xref: /openbsd-src/sys/uvm/uvm_vnode.c (revision b2ea75c1b17e1a9a339660e7ed45cd24946b230e)
1 /*	$OpenBSD: uvm_vnode.c,v 1.18 2001/08/11 10:57:22 art Exp $	*/
2 /*	$NetBSD: uvm_vnode.c,v 1.33 2000/05/19 03:45:05 thorpej 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 
63 #include <miscfs/specfs/specdev.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_kern.h>
68 
69 #include <uvm/uvm.h>
70 #include <uvm/uvm_vnode.h>
71 
72 /*
73  * private global data structure
74  *
75  * we keep a list of writeable active vnode-backed VM objects for sync op.
76  * we keep a simpleq of vnodes that are currently being sync'd.
77  */
78 
79 LIST_HEAD(uvn_list_struct, uvm_vnode);
80 static struct uvn_list_struct uvn_wlist;	/* writeable uvns */
81 static simple_lock_data_t uvn_wl_lock;		/* locks uvn_wlist */
82 
83 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
84 static struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
85 lock_data_t uvn_sync_lock;			/* locks sync operation */
86 
87 /*
88  * functions
89  */
90 
91 static int		   uvn_asyncget __P((struct uvm_object *, voff_t,
92 					    int));
93 static void		   uvn_cluster __P((struct uvm_object *, voff_t,
94 					   voff_t *, voff_t *));
95 static void                uvn_detach __P((struct uvm_object *));
96 static boolean_t           uvn_flush __P((struct uvm_object *, voff_t,
97 					 voff_t, int));
98 static int                 uvn_get __P((struct uvm_object *, voff_t,
99 					vm_page_t *, int *, int,
100 					vm_prot_t, int, int));
101 static void		   uvn_init __P((void));
102 static int		   uvn_io __P((struct uvm_vnode *, vm_page_t *,
103 				      int, int, int));
104 static int		   uvn_put __P((struct uvm_object *, vm_page_t *,
105 					int, boolean_t));
106 static void                uvn_reference __P((struct uvm_object *));
107 static boolean_t	   uvn_releasepg __P((struct vm_page *,
108 					      struct vm_page **));
109 
110 /*
111  * master pager structure
112  */
113 
114 struct uvm_pagerops uvm_vnodeops = {
115 	uvn_init,
116 	uvn_reference,
117 	uvn_detach,
118 	NULL,			/* no specialized fault routine required */
119 	uvn_flush,
120 	uvn_get,
121 	uvn_asyncget,
122 	uvn_put,
123 	uvn_cluster,
124 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
125 	NULL,		 /* AIO-DONE function (not until we have asyncio) */
126 	uvn_releasepg,
127 };
128 
129 /*
130  * the ops!
131  */
132 
133 /*
134  * uvn_init
135  *
136  * init pager private data structures.
137  */
138 
139 static void
140 uvn_init()
141 {
142 
143 	LIST_INIT(&uvn_wlist);
144 	simple_lock_init(&uvn_wl_lock);
145 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
146 	lockinit(&uvn_sync_lock, PVM, "uvnsync", 0, 0);
147 }
148 
149 /*
150  * uvn_attach
151  *
152  * attach a vnode structure to a VM object.  if the vnode is already
153  * attached, then just bump the reference count by one and return the
154  * VM object.   if not already attached, attach and return the new VM obj.
155  * the "accessprot" tells the max access the attaching thread wants to
156  * our pages.
157  *
158  * => caller must _not_ already be holding the lock on the uvm_object.
159  * => in fact, nothing should be locked so that we can sleep here.
160  * => note that uvm_object is first thing in vnode structure, so their
161  *    pointers are equiv.
162  */
163 
164 struct uvm_object *
165 uvn_attach(arg, accessprot)
166 	void *arg;
167 	vm_prot_t accessprot;
168 {
169 	struct vnode *vp = arg;
170 	struct uvm_vnode *uvn = &vp->v_uvm;
171 	struct vattr vattr;
172 	int oldflags, result;
173 	struct partinfo pi;
174 	u_quad_t used_vnode_size;
175 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
176 
177 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
178 
179 	used_vnode_size = (u_quad_t)0;	/* XXX gcc -Wuninitialized */
180 
181 	/*
182 	 * first get a lock on the uvn.
183 	 */
184 	simple_lock(&uvn->u_obj.vmobjlock);
185 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
186 		printf("uvn_attach: blocked at 0x%p flags 0x%x\n",
187 		    uvn, uvn->u_flags);
188 		uvn->u_flags |= UVM_VNODE_WANTED;
189 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
190 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
191 		    "uvn_attach", 0);
192 		simple_lock(&uvn->u_obj.vmobjlock);
193 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
194 	}
195 
196 	/*
197 	 * if we're mapping a BLK device, make sure it is a disk.
198 	 */
199 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
200 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
201 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
202 		return(NULL);
203 	}
204 
205 	/*
206 	 * now we have lock and uvn must not be in a blocked state.
207 	 * first check to see if it is already active, in which case
208 	 * we can bump the reference count, check to see if we need to
209 	 * add it to the writeable list, and then return.
210 	 */
211 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
212 
213 		/* regain VREF if we were persisting */
214 		if (uvn->u_obj.uo_refs == 0) {
215 			VREF(vp);
216 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
217 			    0,0,0,0);
218 		}
219 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
220 
221 		/* check for new writeable uvn */
222 		if ((accessprot & VM_PROT_WRITE) != 0 &&
223 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
224 			simple_lock(&uvn_wl_lock);
225 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
226 			simple_unlock(&uvn_wl_lock);
227 			/* we are now on wlist! */
228 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
229 		}
230 
231 		/* unlock and return */
232 		simple_unlock(&uvn->u_obj.vmobjlock);
233 		UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
234 		    0, 0, 0);
235 		return (&uvn->u_obj);
236 	}
237 
238 	/*
239 	 * need to call VOP_GETATTR() to get the attributes, but that could
240 	 * block (due to I/O), so we want to unlock the object before calling.
241 	 * however, we want to keep anyone else from playing with the object
242 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
243 	 * prevents anyone from attaching to the vnode until we are done with
244 	 * it.
245 	 */
246 	uvn->u_flags = UVM_VNODE_ALOCK;
247 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
248 		/* XXX: curproc? */
249 
250 	if (vp->v_type == VBLK) {
251 		/*
252 		 * We could implement this as a specfs getattr call, but:
253 		 *
254 		 *	(1) VOP_GETATTR() would get the file system
255 		 *	    vnode operation, not the specfs operation.
256 		 *
257 		 *	(2) All we want is the size, anyhow.
258 		 */
259 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
260 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
261 		if (result == 0) {
262 			/* XXX should remember blocksize */
263 			used_vnode_size = (u_quad_t)pi.disklab->d_secsize *
264 			    (u_quad_t)pi.part->p_size;
265 		}
266 	} else {
267 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
268 		if (result == 0)
269 			used_vnode_size = vattr.va_size;
270 	}
271 
272 	/* relock object */
273 	simple_lock(&uvn->u_obj.vmobjlock);
274 
275 	if (result != 0) {
276 		if (uvn->u_flags & UVM_VNODE_WANTED)
277 			wakeup(uvn);
278 		uvn->u_flags = 0;
279 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
280 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
281 		return(NULL);
282 	}
283 
284 	/*
285 	 * make sure that the newsize fits within a vaddr_t
286 	 * XXX: need to revise addressing data types
287 	 */
288 #ifdef DEBUG
289 	if (vp->v_type == VBLK)
290 		printf("used_vnode_size = %qu\n", (long long)used_vnode_size);
291 #endif
292 
293 	/*
294 	 * now set up the uvn.
