FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_lwp.c
1 /* $NetBSD: kern_lwp.c,v 1.126.2.2 2009/03/08 03:15:36 snj Exp $ */
2
3 /*-
4 * Copyright (c) 2001, 2006, 2007, 2008, 2009 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Nathan J. Williams, and Andrew Doran.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
30 */
31
32 /*
33 * Overview
34 *
35 * Lightweight processes (LWPs) are the basic unit or thread of
36 * execution within the kernel. The core state of an LWP is described
37 * by "struct lwp", also known as lwp_t.
38 *
39 * Each LWP is contained within a process (described by "struct proc"),
40 * Every process contains at least one LWP, but may contain more. The
41 * process describes attributes shared among all of its LWPs such as a
42 * private address space, global execution state (stopped, active,
43 * zombie, ...), signal disposition and so on. On a multiprocessor
44 * machine, multiple LWPs be executing concurrently in the kernel.
45 *
46 * Execution states
47 *
48 * At any given time, an LWP has overall state that is described by
49 * lwp::l_stat. The states are broken into two sets below. The first
50 * set is guaranteed to represent the absolute, current state of the
51 * LWP:
52 *
53 * LSONPROC
54 *
55 * On processor: the LWP is executing on a CPU, either in the
56 * kernel or in user space.
57 *
58 * LSRUN
59 *
60 * Runnable: the LWP is parked on a run queue, and may soon be
61 * chosen to run by an idle processor, or by a processor that
62 * has been asked to preempt a currently runnning but lower
63 * priority LWP. If the LWP is not swapped in (LW_INMEM == 0)
64 * then the LWP is not on a run queue, but may be soon.
65 *
66 * LSIDL
67 *
68 * Idle: the LWP has been created but has not yet executed,
69 * or it has ceased executing a unit of work and is waiting
70 * to be started again.
71 *
72 * LSSUSPENDED:
73 *
74 * Suspended: the LWP has had its execution suspended by
75 * another LWP in the same process using the _lwp_suspend()
76 * system call. User-level LWPs also enter the suspended
77 * state when the system is shutting down.
78 *
79 * The second set represent a "statement of intent" on behalf of the
80 * LWP. The LWP may in fact be executing on a processor, may be
81 * sleeping or idle. It is expected to take the necessary action to
82 * stop executing or become "running" again within a short timeframe.
83 * The LP_RUNNING flag in lwp::l_pflag indicates that an LWP is running.
84 * Importantly, it indicates that its state is tied to a CPU.
85 *
86 * LSZOMB:
87 *
88 * Dead or dying: the LWP has released most of its resources
89 * and is: a) about to switch away into oblivion b) has already
90 * switched away. When it switches away, its few remaining
91 * resources can be collected.
92 *
93 * LSSLEEP:
94 *
95 * Sleeping: the LWP has entered itself onto a sleep queue, and
96 * has switched away or will switch away shortly to allow other
97 * LWPs to run on the CPU.
98 *
99 * LSSTOP:
100 *
101 * Stopped: the LWP has been stopped as a result of a job
102 * control signal, or as a result of the ptrace() interface.
103 *
104 * Stopped LWPs may run briefly within the kernel to handle
105 * signals that they receive, but will not return to user space
106 * until their process' state is changed away from stopped.
107 *
108 * Single LWPs within a process can not be set stopped
109 * selectively: all actions that can stop or continue LWPs
110 * occur at the process level.
111 *
112 * State transitions
113 *
114 * Note that the LSSTOP state may only be set when returning to
115 * user space in userret(), or when sleeping interruptably. The
116 * LSSUSPENDED state may only be set in userret(). Before setting
117 * those states, we try to ensure that the LWPs will release all
118 * locks that they hold, and at a minimum try to ensure that the
119 * LWP can be set runnable again by a signal.
120 *
121 * LWPs may transition states in the following ways:
122 *
123 * RUN -------> ONPROC ONPROC -----> RUN
124 * > STOPPED > SLEEP
125 * > SUSPENDED > STOPPED
126 * > SUSPENDED
127 * > ZOMB
128 *
129 * STOPPED ---> RUN SUSPENDED --> RUN
130 * > SLEEP > SLEEP
131 *
132 * SLEEP -----> ONPROC IDL --------> RUN
133 * > RUN > SUSPENDED
134 * > STOPPED > STOPPED
135 * > SUSPENDED
136 *
137 * Other state transitions are possible with kernel threads (eg
138 * ONPROC -> IDL), but only happen under tightly controlled
139 * circumstances the side effects are understood.
140 *
141 * Migration
142 *
143 * Migration of threads from one CPU to another could be performed
144 * internally by the scheduler via sched_takecpu() or sched_catchlwp()
145 * functions. The universal lwp_migrate() function should be used for
146 * any other cases. Subsystems in the kernel must be aware that CPU
147 * of LWP may change, while it is not locked.
148 *
149 * Locking
150 *
151 * The majority of fields in 'struct lwp' are covered by a single,
152 * general spin lock pointed to by lwp::l_mutex. The locks covering
153 * each field are documented in sys/lwp.h.
154 *
155 * State transitions must be made with the LWP's general lock held,
156 * and may cause the LWP's lock pointer to change. Manipulation of
157 * the general lock is not performed directly, but through calls to
158 * lwp_lock(), lwp_relock() and similar.
159 *
160 * States and their associated locks:
161 *
162 * LSONPROC, LSZOMB:
163 *
164 * Always covered by spc_lwplock, which protects running LWPs.
165 * This is a per-CPU lock.
166 *
167 * LSIDL, LSRUN:
168 *
169 * Always covered by spc_mutex, which protects the run queues.
170 * This is a per-CPU lock.
171 *
172 * LSSLEEP:
173 *
174 * Covered by a lock associated with the sleep queue that the
175 * LWP resides on.
176 *
177 * LSSTOP, LSSUSPENDED:
178 *
179 * If the LWP was previously sleeping (l_wchan != NULL), then
180 * l_mutex references the sleep queue lock. If the LWP was
181 * runnable or on the CPU when halted, or has been removed from
182 * the sleep queue since halted, then the lock is spc_lwplock.
183 *
184 * The lock order is as follows:
185 *
186 * spc::spc_lwplock ->
187 * sleeptab::st_mutex ->
188 * tschain_t::tc_mutex ->
189 * spc::spc_mutex
190 *
191 * Each process has an scheduler state lock (proc::p_lock), and a
192 * number of counters on LWPs and their states: p_nzlwps, p_nrlwps, and
193 * so on. When an LWP is to be entered into or removed from one of the
194 * following states, p_lock must be held and the process wide counters
195 * adjusted:
196 *
197 * LSIDL, LSZOMB, LSSTOP, LSSUSPENDED
198 *
199 * Note that an LWP is considered running or likely to run soon if in
200 * one of the following states. This affects the value of p_nrlwps:
201 *
202 * LSRUN, LSONPROC, LSSLEEP
203 *
204 * p_lock does not need to be held when transitioning among these
205 * three states.
