xref: /netbsd-src/sys/uvm/uvm_glue.c (revision 93bf6008f8b7982c1d1a9486e4a4a0e687fe36eb)
1 /*	$NetBSD: uvm_glue.c,v 1.137 2009/04/16 00:17:19 rmind Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993, The Regents of the University of California.
6  *
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by Charles D. Cranor,
23  *      Washington University, the University of California, Berkeley and
24  *      its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)vm_glue.c	8.6 (Berkeley) 1/5/94
42  * from: Id: uvm_glue.c,v 1.1.2.8 1998/02/07 01:16:54 chs Exp
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 #include <sys/cdefs.h>
70 __KERNEL_RCSID(0, "$NetBSD: uvm_glue.c,v 1.137 2009/04/16 00:17:19 rmind Exp $");
71 
72 #include "opt_kgdb.h"
73 #include "opt_kstack.h"
74 #include "opt_uvmhist.h"
75 
76 /*
77  * uvm_glue.c: glue functions
78  */
79 
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/proc.h>
83 #include <sys/resourcevar.h>
84 #include <sys/buf.h>
85 #include <sys/user.h>
86 #include <sys/syncobj.h>
87 #include <sys/cpu.h>
88 #include <sys/atomic.h>
89 
90 #include <uvm/uvm.h>
91 
92 /*
93  * local prototypes
94  */
95 
96 static void uvm_swapout(struct lwp *);
97 static int uarea_swapin(vaddr_t);
98 
99 /*
100  * XXXCDC: do these really belong here?
101  */
102 
103 /*
104  * uvm_kernacc: can the kernel access a region of memory
105  *
106  * - used only by /dev/kmem driver (mem.c)
107  */
108 
109 bool
110 uvm_kernacc(void *addr, size_t len, int rw)
111 {
112 	bool rv;
113 	vaddr_t saddr, eaddr;
114 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
115 
116 	saddr = trunc_page((vaddr_t)addr);
117 	eaddr = round_page((vaddr_t)addr + len);
118 	vm_map_lock_read(kernel_map);
119 	rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot);
120 	vm_map_unlock_read(kernel_map);
121 
122 	return(rv);
123 }
124 
125 #ifdef KGDB
126 /*
127  * Change protections on kernel pages from addr to addr+len
128  * (presumably so debugger can plant a breakpoint).
129  *
130  * We force the protection change at the pmap level.  If we were
131  * to use vm_map_protect a change to allow writing would be lazily-
132  * applied meaning we would still take a protection fault, something
133  * we really don't want to do.  It would also fragment the kernel
134  * map unnecessarily.  We cannot use pmap_protect since it also won't
135  * enforce a write-enable request.  Using pmap_enter is the only way
136  * we can ensure the change takes place properly.
137  */
138 void
139 uvm_chgkprot(void *addr, size_t len, int rw)
140 {
141 	vm_prot_t prot;
142 	paddr_t pa;
143 	vaddr_t sva, eva;
144 
145 	prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE;
146 	eva = round_page((vaddr_t)addr + len);
147 	for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) {
148 		/*
149 		 * Extract physical address for the page.
150 		 */
151 		if (pmap_extract(pmap_kernel(), sva, &pa) == false)
152 			panic("%s: invalid page", __func__);
153 		pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED);
154 	}
155 	pmap_update(pmap_kernel());
156 }
157 #endif
158 
159 /*
160  * uvm_vslock: wire user memory for I/O
161  *
162  * - called from physio and sys___sysctl
163  * - XXXCDC: consider nuking this (or making it a macro?)
164  */
165 
166 int
167 uvm_vslock(struct vmspace *vs, void *addr, size_t len, vm_prot_t access_type)
168 {
169 	struct vm_map *map;
170 	vaddr_t start, end;
171 	int error;
172 
173 	map = &vs->vm_map;
174 	start = trunc_page((vaddr_t)addr);
175 	end = round_page((vaddr_t)addr + len);
176 	error = uvm_fault_wire(map, start, end, access_type, 0);
177 	return error;
178 }
179 
180 /*
181  * uvm_vsunlock: unwire user memory wired by uvm_vslock()
182  *
183  * - called from physio and sys___sysctl
184  * - XXXCDC: consider nuking this (or making it a macro?)
