1 /* $NetBSD: kern_lwp.c,v 1.252 2023/04/09 09:18:09 riastradh Exp $ */
2
3 /*-
4 * Copyright (c) 2001, 2006, 2007, 2008, 2009, 2019, 2020
5 * The NetBSD Foundation, Inc.
6 * All rights reserved.
7 *
8 * This code is derived from software contributed to The NetBSD Foundation
9 * by Nathan J. Williams, and Andrew Doran.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30 * POSSIBILITY OF SUCH DAMAGE.
31 */
32
33 /*
34 * Overview
35 *
36 * Lightweight processes (LWPs) are the basic unit or thread of
37 * execution within the kernel. The core state of an LWP is described
38 * by "struct lwp", also known as lwp_t.
39 *
40 * Each LWP is contained within a process (described by "struct proc"),
41 * Every process contains at least one LWP, but may contain more. The
42 * process describes attributes shared among all of its LWPs such as a
43 * private address space, global execution state (stopped, active,
44 * zombie, ...), signal disposition and so on. On a multiprocessor
45 * machine, multiple LWPs be executing concurrently in the kernel.
46 *
47 * Execution states
48 *
49 * At any given time, an LWP has overall state that is described by
50 * lwp::l_stat. The states are broken into two sets below. The first
51 * set is guaranteed to represent the absolute, current state of the
52 * LWP:
53 *
54 * LSONPROC
55 *
56 * On processor: the LWP is executing on a CPU, either in the
57 * kernel or in user space.
58 *
59 * LSRUN
60 *
61 * Runnable: the LWP is parked on a run queue, and may soon be
62 * chosen to run by an idle processor, or by a processor that
63 * has been asked to preempt a currently runnning but lower
64 * priority LWP.
65 *
66 * LSIDL
67 *
68 * Idle: the LWP has been created but has not yet executed, or
69 * it has ceased executing a unit of work and is waiting to be
70 * started again. This state exists so that the LWP can occupy
71 * a slot in the process & PID table, but without having to
72 * worry about being touched; lookups of the LWP by ID will
73 * fail while in this state. The LWP will become visible for
74 * lookup once its state transitions further. Some special
75 * kernel threads also (ab)use this state to indicate that they
76 * are idle (soft interrupts and idle LWPs).
77 *
78 * LSSUSPENDED:
79 *
80 * Suspended: the LWP has had its execution suspended by
81 * another LWP in the same process using the _lwp_suspend()
82 * system call. User-level LWPs also enter the suspended
83 * state when the system is shutting down.
84 *
85 * The second set represent a "statement of intent" on behalf of the
86 * LWP. The LWP may in fact be executing on a processor, may be
87 * sleeping or idle. It is expected to take the necessary action to
88 * stop executing or become "running" again within a short timeframe.
89 * The LP_RUNNING flag in lwp::l_pflag indicates that an LWP is running.
90 * Importantly, it indicates that its state is tied to a CPU.
91 *
92 * LSZOMB:
93 *
94 * Dead or dying: the LWP has released most of its resources
95 * and is about to switch away into oblivion, or has already
96 * switched away. When it switches away, its few remaining
97 * resources can be collected.
98 *
99 * LSSLEEP:
100 *
101 * Sleeping: the LWP has entered itself onto a sleep queue, and
102 * has switched away or will switch away shortly to allow other
103 * LWPs to run on the CPU.
104 *
105 * LSSTOP:
106 *
107 * Stopped: the LWP has been stopped as a result of a job
108 * control signal, or as a result of the ptrace() interface.
109 *
110 * Stopped LWPs may run briefly within the kernel to handle
111 * signals that they receive, but will not return to user space
112 * until their process' state is changed away from stopped.
113 *
114 * Single LWPs within a process can not be set stopped
115 * selectively: all actions that can stop or continue LWPs
116 * occur at the process level.
117 *
118 * State transitions
119 *
120 * Note that the LSSTOP state may only be set when returning to
121 * user space in userret(), or when sleeping interruptably. The
122 * LSSUSPENDED state may only be set in userret(). Before setting
123 * those states, we try to ensure that the LWPs will release all
124 * locks that they hold, and at a minimum try to ensure that the
125 * LWP can be set runnable again by a signal.
126 *
127 * LWPs may transition states in the following ways:
128 *
129 * RUN -------> ONPROC ONPROC -----> RUN
130 * > SLEEP
131 * > STOPPED
132 * > SUSPENDED
133 * > ZOMB
134 * > IDL (special cases)
135 *
136 * STOPPED ---> RUN SUSPENDED --> RUN
137 * > SLEEP
138 *
139 * SLEEP -----> ONPROC IDL --------> RUN
140 * > RUN > SUSPENDED
141 * > STOPPED > STOPPED
142 * > ONPROC (special cases)
143 *
144 * Some state transitions are only possible with kernel threads (eg
145 * ONPROC -> IDL) and happen under tightly controlled circumstances
146 * free of unwanted side effects.
147 *
148 * Migration
149 *
150 * Migration of threads from one CPU to another could be performed
151 * internally by the scheduler via sched_takecpu() or sched_catchlwp()
152 * functions. The universal lwp_migrate() function should be used for
153 * any other cases. Subsystems in the kernel must be aware that CPU
154 * of LWP may change, while it is not locked.
155 *
156 * Locking
157 *
158 * The majority of fields in 'struct lwp' are covered by a single,
159 * general spin lock pointed to by lwp::l_mutex. The locks covering
160 * each field are documented in sys/lwp.h.
161 *
162 * State transitions must be made with the LWP's general lock held,
163 * and may cause the LWP's lock pointer to change. Manipulation of
164 * the general lock is not performed directly, but through calls to
165 * lwp_lock(), lwp_unlock() and others. It should be noted that the
166 * adaptive locks are not allowed to be released while the LWP's lock
167 * is being held (unlike for other spin-locks).
168 *
169 * States and their associated locks:
170 *
171 * LSIDL, LSONPROC, LSZOMB, LSSUPENDED:
172 *
173 * Always covered by spc_lwplock, which protects LWPs not
174 * associated with any other sync object. This is a per-CPU
175 * lock and matches lwp::l_cpu.
176 *
177 * LSRUN:
178 *
179 * Always covered by spc_mutex, which protects the run queues.
180 * This is a per-CPU lock and matches lwp::l_cpu.
181 *
182 * LSSLEEP:
183 *
184 * Covered by a lock associated with the sleep queue (sometimes
185 * a turnstile sleep queue) that the LWP resides on. This can
186 * be spc_lwplock for SOBJ_SLEEPQ_NULL (an "untracked" sleep).
187 *
188 * LSSTOP:
189 *
190 * If the LWP was previously sleeping (l_wchan != NULL), then
191 * l_mutex references the sleep queue lock. If the LWP was
192 * runnable or on the CPU when halted, or has been removed from
193 * the sleep queue since halted, then the lock is spc_lwplock.
194 *
195 * The lock order is as follows:
196 *
197 * sleepq -> turnstile -> spc_lwplock -> spc_mutex
198 *
199 * Each process has a scheduler state lock (proc::p_lock), and a
200 * number of counters on LWPs and their states: p_nzlwps, p_nrlwps, and
201 * so on. When an LWP is to be entered into or removed from one of the
202 * following states, p_lock must be held and the process wide counters
203 * adjusted:
204 *
205 * LSIDL, LSZOMB, LSSTOP, LSSUSPENDED
206 *
207 * (But not always for kernel threads. There are some special cases
208 * as mentioned above: soft interrupts, and the idle loops.)
209 *
210 * Note that an LWP is considered running or likely to run soon if in
211 * one of the following states. This affects the value of p_nrlwps:
212 *
213 * LSRUN, LSONPROC, LSSLEEP
214 *
215 * p_lock does not need to be held when transitioning among these
216 * three states, hence p_lock is rarely taken for state transitions.
217 */
218
219 #include <sys/cdefs.h>
220 __KERNEL_RCSID(0, "$NetBSD: kern_lwp.c,v 1.252 2023/04/09 09:18:09 riastradh Exp $");
221
222 #include "opt_ddb.h"
223 #include "opt_lockdebug.h"
224 #include "opt_dtrace.h"
225
226 #define _LWP_API_PRIVATE
227
228 #include <sys/param.h>
229 #include <sys/systm.h>
230 #include <sys/cpu.h>
231 #include <sys/pool.h>
232 #include <sys/proc.h>
233 #include <sys/syscallargs.h>
234 #include <sys/syscall_stats.h>
235 #include <sys/kauth.h>
236 #include <sys/sleepq.h>
237 #include <sys/lockdebug.h>
238 #include <sys/kmem.h>
239 #include <sys/pset.h>
240 #include <sys/intr.h>
241 #include <sys/lwpctl.h>
242 #include <sys/atomic.h>
243 #include <sys/filedesc.h>
244 #include <sys/fstrans.h>
245 #include <sys/dtrace_bsd.h>
246 #include <sys/sdt.h>
247 #include <sys/ptrace.h>
248 #include <sys/xcall.h>
249 #include <sys/uidinfo.h>
250 #include <sys/sysctl.h>
251 #include <sys/psref.h>
252 #include <sys/msan.h>
253 #include <sys/kcov.h>
254 #include <sys/cprng.h>
255 #include <sys/futex.h>
256
257 #include <uvm/uvm_extern.h>
258 #include <uvm/uvm_object.h>
259
260 static pool_cache_t lwp_cache __read_mostly;
261 struct lwplist alllwp __cacheline_aligned;
262
263 static int lwp_ctor(void *, void *, int);
264 static void lwp_dtor(void *, void *);
265
266 /* DTrace proc provider probes */
267 SDT_PROVIDER_DEFINE(proc);
268
269 SDT_PROBE_DEFINE1(proc, kernel, , lwp__create, "struct lwp *");
270 SDT_PROBE_DEFINE1(proc, kernel, , lwp__start, "struct lwp *");
271 SDT_PROBE_DEFINE1(proc, kernel, , lwp__exit, "struct lwp *");
272
273 struct turnstile turnstile0 __cacheline_aligned;
274 struct lwp lwp0 __aligned(MIN_LWP_ALIGNMENT) = {
275 #ifdef LWP0_CPU_INFO
276 .l_cpu = LWP0_CPU_INFO,
277 #endif
278 #ifdef LWP0_MD_INITIALIZER
279 .l_md = LWP0_MD_INITIALIZER,
280 #endif
281 .l_proc = &proc0,
282 .l_lid = 0, /* we own proc0's slot in the pid table */
283 .l_flag = LW_SYSTEM,
284 .l_stat = LSONPROC,
285 .l_ts = &turnstile0,
286 .l_syncobj = &sched_syncobj,
287 .l_refcnt = 0,
288 .l_priority = PRI_USER + NPRI_USER - 1,
289 .l_inheritedprio = -1,
290 .l_class = SCHED_OTHER,
291 .l_psid = PS_NONE,
292 .l_pi_lenders = SLIST_HEAD_INITIALIZER(&lwp0.l_pi_lenders),
293 .l_name = __UNCONST("swapper"),
294 .l_fd = &filedesc0,
295 };
296
297 static int
lwp_maxlwp(void)298 lwp_maxlwp(void)
299 {
300 /* Assume 1 LWP per 1MiB. */
301 uint64_t lwps_per = ctob(physmem) / (1024 * 1024);
302
303 return MAX(MIN(MAXMAXLWP, lwps_per), MAXLWP);
304 }
305
306 static int sysctl_kern_maxlwp(SYSCTLFN_PROTO);
307
308 /*
309 * sysctl helper routine for kern.maxlwp. Ensures that the new
310 * values are not too low or too high.