295 	 */
296 	uvn->u_obj.pgops = &uvm_vnodeops;
297 	TAILQ_INIT(&uvn->u_obj.memq);
298 	uvn->u_obj.uo_npages = 0;
299 	uvn->u_obj.uo_refs = 1;			/* just us... */
300 	oldflags = uvn->u_flags;
301 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
302 	uvn->u_nio = 0;
303 	uvn->u_size = used_vnode_size;
304 
305 	/* if write access, we need to add it to the wlist */
306 	if (accessprot & VM_PROT_WRITE) {
307 		simple_lock(&uvn_wl_lock);
308 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
309 		simple_unlock(&uvn_wl_lock);
310 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
311 	}
312 
313 	/*
314 	 * add a reference to the vnode.   this reference will stay as long
315 	 * as there is a valid mapping of the vnode.   dropped when the
316 	 * reference count goes to zero [and we either free or persist].
317 	 */
318 	VREF(vp);
319 	simple_unlock(&uvn->u_obj.vmobjlock);
320 	if (oldflags & UVM_VNODE_WANTED)
321 		wakeup(uvn);
322 
323 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
324 	return(&uvn->u_obj);
325 }
326 
327 
328 /*
329  * uvn_reference
330  *
331  * duplicate a reference to a VM object.  Note that the reference
332  * count must already be at least one (the passed in reference) so
333  * there is no chance of the uvn being killed or locked out here.
334  *
335  * => caller must call with object unlocked.
336  * => caller must be using the same accessprot as was used at attach time
337  */
338 
339 
340 static void
341 uvn_reference(uobj)
342 	struct uvm_object *uobj;
343 {
344 #ifdef DEBUG
345 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
346 #endif
347 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
348 
349 	simple_lock(&uobj->vmobjlock);
350 #ifdef DEBUG
351 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
352 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
353 		    uobj->uo_refs);
354 		panic("uvn_reference: invalid state");
355 	}
356 #endif
357 	uobj->uo_refs++;
358 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
359 	uobj, uobj->uo_refs,0,0);
360 	simple_unlock(&uobj->vmobjlock);
361 }
362 
363 /*
364  * uvn_detach
365  *
366  * remove a reference to a VM object.
367  *
368  * => caller must call with object unlocked and map locked.
369  * => this starts the detach process, but doesn't have to finish it
370  *    (async i/o could still be pending).
371  */
372 static void
373 uvn_detach(uobj)
374 	struct uvm_object *uobj;
375 {
376 	struct uvm_vnode *uvn;
377 	struct vnode *vp;
378 	int oldflags;
379 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
380 
381 	simple_lock(&uobj->vmobjlock);
382 
383 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
384 	uobj->uo_refs--;			/* drop ref! */
385 	if (uobj->uo_refs) {			/* still more refs */
386 		simple_unlock(&uobj->vmobjlock);
387 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
388 		return;
389 	}
390 
391 	/*
392 	 * get other pointers ...
393 	 */
394 
395 	uvn = (struct uvm_vnode *) uobj;
396 	vp = (struct vnode *) uobj;
397 
398 	/*
399 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
400 	 * to keep an active text file from being overwritten).
401 	 */
402 	vp->v_flag &= ~VTEXT;
403 
404 	/*
405 	 * we just dropped the last reference to the uvn.   see if we can
406 	 * let it "stick around".
407 	 */
408 
409 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
410 		/* won't block */
411 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
412 		simple_unlock(&uobj->vmobjlock);
413 		vrele(vp);			/* drop vnode reference */
414 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
415 		return;
416 	}
417 
418 	/*
419 	 * its a goner!
420 	 */
421 
422 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
423 
424 	uvn->u_flags |= UVM_VNODE_DYING;
425 
426 	/*
427 	 * even though we may unlock in flush, no one can gain a reference
428 	 * to us until we clear the "dying" flag [because it blocks
429 	 * attaches].  we will not do that until after we've disposed of all
430 	 * the pages with uvn_flush().  note that before the flush the only
431 	 * pages that could be marked PG_BUSY are ones that are in async
432 	 * pageout by the daemon.  (there can't be any pending "get"'s
433 	 * because there are no references to the object).
434 	 */
435 
436 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
437 
438 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
439 
440 	/*
441 	 * given the structure of this pager, the above flush request will
442 	 * create the following state: all the pages that were in the object
443 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
444 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
445 	 * in either case, if we have pages left we can't kill the object
446 	 * yet because i/o is pending.  in this case we set the "relkill"
447 	 * flag which will cause pgo_releasepg to kill the object once all
448 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
449 	 * routine or from the page daemon].
450 	 */
451 
452 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
453 #ifdef DEBUG
454 		/*
455 		 * XXXCDC: very unlikely to happen until we have async i/o
456 		 * so print a little info message in case it does.
457 		 */
458 		printf("uvn_detach: vn %p has pages left after flush - "
459 		    "relkill mode\n", uobj);
460 #endif
461 		uvn->u_flags |= UVM_VNODE_RELKILL;
462 		simple_unlock(&uobj->vmobjlock);
463 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
464 		    0, 0);
465 		return;
466 	}
467 
468 	/*
469 	 * kill object now.   note that we can't be on the sync q because
470 	 * all references are gone.
471 	 */
472 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
473 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
474 		LIST_REMOVE(uvn, u_wlist);
475 		simple_unlock(&uvn_wl_lock);
476 	}
477 #ifdef DIAGNOSTIC
478 	if (uobj->memq.tqh_first != NULL)
479 		panic("uvn_deref: vnode VM object still has pages afer "
480 		    "syncio/free flush");
481 #endif
482 	oldflags = uvn->u_flags;
483 	uvn->u_flags = 0;
484 	simple_unlock(&uobj->vmobjlock);
485 
486 	/* wake up any sleepers */
487 	if (oldflags & UVM_VNODE_WANTED)
488 		wakeup(uvn);
489 
490 	/*
491 	 * drop our reference to the vnode.
492 	 */
493 	vrele(vp);
494 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
495 
496 	return;
497 }
498 
499 /*
500  * uvm_vnp_terminate: external hook to clear out a vnode's VM
501  *
502  * called in two cases:
503  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
504  *      count) needs to be freed so that a vnode can be reused.  this
505  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
506  *      the free list is still attached (i.e. not VBAD) then vgone is
507  *	called.   as part of the vgone trace this should get called to
508  *	free the vm object.   this is the common case.
509  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
510  *	"umount -f") the vgone() function is called on active vnodes
511  *	on the mounted file systems to kill their data (the vnodes become
512  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
513  *	call here (even if the uvn is still in use -- i.e. has a non-zero
514  *	reference count).  this case happens at "umount -f" and during a
515  *	"reboot/halt" operation.
516  *
517  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
518  *	[protects us from getting a vnode that is already in the DYING
519  *	 state...]
520  * => unlike uvn_detach, this function must not return until all the
521  *	uvn's pages are disposed of.
522  * => in case [2] the uvn is still alive after this call, but all I/O
523  *	ops will fail (due to the backing vnode now being "dead").  this
524  *	will prob. kill any process using the uvn due to pgo_get failing.
525  */
526 
527 void
528 uvm_vnp_terminate(vp)
529 	struct vnode *vp;
530 {
531 	struct uvm_vnode *uvn = &vp->v_uvm;
532 	int oldflags;
533 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
534 
535 	/*
536 	 * lock object and check if it is valid
537 	 */
538 	simple_lock(&uvn->u_obj.vmobjlock);
539 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
540 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
541 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
542 		simple_unlock(&uvn->u_obj.vmobjlock);
543 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
544 		return;
545 	}
546 
547 	/*
548 	 * must be a valid uvn that is not already dying (because XLOCK
549 	 * protects us from that).   the uvn can't in the ALOCK state
550 	 * because it is valid, and uvn's that are in the ALOCK state haven't
551 	 * been marked valid yet.
552 	 */
553 
554 #ifdef DEBUG
555 	/*
556 	 * debug check: are we yanking the vnode out from under our uvn?
