1 /*- 2 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>. 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice(s), this list of conditions and the following disclaimer as 10 * the first lines of this file unmodified other than the possible 11 * addition of one or more copyright notices. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice(s), this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY 17 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 18 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 19 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY 20 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 21 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 22 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 23 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH 26 * DAMAGE. 27 */ 28 29 #include "opt_witness.h" 30 #include "opt_hwpmc_hooks.h" 31 32 #include <sys/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/kernel.h> 38 #include <sys/lock.h> 39 #include <sys/mutex.h> 40 #include <sys/proc.h> 41 #include <sys/rangelock.h> 42 #include <sys/resourcevar.h> 43 #include <sys/sdt.h> 44 #include <sys/smp.h> 45 #include <sys/sched.h> 46 #include <sys/sleepqueue.h> 47 #include <sys/selinfo.h> 48 #include <sys/syscallsubr.h> 49 #include <sys/sysent.h> 50 #include <sys/turnstile.h> 51 #include <sys/ktr.h> 52 #include <sys/rwlock.h> 53 #include <sys/umtx.h> 54 #include <sys/vmmeter.h> 55 #include <sys/cpuset.h> 56 #ifdef HWPMC_HOOKS 57 #include <sys/pmckern.h> 58 #endif 59 60 #include <security/audit/audit.h> 61 62 #include <vm/vm.h> 63 #include <vm/vm_extern.h> 64 #include <vm/uma.h> 65 #include <vm/vm_domain.h> 66 #include <sys/eventhandler.h> 67 68 /* 69 * Asserts below verify the stability of struct thread and struct proc 70 * layout, as exposed by KBI to modules. On head, the KBI is allowed 71 * to drift, change to the structures must be accompanied by the 72 * assert update. 73 * 74 * On the stable branches after KBI freeze, conditions must not be 75 * violated. Typically new fields are moved to the end of the 76 * structures. 77 */ 78 #ifdef __amd64__ 79 _Static_assert(offsetof(struct thread, td_flags) == 0xf4, 80 "struct thread KBI td_flags"); 81 _Static_assert(offsetof(struct thread, td_pflags) == 0xfc, 82 "struct thread KBI td_pflags"); 83 _Static_assert(offsetof(struct thread, td_frame) == 0x460, 84 "struct thread KBI td_frame"); 85 _Static_assert(offsetof(struct thread, td_emuldata) == 0x508, 86 "struct thread KBI td_emuldata"); 87 _Static_assert(offsetof(struct proc, p_flag) == 0xb0, 88 "struct proc KBI p_flag"); 89 _Static_assert(offsetof(struct proc, p_pid) == 0xbc, 90 "struct proc KBI p_pid"); 91 _Static_assert(offsetof(struct proc, p_filemon) == 0x3d0, 92 "struct proc KBI p_filemon"); 93 _Static_assert(offsetof(struct proc, p_comm) == 0x3e0, 94 "struct proc KBI p_comm"); 95 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4b8, 96 "struct proc KBI p_emuldata"); 97 #endif 98 #ifdef __i386__ 99 _Static_assert(offsetof(struct thread, td_flags) == 0x9c, 100 "struct thread KBI td_flags"); 101 _Static_assert(offsetof(struct thread, td_pflags) == 0xa4, 102 "struct thread KBI td_pflags"); 103 _Static_assert(offsetof(struct thread, td_frame) == 0x2ec, 104 "struct thread KBI td_frame"); 105 _Static_assert(offsetof(struct thread, td_emuldata) == 0x338, 106 "struct thread KBI td_emuldata"); 107 _Static_assert(offsetof(struct proc, p_flag) == 0x68, 108 "struct proc KBI p_flag"); 109 _Static_assert(offsetof(struct proc, p_pid) == 0x74, 110 "struct proc KBI p_pid"); 111 _Static_assert(offsetof(struct proc, p_filemon) == 0x27c, 112 "struct proc KBI p_filemon"); 113 _Static_assert(offsetof(struct proc, p_comm) == 0x288, 114 "struct proc KBI p_comm"); 115 _Static_assert(offsetof(struct proc, p_emuldata) == 0x314, 116 "struct proc KBI p_emuldata"); 117 #endif 118 119 SDT_PROVIDER_DECLARE(proc); 120 SDT_PROBE_DEFINE(proc, , , lwp__exit); 121 122 /* 123 * thread related storage. 