch); 633 tdcount++; 634 if (tdcount == 32) { 635 thread_count_sub(tdcount); 636 tdcount = 0; 637 } 638 itd = ntd; 639 } 640 641 tidbatch_final(&tidbatch); 642 credbatch_final(&credbatch); 643 if (tdcount != 0) { 644 thread_count_sub(tdcount); 645 } 646 MPASS(limcount != 0); 647 lim_freen(lim, limcount); 648 } 649 650 /* 651 * Reap zombies from all domains. 652 */ 653 static void 654 thread_reap_all(void) 655 { 656 struct thread_domain_data *tdd; 657 int i, domain; 658 659 domain = PCPU_GET(domain); 660 for (i = 0; i < vm_ndomains; i++) { 661 tdd = &thread_domain_data[(i + domain) % vm_ndomains]; 662 thread_reap_domain(tdd); 663 } 664 } 665 666 /* 667 * Reap zombies from local domain. 668 */ 669 static void 670 thread_reap(void) 671 { 672 struct thread_domain_data *tdd; 673 int domain; 674 675 domain = PCPU_GET(domain); 676 tdd = &thread_domain_data[domain]; 677 678 thread_reap_domain(tdd); 679 } 680 681 static void 682 thread_reap_task_cb(void *arg __unused, int pending __unused) 683 { 684 685 thread_reap_all(); 686 } 687 688 static void 689 thread_reap_callout_cb(void *arg __unused) 690 { 691 struct thread_domain_data *tdd; 692 int i, cticks, lticks; 693 bool wantreap; 694 695 wantreap = false; 696 cticks = atomic_load_int(&ticks); 697 for (i = 0; i < vm_ndomains; i++) { 698 tdd = &thread_domain_data[i]; 699 lticks = tdd->tdd_reapticks; 700 if (tdd->tdd_zombies != NULL && 701 (u_int)(cticks - lticks) > 5 * hz) { 702 wantreap = true; 703 break; 704 } 705 } 706 707 if (wantreap) 708 taskqueue_enqueue(taskqueue_thread, &thread_reap_task); 709 callout_reset(&thread_reap_callout, 5 * hz, 710 thread_reap_callout_cb, NULL); 711 } 712 713 /* 714 * Calling this function guarantees that any thread that exited before 715 * the call is reaped when the function returns. By 'exited' we mean 716 * a thread removed from the process linkage with thread_unlink(). 717 * Practically this means that caller must lock/unlock corresponding 718 * process lock before the call, to synchronize with thread_exit(). 719 */ 720 void 721 thread_reap_barrier(void) 722 { 723 struct task *t; 724 725 /* 726 * First do context switches to each CPU to ensure that all 727 * PCPU pc_deadthreads are moved to zombie list. 728 */ 729 quiesce_all_cpus("", PDROP); 730 731 /* 732 * Second, fire the task in the same thread as normal 733 * thread_reap() is done, to serialize reaping. 734 */ 735 t = malloc(sizeof(*t), M_TEMP, M_WAITOK); 736 TASK_INIT(t, 0, thread_reap_task_cb, t); 737 taskqueue_enqueue(taskqueue_thread, t); 738 taskqueue_drain(taskqueue_thread, t); 739 free(t, M_TEMP); 740 } 741 742 /* 743 * Allocate a thread. 744 */ 745 struct thread * 746 thread_alloc(int pages) 747 { 748 struct thread *td; 749 lwpid_t tid; 750 751 if (!thread_count_inc()) { 752 return (NULL); 753 } 754 755 tid = tid_alloc(); 756 td = uma_zalloc(thread_zone, M_WAITOK); 757 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack")); 758 if (!vm_thread_new(td, pages)) { 759 uma_zfree(thread_zone, td); 760 tid_free(tid); 761 thread_count_dec(); 762 return (NULL); 763 } 764 td->td_tid = tid; 765 bzero(&td->td_sa.args, sizeof(td->td_sa.args)); 766 kmsan_thread_alloc(td); 767 cpu_thread_alloc(td); 768 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td); 769 return (td); 770 } 771 772 int 773 thread_alloc_stack(struct thread *td, int pages) 774 { 775 776 KASSERT(td->td_kstack == 0, 777 ("thread_alloc_stack called on a thread with kstack")); 778 if (!vm_thread_new(td, pages)) 779 return (0); 780 cpu_thread_alloc(td); 781 return (1); 782 } 783 784 /* 785 * Deallocate a thread. 786 */ 787 static void 788 thread_free_batched(struct thread *td) 789 { 790 791 lock_profile_thread_exit(td); 792 if (td->td_cpuset) 793 cpuset_rel(td->td_cpuset); 794 td->td_cpuset = NULL; 795 cpu_thread_free(td); 796 if (td->td_kstack != 0) 797 vm_thread_dispose(td); 798 callout_drain(&td->td_slpcallout); 799 /* 800 * Freeing handled by the caller. 