Home | History | Annotate | Line # | Download | only in uvm
uvm_glue.c revision 1.71
      1 /*	$NetBSD: uvm_glue.c,v 1.71 2003/11/02 16:53:43 jdolecek 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 (at) 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 1.71 2003/11/02 16:53:43 jdolecek 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 
     87 #include <uvm/uvm.h>
     88 
     89 #include <machine/cpu.h>
     90 
     91 /*
     92  * local prototypes
     93  */
     94 
     95 static void uvm_swapout __P((struct lwp *));
     96 
     97 #define UVM_NUAREA_MAX 16
     98 void *uvm_uareas;
     99 int uvm_nuarea;
    100 struct simplelock uvm_uareas_slock = SIMPLELOCK_INITIALIZER;
    101 
    102 /*
    103  * XXXCDC: do these really belong here?
    104  */
    105 
    106 int readbuffers = 0;		/* allow KGDB to read kern buffer pool */
    107 				/* XXX: see uvm_kernacc */
    108 
    109 
    110 /*
    111  * uvm_kernacc: can the kernel access a region of memory
    112  *
    113  * - called from malloc [DIAGNOSTIC], and /dev/kmem driver (mem.c)
    114  */
    115 
    116 boolean_t
    117 uvm_kernacc(addr, len, rw)
    118 	caddr_t addr;
    119 	size_t len;
    120 	int rw;
    121 {
    122 	boolean_t rv;
    123 	vaddr_t saddr, eaddr;
    124 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
    125 
    126 	saddr = trunc_page((vaddr_t)addr);
    127 	eaddr = round_page((vaddr_t)addr + len);
    128 	vm_map_lock_read(kernel_map);
    129 	rv = uvm_map_checkprot(kernel_map, saddr, eaddr, prot);
    130 	vm_map_unlock_read(kernel_map);
    131 
    132 	/*
    133 	 * XXX there are still some things (e.g. the buffer cache) that
    134 	 * are managed behind the VM system's back so even though an
    135 	 * address is accessible in the mind of the VM system, there may
    136 	 * not be physical pages where the VM thinks there is.  This can
    137 	 * lead to bogus allocation of pages in the kernel address space
    138 	 * or worse, inconsistencies at the pmap level.  We only worry
    139 	 * about the buffer cache for now.
    140 	 */
    141 	if (!readbuffers && rv && (eaddr > (vaddr_t)buffers &&
    142 			     saddr < (vaddr_t)buffers + MAXBSIZE * nbuf))
    143 		rv = FALSE;
    144 	return(rv);
    145 }
    146 
    147 /*
    148  * uvm_useracc: can the user access it?
    149  *
    150  * - called from physio() and sys___sysctl().
    151  */
    152 
    153 boolean_t
    154 uvm_useracc(addr, len, rw)
    155 	caddr_t addr;
    156 	size_t len;
    157 	int rw;
    158 {
    159 	struct vm_map *map;
    160 	boolean_t rv;
    161 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
    162 
    163 	/* XXX curproc */
    164 	map = &curproc->p_vmspace->vm_map;
    165 
    166 	vm_map_lock_read(map);
    167 	rv = uvm_map_checkprot(map, trunc_page((vaddr_t)addr),
    168 	    round_page((vaddr_t)addr + len), prot);
    169 	vm_map_unlock_read(map);
    170 
    171 	return(rv);
    172 }
    173 
    174 #ifdef KGDB
    175 /*
    176  * Change protections on kernel pages from addr to addr+len
    177  * (presumably so debugger can plant a breakpoint).
    178  *
    179  * We force the protection change at the pmap level.  If we were
    180  * to use vm_map_protect a change to allow writing would be lazily-
    181  * applied meaning we would still take a protection fault, something
    182  * we really don't want to do.  It would also fragment the kernel
    183  * map unnecessarily.  We cannot use pmap_protect since it also won't
    184  * enforce a write-enable request.  Using pmap_enter is the only way
    185  * we can ensure the change takes place properly.
    186  */
    187 void
    188 uvm_chgkprot(addr, len, rw)
    189 	caddr_t addr;
    190 	size_t len;
    191 	int rw;
    192 {
    193 	vm_prot_t prot;
    194 	paddr_t pa;
    195 	vaddr_t sva, eva;
    196 
    197 	prot = rw == B_READ ? VM_PROT_READ : VM_PROT_READ|VM_PROT_WRITE;
    198 	eva = round_page((vaddr_t)addr + len);
    199 	for (sva = trunc_page((vaddr_t)addr); sva < eva; sva += PAGE_SIZE) {
    200 		/*
    201 		 * Extract physical address for the page.