124 */ 125 static uma_zone_t thread_zone; 126 127 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads); 128 static struct mtx zombie_lock; 129 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN); 130 131 static void thread_zombie(struct thread *); 132 static int thread_unsuspend_one(struct thread *td, struct proc *p, 133 bool boundary); 134 135 #define TID_BUFFER_SIZE 1024 136 137 struct mtx tid_lock; 138 static struct unrhdr *tid_unrhdr; 139 static lwpid_t tid_buffer[TID_BUFFER_SIZE]; 140 static int tid_head, tid_tail; 141 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash"); 142 143 struct tidhashhead *tidhashtbl; 144 u_long tidhash; 145 struct rwlock tidhash_lock; 146 147 EVENTHANDLER_LIST_DEFINE(thread_ctor); 148 EVENTHANDLER_LIST_DEFINE(thread_dtor); 149 EVENTHANDLER_LIST_DEFINE(thread_init); 150 EVENTHANDLER_LIST_DEFINE(thread_fini); 151 152 static lwpid_t 153 tid_alloc(void) 154 { 155 lwpid_t tid; 156 157 tid = alloc_unr(tid_unrhdr); 158 if (tid != -1) 159 return (tid); 160 mtx_lock(&tid_lock); 161 if (tid_head == tid_tail) { 162 mtx_unlock(&tid_lock); 163 return (-1); 164 } 165 tid = tid_buffer[tid_head]; 166 tid_head = (tid_head + 1) % TID_BUFFER_SIZE; 167 mtx_unlock(&tid_lock); 168 return (tid); 169 } 170 171 static void 172 tid_free(lwpid_t tid) 173 { 174 lwpid_t tmp_tid = -1; 175 176 mtx_lock(&tid_lock); 177 if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) { 178 tmp_tid = tid_buffer[tid_head]; 179 tid_head = (tid_head + 1) % TID_BUFFER_SIZE; 180 } 181 tid_buffer[tid_tail] = tid; 182 tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE; 183 mtx_unlock(&tid_lock); 184 if (tmp_tid != -1) 185 free_unr(tid_unrhdr, tmp_tid); 186 } 187 188 /* 189 * Prepare a thread for use. 190 */ 191 static int 192 thread_ctor(void *mem, int size, void *arg, int flags) 193 { 194 struct thread *td; 195 196 td = (struct thread *)mem; 197 td->td_state = TDS_INACTIVE; 198 td->td_oncpu = NOCPU; 199 200 td->td_tid = tid_alloc(); 201 202 /* 203 * Note that td_critnest begins life as 1 because the thread is not 204 * running and is thereby implicitly waiting to be on the receiving 205 * end of a context switch. 206 */ 207 td->td_critnest = 1; 208 td->td_lend_user_pri = PRI_MAX; 209 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td); 210 #ifdef AUDIT 211 audit_thread_alloc(td); 212 #endif 213 umtx_thread_alloc(td); 214 return (0); 215 } 216 217 /* 218 * Reclaim a thread after use. 219 */ 220 static void 221 thread_dtor(void *mem, int size, void *arg) 222 { 223 struct thread *td; 224 225 td = (struct thread *)mem; 226 227 #ifdef INVARIANTS 228 /* Verify that this thread is in a safe state to free. */ 229 switch (td->td_state) { 230 case TDS_INHIBITED: 231 case TDS_RUNNING: 232 case TDS_CAN_RUN: 233 case TDS_RUNQ: 234 /* 235 * We must never unlink a thread that is in one of 236 * these states, because it is currently active. 237 */ 238 panic("bad state for thread unlinking"); 239 /* NOTREACHED */ 240 case TDS_INACTIVE: 241 break; 242 default: 243 panic("bad thread state"); 244 /* NOTREACHED */ 245 } 246 #endif 247 #ifdef AUDIT 248 audit_thread_free(td); 249 #endif 250 /* Free all OSD associated to this thread. */ 251 osd_thread_exit(td); 252 td_softdep_cleanup(td); 253 MPASS(td->td_su == NULL); 254 255 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td); 256 tid_free(td->td_tid); 257 } 258 259 /* 260 * Initialize type-stable parts of a thread (when newly created). 261 */ 262 static int 263 thread_init(void *mem, int size, int flags) 264 { 265 struct thread *td; 266 267 td = (struct thread *)mem; 268 269 td->td_sleepqueue = sleepq_alloc(); 270 td->td_turnstile = turnstile_alloc(); 271 td->td_rlqe = NULL; 272 EVENTHANDLER_DIRECT_INVOKE(thread_init, td); 273 umtx_thread_init(td); 274 td->td_kstack = 0; 275 td->td_sel = NULL; 276 return (0); 277 } 278 279 /* 280 * Tear down type-stable parts of a thread (just before being discarded). 281 */ 282 static void 283 thread_fini(void *mem, int size) 284 { 285 struct thread *td; 286 287 td = (struct thread *)mem; 288 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td); 289 rlqentry_free(td->td_rlqe); 290 turnstile_free(td->td_turnstile); 291 sleepq_free(td->td_sleepqueue); 292 umtx_thread_fini(td); 293 seltdfini(td); 294 } 295 296 /* 297 * For a newly created process, 298 * link up all the structures and its initial threads etc. 299 * called from: 300 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc. 301 * proc_dtor() (should go away) 302 * proc_init() 303 */ 304 void 305 proc_linkup0(struct proc *p, struct thread *td) 306 { 307 TAILQ_INIT(&p->p_threads); /* all threads in proc */ 308 proc_linkup(p, td); 309 } 310 311 void 312 proc_linkup(struct proc *p, struct thread *td) 313 { 314 315 sigqueue_init(&p->p_sigqueue, p); 316 p->p_ksi = ksiginfo_alloc(1); 317 if (p->p_ksi != NULL) { 318 /* XXX p_ksi may be null if ksiginfo zone is not ready */ 319 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS; 320 } 321 LIST_INIT(&p->p_mqnotifier); 322 p->p_numthreads = 0; 323 thread_link(td, p); 324 } 325 326 /* 327 * Initialize global thread allocation resources. 