801 */ 802 td->td_tid = -1; 803 kmsan_thread_free(td); 804 uma_zfree(thread_zone, td); 805 } 806 807 void 808 thread_free(struct thread *td) 809 { 810 lwpid_t tid; 811 812 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td); 813 tid = td->td_tid; 814 thread_free_batched(td); 815 tid_free(tid); 816 thread_count_dec(); 817 } 818 819 void 820 thread_cow_get_proc(struct thread *newtd, struct proc *p) 821 { 822 823 PROC_LOCK_ASSERT(p, MA_OWNED); 824 newtd->td_realucred = crcowget(p->p_ucred); 825 newtd->td_ucred = newtd->td_realucred; 826 newtd->td_limit = lim_hold(p->p_limit); 827 newtd->td_cowgen = p->p_cowgen; 828 } 829 830 void 831 thread_cow_get(struct thread *newtd, struct thread *td) 832 { 833 834 MPASS(td->td_realucred == td->td_ucred); 835 newtd->td_realucred = crcowget(td->td_realucred); 836 newtd->td_ucred = newtd->td_realucred; 837 newtd->td_limit = lim_hold(td->td_limit); 838 newtd->td_cowgen = td->td_cowgen; 839 } 840 841 void 842 thread_cow_free(struct thread *td) 843 { 844 845 if (td->td_realucred != NULL) 846 crcowfree(td); 847 if (td->td_limit != NULL) 848 lim_free(td->td_limit); 849 } 850 851 void 852 thread_cow_update(struct thread *td) 853 { 854 struct proc *p; 855 struct ucred *oldcred; 856 struct plimit *oldlimit; 857 858 p = td->td_proc; 859 PROC_LOCK(p); 860 oldcred = crcowsync(); 861 oldlimit = lim_cowsync(); 862 td->td_cowgen = p->p_cowgen; 863 PROC_UNLOCK(p); 864 if (oldcred != NULL) 865 crfree(oldcred); 866 if (oldlimit != NULL) 867 lim_free(oldlimit); 868 } 869 870 void 871 thread_cow_synced(struct thread *td) 872 { 873 struct proc *p; 874 875 p = td->td_proc; 876 PROC_LOCK_ASSERT(p, MA_OWNED); 877 MPASS(td->td_cowgen != p->p_cowgen); 878 MPASS(td->td_ucred == p->p_ucred); 879 MPASS(td->td_limit == p->p_limit); 880 td->td_cowgen = p->p_cowgen; 881 } 882 883 /* 884 * Discard the current thread and exit from its context. 885 * Always called with scheduler locked. 886 * 887 * Because we can't free a thread while we're operating under its context, 888 * push the current thread into our CPU's deadthread holder. This means 889 * we needn't worry about someone else grabbing our context before we 890 * do a cpu_throw(). 891 */ 892 void 893 thread_exit(void) 894 { 895 uint64_t runtime, new_switchtime; 896 struct thread *td; 897 struct thread *td2; 898 struct proc *p; 899 int wakeup_swapper; 900 901 td = curthread; 902 p = td->td_proc; 903 904 PROC_SLOCK_ASSERT(p, MA_OWNED); 905 mtx_assert(&Giant, MA_NOTOWNED); 906 907 PROC_LOCK_ASSERT(p, MA_OWNED); 908 KASSERT(p != NULL, ("thread exiting without a process")); 909 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td, 910 (long)p->p_pid, td->td_name); 911 SDT_PROBE0(proc, , , lwp__exit); 912 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending")); 913 MPASS(td->td_realucred == td->td_ucred); 914 915 /* 916 * drop FPU & debug register state storage, or any other 917 * architecture specific resources that 918 * would not be on a new untouched process. 919 */ 920 cpu_thread_exit(td); 921 922 /* 923 * The last thread is left attached to the process 924 * So that the whole bundle gets recycled. Skip 925 * all this stuff if we never had threads. 926 * EXIT clears all sign of other threads when 927 * it goes to single threading, so the last thread always 928 * takes the short path. 929 */ 930 if (p->p_flag & P_HADTHREADS) { 931 if (p->p_numthreads > 1) { 932 atomic_add_int(&td->td_proc->p_exitthreads, 1); 933 thread_unlink(td); 934 td2 = FIRST_THREAD_IN_PROC(p); 935 sched_exit_thread(td2, td); 936 937 /* 938 * The test below is NOT true if we are the 939 * sole exiting thread. P_STOPPED_SINGLE is unset 940 * in exit1() after it is the only survivor. 