xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision 481fca6e59249d8ffcf24fef7cfbe7b131bfb080)
1 /*	$NetBSD: uvm_vnode.c,v 1.35 2000/06/27 17:29:37 mrg Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993
6  *      The Regents of the University of California.
7  * Copyright (c) 1990 University of Utah.
8  *
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * the Systems Programming Group of the University of Utah Computer
13  * Science Department.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. All advertising materials mentioning features or use of this software
24  *    must display the following acknowledgement:
25  *      This product includes software developed by Charles D. Cranor,
26  *	Washington University, the University of California, Berkeley and
27  *	its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  *
44  *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
45  * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
46  */
47 
48 #include "fs_nfs.h"
49 #include "opt_uvmhist.h"
50 
51 /*
52  * uvm_vnode.c: the vnode pager.
53  */
54 
55 #include <sys/param.h>
56 #include <sys/systm.h>
57 #include <sys/proc.h>
58 #include <sys/malloc.h>
59 #include <sys/vnode.h>
60 #include <sys/disklabel.h>
61 #include <sys/ioctl.h>
62 #include <sys/fcntl.h>
63 #include <sys/conf.h>
64 
65 #include <miscfs/specfs/specdev.h>
66 
67 #include <uvm/uvm.h>
68 #include <uvm/uvm_vnode.h>
69 
70 /*
71  * private global data structure
72  *
73  * we keep a list of writeable active vnode-backed VM objects for sync op.
74  * we keep a simpleq of vnodes that are currently being sync'd.
75  */
76 
77 LIST_HEAD(uvn_list_struct, uvm_vnode);
78 static struct uvn_list_struct uvn_wlist;	/* writeable uvns */
79 static simple_lock_data_t uvn_wl_lock;		/* locks uvn_wlist */
80 
81 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
82 static struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
83 lock_data_t uvn_sync_lock;			/* locks sync operation */
84 
85 /*
86  * functions
87  */
88 
89 static int		   uvn_asyncget __P((struct uvm_object *, voff_t,
90 					    int));
91 static void		   uvn_cluster __P((struct uvm_object *, voff_t,
92 					   voff_t *, voff_t *));
93 static void                uvn_detach __P((struct uvm_object *));
94 static boolean_t           uvn_flush __P((struct uvm_object *, voff_t,
95 					 voff_t, int));
96 static int                 uvn_get __P((struct uvm_object *, voff_t,
97 					vm_page_t *, int *, int,
98 					vm_prot_t, int, int));
99 static void		   uvn_init __P((void));
100 static int		   uvn_io __P((struct uvm_vnode *, vm_page_t *,
101 				      int, int, int));
102 static int		   uvn_put __P((struct uvm_object *, vm_page_t *,
103 					int, boolean_t));
104 static void                uvn_reference __P((struct uvm_object *));
105 static boolean_t	   uvn_releasepg __P((struct vm_page *,
106 					      struct vm_page **));
107 
108 /*
109  * master pager structure
110  */
111 
112 struct uvm_pagerops uvm_vnodeops = {
113 	uvn_init,
114 	uvn_reference,
115 	uvn_detach,
116 	NULL,			/* no specialized fault routine required */
117 	uvn_flush,
118 	uvn_get,
119 	uvn_asyncget,
120 	uvn_put,
121 	uvn_cluster,
122 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
123 	NULL,		 /* AIO-DONE function (not until we have asyncio) */
124 	uvn_releasepg,
125 };
126 
127 /*
128  * the ops!
129  */
130 
131 /*
132  * uvn_init
133  *
134  * init pager private data structures.
135  */
136 
137 static void
138 uvn_init()
139 {
140 
141 	LIST_INIT(&uvn_wlist);
142 	simple_lock_init(&uvn_wl_lock);
143 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
144 	lockinit(&uvn_sync_lock, PVM, "uvnsync", 0, 0);
145 }
146 
147 /*
148  * uvn_attach
149  *
150  * attach a vnode structure to a VM object.  if the vnode is already
151  * attached, then just bump the reference count by one and return the
152  * VM object.   if not already attached, attach and return the new VM obj.
153  * the "accessprot" tells the max access the attaching thread wants to
154  * our pages.
155  *
156  * => caller must _not_ already be holding the lock on the uvm_object.
157  * => in fact, nothing should be locked so that we can sleep here.
158  * => note that uvm_object is first thing in vnode structure, so their
159  *    pointers are equiv.
160  */
161 
162 struct uvm_object *
163 uvn_attach(arg, accessprot)
164 	void *arg;
165 	vm_prot_t accessprot;
166 {
167 	struct vnode *vp = arg;
168 	struct uvm_vnode *uvn = &vp->v_uvm;
169 	struct vattr vattr;
170 	int oldflags, result;
171 	struct partinfo pi;
172 	u_quad_t used_vnode_size;
173 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
174 
175 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
176 
177 	used_vnode_size = (u_quad_t)0;	/* XXX gcc -Wuninitialized */
178 
179 	/*
180 	 * first get a lock on the uvn.
181 	 */
182 	simple_lock(&uvn->u_obj.vmobjlock);
183 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
184 		printf("uvn_attach: blocked at 0x%p flags 0x%x\n",
185 		    uvn, uvn->u_flags);
186 		uvn->u_flags |= UVM_VNODE_WANTED;
187 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
188 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
189 		    "uvn_attach", 0);
190 		simple_lock(&uvn->u_obj.vmobjlock);
191 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
192 	}
193 
194 	/*
195 	 * if we're mapping a BLK device, make sure it is a disk.
196 	 */
197 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
198 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
199 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
200 		return(NULL);
201 	}
202 
203 	/*
204 	 * now we have lock and uvn must not be in a blocked state.
205 	 * first check to see if it is already active, in which case
206 	 * we can bump the reference count, check to see if we need to
207 	 * add it to the writeable list, and then return.
208 	 */
209 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
210 
211 		/* regain VREF if we were persisting */
212 		if (uvn->u_obj.uo_refs == 0) {
213 			VREF(vp);
214 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
215 			    0,0,0,0);
216 		}
217 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
218 
219 		/* check for new writeable uvn */
220 		if ((accessprot & VM_PROT_WRITE) != 0 &&
221 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
222 			simple_lock(&uvn_wl_lock);
223 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
224 			simple_unlock(&uvn_wl_lock);
225 			/* we are now on wlist! */
226 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
227 		}
228 
229 		/* unlock and return */
230 		simple_unlock(&uvn->u_obj.vmobjlock);
231 		UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
232 		    0, 0, 0);
233 		return (&uvn->u_obj);
234 	}
235 
236 	/*
237 	 * need to call VOP_GETATTR() to get the attributes, but that could
238 	 * block (due to I/O), so we want to unlock the object before calling.
239 	 * however, we want to keep anyone else from playing with the object
240 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
241 	 * prevents anyone from attaching to the vnode until we are done with
242 	 * it.
243 	 */
244 	uvn->u_flags = UVM_VNODE_ALOCK;
245 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
246 		/* XXX: curproc? */
247 
248 	if (vp->v_type == VBLK) {
249 		/*
250 		 * We could implement this as a specfs getattr call, but:
251 		 *
252 		 *	(1) VOP_GETATTR() would get the file system
253 		 *	    vnode operation, not the specfs operation.
254 		 *
255 		 *	(2) All we want is the size, anyhow.
256 		 */
257 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
258 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
259 		if (result == 0) {
260 			/* XXX should remember blocksize */
261 			used_vnode_size = (u_quad_t)pi.disklab->d_secsize *
262 			    (u_quad_t)pi.part->p_size;
263 		}
264 	} else {
265 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
266 		if (result == 0)
267 			used_vnode_size = vattr.va_size;
268 	}
269 
270 	/* relock object */
271 	simple_lock(&uvn->u_obj.vmobjlock);
272 
273 	if (result != 0) {
274 		if (uvn->u_flags & UVM_VNODE_WANTED)
275 			wakeup(uvn);
276 		uvn->u_flags = 0;
277 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
278 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
279 		return(NULL);
280 	}
281 
282 	/*
283 	 * make sure that the newsize fits within a vaddr_t
284 	 * XXX: need to revise addressing data types
285 	 */
286 #ifdef DEBUG
287 	if (vp->v_type == VBLK)
288 		printf("used_vnode_size = %qu\n", (long long)used_vnode_size);
289 #endif
290 
291 	/*
292 	 * now set up the uvn.