557 	 */
558 	if (uvn->u_obj.uo_refs) {
559 		printf("uvm_vnp_terminate(%p): terminating active vnode "
560 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
561 	}
562 #endif
563 
564 	/*
565 	 * it is possible that the uvn was detached and is in the relkill
566 	 * state [i.e. waiting for async i/o to finish so that releasepg can
567 	 * kill object].  we take over the vnode now and cancel the relkill.
568 	 * we want to know when the i/o is done so we can recycle right
569 	 * away.   note that a uvn can only be in the RELKILL state if it
570 	 * has a zero reference count.
571 	 */
572 
573 	if (uvn->u_flags & UVM_VNODE_RELKILL)
574 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
575 
576 	/*
577 	 * block the uvn by setting the dying flag, and then flush the
578 	 * pages.  (note that flush may unlock object while doing I/O, but
579 	 * it will re-lock it before it returns control here).
580 	 *
581 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
582 	 * that uvn_io doesn't attempt to VOP_LOCK again.
583 	 *
584 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
585 	 *	due to a forceful unmount might not be a good idea.  maybe we
586 	 *	need a way to pass in this info to uvn_flush through a
587 	 *	pager-defined PGO_ constant [currently there are none].
588 	 */
589 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
590 
591 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
592 
593 	/*
594 	 * as we just did a flush we expect all the pages to be gone or in
595 	 * the process of going.  sleep to wait for the rest to go [via iosync].
596 	 */
597 
598 	while (uvn->u_obj.uo_npages) {
599 #ifdef DEBUG
600 		struct vm_page *pp;
601 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
602 		     pp = pp->listq.tqe_next) {
603 			if ((pp->flags & PG_BUSY) == 0)
604 				panic("uvm_vnp_terminate: detected unbusy pg");
605 		}
606 		if (uvn->u_nio == 0)
607 			panic("uvm_vnp_terminate: no I/O to wait for?");
608 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
609 		/*
610 		 * XXXCDC: this is unlikely to happen without async i/o so we
611 		 * put a printf in just to keep an eye on it.
612 		 */
613 #endif
614 		uvn->u_flags |= UVM_VNODE_IOSYNC;
615 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
616 		    "uvn_term",0);
617 		simple_lock(&uvn->u_obj.vmobjlock);
618 	}
619 
620 	/*
621 	 * done.   now we free the uvn if its reference count is zero
622 	 * (true if we are zapping a persisting uvn).   however, if we are
623 	 * terminating a uvn with active mappings we let it live ... future
624 	 * calls down to the vnode layer will fail.
625 	 */
626 
627 	oldflags = uvn->u_flags;
628 	if (uvn->u_obj.uo_refs) {
629 
630 		/*
631 		 * uvn must live on it is dead-vnode state until all references
632 		 * are gone.   restore flags.    clear CANPERSIST state.
633 		 */
634 
635 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
636 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
637 
638 	} else {
639 
640 		/*
641 		 * free the uvn now.   note that the VREF reference is already
642 		 * gone [it is dropped when we enter the persist state].
643 		 */
644 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
645 			panic("uvm_vnp_terminate: io sync wanted bit set");
646 
647 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
648 			simple_lock(&uvn_wl_lock);
649 			LIST_REMOVE(uvn, u_wlist);
650 			simple_unlock(&uvn_wl_lock);
651 		}
652 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
653 	}
654 
655 	if (oldflags & UVM_VNODE_WANTED)
656 		wakeup(uvn);		/* object lock still held */
657 
658 	simple_unlock(&uvn->u_obj.vmobjlock);
659 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
660 
661 }
662 
663 /*
664  * uvn_releasepg: handled a released page in a uvn
665  *
666  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
667  *	to dispose of.
668  * => caller must handled PG_WANTED case
669  * => called with page's object locked, pageq's unlocked
670  * => returns TRUE if page's object is still alive, FALSE if we
671  *	killed the page's object.    if we return TRUE, then we
672  *	return with the object locked.
673  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
674  *				with the page queues locked [for pagedaemon]
675  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
676  * => we kill the uvn if it is not referenced and we are suppose to
677  *	kill it ("relkill").
678  */
679 
680 boolean_t
681 uvn_releasepg(pg, nextpgp)
682 	struct vm_page *pg;
683 	struct vm_page **nextpgp;	/* OUT */
684 {
685 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
686 #ifdef DIAGNOSTIC
687 	if ((pg->flags & PG_RELEASED) == 0)
688 		panic("uvn_releasepg: page not released!");
689 #endif
690 
691 	/*
692 	 * dispose of the page [caller handles PG_WANTED]
693 	 */
694 	pmap_page_protect(pg, VM_PROT_NONE);
695 	uvm_lock_pageq();
696 	if (nextpgp)
697 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
698 	uvm_pagefree(pg);
699 	if (!nextpgp)
700 		uvm_unlock_pageq();
701 
702 	/*
703 	 * now see if we need to kill the object
704 	 */
705 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
706 		if (uvn->u_obj.uo_refs)
707 			panic("uvn_releasepg: kill flag set on referenced "
708 			    "object!");
709 		if (uvn->u_obj.uo_npages == 0) {
710 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
711 				simple_lock(&uvn_wl_lock);
712 				LIST_REMOVE(uvn, u_wlist);
713 				simple_unlock(&uvn_wl_lock);
714 			}
715 #ifdef DIAGNOSTIC
716 			if (uvn->u_obj.memq.tqh_first)
717 	panic("uvn_releasepg: pages in object with npages == 0");
718 #endif
719 			if (uvn->u_flags & UVM_VNODE_WANTED)
720 				/* still holding object lock */
721 				wakeup(uvn);
722 
723 			uvn->u_flags = 0;		/* DEAD! */
724 			simple_unlock(&uvn->u_obj.vmobjlock);
725 			return (FALSE);
726 		}
727 	}
728 	return (TRUE);
729 }
730 
731 /*
732  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
733  * through the buffer cache and allow I/O in any size.  These VOPs use
734  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
735  * go through the buffer cache or allow I/O sizes larger than a
736  * block].  we will eventually want to change this.
737  *
738  * issues to consider:
739  *   uvm provides the uvm_aiodesc structure for async i/o management.
740  * there are two tailq's in the uvm. structure... one for pending async
741  * i/o and one for "done" async i/o.   to do an async i/o one puts
742  * an aiodesc on the "pending" list (protected by splbio()), starts the
743  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
744  * some sort of "i/o done" function to be called (at splbio(), interrupt
745  * time).   this function should remove the aiodesc from the pending list
746  * and place it on the "done" list and wakeup the daemon.   the daemon
747  * will run at normal spl() and will remove all items from the "done"
748  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
749  * [in the old vm code, this was done by calling the "put" routine with
750  * null arguments which made the code harder to read and understand because
751  * you had one function ("put") doing two things.]
752  *
753  * so the current pager needs:
754  *   int uvn_aiodone(struct uvm_aiodesc *)
755  *
756  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
757  *	later collection.
758  * => called with pageq's locked by the daemon.
759  *
760  * general outline:
761  * - "try" to lock object.   if fail, just return (will try again later)
762  * - drop "u_nio" (this req is done!)
763  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
764  * - get "page" structures (atop?).
765  * - handle "wanted" pages
766  * - handle "released" pages [using pgo_releasepg]
767  *   >>> pgo_releasepg may kill the object
768  * dont forget to look at "object" wanted flag in all cases.
769  */
770 
771 
772 /*
773  * uvn_flush: flush pages out of a uvm object.
774  *
775  * => object should be locked by caller.   we may _unlock_ the object
776  *	if (and only if) we need to clean a page (PGO_CLEANIT).
777  *	we return with the object locked.
778  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
779  *	might want to unlock higher level resources (e.g. vm_map)
780  *	before calling flush.
