xref: /openbsd-src/sys/uvm/uvm_vnode.c (revision 850e275390052b330d93020bf619a739a3c277ac)
1 /*	$OpenBSD: uvm_vnode.c,v 1.51 2007/09/17 20:29:55 thib Exp $	*/
2 /*	$NetBSD: uvm_vnode.c,v 1.36 2000/11/24 20:34:01 chs Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * Copyright (c) 1991, 1993
7  *      The Regents of the University of California.
8  * Copyright (c) 1990 University of Utah.
9  *
10  * All rights reserved.
11  *
12  * This code is derived from software contributed to Berkeley by
13  * the Systems Programming Group of the University of Utah Computer
14  * Science Department.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. All advertising materials mentioning features or use of this software
25  *    must display the following acknowledgement:
26  *      This product includes software developed by Charles D. Cranor,
27  *	Washington University, the University of California, Berkeley and
28  *	its contributors.
29  * 4. Neither the name of the University nor the names of its contributors
30  *    may be used to endorse or promote products derived from this software
31  *    without specific prior written permission.
32  *
33  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
34  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
35  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
36  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
37  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
38  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
39  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
41  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
42  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43  * SUCH DAMAGE.
44  *
45  *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
46  * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
47  */
48 
49 /*
50  * uvm_vnode.c: the vnode pager.
51  */
52 
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/proc.h>
56 #include <sys/malloc.h>
57 #include <sys/vnode.h>
58 #include <sys/disklabel.h>
59 #include <sys/ioctl.h>
60 #include <sys/fcntl.h>
61 #include <sys/conf.h>
62 #include <sys/rwlock.h>
63 
64 #include <miscfs/specfs/specdev.h>
65 
66 #include <uvm/uvm.h>
67 #include <uvm/uvm_vnode.h>
68 
69 /*
70  * private global data structure
71  *
72  * we keep a list of writeable active vnode-backed VM objects for sync op.
73  * we keep a simpleq of vnodes that are currently being sync'd.
74  */
75 
76 LIST_HEAD(uvn_list_struct, uvm_vnode);
77 struct uvn_list_struct uvn_wlist;	/* writeable uvns */
78 
79 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
80 struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
81 struct rwlock uvn_sync_lock;			/* locks sync operation */
82 
83 /*
84  * functions
85  */
86 
87 void		   uvn_cluster(struct uvm_object *, voff_t,
88 					   voff_t *, voff_t *);
89 void                uvn_detach(struct uvm_object *);
90 boolean_t           uvn_flush(struct uvm_object *, voff_t,
91 					 voff_t, int);
92 int                 uvn_get(struct uvm_object *, voff_t,
93 					vm_page_t *, int *, int,
94 					vm_prot_t, int, int);
95 void		   uvn_init(void);
96 int		   uvn_io(struct uvm_vnode *, vm_page_t *,
97 				      int, int, int);
98 int		   uvn_put(struct uvm_object *, vm_page_t *,
99 					int, boolean_t);
100 void                uvn_reference(struct uvm_object *);
101 boolean_t	   uvn_releasepg(struct vm_page *,
102 					      struct vm_page **);
103 
104 /*
105  * master pager structure
106  */
107 
108 struct uvm_pagerops uvm_vnodeops = {
109 	uvn_init,
110 	uvn_reference,
111 	uvn_detach,
112 	NULL,			/* no specialized fault routine required */
113 	uvn_flush,
114 	uvn_get,
115 	uvn_put,
116 	uvn_cluster,
117 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
118 	uvn_releasepg,
119 };
120 
121 /*
122  * the ops!
123  */
124 
125 /*
126  * uvn_init
127  *
128  * init pager private data structures.
129  */
130 
131 void
132 uvn_init(void)
133 {
134 
135 	LIST_INIT(&uvn_wlist);
136 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
137 	rw_init(&uvn_sync_lock, "uvnsync");
138 }
139 
140 /*
141  * uvn_attach
142  *
143  * attach a vnode structure to a VM object.  if the vnode is already
144  * attached, then just bump the reference count by one and return the
145  * VM object.   if not already attached, attach and return the new VM obj.
146  * the "accessprot" tells the max access the attaching thread wants to
147  * our pages.
148  *
149  * => caller must _not_ already be holding the lock on the uvm_object.
150  * => in fact, nothing should be locked so that we can sleep here.
151  * => note that uvm_object is first thing in vnode structure, so their
152  *    pointers are equiv.
153  */
154 
155 struct uvm_object *
156 uvn_attach(arg, accessprot)
157 	void *arg;
158 	vm_prot_t accessprot;
159 {
160 	struct vnode *vp = arg;
161 	struct uvm_vnode *uvn = &vp->v_uvm;
162 	struct vattr vattr;
163 	int oldflags, result;
164 	struct partinfo pi;
165 	u_quad_t used_vnode_size;
166 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
167 
168 	UVMHIST_LOG(maphist, "(vn=%p)", arg,0,0,0);
169 
170 	used_vnode_size = (u_quad_t)0;	/* XXX gcc -Wuninitialized */
171 
172 	/*
173 	 * first get a lock on the uvn.
174 	 */
175 	simple_lock(&uvn->u_obj.vmobjlock);
176 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
177 		printf("uvn_attach: blocked at %p flags 0x%x\n",
178 		    uvn, uvn->u_flags);
179 		uvn->u_flags |= UVM_VNODE_WANTED;
180 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
181 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
182 		    "uvn_attach", 0);
183 		simple_lock(&uvn->u_obj.vmobjlock);
184 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
185 	}
186 
187 	/*
188 	 * if we're mapping a BLK device, make sure it is a disk.
189 	 */
190 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
191 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
192 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
193 		return(NULL);
194 	}
195 
196 	/*
197 	 * now we have lock and uvn must not be in a blocked state.
198 	 * first check to see if it is already active, in which case
199 	 * we can bump the reference count, check to see if we need to
200 	 * add it to the writeable list, and then return.
201 	 */
202 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
203 
204 		/* regain VREF if we were persisting */
205 		if (uvn->u_obj.uo_refs == 0) {
206 			VREF(vp);
207 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
208 			    0,0,0,0);
209 		}
210 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
211 
212 		/* check for new writeable uvn */
213 		if ((accessprot & VM_PROT_WRITE) != 0 &&
214 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
215 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
216 			/* we are now on wlist! */
217 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
218 		}
219 
220 		/* unlock and return */
221 		simple_unlock(&uvn->u_obj.vmobjlock);
222 		UVMHIST_LOG(maphist,"<- done, refcnt=%ld", uvn->u_obj.uo_refs,
223 		    0, 0, 0);
224 		return (&uvn->u_obj);
225 	}
226 
227 	/*
228 	 * need to call VOP_GETATTR() to get the attributes, but that could
229 	 * block (due to I/O), so we want to unlock the object before calling.
230 	 * however, we want to keep anyone else from playing with the object
231 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
232 	 * prevents anyone from attaching to the vnode until we are done with
233 	 * it.
234 	 */
235 	uvn->u_flags = UVM_VNODE_ALOCK;
236 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
237 		/* XXX: curproc? */
238 
239 	if (vp->v_type == VBLK) {
240 		/*
241 		 * We could implement this as a specfs getattr call, but:
242 		 *
243 		 *	(1) VOP_GETATTR() would get the file system
244 		 *	    vnode operation, not the specfs operation.
245 		 *
246 		 *	(2) All we want is the size, anyhow.
247 		 */
248 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
249 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
250 		if (result == 0) {
251 			/* XXX should remember blocksize */
252 			used_vnode_size = (u_quad_t)pi.disklab->d_secsize *
253 			    (u_quad_t)DL_GETPSIZE(pi.part);
254 		}
255 	} else {
256 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
257 		if (result == 0)
258 			used_vnode_size = vattr.va_size;
259 	}
260 
261 	/* relock object */
262 	simple_lock(&uvn->u_obj.vmobjlock);
263 
264 	if (result != 0) {
265 		if (uvn->u_flags & UVM_VNODE_WANTED)
266 			wakeup(uvn);
267 		uvn->u_flags = 0;
268 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
269 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
270 		return(NULL);
271 	}
272 
273 	/*
274 	 * make sure that the newsize fits within a vaddr_t
275 	 * XXX: need to revise addressing data types
276 	 */
277 #ifdef DEBUG
278 	if (vp->v_type == VBLK)
279 		printf("used_vnode_size = %llu\n", (long long)used_vnode_size);
280 #endif
281 
282 	/*
283 	 * now set up the uvn.