311 */
312 static int
sysctl_kern_maxlwp(SYSCTLFN_ARGS)313 sysctl_kern_maxlwp(SYSCTLFN_ARGS)
314 {
315 int error, nmaxlwp;
316 struct sysctlnode node;
317
318 nmaxlwp = maxlwp;
319 node = *rnode;
320 node.sysctl_data = &nmaxlwp;
321 error = sysctl_lookup(SYSCTLFN_CALL(&node));
322 if (error || newp == NULL)
323 return error;
324
325 if (nmaxlwp < 0 || nmaxlwp >= MAXMAXLWP)
326 return EINVAL;
327 if (nmaxlwp > lwp_maxlwp())
328 return EINVAL;
329 maxlwp = nmaxlwp;
330
331 return 0;
332 }
333
334 static void
sysctl_kern_lwp_setup(void)335 sysctl_kern_lwp_setup(void)
336 {
337 sysctl_createv(NULL, 0, NULL, NULL,
338 CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
339 CTLTYPE_INT, "maxlwp",
340 SYSCTL_DESCR("Maximum number of simultaneous threads"),
341 sysctl_kern_maxlwp, 0, NULL, 0,
342 CTL_KERN, CTL_CREATE, CTL_EOL);
343 }
344
345 void
lwpinit(void)346 lwpinit(void)
347 {
348
349 LIST_INIT(&alllwp);
350 lwpinit_specificdata();
351 /*
352 * Provide a barrier to ensure that all mutex_oncpu() and rw_oncpu()
353 * calls will exit before memory of LWPs is returned to the pool, where
354 * KVA of LWP structure might be freed and re-used for other purposes.
355 * Kernel preemption is disabled around mutex_oncpu() and rw_oncpu()
356 * callers, therefore a regular passive serialization barrier will
357 * do the job.
358 */
359 lwp_cache = pool_cache_init(sizeof(lwp_t), MIN_LWP_ALIGNMENT, 0,
360 PR_PSERIALIZE, "lwppl", NULL, IPL_NONE, lwp_ctor, lwp_dtor, NULL);
361
362 maxlwp = lwp_maxlwp();
363 sysctl_kern_lwp_setup();
364 }
365
366 void
lwp0_init(void)367 lwp0_init(void)
368 {
369 struct lwp *l = &lwp0;
370
371 KASSERT((void *)uvm_lwp_getuarea(l) != NULL);
372
373 LIST_INSERT_HEAD(&alllwp, l, l_list);
374
375 callout_init(&l->l_timeout_ch, CALLOUT_MPSAFE);
376 callout_setfunc(&l->l_timeout_ch, sleepq_timeout, l);
377 cv_init(&l->l_sigcv, "sigwait");
378 cv_init(&l->l_waitcv, "vfork");
379
380 kauth_cred_hold(proc0.p_cred);
381 l->l_cred = proc0.p_cred;
382
383 kdtrace_thread_ctor(NULL, l);
384 lwp_initspecific(l);
385
386 SYSCALL_TIME_LWP_INIT(l);
387 }
388
389 /*
390 * Initialize the non-zeroed portion of an lwp_t.
391 */
392 static int
lwp_ctor(void * arg,void * obj,int flags)393 lwp_ctor(void *arg, void *obj, int flags)
394 {
395 lwp_t *l = obj;
396
397 l->l_stat = LSIDL;
398 l->l_cpu = curcpu();
399 l->l_mutex = l->l_cpu->ci_schedstate.spc_lwplock;
400 l->l_ts = pool_get(&turnstile_pool, flags);
401
402 if (l->l_ts == NULL) {
403 return ENOMEM;
404 } else {
405 turnstile_ctor(l->l_ts);
406 return 0;
407 }
408 }
409
410 static void
lwp_dtor(void * arg,void * obj)411 lwp_dtor(void *arg, void *obj)
412 {
413 lwp_t *l = obj;
414
415 /*
416 * The value of l->l_cpu must still be valid at this point.
417 */
418 KASSERT(l->l_cpu != NULL);
419
420 /*
421 * We can't return turnstile0 to the pool (it didn't come from it),
422 * so if it comes up just drop it quietly and move on.
423 */
424 if (l->l_ts != &turnstile0)
425 pool_put(&turnstile_pool, l->l_ts);
426 }
427
428 /*
429 * Set an LWP suspended.
430 *
431 * Must be called with p_lock held, and the LWP locked. Will unlock the
432 * LWP before return.
433 */
434 int
lwp_suspend(struct lwp * curl,struct lwp * t)435 lwp_suspend(struct lwp *curl, struct lwp *t)
436 {
437 int error;
438
439 KASSERT(mutex_owned(t->l_proc->p_lock));
440 KASSERT(lwp_locked(t, NULL));
441
442 KASSERT(curl != t || curl->l_stat == LSONPROC);
443
444 /*
445 * If the current LWP has been told to exit, we must not suspend anyone
446 * else or deadlock could occur. We won't return to userspace.
447 */
448 if ((curl->l_flag & (LW_WEXIT | LW_WCORE)) != 0) {
449 lwp_unlock(t);
450 return (EDEADLK);
451 }
452
453 if ((t->l_flag & LW_DBGSUSPEND) != 0) {
454 lwp_unlock(t);
455 return 0;
456 }
457
458 error = 0;
459
460 switch (t->l_stat) {
461 case LSRUN:
462 case LSONPROC:
463 t->l_flag |= LW_WSUSPEND;
464 lwp_need_userret(t);
465 lwp_unlock(t);
466 break;
467
468 case LSSLEEP:
469 t->l_flag |= LW_WSUSPEND;
470
471 /*
472 * Kick the LWP and try to get it to the kernel boundary
473 * so that it will release any locks that it holds.
474 * setrunnable() will release the lock.
475 */
476 if ((t->l_flag & LW_SINTR) != 0)
477 setrunnable(t);
478 else
479 lwp_unlock(t);
480 break;
481
482 case LSSUSPENDED:
483 lwp_unlock(t);
484 break;
485
486 case LSSTOP:
487 t->l_flag |= LW_WSUSPEND;
488 setrunnable(t);
489 break;
490
491 case LSIDL:
492 case LSZOMB:
493 error = EINTR; /* It's what Solaris does..... */
494 lwp_unlock(t);
495 break;
496 }
497
498 return (error);
499 }
500
501 /*
502 * Restart a suspended LWP.
503 *
504 * Must be called with p_lock held, and the LWP locked. Will unlock the
505 * LWP before return.
506 */
507 void
lwp_continue(struct lwp * l)508 lwp_continue(struct lwp *l)
509 {
510
511 KASSERT(mutex_owned(l->l_proc->p_lock));
512 KASSERT(lwp_locked(l, NULL));
513
514 /* If rebooting or not suspended, then just bail out. */
515 if ((l->l_flag & LW_WREBOOT) != 0) {
516 lwp_unlock(l);
517 return;
518 }
519
520 l->l_flag &= ~LW_WSUSPEND;
521
522 if (l->l_stat != LSSUSPENDED || (l->l_flag & LW_DBGSUSPEND) != 0) {
523 lwp_unlock(l);
524 return;
525 }
526
527 /* setrunnable() will release the lock. */
528 setrunnable(l);
529 }
530
531 /*
532 * Restart a stopped LWP.
533 *
534 * Must be called with p_lock held, and the LWP NOT locked. Will unlock the
535 * LWP before return.