    202 		 */
    203 		if (pmap_extract(pmap_kernel(), sva, &pa) == FALSE)
    204 			panic("chgkprot: invalid page");
    205 		pmap_enter(pmap_kernel(), sva, pa, prot, PMAP_WIRED);
    206 	}
    207 	pmap_update(pmap_kernel());
    208 }
    209 #endif
    210 
    211 /*
    212  * uvm_vslock: wire user memory for I/O
    213  *
    214  * - called from physio and sys___sysctl
    215  * - XXXCDC: consider nuking this (or making it a macro?)
    216  */
    217 
    218 int
    219 uvm_vslock(p, addr, len, access_type)
    220 	struct proc *p;
    221 	caddr_t	addr;
    222 	size_t	len;
    223 	vm_prot_t access_type;
    224 {
    225 	struct vm_map *map;
    226 	vaddr_t start, end;
    227 	int error;
    228 
    229 	map = &p->p_vmspace->vm_map;
    230 	start = trunc_page((vaddr_t)addr);
    231 	end = round_page((vaddr_t)addr + len);
    232 	error = uvm_fault_wire(map, start, end, VM_FAULT_WIRE, access_type);
    233 	return error;
    234 }
    235 
    236 /*
    237  * uvm_vsunlock: unwire user memory wired by uvm_vslock()
    238  *
    239  * - called from physio and sys___sysctl
    240  * - XXXCDC: consider nuking this (or making it a macro?)
    241  */
    242 
    243 void
    244 uvm_vsunlock(p, addr, len)
    245 	struct proc *p;
    246 	caddr_t	addr;
    247 	size_t	len;
    248 {
    249 	uvm_fault_unwire(&p->p_vmspace->vm_map, trunc_page((vaddr_t)addr),
    250 		round_page((vaddr_t)addr + len));
    251 }
    252 
    253 /*
    254  * uvm_proc_fork: fork a virtual address space
    255  *
    256  * - the address space is copied as per parent map's inherit values
    257  */
    258 void
    259 uvm_proc_fork(p1, p2, shared)
    260 	struct proc *p1, *p2;
    261 	boolean_t shared;
    262 {
    263 
    264 	if (shared == TRUE) {
    265 		p2->p_vmspace = NULL;
    266 		uvmspace_share(p1, p2);
    267 	} else {
    268 		p2->p_vmspace = uvmspace_fork(p1->p_vmspace);
    269 	}
    270 
    271 	cpu_proc_fork(p1, p2);
    272 }
    273 
    274 
    275 /*
    276  * uvm_lwp_fork: fork a thread
    277  *
    278  * - a new "user" structure is allocated for the child process
    279  *	[filled in by MD layer...]
    280  * - if specified, the child gets a new user stack described by
    281  *	stack and stacksize
    282  * - NOTE: the kernel stack may be at a different location in the child
    283  *	process, and thus addresses of automatic variables may be invalid
    284  *	after cpu_lwp_fork returns in the child process.  We do nothing here
    285  *	after cpu_lwp_fork returns.
    286  * - XXXCDC: we need a way for this to return a failure value rather
    287  *   than just hang
    288  */
    289 void
    290 uvm_lwp_fork(l1, l2, stack, stacksize, func, arg)
    291 	struct lwp *l1, *l2;
    292 	void *stack;
    293 	size_t stacksize;
    294 	void (*func) __P((void *));
    295 	void *arg;
    296 {
    297 	struct user *up = l2->l_addr;
    298 	int error;
    299 
    300 	/*
    301 	 * Wire down the U-area for the process, which contains the PCB
    302 	 * and the kernel stack.  Wired state is stored in l->l_flag's
    303 	 * L_INMEM bit rather than in the vm_map_entry's wired count
    304 	 * to prevent kernel_map fragmentation.  If we reused a cached U-area,
    305 	 * L_INMEM will already be set and we don't need to do anything.
    306 	 *
    307 	 * Note the kernel stack gets read/write accesses right off the bat.