206 */
207
208 #include <sys/cdefs.h>
209 __KERNEL_RCSID(0, "$NetBSD: kern_lwp.c,v 1.126.2.2 2009/03/08 03:15:36 snj Exp $");
210
211 #include "opt_ddb.h"
212 #include "opt_lockdebug.h"
213 #include "opt_sa.h"
214
215 #define _LWP_API_PRIVATE
216
217 #include <sys/param.h>
218 #include <sys/systm.h>
219 #include <sys/cpu.h>
220 #include <sys/pool.h>
221 #include <sys/proc.h>
222 #include <sys/sa.h>
223 #include <sys/savar.h>
224 #include <sys/syscallargs.h>
225 #include <sys/syscall_stats.h>
226 #include <sys/kauth.h>
227 #include <sys/sleepq.h>
228 #include <sys/user.h>
229 #include <sys/lockdebug.h>
230 #include <sys/kmem.h>
231 #include <sys/pset.h>
232 #include <sys/intr.h>
233 #include <sys/lwpctl.h>
234 #include <sys/atomic.h>
235
236 #include <uvm/uvm_extern.h>
237 #include <uvm/uvm_object.h>
238
239 struct lwplist alllwp = LIST_HEAD_INITIALIZER(alllwp);
240
241 POOL_INIT(lwp_uc_pool, sizeof(ucontext_t), 0, 0, 0, "lwpucpl",
242 &pool_allocator_nointr, IPL_NONE);
243
244 static pool_cache_t lwp_cache;
245 static specificdata_domain_t lwp_specificdata_domain;
246
247 void
248 lwpinit(void)
249 {
250
251 lwp_specificdata_domain = specificdata_domain_create();
252 KASSERT(lwp_specificdata_domain != NULL);
253 lwp_sys_init();
254 lwp_cache = pool_cache_init(sizeof(lwp_t), MIN_LWP_ALIGNMENT, 0, 0,
255 "lwppl", NULL, IPL_NONE, NULL, NULL, NULL);
256 }
257
258 /*
259 * Set an suspended.
260 *
261 * Must be called with p_lock held, and the LWP locked. Will unlock the
262 * LWP before return.
263 */
264 int
265 lwp_suspend(struct lwp *curl, struct lwp *t)
266 {
267 int error;
268
269 KASSERT(mutex_owned(t->l_proc->p_lock));
270 KASSERT(lwp_locked(t, NULL));
271
272 KASSERT(curl != t || curl->l_stat == LSONPROC);
273
274 /*
275 * If the current LWP has been told to exit, we must not suspend anyone
276 * else or deadlock could occur. We won't return to userspace.
277 */
278 if ((curl->l_flag & (LW_WEXIT | LW_WCORE)) != 0) {
279 lwp_unlock(t);
280 return (EDEADLK);
281 }
282
283 error = 0;
284
285 switch (t->l_stat) {
286 case LSRUN:
287 case LSONPROC:
288 t->l_flag |= LW_WSUSPEND;
289 lwp_need_userret(t);
290 lwp_unlock(t);
291 break;
292
293 case LSSLEEP:
294 t->l_flag |= LW_WSUSPEND;
295
296 /*
297 * Kick the LWP and try to get it to the kernel boundary
298 * so that it will release any locks that it holds.
299 * setrunnable() will release the lock.
300 */
301 if ((t->l_flag & LW_SINTR) != 0)
302 setrunnable(t);
303 else
304 lwp_unlock(t);
305 break;
306
307 case LSSUSPENDED:
308 lwp_unlock(t);
309 break;
310
311 case LSSTOP:
312 t->l_flag |= LW_WSUSPEND;
313 setrunnable(t);
314 break;
315
316 case LSIDL:
317 case LSZOMB:
318 error = EINTR; /* It's what Solaris does..... */
319 lwp_unlock(t);
320 break;
321 }
322
323 return (error);
324 }
325
326 /*
327 * Restart a suspended LWP.
328 *
329 * Must be called with p_lock held, and the LWP locked. Will unlock the
330 * LWP before return.
331 */
332 void
333 lwp_continue(struct lwp *l)
334 {
335
336 KASSERT(mutex_owned(l->l_proc->p_lock));
337 KASSERT(lwp_locked(l, NULL));
338
339 /* If rebooting or not suspended, then just bail out. */
340 if ((l->l_flag & LW_WREBOOT) != 0) {
341 lwp_unlock(l);
342 return;
343 }
344
345 l->l_flag &= ~LW_WSUSPEND;
346
347 if (l->l_stat != LSSUSPENDED) {
348 lwp_unlock(l);
349 return;
350 }
351
352 /* setrunnable() will release the lock. */
353 setrunnable(l);
354 }
355
356 /*
357 * Wait for an LWP within the current process to exit. If 'lid' is
358 * non-zero, we are waiting for a specific LWP.
359 *
360 * Must be called with p->p_lock held.
361 */
362 int
363 lwp_wait1(struct lwp *l, lwpid_t lid, lwpid_t *departed, int flags)
364 {
365 struct proc *p = l->l_proc;
366 struct lwp *l2;
367 int nfound, error;
368 lwpid_t curlid;
369 bool exiting;
370
371 KASSERT(mutex_owned(p->p_lock));
372
373 p->p_nlwpwait++;
374 l->l_waitingfor = lid;
375 curlid = l->l_lid;
376 exiting = ((flags & LWPWAIT_EXITCONTROL) != 0);
377
378 for (;;) {
379 /*
380 * Avoid a race between exit1() and sigexit(): if the
381 * process is dumping core, then we need to bail out: call
382 * into lwp_userret() where we will be suspended until the
383 * deed is done.
384 */
385 if ((p->p_sflag & PS_WCORE) != 0) {
386 mutex_exit(p->p_lock);
387 lwp_userret(l);
388 #ifdef DIAGNOSTIC
389 panic("lwp_wait1");
390 #endif
391 /* NOTREACHED */
392 }
393
394 /*
395 * First off, drain any detached LWP that is waiting to be
396 * reaped.
397 */
398 while ((l2 = p->p_zomblwp) != NULL) {
399 p->p_zomblwp = NULL;
400 lwp_free(l2, false, false);/* releases proc mutex */
401 mutex_enter(p->p_lock);
402 }
403
404 /*
405 * Now look for an LWP to collect. If the whole process is
406 * exiting, count detached LWPs as eligible to be collected,
407 * but don't drain them here.
408 */
409 nfound = 0;
410 error = 0;
411 LIST_FOREACH(l2, &p->p_lwps, l_sibling) {
412 /*
413 * If a specific wait and the target is waiting on
414 * us, then avoid deadlock. This also traps LWPs
415 * that try to wait on themselves.
416 *
417 * Note that this does not handle more complicated
418 * cycles, like: t1 -> t2 -> t3 -> t1. The process
419 * can still be killed so it is not a major problem.
420 */
421 if (l2->l_lid == lid && l2->l_waitingfor == curlid) {
422 error = EDEADLK;
423 break;
424 }
425 if (l2 == l)
426 continue;
427 if ((l2->l_prflag & LPR_DETACHED) != 0) {
428 nfound += exiting;
429 continue;
430 }
431 if (lid != 0) {
432 if (l2->l_lid != lid)
433 continue;
434 /*
435 * Mark this LWP as the first waiter, if there
436 * is no other.
437 */
438 if (l2->l_waiter == 0)
439 l2->l_waiter = curlid;
440 } else if (l2->l_waiter != 0) {
441 /*
442 * It already has a waiter - so don't
443 * collect it. If the waiter doesn't
444 * grab it we'll get another chance
445 * later.
446 */
447 nfound++;
448 continue;
449 }
450 nfound++;
451
452 /* No need to lock the LWP in order to see LSZOMB. */
453 if (l2->l_stat != LSZOMB)
454 continue;
455
456 /*
457 * We're no longer waiting. Reset the "first waiter"
458 * pointer on the target, in case it was us.