781  * => if PGO_CLEANIT is not set, then we will neither unlock the object
782  *	or block.
783  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
784  *	for flushing.
785  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
786  *	that new pages are inserted on the tail end of the list.   thus,
787  *	we can make a complete pass through the object in one go by starting
788  *	at the head and working towards the tail (new pages are put in
789  *	front of us).
790  * => NOTE: we are allowed to lock the page queues, so the caller
791  *	must not be holding the lock on them [e.g. pagedaemon had
792  *	better not call us with the queues locked]
793  * => we return TRUE unless we encountered some sort of I/O error
794  *
795  * comment on "cleaning" object and PG_BUSY pages:
796  *	this routine is holding the lock on the object.   the only time
797  *	that it can run into a PG_BUSY page that it does not own is if
798  *	some other process has started I/O on the page (e.g. either
799  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
800  *	in, then it can not be dirty (!PG_CLEAN) because no one has
801  *	had a chance to modify it yet.    if the PG_BUSY page is being
802  *	paged out then it means that someone else has already started
803  *	cleaning the page for us (how nice!).    in this case, if we
804  *	have syncio specified, then after we make our pass through the
805  *	object we need to wait for the other PG_BUSY pages to clear
806  *	off (i.e. we need to do an iosync).   also note that once a
807  *	page is PG_BUSY it must stay in its object until it is un-busyed.
808  *
809  * note on page traversal:
810  *	we can traverse the pages in an object either by going down the
811  *	linked list in "uobj->memq", or we can go over the address range
812  *	by page doing hash table lookups for each address.    depending
813  *	on how many pages are in the object it may be cheaper to do one
814  *	or the other.   we set "by_list" to true if we are using memq.
815  *	if the cost of a hash lookup was equal to the cost of the list
816  *	traversal we could compare the number of pages in the start->stop
817  *	range to the total number of pages in the object.   however, it
818  *	seems that a hash table lookup is more expensive than the linked
819  *	list traversal, so we multiply the number of pages in the
820  *	start->stop range by a penalty which we define below.
821  */
822 
823 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
824 
825 static boolean_t
826 uvn_flush(uobj, start, stop, flags)
827 	struct uvm_object *uobj;
828 	voff_t start, stop;
829 	int flags;
830 {
831 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
832 	struct vm_page *pp, *ppnext, *ptmp;
833 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
834 	int npages, result, lcv;
835 	boolean_t retval, need_iosync, by_list, needs_clean, all;
836 	voff_t curoff;
837 	u_short pp_version;
838 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
839 
840 	curoff = 0;	/* XXX: shut up gcc */
841 	/*
842 	 * get init vals and determine how we are going to traverse object
843 	 */
844 
845 	need_iosync = FALSE;
846 	retval = TRUE;		/* return value */
847 	if (flags & PGO_ALLPAGES) {
848 		all = TRUE;
849 		by_list = TRUE;		/* always go by the list */
850 	} else {
851 		start = trunc_page(start);
852 		stop = round_page(stop);
853 #ifdef DEBUG
854 		if (stop > round_page(uvn->u_size))
855 			printf("uvn_flush: strange, got an out of range "
856 			    "flush (fixed)\n");
857 #endif
858 		all = FALSE;
859 		by_list = (uobj->uo_npages <=
860 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
861 	}
862 
863 	UVMHIST_LOG(maphist,
864 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
865 	    start, stop, by_list, flags);
866 
867 	/*
868 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
869 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
870 	 * is wrong it will only prevent us from clustering... it won't break
871 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
872 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
873 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
874 	 * bit is always up to date since there are no mappings).
875 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
876 	 */
877 
878 	if ((flags & PGO_CLEANIT) != 0 &&
879 	    uobj->pgops->pgo_mk_pcluster != NULL) {
880 		if (by_list) {
881 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
882 			    pp = pp->listq.tqe_next) {
883 				if (!all &&
884 				    (pp->offset < start || pp->offset >= stop))
885 					continue;
886 				pp->flags &= ~PG_CLEANCHK;
887 			}
888 
889 		} else {   /* by hash */
890 			for (curoff = start ; curoff < stop;
891 			    curoff += PAGE_SIZE) {
892 				pp = uvm_pagelookup(uobj, curoff);
893 				if (pp)
894 					pp->flags &= ~PG_CLEANCHK;
895 			}
896 		}
897 	}
898 
899 	/*
900 	 * now do it.   note: we must update ppnext in body of loop or we
901 	 * will get stuck.  we need to use ppnext because we may free "pp"
902 	 * before doing the next loop.
903 	 */
904 
905 	if (by_list) {
906 		pp = uobj->memq.tqh_first;
907 	} else {
908 		curoff = start;
909 		pp = uvm_pagelookup(uobj, curoff);
910 	}
911 
912 	ppnext = NULL;	/* XXX: shut up gcc */
913 	ppsp = NULL;		/* XXX: shut up gcc */
914 	uvm_lock_pageq();	/* page queues locked */
915 
916 	/* locked: both page queues and uobj */
917 	for ( ; (by_list && pp != NULL) ||
918 	  (!by_list && curoff < stop) ; pp = ppnext) {
919 
920 		if (by_list) {
921 
922 			/*
923 			 * range check
924 			 */
925 
926 			if (!all &&
927 			    (pp->offset < start || pp->offset >= stop)) {
928 				ppnext = pp->listq.tqe_next;
929 				continue;
930 			}
931 
932 		} else {
933 
934 			/*
935 			 * null check
936 			 */
937 
938 			curoff += PAGE_SIZE;
939 			if (pp == NULL) {
940 				if (curoff < stop)
941 					ppnext = uvm_pagelookup(uobj, curoff);
942 				continue;
943 			}
944 
945 		}
946 
947 		/*
948 		 * handle case where we do not need to clean page (either
949 		 * because we are not clean or because page is not dirty or
950 		 * is busy):
951 		 *
952 		 * NOTE: we are allowed to deactivate a non-wired active
953 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
954 		 * queue it must stay put until it is !PG_BUSY (so as not to
955 		 * confuse pagedaemon).
956 		 */
957 
958 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
959 			needs_clean = FALSE;
960 			if ((pp->flags & PG_BUSY) != 0 &&
961 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
962 			             (PGO_CLEANIT|PGO_SYNCIO))
963 				need_iosync = TRUE;
964 		} else {
965 			/*
966 			 * freeing: nuke all mappings so we can sync
967 			 * PG_CLEAN bit with no race
968 			 */
969 			if ((pp->flags & PG_CLEAN) != 0 &&
970 			    (flags & PGO_FREE) != 0 &&
971 			    (pp->pqflags & PQ_ACTIVE) != 0)
972 				pmap_page_protect(pp, VM_PROT_NONE);
973 			if ((pp->flags & PG_CLEAN) != 0 &&
974 			    pmap_is_modified(pp))
975 				pp->flags &= ~(PG_CLEAN);
976 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
977 
978 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
979 		}
980 
981 		/*
982 		 * if we don't need a clean... load ppnext and dispose of pp
983 		 */
984 		if (!needs_clean) {
985 			/* load ppnext */
986 			if (by_list)
987 				ppnext = pp->listq.tqe_next;
988 			else {
989 				if (curoff < stop)
990 					ppnext = uvm_pagelookup(uobj, curoff);
991 			}
992 
993 			/* now dispose of pp */
994 			if (flags & PGO_DEACTIVATE) {
995 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
996 				    pp->wire_count == 0) {
997 					pmap_page_protect(pp, VM_PROT_NONE);
998 					uvm_pagedeactivate(pp);
999 				}
1000 
1001 			} else if (flags & PGO_FREE) {
1002 				if (pp->flags & PG_BUSY) {
1003 					/* release busy pages */
1004 					pp->flags |= PG_RELEASED;
1005 				} else {
1006 					pmap_page_protect(pp, VM_PROT_NONE);
1007 					/* removed page from object */
1008 					uvm_pagefree(pp);
1009 				}
1010 			}
1011 			/* ppnext is valid so we can continue... */
1012 			continue;
1013 		}
1014 
1015 		/*
1016 		 * pp points to a page in the locked object that we are
1017 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
1018 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1019 		 *
1020 		 * let uvm_pager_put attempted a clustered page out.