    308 	 */
    309 
    310 	if ((l2->l_flag & L_INMEM) == 0) {
    311 		error = uvm_fault_wire(kernel_map, (vaddr_t)up,
    312 		    (vaddr_t)up + USPACE, VM_FAULT_WIRE,
    313 		    VM_PROT_READ | VM_PROT_WRITE);
    314 		if (error)
    315 			panic("uvm_lwp_fork: uvm_fault_wire failed: %d", error);
    316 #ifdef PMAP_UAREA
    317 		/* Tell the pmap this is a u-area mapping */
    318 		PMAP_UAREA((vaddr_t)up);
    319 #endif
    320 		l2->l_flag |= L_INMEM;
    321 	}
    322 
    323 #ifdef KSTACK_CHECK_MAGIC
    324 	/*
    325 	 * fill stack with magic number
    326 	 */
    327 	kstack_setup_magic(l2);
    328 #endif
    329 
    330 	/*
    331 	 * cpu_lwp_fork() copy and update the pcb, and make the child ready
    332  	 * to run.  If this is a normal user fork, the child will exit
    333 	 * directly to user mode via child_return() on its first time
    334 	 * slice and will not return here.  If this is a kernel thread,
    335 	 * the specified entry point will be executed.
    336 	 */
    337 	cpu_lwp_fork(l1, l2, stack, stacksize, func, arg);
    338 }
    339 
    340 /*
    341  * uvm_exit: exit a virtual address space
    342  *
    343  * - the process passed to us is a dead (pre-zombie) process; we
    344  *   are running on a different context now (the reaper).
    345  * - we must run in a separate thread because freeing the vmspace
    346  *   of the dead process may block.
    347  */
    348 
    349 void
    350 uvm_proc_exit(p)
    351 	struct proc *p;
    352 {
    353 	uvmspace_free(p->p_vmspace);
    354 }
    355 
    356 void
    357 uvm_lwp_exit(struct lwp *l)
    358 {
    359 	vaddr_t va = (vaddr_t)l->l_addr;
    360 
    361 	l->l_flag &= ~L_INMEM;
    362 	uvm_uarea_free(va);
    363 	l->l_addr = NULL;
    364 }
    365 
    366 /*
    367  * uvm_uarea_alloc: allocate a u-area
    368  */
    369 
    370 boolean_t
    371 uvm_uarea_alloc(vaddr_t *uaddrp)
    372 {
    373 	vaddr_t uaddr;
    374 
    375 #ifndef USPACE_ALIGN
    376 #define USPACE_ALIGN    0
    377 #endif
    378 
    379 	simple_lock(&uvm_uareas_slock);
    380 	uaddr = (vaddr_t)uvm_uareas;
    381 	if (uaddr) {
    382 		uvm_uareas = *(void **)uvm_uareas;
    383 		uvm_nuarea--;
    384 		simple_unlock(&uvm_uareas_slock);
    385 		*uaddrp = uaddr;
    386 		return TRUE;
    387 	} else {
    388 		simple_unlock(&uvm_uareas_slock);
    389 		*uaddrp = uvm_km_valloc_align(kernel_map, USPACE, USPACE_ALIGN);
    390 		return FALSE;
    391 	}
    392 }
    393 
    394 /*
    395  * uvm_uarea_free: free a u-area
    396  */
    397 
    398 void
    399 uvm_uarea_free(vaddr_t uaddr)
    400 {
    401 
    402 	simple_lock(&uvm_uareas_slock);
    403 	if (uvm_nuarea < UVM_NUAREA_MAX) {
    404 		*(void **)uaddr = uvm_uareas;
    405 		uvm_uareas = (void *)uaddr;
    406 		uvm_nuarea++;
    407 		simple_unlock(&uvm_uareas_slock);
    408 	} else {
    409 		simple_unlock(&uvm_uareas_slock);
    410 		uvm_km_free(kernel_map, uaddr, USPACE);
    411 	}
    412 }
    413 
    414 /*
    415  * uvm_init_limit: init per-process VM limits
    416  *
    417  * - called for process 0 and then inherited by all others.
    418  */
    419 
    420 void
    421 uvm_init_limits(p)
    422 	struct proc *p;
    423 {
    424 
    425 	/*
    426 	 * Set up the initial limits on process VM.  Set the maximum
    427 	 * resident set size to be all of (reasonably) available memory.
    428 	 * This causes any single, large process to start random page
    429 	 * replacement once it fills memory.