293 	 */
294 	uvn->u_obj.pgops = &uvm_vnodeops;
295 	TAILQ_INIT(&uvn->u_obj.memq);
296 	uvn->u_obj.uo_npages = 0;
297 	uvn->u_obj.uo_refs = 1;			/* just us... */
298 	oldflags = uvn->u_flags;
299 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
300 	uvn->u_nio = 0;
301 	uvn->u_size = used_vnode_size;
302 
303 	/* if write access, we need to add it to the wlist */
304 	if (accessprot & VM_PROT_WRITE) {
305 		simple_lock(&uvn_wl_lock);
306 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
307 		simple_unlock(&uvn_wl_lock);
308 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
309 	}
310 
311 	/*
312 	 * add a reference to the vnode.   this reference will stay as long
313 	 * as there is a valid mapping of the vnode.   dropped when the
314 	 * reference count goes to zero [and we either free or persist].
315 	 */
316 	VREF(vp);
317 	simple_unlock(&uvn->u_obj.vmobjlock);
318 	if (oldflags & UVM_VNODE_WANTED)
319 		wakeup(uvn);
320 
321 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
322 	return(&uvn->u_obj);
323 }
324 
325 
326 /*
327  * uvn_reference
328  *
329  * duplicate a reference to a VM object.  Note that the reference
330  * count must already be at least one (the passed in reference) so
331  * there is no chance of the uvn being killed or locked out here.
332  *
333  * => caller must call with object unlocked.
334  * => caller must be using the same accessprot as was used at attach time
335  */
336 
337 
338 static void
339 uvn_reference(uobj)
340 	struct uvm_object *uobj;
341 {
342 #ifdef DEBUG
343 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
344 #endif
345 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
346 
347 	simple_lock(&uobj->vmobjlock);
348 #ifdef DEBUG
349 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
350 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
351 		    uobj->uo_refs);
352 		panic("uvn_reference: invalid state");
353 	}
354 #endif
355 	uobj->uo_refs++;
356 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
357 	uobj, uobj->uo_refs,0,0);
358 	simple_unlock(&uobj->vmobjlock);
359 }
360 
361 /*
362  * uvn_detach
363  *
364  * remove a reference to a VM object.
365  *
366  * => caller must call with object unlocked and map locked.
367  * => this starts the detach process, but doesn't have to finish it
368  *    (async i/o could still be pending).
369  */
370 static void
371 uvn_detach(uobj)
372 	struct uvm_object *uobj;
373 {
374 	struct uvm_vnode *uvn;
375 	struct vnode *vp;
376 	int oldflags;
377 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
378 
379 	simple_lock(&uobj->vmobjlock);
380 
381 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
382 	uobj->uo_refs--;			/* drop ref! */
383 	if (uobj->uo_refs) {			/* still more refs */
384 		simple_unlock(&uobj->vmobjlock);
385 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
386 		return;
387 	}
388 
389 	/*
390 	 * get other pointers ...
391 	 */
392 
393 	uvn = (struct uvm_vnode *) uobj;
394 	vp = (struct vnode *) uobj;
395 
396 	/*
397 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
398 	 * to keep an active text file from being overwritten).
399 	 */
400 	vp->v_flag &= ~VTEXT;
401 
402 	/*
403 	 * we just dropped the last reference to the uvn.   see if we can
404 	 * let it "stick around".
405 	 */
406 
407 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
408 		/* won't block */
409 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
410 		simple_unlock(&uobj->vmobjlock);
411 		vrele(vp);			/* drop vnode reference */
412 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
413 		return;
414 	}
415 
416 	/*
417 	 * its a goner!
418 	 */
419 
420 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
421 
422 	uvn->u_flags |= UVM_VNODE_DYING;
423 
424 	/*
425 	 * even though we may unlock in flush, no one can gain a reference
426 	 * to us until we clear the "dying" flag [because it blocks
427 	 * attaches].  we will not do that until after we've disposed of all
428 	 * the pages with uvn_flush().  note that before the flush the only
429 	 * pages that could be marked PG_BUSY are ones that are in async
430 	 * pageout by the daemon.  (there can't be any pending "get"'s
431 	 * because there are no references to the object).
432 	 */
433 
434 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
435 
436 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
437 
438 	/*
439 	 * given the structure of this pager, the above flush request will
440 	 * create the following state: all the pages that were in the object
441 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
442 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
443 	 * in either case, if we have pages left we can't kill the object
444 	 * yet because i/o is pending.  in this case we set the "relkill"
445 	 * flag which will cause pgo_releasepg to kill the object once all
446 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
447 	 * routine or from the page daemon].
448 	 */
449 
450 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
451 #ifdef DEBUG
452 		/*
453 		 * XXXCDC: very unlikely to happen until we have async i/o
454 		 * so print a little info message in case it does.
455 		 */
456 		printf("uvn_detach: vn %p has pages left after flush - "
457 		    "relkill mode\n", uobj);
458 #endif
459 		uvn->u_flags |= UVM_VNODE_RELKILL;
460 		simple_unlock(&uobj->vmobjlock);
461 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
462 		    0, 0);
463 		return;
464 	}
465 
466 	/*
467 	 * kill object now.   note that we can't be on the sync q because
468 	 * all references are gone.
469 	 */
470 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
471 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
472 		LIST_REMOVE(uvn, u_wlist);
473 		simple_unlock(&uvn_wl_lock);
474 	}
475 #ifdef DIAGNOSTIC
476 	if (uobj->memq.tqh_first != NULL)
477 		panic("uvn_deref: vnode VM object still has pages afer "
478 		    "syncio/free flush");
479 #endif
480 	oldflags = uvn->u_flags;
481 	uvn->u_flags = 0;
482 	simple_unlock(&uobj->vmobjlock);
483 
484 	/* wake up any sleepers */
485 	if (oldflags & UVM_VNODE_WANTED)
486 		wakeup(uvn);
487 
488 	/*
489 	 * drop our reference to the vnode.
490 	 */
491 	vrele(vp);
492 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
493 
494 	return;
495 }
496 
497 /*
498  * uvm_vnp_terminate: external hook to clear out a vnode's VM
499  *
500  * called in two cases:
501  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
502  *      count) needs to be freed so that a vnode can be reused.  this
503  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
504  *      the free list is still attached (i.e. not VBAD) then vgone is
505  *	called.   as part of the vgone trace this should get called to
506  *	free the vm object.   this is the common case.
507  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
508  *	"umount -f") the vgone() function is called on active vnodes
509  *	on the mounted file systems to kill their data (the vnodes become
510  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
511  *	call here (even if the uvn is still in use -- i.e. has a non-zero
512  *	reference count).  this case happens at "umount -f" and during a
513  *	"reboot/halt" operation.
514  *
515  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
516  *	[protects us from getting a vnode that is already in the DYING
517  *	 state...]
518  * => unlike uvn_detach, this function must not return until all the
519  *	uvn's pages are disposed of.
520  * => in case [2] the uvn is still alive after this call, but all I/O
521  *	ops will fail (due to the backing vnode now being "dead").  this
522  *	will prob. kill any process using the uvn due to pgo_get failing.
523  */
524 
525 void
526 uvm_vnp_terminate(vp)
527 	struct vnode *vp;
528 {
529 	struct uvm_vnode *uvn = &vp->v_uvm;
530 	int oldflags;
531 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
532 
533 	/*
534 	 * lock object and check if it is valid
535 	 */
536 	simple_lock(&uvn->u_obj.vmobjlock);
537 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
538 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
539 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
540 		simple_unlock(&uvn->u_obj.vmobjlock);
541 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
542 		return;
543 	}
544 
545 	/*
546 	 * must be a valid uvn that is not already dying (because XLOCK
547 	 * protects us from that).   the uvn can't in the ALOCK state
548 	 * because it is valid, and uvn's that are in the ALOCK state haven't
549 	 * been marked valid yet.
550 	 */
551 
552 #ifdef DEBUG
553 	/*
554 	 * debug check: are we yanking the vnode out from under our uvn?
555 	 */
556 	if (uvn->u_obj.uo_refs) {
557 		printf("uvm_vnp_terminate(%p): terminating active vnode "
558 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
559 	}
560 #endif
561 
562 	/*
563 	 * it is possible that the uvn was detached and is in the relkill
564 	 * state [i.e. waiting for async i/o to finish so that releasepg can
565 	 * kill object].  we take over the vnode now and cancel the relkill.