536 */
537 void
lwp_unstop(struct lwp * l)538 lwp_unstop(struct lwp *l)
539 {
540 struct proc *p = l->l_proc;
541
542 KASSERT(mutex_owned(&proc_lock));
543 KASSERT(mutex_owned(p->p_lock));
544
545 lwp_lock(l);
546
547 KASSERT((l->l_flag & LW_DBGSUSPEND) == 0);
548
549 /* If not stopped, then just bail out. */
550 if (l->l_stat != LSSTOP) {
551 lwp_unlock(l);
552 return;
553 }
554
555 p->p_stat = SACTIVE;
556 p->p_sflag &= ~PS_STOPPING;
557
558 if (!p->p_waited)
559 p->p_pptr->p_nstopchild--;
560
561 if (l->l_wchan == NULL) {
562 /* setrunnable() will release the lock. */
563 setrunnable(l);
564 } else if (p->p_xsig && (l->l_flag & LW_SINTR) != 0) {
565 /* setrunnable() so we can receive the signal */
566 setrunnable(l);
567 } else {
568 l->l_stat = LSSLEEP;
569 p->p_nrlwps++;
570 lwp_unlock(l);
571 }
572 }
573
574 /*
575 * Wait for an LWP within the current process to exit. If 'lid' is
576 * non-zero, we are waiting for a specific LWP.
577 *
578 * Must be called with p->p_lock held.
579 */
580 int
lwp_wait(struct lwp * l,lwpid_t lid,lwpid_t * departed,bool exiting)581 lwp_wait(struct lwp *l, lwpid_t lid, lwpid_t *departed, bool exiting)
582 {
583 const lwpid_t curlid = l->l_lid;
584 proc_t *p = l->l_proc;
585 lwp_t *l2, *next;
586 int error;
587
588 KASSERT(mutex_owned(p->p_lock));
589
590 p->p_nlwpwait++;
591 l->l_waitingfor = lid;
592
593 for (;;) {
594 int nfound;
595
596 /*
597 * Avoid a race between exit1() and sigexit(): if the
598 * process is dumping core, then we need to bail out: call
599 * into lwp_userret() where we will be suspended until the
600 * deed is done.
601 */
602 if ((p->p_sflag & PS_WCORE) != 0) {
603 mutex_exit(p->p_lock);
604 lwp_userret(l);
605 KASSERT(false);
606 }
607
608 /*
609 * First off, drain any detached LWP that is waiting to be
610 * reaped.
611 */
612 while ((l2 = p->p_zomblwp) != NULL) {
613 p->p_zomblwp = NULL;
614 lwp_free(l2, false, false);/* releases proc mutex */
615 mutex_enter(p->p_lock);
616 }
617
618 /*
619 * Now look for an LWP to collect. If the whole process is
620 * exiting, count detached LWPs as eligible to be collected,
621 * but don't drain them here.
622 */
623 nfound = 0;
624 error = 0;
625
626 /*
627 * If given a specific LID, go via pid_table and make sure
628 * it's not detached.
629 */
630 if (lid != 0) {
631 l2 = proc_find_lwp(p, lid);
632 if (l2 == NULL) {
633 error = ESRCH;
634 break;
635 }
636 KASSERT(l2->l_lid == lid);
637 if ((l2->l_prflag & LPR_DETACHED) != 0) {
638 error = EINVAL;
639 break;
640 }
641 } else {
642 l2 = LIST_FIRST(&p->p_lwps);
643 }
644 for (; l2 != NULL; l2 = next) {
645 next = (lid != 0 ? NULL : LIST_NEXT(l2, l_sibling));
646
647 /*
648 * If a specific wait and the target is waiting on
649 * us, then avoid deadlock. This also traps LWPs
650 * that try to wait on themselves.
651 *
652 * Note that this does not handle more complicated
653 * cycles, like: t1 -> t2 -> t3 -> t1. The process
654 * can still be killed so it is not a major problem.
655 */
656 if (l2->l_lid == lid && l2->l_waitingfor == curlid) {
657 error = EDEADLK;
658 break;
659 }
660 if (l2 == l)
661 continue;
662 if ((l2->l_prflag & LPR_DETACHED) != 0) {
663 nfound += exiting;
664 continue;
665 }
666 if (lid != 0) {
667 /*
668 * Mark this LWP as the first waiter, if there
669 * is no other.
670 */
671 if (l2->l_waiter == 0)
672 l2->l_waiter = curlid;
673 } else if (l2->l_waiter != 0) {
674 /*
675 * It already has a waiter - so don't
676 * collect it. If the waiter doesn't
677 * grab it we'll get another chance
678 * later.
679 */
680 nfound++;
681 continue;
682 }
683 nfound++;
684
685 /* No need to lock the LWP in order to see LSZOMB. */
686 if (l2->l_stat != LSZOMB)
687 continue;
688
689 /*
690 * We're no longer waiting. Reset the "first waiter"
691 * pointer on the target, in case it was us.
692 */
693 l->l_waitingfor = 0;
694 l2->l_waiter = 0;
695 p->p_nlwpwait--;
696 if (departed)
697 *departed = l2->l_lid;
698 sched_lwp_collect(l2);
699
700 /* lwp_free() releases the proc lock. */
701 lwp_free(l2, false, false);
702 mutex_enter(p->p_lock);
703 return 0;
704 }
705
706 if (error != 0)
707 break;
708 if (nfound == 0) {
709 error = ESRCH;
710 break;
711 }
712
713 /*
714 * Note: since the lock will be dropped, need to restart on
715 * wakeup to run all LWPs again, e.g. there may be new LWPs.
716 */
717 if (exiting) {
718 KASSERT(p->p_nlwps > 1);
719 error = cv_timedwait(&p->p_lwpcv, p->p_lock, 1);
720 break;
721 }
722
723 /*
724 * Break out if all LWPs are in _lwp_wait(). There are
725 * other ways to hang the process with _lwp_wait(), but the
726 * sleep is interruptable so little point checking for them.
727 */
728 if (p->p_nlwpwait == p->p_nlwps) {
729 error = EDEADLK;
730 break;
731 }
732
733 /*
734 * Sit around and wait for something to happen. We'll be
735 * awoken if any of the conditions examined change: if an
736 * LWP exits, is collected, or is detached.
737 */
738 if ((error = cv_wait_sig(&p->p_lwpcv, p->p_lock)) != 0)
739 break;
740 }
741
742 /*
743 * We didn't find any LWPs to collect, we may have received a
744 * signal, or some other condition has caused us to bail out.
745 *
746 * If waiting on a specific LWP, clear the waiters marker: some
747 * other LWP may want it. Then, kick all the remaining waiters
748 * so that they can re-check for zombies and for deadlock.
749 */
750 if (lid != 0) {
751 l2 = proc_find_lwp(p, lid);
752 KASSERT(l2 == NULL || l2->l_lid == lid);
753
754 if (l2 != NULL && l2->l_waiter == curlid)
755 l2->l_waiter = 0;
756 }
757 p->p_nlwpwait--;
758 l->l_waitingfor = 0;
759 cv_broadcast(&p->p_lwpcv);
760
761 return error;
762 }
763
764 /*
765 * Create a new LWP within process 'p2', using LWP 'l1' as a template.
766 * The new LWP is created in state LSIDL and must be set running,
767 * suspended, or stopped by the caller.
768 */
769 int
lwp_create(lwp_t * l1,proc_t * p2,vaddr_t uaddr,int flags,void * stack,size_t stacksize,void (* func)(void *),void * arg,lwp_t ** rnewlwpp,int sclass,const sigset_t * sigmask,const stack_t * sigstk)770 lwp_create(lwp_t *l1, proc_t *p2, vaddr_t uaddr, int flags,
771 void *stack, size_t stacksize, void (*func)(void *), void *arg,
772 lwp_t **rnewlwpp, int sclass, const sigset_t *sigmask,
773 const stack_t *sigstk)
774 {
775 struct lwp *l2;
776
777 KASSERT(l1 == curlwp || l1->l_proc == &proc0);
778
779 /*
780 * Enforce limits, excluding the first lwp and kthreads. We must
781 * use the process credentials here when adjusting the limit, as
782 * they are what's tied to the accounting entity. However for
783 * authorizing the action, we'll use the LWP's credentials.
784 */
785 mutex_enter(p2->p_lock);
786 if (p2->p_nlwps != 0 && p2 != &proc0) {
787 uid_t uid = kauth_cred_getuid(p2->p_cred);
788 int count = chglwpcnt(uid, 1);
789 if (__predict_false(count >
790 p2->p_rlimit[RLIMIT_NTHR].rlim_cur)) {
791 if (kauth_authorize_process(l1->l_cred,
792 KAUTH_PROCESS_RLIMIT, p2,
793 KAUTH_ARG(KAUTH_REQ_PROCESS_RLIMIT_BYPASS),
794 &p2->p_rlimit[RLIMIT_NTHR], KAUTH_ARG(RLIMIT_NTHR))
795 != 0) {
796 (void)chglwpcnt(uid, -1);
797 mutex_exit(p2->p_lock);
798 return EAGAIN;
799 }
800 }
801 }
802
803 /*
804 * First off, reap any detached LWP waiting to be collected.
805 * We can re-use its LWP structure and turnstile.
806 */
807 if ((l2 = p2->p_zomblwp) != NULL) {
808 p2->p_zomblwp = NULL;
809 lwp_free(l2, true, false);
810 /* p2 now unlocked by lwp_free() */
811 KASSERT(l2->l_ts != NULL);
812 KASSERT(l2->l_inheritedprio == -1);
813 KASSERT(SLIST_EMPTY(&l2->l_pi_lenders));
814 memset(&l2->l_startzero, 0, sizeof(*l2) -
815 offsetof(lwp_t, l_startzero));
816 } else {
817 mutex_exit(p2->p_lock);
818 l2 = pool_cache_get(lwp_cache, PR_WAITOK);
819 memset(&l2->l_startzero, 0, sizeof(*l2) -
820 offsetof(lwp_t, l_startzero));
821 SLIST_INIT(&l2->l_pi_lenders);
822 }
823
824 /*
825 * Because of lockless lookup via pid_table, the LWP can be locked
826 * and inspected briefly even after it's freed, so a few fields are
827 * kept stable.