328 */ 329 void 330 threadinit(void) 331 { 332 333 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF); 334 335 /* 336 * pid_max cannot be greater than PID_MAX. 337 * leave one number for thread0. 338 */ 339 tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock); 340 341 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(), 342 thread_ctor, thread_dtor, thread_init, thread_fini, 343 32 - 1, UMA_ZONE_NOFREE); 344 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash); 345 rw_init(&tidhash_lock, "tidhash"); 346 } 347 348 /* 349 * Place an unused thread on the zombie list. 350 * Use the slpq as that must be unused by now. 351 */ 352 void 353 thread_zombie(struct thread *td) 354 { 355 mtx_lock_spin(&zombie_lock); 356 TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq); 357 mtx_unlock_spin(&zombie_lock); 358 } 359 360 /* 361 * Release a thread that has exited after cpu_throw(). 362 */ 363 void 364 thread_stash(struct thread *td) 365 { 366 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1); 367 thread_zombie(td); 368 } 369 370 /* 371 * Reap zombie resources. 372 */ 373 void 374 thread_reap(void) 375 { 376 struct thread *td_first, *td_next; 377 378 /* 379 * Don't even bother to lock if none at this instant, 380 * we really don't care about the next instant. 381 */ 382 if (!TAILQ_EMPTY(&zombie_threads)) { 383 mtx_lock_spin(&zombie_lock); 384 td_first = TAILQ_FIRST(&zombie_threads); 385 if (td_first) 386 TAILQ_INIT(&zombie_threads); 387 mtx_unlock_spin(&zombie_lock); 388 while (td_first) { 389 td_next = TAILQ_NEXT(td_first, td_slpq); 390 thread_cow_free(td_first); 391 thread_free(td_first); 392 td_first = td_next; 393 } 394 } 395 } 396 397 /* 398 * Allocate a thread. 399 */ 400 struct thread * 401 thread_alloc(int pages) 402 { 403 struct thread *td; 404 405 thread_reap(); /* check if any zombies to get */ 406 407 td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK); 408 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack")); 409 if (!vm_thread_new(td, pages)) { 410 uma_zfree(thread_zone, td); 411 return (NULL); 412 } 413 cpu_thread_alloc(td); 414 vm_domain_policy_init(&td->td_vm_dom_policy); 415 return (td); 416 } 417 418 int 419 thread_alloc_stack(struct thread *td, int pages) 420 { 421 422 KASSERT(td->td_kstack == 0, 423 ("thread_alloc_stack called on a thread with kstack")); 424 if (!vm_thread_new(td, pages)) 425 return (0); 426 cpu_thread_alloc(td); 427 return (1); 428 } 429 430 /* 431 * Deallocate a thread. 432 */ 433 void 434 thread_free(struct thread *td) 435 { 436 437 lock_profile_thread_exit(td); 438 if (td->td_cpuset) 439 cpuset_rel(td->td_cpuset); 440 td->td_cpuset = NULL; 441 cpu_thread_free(td); 442 if (td->td_kstack != 0) 443 vm_thread_dispose(td); 444 vm_domain_policy_cleanup(&td->td_vm_dom_policy); 445 callout_drain(&td->td_slpcallout); 446 uma_zfree(thread_zone, td); 447 } 448 449 void 450 thread_cow_get_proc(struct thread *newtd, struct proc *p) 451 { 452 453 PROC_LOCK_ASSERT(p, MA_OWNED); 454 newtd->td_ucred = crhold(p->p_ucred); 455 newtd->td_limit = lim_hold(p->p_limit); 456 newtd->td_cowgen = p->p_cowgen; 457 } 458 459 void 460 thread_cow_get(struct thread *newtd, struct thread *td) 461 { 462 463 newtd->td_ucred = crhold(td->td_ucred); 464 newtd->td_limit = lim_hold(td->td_limit); 465 newtd->td_cowgen = td->td_cowgen; 466 } 467 468 void 469 thread_cow_free(struct thread *td) 470 { 471 472 if (td->td_ucred != NULL) 473 crfree(td->td_ucred); 474 if (td->td_limit != NULL) 475 lim_free(td->td_limit); 476 } 477 478 void 479 thread_cow_update(struct thread *td) 480 { 481 struct proc *p; 482 struct ucred *oldcred; 483 struct plimit *oldlimit; 484 485 p = td->td_proc; 486 oldcred = NULL; 487 oldlimit = NULL; 488 PROC_LOCK(p); 489 if (td->td_ucred != p->p_ucred) { 490 oldcred = td->td_ucred; 491 td->td_ucred = crhold(p->p_ucred); 492 } 493 if (td->td_limit != p->p_limit) { 494 oldlimit = td->td_limit; 495 td->td_limit = lim_hold(p->p_limit); 496 } 497 td->td_cowgen = p->p_cowgen; 498 PROC_UNLOCK(p); 499 if (oldcred != NULL) 500 crfree(oldcred); 501 if (oldlimit != NULL) 502 lim_free(oldlimit); 503 } 504 505 /* 506 * Discard the current thread and exit from its context. 507 * Always called with scheduler locked. 