459 */
460 l->l_waitingfor = 0;
461 l2->l_waiter = 0;
462 p->p_nlwpwait--;
463 if (departed)
464 *departed = l2->l_lid;
465 sched_lwp_collect(l2);
466
467 /* lwp_free() releases the proc lock. */
468 lwp_free(l2, false, false);
469 mutex_enter(p->p_lock);
470 return 0;
471 }
472
473 if (error != 0)
474 break;
475 if (nfound == 0) {
476 error = ESRCH;
477 break;
478 }
479
480 /*
481 * The kernel is careful to ensure that it can not deadlock
482 * when exiting - just keep waiting.
483 */
484 if (exiting) {
485 KASSERT(p->p_nlwps > 1);
486 cv_wait(&p->p_lwpcv, p->p_lock);
487 continue;
488 }
489
490 /*
491 * If all other LWPs are waiting for exits or suspends
492 * and the supply of zombies and potential zombies is
493 * exhausted, then we are about to deadlock.
494 *
495 * If the process is exiting (and this LWP is not the one
496 * that is coordinating the exit) then bail out now.
497 */
498 if ((p->p_sflag & PS_WEXIT) != 0 ||
499 p->p_nrlwps + p->p_nzlwps - p->p_ndlwps <= p->p_nlwpwait) {
500 error = EDEADLK;
501 break;
502 }
503
504 /*
505 * Sit around and wait for something to happen. We'll be
506 * awoken if any of the conditions examined change: if an
507 * LWP exits, is collected, or is detached.
508 */
509 if ((error = cv_wait_sig(&p->p_lwpcv, p->p_lock)) != 0)
510 break;
511 }
512
513 /*
514 * We didn't find any LWPs to collect, we may have received a
515 * signal, or some other condition has caused us to bail out.
516 *
517 * If waiting on a specific LWP, clear the waiters marker: some
518 * other LWP may want it. Then, kick all the remaining waiters
519 * so that they can re-check for zombies and for deadlock.
520 */
521 if (lid != 0) {
522 LIST_FOREACH(l2, &p->p_lwps, l_sibling) {
523 if (l2->l_lid == lid) {
524 if (l2->l_waiter == curlid)
525 l2->l_waiter = 0;
526 break;
527 }
528 }
529 }
530 p->p_nlwpwait--;
531 l->l_waitingfor = 0;
532 cv_broadcast(&p->p_lwpcv);
533
534 return error;
535 }
536
537 /*
538 * Create a new LWP within process 'p2', using LWP 'l1' as a template.
539 * The new LWP is created in state LSIDL and must be set running,
540 * suspended, or stopped by the caller.
541 */
542 int
543 lwp_create(lwp_t *l1, proc_t *p2, vaddr_t uaddr, bool inmem, int flags,
544 void *stack, size_t stacksize, void (*func)(void *), void *arg,
545 lwp_t **rnewlwpp, int sclass)
546 {
547 struct lwp *l2, *isfree;
548 turnstile_t *ts;
549
550 KASSERT(l1 == curlwp || l1->l_proc == &proc0);
551
552 /*
553 * First off, reap any detached LWP waiting to be collected.
554 * We can re-use its LWP structure and turnstile.
555 */
556 isfree = NULL;
557 if (p2->p_zomblwp != NULL) {
558 mutex_enter(p2->p_lock);
559 if ((isfree = p2->p_zomblwp) != NULL) {
560 p2->p_zomblwp = NULL;
561 lwp_free(isfree, true, false);/* releases proc mutex */
562 } else
563 mutex_exit(p2->p_lock);
564 }
565 if (isfree == NULL) {
566 l2 = pool_cache_get(lwp_cache, PR_WAITOK);
567 memset(l2, 0, sizeof(*l2));
568 l2->l_ts = pool_cache_get(turnstile_cache, PR_WAITOK);
569 SLIST_INIT(&l2->l_pi_lenders);
570 } else {
571 l2 = isfree;
572 ts = l2->l_ts;
573 KASSERT(l2->l_inheritedprio == -1);
574 KASSERT(SLIST_EMPTY(&l2->l_pi_lenders));
575 memset(l2, 0, sizeof(*l2));
576 l2->l_ts = ts;
577 }
578
579 l2->l_stat = LSIDL;
580 l2->l_proc = p2;
581 l2->l_refcnt = 1;
582 l2->l_class = sclass;
583
584 /*
585 * If vfork(), we want the LWP to run fast and on the same CPU
586 * as its parent, so that it can reuse the VM context and cache
587 * footprint on the local CPU.
588 */
589 l2->l_kpriority = ((flags & LWP_VFORK) ? true : false);
590 l2->l_kpribase = PRI_KERNEL;
591 l2->l_priority = l1->l_priority;
592 l2->l_inheritedprio = -1;
593 l2->l_flag = inmem ? LW_INMEM : 0;
594 l2->l_pflag = LP_MPSAFE;
595 l2->l_fd = p2->p_fd;
596 TAILQ_INIT(&l2->l_ld_locks);
597
598 if (p2->p_flag & PK_SYSTEM) {
599 /* Mark it as a system LWP and not a candidate for swapping */
600 l2->l_flag |= LW_SYSTEM;
601 }
602
603 kpreempt_disable();
604 l2->l_mutex = l1->l_cpu->ci_schedstate.spc_mutex;
605 l2->l_cpu = l1->l_cpu;
606 kpreempt_enable();
607
608 lwp_initspecific(l2);
609 sched_lwp_fork(l1, l2);
610 lwp_update_creds(l2);
611 callout_init(&l2->l_timeout_ch, CALLOUT_MPSAFE);
612 callout_setfunc(&l2->l_timeout_ch, sleepq_timeout, l2);
613 mutex_init(&l2->l_swaplock, MUTEX_DEFAULT, IPL_NONE);
614 cv_init(&l2->l_sigcv, "sigwait");
615 l2->l_syncobj = &sched_syncobj;
616
617 if (rnewlwpp != NULL)
618 *rnewlwpp = l2;
619
620 l2->l_addr = UAREA_TO_USER(uaddr);
621 uvm_lwp_fork(l1, l2, stack, stacksize, func,
622 (arg != NULL) ? arg : l2);
623
624 mutex_enter(p2->p_lock);
625
626 if ((flags & LWP_DETACHED) != 0) {
627 l2->l_prflag = LPR_DETACHED;
628 p2->p_ndlwps++;
629 } else
630 l2->l_prflag = 0;
631
632 l2->l_sigmask = l1->l_sigmask;
633 CIRCLEQ_INIT(&l2->l_sigpend.sp_info);
634 sigemptyset(&l2->l_sigpend.sp_set);
635
636 p2->p_nlwpid++;
637 if (p2->p_nlwpid == 0)
638 p2->p_nlwpid++;
639 l2->l_lid = p2->p_nlwpid;
640 LIST_INSERT_HEAD(&p2->p_lwps, l2, l_sibling);
641 p2->p_nlwps++;
642
643 if ((p2->p_flag & PK_SYSTEM) == 0) {
644 /* Inherit an affinity */
645 if (l1->l_flag & LW_AFFINITY) {
646 /*
647 * Note that we hold the state lock while inheriting
648 * the affinity to avoid race with sched_setaffinity().
649 */
650 lwp_lock(l1);
651 if (l1->l_flag & LW_AFFINITY) {
652 kcpuset_use(l1->l_affinity);
653 l2->l_affinity = l1->l_affinity;
654 l2->l_flag |= LW_AFFINITY;
655 }
656 lwp_unlock(l1);
657 }
658 lwp_lock(l2);
659 /* Inherit a processor-set */
660 l2->l_psid = l1->l_psid;
661 /* Look for a CPU to start */
662 l2->l_cpu = sched_takecpu(l2);
663 lwp_unlock_to(l2, l2->l_cpu->ci_schedstate.spc_mutex);
664 }
665 mutex_exit(p2->p_lock);
666
667 mutex_enter(proc_lock);
668 LIST_INSERT_HEAD(&alllwp, l2, l_list);
669 mutex_exit(proc_lock);
670
671 SYSCALL_TIME_LWP_INIT(l2);
672
673 if (p2->p_emul->e_lwp_fork)
674 (*p2->p_emul->e_lwp_fork)(l1, l2);
675
676 return (0);
677 }
678
679 /*
680 * Called by MD code when a new LWP begins execution. Must be called
681 * with the previous LWP locked (so at splsched), or if there is no
682 * previous LWP, at splsched.