    430 	 */
    431 
    432 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
    433 	p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
    434 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
    435 	p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
    436 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(uvmexp.free);
    437 }
    438 
    439 #ifdef DEBUG
    440 int	enableswap = 1;
    441 int	swapdebug = 0;
    442 #define	SDB_FOLLOW	1
    443 #define SDB_SWAPIN	2
    444 #define SDB_SWAPOUT	4
    445 #endif
    446 
    447 /*
    448  * uvm_swapin: swap in a process's u-area.
    449  */
    450 
    451 void
    452 uvm_swapin(l)
    453 	struct lwp *l;
    454 {
    455 	vaddr_t addr;
    456 	int s, error;
    457 
    458 	addr = (vaddr_t)l->l_addr;
    459 	/* make L_INMEM true */
    460 	error = uvm_fault_wire(kernel_map, addr, addr + USPACE, VM_FAULT_WIRE,
    461 	    VM_PROT_READ | VM_PROT_WRITE);
    462 	if (error) {
    463 		panic("uvm_swapin: rewiring stack failed: %d", error);
    464 	}
    465 
    466 	/*
    467 	 * Some architectures need to be notified when the user area has
    468 	 * moved to new physical page(s) (e.g.  see mips/mips/vm_machdep.c).
    469 	 */
    470 	cpu_swapin(l);
    471 	SCHED_LOCK(s);
    472 	if (l->l_stat == LSRUN)
    473 		setrunqueue(l);
    474 	l->l_flag |= L_INMEM;
    475 	SCHED_UNLOCK(s);
    476 	l->l_swtime = 0;
    477 	++uvmexp.swapins;
    478 }
    479 
    480 /*
    481  * uvm_scheduler: process zero main loop
    482  *
    483  * - attempt to swapin every swaped-out, runnable process in order of
    484  *	priority.
    485  * - if not enough memory, wake the pagedaemon and let it clear space.
    486  */
    487 
    488 void
    489 uvm_scheduler()
    490 {
    491 	struct lwp *l, *ll;
    492 	int pri;
    493 	int ppri;
    494 
    495 loop:
    496 #ifdef DEBUG
    497 	while (!enableswap)
    498 		tsleep(&proc0, PVM, "noswap", 0);
    499 #endif
    500 	ll = NULL;		/* process to choose */
    501 	ppri = INT_MIN;	/* its priority */
    502 	proclist_lock_read();
    503 
    504 	LIST_FOREACH(l, &alllwp, l_list) {
    505 		/* is it a runnable swapped out process? */
    506 		if (l->l_stat == LSRUN && (l->l_flag & L_INMEM) == 0) {
    507 			pri = l->l_swtime + l->l_slptime -
    508 			    (l->l_proc->p_nice - NZERO) * 8;
    509 			if (pri > ppri) {   /* higher priority?  remember it. */
    510 				ll = l;
    511 				ppri = pri;
    512 			}
    513 		}
    514 	}
    515 	/*
    516 	 * XXXSMP: possible unlock/sleep race between here and the
    517 	 * "scheduler" tsleep below..
    518 	 */
    519 	proclist_unlock_read();
    520 
    521 #ifdef DEBUG
    522 	if (swapdebug & SDB_FOLLOW)
    523 		printf("scheduler: running, procp %p pri %d\n", ll, ppri);
    524 #endif
    525 	/*
    526 	 * Nothing to do, back to sleep
    527 	 */
    528 	if ((l = ll) == NULL) {
    529 		tsleep(&proc0, PVM, "scheduler", 0);
    530 		goto loop;
    531 	}
    532 
    533 	/*
    534 	 * we have found swapped out process which we would like to bring
    535 	 * back in.
    536 	 *
    537 	 * XXX: this part is really bogus cuz we could deadlock on memory
    538 	 * despite our feeble check
    539 	 */
    540 	if (uvmexp.free > atop(USPACE)) {
    541 #ifdef DEBUG
    542 		if (swapdebug & SDB_SWAPIN)
    543 			printf("swapin: pid %d(%s)@%p, pri %d free %d\n",
    544 	     l->l_proc->p_pid, l->l_proc->p_comm, l->l_addr, ppri, uvmexp.free);
    545 #endif
    546 		uvm_swapin(l);
    547 		goto loop;
    548 	}
    549 	/*
    550 	 * not enough memory, jab the pageout daemon and wait til the coast
    551 	 * is clear
    552 	 */
    553 #ifdef DEBUG
    554 	if (swapdebug & SDB_FOLLOW)
    555 		printf("scheduler: no room for pid %d(%s), free %d\n",
    556 	   l->l_proc->p_pid, l->l_proc->p_comm, uvmexp.free);
    557 #endif
    558 	uvm_wait("schedpwait");
    559 #ifdef DEBUG
    560 	if (swapdebug & SDB_FOLLOW)
    561 		printf("scheduler: room again, free %d\n", uvmexp.free);
    562 #endif
    563 	goto loop;
    564 }
    565 
    566 /*
    567  * swappable: is LWP "l" swappable?