508 * 509 * Because we can't free a thread while we're operating under its context, 510 * push the current thread into our CPU's deadthread holder. This means 511 * we needn't worry about someone else grabbing our context before we 512 * do a cpu_throw(). 513 */ 514 void 515 thread_exit(void) 516 { 517 uint64_t runtime, new_switchtime; 518 struct thread *td; 519 struct thread *td2; 520 struct proc *p; 521 int wakeup_swapper; 522 523 td = curthread; 524 p = td->td_proc; 525 526 PROC_SLOCK_ASSERT(p, MA_OWNED); 527 mtx_assert(&Giant, MA_NOTOWNED); 528 529 PROC_LOCK_ASSERT(p, MA_OWNED); 530 KASSERT(p != NULL, ("thread exiting without a process")); 531 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td, 532 (long)p->p_pid, td->td_name); 533 SDT_PROBE0(proc, , , lwp__exit); 534 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending")); 535 536 #ifdef AUDIT 537 AUDIT_SYSCALL_EXIT(0, td); 538 #endif 539 /* 540 * drop FPU & debug register state storage, or any other 541 * architecture specific resources that 542 * would not be on a new untouched process. 543 */ 544 cpu_thread_exit(td); 545 546 /* 547 * The last thread is left attached to the process 548 * So that the whole bundle gets recycled. Skip 549 * all this stuff if we never had threads. 550 * EXIT clears all sign of other threads when 551 * it goes to single threading, so the last thread always 552 * takes the short path. 553 */ 554 if (p->p_flag & P_HADTHREADS) { 555 if (p->p_numthreads > 1) { 556 atomic_add_int(&td->td_proc->p_exitthreads, 1); 557 thread_unlink(td); 558 td2 = FIRST_THREAD_IN_PROC(p); 559 sched_exit_thread(td2, td); 560 561 /* 562 * The test below is NOT true if we are the 563 * sole exiting thread. P_STOPPED_SINGLE is unset 564 * in exit1() after it is the only survivor. 565 */ 566 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 567 if (p->p_numthreads == p->p_suspcount) { 568 thread_lock(p->p_singlethread); 569 wakeup_swapper = thread_unsuspend_one( 570 p->p_singlethread, p, false); 571 thread_unlock(p->p_singlethread); 572 if (wakeup_swapper) 573 kick_proc0(); 574 } 575 } 576 577 PCPU_SET(deadthread, td); 578 } else { 579 /* 580 * The last thread is exiting.. but not through exit() 581 */ 582 panic ("thread_exit: Last thread exiting on its own"); 583 } 584 } 585 #ifdef HWPMC_HOOKS 586 /* 587 * If this thread is part of a process that is being tracked by hwpmc(4), 588 * inform the module of the thread's impending exit. 589 */ 590 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 591 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 592 #endif 593 PROC_UNLOCK(p); 594 PROC_STATLOCK(p); 595 thread_lock(td); 596 PROC_SUNLOCK(p); 597 598 /* Do the same timestamp bookkeeping that mi_switch() would do. */ 599 new_switchtime = cpu_ticks(); 600 runtime = new_switchtime - PCPU_GET(switchtime); 601 td->td_runtime += runtime; 602 td->td_incruntime += runtime; 603 PCPU_SET(switchtime, new_switchtime); 604 PCPU_SET(switchticks, ticks); 605 VM_CNT_INC(v_swtch); 606 607 /* Save our resource usage in our process. */ 608 td->td_ru.ru_nvcsw++; 609 ruxagg(p, td); 610 rucollect(&p->p_ru, &td->td_ru); 611 PROC_STATUNLOCK(p); 612 613 td->td_state = TDS_INACTIVE; 614 #ifdef WITNESS 615 witness_thread_exit(td); 616 #endif 617 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td); 618 sched_throw(td); 619 panic("I'm a teapot!"); 620 /* NOTREACHED */ 621 } 622 623 /* 624 * Do any thread specific cleanups that may be needed in wait() 625 * called with Giant, proc and schedlock not held. 626 */ 627 void 628 thread_wait(struct proc *p) 629 { 630 struct thread *td; 631 632 mtx_assert(&Giant, MA_NOTOWNED); 633 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()")); 634 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking")); 635 td = FIRST_THREAD_IN_PROC(p); 636 /* Lock the last thread so we spin until it exits cpu_throw(). */ 637 thread_lock(td); 638 thread_unlock(td); 639 lock_profile_thread_exit(td); 640 cpuset_rel(td->td_cpuset); 641 td->td_cpuset = NULL; 642 cpu_thread_clean(td); 643 thread_cow_free(td); 644 callout_drain(&td->td_slpcallout); 645 thread_reap(); /* check for zombie threads etc. */ 646 } 647 648 /* 649 * Link a thread to a process. 650 * set up anything that needs to be initialized for it to 651 * be used by the process. 652 */ 653 void 654 thread_link(struct thread *td, struct proc *p) 655 { 656 657 /* 658 * XXX This can't be enabled because it's called for proc0 before 659 * its lock has been created. 