683 */
684 void
685 lwp_startup(struct lwp *prev, struct lwp *new)
686 {
687
688 KASSERT(kpreempt_disabled());
689 if (prev != NULL) {
690 /*
691 * Normalize the count of the spin-mutexes, it was
692 * increased in mi_switch(). Unmark the state of
693 * context switch - it is finished for previous LWP.
694 */
695 curcpu()->ci_mtx_count++;
696 membar_exit();
697 prev->l_ctxswtch = 0;
698 }
699 KPREEMPT_DISABLE(new);
700 spl0();
701 pmap_activate(new);
702 LOCKDEBUG_BARRIER(NULL, 0);
703 KPREEMPT_ENABLE(new);
704 if ((new->l_pflag & LP_MPSAFE) == 0) {
705 KERNEL_LOCK(1, new);
706 }
707 }
708
709 /*
710 * Exit an LWP.
711 */
712 void
713 lwp_exit(struct lwp *l)
714 {
715 struct proc *p = l->l_proc;
716 struct lwp *l2;
717 bool current;
718
719 current = (l == curlwp);
720
721 KASSERT(current || (l->l_stat == LSIDL && l->l_target_cpu == NULL));
722
723 /*
724 * Verify that we hold no locks other than the kernel lock.
725 */
726 LOCKDEBUG_BARRIER(&kernel_lock, 0);
727
728 /*
729 * If we are the last live LWP in a process, we need to exit the
730 * entire process. We do so with an exit status of zero, because
731 * it's a "controlled" exit, and because that's what Solaris does.
732 *
733 * We are not quite a zombie yet, but for accounting purposes we
734 * must increment the count of zombies here.
735 *
736 * Note: the last LWP's specificdata will be deleted here.
737 */
738 mutex_enter(p->p_lock);
739 if (p->p_nlwps - p->p_nzlwps == 1) {
740 KASSERT(current == true);
741 /* XXXSMP kernel_lock not held */
742 exit1(l, 0);
743 /* NOTREACHED */
744 }
745 p->p_nzlwps++;
746 mutex_exit(p->p_lock);
747
748 if (p->p_emul->e_lwp_exit)
749 (*p->p_emul->e_lwp_exit)(l);
750
751 /* Delete the specificdata while it's still safe to sleep. */
752 specificdata_fini(lwp_specificdata_domain, &l->l_specdataref);
753
754 /*
755 * Release our cached credentials.
756 */
757 kauth_cred_free(l->l_cred);
758 callout_destroy(&l->l_timeout_ch);
759
760 /*
761 * While we can still block, mark the LWP as unswappable to
762 * prevent conflicts with the with the swapper.
763 */
764 if (current)
765 uvm_lwp_hold(l);
766
767 /*
768 * Remove the LWP from the global list.
769 */
770 mutex_enter(proc_lock);
771 LIST_REMOVE(l, l_list);
772 mutex_exit(proc_lock);
773
774 /*
775 * Get rid of all references to the LWP that others (e.g. procfs)
776 * may have, and mark the LWP as a zombie. If the LWP is detached,
777 * mark it waiting for collection in the proc structure. Note that
778 * before we can do that, we need to free any other dead, deatched
779 * LWP waiting to meet its maker.
780 */
781 mutex_enter(p->p_lock);
782 lwp_drainrefs(l);
783
784 if ((l->l_prflag & LPR_DETACHED) != 0) {
785 while ((l2 = p->p_zomblwp) != NULL) {
786 p->p_zomblwp = NULL;
787 lwp_free(l2, false, false);/* releases proc mutex */
788 mutex_enter(p->p_lock);
789 l->l_refcnt++;
790 lwp_drainrefs(l);
791 }
792 p->p_zomblwp = l;
793 }
794
795 /*
796 * If we find a pending signal for the process and we have been
797 * asked to check for signals, then we loose: arrange to have
798 * all other LWPs in the process check for signals.
799 */
800 if ((l->l_flag & LW_PENDSIG) != 0 &&
801 firstsig(&p->p_sigpend.sp_set) != 0) {
802 LIST_FOREACH(l2, &p->p_lwps, l_sibling) {
803 lwp_lock(l2);
804 l2->l_flag |= LW_PENDSIG;
805 lwp_unlock(l2);
806 }
807 }
808
809 lwp_lock(l);
810 l->l_stat = LSZOMB;
811 if (l->l_name != NULL)
812 strcpy(l->l_name, "(zombie)");
813 if (l->l_flag & LW_AFFINITY) {
814 l->l_flag &= ~LW_AFFINITY;
815 } else {
816 KASSERT(l->l_affinity == NULL);
817 }
818 lwp_unlock(l);
819 p->p_nrlwps--;
820 cv_broadcast(&p->p_lwpcv);
821 if (l->l_lwpctl != NULL)
822 l->l_lwpctl->lc_curcpu = LWPCTL_CPU_EXITED;
823 mutex_exit(p->p_lock);
824
825 /* Safe without lock since LWP is in zombie state */
826 if (l->l_affinity) {
827 kcpuset_unuse(l->l_affinity, NULL);
828 l->l_affinity = NULL;
829 }
830
831 /*
832 * We can no longer block. At this point, lwp_free() may already
833 * be gunning for us. On a multi-CPU system, we may be off p_lwps.
834 *
835 * Free MD LWP resources.
836 */
837 #ifndef __NO_CPU_LWP_FREE
838 cpu_lwp_free(l, 0);
839 #endif
840
841 if (current) {
842 pmap_deactivate(l);
843
844 /*
845 * Release the kernel lock, and switch away into
846 * oblivion.
847 */
848 #ifdef notyet
849 /* XXXSMP hold in lwp_userret() */
850 KERNEL_UNLOCK_LAST(l);
851 #else
852 KERNEL_UNLOCK_ALL(l, NULL);
853 #endif
854 lwp_exit_switchaway(l);
855 }
856 }
857
858 /*
859 * Free a dead LWP's remaining resources.
860 *
861 * XXXLWP limits.
862 */
863 void
864 lwp_free(struct lwp *l, bool recycle, bool last)
865 {
866 struct proc *p = l->l_proc;
867 struct rusage *ru;
868 ksiginfoq_t kq;
869
870 KASSERT(l != curlwp);
871
872 /*
873 * If this was not the last LWP in the process, then adjust
874 * counters and unlock.
875 */
876 if (!last) {
877 /*
878 * Add the LWP's run time to the process' base value.
879 * This needs to co-incide with coming off p_lwps.
880 */
881 bintime_add(&p->p_rtime, &l->l_rtime);
882 p->p_pctcpu += l->l_pctcpu;
883 ru = &p->p_stats->p_ru;
884 ruadd(ru, &l->l_ru);
885 ru->ru_nvcsw += (l->l_ncsw - l->l_nivcsw);
886 ru->ru_nivcsw += l->l_nivcsw;
887 LIST_REMOVE(l, l_sibling);
888 p->p_nlwps--;
889 p->p_nzlwps--;
890 if ((l->l_prflag & LPR_DETACHED) != 0)
891 p->p_ndlwps--;
892
893 /*
894 * Have any LWPs sleeping in lwp_wait() recheck for
895 * deadlock.