185  */
186 
187 void
188 uvm_vsunlock(struct vmspace *vs, void *addr, size_t len)
189 {
190 	uvm_fault_unwire(&vs->vm_map, trunc_page((vaddr_t)addr),
191 		round_page((vaddr_t)addr + len));
192 }
193 
194 /*
195  * uvm_proc_fork: fork a virtual address space
196  *
197  * - the address space is copied as per parent map's inherit values
198  */
199 void
200 uvm_proc_fork(struct proc *p1, struct proc *p2, bool shared)
201 {
202 
203 	if (shared == true) {
204 		p2->p_vmspace = NULL;
205 		uvmspace_share(p1, p2);
206 	} else {
207 		p2->p_vmspace = uvmspace_fork(p1->p_vmspace);
208 	}
209 
210 	cpu_proc_fork(p1, p2);
211 }
212 
213 
214 /*
215  * uvm_lwp_fork: fork a thread
216  *
217  * - a new "user" structure is allocated for the child process
218  *	[filled in by MD layer...]
219  * - if specified, the child gets a new user stack described by
220  *	stack and stacksize
221  * - NOTE: the kernel stack may be at a different location in the child
222  *	process, and thus addresses of automatic variables may be invalid
223  *	after cpu_lwp_fork returns in the child process.  We do nothing here
224  *	after cpu_lwp_fork returns.
225  * - XXXCDC: we need a way for this to return a failure value rather
226  *   than just hang
227  */
228 void
229 uvm_lwp_fork(struct lwp *l1, struct lwp *l2, void *stack, size_t stacksize,
230     void (*func)(void *), void *arg)
231 {
232 	int error;
233 
234 	/*
235 	 * Wire down the U-area for the process, which contains the PCB
236 	 * and the kernel stack.  Wired state is stored in l->l_flag's
237 	 * L_INMEM bit rather than in the vm_map_entry's wired count
238 	 * to prevent kernel_map fragmentation.  If we reused a cached U-area,
239 	 * L_INMEM will already be set and we don't need to do anything.
240 	 *
241 	 * Note the kernel stack gets read/write accesses right off the bat.
242 	 */
243 
244 	if ((l2->l_flag & LW_INMEM) == 0) {
245 		vaddr_t uarea = USER_TO_UAREA(l2->l_addr);
246 
247 		if ((error = uarea_swapin(uarea)) != 0)
248 			panic("%s: uvm_fault_wire failed: %d", __func__, error);
249 #ifdef PMAP_UAREA
250 		/* Tell the pmap this is a u-area mapping */
251 		PMAP_UAREA(uarea);
252 #endif
253 		l2->l_flag |= LW_INMEM;
254 	}
255 
256 	/* Fill stack with magic number. */
257 	kstack_setup_magic(l2);
258 
259 	/*
260 	 * cpu_lwp_fork() copy and update the pcb, and make the child ready
261  	 * to run.  If this is a normal user fork, the child will exit
262 	 * directly to user mode via child_return() on its first time
263 	 * slice and will not return here.  If this is a kernel thread,
264 	 * the specified entry point will be executed.
265 	 */
266 	cpu_lwp_fork(l1, l2, stack, stacksize, func, arg);
267 }
268 
269 static int
270 uarea_swapin(vaddr_t addr)
271 {
272 
273 	return uvm_fault_wire(kernel_map, addr, addr + USPACE,
274 	    VM_PROT_READ | VM_PROT_WRITE, 0);
275 }
276 
277 static void
278 uarea_swapout(vaddr_t addr)
279 {
280 
281 	uvm_fault_unwire(kernel_map, addr, addr + USPACE);
282 }
283 
284 #ifndef USPACE_ALIGN
285 #define	USPACE_ALIGN	0
286 #endif
287 
288 static pool_cache_t uvm_uarea_cache;
289 
290 static int
291 uarea_ctor(void *arg, void *obj, int flags)
292 {
293 
294 	KASSERT((flags & PR_WAITOK) != 0);
295 	return uarea_swapin((vaddr_t)obj);
296 }
297 
298 static void *
299 uarea_poolpage_alloc(struct pool *pp, int flags)
300 {
301 
302 	return (void *)uvm_km_alloc(kernel_map, pp->pr_alloc->pa_pagesz,
303 	    USPACE_ALIGN, UVM_KMF_PAGEABLE |
304 	    ((flags & PR_WAITOK) != 0 ? UVM_KMF_WAITVA :
305 	    (UVM_KMF_NOWAIT | UVM_KMF_TRYLOCK)));
306 }
307 
308 static void
309 uarea_poolpage_free(struct pool *pp, void *addr)
310 {
311 
312 	uvm_km_free(kernel_map, (vaddr_t)addr, pp->pr_alloc->pa_pagesz,
313 	    UVM_KMF_PAGEABLE);
314 }
315 
316 static struct pool_allocator uvm_uarea_allocator = {
317 	.pa_alloc = uarea_poolpage_alloc,
318 	.pa_free = uarea_poolpage_free,
319 	.pa_pagesz = USPACE,
320 };
321 
322 void
323 uvm_uarea_init(void)
324 {
325 	int flags = PR_NOTOUCH;
326 
327 	/*
328 	 * specify PR_NOALIGN unless the alignment provided by
329 	 * the backend (USPACE_ALIGN) is sufficient to provide
330 	 * pool page size (UPSACE) alignment.