566 	 * we want to know when the i/o is done so we can recycle right
567 	 * away.   note that a uvn can only be in the RELKILL state if it
568 	 * has a zero reference count.
569 	 */
570 
571 	if (uvn->u_flags & UVM_VNODE_RELKILL)
572 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
573 
574 	/*
575 	 * block the uvn by setting the dying flag, and then flush the
576 	 * pages.  (note that flush may unlock object while doing I/O, but
577 	 * it will re-lock it before it returns control here).
578 	 *
579 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
580 	 * that uvn_io doesn't attempt to VOP_LOCK again.
581 	 *
582 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
583 	 *	due to a forceful unmount might not be a good idea.  maybe we
584 	 *	need a way to pass in this info to uvn_flush through a
585 	 *	pager-defined PGO_ constant [currently there are none].
586 	 */
587 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
588 
589 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
590 
591 	/*
592 	 * as we just did a flush we expect all the pages to be gone or in
593 	 * the process of going.  sleep to wait for the rest to go [via iosync].
594 	 */
595 
596 	while (uvn->u_obj.uo_npages) {
597 #ifdef DEBUG
598 		struct vm_page *pp;
599 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
600 		     pp = pp->listq.tqe_next) {
601 			if ((pp->flags & PG_BUSY) == 0)
602 				panic("uvm_vnp_terminate: detected unbusy pg");
603 		}
604 		if (uvn->u_nio == 0)
605 			panic("uvm_vnp_terminate: no I/O to wait for?");
606 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
607 		/*
608 		 * XXXCDC: this is unlikely to happen without async i/o so we
609 		 * put a printf in just to keep an eye on it.
610 		 */
611 #endif
612 		uvn->u_flags |= UVM_VNODE_IOSYNC;
613 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
614 		    "uvn_term",0);
615 		simple_lock(&uvn->u_obj.vmobjlock);
616 	}
617 
618 	/*
619 	 * done.   now we free the uvn if its reference count is zero
620 	 * (true if we are zapping a persisting uvn).   however, if we are
621 	 * terminating a uvn with active mappings we let it live ... future
622 	 * calls down to the vnode layer will fail.
623 	 */
624 
625 	oldflags = uvn->u_flags;
626 	if (uvn->u_obj.uo_refs) {
627 
628 		/*
629 		 * uvn must live on it is dead-vnode state until all references
630 		 * are gone.   restore flags.    clear CANPERSIST state.
631 		 */
632 
633 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
634 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
635 
636 	} else {
637 
638 		/*
639 		 * free the uvn now.   note that the VREF reference is already
640 		 * gone [it is dropped when we enter the persist state].
641 		 */
642 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
643 			panic("uvm_vnp_terminate: io sync wanted bit set");
644 
645 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
646 			simple_lock(&uvn_wl_lock);
647 			LIST_REMOVE(uvn, u_wlist);
648 			simple_unlock(&uvn_wl_lock);
649 		}
650 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
651 	}
652 
653 	if (oldflags & UVM_VNODE_WANTED)
654 		wakeup(uvn);		/* object lock still held */
655 
656 	simple_unlock(&uvn->u_obj.vmobjlock);
657 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
658 
659 }
660 
661 /*
662  * uvn_releasepg: handled a released page in a uvn
663  *
664  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
665  *	to dispose of.
666  * => caller must handled PG_WANTED case
667  * => called with page's object locked, pageq's unlocked
668  * => returns TRUE if page's object is still alive, FALSE if we
669  *	killed the page's object.    if we return TRUE, then we
670  *	return with the object locked.
671  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
672  *				with the page queues locked [for pagedaemon]
673  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
674  * => we kill the uvn if it is not referenced and we are suppose to
675  *	kill it ("relkill").
676  */
677 
678 boolean_t
679 uvn_releasepg(pg, nextpgp)
680 	struct vm_page *pg;
681 	struct vm_page **nextpgp;	/* OUT */
682 {
683 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
684 #ifdef DIAGNOSTIC
685 	if ((pg->flags & PG_RELEASED) == 0)
686 		panic("uvn_releasepg: page not released!");
687 #endif
688 
689 	/*
690 	 * dispose of the page [caller handles PG_WANTED]
691 	 */
692 	pmap_page_protect(pg, VM_PROT_NONE);
693 	uvm_lock_pageq();
694 	if (nextpgp)
695 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
696 	uvm_pagefree(pg);
697 	if (!nextpgp)
698 		uvm_unlock_pageq();
699 
700 	/*
701 	 * now see if we need to kill the object
702 	 */
703 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
704 		if (uvn->u_obj.uo_refs)
705 			panic("uvn_releasepg: kill flag set on referenced "
706 			    "object!");
707 		if (uvn->u_obj.uo_npages == 0) {
708 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
709 				simple_lock(&uvn_wl_lock);
710 				LIST_REMOVE(uvn, u_wlist);
711 				simple_unlock(&uvn_wl_lock);
712 			}
713 #ifdef DIAGNOSTIC
714 			if (uvn->u_obj.memq.tqh_first)
715 	panic("uvn_releasepg: pages in object with npages == 0");
716 #endif
717 			if (uvn->u_flags & UVM_VNODE_WANTED)
718 				/* still holding object lock */
719 				wakeup(uvn);
720 
721 			uvn->u_flags = 0;		/* DEAD! */
722 			simple_unlock(&uvn->u_obj.vmobjlock);
723 			return (FALSE);
724 		}
725 	}
726 	return (TRUE);
727 }
728 
729 /*
730  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
731  * through the buffer cache and allow I/O in any size.  These VOPs use
732  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
733  * go through the buffer cache or allow I/O sizes larger than a
734  * block].  we will eventually want to change this.
735  *
736  * issues to consider:
737  *   uvm provides the uvm_aiodesc structure for async i/o management.
738  * there are two tailq's in the uvm. structure... one for pending async
739  * i/o and one for "done" async i/o.   to do an async i/o one puts
740  * an aiodesc on the "pending" list (protected by splbio()), starts the
741  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
742  * some sort of "i/o done" function to be called (at splbio(), interrupt
743  * time).   this function should remove the aiodesc from the pending list
744  * and place it on the "done" list and wakeup the daemon.   the daemon
745  * will run at normal spl() and will remove all items from the "done"
746  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
747  * [in the old vm code, this was done by calling the "put" routine with
748  * null arguments which made the code harder to read and understand because
749  * you had one function ("put") doing two things.]
750  *
751  * so the current pager needs:
752  *   int uvn_aiodone(struct uvm_aiodesc *)
753  *
754  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
755  *	later collection.
756  * => called with pageq's locked by the daemon.
757  *
758  * general outline:
759  * - "try" to lock object.   if fail, just return (will try again later)
760  * - drop "u_nio" (this req is done!)
761  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
762  * - get "page" structures (atop?).
763  * - handle "wanted" pages
764  * - handle "released" pages [using pgo_releasepg]
765  *   >>> pgo_releasepg may kill the object
766  * dont forget to look at "object" wanted flag in all cases.
767  */
768 
769 
770 /*
771  * uvn_flush: flush pages out of a uvm object.
772  *
773  * => object should be locked by caller.   we may _unlock_ the object
774  *	if (and only if) we need to clean a page (PGO_CLEANIT).
775  *	we return with the object locked.
776  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
777  *	might want to unlock higher level resources (e.g. vm_map)
778  *	before calling flush.
779  * => if PGO_CLEANIT is not set, then we will neither unlock the object
780  *	or block.
781  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
782  *	for flushing.
783  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
784  *	that new pages are inserted on the tail end of the list.   thus,
785  *	we can make a complete pass through the object in one go by starting
786  *	at the head and working towards the tail (new pages are put in
787  *	front of us).
788  * => NOTE: we are allowed to lock the page queues, so the caller
789  *	must not be holding the lock on them [e.g. pagedaemon had
790  *	better not call us with the queues locked]
791  * => we return TRUE unless we encountered some sort of I/O error
792  *
793  * comment on "cleaning" object and PG_BUSY pages:
794  *	this routine is holding the lock on the object.   the only time
795  *	that it can run into a PG_BUSY page that it does not own is if
796  *	some other process has started I/O on the page (e.g. either
797  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
798  *	in, then it can not be dirty (!PG_CLEAN) because no one has
799  *	had a chance to modify it yet.    if the PG_BUSY page is being
800  *	paged out then it means that someone else has already started
801  *	cleaning the page for us (how nice!).    in this case, if we
802  *	have syncio specified, then after we make our pass through the
803  *	object we need to wait for the other PG_BUSY pages to clear
804  *	off (i.e. we need to do an iosync).   also note that once a
805  *	page is PG_BUSY it must stay in its object until it is un-busyed.