896 */
897 cv_broadcast(&p->p_lwpcv);
898 mutex_exit(p->p_lock);
899 }
900
901 #ifdef MULTIPROCESSOR
902 /*
903 * In the unlikely event that the LWP is still on the CPU,
904 * then spin until it has switched away. We need to release
905 * all locks to avoid deadlock against interrupt handlers on
906 * the target CPU.
907 */
908 if ((l->l_pflag & LP_RUNNING) != 0 || l->l_cpu->ci_curlwp == l) {
909 int count;
910 (void)count; /* XXXgcc */
911 KERNEL_UNLOCK_ALL(curlwp, &count);
912 while ((l->l_pflag & LP_RUNNING) != 0 ||
913 l->l_cpu->ci_curlwp == l)
914 SPINLOCK_BACKOFF_HOOK;
915 KERNEL_LOCK(count, curlwp);
916 }
917 #endif
918
919 /*
920 * Destroy the LWP's remaining signal information.
921 */
922 ksiginfo_queue_init(&kq);
923 sigclear(&l->l_sigpend, NULL, &kq);
924 ksiginfo_queue_drain(&kq);
925 cv_destroy(&l->l_sigcv);
926 mutex_destroy(&l->l_swaplock);
927
928 /*
929 * Free the LWP's turnstile and the LWP structure itself unless the
930 * caller wants to recycle them. Also, free the scheduler specific
931 * data.
932 *
933 * We can't return turnstile0 to the pool (it didn't come from it),
934 * so if it comes up just drop it quietly and move on.
935 *
936 * We don't recycle the VM resources at this time.
937 */
938 if (l->l_lwpctl != NULL)
939 lwp_ctl_free(l);
940
941 if (!recycle && l->l_ts != &turnstile0)
942 pool_cache_put(turnstile_cache, l->l_ts);
943 if (l->l_name != NULL)
944 kmem_free(l->l_name, MAXCOMLEN);
945 #ifndef __NO_CPU_LWP_FREE
946 cpu_lwp_free2(l);
947 #endif
948 KASSERT((l->l_flag & LW_INMEM) != 0);
949 uvm_lwp_exit(l);
950 KASSERT(SLIST_EMPTY(&l->l_pi_lenders));
951 KASSERT(l->l_inheritedprio == -1);
952 if (!recycle)
953 pool_cache_put(lwp_cache, l);
954 }
955
956 /*
957 * Migrate the LWP to the another CPU. Unlocks the LWP.
958 */
959 void
960 lwp_migrate(lwp_t *l, struct cpu_info *tci)
961 {
962 struct schedstate_percpu *tspc;
963 int lstat = l->l_stat;
964
965 KASSERT(lwp_locked(l, NULL));
966 KASSERT(tci != NULL);
967
968 /* If LWP is still on the CPU, it must be handled like LSONPROC */
969 if ((l->l_pflag & LP_RUNNING) != 0) {
970 lstat = LSONPROC;
971 }
972
973 /*
974 * The destination CPU could be changed while previous migration
975 * was not finished.
976 */
977 if (l->l_target_cpu != NULL) {
978 l->l_target_cpu = tci;
979 lwp_unlock(l);
980 return;
981 }
982
983 /* Nothing to do if trying to migrate to the same CPU */
984 if (l->l_cpu == tci) {
985 lwp_unlock(l);
986 return;
987 }
988
989 KASSERT(l->l_target_cpu == NULL);
990 tspc = &tci->ci_schedstate;
991 switch (lstat) {
992 case LSRUN:
993 if (l->l_flag & LW_INMEM) {
994 l->l_target_cpu = tci;
995 lwp_unlock(l);
996 return;
997 }
998 case LSIDL:
999 l->l_cpu = tci;
1000 lwp_unlock_to(l, tspc->spc_mutex);
1001 return;
1002 case LSSLEEP:
1003 l->l_cpu = tci;
1004 break;
1005 case LSSTOP:
1006 case LSSUSPENDED:
1007 l->l_cpu = tci;
1008 if (l->l_wchan == NULL) {
1009 lwp_unlock_to(l, tspc->spc_lwplock);
1010 return;
1011 }
1012 break;
1013 case LSONPROC:
1014 l->l_target_cpu = tci;
1015 spc_lock(l->l_cpu);
1016 cpu_need_resched(l->l_cpu, RESCHED_KPREEMPT);
1017 spc_unlock(l->l_cpu);
1018 break;
1019 }
1020 lwp_unlock(l);
1021 }
1022
1023 /*
1024 * Find the LWP in the process. Arguments may be zero, in such case,
1025 * the calling process and first LWP in the list will be used.
1026 * On success - returns proc locked.
1027 */
1028 struct lwp *
1029 lwp_find2(pid_t pid, lwpid_t lid)
1030 {
1031 proc_t *p;
1032 lwp_t *l;
1033
1034 /* Find the process */
1035 p = (pid == 0) ? curlwp->l_proc : p_find(pid, PFIND_UNLOCK_FAIL);
1036 if (p == NULL)
1037 return NULL;
1038 mutex_enter(p->p_lock);
1039 if (pid != 0) {
1040 /* Case of p_find */
1041 mutex_exit(proc_lock);
1042 }
1043
1044 /* Find the thread */
1045 l = (lid == 0) ? LIST_FIRST(&p->p_lwps) : lwp_find(p, lid);
1046 if (l == NULL) {
1047 mutex_exit(p->p_lock);
1048 }
1049
1050 return l;
1051 }
1052
1053 /*
1054 * Look up a live LWP within the speicifed process, and return it locked.
1055 *
1056 * Must be called with p->p_lock held.
1057 */
1058 struct lwp *
1059 lwp_find(struct proc *p, int id)
1060 {
1061 struct lwp *l;
1062
1063 KASSERT(mutex_owned(p->p_lock));
1064
1065 LIST_FOREACH(l, &p->p_lwps, l_sibling) {
1066 if (l->l_lid == id)
1067 break;
1068 }
1069
1070 /*
1071 * No need to lock - all of these conditions will
1072 * be visible with the process level mutex held.
1073 */
1074 if (l != NULL && (l->l_stat == LSIDL || l->l_stat == LSZOMB))
1075 l = NULL;
1076
1077 return l;
1078 }
1079
1080 /*
1081 * Update an LWP's cached credentials to mirror the process' master copy.
1082 *
1083 * This happens early in the syscall path, on user trap, and on LWP
1084 * creation. A long-running LWP can also voluntarily choose to update
1085 * it's credentials by calling this routine. This may be called from
1086 * LWP_CACHE_CREDS(), which checks l->l_cred != p->p_cred beforehand.
1087 */
1088 void
1089 lwp_update_creds(struct lwp *l)
1090 {
1091 kauth_cred_t oc;
1092 struct proc *p;
1093
1094 p = l->l_proc;
1095 oc = l->l_cred;
1096
1097 mutex_enter(p->p_lock);
1098 kauth_cred_hold(p->p_cred);
1099 l->l_cred = p->p_cred;
1100 l->l_prflag &= ~LPR_CRMOD;
1101 mutex_exit(p->p_lock);
1102 if (oc != NULL)
1103 kauth_cred_free(oc);
1104 }
1105
1106 /*
1107 * Verify that an LWP is locked, and optionally verify that the lock matches
1108 * one we specify.
1109 */
1110 int
1111 lwp_locked(struct lwp *l, kmutex_t *mtx)
1112 {
1113 kmutex_t *cur = l->l_mutex;
1114
1115 return mutex_owned(cur) && (mtx == cur || mtx == NULL);
1116 }
1117
1118 /*
1119 * Lock an LWP.