660 * PROC_LOCK_ASSERT(p, MA_OWNED); 661 */ 662 td->td_state = TDS_INACTIVE; 663 td->td_proc = p; 664 td->td_flags = TDF_INMEM; 665 666 LIST_INIT(&td->td_contested); 667 LIST_INIT(&td->td_lprof[0]); 668 LIST_INIT(&td->td_lprof[1]); 669 sigqueue_init(&td->td_sigqueue, p); 670 callout_init(&td->td_slpcallout, 1); 671 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist); 672 p->p_numthreads++; 673 } 674 675 /* 676 * Called from: 677 * thread_exit() 678 */ 679 void 680 thread_unlink(struct thread *td) 681 { 682 struct proc *p = td->td_proc; 683 684 PROC_LOCK_ASSERT(p, MA_OWNED); 685 TAILQ_REMOVE(&p->p_threads, td, td_plist); 686 p->p_numthreads--; 687 /* could clear a few other things here */ 688 /* Must NOT clear links to proc! */ 689 } 690 691 static int 692 calc_remaining(struct proc *p, int mode) 693 { 694 int remaining; 695 696 PROC_LOCK_ASSERT(p, MA_OWNED); 697 PROC_SLOCK_ASSERT(p, MA_OWNED); 698 if (mode == SINGLE_EXIT) 699 remaining = p->p_numthreads; 700 else if (mode == SINGLE_BOUNDARY) 701 remaining = p->p_numthreads - p->p_boundary_count; 702 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC) 703 remaining = p->p_numthreads - p->p_suspcount; 704 else 705 panic("calc_remaining: wrong mode %d", mode); 706 return (remaining); 707 } 708 709 static int 710 remain_for_mode(int mode) 711 { 712 713 return (mode == SINGLE_ALLPROC ? 0 : 1); 714 } 715 716 static int 717 weed_inhib(int mode, struct thread *td2, struct proc *p) 718 { 719 int wakeup_swapper; 720 721 PROC_LOCK_ASSERT(p, MA_OWNED); 722 PROC_SLOCK_ASSERT(p, MA_OWNED); 723 THREAD_LOCK_ASSERT(td2, MA_OWNED); 724 725 wakeup_swapper = 0; 726 switch (mode) { 727 case SINGLE_EXIT: 728 if (TD_IS_SUSPENDED(td2)) 729 wakeup_swapper |= thread_unsuspend_one(td2, p, true); 730 if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0) 731 wakeup_swapper |= sleepq_abort(td2, EINTR); 732 break; 733 case SINGLE_BOUNDARY: 734 case SINGLE_NO_EXIT: 735 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_BOUNDARY) == 0) 736 wakeup_swapper |= thread_unsuspend_one(td2, p, false); 737 if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0) 738 wakeup_swapper |= sleepq_abort(td2, ERESTART); 739 break; 740 case SINGLE_ALLPROC: 741 /* 742 * ALLPROC suspend tries to avoid spurious EINTR for 743 * threads sleeping interruptable, by suspending the 744 * thread directly, similarly to sig_suspend_threads(). 745 * Since such sleep is not performed at the user 746 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP 747 * is used to avoid immediate un-suspend. 748 */ 749 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY | 750 TDF_ALLPROCSUSP)) == 0) 751 wakeup_swapper |= thread_unsuspend_one(td2, p, false); 752 if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0) { 753 if ((td2->td_flags & TDF_SBDRY) == 0) { 754 thread_suspend_one(td2); 755 td2->td_flags |= TDF_ALLPROCSUSP; 756 } else { 757 wakeup_swapper |= sleepq_abort(td2, ERESTART); 758 } 759 } 760 break; 761 } 762 return (wakeup_swapper); 763 } 764 765 /* 766 * Enforce single-threading. 767 * 768 * Returns 1 if the caller must abort (another thread is waiting to 769 * exit the process or similar). Process is locked! 770 * Returns 0 when you are successfully the only thread running. 771 * A process has successfully single threaded in the suspend mode when 772 * There are no threads in user mode. Threads in the kernel must be 773 * allowed to continue until they get to the user boundary. They may even 774 * copy out their return values and data before suspending. They may however be 775 * accelerated in reaching the user boundary as we will wake up 776 * any sleeping threads that are interruptable. (PCATCH). 777 */ 778 int 779 thread_single(struct proc *p, int mode) 780 { 781 struct thread *td; 782 struct thread *td2; 783 int remaining, wakeup_swapper; 784 785 td = curthread; 786 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 787 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 788 ("invalid mode %d", mode)); 789 /* 790 * If allowing non-ALLPROC singlethreading for non-curproc 791 * callers, calc_remaining() and remain_for_mode() should be 792 * adjusted to also account for td->td_proc != p. For now 793 * this is not implemented because it is not used. 794 */ 795 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) || 796 (mode != SINGLE_ALLPROC && td->td_proc == p), 797 ("mode %d proc %p curproc %p", mode, p, td->td_proc)); 798 mtx_assert(&Giant, MA_NOTOWNED); 799 PROC_LOCK_ASSERT(p, MA_OWNED); 800 801 