828 */
829 KASSERT(l2->l_stat == LSIDL);
830 KASSERT(l2->l_cpu != NULL);
831 KASSERT(l2->l_ts != NULL);
832 KASSERT(l2->l_mutex == l2->l_cpu->ci_schedstate.spc_lwplock);
833
834 l2->l_proc = p2;
835 l2->l_refcnt = 0;
836 l2->l_class = sclass;
837
838 /*
839 * Allocate a process ID for this LWP. We need to do this now
840 * while we can still unwind if it fails. Because we're marked
841 * as LSIDL, no lookups by the ID will succeed.
842 *
843 * N.B. this will always succeed for the first LWP in a process,
844 * because proc_alloc_lwpid() will usurp the slot. Also note
845 * that l2->l_proc MUST be valid so that lookups of the proc
846 * will succeed, even if the LWP itself is not visible.
847 */
848 if (__predict_false(proc_alloc_lwpid(p2, l2) == -1)) {
849 pool_cache_put(lwp_cache, l2);
850 return EAGAIN;
851 }
852
853 /*
854 * If vfork(), we want the LWP to run fast and on the same CPU
855 * as its parent, so that it can reuse the VM context and cache
856 * footprint on the local CPU.
857 */
858 l2->l_kpriority = ((flags & LWP_VFORK) ? true : false);
859 l2->l_kpribase = PRI_KERNEL;
860 l2->l_priority = l1->l_priority;
861 l2->l_inheritedprio = -1;
862 l2->l_protectprio = -1;
863 l2->l_auxprio = -1;
864 l2->l_flag = 0;
865 l2->l_pflag = LP_MPSAFE;
866 TAILQ_INIT(&l2->l_ld_locks);
867 l2->l_psrefs = 0;
868 kmsan_lwp_alloc(l2);
869
870 /*
871 * For vfork, borrow parent's lwpctl context if it exists.
872 * This also causes us to return via lwp_userret.
873 */
874 if (flags & LWP_VFORK && l1->l_lwpctl) {
875 l2->l_lwpctl = l1->l_lwpctl;
876 l2->l_flag |= LW_LWPCTL;
877 }
878
879 /*
880 * If not the first LWP in the process, grab a reference to the
881 * descriptor table.
882 */
883 l2->l_fd = p2->p_fd;
884 if (p2->p_nlwps != 0) {
885 KASSERT(l1->l_proc == p2);
886 fd_hold(l2);
887 } else {
888 KASSERT(l1->l_proc != p2);
889 }
890
891 if (p2->p_flag & PK_SYSTEM) {
892 /* Mark it as a system LWP. */
893 l2->l_flag |= LW_SYSTEM;
894 }
895
896 kdtrace_thread_ctor(NULL, l2);
897 lwp_initspecific(l2);
898 sched_lwp_fork(l1, l2);
899 lwp_update_creds(l2);
900 callout_init(&l2->l_timeout_ch, CALLOUT_MPSAFE);
901 callout_setfunc(&l2->l_timeout_ch, sleepq_timeout, l2);
902 cv_init(&l2->l_sigcv, "sigwait");
903 cv_init(&l2->l_waitcv, "vfork");
904 l2->l_syncobj = &sched_syncobj;
905 PSREF_DEBUG_INIT_LWP(l2);
906
907 if (rnewlwpp != NULL)
908 *rnewlwpp = l2;
909
910 /*
911 * PCU state needs to be saved before calling uvm_lwp_fork() so that
912 * the MD cpu_lwp_fork() can copy the saved state to the new LWP.
913 */
914 pcu_save_all(l1);
915 #if PCU_UNIT_COUNT > 0
916 l2->l_pcu_valid = l1->l_pcu_valid;
917 #endif
918
919 uvm_lwp_setuarea(l2, uaddr);
920 uvm_lwp_fork(l1, l2, stack, stacksize, func, (arg != NULL) ? arg : l2);
921
922 mutex_enter(p2->p_lock);
923 if ((flags & LWP_DETACHED) != 0) {
924 l2->l_prflag = LPR_DETACHED;
925 p2->p_ndlwps++;
926 } else
927 l2->l_prflag = 0;
928
929 if (l1->l_proc == p2) {
930 /*
931 * These flags are set while p_lock is held. Copy with
932 * p_lock held too, so the LWP doesn't sneak into the
933 * process without them being set.
934 */
935 l2->l_flag |= (l1->l_flag & (LW_WEXIT | LW_WREBOOT | LW_WCORE));
936 } else {
937 /* fork(): pending core/exit doesn't apply to child. */
938 l2->l_flag |= (l1->l_flag & LW_WREBOOT);
939 }
940
941 l2->l_sigstk = *sigstk;
942 l2->l_sigmask = *sigmask;
943 TAILQ_INIT(&l2->l_sigpend.sp_info);
944 sigemptyset(&l2->l_sigpend.sp_set);
945 LIST_INSERT_HEAD(&p2->p_lwps, l2, l_sibling);
946 p2->p_nlwps++;
947 p2->p_nrlwps++;
948
949 KASSERT(l2->l_affinity == NULL);
950
951 /* Inherit the affinity mask. */
952 if (l1->l_affinity) {
953 /*
954 * Note that we hold the state lock while inheriting
955 * the affinity to avoid race with sched_setaffinity().
956 */
957 lwp_lock(l1);
958 if (l1->l_affinity) {
959 kcpuset_use(l1->l_affinity);
960 l2->l_affinity = l1->l_affinity;
961 }
962 lwp_unlock(l1);
963 }
964
965 /* This marks the end of the "must be atomic" section. */
966 mutex_exit(p2->p_lock);
967
968 SDT_PROBE(proc, kernel, , lwp__create, l2, 0, 0, 0, 0);
969
970 mutex_enter(&proc_lock);
971 LIST_INSERT_HEAD(&alllwp, l2, l_list);
972 /* Inherit a processor-set */
973 l2->l_psid = l1->l_psid;
974 mutex_exit(&proc_lock);
975
976 SYSCALL_TIME_LWP_INIT(l2);
977
978 if (p2->p_emul->e_lwp_fork)
979 (*p2->p_emul->e_lwp_fork)(l1, l2);
980
981 return (0);
982 }
983
984 /*
985 * Set a new LWP running. If the process is stopping, then the LWP is
986 * created stopped.
987 */
988 void
lwp_start(lwp_t * l,int flags)989 lwp_start(lwp_t *l, int flags)
990 {
991 proc_t *p = l->l_proc;
992
993 mutex_enter(p->p_lock);
994 lwp_lock(l);
995 KASSERT(l->l_stat == LSIDL);
996 if ((flags & LWP_SUSPENDED) != 0) {
997 /* It'll suspend itself in lwp_userret(). */
998 l->l_flag |= LW_WSUSPEND;
999 }
1000 if (p->p_stat == SSTOP || (p->p_sflag & PS_STOPPING) != 0) {
1001 KASSERT(l->l_wchan == NULL);
1002 l->l_stat = LSSTOP;
1003 p->p_nrlwps--;
1004 lwp_unlock(l);
1005 } else {
1006 setrunnable(l);
1007 /* LWP now unlocked */
1008 }
1009 mutex_exit(p->p_lock);
1010 }
1011
1012 /*
1013 * Called by MD code when a new LWP begins execution. Must be called
1014 * with the previous LWP locked (so at splsched), or if there is no
1015 * previous LWP, at splsched.
1016 */
1017 void
lwp_startup(struct lwp * prev,struct lwp * new_lwp)1018 lwp_startup(struct lwp *prev, struct lwp *new_lwp)
1019 {
1020 kmutex_t *lock;
1021
1022 KASSERTMSG(new_lwp == curlwp, "l %p curlwp %p prevlwp %p", new_lwp, curlwp, prev);
1023 KASSERT(kpreempt_disabled());
1024 KASSERT(prev != NULL);
1025 KASSERT((prev->l_pflag & LP_RUNNING) != 0);
1026 KASSERT(curcpu()->ci_mtx_count == -2);
1027
1028 /*
1029 * Immediately mark the previous LWP as no longer running and
1030 * unlock (to keep lock wait times short as possible). If a
1031 * zombie, don't touch after clearing LP_RUNNING as it could be
1032 * reaped by another CPU. Use atomic_store_release to ensure
1033 * this -- matches atomic_load_acquire in lwp_free.
1034 */
1035 lock = prev->l_mutex;
1036 if (__predict_false(prev->l_stat == LSZOMB)) {
1037 atomic_store_release(&prev->l_pflag,
1038 prev->l_pflag & ~LP_RUNNING);
1039 } else {
1040 prev->l_pflag &= ~LP_RUNNING;
1041 }
1042 mutex_spin_exit(lock);
1043
1044 /* Correct spin mutex count after mi_switch(). */
1045 curcpu()->ci_mtx_count = 0;
1046
1047 /* Install new VM context. */
1048 if (__predict_true(new_lwp->l_proc->p_vmspace)) {
1049 pmap_activate(new_lwp);
1050 }
1051
1052 /* We remain at IPL_SCHED from mi_switch() - reset it. */
1053 spl0();
1054
1055 LOCKDEBUG_BARRIER(NULL, 0);
1056 SDT_PROBE(proc, kernel, , lwp__start, new_lwp, 0, 0, 0, 0);
1057
1058 /* For kthreads, acquire kernel lock if not MPSAFE. */
1059 if (__predict_false((new_lwp->l_pflag & LP_MPSAFE) == 0)) {
1060 KERNEL_LOCK(1, new_lwp);
1061 }
1062 }
1063
1064 /*
1065 * Exit an LWP.
1066 *
1067 * *** WARNING *** This can be called with (l != curlwp) in error paths.