1021 		 * note: locked: uobj and page queues.
1022 		 */
1023 
1024 		pp->flags |= PG_BUSY;	/* we 'own' page now */
1025 		UVM_PAGE_OWN(pp, "uvn_flush");
1026 		pmap_page_protect(pp, VM_PROT_READ);
1027 		pp_version = pp->version;
1028 ReTry:
1029 		ppsp = pps;
1030 		npages = sizeof(pps) / sizeof(struct vm_page *);
1031 
1032 		/* locked: page queues, uobj */
1033 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1034 			   flags | PGO_DOACTCLUST, start, stop);
1035 		/* unlocked: page queues, uobj */
1036 
1037 		/*
1038 		 * at this point nothing is locked.   if we did an async I/O
1039 		 * it is remotely possible for the async i/o to complete and
1040 		 * the page "pp" be freed or what not before we get a chance
1041 		 * to relock the object.   in order to detect this, we have
1042 		 * saved the version number of the page in "pp_version".
1043 		 */
1044 
1045 		/* relock! */
1046 		simple_lock(&uobj->vmobjlock);
1047 		uvm_lock_pageq();
1048 
1049 		/*
1050 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1051 		 * can only happen when  we are doing async I/O and can't
1052 		 * map the pages into kernel memory (pager_map) due to lack
1053 		 * of vm space.   if this happens we drop back to sync I/O.
1054 		 */
1055 
1056 		if (result == VM_PAGER_AGAIN) {
1057 			/*
1058 			 * it is unlikely, but page could have been released
1059 			 * while we had the object lock dropped.   we ignore
1060 			 * this now and retry the I/O.  we will detect and
1061 			 * handle the released page after the syncio I/O
1062 			 * completes.
1063 			 */
1064 #ifdef DIAGNOSTIC
1065 			if (flags & PGO_SYNCIO)
1066 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1067 #endif
1068 			flags |= PGO_SYNCIO;
1069 			goto ReTry;
1070 		}
1071 
1072 		/*
1073 		 * the cleaning operation is now done.   finish up.  note that
1074 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1075 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1076 		 * to us in ppsp/npages.
1077 		 */
1078 
1079 		/*
1080 		 * for pending async i/o if we are not deactivating/freeing
1081 		 * we can move on to the next page.
1082 		 */
1083 
1084 		if (result == VM_PAGER_PEND) {
1085 
1086 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1087 				/*
1088 				 * no per-page ops: refresh ppnext and continue
1089 				 */
1090 				if (by_list) {
1091 					if (pp->version == pp_version)
1092 						ppnext = pp->listq.tqe_next;
1093 					else
1094 						/* reset */
1095 						ppnext = uobj->memq.tqh_first;
1096 				} else {
1097 					if (curoff < stop)
1098 						ppnext = uvm_pagelookup(uobj,
1099 						    curoff);
1100 				}
1101 				continue;
1102 			}
1103 
1104 			/* need to do anything here? */
1105 		}
1106 
1107 		/*
1108 		 * need to look at each page of the I/O operation.  we defer
1109 		 * processing "pp" until the last trip through this "for" loop
1110 		 * so that we can load "ppnext" for the main loop after we
1111 		 * play with the cluster pages [thus the "npages + 1" in the
1112 		 * loop below].
1113 		 */
1114 
1115 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1116 
1117 			/*
1118 			 * handle ppnext for outside loop, and saving pp
1119 			 * until the end.
1120 			 */
1121 			if (lcv < npages) {
1122 				if (ppsp[lcv] == pp)
1123 					continue; /* skip pp until the end */
1124 				ptmp = ppsp[lcv];
1125 			} else {
1126 				ptmp = pp;
1127 
1128 				/* set up next page for outer loop */
1129 				if (by_list) {
1130 					if (pp->version == pp_version)
1131 						ppnext = pp->listq.tqe_next;
1132 					else
1133 						/* reset */
1134 						ppnext = uobj->memq.tqh_first;
1135 				} else {
1136 					if (curoff < stop)
1137 					ppnext = uvm_pagelookup(uobj, curoff);
1138 				}
1139 			}
1140 
1141 			/*
1142 			 * verify the page didn't get moved while obj was
1143 			 * unlocked
1144 			 */
1145 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1146 				continue;
1147 
1148 			/*
1149 			 * unbusy the page if I/O is done.   note that for
1150 			 * pending I/O it is possible that the I/O op
1151 			 * finished before we relocked the object (in
1152 			 * which case the page is no longer busy).
1153 			 */
1154 
1155 			if (result != VM_PAGER_PEND) {
1156 				if (ptmp->flags & PG_WANTED)
1157 					/* still holding object lock */
1158 					wakeup(ptmp);
1159 
1160 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1161 				UVM_PAGE_OWN(ptmp, NULL);
1162 				if (ptmp->flags & PG_RELEASED) {
1163 
1164 					/* pgo_releasepg wants this */
1165 					uvm_unlock_pageq();
1166 					if (!uvn_releasepg(ptmp, NULL))
1167 						return (TRUE);
1168 
1169 					uvm_lock_pageq();	/* relock */
1170 					continue;		/* next page */
1171 
1172 				} else {
1173 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1174 					if ((flags & PGO_FREE) == 0)
1175 						pmap_clear_modify(ptmp);
1176 				}
1177 			}
1178 
1179 			/*
1180 			 * dispose of page
1181 			 */
1182 
1183 			if (flags & PGO_DEACTIVATE) {
1184 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1185 				    pp->wire_count == 0) {
1186 					pmap_page_protect(ptmp, VM_PROT_NONE);
1187 					uvm_pagedeactivate(ptmp);
1188 				}
1189 
1190 			} else if (flags & PGO_FREE) {
1191 				if (result == VM_PAGER_PEND) {
1192 					if ((ptmp->flags & PG_BUSY) != 0)
1193 						/* signal for i/o done */
1194 						ptmp->flags |= PG_RELEASED;
1195 				} else {
1196 					if (result != VM_PAGER_OK) {
1197 						printf("uvn_flush: obj=%p, "
1198 						   "offset=0x%llx.  error "
1199 						   "during pageout.\n",
1200 						    pp->uobject,
1201 						    (long long)pp->offset);
1202 						printf("uvn_flush: WARNING: "
1203 						    "changes to page may be "
1204 						    "lost!\n");
1205 						retval = FALSE;
1206 					}
1207 					pmap_page_protect(ptmp, VM_PROT_NONE);
1208 					uvm_pagefree(ptmp);
1209 				}
1210 			}
1211 
1212 		}		/* end of "lcv" for loop */
1213 
1214 	}		/* end of "pp" for loop */
1215 
1216 	/*
1217 	 * done with pagequeues: unlock
1218 	 */
1219 	uvm_unlock_pageq();
1220 
1221 	/*
1222 	 * now wait for all I/O if required.
1223 	 */
1224 	if (need_iosync) {
1225 
1226 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1227 		while (uvn->u_nio != 0) {
1228 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1229 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1230 			  FALSE, "uvn_flush",0);
1231 			simple_lock(&uvn->u_obj.vmobjlock);
1232 		}
1233 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1234 			wakeup(&uvn->u_flags);
1235 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1236 	}
1237 
1238 	/* return, with object locked! */
1239 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1240 	return(retval);
1241 }
1242 
1243 /*
1244  * uvn_cluster
1245  *
1246  * we are about to do I/O in an object at offset.   this function is called
1247  * to establish a range of offsets around "offset" in which we can cluster
1248  * I/O.