284 	 */
285 	uvn->u_obj.pgops = &uvm_vnodeops;
286 	TAILQ_INIT(&uvn->u_obj.memq);
287 	uvn->u_obj.uo_npages = 0;
288 	uvn->u_obj.uo_refs = 1;			/* just us... */
289 	oldflags = uvn->u_flags;
290 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
291 	uvn->u_nio = 0;
292 	uvn->u_size = used_vnode_size;
293 
294 	/* if write access, we need to add it to the wlist */
295 	if (accessprot & VM_PROT_WRITE) {
296 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
297 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
298 	}
299 
300 	/*
301 	 * add a reference to the vnode.   this reference will stay as long
302 	 * as there is a valid mapping of the vnode.   dropped when the
303 	 * reference count goes to zero [and we either free or persist].
304 	 */
305 	VREF(vp);
306 	simple_unlock(&uvn->u_obj.vmobjlock);
307 	if (oldflags & UVM_VNODE_WANTED)
308 		wakeup(uvn);
309 
310 	UVMHIST_LOG(maphist,"<- done/VREF, ret %p", &uvn->u_obj,0,0,0);
311 	return(&uvn->u_obj);
312 }
313 
314 
315 /*
316  * uvn_reference
317  *
318  * duplicate a reference to a VM object.  Note that the reference
319  * count must already be at least one (the passed in reference) so
320  * there is no chance of the uvn being killed or locked out here.
321  *
322  * => caller must call with object unlocked.
323  * => caller must be using the same accessprot as was used at attach time
324  */
325 
326 
327 void
328 uvn_reference(uobj)
329 	struct uvm_object *uobj;
330 {
331 #ifdef DEBUG
332 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
333 #endif
334 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
335 
336 	simple_lock(&uobj->vmobjlock);
337 #ifdef DEBUG
338 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
339 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
340 		    uobj->uo_refs);
341 		panic("uvn_reference: invalid state");
342 	}
343 #endif
344 	uobj->uo_refs++;
345 	UVMHIST_LOG(maphist, "<- done (uobj=%p, ref = %ld)",
346 	    uobj, uobj->uo_refs,0,0);
347 	simple_unlock(&uobj->vmobjlock);
348 }
349 
350 /*
351  * uvn_detach
352  *
353  * remove a reference to a VM object.
354  *
355  * => caller must call with object unlocked and map locked.
356  * => this starts the detach process, but doesn't have to finish it
357  *    (async i/o could still be pending).
358  */
359 void
360 uvn_detach(uobj)
361 	struct uvm_object *uobj;
362 {
363 	struct uvm_vnode *uvn;
364 	struct vnode *vp;
365 	int oldflags;
366 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
367 
368 	simple_lock(&uobj->vmobjlock);
369 
370 	UVMHIST_LOG(maphist,"  (uobj=%p)  ref=%ld", uobj,uobj->uo_refs,0,0);
371 	uobj->uo_refs--;			/* drop ref! */
372 	if (uobj->uo_refs) {			/* still more refs */
373 		simple_unlock(&uobj->vmobjlock);
374 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
375 		return;
376 	}
377 
378 	/*
379 	 * get other pointers ...
380 	 */
381 
382 	uvn = (struct uvm_vnode *) uobj;
383 	vp = (struct vnode *) uobj;
384 
385 	/*
386 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
387 	 * to keep an active text file from being overwritten).
388 	 */
389 	vp->v_flag &= ~VTEXT;
390 
391 	/*
392 	 * we just dropped the last reference to the uvn.   see if we can
393 	 * let it "stick around".
394 	 */
395 
396 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
397 		/* won't block */
398 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
399 		simple_unlock(&uobj->vmobjlock);
400 		vrele(vp);			/* drop vnode reference */
401 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
402 		return;
403 	}
404 
405 	/*
406 	 * its a goner!
407 	 */
408 
409 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
410 
411 	uvn->u_flags |= UVM_VNODE_DYING;
412 
413 	/*
414 	 * even though we may unlock in flush, no one can gain a reference
415 	 * to us until we clear the "dying" flag [because it blocks
416 	 * attaches].  we will not do that until after we've disposed of all
417 	 * the pages with uvn_flush().  note that before the flush the only
418 	 * pages that could be marked PG_BUSY are ones that are in async
419 	 * pageout by the daemon.  (there can't be any pending "get"'s
420 	 * because there are no references to the object).
421 	 */
422 
423 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
424 
425 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
426 
427 	/*
428 	 * given the structure of this pager, the above flush request will
429 	 * create the following state: all the pages that were in the object
430 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
431 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
432 	 * in either case, if we have pages left we can't kill the object
433 	 * yet because i/o is pending.  in this case we set the "relkill"
434 	 * flag which will cause pgo_releasepg to kill the object once all
435 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
436 	 * routine or from the page daemon].
437 	 */
438 
439 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
440 #ifdef DEBUG
441 		/*
442 		 * XXXCDC: very unlikely to happen until we have async i/o
443 		 * so print a little info message in case it does.
444 		 */
445 		printf("uvn_detach: vn %p has pages left after flush - "
446 		    "relkill mode\n", uobj);
447 #endif
448 		uvn->u_flags |= UVM_VNODE_RELKILL;
449 		simple_unlock(&uobj->vmobjlock);
450 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
451 		    0, 0);
452 		return;
453 	}
454 
455 	/*
456 	 * kill object now.   note that we can't be on the sync q because
457 	 * all references are gone.
458 	 */
459 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
460 		LIST_REMOVE(uvn, u_wlist);
461 	}
462 #ifdef DIAGNOSTIC
463 	if (!TAILQ_EMPTY(&uobj->memq))
464 		panic("uvn_deref: vnode VM object still has pages afer "
465 		    "syncio/free flush");
466 #endif
467 	oldflags = uvn->u_flags;
468 	uvn->u_flags = 0;
469 	simple_unlock(&uobj->vmobjlock);
470 
471 	/* wake up any sleepers */
472 	if (oldflags & UVM_VNODE_WANTED)
473 		wakeup(uvn);
474 
475 	/*
476 	 * drop our reference to the vnode.
477 	 */
478 	vrele(vp);
479 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
480 
481 	return;
482 }
483 
484 /*
485  * uvm_vnp_terminate: external hook to clear out a vnode's VM
486  *
487  * called in two cases:
488  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
489  *      count) needs to be freed so that a vnode can be reused.  this
490  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
491  *      the free list is still attached (i.e. not VBAD) then vgone is
492  *	called.   as part of the vgone trace this should get called to
493  *	free the vm object.   this is the common case.
494  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
495  *	"umount -f") the vgone() function is called on active vnodes
496  *	on the mounted file systems to kill their data (the vnodes become
497  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
498  *	call here (even if the uvn is still in use -- i.e. has a non-zero
499  *	reference count).  this case happens at "umount -f" and during a
500  *	"reboot/halt" operation.
501  *
502  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
503  *	[protects us from getting a vnode that is already in the DYING
504  *	 state...]
505  * => unlike uvn_detach, this function must not return until all the
506  *	uvn's pages are disposed of.
507  * => in case [2] the uvn is still alive after this call, but all I/O
508  *	ops will fail (due to the backing vnode now being "dead").  this
509  *	will prob. kill any process using the uvn due to pgo_get failing.
510  */
511 
512 void
513 uvm_vnp_terminate(vp)
514 	struct vnode *vp;
515 {
516 	struct uvm_vnode *uvn = &vp->v_uvm;
517 	int oldflags;
518 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
519 
520 	/*
521 	 * lock object and check if it is valid
522 	 */
523 	simple_lock(&uvn->u_obj.vmobjlock);
524 	UVMHIST_LOG(maphist, "  vp=%p, ref=%ld, flag=0x%lx", vp,
525 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
526 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
527 		simple_unlock(&uvn->u_obj.vmobjlock);
528 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
529 		return;
530 	}
531 
532 	/*
533 	 * must be a valid uvn that is not already dying (because XLOCK
534 	 * protects us from that).   the uvn can't in the ALOCK state
535 	 * because it is valid, and uvn's that are in the ALOCK state haven't
536 	 * been marked valid yet.