1068 */
1069 void
lwp_exit(struct lwp * l)1070 lwp_exit(struct lwp *l)
1071 {
1072 struct proc *p = l->l_proc;
1073 struct lwp *l2;
1074 bool current;
1075
1076 current = (l == curlwp);
1077
1078 KASSERT(current || l->l_stat == LSIDL);
1079 KASSERT(current || l->l_target_cpu == NULL);
1080 KASSERT(p == curproc);
1081
1082 SDT_PROBE(proc, kernel, , lwp__exit, l, 0, 0, 0, 0);
1083
1084 /* Verify that we hold no locks; for DIAGNOSTIC check kernel_lock. */
1085 LOCKDEBUG_BARRIER(NULL, 0);
1086 KASSERTMSG(curcpu()->ci_biglock_count == 0, "kernel_lock leaked");
1087
1088 /*
1089 * If we are the last live LWP in a process, we need to exit the
1090 * entire process. We do so with an exit status of zero, because
1091 * it's a "controlled" exit, and because that's what Solaris does.
1092 *
1093 * We are not quite a zombie yet, but for accounting purposes we
1094 * must increment the count of zombies here.
1095 *
1096 * Note: the last LWP's specificdata will be deleted here.
1097 */
1098 mutex_enter(p->p_lock);
1099 if (p->p_nlwps - p->p_nzlwps == 1) {
1100 KASSERT(current == true);
1101 KASSERT(p != &proc0);
1102 exit1(l, 0, 0);
1103 /* NOTREACHED */
1104 }
1105 p->p_nzlwps++;
1106
1107 /*
1108 * Perform any required thread cleanup. Do this early so
1109 * anyone wanting to look us up with lwp_getref_lwpid() will
1110 * fail to find us before we become a zombie.
1111 *
1112 * N.B. this will unlock p->p_lock on our behalf.
1113 */
1114 lwp_thread_cleanup(l);
1115
1116 if (p->p_emul->e_lwp_exit)
1117 (*p->p_emul->e_lwp_exit)(l);
1118
1119 /* Drop filedesc reference. */
1120 fd_free();
1121
1122 /* Release fstrans private data. */
1123 fstrans_lwp_dtor(l);
1124
1125 /* Delete the specificdata while it's still safe to sleep. */
1126 lwp_finispecific(l);
1127
1128 /*
1129 * Release our cached credentials.
1130 */
1131 kauth_cred_free(l->l_cred);
1132 callout_destroy(&l->l_timeout_ch);
1133
1134 /*
1135 * If traced, report LWP exit event to the debugger.
1136 *
1137 * Remove the LWP from the global list.
1138 * Free its LID from the PID namespace if needed.
1139 */
1140 mutex_enter(&proc_lock);
1141
1142 if ((p->p_slflag & (PSL_TRACED|PSL_TRACELWP_EXIT)) ==
1143 (PSL_TRACED|PSL_TRACELWP_EXIT)) {
1144 mutex_enter(p->p_lock);
1145 if (ISSET(p->p_sflag, PS_WEXIT)) {
1146 mutex_exit(p->p_lock);
1147 /*
1148 * We are exiting, bail out without informing parent
1149 * about a terminating LWP as it would deadlock.
1150 */
1151 } else {
1152 eventswitch(TRAP_LWP, PTRACE_LWP_EXIT, l->l_lid);
1153 mutex_enter(&proc_lock);
1154 }
1155 }
1156
1157 LIST_REMOVE(l, l_list);
1158 mutex_exit(&proc_lock);
1159
1160 /*
1161 * Get rid of all references to the LWP that others (e.g. procfs)
1162 * may have, and mark the LWP as a zombie. If the LWP is detached,
1163 * mark it waiting for collection in the proc structure. Note that
1164 * before we can do that, we need to free any other dead, deatched
1165 * LWP waiting to meet its maker.
1166 *
1167 * All conditions need to be observed upon under the same hold of
1168 * p_lock, because if the lock is dropped any of them can change.
1169 */
1170 mutex_enter(p->p_lock);
1171 for (;;) {
1172 if (lwp_drainrefs(l))
1173 continue;
1174 if ((l->l_prflag & LPR_DETACHED) != 0) {
1175 if ((l2 = p->p_zomblwp) != NULL) {
1176 p->p_zomblwp = NULL;
1177 lwp_free(l2, false, false);
1178 /* proc now unlocked */
1179 mutex_enter(p->p_lock);
1180 continue;
1181 }
1182 p->p_zomblwp = l;
1183 }
1184 break;
1185 }
1186
1187 /*
1188 * If we find a pending signal for the process and we have been
1189 * asked to check for signals, then we lose: arrange to have
1190 * all other LWPs in the process check for signals.
1191 */
1192 if ((l->l_flag & LW_PENDSIG) != 0 &&
1193 firstsig(&p->p_sigpend.sp_set) != 0) {
1194 LIST_FOREACH(l2, &p->p_lwps, l_sibling) {
1195 lwp_lock(l2);
1196 signotify(l2);
1197 lwp_unlock(l2);
1198 }
1199 }
1200
1201 /*
1202 * Release any PCU resources before becoming a zombie.
1203 */
1204 pcu_discard_all(l);
1205
1206 lwp_lock(l);
1207 l->l_stat = LSZOMB;
1208 if (l->l_name != NULL) {
1209 strcpy(l->l_name, "(zombie)");
1210 }
1211 lwp_unlock(l);
1212 p->p_nrlwps--;
1213 cv_broadcast(&p->p_lwpcv);
1214 if (l->l_lwpctl != NULL)
1215 l->l_lwpctl->lc_curcpu = LWPCTL_CPU_EXITED;
1216 mutex_exit(p->p_lock);
1217
1218 /*
1219 * We can no longer block. At this point, lwp_free() may already
1220 * be gunning for us. On a multi-CPU system, we may be off p_lwps.
1221 *
1222 * Free MD LWP resources.
1223 */
1224 cpu_lwp_free(l, 0);
1225
1226 if (current) {
1227 /* Switch away into oblivion. */
1228 lwp_lock(l);
1229 spc_lock(l->l_cpu);
1230 mi_switch(l);
1231 panic("lwp_exit");
1232 }
1233 }
1234
1235 /*
1236 * Free a dead LWP's remaining resources.
1237 *
1238 * XXXLWP limits.
1239 */
1240 void
lwp_free(struct lwp * l,bool recycle,bool last)1241 lwp_free(struct lwp *l, bool recycle, bool last)
1242 {
1243 struct proc *p = l->l_proc;
1244 struct rusage *ru;
1245 ksiginfoq_t kq;
1246
1247 KASSERT(l != curlwp);
1248 KASSERT(last || mutex_owned(p->p_lock));
1249
1250 /*
1251 * We use the process credentials instead of the lwp credentials here
1252 * because the lwp credentials maybe cached (just after a setuid call)
1253 * and we don't want pay for syncing, since the lwp is going away
1254 * anyway
1255 */
1256 if (p != &proc0 && p->p_nlwps != 1)
1257 (void)chglwpcnt(kauth_cred_getuid(p->p_cred), -1);
1258
1259 /*
1260 * In the unlikely event that the LWP is still on the CPU,
1261 * then spin until it has switched away.
1262 *
1263 * atomic_load_acquire matches atomic_store_release in
1264 * lwp_startup and mi_switch.
1265 */
1266 while (__predict_false((atomic_load_acquire(&l->l_pflag) & LP_RUNNING)
1267 != 0)) {
1268 SPINLOCK_BACKOFF_HOOK;
1269 }
1270
1271 /*
1272 * Now that the LWP's known off the CPU, reset its state back to
1273 * LSIDL, which defeats anything that might have gotten a hold on
1274 * the LWP via pid_table before the ID was freed. It's important
1275 * to do this with both the LWP locked and p_lock held.
1276 *
1277 * Also reset the CPU and lock pointer back to curcpu(), since the
1278 * LWP will in all likelyhood be cached with the current CPU in
1279 * lwp_cache when we free it and later allocated from there again
1280 * (avoid incidental lock contention).
1281 */
1282 lwp_lock(l);
1283 l->l_stat = LSIDL;
1284 l->l_cpu = curcpu();
1285 lwp_unlock_to(l, l->l_cpu->ci_schedstate.spc_lwplock);
1286
1287 /*
1288 * If this was not the last LWP in the process, then adjust counters
1289 * and unlock. This is done differently for the last LWP in exit1().
1290 */
1291 if (!last) {
1292 /*
1293 * Add the LWP's run time to the process' base value.
1294 * This needs to co-incide with coming off p_lwps.
1295 */
1296 bintime_add(&p->p_rtime, &l->l_rtime);
1297 p->p_pctcpu += l->l_pctcpu;
1298 ru = &p->p_stats->p_ru;
1299 ruadd(ru, &l->l_ru);
1300 ru->ru_nvcsw += (l->l_ncsw - l->l_nivcsw);
1301 ru->ru_nivcsw += l->l_nivcsw;
1302 LIST_REMOVE(l, l_sibling);
1303 p->p_nlwps--;
1304 p->p_nzlwps--;
1305 if ((l->l_prflag & LPR_DETACHED) != 0)
1306 p->p_ndlwps--;
1307
1308 /*
1309 * Have any LWPs sleeping in lwp_wait() recheck for
1310 * deadlock.
1311 */
1312 cv_broadcast(&p->p_lwpcv);
1313 mutex_exit(p->p_lock);
1314
1315 /* Free the LWP ID. */
1316 mutex_enter(&proc_lock);
1317 proc_free_lwpid(p, l->l_lid);
1318 mutex_exit(&proc_lock);
1319 }
1320
1321 /*
1322 * Destroy the LWP's remaining signal information.