1249  *
1250  * - currently doesn't matter if obj locked or not.
1251  */
1252 
1253 static void
1254 uvn_cluster(uobj, offset, loffset, hoffset)
1255 	struct uvm_object *uobj;
1256 	voff_t offset;
1257 	voff_t *loffset, *hoffset; /* OUT */
1258 {
1259 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1260 	*loffset = offset;
1261 
1262 	if (*loffset >= uvn->u_size)
1263 		panic("uvn_cluster: offset out of range");
1264 
1265 	/*
1266 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1267 	 */
1268 	*hoffset = *loffset + MAXBSIZE;
1269 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1270 		*hoffset = round_page(uvn->u_size);
1271 
1272 	return;
1273 }
1274 
1275 /*
1276  * uvn_put: flush page data to backing store.
1277  *
1278  * => prefer map unlocked (not required)
1279  * => object must be locked!   we will _unlock_ it before starting I/O.
1280  * => flags: PGO_SYNCIO -- use sync. I/O
1281  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1282  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1283  *	[thus we never do async i/o!  see iodone comment]
1284  */
1285 
1286 static int
1287 uvn_put(uobj, pps, npages, flags)
1288 	struct uvm_object *uobj;
1289 	struct vm_page **pps;
1290 	int npages, flags;
1291 {
1292 	int retval;
1293 
1294 	/* note: object locked */
1295 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1296 	/* note: object unlocked */
1297 
1298 	return(retval);
1299 }
1300 
1301 
1302 /*
1303  * uvn_get: get pages (synchronously) from backing store
1304  *
1305  * => prefer map unlocked (not required)
1306  * => object must be locked!  we will _unlock_ it before starting any I/O.
1307  * => flags: PGO_ALLPAGES: get all of the pages
1308  *           PGO_LOCKED: fault data structures are locked
1309  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1310  * => NOTE: caller must check for released pages!!
1311  */
1312 
1313 static int
1314 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1315 	struct uvm_object *uobj;
1316 	voff_t offset;
1317 	struct vm_page **pps;		/* IN/OUT */
1318 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1319 	int centeridx, advice, flags;
1320 	vm_prot_t access_type;
1321 {
1322 	voff_t current_offset;
1323 	struct vm_page *ptmp;
1324 	int lcv, result, gotpages;
1325 	boolean_t done;
1326 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1327 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1328 
1329 	/*
1330 	 * step 1: handled the case where fault data structures are locked.
1331 	 */
1332 
1333 	if (flags & PGO_LOCKED) {
1334 
1335 		/*
1336 		 * gotpages is the current number of pages we've gotten (which
1337 		 * we pass back up to caller via *npagesp.
1338 		 */
1339 
1340 		gotpages = 0;
1341 
1342 		/*
1343 		 * step 1a: get pages that are already resident.   only do this
1344 		 * if the data structures are locked (i.e. the first time
1345 		 * through).
1346 		 */
1347 
1348 		done = TRUE;	/* be optimistic */
1349 
1350 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1351 		    lcv++, current_offset += PAGE_SIZE) {
1352 
1353 			/* do we care about this page?  if not, skip it */
1354 			if (pps[lcv] == PGO_DONTCARE)
1355 				continue;
1356 
1357 			/* lookup page */
1358 			ptmp = uvm_pagelookup(uobj, current_offset);
1359 
1360 			/* to be useful must get a non-busy, non-released pg */
1361 			if (ptmp == NULL ||
1362 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1363 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1364 				    != 0)
1365 				done = FALSE;	/* need to do a wait or I/O! */
1366 				continue;
1367 			}
1368 
1369 			/*
1370 			 * useful page: busy/lock it and plug it in our
1371 			 * result array
1372 			 */
1373 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1374 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1375 			pps[lcv] = ptmp;
1376 			gotpages++;
1377 
1378 		}	/* "for" lcv loop */
1379 
1380 		/*
1381 		 * XXX: given the "advice", should we consider async read-ahead?
1382 		 * XXX: fault current does deactive of pages behind us.  is
1383 		 * this good (other callers might now).
1384 		 */
1385 		/*
1386 		 * XXX: read-ahead currently handled by buffer cache (bread)
1387 		 * level.
1388 		 * XXX: no async i/o available.
1389 		 * XXX: so we don't do anything now.
1390 		 */
1391 
1392 		/*
1393 		 * step 1c: now we've either done everything needed or we to
1394 		 * unlock and do some waiting or I/O.
1395 		 */
1396 
1397 		*npagesp = gotpages;		/* let caller know */
1398 		if (done)
1399 			return(VM_PAGER_OK);		/* bingo! */
1400 		else
1401 			/* EEK!   Need to unlock and I/O */
1402 			return(VM_PAGER_UNLOCK);
1403 	}
1404 
1405 	/*
1406 	 * step 2: get non-resident or busy pages.
1407 	 * object is locked.   data structures are unlocked.
1408 	 *
1409 	 * XXX: because we can't do async I/O at this level we get things
1410 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1411 	 * async-read-ahead for us at a lower level.
1412 	 */
1413 
1414 	for (lcv = 0, current_offset = offset ;
1415 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1416 
1417 		/* skip over pages we've already gotten or don't want */
1418 		/* skip over pages we don't _have_ to get */
1419 		if (pps[lcv] != NULL || (lcv != centeridx &&
1420 		    (flags & PGO_ALLPAGES) == 0))
1421 			continue;
1422 
1423 		/*
1424 		 * we have yet to locate the current page (pps[lcv]).   we first
1425 		 * look for a page that is already at the current offset.   if
1426 		 * we fine a page, we check to see if it is busy or released.
1427 		 * if that is the case, then we sleep on the page until it is
1428 		 * no longer busy or released and repeat the lookup.    if the
1429 		 * page we found is neither busy nor released, then we busy it
1430 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1431 		 * following while loop and indicates we are ready to move on
1432 		 * to the next page in the "lcv" loop above.
1433 		 *
1434 		 * if we exit the while loop with pps[lcv] still set to NULL,
1435 		 * then it means that we allocated a new busy/fake/clean page
1436 		 * ptmp in the object and we need to do I/O to fill in the data.
1437 		 */
1438 
1439 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1440 
1441 			/* look for a current page */
1442 			ptmp = uvm_pagelookup(uobj, current_offset);
1443 
1444 			/* nope?   allocate one now (if we can) */
1445 			if (ptmp == NULL) {
1446 
1447 				ptmp = uvm_pagealloc(uobj, current_offset,
1448 				    NULL, 0);
1449 
1450 				/* out of RAM? */
1451 				if (ptmp == NULL) {
1452 					simple_unlock(&uobj->vmobjlock);
1453 					uvm_wait("uvn_getpage");
1454 					simple_lock(&uobj->vmobjlock);
1455 
1456 					/* goto top of pps while loop */
1457 					continue;
1458 				}
1459 
1460 				/*
1461 				 * got new page ready for I/O.  break pps
1462 				 * while loop.  pps[lcv] is still NULL.
1463 				 */
1464 				break;
1465 			}
1466 
1467 			/* page is there, see if we need to wait on it */
1468 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1469 				ptmp->flags |= PG_WANTED;
1470 				UVM_UNLOCK_AND_WAIT(ptmp,
1471 				    &uobj->vmobjlock, FALSE, "uvn_get",0);
1472 				simple_lock(&uobj->vmobjlock);
1473 				continue;	/* goto top of pps while loop */
1474 			}
1475 
1476 			/*
1477 			 * if we get here then the page has become resident
1478 			 * and unbusy between steps 1 and 2.  we busy it
1479 			 * now (so we own it) and set pps[lcv] (so that we
1480 			 * exit the while loop).