331 	 */
332 
333 	if ((USPACE_ALIGN == 0 && USPACE != PAGE_SIZE) ||
334 	    (USPACE_ALIGN % USPACE) != 0) {
335 		flags |= PR_NOALIGN;
336 	}
337 
338 	uvm_uarea_cache = pool_cache_init(USPACE, USPACE_ALIGN, 0, flags,
339 	    "uarea", &uvm_uarea_allocator, IPL_NONE, uarea_ctor, NULL, NULL);
340 }
341 
342 /*
343  * uvm_uarea_alloc: allocate a u-area
344  */
345 
346 bool
347 uvm_uarea_alloc(vaddr_t *uaddrp)
348 {
349 
350 	*uaddrp = (vaddr_t)pool_cache_get(uvm_uarea_cache, PR_WAITOK);
351 	return true;
352 }
353 
354 /*
355  * uvm_uarea_free: free a u-area
356  */
357 
358 void
359 uvm_uarea_free(vaddr_t uaddr, struct cpu_info *ci)
360 {
361 
362 	pool_cache_put(uvm_uarea_cache, (void *)uaddr);
363 }
364 
365 /*
366  * uvm_proc_exit: exit a virtual address space
367  *
368  * - borrow proc0's address space because freeing the vmspace
369  *   of the dead process may block.
370  */
371 
372 void
373 uvm_proc_exit(struct proc *p)
374 {
375 	struct lwp *l = curlwp; /* XXX */
376 	struct vmspace *ovm;
377 
378 	KASSERT(p == l->l_proc);
379 	ovm = p->p_vmspace;
380 
381 	/*
382 	 * borrow proc0's address space.
383 	 */
384 	KPREEMPT_DISABLE(l);
385 	pmap_deactivate(l);
386 	p->p_vmspace = proc0.p_vmspace;
387 	pmap_activate(l);
388 	KPREEMPT_ENABLE(l);
389 
390 	uvmspace_free(ovm);
391 }
392 
393 void
394 uvm_lwp_exit(struct lwp *l)
395 {
396 	vaddr_t va = USER_TO_UAREA(l->l_addr);
397 
398 	l->l_flag &= ~LW_INMEM;
399 	uvm_uarea_free(va, l->l_cpu);
400 	l->l_addr = NULL;
401 }
402 
403 /*
404  * uvm_init_limit: init per-process VM limits
405  *
406  * - called for process 0 and then inherited by all others.
407  */
408 
409 void
410 uvm_init_limits(struct proc *p)
411 {
412 
413 	/*
414 	 * Set up the initial limits on process VM.  Set the maximum
415 	 * resident set size to be all of (reasonably) available memory.
416 	 * This causes any single, large process to start random page
417 	 * replacement once it fills memory.
418 	 */
419 
420 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
421 	p->p_rlimit[RLIMIT_STACK].rlim_max = maxsmap;
422 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
423 	p->p_rlimit[RLIMIT_DATA].rlim_max = maxdmap;
424 	p->p_rlimit[RLIMIT_AS].rlim_cur = RLIM_INFINITY;
425 	p->p_rlimit[RLIMIT_AS].rlim_max = RLIM_INFINITY;
426 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free);
427 }
428 
429 #ifdef DEBUG
430 int	enableswap = 1;
431 int	swapdebug = 0;
432 #define	SDB_FOLLOW	1
433 #define SDB_SWAPIN	2
434 #define SDB_SWAPOUT	4
435 #endif
436 
437 /*
438  * uvm_swapin: swap in an lwp's u-area.