537 	 */
538 
539 #ifdef DEBUG
540 	/*
541 	 * debug check: are we yanking the vnode out from under our uvn?
542 	 */
543 	if (uvn->u_obj.uo_refs) {
544 		printf("uvm_vnp_terminate(%p): terminating active vnode "
545 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
546 	}
547 #endif
548 
549 	/*
550 	 * it is possible that the uvn was detached and is in the relkill
551 	 * state [i.e. waiting for async i/o to finish so that releasepg can
552 	 * kill object].  we take over the vnode now and cancel the relkill.
553 	 * we want to know when the i/o is done so we can recycle right
554 	 * away.   note that a uvn can only be in the RELKILL state if it
555 	 * has a zero reference count.
556 	 */
557 
558 	if (uvn->u_flags & UVM_VNODE_RELKILL)
559 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
560 
561 	/*
562 	 * block the uvn by setting the dying flag, and then flush the
563 	 * pages.  (note that flush may unlock object while doing I/O, but
564 	 * it will re-lock it before it returns control here).
565 	 *
566 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
567 	 * that uvn_io doesn't attempt to VOP_LOCK again.
568 	 *
569 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
570 	 *	due to a forceful unmount might not be a good idea.  maybe we
571 	 *	need a way to pass in this info to uvn_flush through a
572 	 *	pager-defined PGO_ constant [currently there are none].
573 	 */
574 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
575 
576 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
577 
578 	/*
579 	 * as we just did a flush we expect all the pages to be gone or in
580 	 * the process of going.  sleep to wait for the rest to go [via iosync].
581 	 */
582 
583 	while (uvn->u_obj.uo_npages) {
584 #ifdef DEBUG
585 		struct vm_page *pp;
586 		TAILQ_FOREACH(pp, &uvn->u_obj.memq, listq) {
587 			if ((pp->pg_flags & PG_BUSY) == 0)
588 				panic("uvm_vnp_terminate: detected unbusy pg");
589 		}
590 		if (uvn->u_nio == 0)
591 			panic("uvm_vnp_terminate: no I/O to wait for?");
592 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
593 		/*
594 		 * XXXCDC: this is unlikely to happen without async i/o so we
595 		 * put a printf in just to keep an eye on it.
596 		 */
597 #endif
598 		uvn->u_flags |= UVM_VNODE_IOSYNC;
599 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
600 		    "uvn_term",0);
601 		simple_lock(&uvn->u_obj.vmobjlock);
602 	}
603 
604 	/*
605 	 * done.   now we free the uvn if its reference count is zero
606 	 * (true if we are zapping a persisting uvn).   however, if we are
607 	 * terminating a uvn with active mappings we let it live ... future
608 	 * calls down to the vnode layer will fail.
609 	 */
610 
611 	oldflags = uvn->u_flags;
612 	if (uvn->u_obj.uo_refs) {
613 
614 		/*
615 		 * uvn must live on it is dead-vnode state until all references
616 		 * are gone.   restore flags.    clear CANPERSIST state.
617 		 */
618 
619 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
620 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
621 
622 	} else {
623 
624 		/*
625 		 * free the uvn now.   note that the VREF reference is already
626 		 * gone [it is dropped when we enter the persist state].
627 		 */
628 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
629 			panic("uvm_vnp_terminate: io sync wanted bit set");
630 
631 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
632 			LIST_REMOVE(uvn, u_wlist);
633 		}
634 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
635 	}
636 
637 	if (oldflags & UVM_VNODE_WANTED)
638 		wakeup(uvn);		/* object lock still held */
639 
640 	simple_unlock(&uvn->u_obj.vmobjlock);
641 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
642 
643 }
644 
645 /*
646  * uvn_releasepg: handled a released page in a uvn
647  *
648  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
649  *	to dispose of.
650  * => caller must handled PG_WANTED case
651  * => called with page's object locked, pageq's unlocked
652  * => returns TRUE if page's object is still alive, FALSE if we
653  *	killed the page's object.    if we return TRUE, then we
654  *	return with the object locked.
655  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
656  *				with the page queues locked [for pagedaemon]
657  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
658  * => we kill the uvn if it is not referenced and we are suppose to
659  *	kill it ("relkill").
660  */
661 
662 boolean_t
663 uvn_releasepg(pg, nextpgp)
664 	struct vm_page *pg;
665 	struct vm_page **nextpgp;	/* OUT */
666 {
667 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
668 #ifdef DIAGNOSTIC
669 	if ((pg->pg_flags & PG_RELEASED) == 0)
670 		panic("uvn_releasepg: page not released!");
671 #endif
672 
673 	/*
674 	 * dispose of the page [caller handles PG_WANTED]
675 	 */
676 	pmap_page_protect(pg, VM_PROT_NONE);
677 	uvm_lock_pageq();
678 	if (nextpgp)
679 		*nextpgp = TAILQ_NEXT(pg, pageq); /* next page for daemon */
680 	uvm_pagefree(pg);
681 	if (!nextpgp)
682 		uvm_unlock_pageq();
683 
684 	/*
685 	 * now see if we need to kill the object
686 	 */
687 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
688 		if (uvn->u_obj.uo_refs)
689 			panic("uvn_releasepg: kill flag set on referenced "
690 			    "object!");
691 		if (uvn->u_obj.uo_npages == 0) {
692 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
693 				LIST_REMOVE(uvn, u_wlist);
694 			}
695 #ifdef DIAGNOSTIC
696 			if (!TAILQ_EMPTY(&uvn->u_obj.memq))
697 	panic("uvn_releasepg: pages in object with npages == 0");
698 #endif
699 			if (uvn->u_flags & UVM_VNODE_WANTED)
700 				/* still holding object lock */
701 				wakeup(uvn);
702 
703 			uvn->u_flags = 0;		/* DEAD! */
704 			simple_unlock(&uvn->u_obj.vmobjlock);
705 			return (FALSE);
706 		}
707 	}
708 	return (TRUE);
709 }
710 
711 /*
712  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
713  * through the buffer cache and allow I/O in any size.  These VOPs use
714  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
715  * go through the buffer cache or allow I/O sizes larger than a
716  * block].  we will eventually want to change this.
717  *
718  * issues to consider:
719  *   uvm provides the uvm_aiodesc structure for async i/o management.
720  * there are two tailq's in the uvm. structure... one for pending async
721  * i/o and one for "done" async i/o.   to do an async i/o one puts
722  * an aiodesc on the "pending" list (protected by splbio()), starts the
723  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
724  * some sort of "i/o done" function to be called (at splbio(), interrupt
725  * time).   this function should remove the aiodesc from the pending list
726  * and place it on the "done" list and wakeup the daemon.   the daemon
727  * will run at normal spl() and will remove all items from the "done"
728  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
729  * [in the old vm code, this was done by calling the "put" routine with
730  * null arguments which made the code harder to read and understand because
731  * you had one function ("put") doing two things.]
732  *
733  * so the current pager needs:
734  *   int uvn_aiodone(struct uvm_aiodesc *)
735  *
736  * => return 0 (aio finished, free it). otherwise requeue for later collection.
737  * => called with pageq's locked by the daemon.
738  *
739  * general outline:
740  * - "try" to lock object.   if fail, just return (will try again later)
741  * - drop "u_nio" (this req is done!)
742  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
743  * - get "page" structures (atop?).
744  * - handle "wanted" pages
745  * - handle "released" pages [using pgo_releasepg]
746  *   >>> pgo_releasepg may kill the object
747  * dont forget to look at "object" wanted flag in all cases.
748  */
749 
750 
751 /*
752  * uvn_flush: flush pages out of a uvm object.
753  *
754  * => object should be locked by caller.   we may _unlock_ the object
755  *	if (and only if) we need to clean a page (PGO_CLEANIT).
756  *	we return with the object locked.
757  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
758  *	might want to unlock higher level resources (e.g. vm_map)
759  *	before calling flush.
760  * => if PGO_CLEANIT is not set, then we will neither unlock the object
761  *	or block.
762  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
763  *	for flushing.
764  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
765  *	that new pages are inserted on the tail end of the list.   thus,
766  *	we can make a complete pass through the object in one go by starting
767  *	at the head and working towards the tail (new pages are put in
768  *	front of us).