1120 */
1121 kmutex_t *
1122 lwp_lock_retry(struct lwp *l, kmutex_t *old)
1123 {
1124
1125 /*
1126 * XXXgcc ignoring kmutex_t * volatile on i386
1127 *
1128 * gcc version 4.1.2 20061021 prerelease (NetBSD nb1 20061021)
1129 */
1130 #if 1
1131 while (l->l_mutex != old) {
1132 #else
1133 for (;;) {
1134 #endif
1135 mutex_spin_exit(old);
1136 old = l->l_mutex;
1137 mutex_spin_enter(old);
1138
1139 /*
1140 * mutex_enter() will have posted a read barrier. Re-test
1141 * l->l_mutex. If it has changed, we need to try again.
1142 */
1143 #if 1
1144 }
1145 #else
1146 } while (__predict_false(l->l_mutex != old));
1147 #endif
1148
1149 return old;
1150 }
1151
1152 /*
1153 * Lend a new mutex to an LWP. The old mutex must be held.
1154 */
1155 void
1156 lwp_setlock(struct lwp *l, kmutex_t *new)
1157 {
1158
1159 KASSERT(mutex_owned(l->l_mutex));
1160
1161 membar_exit();
1162 l->l_mutex = new;
1163 }
1164
1165 /*
1166 * Lend a new mutex to an LWP, and release the old mutex. The old mutex
1167 * must be held.
1168 */
1169 void
1170 lwp_unlock_to(struct lwp *l, kmutex_t *new)
1171 {
1172 kmutex_t *old;
1173
1174 KASSERT(mutex_owned(l->l_mutex));
1175
1176 old = l->l_mutex;
1177 membar_exit();
1178 l->l_mutex = new;
1179 mutex_spin_exit(old);
1180 }
1181
1182 /*
1183 * Acquire a new mutex, and donate it to an LWP. The LWP must already be
1184 * locked.
1185 */
1186 void
1187 lwp_relock(struct lwp *l, kmutex_t *new)
1188 {
1189 kmutex_t *old;
1190
1191 KASSERT(mutex_owned(l->l_mutex));
1192
1193 old = l->l_mutex;
1194 if (old != new) {
1195 mutex_spin_enter(new);
1196 l->l_mutex = new;
1197 mutex_spin_exit(old);
1198 }
1199 }
1200
1201 int
1202 lwp_trylock(struct lwp *l)
1203 {
1204 kmutex_t *old;
1205
1206 for (;;) {
1207 if (!mutex_tryenter(old = l->l_mutex))
1208 return 0;
1209 if (__predict_true(l->l_mutex == old))
1210 return 1;
1211 mutex_spin_exit(old);
1212 }
1213 }
1214
1215 u_int
1216 lwp_unsleep(lwp_t *l, bool cleanup)
1217 {
1218
1219 KASSERT(mutex_owned(l->l_mutex));
1220
1221 return (*l->l_syncobj->sobj_unsleep)(l, cleanup);
1222 }
1223
1224
1225 /*
1226 * Handle exceptions for mi_userret(). Called if a member of LW_USERRET is
1227 * set.
1228 */
1229 void
1230 lwp_userret(struct lwp *l)
1231 {
1232 struct proc *p;
1233 void (*hook)(void);
1234 int sig;
1235
1236 KASSERT(l == curlwp);
1237 KASSERT(l->l_stat == LSONPROC);
1238 p = l->l_proc;
1239
1240 #ifndef __HAVE_FAST_SOFTINTS
1241 /* Run pending soft interrupts. */
1242 if (l->l_cpu->ci_data.cpu_softints != 0)
1243 softint_overlay();
1244 #endif
1245
1246 #ifdef KERN_SA
1247 /* Generate UNBLOCKED upcall if needed */
1248 if (l->l_flag & LW_SA_BLOCKING) {
1249 sa_unblock_userret(l);
1250 /* NOTREACHED */
1251 }
1252 #endif
1253
1254 /*
1255 * It should be safe to do this read unlocked on a multiprocessor
1256 * system..
1257 *
1258 * LW_SA_UPCALL will be handled after the while() loop, so don't
1259 * consider it now.
1260 */
1261 while ((l->l_flag & (LW_USERRET & ~(LW_SA_UPCALL))) != 0) {
1262 /*
1263 * Process pending signals first, unless the process
1264 * is dumping core or exiting, where we will instead
1265 * enter the LW_WSUSPEND case below.
1266 */
1267 if ((l->l_flag & (LW_PENDSIG | LW_WCORE | LW_WEXIT)) ==
1268 LW_PENDSIG) {
1269 mutex_enter(p->p_lock);
1270 while ((sig = issignal(l)) != 0)
1271 postsig(sig);
1272 mutex_exit(p->p_lock);
1273 }
1274
1275 /*
1276 * Core-dump or suspend pending.
1277 *
1278 * In case of core dump, suspend ourselves, so that the
1279 * kernel stack and therefore the userland registers saved
1280 * in the trapframe are around for coredump() to write them
1281 * out. We issue a wakeup on p->p_lwpcv so that sigexit()
1282 * will write the core file out once all other LWPs are
1283 * suspended.
1284 */
1285 if ((l->l_flag & LW_WSUSPEND) != 0) {
1286 mutex_enter(p->p_lock);
1287 p->p_nrlwps--;
1288 cv_broadcast(&p->p_lwpcv);
1289 lwp_lock(l);
1290 l->l_stat = LSSUSPENDED;
1291 lwp_unlock(l);
1292 mutex_exit(p->p_lock);
1293 lwp_lock(l);
1294 mi_switch(l);
1295 }
1296
1297 /* Process is exiting. */
1298 if ((l->l_flag & LW_WEXIT) != 0) {
1299 lwp_exit(l);
1300 KASSERT(0);
1301 /* NOTREACHED */
1302 }
1303
1304 /* Call userret hook; used by Linux emulation. */
1305 if ((l->l_flag & LW_WUSERRET) != 0) {
1306 lwp_lock(l);
1307 l->l_flag &= ~LW_WUSERRET;
1308 lwp_unlock(l);
1309 hook = p->p_userret;
1310 p->p_userret = NULL;
1311 (*hook)();
1312 }
1313 }
1314
1315 #ifdef KERN_SA
1316 /*
1317 * Timer events are handled specially. We only try once to deliver
1318 * pending timer upcalls; if if fails, we can try again on the next
1319 * loop around. If we need to re-enter lwp_userret(), MD code will
1320 * bounce us back here through the trap path after we return.
1321 */
1322 if (p->p_timerpend)
1323 timerupcall(l);
1324 if (l->l_flag & LW_SA_UPCALL)
1325 sa_upcall_userret(l);
1326 #endif /* KERN_SA */
1327 }
1328
1329 /*
1330 * Force an LWP to enter the kernel, to take a trip through lwp_userret().
1331 */
1332 void
1333 lwp_need_userret(struct lwp *l)
1334 {
1335 KASSERT(lwp_locked(l, NULL));
1336
1337 /*
1338 * Since the tests in lwp_userret() are done unlocked, make sure
1339 * that the condition will be seen before forcing the LWP to enter
1340 * kernel mode.
1341 */
1342 membar_producer();
1343 cpu_signotify(l);
1344 }
1345
1346 /*
1347 * Add one reference to an LWP. This will prevent the LWP from
1348 * exiting, thus keep the lwp structure and PCB around to inspect.