806  *
807  * note on page traversal:
808  *	we can traverse the pages in an object either by going down the
809  *	linked list in "uobj->memq", or we can go over the address range
810  *	by page doing hash table lookups for each address.    depending
811  *	on how many pages are in the object it may be cheaper to do one
812  *	or the other.   we set "by_list" to true if we are using memq.
813  *	if the cost of a hash lookup was equal to the cost of the list
814  *	traversal we could compare the number of pages in the start->stop
815  *	range to the total number of pages in the object.   however, it
816  *	seems that a hash table lookup is more expensive than the linked
817  *	list traversal, so we multiply the number of pages in the
818  *	start->stop range by a penalty which we define below.
819  */
820 
821 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
822 
823 static boolean_t
824 uvn_flush(uobj, start, stop, flags)
825 	struct uvm_object *uobj;
826 	voff_t start, stop;
827 	int flags;
828 {
829 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
830 	struct vm_page *pp, *ppnext, *ptmp;
831 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
832 	int npages, result, lcv;
833 	boolean_t retval, need_iosync, by_list, needs_clean, all;
834 	voff_t curoff;
835 	u_short pp_version;
836 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
837 
838 	curoff = 0;	/* XXX: shut up gcc */
839 	/*
840 	 * get init vals and determine how we are going to traverse object
841 	 */
842 
843 	need_iosync = FALSE;
844 	retval = TRUE;		/* return value */
845 	if (flags & PGO_ALLPAGES) {
846 		all = TRUE;
847 		by_list = TRUE;		/* always go by the list */
848 	} else {
849 		start = trunc_page(start);
850 		stop = round_page(stop);
851 #ifdef DEBUG
852 		if (stop > round_page(uvn->u_size))
853 			printf("uvn_flush: strange, got an out of range "
854 			    "flush (fixed)\n");
855 #endif
856 		all = FALSE;
857 		by_list = (uobj->uo_npages <=
858 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
859 	}
860 
861 	UVMHIST_LOG(maphist,
862 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
863 	    start, stop, by_list, flags);
864 
865 	/*
866 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
867 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
868 	 * is wrong it will only prevent us from clustering... it won't break
869 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
870 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
871 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
872 	 * bit is always up to date since there are no mappings).
873 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
874 	 */
875 
876 	if ((flags & PGO_CLEANIT) != 0 &&
877 	    uobj->pgops->pgo_mk_pcluster != NULL) {
878 		if (by_list) {
879 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
880 			    pp = pp->listq.tqe_next) {
881 				if (!all &&
882 				    (pp->offset < start || pp->offset >= stop))
883 					continue;
884 				pp->flags &= ~PG_CLEANCHK;
885 			}
886 
887 		} else {   /* by hash */
888 			for (curoff = start ; curoff < stop;
889 			    curoff += PAGE_SIZE) {
890 				pp = uvm_pagelookup(uobj, curoff);
891 				if (pp)
892 					pp->flags &= ~PG_CLEANCHK;
893 			}
894 		}
895 	}
896 
897 	/*
898 	 * now do it.   note: we must update ppnext in body of loop or we
899 	 * will get stuck.  we need to use ppnext because we may free "pp"
900 	 * before doing the next loop.
901 	 */
902 
903 	if (by_list) {
904 		pp = uobj->memq.tqh_first;
905 	} else {
906 		curoff = start;
907 		pp = uvm_pagelookup(uobj, curoff);
908 	}
909 
910 	ppnext = NULL;	/* XXX: shut up gcc */
911 	ppsp = NULL;		/* XXX: shut up gcc */
912 	uvm_lock_pageq();	/* page queues locked */
913 
914 	/* locked: both page queues and uobj */
915 	for ( ; (by_list && pp != NULL) ||
916 	  (!by_list && curoff < stop) ; pp = ppnext) {
917 
918 		if (by_list) {
919 
920 			/*
921 			 * range check
922 			 */
923 
924 			if (!all &&
925 			    (pp->offset < start || pp->offset >= stop)) {
926 				ppnext = pp->listq.tqe_next;
927 				continue;
928 			}
929 
930 		} else {
931 
932 			/*
933 			 * null check
934 			 */
935 
936 			curoff += PAGE_SIZE;
937 			if (pp == NULL) {
938 				if (curoff < stop)
939 					ppnext = uvm_pagelookup(uobj, curoff);
940 				continue;
941 			}
942 
943 		}
944 
945 		/*
946 		 * handle case where we do not need to clean page (either
947 		 * because we are not clean or because page is not dirty or
948 		 * is busy):
949 		 *
950 		 * NOTE: we are allowed to deactivate a non-wired active
951 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
952 		 * queue it must stay put until it is !PG_BUSY (so as not to
953 		 * confuse pagedaemon).
954 		 */
955 
956 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
957 			needs_clean = FALSE;
958 			if ((pp->flags & PG_BUSY) != 0 &&
959 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
960 			             (PGO_CLEANIT|PGO_SYNCIO))
961 				need_iosync = TRUE;
962 		} else {
963 			/*
964 			 * freeing: nuke all mappings so we can sync
965 			 * PG_CLEAN bit with no race
966 			 */
967 			if ((pp->flags & PG_CLEAN) != 0 &&
968 			    (flags & PGO_FREE) != 0 &&
969 			    (pp->pqflags & PQ_ACTIVE) != 0)
970 				pmap_page_protect(pp, VM_PROT_NONE);
971 			if ((pp->flags & PG_CLEAN) != 0 &&
972 			    pmap_is_modified(pp))
973 				pp->flags &= ~(PG_CLEAN);
974 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
975 
976 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
977 		}
978 
979 		/*
980 		 * if we don't need a clean... load ppnext and dispose of pp
981 		 */
982 		if (!needs_clean) {
983 			/* load ppnext */
984 			if (by_list)
985 				ppnext = pp->listq.tqe_next;
986 			else {
987 				if (curoff < stop)
988 					ppnext = uvm_pagelookup(uobj, curoff);
989 			}
990 
991 			/* now dispose of pp */
992 			if (flags & PGO_DEACTIVATE) {
993 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
994 				    pp->wire_count == 0) {
995 					pmap_page_protect(pp, VM_PROT_NONE);
996 					uvm_pagedeactivate(pp);
997 				}
998 
999 			} else if (flags & PGO_FREE) {
1000 				if (pp->flags & PG_BUSY) {
1001 					/* release busy pages */
1002 					pp->flags |= PG_RELEASED;
1003 				} else {
1004 					pmap_page_protect(pp, VM_PROT_NONE);
1005 					/* removed page from object */
1006 					uvm_pagefree(pp);
1007 				}
1008 			}
1009 			/* ppnext is valid so we can continue... */
1010 			continue;
1011 		}
1012 
1013 		/*
1014 		 * pp points to a page in the locked object that we are
1015 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
1016 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1017 		 *
1018 		 * let uvm_pager_put attempted a clustered page out.
1019 		 * note: locked: uobj and page queues.
1020 		 */
1021 
1022 		pp->flags |= PG_BUSY;	/* we 'own' page now */
1023 		UVM_PAGE_OWN(pp, "uvn_flush");
1024 		pmap_page_protect(pp, VM_PROT_READ);
1025 		pp_version = pp->version;
1026 ReTry:
1027 		ppsp = pps;
1028 		npages = sizeof(pps) / sizeof(struct vm_page *);
1029 
1030 		/* locked: page queues, uobj */
1031 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1032 			   flags | PGO_DOACTCLUST, start, stop);
1033 		/* unlocked: page queues, uobj */
1034 
1035 		/*
1036 		 * at this point nothing is locked.   if we did an async I/O
1037 		 * it is remotely possible for the async i/o to complete and
1038 		 * the page "pp" be freed or what not before we get a chance
1039 		 * to relock the object.   in order to detect this, we have
1040 		 * saved the version number of the page in "pp_version".