1481 			 */
1482 			ptmp->flags |= PG_BUSY;
1483 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1484 			pps[lcv] = ptmp;
1485 		}
1486 
1487 		/*
1488 		 * if we own the a valid page at the correct offset, pps[lcv]
1489 		 * will point to it.   nothing more to do except go to the
1490 		 * next page.
1491 		 */
1492 
1493 		if (pps[lcv])
1494 			continue;			/* next lcv */
1495 
1496 		/*
1497 		 * we have a "fake/busy/clean" page that we just allocated.  do
1498 		 * I/O to fill it with valid data.  note that object must be
1499 		 * locked going into uvn_io, but will be unlocked afterwards.
1500 		 */
1501 
1502 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1503 		    PGO_SYNCIO, UIO_READ);
1504 
1505 		/*
1506 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1507 		 * syncio the result can not be PEND or AGAIN.   we must relock
1508 		 * and check for errors.
1509 		 */
1510 
1511 		/* lock object.   check for errors.   */
1512 		simple_lock(&uobj->vmobjlock);
1513 		if (result != VM_PAGER_OK) {
1514 			if (ptmp->flags & PG_WANTED)
1515 				/* object lock still held */
1516 				wakeup(ptmp);
1517 
1518 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1519 			UVM_PAGE_OWN(ptmp, NULL);
1520 			uvm_lock_pageq();
1521 			uvm_pagefree(ptmp);
1522 			uvm_unlock_pageq();
1523 			simple_unlock(&uobj->vmobjlock);
1524 			return(result);
1525 		}
1526 
1527 		/*
1528 		 * we got the page!   clear the fake flag (indicates valid
1529 		 * data now in page) and plug into our result array.   note
1530 		 * that page is still busy.
1531 		 *
1532 		 * it is the callers job to:
1533 		 * => check if the page is released
1534 		 * => unbusy the page
1535 		 * => activate the page
1536 		 */
1537 
1538 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1539 		pmap_clear_modify(ptmp);		/* ... and clean */
1540 		pps[lcv] = ptmp;
1541 
1542 	}	/* lcv loop */
1543 
1544 	/*
1545 	 * finally, unlock object and return.
1546 	 */
1547 
1548 	simple_unlock(&uobj->vmobjlock);
1549 	return (VM_PAGER_OK);
1550 }
1551 
1552 /*
1553  * uvn_asyncget: start async I/O to bring pages into ram
1554  *
1555  * => caller must lock object(???XXX: see if this is best)
1556  * => could be called from uvn_get or a madvise() fault-ahead.
1557  * => if it fails, it doesn't matter.
1558  */
1559 
1560 static int
1561 uvn_asyncget(uobj, offset, npages)
1562 	struct uvm_object *uobj;
1563 	voff_t offset;
1564 	int npages;
1565 {
1566 
1567 	/*
1568 	 * XXXCDC: we can't do async I/O yet
1569 	 */
1570 	printf("uvn_asyncget called\n");
1571 	return (KERN_SUCCESS);
1572 }
1573 
1574 /*
1575  * uvn_io: do I/O to a vnode
1576  *
1577  * => prefer map unlocked (not required)
1578  * => object must be locked!   we will _unlock_ it before starting I/O.
1579  * => flags: PGO_SYNCIO -- use sync. I/O
1580  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1581  *	[thus we never do async i/o!  see iodone comment]
1582  */
1583 
1584 static int
1585 uvn_io(uvn, pps, npages, flags, rw)
1586 	struct uvm_vnode *uvn;
1587 	vm_page_t *pps;
1588 	int npages, flags, rw;
1589 {
1590 	struct vnode *vn;
1591 	struct uio uio;
1592 	struct iovec iov;
1593 	vaddr_t kva;
1594 	off_t file_offset;
1595 	int waitf, result, mapinflags;
1596 	size_t got, wanted;
1597 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1598 
1599 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1600 
1601 	/*
1602 	 * init values
1603 	 */
1604 
1605 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1606 	vn = (struct vnode *) uvn;
1607 	file_offset = pps[0]->offset;
1608 
1609 	/*
1610 	 * check for sync'ing I/O.
1611 	 */
1612 
1613 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1614 		if (waitf == M_NOWAIT) {
1615 			simple_unlock(&uvn->u_obj.vmobjlock);
1616 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1617 			return(VM_PAGER_AGAIN);
1618 		}
1619 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1620 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1621 			FALSE, "uvn_iosync",0);
1622 		simple_lock(&uvn->u_obj.vmobjlock);
1623 	}
1624 
1625 	/*
1626 	 * check size
1627 	 */
1628 
1629 	if (file_offset >= uvn->u_size) {
1630 			simple_unlock(&uvn->u_obj.vmobjlock);
1631 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1632 			return(VM_PAGER_BAD);
1633 	}
1634 
1635 	/*
1636 	 * first try and map the pages in (without waiting)
1637 	 */
1638 
1639 	mapinflags = (rw == UIO_READ) ?
1640 	    UVMPAGER_MAPIN_READ : UVMPAGER_MAPIN_WRITE;
1641 
1642 	kva = uvm_pagermapin(pps, npages, NULL, mapinflags);
1643 	if (kva == 0 && waitf == M_NOWAIT) {
1644 		simple_unlock(&uvn->u_obj.vmobjlock);
1645 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1646 		return(VM_PAGER_AGAIN);
1647 	}
1648 
1649 	/*
1650 	 * ok, now bump u_nio up.   at this point we are done with uvn
1651 	 * and can unlock it.   if we still don't have a kva, try again
1652 	 * (this time with sleep ok).
1653 	 */
1654 
1655 	uvn->u_nio++;			/* we have an I/O in progress! */
1656 	simple_unlock(&uvn->u_obj.vmobjlock);
1657 	/* NOTE: object now unlocked */
1658 	if (kva == 0)
1659 		kva = uvm_pagermapin(pps, npages, NULL,
1660 		    mapinflags | UVMPAGER_MAPIN_WAITOK);
1661 
1662 	/*
1663 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1664 	 * get touched (so we can look at "offset" without having to lock
1665 	 * the object).  set up for I/O.
1666 	 */
1667 
1668 	/*
1669 	 * fill out uio/iov
1670 	 */
1671 
1672 	iov.iov_base = (caddr_t) kva;
1673 	wanted = npages << PAGE_SHIFT;
1674 	if (file_offset + wanted > uvn->u_size)
1675 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1676 	iov.iov_len = wanted;
1677 	uio.uio_iov = &iov;
1678 	uio.uio_iovcnt = 1;
1679 	uio.uio_offset = file_offset;
1680 	uio.uio_segflg = UIO_SYSSPACE;
1681 	uio.uio_rw = rw;
1682 	uio.uio_resid = wanted;
1683 	uio.uio_procp = curproc;
1684 
1685 	/*
1686 	 * do the I/O!  (XXX: curproc?)
1687 	 */
1688 
1689 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1690 
1691 	/*
1692 	 * This process may already have this vnode locked, if we faulted in
1693 	 * copyin() or copyout() on a region backed by this vnode
1694 	 * while doing I/O to the vnode.  If this is the case, don't
1695 	 * panic.. instead, return the error to the user.
1696 	 *
1697 	 * XXX this is a stopgap to prevent a panic.
1698 	 * Ideally, this kind of operation *should* work.
1699 	 */
1700 	result = 0;
1701 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1702 		result = vn_lock(vn, LK_EXCLUSIVE | LK_RETRY, curproc /*XXX*/);
1703 
1704 	if (result == 0) {
1705 		/* NOTE: vnode now locked! */
1706 
1707 		if (rw == UIO_READ)
1708 			result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1709 		else
1710 			result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1711 
1712 		if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1713 			VOP_UNLOCK(vn, 0, curproc /*XXX*/);
1714 	}
1715 
1716 	/* NOTE: vnode now unlocked (unless vnislocked) */
1717 
1718 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1719 
1720 	/*
1721 	 * result == unix style errno (0 == OK!)