439  *
440  * - must be called with the LWP's swap lock held.
441  * - naturally, must not be called with l == curlwp
442  */
443 
444 void
445 uvm_swapin(struct lwp *l)
446 {
447 	int error;
448 
449 	KASSERT(mutex_owned(&l->l_swaplock));
450 	KASSERT(l != curlwp);
451 
452 	error = uarea_swapin(USER_TO_UAREA(l->l_addr));
453 	if (error) {
454 		panic("%s: rewiring stack failed: %d", __func__, error);
455 	}
456 
457 	/*
458 	 * Some architectures need to be notified when the user area has
459 	 * moved to new physical page(s) (e.g.  see mips/mips/vm_machdep.c).
460 	 */
461 	cpu_swapin(l);
462 	lwp_lock(l);
463 	if (l->l_stat == LSRUN)
464 		sched_enqueue(l, false);
465 	l->l_flag |= LW_INMEM;
466 	l->l_swtime = 0;
467 	lwp_unlock(l);
468 	++uvmexp.swapins;
469 }
470 
471 /*
472  * uvm_kick_scheduler: kick the scheduler into action if not running.
473  *
474  * - called when swapped out processes have been awoken.
475  */
476 
477 void
478 uvm_kick_scheduler(void)
479 {
480 
481 	if (uvm.swap_running == false)
482 		return;
483 
484 	mutex_enter(&uvm_scheduler_mutex);
485 	uvm.scheduler_kicked = true;
486 	cv_signal(&uvm.scheduler_cv);
487 	mutex_exit(&uvm_scheduler_mutex);
488 }
489 
490 /*
491  * uvm_scheduler: process zero main loop
492  *
493  * - attempt to swapin every swaped-out, runnable process in order of
494  *	priority.
495  * - if not enough memory, wake the pagedaemon and let it clear space.
496  */
497 
498 void
499 uvm_scheduler(void)
500 {
501 	struct lwp *l, *ll;
502 	int pri;
503 	int ppri;
504 
505 	l = curlwp;
506 	lwp_lock(l);
507 	l->l_priority = PRI_VM;
508 	l->l_class = SCHED_FIFO;
509 	lwp_unlock(l);
510 
511 	for (;;) {
512 #ifdef DEBUG
513 		mutex_enter(&uvm_scheduler_mutex);
514 		while (!enableswap)
515 			cv_wait(&uvm.scheduler_cv, &uvm_scheduler_mutex);
516 		mutex_exit(&uvm_scheduler_mutex);
517 #endif
518 		ll = NULL;		/* process to choose */
519 		ppri = INT_MIN;		/* its priority */
520 
521 		mutex_enter(proc_lock);
522 		LIST_FOREACH(l, &alllwp, l_list) {
523 			/* is it a runnable swapped out process? */
524 			if (l->l_stat == LSRUN && !(l->l_flag & LW_INMEM)) {
525 				pri = l->l_swtime + l->l_slptime -
526 				    (l->l_proc->p_nice - NZERO) * 8;
527 				if (pri > ppri) {   /* higher priority? */
528 					ll = l;
529 					ppri = pri;
530 				}
531 			}
532 		}
533 #ifdef DEBUG
534 		if (swapdebug & SDB_FOLLOW)
535 			printf("%s: running, procp %p pri %d\n", __func__, ll,
536 			    ppri);
537 #endif
538 		/*
539 		 * Nothing to do, back to sleep
540 		 */
541 		if ((l = ll) == NULL) {
542 			mutex_exit(proc_lock);
543 			mutex_enter(&uvm_scheduler_mutex);
544 			if (uvm.scheduler_kicked == false)
545 				cv_wait(&uvm.scheduler_cv,
546 				    &uvm_scheduler_mutex);
547 			uvm.scheduler_kicked = false;
548 			mutex_exit(&uvm_scheduler_mutex);
549 			continue;
550 		}
551 
552 		/*
553 		 * we have found swapped out process which we would like
554 		 * to bring back in.