    568  */
    569 
    570 #define	swappable(l)							\
    571 	(((l)->l_flag & (L_INMEM)) &&					\
    572 	 ((((l)->l_proc->p_flag) & (P_SYSTEM | P_WEXIT)) == 0) &&	\
    573 	 (l)->l_holdcnt == 0)
    574 
    575 /*
    576  * swapout_threads: find threads that can be swapped and unwire their
    577  *	u-areas.
    578  *
    579  * - called by the pagedaemon
    580  * - try and swap at least one processs
    581  * - processes that are sleeping or stopped for maxslp or more seconds
    582  *   are swapped... otherwise the longest-sleeping or stopped process
    583  *   is swapped, otherwise the longest resident process...
    584  */
    585 
    586 void
    587 uvm_swapout_threads()
    588 {
    589 	struct lwp *l;
    590 	struct lwp *outl, *outl2;
    591 	int outpri, outpri2;
    592 	int didswap = 0;
    593 	extern int maxslp;
    594 	/* XXXCDC: should move off to uvmexp. or uvm., also in uvm_meter */
    595 
    596 #ifdef DEBUG
    597 	if (!enableswap)
    598 		return;
    599 #endif
    600 
    601 	/*
    602 	 * outl/outpri  : stop/sleep thread with largest sleeptime < maxslp
    603 	 * outl2/outpri2: the longest resident thread (its swap time)
    604 	 */
    605 	outl = outl2 = NULL;
    606 	outpri = outpri2 = 0;
    607 	proclist_lock_read();
    608 	LIST_FOREACH(l, &alllwp, l_list) {
    609 		if (!swappable(l))
    610 			continue;
    611 		switch (l->l_stat) {
    612 		case LSONPROC:
    613 			KDASSERT(l->l_cpu != curcpu());
    614 			continue;
    615 
    616 		case LSRUN:
    617 			if (l->l_swtime > outpri2) {
    618 				outl2 = l;
    619 				outpri2 = l->l_swtime;
    620 			}
    621 			continue;
    622 
    623 		case LSSLEEP:
    624 		case LSSTOP:
    625 			if (l->l_slptime >= maxslp) {
    626 				uvm_swapout(l);
    627 				didswap++;
    628 			} else if (l->l_slptime > outpri) {
    629 				outl = l;
    630 				outpri = l->l_slptime;
    631 			}
    632 			continue;
    633 		}
    634 	}
    635 	proclist_unlock_read();
    636 
    637 	/*
    638 	 * If we didn't get rid of any real duds, toss out the next most
    639 	 * likely sleeping/stopped or running candidate.  We only do this
    640 	 * if we are real low on memory since we don't gain much by doing
    641 	 * it (USPACE bytes).
    642 	 */
    643 	if (didswap == 0 && uvmexp.free <= atop(round_page(USPACE))) {
    644 		if ((l = outl) == NULL)
    645 			l = outl2;
    646 #ifdef DEBUG
    647 		if (swapdebug & SDB_SWAPOUT)
    648 			printf("swapout_threads: no duds, try procp %p\n", l);
    649 #endif
    650 		if (l)
    651 			uvm_swapout(l);
    652 	}
    653 }
    654 
    655 /*
    656  * uvm_swapout: swap out lwp "l"
    657  *
    658  * - currently "swapout" means "unwire U-area" and "pmap_collect()"
    659  *   the pmap.
    660  * - XXXCDC: should deactivate all process' private anonymous memory
    661  */
    662 
    663 static void
    664 uvm_swapout(l)
    665 	struct lwp *l;
    666 {
    667 	vaddr_t addr;
    668 	int s;
    669 	struct proc *p = l->l_proc;
    670 
    671 #ifdef DEBUG
    672 	if (swapdebug & SDB_SWAPOUT)
    673 		printf("swapout: lid %d.%d(%s)@%p, stat %x pri %d free %d\n",
    674 	   p->p_pid, l->l_lid, p->p_comm, l->l_addr, l->l_stat,
    675 	   l->l_slptime, uvmexp.free);
    676 #endif
    677 
    678 	/*
    679 	 * Mark it as (potentially) swapped out.