941 */ 942 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 943 if (p->p_numthreads == p->p_suspcount) { 944 thread_lock(p->p_singlethread); 945 wakeup_swapper = thread_unsuspend_one( 946 p->p_singlethread, p, false); 947 if (wakeup_swapper) 948 kick_proc0(); 949 } 950 } 951 952 PCPU_SET(deadthread, td); 953 } else { 954 /* 955 * The last thread is exiting.. but not through exit() 956 */ 957 panic ("thread_exit: Last thread exiting on its own"); 958 } 959 } 960 #ifdef HWPMC_HOOKS 961 /* 962 * If this thread is part of a process that is being tracked by hwpmc(4), 963 * inform the module of the thread's impending exit. 964 */ 965 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) { 966 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 967 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL); 968 } else if (PMC_SYSTEM_SAMPLING_ACTIVE()) 969 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL); 970 #endif 971 PROC_UNLOCK(p); 972 PROC_STATLOCK(p); 973 thread_lock(td); 974 PROC_SUNLOCK(p); 975 976 /* Do the same timestamp bookkeeping that mi_switch() would do. */ 977 new_switchtime = cpu_ticks(); 978 runtime = new_switchtime - PCPU_GET(switchtime); 979 td->td_runtime += runtime; 980 td->td_incruntime += runtime; 981 PCPU_SET(switchtime, new_switchtime); 982 PCPU_SET(switchticks, ticks); 983 VM_CNT_INC(v_swtch); 984 985 /* Save our resource usage in our process. */ 986 td->td_ru.ru_nvcsw++; 987 ruxagg_locked(p, td); 988 rucollect(&p->p_ru, &td->td_ru); 989 PROC_STATUNLOCK(p); 990 991 TD_SET_STATE(td, TDS_INACTIVE); 992 #ifdef WITNESS 993 witness_thread_exit(td); 994 #endif 995 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td); 996 sched_throw(td); 997 panic("I'm a teapot!"); 998 /* NOTREACHED */ 999 } 1000 1001 /* 1002 * Do any thread specific cleanups that may be needed in wait() 1003 * called with Giant, proc and schedlock not held. 1004 */ 1005 void 1006 thread_wait(struct proc *p) 1007 { 1008 struct thread *td; 1009 1010 mtx_assert(&Giant, MA_NOTOWNED); 1011 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()")); 1012 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking")); 1013 td = FIRST_THREAD_IN_PROC(p); 1014 /* Lock the last thread so we spin until it exits cpu_throw(). */ 1015 thread_lock(td); 1016 thread_unlock(td); 1017 lock_profile_thread_exit(td); 1018 cpuset_rel(td->td_cpuset); 1019 td->td_cpuset = NULL; 1020 cpu_thread_clean(td); 1021 thread_cow_free(td); 1022 callout_drain(&td->td_slpcallout); 1023 thread_reap(); /* check for zombie threads etc. */ 1024 } 1025 1026 /* 1027 * Link a thread to a process. 1028 * set up anything that needs to be initialized for it to 1029 * be used by the process. 1030 */ 1031 void 1032 thread_link(struct thread *td, struct proc *p) 1033 { 1034 1035 /* 1036 * XXX This can't be enabled because it's called for proc0 before 1037 * its lock has been created. 1038 * PROC_LOCK_ASSERT(p, MA_OWNED); 1039 */ 1040 TD_SET_STATE(td, TDS_INACTIVE); 1041 td->td_proc = p; 1042 td->td_flags = TDF_INMEM; 1043 1044 LIST_INIT(&td->td_contested); 1045 LIST_INIT(&td->td_lprof[0]); 1046 LIST_INIT(&td->td_lprof[1]); 1047 #ifdef EPOCH_TRACE 1048 SLIST_INIT(&td->td_epochs); 1049 #endif 1050 sigqueue_init(&td->td_sigqueue, p); 1051 callout_init(&td->td_slpcallout, 1); 1052 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist); 1053 p->p_numthreads++; 1054 } 1055 1056 /* 1057 * Called from: 1058 * thread_exit() 1059 */ 1060 void 1061 thread_unlink(struct thread *td) 1062 { 1063 struct proc *p = td->td_proc; 1064 1065 PROC_LOCK_ASSERT(p, MA_OWNED); 1066 #ifdef EPOCH_TRACE 1067 MPASS(SLIST_EMPTY(&td->td_epochs)); 1068 #endif 1069 1070 TAILQ_REMOVE(&p->p_threads, td, td_plist); 1071 p->p_numthreads--; 1072 /* could clear a few other things here */ 1073 /* Must NOT clear links to proc! */ 1074 } 1075 1076 static int 1077 calc_remaining(struct proc *p, int mode) 1078 { 1079 int remaining; 1080 1081 PROC_LOCK_ASSERT(p, MA_OWNED); 1082 PROC_SLOCK_ASSERT(p, MA_OWNED); 1083 if (mode == SINGLE_EXIT) 1084 remaining = p->p_numthreads; 1085 else if (mode == SINGLE_BOUNDARY) 1086 remaining = p->p_numthreads - p->p_boundary_count; 1087 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC) 1088 remaining = p->p_numthreads - p->p_suspcount; 1089 else 1090 panic("calc_remaining: wrong mode %d", mode); 1091 return (remaining); 1092 } 1093 1094 static int 1095 remain_for_mode(int mode) 1096 { 1097 1098 return (mode == SINGLE_ALLPROC ? 