555 		 *
556 		 * XXX: this part is really bogus cuz we could deadlock
557 		 * on memory despite our feeble check
558 		 */
559 		if (uvmexp.free > atop(USPACE)) {
560 #ifdef DEBUG
561 			if (swapdebug & SDB_SWAPIN)
562 				printf("swapin: pid %d(%s)@%p, pri %d "
563 				    "free %d\n", l->l_proc->p_pid,
564 				    l->l_proc->p_comm, l->l_addr, ppri,
565 				    uvmexp.free);
566 #endif
567 			mutex_enter(&l->l_swaplock);
568 			mutex_exit(proc_lock);
569 			uvm_swapin(l);
570 			mutex_exit(&l->l_swaplock);
571 			continue;
572 		} else {
573 			/*
574 			 * not enough memory, jab the pageout daemon and
575 			 * wait til the coast is clear
576 			 */
577 			mutex_exit(proc_lock);
578 #ifdef DEBUG
579 			if (swapdebug & SDB_FOLLOW)
580 				printf("%s: no room for pid %d(%s),"
581 				    " free %d\n", __func__, l->l_proc->p_pid,
582 				    l->l_proc->p_comm, uvmexp.free);
583 #endif
584 			uvm_wait("schedpwait");
585 #ifdef DEBUG
586 			if (swapdebug & SDB_FOLLOW)
587 				printf("%s: room again, free %d\n", __func__,
588 				    uvmexp.free);
589 #endif
590 		}
591 	}
592 }
593 
594 /*
595  * swappable: is LWP "l" swappable?
596  */
597 
598 static bool
599 swappable(struct lwp *l)
600 {
601 
602 	if ((l->l_flag & (LW_INMEM|LW_SYSTEM|LW_WEXIT)) != LW_INMEM)
603 		return false;
604 	if ((l->l_pflag & LP_RUNNING) != 0)
605 		return false;
606 	if (l->l_holdcnt != 0)
607 		return false;
608 	if (l->l_class != SCHED_OTHER)
609 		return false;
610 	if (l->l_syncobj == &rw_syncobj || l->l_syncobj == &mutex_syncobj)
611 		return false;
612 	if (l->l_proc->p_stat != SACTIVE && l->l_proc->p_stat != SSTOP)
613 		return false;
614 	return true;
615 }
616 
617 /*
618  * swapout_threads: find threads that can be swapped and unwire their
619  *	u-areas.
620  *
621  * - called by the pagedaemon
622  * - try and swap at least one processs
623  * - processes that are sleeping or stopped for maxslp or more seconds
624  *   are swapped... otherwise the longest-sleeping or stopped process
625  *   is swapped, otherwise the longest resident process...
626  */
627 
628 void
629 uvm_swapout_threads(void)
630 {
631 	struct lwp *l;
632 	struct lwp *outl, *outl2;
633 	int outpri, outpri2;
634 	int didswap = 0;
635 	extern int maxslp;
636 	bool gotit;
637 
638 	/* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
639 
640 #ifdef DEBUG
641 	if (!enableswap)
642 		return;
643 #endif
644 
645 	/*
646 	 * outl/outpri  : stop/sleep thread with largest sleeptime < maxslp
647 	 * outl2/outpri2: the longest resident thread (its swap time)
648 	 */
649 	outl = outl2 = NULL;
650 	outpri = outpri2 = 0;
651 
652  restart:
653 	mutex_enter(proc_lock);
654 	LIST_FOREACH(l, &alllwp, l_list) {
655 		KASSERT(l->l_proc != NULL);
656 		if (!mutex_tryenter(&l->l_swaplock))
657 			continue;
658 		if (!swappable(l)) {
659 			mutex_exit(&l->l_swaplock);
660 			continue;
661 		}
662 		switch (l->l_stat) {
663 		case LSONPROC:
664 			break;
665 
666 		case LSRUN:
667 			if (l->l_swtime > outpri2) {
668 				outl2 = l;
669 				outpri2 = l->l_swtime;
670 			}
671 			break;
672 
673 		case LSSLEEP:
674 		case LSSTOP:
675 			if (l->l_slptime >= maxslp) {
676 				mutex_exit(proc_lock);
677 				uvm_swapout(l);
678 				/*
679 				 * Locking in the wrong direction -
680 				 * try to prevent the LWP from exiting.