if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC) 802 return (0); 803 804 /* Is someone already single threading? */ 805 if (p->p_singlethread != NULL && p->p_singlethread != td) 806 return (1); 807 808 if (mode == SINGLE_EXIT) { 809 p->p_flag |= P_SINGLE_EXIT; 810 p->p_flag &= ~P_SINGLE_BOUNDARY; 811 } else { 812 p->p_flag &= ~P_SINGLE_EXIT; 813 if (mode == SINGLE_BOUNDARY) 814 p->p_flag |= P_SINGLE_BOUNDARY; 815 else 816 p->p_flag &= ~P_SINGLE_BOUNDARY; 817 } 818 if (mode == SINGLE_ALLPROC) 819 p->p_flag |= P_TOTAL_STOP; 820 p->p_flag |= P_STOPPED_SINGLE; 821 PROC_SLOCK(p); 822 p->p_singlethread = td; 823 remaining = calc_remaining(p, mode); 824 while (remaining != remain_for_mode(mode)) { 825 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE) 826 goto stopme; 827 wakeup_swapper = 0; 828 FOREACH_THREAD_IN_PROC(p, td2) { 829 if (td2 == td) 830 continue; 831 thread_lock(td2); 832 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK; 833 if (TD_IS_INHIBITED(td2)) { 834 wakeup_swapper |= weed_inhib(mode, td2, p); 835 #ifdef SMP 836 } else if (TD_IS_RUNNING(td2) && td != td2) { 837 forward_signal(td2); 838 #endif 839 } 840 thread_unlock(td2); 841 } 842 if (wakeup_swapper) 843 kick_proc0(); 844 remaining = calc_remaining(p, mode); 845 846 /* 847 * Maybe we suspended some threads.. was it enough? 848 */ 849 if (remaining == remain_for_mode(mode)) 850 break; 851 852 stopme: 853 /* 854 * Wake us up when everyone else has suspended. 855 * In the mean time we suspend as well. 856 */ 857 thread_suspend_switch(td, p); 858 remaining = calc_remaining(p, mode); 859 } 860 if (mode == SINGLE_EXIT) { 861 /* 862 * Convert the process to an unthreaded process. The 863 * SINGLE_EXIT is called by exit1() or execve(), in 864 * both cases other threads must be retired. 865 */ 866 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads")); 867 p->p_singlethread = NULL; 868 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS); 869 870 /* 871 * Wait for any remaining threads to exit cpu_throw(). 872 */ 873 while (p->p_exitthreads != 0) { 874 PROC_SUNLOCK(p); 875 PROC_UNLOCK(p); 876 sched_relinquish(td); 877 PROC_LOCK(p); 878 PROC_SLOCK(p); 879 } 880 } else if (mode == SINGLE_BOUNDARY) { 881 /* 882 * Wait until all suspended threads are removed from 883 * the processors. The thread_suspend_check() 884 * increments p_boundary_count while it is still 885 * running, which makes it possible for the execve() 886 * to destroy vmspace while our other threads are 887 * still using the address space. 888 * 889 * We lock the thread, which is only allowed to 890 * succeed after context switch code finished using 891 * the address space. 892 */ 893 FOREACH_THREAD_IN_PROC(p, td2) { 894 if (td2 == td) 895 continue; 896 thread_lock(td2); 897 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0, 898 ("td %p not on boundary", td2)); 899 KASSERT(TD_IS_SUSPENDED(td2), 900 ("td %p is not suspended", td2)); 901 thread_unlock(td2); 902 } 903 } 904 PROC_SUNLOCK(p); 905 return (0); 906 } 907 908 bool 909 thread_suspend_check_needed(void) 910 { 911 struct proc *p; 912 struct thread *td; 913 914 td = curthread; 915 p = td->td_proc; 916 PROC_LOCK_ASSERT(p, MA_OWNED); 917 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 && 918 (td->td_dbgflags & TDB_SUSPEND) != 0)); 919 } 920 921 /* 922 * Called in from locations that can safely check to see 923 * whether we have to suspend or at least throttle for a 924 * single-thread event (e.g. fork). 925 * 926 * Such locations include userret(). 927 * If the "return_instead" argument is non zero, the thread must be able to 928 * accept 0 (caller may continue), or 1 (caller must abort) as a result. 929 * 930 * The 'return_instead' argument tells the function if it may do a 931 * thread_exit() or suspend, or whether the caller must abort and back 932 * out instead. 933 * 934 * If the thread that set the single_threading request has set the 935 * P_SINGLE_EXIT bit in the process flags then this call will never return 936 * if 'return_instead' is false, but will exit. 937 * 938 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0 939 *---------------+--------------------+--------------------- 940 * 0 | returns 0 | returns 0 or 1 941 * | when ST ends | immediately 942 *---------------+--------------------+--------------------- 943 * 1 | thread exits | returns 1 944 * | | immediately 945 * 0 = thread_exit() or suspension ok, 946 * other = return error instead of stopping the thread. 