1041 		 */
1042 
1043 		/* relock! */
1044 		simple_lock(&uobj->vmobjlock);
1045 		uvm_lock_pageq();
1046 
1047 		/*
1048 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1049 		 * can only happen when  we are doing async I/O and can't
1050 		 * map the pages into kernel memory (pager_map) due to lack
1051 		 * of vm space.   if this happens we drop back to sync I/O.
1052 		 */
1053 
1054 		if (result == VM_PAGER_AGAIN) {
1055 			/*
1056 			 * it is unlikely, but page could have been released
1057 			 * while we had the object lock dropped.   we ignore
1058 			 * this now and retry the I/O.  we will detect and
1059 			 * handle the released page after the syncio I/O
1060 			 * completes.
1061 			 */
1062 #ifdef DIAGNOSTIC
1063 			if (flags & PGO_SYNCIO)
1064 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1065 #endif
1066 			flags |= PGO_SYNCIO;
1067 			goto ReTry;
1068 		}
1069 
1070 		/*
1071 		 * the cleaning operation is now done.   finish up.  note that
1072 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1073 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1074 		 * to us in ppsp/npages.
1075 		 */
1076 
1077 		/*
1078 		 * for pending async i/o if we are not deactivating/freeing
1079 		 * we can move on to the next page.
1080 		 */
1081 
1082 		if (result == VM_PAGER_PEND) {
1083 
1084 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1085 				/*
1086 				 * no per-page ops: refresh ppnext and continue
1087 				 */
1088 				if (by_list) {
1089 					if (pp->version == pp_version)
1090 						ppnext = pp->listq.tqe_next;
1091 					else
1092 						/* reset */
1093 						ppnext = uobj->memq.tqh_first;
1094 				} else {
1095 					if (curoff < stop)
1096 						ppnext = uvm_pagelookup(uobj,
1097 						    curoff);
1098 				}
1099 				continue;
1100 			}
1101 
1102 			/* need to do anything here? */
1103 		}
1104 
1105 		/*
1106 		 * need to look at each page of the I/O operation.  we defer
1107 		 * processing "pp" until the last trip through this "for" loop
1108 		 * so that we can load "ppnext" for the main loop after we
1109 		 * play with the cluster pages [thus the "npages + 1" in the
1110 		 * loop below].
1111 		 */
1112 
1113 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1114 
1115 			/*
1116 			 * handle ppnext for outside loop, and saving pp
1117 			 * until the end.
1118 			 */
1119 			if (lcv < npages) {
1120 				if (ppsp[lcv] == pp)
1121 					continue; /* skip pp until the end */
1122 				ptmp = ppsp[lcv];
1123 			} else {
1124 				ptmp = pp;
1125 
1126 				/* set up next page for outer loop */
1127 				if (by_list) {
1128 					if (pp->version == pp_version)
1129 						ppnext = pp->listq.tqe_next;
1130 					else
1131 						/* reset */
1132 						ppnext = uobj->memq.tqh_first;
1133 				} else {
1134 					if (curoff < stop)
1135 					ppnext = uvm_pagelookup(uobj, curoff);
1136 				}
1137 			}
1138 
1139 			/*
1140 			 * verify the page didn't get moved while obj was
1141 			 * unlocked
1142 			 */
1143 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1144 				continue;
1145 
1146 			/*
1147 			 * unbusy the page if I/O is done.   note that for
1148 			 * pending I/O it is possible that the I/O op
1149 			 * finished before we relocked the object (in
1150 			 * which case the page is no longer busy).
1151 			 */
1152 
1153 			if (result != VM_PAGER_PEND) {
1154 				if (ptmp->flags & PG_WANTED)
1155 					/* still holding object lock */
1156 					wakeup(ptmp);
1157 
1158 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1159 				UVM_PAGE_OWN(ptmp, NULL);
1160 				if (ptmp->flags & PG_RELEASED) {
1161 
1162 					/* pgo_releasepg wants this */
1163 					uvm_unlock_pageq();
1164 					if (!uvn_releasepg(ptmp, NULL))
1165 						return (TRUE);
1166 
1167 					uvm_lock_pageq();	/* relock */
1168 					continue;		/* next page */
1169 
1170 				} else {
1171 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1172 					if ((flags & PGO_FREE) == 0)
1173 						pmap_clear_modify(ptmp);
1174 				}
1175 			}
1176 
1177 			/*
1178 			 * dispose of page
1179 			 */
1180 
1181 			if (flags & PGO_DEACTIVATE) {
1182 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1183 				    pp->wire_count == 0) {
1184 					pmap_page_protect(ptmp, VM_PROT_NONE);
1185 					uvm_pagedeactivate(ptmp);
1186 				}
1187 
1188 			} else if (flags & PGO_FREE) {
1189 				if (result == VM_PAGER_PEND) {
1190 					if ((ptmp->flags & PG_BUSY) != 0)
1191 						/* signal for i/o done */
1192 						ptmp->flags |= PG_RELEASED;
1193 				} else {
1194 					if (result != VM_PAGER_OK) {
1195 						printf("uvn_flush: obj=%p, "
1196 						   "offset=0x%llx.  error "
1197 						   "during pageout.\n",
1198 						    pp->uobject,
1199 						    (long long)pp->offset);
1200 						printf("uvn_flush: WARNING: "
1201 						    "changes to page may be "
1202 						    "lost!\n");
1203 						retval = FALSE;
1204 					}
1205 					pmap_page_protect(ptmp, VM_PROT_NONE);
1206 					uvm_pagefree(ptmp);
1207 				}
1208 			}
1209 
1210 		}		/* end of "lcv" for loop */
1211 
1212 	}		/* end of "pp" for loop */
1213 
1214 	/*
1215 	 * done with pagequeues: unlock
1216 	 */
1217 	uvm_unlock_pageq();
1218 
1219 	/*
1220 	 * now wait for all I/O if required.
1221 	 */
1222 	if (need_iosync) {
1223 
1224 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1225 		while (uvn->u_nio != 0) {
1226 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1227 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1228 			  FALSE, "uvn_flush",0);
1229 			simple_lock(&uvn->u_obj.vmobjlock);
1230 		}
1231 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1232 			wakeup(&uvn->u_flags);
1233 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1234 	}
1235 
1236 	/* return, with object locked! */
1237 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1238 	return(retval);
1239 }
1240 
1241 /*
1242  * uvn_cluster
1243  *
1244  * we are about to do I/O in an object at offset.   this function is called
1245  * to establish a range of offsets around "offset" in which we can cluster
1246  * I/O.
1247  *
1248  * - currently doesn't matter if obj locked or not.
1249  */
1250 
1251 static void
1252 uvn_cluster(uobj, offset, loffset, hoffset)
1253 	struct uvm_object *uobj;
1254 	voff_t offset;
1255 	voff_t *loffset, *hoffset; /* OUT */
1256 {
1257 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1258 	*loffset = offset;
1259 
1260 	if (*loffset >= uvn->u_size)
1261 		panic("uvn_cluster: offset out of range");
1262 
1263 	/*
1264 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1265 	 */
1266 	*hoffset = *loffset + MAXBSIZE;
1267 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1268 		*hoffset = round_page(uvn->u_size);
1269 
1270 	return;
1271 }
1272 
1273 /*
1274  * uvn_put: flush page data to backing store.
1275  *
1276  * => prefer map unlocked (not required)
1277  * => object must be locked!   we will _unlock_ it before starting I/O.
1278  * => flags: PGO_SYNCIO -- use sync. I/O
1279  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1280  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1281  *	[thus we never do async i/o!  see iodone comment]
1282  */
1283 
1284 static int
1285 uvn_put(uobj, pps, npages, flags)
1286 	struct uvm_object *uobj;
1287 	struct vm_page **pps;
1288 	int npages, flags;
1289 {
1290 	int retval;
1291 
1292 	/* note: object locked */
1293 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1294 	/* note: object unlocked */
1295 
1296 	return(retval);
1297 }
1298 
1299 
1300 /*
1301  * uvn_get: get pages (synchronously) from backing store
1302  *
1303  * => prefer map unlocked (not required)
1304  * => object must be locked!  we will _unlock_ it before starting any I/O.
1305  * => flags: PGO_ALLPAGES: get all of the pages
1306  *           PGO_LOCKED: fault data structures are locked
1307  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1308  * => NOTE: caller must check for released pages!!