1349 */
1350 void
1351 lwp_addref(struct lwp *l)
1352 {
1353
1354 KASSERT(mutex_owned(l->l_proc->p_lock));
1355 KASSERT(l->l_stat != LSZOMB);
1356 KASSERT(l->l_refcnt != 0);
1357
1358 l->l_refcnt++;
1359 }
1360
1361 /*
1362 * Remove one reference to an LWP. If this is the last reference,
1363 * then we must finalize the LWP's death.
1364 */
1365 void
1366 lwp_delref(struct lwp *l)
1367 {
1368 struct proc *p = l->l_proc;
1369
1370 mutex_enter(p->p_lock);
1371 KASSERT(l->l_stat != LSZOMB);
1372 KASSERT(l->l_refcnt > 0);
1373 if (--l->l_refcnt == 0)
1374 cv_broadcast(&p->p_lwpcv);
1375 mutex_exit(p->p_lock);
1376 }
1377
1378 /*
1379 * Drain all references to the current LWP.
1380 */
1381 void
1382 lwp_drainrefs(struct lwp *l)
1383 {
1384 struct proc *p = l->l_proc;
1385
1386 KASSERT(mutex_owned(p->p_lock));
1387 KASSERT(l->l_refcnt != 0);
1388
1389 l->l_refcnt--;
1390 while (l->l_refcnt != 0)
1391 cv_wait(&p->p_lwpcv, p->p_lock);
1392 }
1393
1394 /*
1395 * Return true if the specified LWP is 'alive'. Only p->p_lock need
1396 * be held.
1397 */
1398 bool
1399 lwp_alive(lwp_t *l)
1400 {
1401
1402 KASSERT(mutex_owned(l->l_proc->p_lock));
1403
1404 switch (l->l_stat) {
1405 case LSSLEEP:
1406 case LSRUN:
1407 case LSONPROC:
1408 case LSSTOP:
1409 case LSSUSPENDED:
1410 return true;
1411 default:
1412 return false;
1413 }
1414 }
1415
1416 /*
1417 * Return first live LWP in the process.
1418 */
1419 lwp_t *
1420 lwp_find_first(proc_t *p)
1421 {
1422 lwp_t *l;
1423
1424 KASSERT(mutex_owned(p->p_lock));
1425
1426 LIST_FOREACH(l, &p->p_lwps, l_sibling) {
1427 if (lwp_alive(l)) {
1428 return l;
1429 }
1430 }
1431
1432 return NULL;
1433 }
1434
1435 /*
1436 * lwp_specific_key_create --
1437 * Create a key for subsystem lwp-specific data.
1438 */
1439 int
1440 lwp_specific_key_create(specificdata_key_t *keyp, specificdata_dtor_t dtor)
1441 {
1442
1443 return (specificdata_key_create(lwp_specificdata_domain, keyp, dtor));
1444 }
1445
1446 /*
1447 * lwp_specific_key_delete --
1448 * Delete a key for subsystem lwp-specific data.
1449 */
1450 void
1451 lwp_specific_key_delete(specificdata_key_t key)
1452 {
1453
1454 specificdata_key_delete(lwp_specificdata_domain, key);
1455 }
1456
1457 /*
1458 * lwp_initspecific --
1459 * Initialize an LWP's specificdata container.
1460 */
1461 void
1462 lwp_initspecific(struct lwp *l)
1463 {
1464 int error;
1465
1466 error = specificdata_init(lwp_specificdata_domain, &l->l_specdataref);
1467 KASSERT(error == 0);
1468 }
1469
1470 /*
1471 * lwp_finispecific --
1472 * Finalize an LWP's specificdata container.
1473 */
1474 void
1475 lwp_finispecific(struct lwp *l)
1476 {
1477
1478 specificdata_fini(lwp_specificdata_domain, &l->l_specdataref);
1479 }
1480
1481 /*
1482 * lwp_getspecific --
1483 * Return lwp-specific data corresponding to the specified key.
1484 *
1485 * Note: LWP specific data is NOT INTERLOCKED. An LWP should access
1486 * only its OWN SPECIFIC DATA. If it is necessary to access another
1487 * LWP's specifc data, care must be taken to ensure that doing so
1488 * would not cause internal data structure inconsistency (i.e. caller
1489 * can guarantee that the target LWP is not inside an lwp_getspecific()
1490 * or lwp_setspecific() call).
1491 */
1492 void *
1493 lwp_getspecific(specificdata_key_t key)
1494 {
1495
1496 return (specificdata_getspecific_unlocked(lwp_specificdata_domain,
1497 &curlwp->l_specdataref, key));
1498 }
1499
1500 void *
1501 _lwp_getspecific_by_lwp(struct lwp *l, specificdata_key_t key)
1502 {
1503
1504 return (specificdata_getspecific_unlocked(lwp_specificdata_domain,
1505 &l->l_specdataref, key));
1506 }
1507
1508 /*
1509 * lwp_setspecific --
1510 * Set lwp-specific data corresponding to the specified key.
1511 */
1512 void
1513 lwp_setspecific(specificdata_key_t key, void *data)
1514 {
1515
1516 specificdata_setspecific(lwp_specificdata_domain,
1517 &curlwp->l_specdataref, key, data);
1518 }
1519
1520 /*
1521 * Allocate a new lwpctl structure for a user LWP.
1522 */
1523 int
1524 lwp_ctl_alloc(vaddr_t *uaddr)
1525 {
1526 lcproc_t *lp;
1527 u_int bit, i, offset;
1528 struct uvm_object *uao;
1529 int error;
1530 lcpage_t *lcp;
1531 proc_t *p;
1532 lwp_t *l;
1533
1534 l = curlwp;
1535 p = l->l_proc;
1536
1537 if (l->l_lcpage != NULL) {
1538 lcp = l->l_lcpage;
1539 *uaddr = lcp->lcp_uaddr + (vaddr_t)l->l_lwpctl - lcp->lcp_kaddr;
1540 return (EINVAL);
1541 }
1542
1543 /* First time around, allocate header structure for the process. */
1544 if ((lp = p->p_lwpctl) == NULL) {
1545 lp = kmem_alloc(sizeof(*lp), KM_SLEEP);
1546 mutex_init(&lp->lp_lock, MUTEX_DEFAULT, IPL_NONE);
1547 lp->lp_uao = NULL;
1548 TAILQ_INIT(&lp->lp_pages);
1549 mutex_enter(p->p_lock);
1550 if (p->p_lwpctl == NULL) {
1551 p->p_lwpctl = lp;
1552 mutex_exit(p->p_lock);
1553 } else {
1554 mutex_exit(p->p_lock);
1555 mutex_destroy(&lp->lp_lock);
1556 kmem_free(lp, sizeof(*lp));
1557 lp = p->p_lwpctl;
1558 }
1559 }
1560
1561 /*
1562 * Set up an anonymous memory region to hold the shared pages.
1563 * Map them into the process' address space. The user vmspace
1564 * gets the first reference on the UAO.