0 : 1); 1099 } 1100 1101 static int 1102 weed_inhib(int mode, struct thread *td2, struct proc *p) 1103 { 1104 int wakeup_swapper; 1105 1106 PROC_LOCK_ASSERT(p, MA_OWNED); 1107 PROC_SLOCK_ASSERT(p, MA_OWNED); 1108 THREAD_LOCK_ASSERT(td2, MA_OWNED); 1109 1110 wakeup_swapper = 0; 1111 1112 /* 1113 * Since the thread lock is dropped by the scheduler we have 1114 * to retry to check for races. 1115 */ 1116 restart: 1117 switch (mode) { 1118 case SINGLE_EXIT: 1119 if (TD_IS_SUSPENDED(td2)) { 1120 wakeup_swapper |= thread_unsuspend_one(td2, p, true); 1121 thread_lock(td2); 1122 goto restart; 1123 } 1124 if (TD_CAN_ABORT(td2)) { 1125 wakeup_swapper |= sleepq_abort(td2, EINTR); 1126 return (wakeup_swapper); 1127 } 1128 break; 1129 case SINGLE_BOUNDARY: 1130 case SINGLE_NO_EXIT: 1131 if (TD_IS_SUSPENDED(td2) && 1132 (td2->td_flags & TDF_BOUNDARY) == 0) { 1133 wakeup_swapper |= thread_unsuspend_one(td2, p, false); 1134 thread_lock(td2); 1135 goto restart; 1136 } 1137 if (TD_CAN_ABORT(td2)) { 1138 wakeup_swapper |= sleepq_abort(td2, ERESTART); 1139 return (wakeup_swapper); 1140 } 1141 break; 1142 case SINGLE_ALLPROC: 1143 /* 1144 * ALLPROC suspend tries to avoid spurious EINTR for 1145 * threads sleeping interruptable, by suspending the 1146 * thread directly, similarly to sig_suspend_threads(). 1147 * Since such sleep is not neccessary performed at the user 1148 * boundary, TDF_ALLPROCSUSP is used to avoid immediate 1149 * un-suspend. 1150 */ 1151 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & 1152 TDF_ALLPROCSUSP) == 0) { 1153 wakeup_swapper |= thread_unsuspend_one(td2, p, false); 1154 thread_lock(td2); 1155 goto restart; 1156 } 1157 if (TD_CAN_ABORT(td2)) { 1158 td2->td_flags |= TDF_ALLPROCSUSP; 1159 wakeup_swapper |= sleepq_abort(td2, ERESTART); 1160 return (wakeup_swapper); 1161 } 1162 break; 1163 default: 1164 break; 1165 } 1166 thread_unlock(td2); 1167 return (wakeup_swapper); 1168 } 1169 1170 /* 1171 * Enforce single-threading. 1172 * 1173 * Returns 1 if the caller must abort (another thread is waiting to 1174 * exit the process or similar). Process is locked! 1175 * Returns 0 when you are successfully the only thread running. 1176 * A process has successfully single threaded in the suspend mode when 1177 * There are no threads in user mode. Threads in the kernel must be 1178 * allowed to continue until they get to the user boundary. They may even 1179 * copy out their return values and data before suspending. They may however be 1180 * accelerated in reaching the user boundary as we will wake up 1181 * any sleeping threads that are interruptable. (PCATCH). 1182 */ 1183 int 1184 thread_single(struct proc *p, int mode) 1185 { 1186 struct thread *td; 1187 struct thread *td2; 1188 int remaining, wakeup_swapper; 1189 1190 td = curthread; 1191 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 1192 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 1193 ("invalid mode %d", mode)); 1194 /* 1195 * If allowing non-ALLPROC singlethreading for non-curproc 1196 * callers, calc_remaining() and remain_for_mode() should be 1197 * adjusted to also account for td->td_proc != p. For now 1198 * this is not implemented because it is not used. 1199 */ 1200 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) || 1201 (mode != SINGLE_ALLPROC && td->td_proc == p), 1202 ("mode %d proc %p curproc %p", mode, p, td->td_proc)); 1203 mtx_assert(&Giant, MA_NOTOWNED); 1204 PROC_LOCK_ASSERT(p, MA_OWNED); 1205 1206 /* 1207 * Is someone already single threading? 1208 * Or may be singlethreading is not needed at all. 1209 */ 1210 if (mode == SINGLE_ALLPROC) { 1211 while ((p->p_flag & P_STOPPED_SINGLE) != 0) { 1212 if ((p->p_flag2 & P2_WEXIT) != 0) 1213 return (1); 1214 msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0); 1215 } 1216 } else if ((p->p_flag & P_HADTHREADS) == 0) 1217 return (0); 1218 if (p->p_singlethread != NULL && p->p_singlethread != td) 1219 return (1); 1220 1221 if (mode == SINGLE_EXIT) { 1222 p->p_flag |= P_SINGLE_EXIT; 1223 p->p_flag &= ~P_SINGLE_BOUNDARY; 1224 } else { 1225 p->p_flag &= ~P_SINGLE_EXIT; 1226 if (mode == SINGLE_BOUNDARY) 1227 p->p_flag |= P_SINGLE_BOUNDARY; 1228 else 1229 p->p_flag &= ~P_SINGLE_BOUNDARY; 1230 } 1231 if (mode == SINGLE_ALLPROC) { 1232 p->p_flag |= P_TOTAL_STOP; 1233 thread_lock(td); 1234 td->td_flags |= TDF_DOING_SA; 1235 thread_unlock(td); 1236 } 1237 p->p_flag |= P_STOPPED_SINGLE; 1238 PROC_SLOCK(p); 1239 p->p_singlethread = td; 1240 remaining = calc_remaining(p, mode); 1241 while (remaining != remain_for_mode(mode)) { 1242 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE) 1243 goto stopme; 1244 wakeup_swapper = 0; 1245 FOREACH_THREAD_IN_PROC(p, td2) { 1246 if (td2 == td) 1247 continue; 1248 thread_lock(td2); 1249 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK; 1250 if (TD_IS_INHIBITED(td2)) { 1251 wakeup_swapper |= weed_inhib(mode, td2, p); 1252 #ifdef SMP 1253 } else if (TD_IS_RUNNING(td2)) { 1254 forward_signal(td2); 1255 thread_unlock(td2); 1256 #endif 1257 } else 1258 thread_unlock(td2); 1259 } 1260 if (wakeup_swapper) 1261 kick_proc0(); 1262 remaining = calc_remaining(p, mode); 1263 1264 /* 1265 * Maybe we suspended some threads.. was it enough? 1266 */ 1267 if (remaining == remain_for_mode(mode)) 1268 break; 1269 1270 stopme: 1271 /* 1272 * Wake us up when everyone else has suspended. 1273 * In the mean time we suspend as well. 1274 */ 1275 thread_suspend_switch(td, p); 1276 remaining = calc_remaining(p, mode); 1277 } 1278 if (mode == SINGLE_EXIT) { 1279 /* 1280 * Convert the process to an unthreaded process. The 1281 * SINGLE_EXIT is called by exit1() or execve(), in 1282 * both cases other threads must be retired. 1283 */ 1284 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads")); 1285 p->p_singlethread = NULL; 1286 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS); 1287 1288 /* 1289 * Wait for any remaining threads to exit cpu_throw(). 1290 */ 1291 while (p->p_exitthreads != 0) { 1292 PROC_SUNLOCK(p); 1293 PROC_UNLOCK(p); 1294 sched_relinquish(td); 1295 PROC_LOCK(p); 1296 PROC_SLOCK(p); 1297 } 1298 } else if (mode == SINGLE_BOUNDARY) { 1299 /* 1300 * Wait until all suspended threads are removed from 1301 * the processors. The thread_suspend_check() 1302 * increments p_boundary_count while it is still 1303 * running, which makes it possible for the execve() 1304 * to destroy vmspace while our other threads are 1305 * still using the address space. 1306 * 1307 * We lock the thread, which is only allowed to 1308 * succeed after context switch code finished using 1309 * the address space. 1310 */ 1311 FOREACH_THREAD_IN_PROC(p, td2) { 1312 if (td2 == td) 1313 continue; 1314 thread_lock(td2); 1315 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0, 1316 ("td %p not on boundary", td2)); 1317 KASSERT(TD_IS_SUSPENDED(td2), 1318 ("td %p is not suspended", td2)); 1319 thread_unlock(td2); 1320 } 1321 } 1322 PROC_SUNLOCK(p); 1323 if (mode == SINGLE_ALLPROC) { 1324 thread_lock(td); 1325 td->td_flags &= ~TDF_DOING_SA; 1326 thread_unlock(td); 1327 } 1328 return (0); 1329 } 1330 1331 bool 1332 thread_suspend_check_needed(void) 1333 { 1334 struct proc *p; 1335 struct thread *td; 1336 1337 td = curthread; 1338 p = td->td_proc; 1339 PROC_LOCK_ASSERT(p, MA_OWNED); 1340 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 && 1341 (td->td_dbgflags & TDB_SUSPEND) != 0)); 1342 } 1343 1344 /* 1345 * Called in from locations that can safely check to see 1346 * whether we have to suspend or at least throttle for a 1347 * single-thread event (e.g. fork). 1348 * 1349 * Such locations include userret(). 1350 * If the "return_instead" argument is non zero, the thread must be able to 1351 * accept 0 (caller may continue), or 1 (caller must abort) as a result. 1352 * 1353 * The 'return_instead' argument tells the function if it may do a 1354 * thread_exit() or suspend, or whether the caller must abort and back 1355 * out instead. 1356 * 1357 * If the thread that set the single_threading request has set the 1358 * P_SINGLE_EXIT bit in the process flags then this call will never return 1359 * if 'return_instead' is false, but will exit. 1360 * 1361 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0 1362 *---------------+--------------------+--------------------- 1363 * 0 | returns 0 | returns 0 or 1 1364 * | when ST ends | immediately 1365 *---------------+--------------------+--------------------- 1366 * 1 | thread exits | returns 1 1367 * | | immediately 1368 * 0 = thread_exit() or suspension ok, 1369 * other = return error instead of stopping the thread. 