1309  */
1310 
1311 static int
1312 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1313 	struct uvm_object *uobj;
1314 	voff_t offset;
1315 	struct vm_page **pps;		/* IN/OUT */
1316 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1317 	int centeridx, advice, flags;
1318 	vm_prot_t access_type;
1319 {
1320 	voff_t current_offset;
1321 	struct vm_page *ptmp;
1322 	int lcv, result, gotpages;
1323 	boolean_t done;
1324 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1325 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1326 
1327 	/*
1328 	 * step 1: handled the case where fault data structures are locked.
1329 	 */
1330 
1331 	if (flags & PGO_LOCKED) {
1332 
1333 		/*
1334 		 * gotpages is the current number of pages we've gotten (which
1335 		 * we pass back up to caller via *npagesp.
1336 		 */
1337 
1338 		gotpages = 0;
1339 
1340 		/*
1341 		 * step 1a: get pages that are already resident.   only do this
1342 		 * if the data structures are locked (i.e. the first time
1343 		 * through).
1344 		 */
1345 
1346 		done = TRUE;	/* be optimistic */
1347 
1348 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1349 		    lcv++, current_offset += PAGE_SIZE) {
1350 
1351 			/* do we care about this page?  if not, skip it */
1352 			if (pps[lcv] == PGO_DONTCARE)
1353 				continue;
1354 
1355 			/* lookup page */
1356 			ptmp = uvm_pagelookup(uobj, current_offset);
1357 
1358 			/* to be useful must get a non-busy, non-released pg */
1359 			if (ptmp == NULL ||
1360 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1361 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1362 				    != 0)
1363 				done = FALSE;	/* need to do a wait or I/O! */
1364 				continue;
1365 			}
1366 
1367 			/*
1368 			 * useful page: busy/lock it and plug it in our
1369 			 * result array
1370 			 */
1371 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1372 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1373 			pps[lcv] = ptmp;
1374 			gotpages++;
1375 
1376 		}	/* "for" lcv loop */
1377 
1378 		/*
1379 		 * XXX: given the "advice", should we consider async read-ahead?
1380 		 * XXX: fault current does deactive of pages behind us.  is
1381 		 * this good (other callers might now).
1382 		 */
1383 		/*
1384 		 * XXX: read-ahead currently handled by buffer cache (bread)
1385 		 * level.
1386 		 * XXX: no async i/o available.
1387 		 * XXX: so we don't do anything now.
1388 		 */
1389 
1390 		/*
1391 		 * step 1c: now we've either done everything needed or we to
1392 		 * unlock and do some waiting or I/O.
1393 		 */
1394 
1395 		*npagesp = gotpages;		/* let caller know */
1396 		if (done)
1397 			return(VM_PAGER_OK);		/* bingo! */
1398 		else
1399 			/* EEK!   Need to unlock and I/O */
1400 			return(VM_PAGER_UNLOCK);
1401 	}
1402 
1403 	/*
1404 	 * step 2: get non-resident or busy pages.
1405 	 * object is locked.   data structures are unlocked.
1406 	 *
1407 	 * XXX: because we can't do async I/O at this level we get things
1408 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1409 	 * async-read-ahead for us at a lower level.
1410 	 */
1411 
1412 	for (lcv = 0, current_offset = offset ;
1413 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1414 
1415 		/* skip over pages we've already gotten or don't want */
1416 		/* skip over pages we don't _have_ to get */
1417 		if (pps[lcv] != NULL || (lcv != centeridx &&
1418 		    (flags & PGO_ALLPAGES) == 0))
1419 			continue;
1420 
1421 		/*
1422 		 * we have yet to locate the current page (pps[lcv]).   we first
1423 		 * look for a page that is already at the current offset.   if
1424 		 * we fine a page, we check to see if it is busy or released.
1425 		 * if that is the case, then we sleep on the page until it is
1426 		 * no longer busy or released and repeat the lookup.    if the
1427 		 * page we found is neither busy nor released, then we busy it
1428 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1429 		 * following while loop and indicates we are ready to move on
1430 		 * to the next page in the "lcv" loop above.
1431 		 *
1432 		 * if we exit the while loop with pps[lcv] still set to NULL,
1433 		 * then it means that we allocated a new busy/fake/clean page
1434 		 * ptmp in the object and we need to do I/O to fill in the data.
1435 		 */
1436 
1437 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1438 
1439 			/* look for a current page */
1440 			ptmp = uvm_pagelookup(uobj, current_offset);
1441 
1442 			/* nope?   allocate one now (if we can) */
1443 			if (ptmp == NULL) {
1444 
1445 				ptmp = uvm_pagealloc(uobj, current_offset,
1446 				    NULL, 0);
1447 
1448 				/* out of RAM? */
1449 				if (ptmp == NULL) {
1450 					simple_unlock(&uobj->vmobjlock);
1451 					uvm_wait("uvn_getpage");
1452 					simple_lock(&uobj->vmobjlock);
1453 
1454 					/* goto top of pps while loop */
1455 					continue;
1456 				}
1457 
1458 				/*
1459 				 * got new page ready for I/O.  break pps
1460 				 * while loop.  pps[lcv] is still NULL.
1461 				 */
1462 				break;
1463 			}
1464 
1465 			/* page is there, see if we need to wait on it */
1466 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1467 				ptmp->flags |= PG_WANTED;
1468 				UVM_UNLOCK_AND_WAIT(ptmp,
1469 				    &uobj->vmobjlock, FALSE, "uvn_get",0);
1470 				simple_lock(&uobj->vmobjlock);
1471 				continue;	/* goto top of pps while loop */
1472 			}
1473 
1474 			/*
1475 			 * if we get here then the page has become resident
1476 			 * and unbusy between steps 1 and 2.  we busy it
1477 			 * now (so we own it) and set pps[lcv] (so that we
1478 			 * exit the while loop).
1479 			 */
1480 			ptmp->flags |= PG_BUSY;
1481 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1482 			pps[lcv] = ptmp;
1483 		}
1484 
1485 		/*
1486 		 * if we own the a valid page at the correct offset, pps[lcv]
1487 		 * will point to it.   nothing more to do except go to the
1488 		 * next page.
1489 		 */
1490 
1491 		if (pps[lcv])
1492 			continue;			/* next lcv */
1493 
1494 		/*
1495 		 * we have a "fake/busy/clean" page that we just allocated.  do
1496 		 * I/O to fill it with valid data.  note that object must be
1497 		 * locked going into uvn_io, but will be unlocked afterwards.
1498 		 */
1499 
1500 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1501 		    PGO_SYNCIO, UIO_READ);
1502 
1503 		/*
1504 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1505 		 * syncio the result can not be PEND or AGAIN.   we must relock
1506 		 * and check for errors.
1507 		 */
1508 
1509 		/* lock object.   check for errors.   */
1510 		simple_lock(&uobj->vmobjlock);
1511 		if (result != VM_PAGER_OK) {
1512 			if (ptmp->flags & PG_WANTED)
1513 				/* object lock still held */
1514 				wakeup(ptmp);
1515 
1516 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1517 			UVM_PAGE_OWN(ptmp, NULL);
1518 			uvm_lock_pageq();
1519 			uvm_pagefree(ptmp);
1520 			uvm_unlock_pageq();
1521 			simple_unlock(&uobj->vmobjlock);
1522 			return(result);
1523 		}
1524 
1525 		/*
1526 		 * we got the page!   clear the fake flag (indicates valid
1527 		 * data now in page) and plug into our result array.   note
1528 		 * that page is still busy.
1529 		 *
1530 		 * it is the callers job to:
1531 		 * => check if the page is released
1532 		 * => unbusy the page
1533 		 * => activate the page
1534 		 */
1535 
1536 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1537 		pmap_clear_modify(ptmp);		/* ... and clean */
1538 		pps[lcv] = ptmp;
1539 
1540 	}	/* lcv loop */
1541 
1542 	/*
1543 	 * finally, unlock object and return.
1544 	 */
1545 
1546 	simple_unlock(&uobj->vmobjlock);
1547 	return (VM_PAGER_OK);
1548 }
1549 
1550 /*
1551  * uvn_asyncget: start async I/O to bring pages into ram
1552  *
1553  * => caller must lock object(???XXX: see if this is best)
1554  * => could be called from uvn_get or a madvise() fault-ahead.