769  * => NOTE: we are allowed to lock the page queues, so the caller
770  *	must not be holding the lock on them [e.g. pagedaemon had
771  *	better not call us with the queues locked]
772  * => we return TRUE unless we encountered some sort of I/O error
773  *
774  * comment on "cleaning" object and PG_BUSY pages:
775  *	this routine is holding the lock on the object.   the only time
776  *	that it can run into a PG_BUSY page that it does not own is if
777  *	some other process has started I/O on the page (e.g. either
778  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
779  *	in, then it can not be dirty (!PG_CLEAN) because no one has
780  *	had a chance to modify it yet.    if the PG_BUSY page is being
781  *	paged out then it means that someone else has already started
782  *	cleaning the page for us (how nice!).    in this case, if we
783  *	have syncio specified, then after we make our pass through the
784  *	object we need to wait for the other PG_BUSY pages to clear
785  *	off (i.e. we need to do an iosync).   also note that once a
786  *	page is PG_BUSY it must stay in its object until it is un-busyed.
787  *
788  * note on page traversal:
789  *	we can traverse the pages in an object either by going down the
790  *	linked list in "uobj->memq", or we can go over the address range
791  *	by page doing hash table lookups for each address.    depending
792  *	on how many pages are in the object it may be cheaper to do one
793  *	or the other.   we set "by_list" to true if we are using memq.
794  *	if the cost of a hash lookup was equal to the cost of the list
795  *	traversal we could compare the number of pages in the start->stop
796  *	range to the total number of pages in the object.   however, it
797  *	seems that a hash table lookup is more expensive than the linked
798  *	list traversal, so we multiply the number of pages in the
799  *	start->stop range by a penalty which we define below.
800  */
801 
802 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
803 
804 boolean_t
805 uvn_flush(uobj, start, stop, flags)
806 	struct uvm_object *uobj;
807 	voff_t start, stop;
808 	int flags;
809 {
810 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
811 	struct vm_page *pp, *ppnext, *ptmp;
812 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
813 	int npages, result, lcv;
814 	boolean_t retval, need_iosync, by_list, needs_clean, all;
815 	voff_t curoff;
816 	u_short pp_version;
817 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
818 
819 	curoff = 0;	/* XXX: shut up gcc */
820 	/*
821 	 * get init vals and determine how we are going to traverse object
822 	 */
823 
824 	need_iosync = FALSE;
825 	retval = TRUE;		/* return value */
826 	if (flags & PGO_ALLPAGES) {
827 		all = TRUE;
828 		by_list = TRUE;		/* always go by the list */
829 	} else {
830 		start = trunc_page(start);
831 		stop = round_page(stop);
832 #ifdef DEBUG
833 		if (stop > round_page(uvn->u_size))
834 			printf("uvn_flush: strange, got an out of range "
835 			    "flush (fixed)\n");
836 #endif
837 		all = FALSE;
838 		by_list = (uobj->uo_npages <=
839 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
840 	}
841 
842 	UVMHIST_LOG(maphist,
843 	    " flush start=0x%lx, stop=0x%lx, by_list=%ld, flags=0x%lx",
844 	    (u_long)start, (u_long)stop, by_list, flags);
845 
846 	/*
847 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
848 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
849 	 * is wrong it will only prevent us from clustering... it won't break
850 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
851 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
852 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
853 	 * bit is always up to date since there are no mappings).
854 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
855 	 */
856 
857 	if ((flags & PGO_CLEANIT) != 0 &&
858 	    uobj->pgops->pgo_mk_pcluster != NULL) {
859 		if (by_list) {
860 			TAILQ_FOREACH(pp, &uobj->memq, listq) {
861 				if (!all &&
862 				    (pp->offset < start || pp->offset >= stop))
863 					continue;
864 				atomic_clearbits_int(&pp->pg_flags,
865 				    PG_CLEANCHK);
866 			}
867 
868 		} else {   /* by hash */
869 			for (curoff = start ; curoff < stop;
870 			    curoff += PAGE_SIZE) {
871 				pp = uvm_pagelookup(uobj, curoff);
872 				if (pp)
873 					atomic_clearbits_int(&pp->pg_flags,
874 					    PG_CLEANCHK);
875 			}
876 		}
877 	}
878 
879 	/*
880 	 * now do it.   note: we must update ppnext in body of loop or we
881 	 * will get stuck.  we need to use ppnext because we may free "pp"
882 	 * before doing the next loop.
883 	 */
884 
885 	if (by_list) {
886 		pp = TAILQ_FIRST(&uobj->memq);
887 	} else {
888 		curoff = start;
889 		pp = uvm_pagelookup(uobj, curoff);
890 	}
891 
892 	ppnext = NULL;	/* XXX: shut up gcc */
893 	ppsp = NULL;		/* XXX: shut up gcc */
894 	uvm_lock_pageq();	/* page queues locked */
895 
896 	/* locked: both page queues and uobj */
897 	for ( ; (by_list && pp != NULL) ||
898 	  (!by_list && curoff < stop) ; pp = ppnext) {
899 
900 		if (by_list) {
901 
902 			/*
903 			 * range check
904 			 */
905 
906 			if (!all &&
907 			    (pp->offset < start || pp->offset >= stop)) {
908 				ppnext = TAILQ_NEXT(pp, listq);
909 				continue;
910 			}
911 
912 		} else {
913 
914 			/*
915 			 * null check
916 			 */
917 
918 			curoff += PAGE_SIZE;
919 			if (pp == NULL) {
920 				if (curoff < stop)
921 					ppnext = uvm_pagelookup(uobj, curoff);
922 				continue;
923 			}
924 
925 		}
926 
927 		/*
928 		 * handle case where we do not need to clean page (either
929 		 * because we are not clean or because page is not dirty or
930 		 * is busy):
931 		 *
932 		 * NOTE: we are allowed to deactivate a non-wired active
933 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
934 		 * queue it must stay put until it is !PG_BUSY (so as not to
935 		 * confuse pagedaemon).
936 		 */
937 
938 		if ((flags & PGO_CLEANIT) == 0 || (pp->pg_flags & PG_BUSY) != 0) {
939 			needs_clean = FALSE;
940 			if ((pp->pg_flags & PG_BUSY) != 0 &&
941 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
942 			             (PGO_CLEANIT|PGO_SYNCIO))
943 				need_iosync = TRUE;
944 		} else {
945 			/*
946 			 * freeing: nuke all mappings so we can sync
947 			 * PG_CLEAN bit with no race
948 			 */
949 			if ((pp->pg_flags & PG_CLEAN) != 0 &&
950 			    (flags & PGO_FREE) != 0 &&
951 			    (pp->pg_flags & PQ_ACTIVE) != 0)
952 				pmap_page_protect(pp, VM_PROT_NONE);
953 			if ((pp->pg_flags & PG_CLEAN) != 0 &&
954 			    pmap_is_modified(pp))
955 				atomic_clearbits_int(&pp->pg_flags, PG_CLEAN);
956 			atomic_setbits_int(&pp->pg_flags, PG_CLEANCHK);
957 
958 			needs_clean = ((pp->pg_flags & PG_CLEAN) == 0);
959 		}
960 
961 		/*
962 		 * if we don't need a clean... load ppnext and dispose of pp
963 		 */
964 		if (!needs_clean) {
965 			/* load ppnext */
966 			if (by_list)
967 				ppnext = TAILQ_NEXT(pp, listq);
968 			else {
969 				if (curoff < stop)
970 					ppnext = uvm_pagelookup(uobj, curoff);
971 			}
972 
973 			/* now dispose of pp */
974 			if (flags & PGO_DEACTIVATE) {
975 				if ((pp->pg_flags & PQ_INACTIVE) == 0 &&
976 				    pp->wire_count == 0) {
977 					pmap_page_protect(pp, VM_PROT_NONE);
978 					uvm_pagedeactivate(pp);
979 				}
980 
981 			} else if (flags & PGO_FREE) {
982 				if (pp->pg_flags & PG_BUSY) {
983 					/* release busy pages */
984 					atomic_setbits_int(&pp->pg_flags,
985 					    PG_RELEASED);
986 				} else {
987 					pmap_page_protect(pp, VM_PROT_NONE);
988 					/* removed page from object */
989 					uvm_pagefree(pp);
990 				}
991 			}
992 			/* ppnext is valid so we can continue... */
993 			continue;
994 		}
995 
996 		/*
997 		 * pp points to a page in the locked object that we are
998 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
999 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1000 		 *
1001 		 * let uvm_pager_put attempted a clustered page out.