1323 */
1324 ksiginfo_queue_init(&kq);
1325 sigclear(&l->l_sigpend, NULL, &kq);
1326 ksiginfo_queue_drain(&kq);
1327 cv_destroy(&l->l_sigcv);
1328 cv_destroy(&l->l_waitcv);
1329
1330 /*
1331 * Free lwpctl structure and affinity.
1332 */
1333 if (l->l_lwpctl) {
1334 lwp_ctl_free(l);
1335 }
1336 if (l->l_affinity) {
1337 kcpuset_unuse(l->l_affinity, NULL);
1338 l->l_affinity = NULL;
1339 }
1340
1341 /*
1342 * Free remaining data structures and the LWP itself unless the
1343 * caller wants to recycle.
1344 */
1345 if (l->l_name != NULL)
1346 kmem_free(l->l_name, MAXCOMLEN);
1347
1348 kmsan_lwp_free(l);
1349 kcov_lwp_free(l);
1350 cpu_lwp_free2(l);
1351 uvm_lwp_exit(l);
1352
1353 KASSERT(SLIST_EMPTY(&l->l_pi_lenders));
1354 KASSERT(l->l_inheritedprio == -1);
1355 KASSERT(l->l_blcnt == 0);
1356 kdtrace_thread_dtor(NULL, l);
1357 if (!recycle)
1358 pool_cache_put(lwp_cache, l);
1359 }
1360
1361 /*
1362 * Migrate the LWP to the another CPU. Unlocks the LWP.
1363 */
1364 void
lwp_migrate(lwp_t * l,struct cpu_info * tci)1365 lwp_migrate(lwp_t *l, struct cpu_info *tci)
1366 {
1367 struct schedstate_percpu *tspc;
1368 int lstat = l->l_stat;
1369
1370 KASSERT(lwp_locked(l, NULL));
1371 KASSERT(tci != NULL);
1372
1373 /* If LWP is still on the CPU, it must be handled like LSONPROC */
1374 if ((l->l_pflag & LP_RUNNING) != 0) {
1375 lstat = LSONPROC;
1376 }
1377
1378 /*
1379 * The destination CPU could be changed while previous migration
1380 * was not finished.
1381 */
1382 if (l->l_target_cpu != NULL) {
1383 l->l_target_cpu = tci;
1384 lwp_unlock(l);
1385 return;
1386 }
1387
1388 /* Nothing to do if trying to migrate to the same CPU */
1389 if (l->l_cpu == tci) {
1390 lwp_unlock(l);
1391 return;
1392 }
1393
1394 KASSERT(l->l_target_cpu == NULL);
1395 tspc = &tci->ci_schedstate;
1396 switch (lstat) {
1397 case LSRUN:
1398 l->l_target_cpu = tci;
1399 break;
1400 case LSSLEEP:
1401 l->l_cpu = tci;
1402 break;
1403 case LSIDL:
1404 case LSSTOP:
1405 case LSSUSPENDED:
1406 l->l_cpu = tci;
1407 if (l->l_wchan == NULL) {
1408 lwp_unlock_to(l, tspc->spc_lwplock);
1409 return;
1410 }
1411 break;
1412 case LSONPROC:
1413 l->l_target_cpu = tci;
1414 spc_lock(l->l_cpu);
1415 sched_resched_cpu(l->l_cpu, PRI_USER_RT, true);
1416 /* spc now unlocked */
1417 break;
1418 }
1419 lwp_unlock(l);
1420 }
1421
1422 #define lwp_find_exclude(l) \
1423 ((l)->l_stat == LSIDL || (l)->l_stat == LSZOMB)
1424
1425 /*
1426 * Find the LWP in the process. Arguments may be zero, in such case,
1427 * the calling process and first LWP in the list will be used.
1428 * On success - returns proc locked.
1429 *
1430 * => pid == 0 -> look in curproc.
1431 * => pid == -1 -> match any proc.
1432 * => otherwise look up the proc.
1433 *
1434 * => lid == 0 -> first LWP in the proc
1435 * => otherwise specific LWP
1436 */
1437 struct lwp *
lwp_find2(pid_t pid,lwpid_t lid)1438 lwp_find2(pid_t pid, lwpid_t lid)
1439 {
1440 proc_t *p;
1441 lwp_t *l;
1442
1443 /* First LWP of specified proc. */
1444 if (lid == 0) {
1445 switch (pid) {
1446 case -1:
1447 /* No lookup keys. */
1448 return NULL;
1449 case 0:
1450 p = curproc;
1451 mutex_enter(p->p_lock);
1452 break;
1453 default:
1454 mutex_enter(&proc_lock);
1455 p = proc_find(pid);
1456 if (__predict_false(p == NULL)) {
1457 mutex_exit(&proc_lock);
1458 return NULL;
1459 }
1460 mutex_enter(p->p_lock);
1461 mutex_exit(&proc_lock);
1462 break;
1463 }
1464 LIST_FOREACH(l, &p->p_lwps, l_sibling) {
1465 if (__predict_true(!lwp_find_exclude(l)))
1466 break;
1467 }
1468 goto out;
1469 }
1470
1471 l = proc_find_lwp_acquire_proc(lid, &p);
1472 if (l == NULL)
1473 return NULL;
1474 KASSERT(p != NULL);
1475 KASSERT(mutex_owned(p->p_lock));
1476
1477 if (__predict_false(lwp_find_exclude(l))) {
1478 l = NULL;
1479 goto out;
1480 }
1481
1482 /* Apply proc filter, if applicable. */
1483 switch (pid) {
1484 case -1:
1485 /* Match anything. */
1486 break;
1487 case 0:
1488 if (p != curproc)
1489 l = NULL;
1490 break;
1491 default:
1492 if (p->p_pid != pid)
1493 l = NULL;
1494 break;
1495 }
1496
1497 out:
1498 if (__predict_false(l == NULL)) {
1499 mutex_exit(p->p_lock);
1500 }
1501 return l;
1502 }
1503
1504 /*
1505 * Look up a live LWP within the specified process.
1506 *
1507 * Must be called with p->p_lock held (as it looks at the radix tree,
1508 * and also wants to exclude idle and zombie LWPs).
1509 */
1510 struct lwp *
lwp_find(struct proc * p,lwpid_t id)1511 lwp_find(struct proc *p, lwpid_t id)
1512 {
1513 struct lwp *l;
1514
1515 KASSERT(mutex_owned(p->p_lock));
1516
1517 l = proc_find_lwp(p, id);
1518 KASSERT(l == NULL || l->l_lid == id);
1519
1520 /*
1521 * No need to lock - all of these conditions will
1522 * be visible with the process level mutex held.
1523 */
1524 if (__predict_false(l != NULL && lwp_find_exclude(l)))
1525 l = NULL;
1526
1527 return l;
1528 }
1529
1530 /*
1531 * Update an LWP's cached credentials to mirror the process' master copy.
1532 *
1533 * This happens early in the syscall path, on user trap, and on LWP
1534 * creation. A long-running LWP can also voluntarily choose to update
1535 * its credentials by calling this routine. This may be called from
1536 * LWP_CACHE_CREDS(), which checks l->l_prflag & LPR_CRMOD beforehand.
1537 */
1538 void
lwp_update_creds(struct lwp * l)1539 lwp_update_creds(struct lwp *l)
1540 {
1541 kauth_cred_t oc;
1542 struct proc *p;
1543
1544 p = l->l_proc;
1545 oc = l->l_cred;
1546
1547 mutex_enter(p->p_lock);
1548 kauth_cred_hold(p->p_cred);
1549 l->l_cred = p->p_cred;
1550 l->l_prflag &= ~LPR_CRMOD;
1551 mutex_exit(p->p_lock);
1552 if (oc != NULL)
1553 kauth_cred_free(oc);
1554 }
1555
1556 /*
1557 * Verify that an LWP is locked, and optionally verify that the lock matches
1558 * one we specify.
1559 */
1560 int
lwp_locked(struct lwp * l,kmutex_t * mtx)1561 lwp_locked(struct lwp *l, kmutex_t *mtx)
1562 {
1563 kmutex_t *cur = l->l_mutex;
1564
1565 return mutex_owned(cur) && (mtx == cur || mtx == NULL);
1566 }
1567
1568 /*
1569 * Lend a new mutex to an LWP. The old mutex must be held.
1570 */
1571 kmutex_t *
lwp_setlock(struct lwp * l,kmutex_t * mtx)1572 lwp_setlock(struct lwp *l, kmutex_t *mtx)
1573 {
1574 kmutex_t *oldmtx = l->l_mutex;
1575
1576 KASSERT(mutex_owned(oldmtx));
1577
1578 atomic_store_release(&l->l_mutex, mtx);
1579 return oldmtx;
1580 }
1581
1582 /*
1583 * Lend a new mutex to an LWP, and release the old mutex. The old mutex
1584 * must be held.
1585 */
1586 void
lwp_unlock_to(struct lwp * l,kmutex_t * mtx)1587 lwp_unlock_to(struct lwp *l, kmutex_t *mtx)
1588 {
1589 kmutex_t *old;
1590
1591 KASSERT(lwp_locked(l, NULL));
1592
1593 old = l->l_mutex;
1594 atomic_store_release(&l->l_mutex, mtx);
1595 mutex_spin_exit(old);
1596 }
1597
1598 int
lwp_trylock(struct lwp * l)1599 lwp_trylock(struct lwp *l)
1600 {
1601 kmutex_t *old;
1602
1603 for (;;) {
1604 if (!mutex_tryenter(old = atomic_load_consume(&l->l_mutex)))
1605 return 0;
1606 if (__predict_true(atomic_load_relaxed(&l->l_mutex) == old))
1607 return 1;
1608 mutex_spin_exit(old);
1609 }
1610 }
1611
1612 void
lwp_unsleep(lwp_t * l,bool unlock)1613 lwp_unsleep(lwp_t *l, bool unlock)
1614 {
1615
1616 KASSERT(mutex_owned(l->l_mutex));
1617 (*l->l_syncobj->sobj_unsleep)(l, unlock);
1618 }
1619
1620 /*
1621 * Handle exceptions for mi_userret(). Called if a member of LW_USERRET is
1622 * set.