1555  * => if it fails, it doesn't matter.
1556  */
1557 
1558 static int
1559 uvn_asyncget(uobj, offset, npages)
1560 	struct uvm_object *uobj;
1561 	voff_t offset;
1562 	int npages;
1563 {
1564 
1565 	/*
1566 	 * XXXCDC: we can't do async I/O yet
1567 	 */
1568 	printf("uvn_asyncget called\n");
1569 	return (KERN_SUCCESS);
1570 }
1571 
1572 /*
1573  * uvn_io: do I/O to a vnode
1574  *
1575  * => prefer map unlocked (not required)
1576  * => object must be locked!   we will _unlock_ it before starting I/O.
1577  * => flags: PGO_SYNCIO -- use sync. I/O
1578  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1579  *	[thus we never do async i/o!  see iodone comment]
1580  */
1581 
1582 static int
1583 uvn_io(uvn, pps, npages, flags, rw)
1584 	struct uvm_vnode *uvn;
1585 	vm_page_t *pps;
1586 	int npages, flags, rw;
1587 {
1588 	struct vnode *vn;
1589 	struct uio uio;
1590 	struct iovec iov;
1591 	vaddr_t kva;
1592 	off_t file_offset;
1593 	int waitf, result, mapinflags;
1594 	size_t got, wanted;
1595 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1596 
1597 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1598 
1599 	/*
1600 	 * init values
1601 	 */
1602 
1603 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1604 	vn = (struct vnode *) uvn;
1605 	file_offset = pps[0]->offset;
1606 
1607 	/*
1608 	 * check for sync'ing I/O.
1609 	 */
1610 
1611 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1612 		if (waitf == M_NOWAIT) {
1613 			simple_unlock(&uvn->u_obj.vmobjlock);
1614 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1615 			return(VM_PAGER_AGAIN);
1616 		}
1617 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1618 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1619 			FALSE, "uvn_iosync",0);
1620 		simple_lock(&uvn->u_obj.vmobjlock);
1621 	}
1622 
1623 	/*
1624 	 * check size
1625 	 */
1626 
1627 	if (file_offset >= uvn->u_size) {
1628 			simple_unlock(&uvn->u_obj.vmobjlock);
1629 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1630 			return(VM_PAGER_BAD);
1631 	}
1632 
1633 	/*
1634 	 * first try and map the pages in (without waiting)
1635 	 */
1636 
1637 	mapinflags = (rw == UIO_READ) ?
1638 	    UVMPAGER_MAPIN_READ : UVMPAGER_MAPIN_WRITE;
1639 
1640 	kva = uvm_pagermapin(pps, npages, NULL, mapinflags);
1641 	if (kva == 0 && waitf == M_NOWAIT) {
1642 		simple_unlock(&uvn->u_obj.vmobjlock);
1643 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1644 		return(VM_PAGER_AGAIN);
1645 	}
1646 
1647 	/*
1648 	 * ok, now bump u_nio up.   at this point we are done with uvn
1649 	 * and can unlock it.   if we still don't have a kva, try again
1650 	 * (this time with sleep ok).
1651 	 */
1652 
1653 	uvn->u_nio++;			/* we have an I/O in progress! */
1654 	simple_unlock(&uvn->u_obj.vmobjlock);
1655 	/* NOTE: object now unlocked */
1656 	if (kva == 0)
1657 		kva = uvm_pagermapin(pps, npages, NULL,
1658 		    mapinflags | UVMPAGER_MAPIN_WAITOK);
1659 
1660 	/*
1661 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1662 	 * get touched (so we can look at "offset" without having to lock
1663 	 * the object).  set up for I/O.
1664 	 */
1665 
1666 	/*
1667 	 * fill out uio/iov
1668 	 */
1669 
1670 	iov.iov_base = (caddr_t) kva;
1671 	wanted = npages << PAGE_SHIFT;
1672 	if (file_offset + wanted > uvn->u_size)
1673 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1674 	iov.iov_len = wanted;
1675 	uio.uio_iov = &iov;
1676 	uio.uio_iovcnt = 1;
1677 	uio.uio_offset = file_offset;
1678 	uio.uio_segflg = UIO_SYSSPACE;
1679 	uio.uio_rw = rw;
1680 	uio.uio_resid = wanted;
1681 	uio.uio_procp = NULL;
1682 
1683 	/*
1684 	 * do the I/O!  (XXX: curproc?)
1685 	 */
1686 
1687 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1688 
1689 	/*
1690 	 * This process may already have this vnode locked, if we faulted in
1691 	 * copyin() or copyout() on a region backed by this vnode
1692 	 * while doing I/O to the vnode.  If this is the case, don't
1693 	 * panic.. instead, return the error to the user.
1694 	 *
1695 	 * XXX this is a stopgap to prevent a panic.
1696 	 * Ideally, this kind of operation *should* work.
1697 	 */
1698 	result = 0;
1699 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1700 		result = vn_lock(vn, LK_EXCLUSIVE | LK_RETRY | LK_RECURSEFAIL);
1701 
1702 	if (result == 0) {
1703 		/* NOTE: vnode now locked! */
1704 
1705 		if (rw == UIO_READ)
1706 			result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1707 		else
1708 			result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1709 
1710 		if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1711 			VOP_UNLOCK(vn, 0);
1712 	}
1713 
1714 	/* NOTE: vnode now unlocked (unless vnislocked) */
1715 
1716 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1717 
1718 	/*
1719 	 * result == unix style errno (0 == OK!)
1720 	 *
1721 	 * zero out rest of buffer (if needed)
1722 	 */
1723 
1724 	if (result == 0) {
1725 		got = wanted - uio.uio_resid;
1726 
1727 		if (wanted && got == 0) {
1728 			result = EIO;		/* XXX: error? */
1729 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1730 			memset((void *) (kva + got), 0,
1731 			       (npages << PAGE_SHIFT) - got);
1732 		}
1733 	}
1734 
1735 	/*
1736 	 * now remove pager mapping
1737 	 */
1738 	uvm_pagermapout(kva, npages);
1739 
1740 	/*
1741 	 * now clean up the object (i.e. drop I/O count)
1742 	 */
1743 
1744 	simple_lock(&uvn->u_obj.vmobjlock);
1745 	/* NOTE: object now locked! */
1746 
1747 	uvn->u_nio--;			/* I/O DONE! */
1748 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1749 		wakeup(&uvn->u_nio);
1750 	}
1751 	simple_unlock(&uvn->u_obj.vmobjlock);
1752 	/* NOTE: object now unlocked! */
1753 
1754 	/*
1755 	 * done!
1756 	 */
1757 
1758 	UVMHIST_LOG(maphist, "<- done (result %d)", result,0,0,0);
1759 	if (result == 0)
1760 		return(VM_PAGER_OK);
1761 	else
1762 		return(VM_PAGER_ERROR);
1763 }
1764 
1765 /*
1766  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1767  * is gone we will kill the object (flushing dirty pages back to the vnode
1768  * if needed).
1769  *
1770  * => returns TRUE if there was no uvm_object attached or if there was
1771  *	one and we killed it [i.e. if there is no active uvn]
1772  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1773  *	needed]
1774  *
1775  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1776  *	having to hold a reference on the vnode) and given a working
1777  *	uvm_vnp_sync(), how does that effect the need for this function?
1778  *      [XXXCDC: seems like it can die?]
1779  *
1780  * => XXX: this function should DIE once we merge the VM and buffer
1781  *	cache.
1782  *
1783  * research shows that this is called in the following places:
1784  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1785  *	changes sizes
1786  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1787  *	are written to
1788  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1789  *	is off
1790  * ffs_realloccg: when we can't extend the current block and have
1791  *	to allocate a new one we call this [XXX: why?]
1792  * nfsrv_rename, rename_files: called when the target filename is there
1793  *	and we want to remove it
1794  * nfsrv_remove, sys_unlink: called on file we are removing
1795  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1796  *	then return "text busy"
1797  * nfs_open: seems to uncache any file opened with nfs
1798  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1799  */
1800 
1801 boolean_t
1802 uvm_vnp_uncache(vp)
1803 	struct vnode *vp;
1804 {
1805 	struct uvm_vnode *uvn = &vp->v_uvm;
1806 
1807 	/*
1808 	 * lock uvn part of the vnode and check to see if we need to do anything
1809 	 */
1810 
1811 	simple_lock(&uvn->u_obj.vmobjlock);
1812 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1813 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1814 		simple_unlock(&uvn->u_obj.vmobjlock);
1815 		return(TRUE);
1816 	}
1817 
1818 	/*
1819 	 * we have a valid, non-blocked uvn.   clear persist flag.