1370 * 1371 * While a full suspension is under effect, even a single threading 1372 * thread would be suspended if it made this call (but it shouldn't). 1373 * This call should only be made from places where 1374 * thread_exit() would be safe as that may be the outcome unless 1375 * return_instead is set. 1376 */ 1377 int 1378 thread_suspend_check(int return_instead) 1379 { 1380 struct thread *td; 1381 struct proc *p; 1382 int wakeup_swapper; 1383 1384 td = curthread; 1385 p = td->td_proc; 1386 mtx_assert(&Giant, MA_NOTOWNED); 1387 PROC_LOCK_ASSERT(p, MA_OWNED); 1388 while (thread_suspend_check_needed()) { 1389 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 1390 KASSERT(p->p_singlethread != NULL, 1391 ("singlethread not set")); 1392 /* 1393 * The only suspension in action is a 1394 * single-threading. Single threader need not stop. 1395 * It is safe to access p->p_singlethread unlocked 1396 * because it can only be set to our address by us. 1397 */ 1398 if (p->p_singlethread == td) 1399 return (0); /* Exempt from stopping. */ 1400 } 1401 if ((p->p_flag & P_SINGLE_EXIT) && return_instead) 1402 return (EINTR); 1403 1404 /* Should we goto user boundary if we didn't come from there? */ 1405 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 1406 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead) 1407 return (ERESTART); 1408 1409 /* 1410 * Ignore suspend requests if they are deferred. 1411 */ 1412 if ((td->td_flags & TDF_SBDRY) != 0) { 1413 KASSERT(return_instead, 1414 ("TDF_SBDRY set for unsafe thread_suspend_check")); 1415 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) != 1416 (TDF_SEINTR | TDF_SERESTART), 1417 ("both TDF_SEINTR and TDF_SERESTART")); 1418 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0); 1419 } 1420 1421 /* 1422 * If the process is waiting for us to exit, 1423 * this thread should just suicide. 1424 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE. 1425 */ 1426 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { 1427 PROC_UNLOCK(p); 1428 1429 /* 1430 * Allow Linux emulation layer to do some work 1431 * before thread suicide. 1432 */ 1433 if (__predict_false(p->p_sysent->sv_thread_detach != NULL)) 1434 (p->p_sysent->sv_thread_detach)(td); 1435 umtx_thread_exit(td); 1436 kern_thr_exit(td); 1437 panic("stopped thread did not exit"); 1438 } 1439 1440 PROC_SLOCK(p); 1441 thread_stopped(p); 1442 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 1443 if (p->p_numthreads == p->p_suspcount + 1) { 1444 thread_lock(p->p_singlethread); 1445 wakeup_swapper = thread_unsuspend_one( 1446 p->p_singlethread, p, false); 1447 if (wakeup_swapper) 1448 kick_proc0(); 1449 } 1450 } 1451 PROC_UNLOCK(p); 1452 thread_lock(td); 1453 /* 1454 * When a thread suspends, it just 1455 * gets taken off all queues. 1456 */ 1457 thread_suspend_one(td); 1458 if (return_instead == 0) { 1459 p->p_boundary_count++; 1460 td->td_flags |= TDF_BOUNDARY; 1461 } 1462 PROC_SUNLOCK(p); 1463 mi_switch(SW_INVOL | SWT_SUSPEND); 1464 PROC_LOCK(p); 1465 } 1466 return (0); 1467 } 1468 1469 /* 1470 * Check for possible stops and suspensions while executing a 1471 * casueword or similar transiently failing operation. 1472 * 1473 * The sleep argument controls whether the function can handle a stop 1474 * request itself or it should return ERESTART and the request is 1475 * proceed at the kernel/user boundary in ast. 1476 * 1477 * Typically, when retrying due to casueword(9) failure (rv == 1), we 1478 * should handle the stop requests there, with exception of cases when 1479 * the thread owns a kernel resource, for instance busied the umtx 1480 * key, or when functions return immediately if thread_check_susp() 1481 * returned non-zero. On the other hand, retrying the whole lock 1482 * operation, we better not stop there but delegate the handling to 1483 * ast. 1484 * 1485 * If the request is for thread termination P_SINGLE_EXIT, we cannot 1486 * handle it at all, and simply return EINTR. 