1002 		 * note: locked: uobj and page queues.
1003 		 */
1004 
1005 		atomic_setbits_int(&pp->pg_flags, PG_BUSY);
1006 		UVM_PAGE_OWN(pp, "uvn_flush");
1007 		pmap_page_protect(pp, VM_PROT_READ);
1008 		pp_version = pp->pg_version;
1009 ReTry:
1010 		ppsp = pps;
1011 		npages = sizeof(pps) / sizeof(struct vm_page *);
1012 
1013 		/* locked: page queues, uobj */
1014 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1015 			   flags | PGO_DOACTCLUST, start, stop);
1016 		/* unlocked: page queues, uobj */
1017 
1018 		/*
1019 		 * at this point nothing is locked.   if we did an async I/O
1020 		 * it is remotely possible for the async i/o to complete and
1021 		 * the page "pp" be freed or what not before we get a chance
1022 		 * to relock the object.   in order to detect this, we have
1023 		 * saved the version number of the page in "pp_version".
1024 		 */
1025 
1026 		/* relock! */
1027 		simple_lock(&uobj->vmobjlock);
1028 		uvm_lock_pageq();
1029 
1030 		/*
1031 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1032 		 * can only happen when  we are doing async I/O and can't
1033 		 * map the pages into kernel memory (pager_map) due to lack
1034 		 * of vm space.   if this happens we drop back to sync I/O.
1035 		 */
1036 
1037 		if (result == VM_PAGER_AGAIN) {
1038 			/*
1039 			 * it is unlikely, but page could have been released
1040 			 * while we had the object lock dropped.   we ignore
1041 			 * this now and retry the I/O.  we will detect and
1042 			 * handle the released page after the syncio I/O
1043 			 * completes.
1044 			 */
1045 #ifdef DIAGNOSTIC
1046 			if (flags & PGO_SYNCIO)
1047 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1048 #endif
1049 			flags |= PGO_SYNCIO;
1050 			goto ReTry;
1051 		}
1052 
1053 		/*
1054 		 * the cleaning operation is now done.   finish up.  note that
1055 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1056 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1057 		 * to us in ppsp/npages.
1058 		 */
1059 
1060 		/*
1061 		 * for pending async i/o if we are not deactivating/freeing
1062 		 * we can move on to the next page.
1063 		 */
1064 
1065 		if (result == VM_PAGER_PEND) {
1066 
1067 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1068 				/*
1069 				 * no per-page ops: refresh ppnext and continue
1070 				 */
1071 				if (by_list) {
1072 					if (pp->pg_version == pp_version)
1073 						ppnext = TAILQ_NEXT(pp, listq);
1074 					else
1075 						/* reset */
1076 						ppnext = TAILQ_FIRST(&uobj->memq);
1077 				} else {
1078 					if (curoff < stop)
1079 						ppnext = uvm_pagelookup(uobj,
1080 						    curoff);
1081 				}
1082 				continue;
1083 			}
1084 
1085 			/* need to do anything here? */
1086 		}
1087 
1088 		/*
1089 		 * need to look at each page of the I/O operation.  we defer
1090 		 * processing "pp" until the last trip through this "for" loop
1091 		 * so that we can load "ppnext" for the main loop after we
1092 		 * play with the cluster pages [thus the "npages + 1" in the
1093 		 * loop below].
1094 		 */
1095 
1096 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1097 
1098 			/*
1099 			 * handle ppnext for outside loop, and saving pp
1100 			 * until the end.
1101 			 */
1102 			if (lcv < npages) {
1103 				if (ppsp[lcv] == pp)
1104 					continue; /* skip pp until the end */
1105 				ptmp = ppsp[lcv];
1106 			} else {
1107 				ptmp = pp;
1108 
1109 				/* set up next page for outer loop */
1110 				if (by_list) {
1111 					if (pp->pg_version == pp_version)
1112 						ppnext = TAILQ_NEXT(pp, listq);
1113 					else
1114 						/* reset */
1115 						ppnext = TAILQ_FIRST(&uobj->memq);
1116 				} else {
1117 					if (curoff < stop)
1118 					ppnext = uvm_pagelookup(uobj, curoff);
1119 				}
1120 			}
1121 
1122 			/*
1123 			 * verify the page didn't get moved while obj was
1124 			 * unlocked
1125 			 */
1126 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1127 				continue;
1128 
1129 			/*
1130 			 * unbusy the page if I/O is done.   note that for
1131 			 * pending I/O it is possible that the I/O op
1132 			 * finished before we relocked the object (in
1133 			 * which case the page is no longer busy).
1134 			 */
1135 
1136 			if (result != VM_PAGER_PEND) {
1137 				if (ptmp->pg_flags & PG_WANTED)
1138 					/* still holding object lock */
1139 					wakeup(ptmp);
1140 
1141 				atomic_clearbits_int(&ptmp->pg_flags,
1142 				    PG_WANTED|PG_BUSY);
1143 				UVM_PAGE_OWN(ptmp, NULL);
1144 				if (ptmp->pg_flags & PG_RELEASED) {
1145 
1146 					/* pgo_releasepg wants this */
1147 					uvm_unlock_pageq();
1148 					if (!uvn_releasepg(ptmp, NULL))
1149 						return (TRUE);
1150 
1151 					uvm_lock_pageq();	/* relock */
1152 					continue;		/* next page */
1153 
1154 				} else {
1155 					atomic_setbits_int(&ptmp->pg_flags,
1156 					    PG_CLEAN|PG_CLEANCHK);
1157 					if ((flags & PGO_FREE) == 0)
1158 						pmap_clear_modify(ptmp);
1159 				}
1160 			}
1161 
1162 			/*
1163 			 * dispose of page
1164 			 */
1165 
1166 			if (flags & PGO_DEACTIVATE) {
1167 				if ((pp->pg_flags & PQ_INACTIVE) == 0 &&
1168 				    pp->wire_count == 0) {
1169 					pmap_page_protect(ptmp, VM_PROT_NONE);
1170 					uvm_pagedeactivate(ptmp);
1171 				}
1172 
1173 			} else if (flags & PGO_FREE) {
1174 				if (result == VM_PAGER_PEND) {
1175 					if ((ptmp->pg_flags & PG_BUSY) != 0)
1176 						/* signal for i/o done */
1177 						atomic_setbits_int(
1178 						    &ptmp->pg_flags,
1179 						    PG_RELEASED);
1180 				} else {
1181 					if (result != VM_PAGER_OK) {
1182 						printf("uvn_flush: obj=%p, "
1183 						   "offset=0x%llx.  error "
1184 						   "during pageout.\n",
1185 						    pp->uobject,
1186 						    (long long)pp->offset);
1187 						printf("uvn_flush: WARNING: "
1188 						    "changes to page may be "
1189 						    "lost!\n");
1190 						retval = FALSE;
1191 					}
1192 					pmap_page_protect(ptmp, VM_PROT_NONE);
1193 					uvm_pagefree(ptmp);
1194 				}
1195 			}
1196 
1197 		}		/* end of "lcv" for loop */
1198 
1199 	}		/* end of "pp" for loop */
1200 
1201 	/*
1202 	 * done with pagequeues: unlock
1203 	 */
1204 	uvm_unlock_pageq();
1205 
1206 	/*
1207 	 * now wait for all I/O if required.
1208 	 */
1209 	if (need_iosync) {
1210 
1211 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1212 		while (uvn->u_nio != 0) {
1213 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1214 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1215 			  FALSE, "uvn_flush",0);
1216 			simple_lock(&uvn->u_obj.vmobjlock);
1217 		}
1218 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1219 			wakeup(&uvn->u_flags);
1220 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1221 	}
1222 
1223 	/* return, with object locked! */
1224 	UVMHIST_LOG(maphist,"<- done (retval=0x%lx)",retval,0,0,0);
1225 	return(retval);
1226 }
1227 
1228 /*
1229  * uvn_cluster
1230  *
1231  * we are about to do I/O in an object at offset.   this function is called
1232  * to establish a range of offsets around "offset" in which we can cluster
1233  * I/O.