1623 */
1624 void
lwp_userret(struct lwp * l)1625 lwp_userret(struct lwp *l)
1626 {
1627 struct proc *p;
1628 int sig;
1629
1630 KASSERT(l == curlwp);
1631 KASSERT(l->l_stat == LSONPROC);
1632 p = l->l_proc;
1633
1634 /*
1635 * It is safe to do this read unlocked on a MP system..
1636 */
1637 while ((l->l_flag & LW_USERRET) != 0) {
1638 /*
1639 * Process pending signals first, unless the process
1640 * is dumping core or exiting, where we will instead
1641 * enter the LW_WSUSPEND case below.
1642 */
1643 if ((l->l_flag & (LW_PENDSIG | LW_WCORE | LW_WEXIT)) ==
1644 LW_PENDSIG) {
1645 mutex_enter(p->p_lock);
1646 while ((sig = issignal(l)) != 0)
1647 postsig(sig);
1648 mutex_exit(p->p_lock);
1649 }
1650
1651 /*
1652 * Core-dump or suspend pending.
1653 *
1654 * In case of core dump, suspend ourselves, so that the kernel
1655 * stack and therefore the userland registers saved in the
1656 * trapframe are around for coredump() to write them out.
1657 * We also need to save any PCU resources that we have so that
1658 * they accessible for coredump(). We issue a wakeup on
1659 * p->p_lwpcv so that sigexit() will write the core file out
1660 * once all other LWPs are suspended.
1661 */
1662 if ((l->l_flag & LW_WSUSPEND) != 0) {
1663 pcu_save_all(l);
1664 mutex_enter(p->p_lock);
1665 p->p_nrlwps--;
1666 cv_broadcast(&p->p_lwpcv);
1667 lwp_lock(l);
1668 l->l_stat = LSSUSPENDED;
1669 lwp_unlock(l);
1670 mutex_exit(p->p_lock);
1671 lwp_lock(l);
1672 spc_lock(l->l_cpu);
1673 mi_switch(l);
1674 }
1675
1676 /* Process is exiting. */
1677 if ((l->l_flag & LW_WEXIT) != 0) {
1678 lwp_exit(l);
1679 KASSERT(0);
1680 /* NOTREACHED */
1681 }
1682
1683 /* update lwpctl processor (for vfork child_return) */
1684 if (l->l_flag & LW_LWPCTL) {
1685 lwp_lock(l);
1686 KASSERT(kpreempt_disabled());
1687 l->l_lwpctl->lc_curcpu = (int)cpu_index(l->l_cpu);
1688 l->l_lwpctl->lc_pctr++;
1689 l->l_flag &= ~LW_LWPCTL;
1690 lwp_unlock(l);
1691 }
1692 }
1693 }
1694
1695 /*
1696 * Force an LWP to enter the kernel, to take a trip through lwp_userret().
1697 */
1698 void
lwp_need_userret(struct lwp * l)1699 lwp_need_userret(struct lwp *l)
1700 {
1701
1702 KASSERT(!cpu_intr_p());
1703 KASSERT(lwp_locked(l, NULL));
1704
1705 /*
1706 * If the LWP is in any state other than LSONPROC, we know that it
1707 * is executing in-kernel and will hit userret() on the way out.
1708 *
1709 * If the LWP is curlwp, then we know we'll be back out to userspace
1710 * soon (can't be called from a hardware interrupt here).
1711 *
1712 * Otherwise, we can't be sure what the LWP is doing, so first make
1713 * sure the update to l_flag will be globally visible, and then
1714 * force the LWP to take a trip through trap() where it will do
1715 * userret().
1716 */
1717 if (l->l_stat == LSONPROC && l != curlwp) {
1718 membar_producer();
1719 cpu_signotify(l);
1720 }
1721 }
1722
1723 /*
1724 * Add one reference to an LWP. This will prevent the LWP from
1725 * exiting, thus keep the lwp structure and PCB around to inspect.
1726 */
1727 void
lwp_addref(struct lwp * l)1728 lwp_addref(struct lwp *l)
1729 {
1730 KASSERT(mutex_owned(l->l_proc->p_lock));
1731 KASSERT(l->l_stat != LSZOMB);
1732 l->l_refcnt++;
1733 }
1734
1735 /*
1736 * Remove one reference to an LWP. If this is the last reference,
1737 * then we must finalize the LWP's death.
1738 */
1739 void
lwp_delref(struct lwp * l)1740 lwp_delref(struct lwp *l)
1741 {
1742 struct proc *p = l->l_proc;
1743
1744 mutex_enter(p->p_lock);
1745 lwp_delref2(l);
1746 mutex_exit(p->p_lock);
1747 }
1748
1749 /*
1750 * Remove one reference to an LWP. If this is the last reference,
1751 * then we must finalize the LWP's death. The proc mutex is held
1752 * on entry.
1753 */
1754 void
lwp_delref2(struct lwp * l)1755 lwp_delref2(struct lwp *l)
1756 {
1757 struct proc *p = l->l_proc;
1758
1759 KASSERT(mutex_owned(p->p_lock));
1760 KASSERT(l->l_stat != LSZOMB);
1761 KASSERT(l->l_refcnt > 0);
1762
1763 if (--l->l_refcnt == 0)
1764 cv_broadcast(&p->p_lwpcv);
1765 }
1766
1767 /*
1768 * Drain all references to the current LWP. Returns true if
1769 * we blocked.
1770 */
1771 bool
lwp_drainrefs(struct lwp * l)1772 lwp_drainrefs(struct lwp *l)
1773 {
1774 struct proc *p = l->l_proc;
1775 bool rv = false;
1776
1777 KASSERT(mutex_owned(p->p_lock));
1778
1779 l->l_prflag |= LPR_DRAINING;
1780
1781 while (l->l_refcnt > 0) {
1782 rv = true;
1783 cv_wait(&p->p_lwpcv, p->p_lock);
1784 }
1785 return rv;
1786 }
1787
1788 /*
1789 * Return true if the specified LWP is 'alive'. Only p->p_lock need
1790 * be held.
1791 */
1792 bool
lwp_alive(lwp_t * l)1793 lwp_alive(lwp_t *l)
1794 {
1795
1796 KASSERT(mutex_owned(l->l_proc->p_lock));
1797
1798 switch (l->l_stat) {
1799 case LSSLEEP:
1800 case LSRUN:
1801 case LSONPROC:
1802 case LSSTOP:
1803 case LSSUSPENDED:
1804 return true;
1805 default:
1806 return false;
1807 }
1808 }
1809
1810 /*
1811 * Return first live LWP in the process.
1812 */
1813 lwp_t *
lwp_find_first(proc_t * p)1814 lwp_find_first(proc_t *p)
1815 {
1816 lwp_t *l;
1817
1818 KASSERT(mutex_owned(p->p_lock));
1819
1820 LIST_FOREACH(l, &p->p_lwps, l_sibling) {
1821 if (lwp_alive(l)) {
1822 return l;
1823 }
1824 }
1825
1826 return NULL;
1827 }
1828
1829 /*
1830 * Allocate a new lwpctl structure for a user LWP.
1831 */
1832 int
lwp_ctl_alloc(vaddr_t * uaddr)1833 lwp_ctl_alloc(vaddr_t *uaddr)
1834 {
1835 lcproc_t *lp;
1836 u_int bit, i, offset;
1837 struct uvm_object *uao;
1838 int error;
1839 lcpage_t *lcp;
1840 proc_t *p;
1841 lwp_t *l;
1842
1843 l = curlwp;
1844 p = l->l_proc;
1845
1846 /* don't allow a vforked process to create lwp ctls */
1847 if (p->p_lflag & PL_PPWAIT)
1848 return EBUSY;
1849
1850 if (l->l_lcpage != NULL) {
1851 lcp = l->l_lcpage;
1852 *uaddr = lcp->lcp_uaddr + (vaddr_t)l->l_lwpctl - lcp->lcp_kaddr;
1853 return 0;
1854 }
1855
1856 /* First time around, allocate header structure for the process. */
1857 if ((lp = p->p_lwpctl) == NULL) {
1858 lp = kmem_alloc(sizeof(*lp), KM_SLEEP);
1859 mutex_init(&lp->lp_lock, MUTEX_DEFAULT, IPL_NONE);
1860 lp->lp_uao = NULL;
1861 TAILQ_INIT(&lp->lp_pages);
1862 mutex_enter(p->p_lock);
1863 if (p->p_lwpctl == NULL) {
1864 p->p_lwpctl = lp;
1865 mutex_exit(p->p_lock);
1866 } else {
1867 mutex_exit(p->p_lock);
1868 mutex_destroy(&lp->lp_lock);
1869 kmem_free(lp, sizeof(*lp));
1870 lp = p->p_lwpctl;
1871 }
1872 }
1873
1874 /*
1875 * Set up an anonymous memory region to hold the shared pages.
1876 * Map them into the process' address space. The user vmspace
1877 * gets the first reference on the UAO.