1565 */
1566 mutex_enter(&lp->lp_lock);
1567 if (lp->lp_uao == NULL) {
1568 lp->lp_uao = uao_create(LWPCTL_UAREA_SZ, 0);
1569 lp->lp_cur = 0;
1570 lp->lp_max = LWPCTL_UAREA_SZ;
1571 lp->lp_uva = p->p_emul->e_vm_default_addr(p,
1572 (vaddr_t)p->p_vmspace->vm_daddr, LWPCTL_UAREA_SZ);
1573 error = uvm_map(&p->p_vmspace->vm_map, &lp->lp_uva,
1574 LWPCTL_UAREA_SZ, lp->lp_uao, 0, 0, UVM_MAPFLAG(UVM_PROT_RW,
1575 UVM_PROT_RW, UVM_INH_NONE, UVM_ADV_NORMAL, 0));
1576 if (error != 0) {
1577 uao_detach(lp->lp_uao);
1578 lp->lp_uao = NULL;
1579 mutex_exit(&lp->lp_lock);
1580 return error;
1581 }
1582 }
1583
1584 /* Get a free block and allocate for this LWP. */
1585 TAILQ_FOREACH(lcp, &lp->lp_pages, lcp_chain) {
1586 if (lcp->lcp_nfree != 0)
1587 break;
1588 }
1589 if (lcp == NULL) {
1590 /* Nothing available - try to set up a free page. */
1591 if (lp->lp_cur == lp->lp_max) {
1592 mutex_exit(&lp->lp_lock);
1593 return ENOMEM;
1594 }
1595 lcp = kmem_alloc(LWPCTL_LCPAGE_SZ, KM_SLEEP);
1596 if (lcp == NULL) {
1597 mutex_exit(&lp->lp_lock);
1598 return ENOMEM;
1599 }
1600 /*
1601 * Wire the next page down in kernel space. Since this
1602 * is a new mapping, we must add a reference.
1603 */
1604 uao = lp->lp_uao;
1605 (*uao->pgops->pgo_reference)(uao);
1606 lcp->lcp_kaddr = vm_map_min(kernel_map);
1607 error = uvm_map(kernel_map, &lcp->lcp_kaddr, PAGE_SIZE,
1608 uao, lp->lp_cur, PAGE_SIZE,
1609 UVM_MAPFLAG(UVM_PROT_RW, UVM_PROT_RW,
1610 UVM_INH_NONE, UVM_ADV_RANDOM, 0));
1611 if (error != 0) {
1612 mutex_exit(&lp->lp_lock);
1613 kmem_free(lcp, LWPCTL_LCPAGE_SZ);
1614 (*uao->pgops->pgo_detach)(uao);
1615 return error;
1616 }
1617 error = uvm_map_pageable(kernel_map, lcp->lcp_kaddr,
1618 lcp->lcp_kaddr + PAGE_SIZE, FALSE, 0);
1619 if (error != 0) {
1620 mutex_exit(&lp->lp_lock);
1621 uvm_unmap(kernel_map, lcp->lcp_kaddr,
1622 lcp->lcp_kaddr + PAGE_SIZE);
1623 kmem_free(lcp, LWPCTL_LCPAGE_SZ);
1624 return error;
1625 }
1626 /* Prepare the page descriptor and link into the list. */
1627 lcp->lcp_uaddr = lp->lp_uva + lp->lp_cur;
1628 lp->lp_cur += PAGE_SIZE;
1629 lcp->lcp_nfree = LWPCTL_PER_PAGE;
1630 lcp->lcp_rotor = 0;
1631 memset(lcp->lcp_bitmap, 0xff, LWPCTL_BITMAP_SZ);
1632 TAILQ_INSERT_HEAD(&lp->lp_pages, lcp, lcp_chain);
1633 }
1634 for (i = lcp->lcp_rotor; lcp->lcp_bitmap[i] == 0;) {
1635 if (++i >= LWPCTL_BITMAP_ENTRIES)
1636 i = 0;
1637 }
1638 bit = ffs(lcp->lcp_bitmap[i]) - 1;
1639 lcp->lcp_bitmap[i] ^= (1 << bit);
1640 lcp->lcp_rotor = i;
1641 lcp->lcp_nfree--;
1642 l->l_lcpage = lcp;
1643 offset = (i << 5) + bit;
1644 l->l_lwpctl = (lwpctl_t *)lcp->lcp_kaddr + offset;
1645 *uaddr = lcp->lcp_uaddr + offset * sizeof(lwpctl_t);
1646 mutex_exit(&lp->lp_lock);
1647
1648 KPREEMPT_DISABLE(l);
1649 l->l_lwpctl->lc_curcpu = (int)curcpu()->ci_data.cpu_index;
1650 KPREEMPT_ENABLE(l);
1651
1652 return 0;
1653 }
1654
1655 /*
1656 * Free an lwpctl structure back to the per-process list.
1657 */
1658 void
1659 lwp_ctl_free(lwp_t *l)
1660 {
1661 lcproc_t *lp;
1662 lcpage_t *lcp;
1663 u_int map, offset;
1664
1665 lp = l->l_proc->p_lwpctl;
1666 KASSERT(lp != NULL);
1667
1668 lcp = l->l_lcpage;
1669 offset = (u_int)((lwpctl_t *)l->l_lwpctl - (lwpctl_t *)lcp->lcp_kaddr);
1670 KASSERT(offset < LWPCTL_PER_PAGE);
1671
1672 mutex_enter(&lp->lp_lock);
1673 lcp->lcp_nfree++;
1674 map = offset >> 5;
1675 lcp->lcp_bitmap[map] |= (1 << (offset & 31));
1676 if (lcp->lcp_bitmap[lcp->lcp_rotor] == 0)
1677 lcp->lcp_rotor = map;
1678 if (TAILQ_FIRST(&lp->lp_pages)->lcp_nfree == 0) {
1679 TAILQ_REMOVE(&lp->lp_pages, lcp, lcp_chain);
1680 TAILQ_INSERT_HEAD(&lp->lp_pages, lcp, lcp_chain);
1681 }
1682 mutex_exit(&lp->lp_lock);
1683 }
1684
1685 /*
1686 * Process is exiting; tear down lwpctl state. This can only be safely
1687 * called by the last LWP in the process.
1688 */
1689 void
1690 lwp_ctl_exit(void)
1691 {
1692 lcpage_t *lcp, *next;
1693 lcproc_t *lp;
1694 proc_t *p;
1695 lwp_t *l;
1696
1697 l = curlwp;
1698 l->l_lwpctl = NULL;
1699 l->l_lcpage = NULL;
1700 p = l->l_proc;
1701 lp = p->p_lwpctl;
1702
1703 KASSERT(lp != NULL);
1704 KASSERT(p->p_nlwps == 1);
1705
1706 for (lcp = TAILQ_FIRST(&lp->lp_pages); lcp != NULL; lcp = next) {
1707 next = TAILQ_NEXT(lcp, lcp_chain);
1708 uvm_unmap(kernel_map, lcp->lcp_kaddr,
1709 lcp->lcp_kaddr + PAGE_SIZE);
1710 kmem_free(lcp, LWPCTL_LCPAGE_SZ);
1711 }
1712
1713 if (lp->lp_uao != NULL) {
1714 uvm_unmap(&p->p_vmspace->vm_map, lp->lp_uva,
1715 lp->lp_uva + LWPCTL_UAREA_SZ);
1716 }
1717
1718 mutex_destroy(&lp->lp_lock);
1719 kmem_free(lp, sizeof(*lp));
1720 p->p_lwpctl = NULL;
1721 }
1722
1723 #if defined(DDB)
1724 void
1725 lwp_whatis(uintptr_t addr, void (*pr)(const char *, ...))
1726 {
1727 lwp_t *l;
1728
1729 LIST_FOREACH(l, &alllwp, l_list) {
1730 uintptr_t stack = (uintptr_t)KSTACK_LOWEST_ADDR(l);
1731
1732 if (addr < stack || stack + KSTACK_SIZE <= addr) {
1733 continue;
1734 }
1735 (*pr)("%p is %p+%zu, LWP %p's stack\n",
1736 (void *)addr, (void *)stack,
1737 (size_t)(addr - stack), l);
1738 }
1739 }
1740 #endif /* defined(DDB) */
Cache object: 27166eb3c283aabea29bfb02603fb1c9
|