1234  *
1235  * - currently doesn't matter if obj locked or not.
1236  */
1237 
1238 void
1239 uvn_cluster(uobj, offset, loffset, hoffset)
1240 	struct uvm_object *uobj;
1241 	voff_t offset;
1242 	voff_t *loffset, *hoffset; /* OUT */
1243 {
1244 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1245 	*loffset = offset;
1246 
1247 	if (*loffset >= uvn->u_size)
1248 		panic("uvn_cluster: offset out of range");
1249 
1250 	/*
1251 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1252 	 */
1253 	*hoffset = *loffset + MAXBSIZE;
1254 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1255 		*hoffset = round_page(uvn->u_size);
1256 
1257 	return;
1258 }
1259 
1260 /*
1261  * uvn_put: flush page data to backing store.
1262  *
1263  * => prefer map unlocked (not required)
1264  * => object must be locked!   we will _unlock_ it before starting I/O.
1265  * => flags: PGO_SYNCIO -- use sync. I/O
1266  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1267  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1268  *	[thus we never do async i/o!  see iodone comment]
1269  */
1270 
1271 int
1272 uvn_put(uobj, pps, npages, flags)
1273 	struct uvm_object *uobj;
1274 	struct vm_page **pps;
1275 	int npages, flags;
1276 {
1277 	int retval;
1278 
1279 	/* note: object locked */
1280 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1281 	/* note: object unlocked */
1282 
1283 	return(retval);
1284 }
1285 
1286 
1287 /*
1288  * uvn_get: get pages (synchronously) from backing store
1289  *
1290  * => prefer map unlocked (not required)
1291  * => object must be locked!  we will _unlock_ it before starting any I/O.
1292  * => flags: PGO_ALLPAGES: get all of the pages
1293  *           PGO_LOCKED: fault data structures are locked
1294  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1295  * => NOTE: caller must check for released pages!!
1296  */
1297 
1298 int
1299 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1300 	struct uvm_object *uobj;
1301 	voff_t offset;
1302 	struct vm_page **pps;		/* IN/OUT */
1303 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1304 	int centeridx, advice, flags;
1305 	vm_prot_t access_type;
1306 {
1307 	voff_t current_offset;
1308 	struct vm_page *ptmp;
1309 	int lcv, result, gotpages;
1310 	boolean_t done;
1311 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1312 	UVMHIST_LOG(maphist, "flags=%ld", flags,0,0,0);
1313 
1314 	/*
1315 	 * step 1: handled the case where fault data structures are locked.
1316 	 */
1317 
1318 	if (flags & PGO_LOCKED) {
1319 
1320 		/*
1321 		 * gotpages is the current number of pages we've gotten (which
1322 		 * we pass back up to caller via *npagesp.
1323 		 */
1324 
1325 		gotpages = 0;
1326 
1327 		/*
1328 		 * step 1a: get pages that are already resident.   only do this
1329 		 * if the data structures are locked (i.e. the first time
1330 		 * through).
1331 		 */
1332 
1333 		done = TRUE;	/* be optimistic */
1334 
1335 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1336 		    lcv++, current_offset += PAGE_SIZE) {
1337 
1338 			/* do we care about this page?  if not, skip it */
1339 			if (pps[lcv] == PGO_DONTCARE)
1340 				continue;
1341 
1342 			/* lookup page */
1343 			ptmp = uvm_pagelookup(uobj, current_offset);
1344 
1345 			/* to be useful must get a non-busy, non-released pg */
1346 			if (ptmp == NULL ||
1347 			    (ptmp->pg_flags & (PG_BUSY|PG_RELEASED)) != 0) {
1348 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1349 				    != 0)
1350 				done = FALSE;	/* need to do a wait or I/O! */
1351 				continue;
1352 			}
1353 
1354 			/*
1355 			 * useful page: busy/lock it and plug it in our
1356 			 * result array
1357 			 */
1358 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1359 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1360 			pps[lcv] = ptmp;
1361 			gotpages++;
1362 
1363 		}	/* "for" lcv loop */
1364 
1365 		/*
1366 		 * XXX: given the "advice", should we consider async read-ahead?
1367 		 * XXX: fault current does deactive of pages behind us.  is
1368 		 * this good (other callers might now).
1369 		 */
1370 		/*
1371 		 * XXX: read-ahead currently handled by buffer cache (bread)
1372 		 * level.
1373 		 * XXX: no async i/o available.
1374 		 * XXX: so we don't do anything now.
1375 		 */
1376 
1377 		/*
1378 		 * step 1c: now we've either done everything needed or we to
1379 		 * unlock and do some waiting or I/O.
1380 		 */
1381 
1382 		*npagesp = gotpages;		/* let caller know */
1383 		if (done)
1384 			return(VM_PAGER_OK);		/* bingo! */
1385 		else
1386 			/* EEK!   Need to unlock and I/O */
1387 			return(VM_PAGER_UNLOCK);
1388 	}
1389 
1390 	/*
1391 	 * step 2: get non-resident or busy pages.
1392 	 * object is locked.   data structures are unlocked.
1393 	 *
1394 	 * XXX: because we can't do async I/O at this level we get things
1395 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1396 	 * async-read-ahead for us at a lower level.
1397 	 */
1398 
1399 	for (lcv = 0, current_offset = offset;
1400 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1401 
1402 		/* skip over pages we've already gotten or don't want */
1403 		/* skip over pages we don't _have_ to get */
1404 		if (pps[lcv] != NULL || (lcv != centeridx &&
1405 		    (flags & PGO_ALLPAGES) == 0))
1406 			continue;
1407 
1408 		/*
1409 		 * we have yet to locate the current page (pps[lcv]).   we first
1410 		 * look for a page that is already at the current offset.   if
1411 		 * we fine a page, we check to see if it is busy or released.
1412 		 * if that is the case, then we sleep on the page until it is
1413 		 * no longer busy or released and repeat the lookup.    if the
1414 		 * page we found is neither busy nor released, then we busy it
1415 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1416 		 * following while loop and indicates we are ready to move on
1417 		 * to the next page in the "lcv" loop above.
1418 		 *
1419 		 * if we exit the while loop with pps[lcv] still set to NULL,
1420 		 * then it means that we allocated a new busy/fake/clean page
1421 		 * ptmp in the object and we need to do I/O to fill in the data.
1422 		 */
1423 
1424 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1425 
1426 			/* look for a current page */
1427 			ptmp = uvm_pagelookup(uobj, current_offset);
1428 
1429 			/* nope?   allocate one now (if we can) */
1430 			if (ptmp == NULL) {
1431 
1432 				ptmp = uvm_pagealloc(uobj, current_offset,
1433 				    NULL, 0);
1434 
1435 				/* out of RAM? */
1436 				if (ptmp == NULL) {
1437 					simple_unlock(&uobj->vmobjlock);
1438 					uvm_wait("uvn_getpage");
1439 					simple_lock(&uobj->vmobjlock);
1440 
1441 					/* goto top of pps while loop */
1442 					continue;
1443 				}
1444 
1445 				/*
1446 				 * got new page ready for I/O.  break pps
1447 				 * while loop.  pps[lcv] is still NULL.
1448 				 */
1449 				break;
1450 			}
1451 
1452 			/* page is there, see if we need to wait on it */
1453 			if ((ptmp->pg_flags & (PG_BUSY|PG_RELEASED)) != 0) {
1454 				atomic_setbits_int(&ptmp->pg_flags, PG_WANTED);
1455 				UVM_UNLOCK_AND_WAIT(ptmp,
1456 				    &uobj->vmobjlock, FALSE, "uvn_get",0);
1457 				simple_lock(&uobj->vmobjlock);
1458 				continue;	/* goto top of pps while loop */
1459 			}
1460 
1461 			/*
1462 			 * if we get here then the page has become resident
1463 			 * and unbusy between steps 1 and 2.  we busy it
1464 			 * now (so we own it) and set pps[lcv] (so that we
1465 			 * exit the while loop).
1466 			 */
1467 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1468 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1469 			pps[lcv] = ptmp;
1470 		}
1471 
1472 		/*
1473 		 * if we own the a valid page at the correct offset, pps[lcv]
1474 		 * will point to it.   nothing more to do except go to the
1475 		 * next page.