947 * 948 * While a full suspension is under effect, even a single threading 949 * thread would be suspended if it made this call (but it shouldn't). 950 * This call should only be made from places where 951 * thread_exit() would be safe as that may be the outcome unless 952 * return_instead is set. 953 */ 954 int 955 thread_suspend_check(int return_instead) 956 { 957 struct thread *td; 958 struct proc *p; 959 int wakeup_swapper; 960 961 td = curthread; 962 p = td->td_proc; 963 mtx_assert(&Giant, MA_NOTOWNED); 964 PROC_LOCK_ASSERT(p, MA_OWNED); 965 while (thread_suspend_check_needed()) { 966 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 967 KASSERT(p->p_singlethread != NULL, 968 ("singlethread not set")); 969 /* 970 * The only suspension in action is a 971 * single-threading. Single threader need not stop. 972 * It is safe to access p->p_singlethread unlocked 973 * because it can only be set to our address by us. 974 */ 975 if (p->p_singlethread == td) 976 return (0); /* Exempt from stopping. */ 977 } 978 if ((p->p_flag & P_SINGLE_EXIT) && return_instead) 979 return (EINTR); 980 981 /* Should we goto user boundary if we didn't come from there? */ 982 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 983 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead) 984 return (ERESTART); 985 986 /* 987 * Ignore suspend requests if they are deferred. 988 */ 989 if ((td->td_flags & TDF_SBDRY) != 0) { 990 KASSERT(return_instead, 991 ("TDF_SBDRY set for unsafe thread_suspend_check")); 992 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) != 993 (TDF_SEINTR | TDF_SERESTART), 994 ("both TDF_SEINTR and TDF_SERESTART")); 995 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0); 996 } 997 998 /* 999 * If the process is waiting for us to exit, 1000 * this thread should just suicide. 1001 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE. 1002 */ 1003 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { 1004 PROC_UNLOCK(p); 1005 1006 /* 1007 * Allow Linux emulation layer to do some work 1008 * before thread suicide. 1009 */ 1010 if (__predict_false(p->p_sysent->sv_thread_detach != NULL)) 1011 (p->p_sysent->sv_thread_detach)(td); 1012 umtx_thread_exit(td); 1013 kern_thr_exit(td); 1014 panic("stopped thread did not exit"); 1015 } 1016 1017 PROC_SLOCK(p); 1018 thread_stopped(p); 1019 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 1020 if (p->p_numthreads == p->p_suspcount + 1) { 1021 thread_lock(p->p_singlethread); 1022 wakeup_swapper = thread_unsuspend_one( 1023 p->p_singlethread, p, false); 1024 thread_unlock(p->p_singlethread); 1025 if (wakeup_swapper) 1026 kick_proc0(); 1027 } 1028 } 1029 PROC_UNLOCK(p); 1030 thread_lock(td); 1031 /* 1032 * When a thread suspends, it just 1033 * gets taken off all queues. 1034 */ 1035 thread_suspend_one(td); 1036 if (return_instead == 0) { 1037 p->p_boundary_count++; 1038 td->td_flags |= TDF_BOUNDARY; 1039 } 1040 PROC_SUNLOCK(p); 1041 mi_switch(SW_INVOL | SWT_SUSPEND, NULL); 1042 thread_unlock(td); 1043 PROC_LOCK(p); 1044 } 1045 return (0); 1046 } 1047 1048 void 1049 thread_suspend_switch(struct thread *td, struct proc *p) 1050 { 1051 1052 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1053 PROC_LOCK_ASSERT(p, MA_OWNED); 1054 PROC_SLOCK_ASSERT(p, MA_OWNED); 1055 /* 1056 * We implement thread_suspend_one in stages here to avoid 1057 * dropping the proc lock while the thread lock is owned. 1058 */ 1059 if (p == td->td_proc) { 1060 thread_stopped(p); 1061 p->p_suspcount++; 1062 } 1063 PROC_UNLOCK(p); 1064 thread_lock(td); 1065 td->td_flags &= ~TDF_NEEDSUSPCHK; 1066 TD_SET_SUSPENDED(td); 1067 sched_sleep(td, 0); 1068 PROC_SUNLOCK(p); 1069 DROP_GIANT(); 1070 mi_switch(SW_VOL | SWT_SUSPEND, NULL); 1071 thread_unlock(td); 1072 PICKUP_GIANT(); 1073 PROC_LOCK(p); 1074 PROC_SLOCK(p); 1075 } 1076 1077 void 1078 thread_suspend_one(struct thread *td) 1079 { 1080 struct proc *p; 1081 1082 p = td->td_proc; 1083 PROC_SLOCK_ASSERT(p, MA_OWNED); 1084 THREAD_LOCK_ASSERT(td, MA_OWNED); 1085 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); 1086 p->p_suspcount++; 1087 td->td_flags &= ~TDF_NEEDSUSPCHK; 1088 TD_SET_SUSPENDED(td); 1089 sched_sleep(td, 0); 1090 } 1091 1092 static int 1093 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary) 1094 { 1095 1096 THREAD_LOCK_ASSERT(td, MA_OWNED); 1097 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); 1098 TD_CLR_SUSPENDED(td); 1099 td->td_flags &= ~TDF_ALLPROCSUSP; 1100 if (td->td_proc == p) { 1101 PROC_SLOCK_ASSERT(p, MA_OWNED); 1102 p->p_suspcount--; 1103 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) { 1104 td->td_flags &= ~TDF_BOUNDARY; 1105 p->p_boundary_count--; 1106 } 1107 } 1108 return (setrunnable(td)); 1109 } 1110 1111 /* 1112 * Allow all threads blocked by single threading to continue running. 