1820 	 * if uvn is currently active we can return now.
1821 	 */
1822 
1823 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1824 	if (uvn->u_obj.uo_refs) {
1825 		simple_unlock(&uvn->u_obj.vmobjlock);
1826 		return(FALSE);
1827 	}
1828 
1829 	/*
1830 	 * uvn is currently persisting!   we have to gain a reference to
1831 	 * it so that we can call uvn_detach to kill the uvn.
1832 	 */
1833 
1834 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1835 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1836 	simple_unlock(&uvn->u_obj.vmobjlock);
1837 
1838 
1839 #ifdef DEBUG
1840 	/*
1841 	 * carry over sanity check from old vnode pager: the vnode should
1842 	 * be VOP_LOCK'd, and we confirm it here.
1843 	 */
1844 	if (!VOP_ISLOCKED(vp)) {
1845 		boolean_t is_ok_anyway = FALSE;
1846 #ifdef NFS
1847 		extern int (**nfsv2_vnodeop_p) __P((void *));
1848 		extern int (**spec_nfsv2nodeop_p) __P((void *));
1849 		extern int (**fifo_nfsv2nodeop_p) __P((void *));
1850 
1851 		/* vnode is NOT VOP_LOCKed: some vnode types _never_ lock */
1852 		if (vp->v_op == nfsv2_vnodeop_p ||
1853 		    vp->v_op == spec_nfsv2nodeop_p) {
1854 			is_ok_anyway = TRUE;
1855 		}
1856 		if (vp->v_op == fifo_nfsv2nodeop_p) {
1857 			is_ok_anyway = TRUE;
1858 		}
1859 #endif	/* NFS */
1860 		if (!is_ok_anyway)
1861 			panic("uvm_vnp_uncache: vnode not locked!");
1862 	}
1863 #endif	/* DEBUG */
1864 
1865 	/*
1866 	 * now drop our reference to the vnode.   if we have the sole
1867 	 * reference to the vnode then this will cause it to die [as we
1868 	 * just cleared the persist flag].   we have to unlock the vnode
1869 	 * while we are doing this as it may trigger I/O.
1870 	 *
1871 	 * XXX: it might be possible for uvn to get reclaimed while we are
1872 	 * unlocked causing us to return TRUE when we should not.   we ignore
1873 	 * this as a false-positive return value doesn't hurt us.
1874 	 */
1875 	VOP_UNLOCK(vp, 0);
1876 	uvn_detach(&uvn->u_obj);
1877 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1878 
1879 	/*
1880 	 * and return...
1881 	 */
1882 
1883 	return(TRUE);
1884 }
1885 
1886 /*
1887  * uvm_vnp_setsize: grow or shrink a vnode uvn
1888  *
1889  * grow   => just update size value
1890  * shrink => toss un-needed pages
1891  *
1892  * => we assume that the caller has a reference of some sort to the
1893  *	vnode in question so that it will not be yanked out from under
1894  *	us.
1895  *
1896  * called from:
1897  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1898  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1899  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1900  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1901  *  => union fs: union_newsize
1902  */
1903 
1904 void
1905 uvm_vnp_setsize(vp, newsize)
1906 	struct vnode *vp;
1907 	voff_t newsize;
1908 {
1909 	struct uvm_vnode *uvn = &vp->v_uvm;
1910 
1911 	/*
1912 	 * lock uvn and check for valid object, and if valid: do it!
1913 	 */
1914 	simple_lock(&uvn->u_obj.vmobjlock);
1915 	if (uvn->u_flags & UVM_VNODE_VALID) {
1916 
1917 		/*
1918 		 * now check if the size has changed: if we shrink we had better
1919 		 * toss some pages...
1920 		 */
1921 
1922 		if (uvn->u_size > newsize) {
1923 			(void)uvn_flush(&uvn->u_obj, newsize,
1924 			    uvn->u_size, PGO_FREE);
1925 		}
1926 		uvn->u_size = newsize;
1927 	}
1928 	simple_unlock(&uvn->u_obj.vmobjlock);
1929 
1930 	/*
1931 	 * done
1932 	 */
1933 	return;
1934 }
1935 
1936 /*
1937  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1938  *
1939  * => called from sys_sync with no VM structures locked
1940  * => only one process can do a sync at a time (because the uvn
1941  *    structure only has one queue for sync'ing).  we ensure this
1942  *    by holding the uvn_sync_lock while the sync is in progress.
1943  *    other processes attempting a sync will sleep on this lock
1944  *    until we are done.
1945  */
1946 
1947 void
1948 uvm_vnp_sync(mp)
1949 	struct mount *mp;
1950 {
1951 	struct uvm_vnode *uvn;
1952 	struct vnode *vp;
1953 	boolean_t got_lock;
1954 
1955 	/*
1956 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1957 	 * our lock...
1958 	 */
1959 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0);
1960 
1961 	/*
1962 	 * step 2: build up a simpleq of uvns of interest based on the
1963 	 * write list.   we gain a reference to uvns of interest.  must
1964 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1965 	 * in the body of the loop.
1966 	 */
1967 	SIMPLEQ_INIT(&uvn_sync_q);
1968 	simple_lock(&uvn_wl_lock);
1969 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1970 	    uvn = uvn->u_wlist.le_next) {
1971 
1972 		vp = (struct vnode *) uvn;
1973 		if (mp && vp->v_mount != mp)
1974 			continue;
1975 
1976 		/* attempt to gain reference */
1977 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1978 		    						FALSE &&
1979 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1980 			/* spin */ ;
1981 
1982 		/*
1983 		 * we will exit the loop if either if the following are true:
1984 		 *  - we got the lock [always true if NCPU == 1]
1985 		 *  - we failed to get the lock but noticed the vnode was
1986 		 * 	"blocked" -- in this case the vnode must be a dying
1987 		 *	vnode, and since dying vnodes are in the process of
1988 		 *	being flushed out, we can safely skip this one
1989 		 *
1990 		 * we want to skip over the vnode if we did not get the lock,
1991 		 * or if the vnode is already dying (due to the above logic).
1992 		 *
1993 		 * note that uvn must already be valid because we found it on
1994 		 * the wlist (this also means it can't be ALOCK'd).
1995 		 */
1996 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1997 			if (got_lock)
1998 				simple_unlock(&uvn->u_obj.vmobjlock);
1999 			continue;		/* skip it */
2000 		}
2001 
2002 		/*
2003 		 * gain reference.   watch out for persisting uvns (need to
2004 		 * regain vnode REF).
2005 		 */
2006 		if (uvn->u_obj.uo_refs == 0)
2007 			VREF(vp);
2008 		uvn->u_obj.uo_refs++;
2009 		simple_unlock(&uvn->u_obj.vmobjlock);
2010 
2011 		/*
2012 		 * got it!
2013 		 */
2014 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
2015 	}
2016 	simple_unlock(&uvn_wl_lock);
2017 
2018 	/*
2019 	 * step 3: we now have a list of uvn's that may need cleaning.
2020 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
2021 	 * (so we can now safely lock uvn's again).
2022 	 */
2023 
2024 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2025 		simple_lock(&uvn->u_obj.vmobjlock);
2026 #ifdef DEBUG
2027 		if (uvn->u_flags & UVM_VNODE_DYING) {
2028 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2029 		}
2030 #endif
2031 		uvn_flush(&uvn->u_obj, 0, 0,
2032 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2033 
2034 		/*
2035 		 * if we have the only reference and we just cleaned the uvn,
2036 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2037 		 * thus allowing us to avoid thinking about flushing it again
2038 		 * on later sync ops.
2039 		 */
2040 		if (uvn->u_obj.uo_refs == 1 &&
2041 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2042 			LIST_REMOVE(uvn, u_wlist);
2043 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2044 		}
2045 
2046 		simple_unlock(&uvn->u_obj.vmobjlock);
2047 
2048 		/* now drop our reference to the uvn */
2049 		uvn_detach(&uvn->u_obj);
2050 	}
2051 
2052 	/*
2053 	 * done!  release sync lock
2054 	 */
2055 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0);
2056 }
2057