1487 */ 1488 int 1489 thread_check_susp(struct thread *td, bool sleep) 1490 { 1491 struct proc *p; 1492 int error; 1493 1494 /* 1495 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to 1496 * eventually break the lockstep loop. 1497 */ 1498 if ((td->td_flags & TDF_NEEDSUSPCHK) == 0) 1499 return (0); 1500 error = 0; 1501 p = td->td_proc; 1502 PROC_LOCK(p); 1503 if (p->p_flag & P_SINGLE_EXIT) 1504 error = EINTR; 1505 else if (P_SHOULDSTOP(p) || 1506 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) 1507 error = sleep ? thread_suspend_check(0) : ERESTART; 1508 PROC_UNLOCK(p); 1509 return (error); 1510 } 1511 1512 void 1513 thread_suspend_switch(struct thread *td, struct proc *p) 1514 { 1515 1516 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1517 PROC_LOCK_ASSERT(p, MA_OWNED); 1518 PROC_SLOCK_ASSERT(p, MA_OWNED); 1519 /* 1520 * We implement thread_suspend_one in stages here to avoid 1521 * dropping the proc lock while the thread lock is owned. 1522 */ 1523 if (p == td->td_proc) { 1524 thread_stopped(p); 1525 p->p_suspcount++; 1526 } 1527 PROC_UNLOCK(p); 1528 thread_lock(td); 1529 td->td_flags &= ~TDF_NEEDSUSPCHK; 1530 TD_SET_SUSPENDED(td); 1531 sched_sleep(td, 0); 1532 PROC_SUNLOCK(p); 1533 DROP_GIANT(); 1534 mi_switch(SW_VOL | SWT_SUSPEND); 1535 PICKUP_GIANT(); 1536 PROC_LOCK(p); 1537 PROC_SLOCK(p); 1538 } 1539 1540 void 1541 thread_suspend_one(struct thread *td) 1542 { 1543 struct proc *p; 1544 1545 p = td->td_proc; 1546 PROC_SLOCK_ASSERT(p, MA_OWNED); 1547 THREAD_LOCK_ASSERT(td, MA_OWNED); 1548 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1549 p->p_suspcount++; 1550 td->td_flags &= ~TDF_NEEDSUSPCHK; 1551 TD_SET_SUSPENDED(td); 1552 sched_sleep(td, 0); 1553 } 1554 1555 static int 1556 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary) 1557 { 1558 1559 THREAD_LOCK_ASSERT(td, MA_OWNED); 1560 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); 1561 TD_CLR_SUSPENDED(td); 1562 td->td_flags &= ~TDF_ALLPROCSUSP; 1563 if (td->td_proc == p) { 1564 PROC_SLOCK_ASSERT(p, MA_OWNED); 1565 p->p_suspcount--; 1566 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) { 1567 td->td_flags &= ~TDF_BOUNDARY; 1568 p->p_boundary_count--; 1569 } 1570 } 1571 return (setrunnable(td, 0)); 1572 } 1573 1574 void 1575 thread_run_flash(struct thread *td) 1576 { 1577 struct proc *p; 1578 1579 p = td->td_proc; 1580 PROC_LOCK_ASSERT(p, MA_OWNED); 1581 1582 if (TD_ON_SLEEPQ(td)) 1583 sleepq_remove_nested(td); 1584 else 1585 thread_lock(td); 1586 1587 THREAD_LOCK_ASSERT(td, MA_OWNED); 1588 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); 1589 1590 TD_CLR_SUSPENDED(td); 1591 PROC_SLOCK(p); 1592 MPASS(p->p_suspcount > 0); 1593 p->p_suspcount--; 1594 PROC_SUNLOCK(p); 1595 if (setrunnable(td, 0)) 1596 kick_proc0(); 1597 } 1598 1599 /* 1600 * Allow all threads blocked by single threading to continue running. 1601 */ 1602 void 1603 thread_unsuspend(struct proc *p) 1604 { 1605 struct thread *td; 1606 int wakeup_swapper; 1607 1608 PROC_LOCK_ASSERT(p, MA_OWNED); 1609 PROC_SLOCK_ASSERT(p, MA_OWNED); 1610 wakeup_swapper = 0; 1611 if (!P_SHOULDSTOP(p)) { 1612 FOREACH_THREAD_IN_PROC(p, td) { 1613 thread_lock(td); 1614 if (TD_IS_SUSPENDED(td) && (td->td_flags & 1615 TDF_DOING_SA) == 0) { 1616 wakeup_swapper |= thread_unsuspend_one(td, p, 1617 true); 1618 } else 1619 thread_unlock(td); 1620 } 1621 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 1622 p->p_numthreads == p->p_suspcount) { 1623 /* 1624 * Stopping everything also did the job for the single 1625 * threading request. Now we've downgraded to single-threaded, 1626 * let it continue. 