1113 */ 1114 void 1115 thread_unsuspend(struct proc *p) 1116 { 1117 struct thread *td; 1118 int wakeup_swapper; 1119 1120 PROC_LOCK_ASSERT(p, MA_OWNED); 1121 PROC_SLOCK_ASSERT(p, MA_OWNED); 1122 wakeup_swapper = 0; 1123 if (!P_SHOULDSTOP(p)) { 1124 FOREACH_THREAD_IN_PROC(p, td) { 1125 thread_lock(td); 1126 if (TD_IS_SUSPENDED(td)) { 1127 wakeup_swapper |= thread_unsuspend_one(td, p, 1128 true); 1129 } 1130 thread_unlock(td); 1131 } 1132 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && 1133 p->p_numthreads == p->p_suspcount) { 1134 /* 1135 * Stopping everything also did the job for the single 1136 * threading request. Now we've downgraded to single-threaded, 1137 * let it continue. 1138 */ 1139 if (p->p_singlethread->td_proc == p) { 1140 thread_lock(p->p_singlethread); 1141 wakeup_swapper = thread_unsuspend_one( 1142 p->p_singlethread, p, false); 1143 thread_unlock(p->p_singlethread); 1144 } 1145 } 1146 if (wakeup_swapper) 1147 kick_proc0(); 1148 } 1149 1150 /* 1151 * End the single threading mode.. 1152 */ 1153 void 1154 thread_single_end(struct proc *p, int mode) 1155 { 1156 struct thread *td; 1157 int wakeup_swapper; 1158 1159 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || 1160 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, 1161 ("invalid mode %d", mode)); 1162 PROC_LOCK_ASSERT(p, MA_OWNED); 1163 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) || 1164 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0), 1165 ("mode %d does not match P_TOTAL_STOP", mode)); 1166 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread, 1167 ("thread_single_end from other thread %p %p", 1168 curthread, p->p_singlethread)); 1169 KASSERT(mode != SINGLE_BOUNDARY || 1170 (p->p_flag & P_SINGLE_BOUNDARY) != 0, 1171 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag)); 1172 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY | 1173 P_TOTAL_STOP); 1174 PROC_SLOCK(p); 1175 p->p_singlethread = NULL; 1176 wakeup_swapper = 0; 1177 /* 1178 * If there are other threads they may now run, 1179 * unless of course there is a blanket 'stop order' 1180 * on the process. The single threader must be allowed 1181 * to continue however as this is a bad place to stop. 1182 */ 1183 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) { 1184 FOREACH_THREAD_IN_PROC(p, td) { 1185 thread_lock(td); 1186 if (TD_IS_SUSPENDED(td)) { 1187 wakeup_swapper |= thread_unsuspend_one(td, p, 1188 mode == SINGLE_BOUNDARY); 1189 } 1190 thread_unlock(td); 1191 } 1192 } 1193 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0, 1194 ("inconsistent boundary count %d", p->p_boundary_count)); 1195 PROC_SUNLOCK(p); 1196 if (wakeup_swapper) 1197 kick_proc0(); 1198 } 1199 1200 struct thread * 1201 thread_find(struct proc *p, lwpid_t tid) 1202 { 1203 struct thread *td; 1204 1205 PROC_LOCK_ASSERT(p, MA_OWNED); 1206 FOREACH_THREAD_IN_PROC(p, td) { 1207 if (td->td_tid == tid) 1208 break; 1209 } 1210 return (td); 1211 } 1212 1213 /* Locate a thread by number; return with proc lock held. */ 1214 struct thread * 1215 tdfind(lwpid_t tid, pid_t pid) 1216 { 1217 #define RUN_THRESH 16 1218 struct thread *td; 1219 int run = 0; 1220 1221 rw_rlock(&tidhash_lock); 1222 LIST_FOREACH(td, TIDHASH(tid), td_hash) { 1223 if (td->td_tid == tid) { 1224 if (pid != -1 && td->td_proc->p_pid != pid) { 1225 td = NULL; 1226 break; 1227 } 1228 PROC_LOCK(td->td_proc); 1229 if (td->td_proc->p_state == PRS_NEW) { 1230 PROC_UNLOCK(td->td_proc); 1231 td = NULL; 1232 break; 1233 } 1234 if (run > RUN_THRESH) { 1235 if (rw_try_upgrade(&tidhash_lock)) { 1236 LIST_REMOVE(td, td_hash); 1237 LIST_INSERT_HEAD(TIDHASH(td->td_tid), 1238 td, td_hash); 1239 rw_wunlock(&tidhash_lock); 1240 return (td); 1241 } 1242 } 1243 break; 1244 } 1245 run++; 1246 } 1247 rw_runlock(&tidhash_lock); 1248 return (td); 1249 } 1250 1251 void 1252 tidhash_add(struct thread *td) 1253 { 1254 rw_wlock(&tidhash_lock); 1255 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash); 1256 rw_wunlock(&tidhash_lock); 1257 } 1258 1259 void 1260 tidhash_remove(struct thread *td) 1261 { 1262 rw_wlock(&tidhash_lock); 1263 LIST_REMOVE(td, td_hash); 1264 rw_wunlock(&tidhash_lock); 1265 } 1266