1476 		 */
1477 
1478 		if (pps[lcv])
1479 			continue;			/* next lcv */
1480 
1481 		/*
1482 		 * we have a "fake/busy/clean" page that we just allocated.  do
1483 		 * I/O to fill it with valid data.  note that object must be
1484 		 * locked going into uvn_io, but will be unlocked afterwards.
1485 		 */
1486 
1487 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1488 		    PGO_SYNCIO, UIO_READ);
1489 
1490 		/*
1491 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1492 		 * syncio the result can not be PEND or AGAIN.   we must relock
1493 		 * and check for errors.
1494 		 */
1495 
1496 		/* lock object.   check for errors.   */
1497 		simple_lock(&uobj->vmobjlock);
1498 		if (result != VM_PAGER_OK) {
1499 			if (ptmp->pg_flags & PG_WANTED)
1500 				/* object lock still held */
1501 				wakeup(ptmp);
1502 
1503 			atomic_clearbits_int(&ptmp->pg_flags,
1504 			    PG_WANTED|PG_BUSY);
1505 			UVM_PAGE_OWN(ptmp, NULL);
1506 			uvm_lock_pageq();
1507 			uvm_pagefree(ptmp);
1508 			uvm_unlock_pageq();
1509 			simple_unlock(&uobj->vmobjlock);
1510 			return(result);
1511 		}
1512 
1513 		/*
1514 		 * we got the page!   clear the fake flag (indicates valid
1515 		 * data now in page) and plug into our result array.   note
1516 		 * that page is still busy.
1517 		 *
1518 		 * it is the callers job to:
1519 		 * => check if the page is released
1520 		 * => unbusy the page
1521 		 * => activate the page
1522 		 */
1523 
1524 		/* data is valid ... */
1525 		atomic_clearbits_int(&ptmp->pg_flags, PG_FAKE);
1526 		pmap_clear_modify(ptmp);		/* ... and clean */
1527 		pps[lcv] = ptmp;
1528 
1529 	}	/* lcv loop */
1530 
1531 	/*
1532 	 * finally, unlock object and return.
1533 	 */
1534 
1535 	simple_unlock(&uobj->vmobjlock);
1536 	return (VM_PAGER_OK);
1537 }
1538 
1539 /*
1540  * uvn_io: do I/O to a vnode
1541  *
1542  * => prefer map unlocked (not required)
1543  * => object must be locked!   we will _unlock_ it before starting I/O.
1544  * => flags: PGO_SYNCIO -- use sync. I/O
1545  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1546  *	[thus we never do async i/o!  see iodone comment]
1547  */
1548 
1549 int
1550 uvn_io(uvn, pps, npages, flags, rw)
1551 	struct uvm_vnode *uvn;
1552 	vm_page_t *pps;
1553 	int npages, flags, rw;
1554 {
1555 	struct vnode *vn;
1556 	struct uio uio;
1557 	struct iovec iov;
1558 	vaddr_t kva;
1559 	off_t file_offset;
1560 	int waitf, result, mapinflags;
1561 	size_t got, wanted;
1562 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1563 
1564 	UVMHIST_LOG(maphist, "rw=%ld", rw,0,0,0);
1565 
1566 	/*
1567 	 * init values
1568 	 */
1569 
1570 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1571 	vn = (struct vnode *) uvn;
1572 	file_offset = pps[0]->offset;
1573 
1574 	/*
1575 	 * check for sync'ing I/O.
1576 	 */
1577 
1578 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1579 		if (waitf == M_NOWAIT) {
1580 			simple_unlock(&uvn->u_obj.vmobjlock);
1581 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1582 			return(VM_PAGER_AGAIN);
1583 		}
1584 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1585 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1586 			FALSE, "uvn_iosync",0);
1587 		simple_lock(&uvn->u_obj.vmobjlock);
1588 	}
1589 
1590 	/*
1591 	 * check size
1592 	 */
1593 
1594 	if (file_offset >= uvn->u_size) {
1595 			simple_unlock(&uvn->u_obj.vmobjlock);
1596 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1597 			return(VM_PAGER_BAD);
1598 	}
1599 
1600 	/*
1601 	 * first try and map the pages in (without waiting)
1602 	 */
1603 
1604 	mapinflags = (rw == UIO_READ) ?
1605 	    UVMPAGER_MAPIN_READ : UVMPAGER_MAPIN_WRITE;
1606 
1607 	kva = uvm_pagermapin(pps, npages, mapinflags);
1608 	if (kva == 0 && waitf == M_NOWAIT) {
1609 		simple_unlock(&uvn->u_obj.vmobjlock);
1610 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1611 		return(VM_PAGER_AGAIN);
1612 	}
1613 
1614 	/*
1615 	 * ok, now bump u_nio up.   at this point we are done with uvn
1616 	 * and can unlock it.   if we still don't have a kva, try again
1617 	 * (this time with sleep ok).
1618 	 */
1619 
1620 	uvn->u_nio++;			/* we have an I/O in progress! */
1621 	simple_unlock(&uvn->u_obj.vmobjlock);
1622 	/* NOTE: object now unlocked */
1623 	if (kva == 0)
1624 		kva = uvm_pagermapin(pps, npages,
1625 		    mapinflags | UVMPAGER_MAPIN_WAITOK);
1626 
1627 	/*
1628 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1629 	 * get touched (so we can look at "offset" without having to lock
1630 	 * the object).  set up for I/O.
1631 	 */
1632 
1633 	/*
1634 	 * fill out uio/iov
1635 	 */
1636 
1637 	iov.iov_base = (caddr_t) kva;
1638 	wanted = npages << PAGE_SHIFT;
1639 	if (file_offset + wanted > uvn->u_size)
1640 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1641 	iov.iov_len = wanted;
1642 	uio.uio_iov = &iov;
1643 	uio.uio_iovcnt = 1;
1644 	uio.uio_offset = file_offset;
1645 	uio.uio_segflg = UIO_SYSSPACE;
1646 	uio.uio_rw = rw;
1647 	uio.uio_resid = wanted;
1648 	uio.uio_procp = curproc;
1649 
1650 	/*
1651 	 * do the I/O!  (XXX: curproc?)
1652 	 */
1653 
1654 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1655 
1656 	/*
1657 	 * This process may already have this vnode locked, if we faulted in
1658 	 * copyin() or copyout() on a region backed by this vnode
1659 	 * while doing I/O to the vnode.  If this is the case, don't
1660 	 * panic.. instead, return the error to the user.
1661 	 *
1662 	 * XXX this is a stopgap to prevent a panic.
1663 	 * Ideally, this kind of operation *should* work.
1664 	 */
1665 	result = 0;
1666 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1667 		result = vn_lock(vn, LK_EXCLUSIVE | LK_RECURSEFAIL, curproc);
1668 
1669 	if (result == 0) {
1670 		/* NOTE: vnode now locked! */
1671 
1672 		if (rw == UIO_READ)
1673 			result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1674 		else
1675 			result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1676 
1677 		if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1678 			VOP_UNLOCK(vn, 0, curproc);
1679 	}
1680 
1681 	/* NOTE: vnode now unlocked (unless vnislocked) */
1682 
1683 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1684 
1685 	/*
1686 	 * result == unix style errno (0 == OK!)
1687 	 *
1688 	 * zero out rest of buffer (if needed)
1689 	 */
1690 
1691 	if (result == 0) {
1692 		got = wanted - uio.uio_resid;
1693 
1694 		if (wanted && got == 0) {
1695 			result = EIO;		/* XXX: error? */
1696 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1697 			memset((void *) (kva + got), 0,
1698 			       (npages << PAGE_SHIFT) - got);
1699 		}
1700 	}
1701 
1702 	/*
1703 	 * now remove pager mapping
1704 	 */
1705 	uvm_pagermapout(kva, npages);
1706 
1707 	/*
1708 	 * now clean up the object (i.e. drop I/O count)
1709 	 */
1710 
1711 	simple_lock(&uvn->u_obj.vmobjlock);
1712 	/* NOTE: object now locked! */
1713 
1714 	uvn->u_nio--;			/* I/O DONE! */
1715 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1716 		wakeup(&uvn->u_nio);
1717 	}
1718 	simple_unlock(&uvn->u_obj.vmobjlock);
1719 	/* NOTE: object now unlocked! */
1720 
1721 	/*
1722 	 * done!