1878 */
1879 mutex_enter(&lp->lp_lock);
1880 if (lp->lp_uao == NULL) {
1881 lp->lp_uao = uao_create(LWPCTL_UAREA_SZ, 0);
1882 lp->lp_cur = 0;
1883 lp->lp_max = LWPCTL_UAREA_SZ;
1884 lp->lp_uva = p->p_emul->e_vm_default_addr(p,
1885 (vaddr_t)p->p_vmspace->vm_daddr, LWPCTL_UAREA_SZ,
1886 p->p_vmspace->vm_map.flags & VM_MAP_TOPDOWN);
1887 error = uvm_map(&p->p_vmspace->vm_map, &lp->lp_uva,
1888 LWPCTL_UAREA_SZ, lp->lp_uao, 0, 0, UVM_MAPFLAG(UVM_PROT_RW,
1889 UVM_PROT_RW, UVM_INH_NONE, UVM_ADV_NORMAL, 0));
1890 if (error != 0) {
1891 uao_detach(lp->lp_uao);
1892 lp->lp_uao = NULL;
1893 mutex_exit(&lp->lp_lock);
1894 return error;
1895 }
1896 }
1897
1898 /* Get a free block and allocate for this LWP. */
1899 TAILQ_FOREACH(lcp, &lp->lp_pages, lcp_chain) {
1900 if (lcp->lcp_nfree != 0)
1901 break;
1902 }
1903 if (lcp == NULL) {
1904 /* Nothing available - try to set up a free page. */
1905 if (lp->lp_cur == lp->lp_max) {
1906 mutex_exit(&lp->lp_lock);
1907 return ENOMEM;
1908 }
1909 lcp = kmem_alloc(LWPCTL_LCPAGE_SZ, KM_SLEEP);
1910
1911 /*
1912 * Wire the next page down in kernel space. Since this
1913 * is a new mapping, we must add a reference.
1914 */
1915 uao = lp->lp_uao;
1916 (*uao->pgops->pgo_reference)(uao);
1917 lcp->lcp_kaddr = vm_map_min(kernel_map);
1918 error = uvm_map(kernel_map, &lcp->lcp_kaddr, PAGE_SIZE,
1919 uao, lp->lp_cur, PAGE_SIZE,
1920 UVM_MAPFLAG(UVM_PROT_RW, UVM_PROT_RW,
1921 UVM_INH_NONE, UVM_ADV_RANDOM, 0));
1922 if (error != 0) {
1923 mutex_exit(&lp->lp_lock);
1924 kmem_free(lcp, LWPCTL_LCPAGE_SZ);
1925 (*uao->pgops->pgo_detach)(uao);
1926 return error;
1927 }
1928 error = uvm_map_pageable(kernel_map, lcp->lcp_kaddr,
1929 lcp->lcp_kaddr + PAGE_SIZE, FALSE, 0);
1930 if (error != 0) {
1931 mutex_exit(&lp->lp_lock);
1932 uvm_unmap(kernel_map, lcp->lcp_kaddr,
1933 lcp->lcp_kaddr + PAGE_SIZE);
1934 kmem_free(lcp, LWPCTL_LCPAGE_SZ);
1935 return error;
1936 }
1937 /* Prepare the page descriptor and link into the list. */
1938 lcp->lcp_uaddr = lp->lp_uva + lp->lp_cur;
1939 lp->lp_cur += PAGE_SIZE;
1940 lcp->lcp_nfree = LWPCTL_PER_PAGE;
1941 lcp->lcp_rotor = 0;
1942 memset(lcp->lcp_bitmap, 0xff, LWPCTL_BITMAP_SZ);
1943 TAILQ_INSERT_HEAD(&lp->lp_pages, lcp, lcp_chain);
1944 }
1945 for (i = lcp->lcp_rotor; lcp->lcp_bitmap[i] == 0;) {
1946 if (++i >= LWPCTL_BITMAP_ENTRIES)
1947 i = 0;
1948 }
1949 bit = ffs(lcp->lcp_bitmap[i]) - 1;
1950 lcp->lcp_bitmap[i] ^= (1U << bit);
1951 lcp->lcp_rotor = i;
1952 lcp->lcp_nfree--;
1953 l->l_lcpage = lcp;
1954 offset = (i << 5) + bit;
1955 l->l_lwpctl = (lwpctl_t *)lcp->lcp_kaddr + offset;
1956 *uaddr = lcp->lcp_uaddr + offset * sizeof(lwpctl_t);
1957 mutex_exit(&lp->lp_lock);
1958
1959 KPREEMPT_DISABLE(l);
1960 l->l_lwpctl->lc_curcpu = (int)cpu_index(curcpu());
1961 KPREEMPT_ENABLE(l);
1962
1963 return 0;
1964 }
1965
1966 /*
1967 * Free an lwpctl structure back to the per-process list.
1968 */
1969 void
lwp_ctl_free(lwp_t * l)1970 lwp_ctl_free(lwp_t *l)
1971 {
1972 struct proc *p = l->l_proc;
1973 lcproc_t *lp;
1974 lcpage_t *lcp;
1975 u_int map, offset;
1976
1977 /* don't free a lwp context we borrowed for vfork */
1978 if (p->p_lflag & PL_PPWAIT) {
1979 l->l_lwpctl = NULL;
1980 return;
1981 }
1982
1983 lp = p->p_lwpctl;
1984 KASSERT(lp != NULL);
1985
1986 lcp = l->l_lcpage;
1987 offset = (u_int)((lwpctl_t *)l->l_lwpctl - (lwpctl_t *)lcp->lcp_kaddr);
1988 KASSERT(offset < LWPCTL_PER_PAGE);
1989
1990 mutex_enter(&lp->lp_lock);
1991 lcp->lcp_nfree++;
1992 map = offset >> 5;
1993 lcp->lcp_bitmap[map] |= (1U << (offset & 31));
1994 if (lcp->lcp_bitmap[lcp->lcp_rotor] == 0)
1995 lcp->lcp_rotor = map;
1996 if (TAILQ_FIRST(&lp->lp_pages)->lcp_nfree == 0) {
1997 TAILQ_REMOVE(&lp->lp_pages, lcp, lcp_chain);
1998 TAILQ_INSERT_HEAD(&lp->lp_pages, lcp, lcp_chain);
1999 }
2000 mutex_exit(&lp->lp_lock);
2001 }
2002
2003 /*
2004 * Process is exiting; tear down lwpctl state. This can only be safely
2005 * called by the last LWP in the process.
2006 */
2007 void
lwp_ctl_exit(void)2008 lwp_ctl_exit(void)
2009 {
2010 lcpage_t *lcp, *next;
2011 lcproc_t *lp;
2012 proc_t *p;
2013 lwp_t *l;
2014
2015 l = curlwp;
2016 l->l_lwpctl = NULL;
2017 l->l_lcpage = NULL;
2018 p = l->l_proc;
2019 lp = p->p_lwpctl;
2020
2021 KASSERT(lp != NULL);
2022 KASSERT(p->p_nlwps == 1);
2023
2024 for (lcp = TAILQ_FIRST(&lp->lp_pages); lcp != NULL; lcp = next) {
2025 next = TAILQ_NEXT(lcp, lcp_chain);
2026 uvm_unmap(kernel_map, lcp->lcp_kaddr,
2027 lcp->lcp_kaddr + PAGE_SIZE);
2028 kmem_free(lcp, LWPCTL_LCPAGE_SZ);
2029 }
2030
2031 if (lp->lp_uao != NULL) {
2032 uvm_unmap(&p->p_vmspace->vm_map, lp->lp_uva,
2033 lp->lp_uva + LWPCTL_UAREA_SZ);
2034 }
2035
2036 mutex_destroy(&lp->lp_lock);
2037 kmem_free(lp, sizeof(*lp));
2038 p->p_lwpctl = NULL;
2039 }
2040
2041 /*
2042 * Return the current LWP's "preemption counter". Used to detect
2043 * preemption across operations that can tolerate preemption without
2044 * crashing, but which may generate incorrect results if preempted.
2045 */
2046 uint64_t
lwp_pctr(void)2047 lwp_pctr(void)
2048 {
2049
2050 return curlwp->l_ncsw;
2051 }
2052
2053 /*
2054 * Set an LWP's private data pointer.
2055 */
2056 int
lwp_setprivate(struct lwp * l,void * ptr)2057 lwp_setprivate(struct lwp *l, void *ptr)
2058 {
2059 int error = 0;
2060
2061 l->l_private = ptr;
2062 #ifdef __HAVE_CPU_LWP_SETPRIVATE
2063 error = cpu_lwp_setprivate(l, ptr);
2064 #endif
2065 return error;
2066 }
2067
2068 /*
2069 * Perform any thread-related cleanup on LWP exit.
2070 * N.B. l->l_proc->p_lock must be HELD on entry but will
2071 * be released before returning!
2072 */
2073 void
lwp_thread_cleanup(struct lwp * l)2074 lwp_thread_cleanup(struct lwp *l)
2075 {
2076
2077 KASSERT(mutex_owned(l->l_proc->p_lock));
2078 mutex_exit(l->l_proc->p_lock);
2079
2080 /*
2081 * If the LWP has robust futexes, release them all
2082 * now.
2083 */
2084 if (__predict_false(l->l_robust_head != 0)) {
2085 futex_release_all_lwp(l);
2086 }
2087 }
2088
2089 #if defined(DDB)
2090 #include <machine/pcb.h>
2091
2092 void
lwp_whatis(uintptr_t addr,void (* pr)(const char *,...))2093 lwp_whatis(uintptr_t addr, void (*pr)(const char *, ...))
2094 {
2095 lwp_t *l;
2096
2097 LIST_FOREACH(l, &alllwp, l_list) {
2098 uintptr_t stack = (uintptr_t)KSTACK_LOWEST_ADDR(l);
2099
2100 if (addr < stack || stack + KSTACK_SIZE <= addr) {
2101 continue;
2102 }
2103 (*pr)("%p is %p+%zu, LWP %p's stack\n",
2104 (void *)addr, (void *)stack,
2105 (size_t)(addr - stack), l);
2106 }
2107 }
2108 #endif /* defined(DDB) */
2109