681 				 */
682 				gotit = mutex_tryenter(proc_lock);
683 				mutex_exit(&l->l_swaplock);
684 				didswap++;
685 				if (!gotit)
686 					goto restart;
687 				continue;
688 			} else if (l->l_slptime > outpri) {
689 				outl = l;
690 				outpri = l->l_slptime;
691 			}
692 			break;
693 		}
694 		mutex_exit(&l->l_swaplock);
695 	}
696 
697 	/*
698 	 * If we didn't get rid of any real duds, toss out the next most
699 	 * likely sleeping/stopped or running candidate.  We only do this
700 	 * if we are real low on memory since we don't gain much by doing
701 	 * it (USPACE bytes).
702 	 */
703 	if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
704 		if ((l = outl) == NULL)
705 			l = outl2;
706 #ifdef DEBUG
707 		if (swapdebug & SDB_SWAPOUT)
708 			printf("%s: no duds, try procp %p\n", __func__, l);
709 #endif
710 		if (l) {
711 			mutex_enter(&l->l_swaplock);
712 			mutex_exit(proc_lock);
713 			if (swappable(l))
714 				uvm_swapout(l);
715 			mutex_exit(&l->l_swaplock);
716 			return;
717 		}
718 	}
719 
720 	mutex_exit(proc_lock);
721 }
722 
723 /*
724  * uvm_swapout: swap out lwp "l"
725  *
726  * - currently "swapout" means "unwire U-area" and "pmap_collect()"
727  *   the pmap.
728  * - must be called with l->l_swaplock held.
729  * - XXXCDC: should deactivate all process' private anonymous memory
730  */
731 
732 static void
733 uvm_swapout(struct lwp *l)
734 {
735 	struct vm_map *map;
736 
737 	KASSERT(mutex_owned(&l->l_swaplock));
738 
739 #ifdef DEBUG
740 	if (swapdebug & SDB_SWAPOUT)
741 		printf("%s: lid %d.%d(%s)@%p, stat %x pri %d free %d\n",
742 		   __func__, l->l_proc->p_pid, l->l_lid, l->l_proc->p_comm,
743 		   l->l_addr, l->l_stat, l->l_slptime, uvmexp.free);
744 #endif
745 
746 	/*
747 	 * Mark it as (potentially) swapped out.
748 	 */
749 	lwp_lock(l);
750 	if (!swappable(l)) {
751 		KDASSERT(l->l_cpu != curcpu());
752 		lwp_unlock(l);
753 		return;
754 	}
755 	l->l_flag &= ~LW_INMEM;
756 	l->l_swtime = 0;
757 	if (l->l_stat == LSRUN)
758 		sched_dequeue(l);
759 	lwp_unlock(l);
760 	l->l_ru.ru_nswap++;
761 	++uvmexp.swapouts;
762 
763 	/*
764 	 * Do any machine-specific actions necessary before swapout.
765 	 * This can include saving floating point state, etc.
766 	 */
767 	cpu_swapout(l);
768 
769 	/*
770 	 * Unwire the to-be-swapped process's user struct and kernel stack.
771 	 */
772 	uarea_swapout(USER_TO_UAREA(l->l_addr));
773 	map = &l->l_proc->p_vmspace->vm_map;
774 	if (vm_map_lock_try(map)) {
775 		pmap_collect(vm_map_pmap(map));
776 		vm_map_unlock(map);
777 	}
778 }
779 
780 /*
781  * uvm_lwp_hold: prevent lwp "l" from being swapped out, and bring
782  * back into memory if it is currently swapped.
783  */
784 
785 void
786 uvm_lwp_hold(struct lwp *l)
787 {
788 
789 	if (l == curlwp) {
790 		atomic_inc_uint(&l->l_holdcnt);
791 	} else {
792 		mutex_enter(&l->l_swaplock);
793 		if (atomic_inc_uint_nv(&l->l_holdcnt) == 1 &&
794 		    (l->l_flag & LW_INMEM) == 0)
795 			uvm_swapin(l);
796 		mutex_exit(&l->l_swaplock);
797 	}
798 }
799 
800 /*
801  * uvm_lwp_rele: release a hold on lwp "l".  when the holdcount
802  * drops to zero, it's eligable to be swapped.
803  */
804 
805 void
806 uvm_lwp_rele(struct lwp *l)
807 {
808 
809 	KASSERT(l->l_holdcnt != 0);
810 
811 	atomic_dec_uint(&l->l_holdcnt);
812 }
813