1723 	 */
1724 
1725 	UVMHIST_LOG(maphist, "<- done (result %ld)", result,0,0,0);
1726 	if (result == 0)
1727 		return(VM_PAGER_OK);
1728 	else
1729 		return(VM_PAGER_ERROR);
1730 }
1731 
1732 /*
1733  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1734  * is gone we will kill the object (flushing dirty pages back to the vnode
1735  * if needed).
1736  *
1737  * => returns TRUE if there was no uvm_object attached or if there was
1738  *	one and we killed it [i.e. if there is no active uvn]
1739  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1740  *	needed]
1741  *
1742  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1743  *	having to hold a reference on the vnode) and given a working
1744  *	uvm_vnp_sync(), how does that effect the need for this function?
1745  *      [XXXCDC: seems like it can die?]
1746  *
1747  * => XXX: this function should DIE once we merge the VM and buffer
1748  *	cache.
1749  *
1750  * research shows that this is called in the following places:
1751  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1752  *	changes sizes
1753  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1754  *	are written to
1755  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1756  *	is off
1757  * ffs_realloccg: when we can't extend the current block and have
1758  *	to allocate a new one we call this [XXX: why?]
1759  * nfsrv_rename, rename_files: called when the target filename is there
1760  *	and we want to remove it
1761  * nfsrv_remove, sys_unlink: called on file we are removing
1762  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1763  *	then return "text busy"
1764  * nfs_open: seems to uncache any file opened with nfs
1765  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1766  */
1767 
1768 boolean_t
1769 uvm_vnp_uncache(struct vnode *vp)
1770 {
1771 	struct uvm_vnode *uvn = &vp->v_uvm;
1772 
1773 	/*
1774 	 * lock uvn part of the vnode and check to see if we need to do anything
1775 	 */
1776 
1777 	simple_lock(&uvn->u_obj.vmobjlock);
1778 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1779 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1780 		simple_unlock(&uvn->u_obj.vmobjlock);
1781 		return(TRUE);
1782 	}
1783 
1784 	/*
1785 	 * we have a valid, non-blocked uvn.   clear persist flag.
1786 	 * if uvn is currently active we can return now.
1787 	 */
1788 
1789 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1790 	if (uvn->u_obj.uo_refs) {
1791 		simple_unlock(&uvn->u_obj.vmobjlock);
1792 		return(FALSE);
1793 	}
1794 
1795 	/*
1796 	 * uvn is currently persisting!   we have to gain a reference to
1797 	 * it so that we can call uvn_detach to kill the uvn.
1798 	 */
1799 
1800 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1801 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1802 	simple_unlock(&uvn->u_obj.vmobjlock);
1803 
1804 #ifdef VFSDEBUG
1805 	/*
1806 	 * carry over sanity check from old vnode pager: the vnode should
1807 	 * be VOP_LOCK'd, and we confirm it here.
1808 	 */
1809 	if ((vp->v_flag & VLOCKSWORK) && !VOP_ISLOCKED(vp))
1810 		panic("uvm_vnp_uncache: vnode not locked!");
1811 #endif	/* VFSDEBUG */
1812 
1813 	/*
1814 	 * now drop our reference to the vnode.   if we have the sole
1815 	 * reference to the vnode then this will cause it to die [as we
1816 	 * just cleared the persist flag].   we have to unlock the vnode
1817 	 * while we are doing this as it may trigger I/O.
1818 	 *
1819 	 * XXX: it might be possible for uvn to get reclaimed while we are
1820 	 * unlocked causing us to return TRUE when we should not.   we ignore
1821 	 * this as a false-positive return value doesn't hurt us.
1822 	 */
1823 	VOP_UNLOCK(vp, 0, curproc);
1824 	uvn_detach(&uvn->u_obj);
1825 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curproc);
1826 
1827 	/*
1828 	 * and return...
1829 	 */
1830 
1831 	return(TRUE);
1832 }
1833 
1834 /*
1835  * uvm_vnp_setsize: grow or shrink a vnode uvn
1836  *
1837  * grow   => just update size value
1838  * shrink => toss un-needed pages
1839  *
1840  * => we assume that the caller has a reference of some sort to the
1841  *	vnode in question so that it will not be yanked out from under
1842  *	us.
1843  *
1844  * called from:
1845  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1846  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1847  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1848  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1849  *  => union fs: union_newsize
1850  */
1851 
1852 void
1853 uvm_vnp_setsize(vp, newsize)
1854 	struct vnode *vp;
1855 	voff_t newsize;
1856 {
1857 	struct uvm_vnode *uvn = &vp->v_uvm;
1858 
1859 	/*
1860 	 * lock uvn and check for valid object, and if valid: do it!
1861 	 */
1862 	simple_lock(&uvn->u_obj.vmobjlock);
1863 	if (uvn->u_flags & UVM_VNODE_VALID) {
1864 
1865 		/*
1866 		 * now check if the size has changed: if we shrink we had better
1867 		 * toss some pages...
1868 		 */
1869 
1870 		if (uvn->u_size > newsize) {
1871 			(void)uvn_flush(&uvn->u_obj, newsize,
1872 			    uvn->u_size, PGO_FREE);
1873 		}
1874 		uvn->u_size = newsize;
1875 	}
1876 	simple_unlock(&uvn->u_obj.vmobjlock);
1877 
1878 	/*
1879 	 * done
1880 	 */
1881 	return;
1882 }
1883 
1884 /*
1885  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1886  *
1887  * => called from sys_sync with no VM structures locked
1888  * => only one process can do a sync at a time (because the uvn
1889  *    structure only has one queue for sync'ing).  we ensure this
1890  *    by holding the uvn_sync_lock while the sync is in progress.
1891  *    other processes attempting a sync will sleep on this lock
1892  *    until we are done.
1893  */
1894 void
1895 uvm_vnp_sync(struct mount *mp)
1896 {
1897 	struct uvm_vnode *uvn;
1898 	struct vnode *vp;
1899 
1900 	/*
1901 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1902 	 * our lock...
1903 	 */
1904 	rw_enter_write(&uvn_sync_lock);
1905 
1906 	/*
1907 	 * step 2: build up a simpleq of uvns of interest based on the
1908 	 * write list.   we gain a reference to uvns of interest.
1909 	 */
1910 	SIMPLEQ_INIT(&uvn_sync_q);
1911 	LIST_FOREACH(uvn, &uvn_wlist, u_wlist) {
1912 
1913 		vp = (struct vnode *)uvn;
1914 		if (mp && vp->v_mount != mp)
1915 			continue;
1916 
1917 		/*
1918 		 * If the vnode is "blocked" it means it must be dying, which
1919 		 * in turn means its in the process of being flushed out so
1920 		 * we can safely skip it.
1921 		 *
1922 		 * note that uvn must already be valid because we found it on
1923 		 * the wlist (this also means it can't be ALOCK'd).
1924 		 */
1925 		if ((uvn->u_flags & UVM_VNODE_BLOCKED) != 0)
1926 			continue;
1927 
1928 
1929 		/*
1930 		 * gain reference.   watch out for persisting uvns (need to
1931 		 * regain vnode REF).
1932 		 */
1933 		if (uvn->u_obj.uo_refs == 0)
1934 			VREF(vp);
1935 		uvn->u_obj.uo_refs++;
1936 
1937 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
1938 	}
1939 
1940 	/* step 3: we now have a list of uvn's that may need cleaning. */
1941 	SIMPLEQ_FOREACH(uvn, &uvn_sync_q, u_syncq) {
1942 #ifdef DEBUG
1943 		if (uvn->u_flags & UVM_VNODE_DYING) {
1944 			printf("uvm_vnp_sync: dying vnode on sync list\n");
1945 		}
1946 #endif
1947 		uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
1948 
1949 		/*
1950 		 * if we have the only reference and we just cleaned the uvn,
1951 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
1952 		 * thus allowing us to avoid thinking about flushing it again
1953 		 * on later sync ops.
1954 		 */
1955 		if (uvn->u_obj.uo_refs == 1 &&
1956 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
1957 			LIST_REMOVE(uvn, u_wlist);
1958 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
1959 		}
1960 
1961 		/* now drop our reference to the uvn */
1962 		uvn_detach(&uvn->u_obj);
1963 	}
1964 
1965 	rw_exit_write(&uvn_sync_lock);
1966 }
1967