1627 */ 1628 if (p->p_singlethread->td_proc == p) { 1629 thread_lock(p->p_singlethread); 1630 wakeup_swapper = thread_unsuspend_one( 1631 p->p_singlethread, p, false); 1632 } 1633 } 1634 if (wakeup_swapper) 1635 kick_proc0(); 1636 } 1637 1638 /* 1639 * End the single threading mode.. 1640 */ 1641 void 1642 thread_single_end(struct proc *p, int mode) 1643 { 1644 struct thread *td; 1645 int wakeup_swapper; 1646 1647 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 1648 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 1649 ("invalid mode %d", mode)); 1650 PROC_LOCK_ASSERT(p, MA_OWNED); 1651 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) || 1652 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0), 1653 ("mode %d does not match P_TOTAL_STOP", mode)); 1654 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread, 1655 ("thread_single_end from other thread %p %p", 1656 curthread, p->p_singlethread)); 1657 KASSERT(mode != SINGLE_BOUNDARY || 1658 (p->p_flag & P_SINGLE_BOUNDARY) != 0, 1659 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag)); 1660 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY | 1661 P_TOTAL_STOP); 1662 PROC_SLOCK(p); 1663 p->p_singlethread = NULL; 1664 wakeup_swapper = 0; 1665 /* 1666 * If there are other threads they may now run, 1667 * unless of course there is a blanket 'stop order' 1668 * on the process. The single threader must be allowed 1669 * to continue however as this is a bad place to stop. 1670 */ 1671 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) { 1672 FOREACH_THREAD_IN_PROC(p, td) { 1673 thread_lock(td); 1674 if (TD_IS_SUSPENDED(td)) { 1675 wakeup_swapper |= thread_unsuspend_one(td, p, 1676 true); 1677 } else 1678 thread_unlock(td); 1679 } 1680 } 1681 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0, 1682 ("inconsistent boundary count %d", p->p_boundary_count)); 1683 PROC_SUNLOCK(p); 1684 if (wakeup_swapper) 1685 kick_proc0(); 1686 wakeup(&p->p_flag); 1687 } 1688 1689 /* 1690 * Locate a thread by number and return with proc lock held. 1691 * 1692 * thread exit establishes proc -> tidhash lock ordering, but lookup 1693 * takes tidhash first and needs to return locked proc. 1694 * 1695 * The problem is worked around by relying on type-safety of both 1696 * structures and doing the work in 2 steps: 1697 * - tidhash-locked lookup which saves both thread and proc pointers 1698 * - proc-locked verification that the found thread still matches 1699 */ 1700 static bool 1701 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp) 1702 { 1703 #define RUN_THRESH 16 1704 struct proc *p; 1705 struct thread *td; 1706 int run; 1707 bool locked; 1708 1709 run = 0; 1710 rw_rlock(TIDHASHLOCK(tid)); 1711 locked = true; 1712 LIST_FOREACH(td, TIDHASH(tid), td_hash) { 1713 if (td->td_tid != tid) { 1714 run++; 1715 continue; 1716 } 1717 p = td->td_proc; 1718 if (pid != -1 && p->p_pid != pid) { 1719 td = NULL; 1720 break; 1721 } 1722 if (run > RUN_THRESH) { 1723 if (rw_try_upgrade(TIDHASHLOCK(tid))) { 1724 LIST_REMOVE(td, td_hash); 1725 LIST_INSERT_HEAD(TIDHASH(td->td_tid), 1726 td, td_hash); 1727 rw_wunlock(TIDHASHLOCK(tid)); 1728 locked = false; 1729 break; 1730 } 1731 } 1732 break; 1733 } 1734 if (locked) 1735 rw_runlock(TIDHASHLOCK(tid)); 1736 if (td == NULL) 1737 return (false); 1738 *pp = p; 1739 *tdp = td; 1740 return (true); 1741 } 1742 1743 struct thread * 1744 tdfind(lwpid_t tid, pid_t pid) 1745 { 1746 struct proc *p; 1747 struct thread *td; 1748 1749 td = curthread; 1750 if (td->td_tid == tid) { 1751 if (pid != -1 && td->td_proc->p_pid != pid) 1752 return (NULL); 1753 PROC_LOCK(td->td_proc); 1754 return (td); 1755 } 1756 1757 for (;;) { 1758 if (!tdfind_hash(tid, pid, &p, &td)) 1759 return (NULL); 1760 PROC_LOCK(p); 1761 if (td->td_tid != tid) { 1762 PROC_UNLOCK(p); 1763 continue; 1764 } 1765 if (td->td_proc != p) { 1766 PROC_UNLOCK(p); 1767 continue; 1768 } 1769 if (p->p_state == PRS_NEW) { 1770 PROC_UNLOCK(p); 1771 return (NULL); 1772 } 1773 return (td); 1774 } 1775 } 1776 1777 void 1778 tidhash_add(struct thread *td) 1779 { 1780 rw_wlock(TIDHASHLOCK(td->td_tid)); 1781 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash); 1782 rw_wunlock(TIDHASHLOCK(td->td_tid)); 1783 } 1784 1785 void 1786 tidhash_remove(struct thread *td) 1787 { 1788 1789 rw_wlock(TIDHASHLOCK(td->td_tid)); 1790 LIST_REMOVE(td, td_hash); 1791 rw_wunlock(TIDHASHLOCK(td->td_tid)); 1792 } 1793