    680 	 */
    681 	SCHED_LOCK(s);
    682 	if (l->l_stat == LSONPROC) {
    683 		KDASSERT(l->l_cpu != curcpu());
    684 		SCHED_UNLOCK(s);
    685 		return;
    686 	}
    687 	l->l_flag &= ~L_INMEM;
    688 	if (l->l_stat == LSRUN)
    689 		remrunqueue(l);
    690 	SCHED_UNLOCK(s);
    691 	l->l_swtime = 0;
    692 	p->p_stats->p_ru.ru_nswap++;
    693 	++uvmexp.swapouts;
    694 
    695 	/*
    696 	 * Do any machine-specific actions necessary before swapout.
    697 	 * This can include saving floating point state, etc.
    698 	 */
    699 	cpu_swapout(l);
    700 
    701 	/*
    702 	 * Unwire the to-be-swapped process's user struct and kernel stack.
    703 	 */
    704 	addr = (vaddr_t)l->l_addr;
    705 	uvm_fault_unwire(kernel_map, addr, addr + USPACE); /* !L_INMEM */
    706 	pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
    707 }
    708 
    709 /*
    710  * uvm_coredump_walkmap: walk a process's map for the purpose of dumping
    711  * a core file.
    712  */
    713 
    714 int
    715 uvm_coredump_walkmap(p, vp, cred, func, cookie)
    716 	struct proc *p;
    717 	struct vnode *vp;
    718 	struct ucred *cred;
    719 	int (*func)(struct proc *, struct vnode *, struct ucred *,
    720 	    struct uvm_coredump_state *);
    721 	void *cookie;
    722 {
    723 	struct uvm_coredump_state state;
    724 	struct vmspace *vm = p->p_vmspace;
    725 	struct vm_map *map = &vm->vm_map;
    726 	struct vm_map_entry *entry;
    727 	vaddr_t maxstack;
    728 	int error;
    729 
    730 	maxstack = trunc_page(USRSTACK - ctob(vm->vm_ssize));
    731 
    732 	entry = NULL;
    733 	vm_map_lock_read(map);
    734 	for (;;) {
    735 		if (entry == NULL)
    736 			entry = map->header.next;
    737 		else if (state.end < vm_map_min(map) ||
    738 		    vm_map_max(map) <= state.end)
    739 			break;
    740 		else if (!uvm_map_lookup_entry(map, state.end, &entry))
    741 			entry = entry->next;
    742 		if (entry == &map->header)
    743 			break;
    744 
    745 		/* Should never happen for a user process. */
    746 		if (UVM_ET_ISSUBMAP(entry))
    747 			panic("uvm_coredump_walkmap: user process with "
    748 			    "submap?");
    749 
    750 		state.cookie = cookie;
    751 		state.start = entry->start;
    752 		state.end = entry->end;
    753 		state.prot = entry->protection;
    754 		state.flags = 0;
    755 
    756 		if (state.start >= VM_MAXUSER_ADDRESS)
    757 			continue;
    758 
    759 		if (state.end > VM_MAXUSER_ADDRESS)
    760 			state.end = VM_MAXUSER_ADDRESS;
    761 
    762 		if (state.start >= (vaddr_t)vm->vm_maxsaddr) {
    763 			if (state.end <= maxstack)
    764 				continue;
    765 			if (state.start < maxstack)
    766 				state.start = maxstack;
    767 			state.flags |= UVM_COREDUMP_STACK;
    768 		}
    769 
    770 		if ((entry->protection & VM_PROT_WRITE) == 0)
    771 			state.flags |= UVM_COREDUMP_NODUMP;
    772 
    773 		if (entry->object.uvm_obj != NULL &&
    774 		    entry->object.uvm_obj->pgops == &uvm_deviceops)
    775 			state.flags |= UVM_COREDUMP_NODUMP;
    776 
    777 		vm_map_unlock_read(map);
    778 		error = (*func)(p, vp, cred, &state);
    779 		if (error)
    780 			return (error);
    781 		vm_map_lock_read(map);
    782 	}
    783 	vm_map_unlock_read(map);
    784 
    785 	return (0);
    786 }
    787