1722 	 *
1723 	 * zero out rest of buffer (if needed)
1724 	 */
1725 
1726 	if (result == 0) {
1727 		got = wanted - uio.uio_resid;
1728 
1729 		if (wanted && got == 0) {
1730 			result = EIO;		/* XXX: error? */
1731 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1732 			memset((void *) (kva + got), 0,
1733 			       (npages << PAGE_SHIFT) - got);
1734 		}
1735 	}
1736 
1737 	/*
1738 	 * now remove pager mapping
1739 	 */
1740 	uvm_pagermapout(kva, npages);
1741 
1742 	/*
1743 	 * now clean up the object (i.e. drop I/O count)
1744 	 */
1745 
1746 	simple_lock(&uvn->u_obj.vmobjlock);
1747 	/* NOTE: object now locked! */
1748 
1749 	uvn->u_nio--;			/* I/O DONE! */
1750 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1751 		wakeup(&uvn->u_nio);
1752 	}
1753 	simple_unlock(&uvn->u_obj.vmobjlock);
1754 	/* NOTE: object now unlocked! */
1755 
1756 	/*
1757 	 * done!
1758 	 */
1759 
1760 	UVMHIST_LOG(maphist, "<- done (result %d)", result,0,0,0);
1761 	if (result == 0)
1762 		return(VM_PAGER_OK);
1763 	else
1764 		return(VM_PAGER_ERROR);
1765 }
1766 
1767 /*
1768  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1769  * is gone we will kill the object (flushing dirty pages back to the vnode
1770  * if needed).
1771  *
1772  * => returns TRUE if there was no uvm_object attached or if there was
1773  *	one and we killed it [i.e. if there is no active uvn]
1774  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1775  *	needed]
1776  *
1777  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1778  *	having to hold a reference on the vnode) and given a working
1779  *	uvm_vnp_sync(), how does that effect the need for this function?
1780  *      [XXXCDC: seems like it can die?]
1781  *
1782  * => XXX: this function should DIE once we merge the VM and buffer
1783  *	cache.
1784  *
1785  * research shows that this is called in the following places:
1786  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1787  *	changes sizes
1788  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1789  *	are written to
1790  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1791  *	is off
1792  * ffs_realloccg: when we can't extend the current block and have
1793  *	to allocate a new one we call this [XXX: why?]
1794  * nfsrv_rename, rename_files: called when the target filename is there
1795  *	and we want to remove it
1796  * nfsrv_remove, sys_unlink: called on file we are removing
1797  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1798  *	then return "text busy"
1799  * nfs_open: seems to uncache any file opened with nfs
1800  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1801  */
1802 
1803 boolean_t
1804 uvm_vnp_uncache(vp)
1805 	struct vnode *vp;
1806 {
1807 	struct uvm_vnode *uvn = &vp->v_uvm;
1808 
1809 	/*
1810 	 * lock uvn part of the vnode and check to see if we need to do anything
1811 	 */
1812 
1813 	simple_lock(&uvn->u_obj.vmobjlock);
1814 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1815 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1816 		simple_unlock(&uvn->u_obj.vmobjlock);
1817 		return(TRUE);
1818 	}
1819 
1820 	/*
1821 	 * we have a valid, non-blocked uvn.   clear persist flag.
1822 	 * if uvn is currently active we can return now.
1823 	 */
1824 
1825 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1826 	if (uvn->u_obj.uo_refs) {
1827 		simple_unlock(&uvn->u_obj.vmobjlock);
1828 		return(FALSE);
1829 	}
1830 
1831 	/*
1832 	 * uvn is currently persisting!   we have to gain a reference to
1833 	 * it so that we can call uvn_detach to kill the uvn.
1834 	 */
1835 
1836 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1837 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1838 	simple_unlock(&uvn->u_obj.vmobjlock);
1839 
1840 
1841 #ifdef DEBUG
1842 	/*
1843 	 * carry over sanity check from old vnode pager: the vnode should
1844 	 * be VOP_LOCK'd, and we confirm it here.
1845 	 */
1846 	if (!VOP_ISLOCKED(vp)) {
1847 		boolean_t is_ok_anyway = FALSE;
1848 #if defined(NFSCLIENT)
1849 		extern int (**nfsv2_vnodeop_p) __P((void *));
1850 		extern int (**spec_nfsv2nodeop_p) __P((void *));
1851 		extern int (**fifo_nfsv2nodeop_p) __P((void *));
1852 
1853 		/* vnode is NOT VOP_LOCKed: some vnode types _never_ lock */
1854 		if (vp->v_op == nfsv2_vnodeop_p ||
1855 		    vp->v_op == spec_nfsv2nodeop_p) {
1856 			is_ok_anyway = TRUE;
1857 		}
1858 		if (vp->v_op == fifo_nfsv2nodeop_p) {
1859 			is_ok_anyway = TRUE;
1860 		}
1861 #endif	/* defined(NFSSERVER) || defined(NFSCLIENT) */
1862 		if (!is_ok_anyway)
1863 			panic("uvm_vnp_uncache: vnode not locked!");
1864 	}
1865 #endif	/* DEBUG */
1866 
1867 	/*
1868 	 * now drop our reference to the vnode.   if we have the sole
1869 	 * reference to the vnode then this will cause it to die [as we
1870 	 * just cleared the persist flag].   we have to unlock the vnode
1871 	 * while we are doing this as it may trigger I/O.
1872 	 *
1873 	 * XXX: it might be possible for uvn to get reclaimed while we are
1874 	 * unlocked causing us to return TRUE when we should not.   we ignore
1875 	 * this as a false-positive return value doesn't hurt us.
1876 	 */
1877 	VOP_UNLOCK(vp, 0, curproc /*XXX*/);
1878 	uvn_detach(&uvn->u_obj);
1879 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curproc/*XXX*/);
1880 
1881 	/*
1882 	 * and return...
1883 	 */
1884 
1885 	return(TRUE);
1886 }
1887 
1888 /*
1889  * uvm_vnp_setsize: grow or shrink a vnode uvn
1890  *
1891  * grow   => just update size value
1892  * shrink => toss un-needed pages
1893  *
1894  * => we assume that the caller has a reference of some sort to the
1895  *	vnode in question so that it will not be yanked out from under
1896  *	us.
1897  *
1898  * called from:
1899  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1900  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1901  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1902  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1903  *  => union fs: union_newsize
1904  */
1905 
1906 void
1907 uvm_vnp_setsize(vp, newsize)
1908 	struct vnode *vp;
1909 	voff_t newsize;
1910 {
1911 	struct uvm_vnode *uvn = &vp->v_uvm;
1912 
1913 	/*
1914 	 * lock uvn and check for valid object, and if valid: do it!
1915 	 */
1916 	simple_lock(&uvn->u_obj.vmobjlock);
1917 	if (uvn->u_flags & UVM_VNODE_VALID) {
1918 
1919 		/*
1920 		 * now check if the size has changed: if we shrink we had better
1921 		 * toss some pages...
1922 		 */
1923 
1924 		if (uvn->u_size > newsize) {
1925 			(void)uvn_flush(&uvn->u_obj, newsize,
1926 			    uvn->u_size, PGO_FREE);
1927 		}
1928 		uvn->u_size = newsize;
1929 	}
1930 	simple_unlock(&uvn->u_obj.vmobjlock);
1931 
1932 	/*
1933 	 * done
1934 	 */
1935 	return;
1936 }
1937 
1938 /*
1939  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1940  *
1941  * => called from sys_sync with no VM structures locked
1942  * => only one process can do a sync at a time (because the uvn
1943  *    structure only has one queue for sync'ing).  we ensure this
1944  *    by holding the uvn_sync_lock while the sync is in progress.
1945  *    other processes attempting a sync will sleep on this lock
1946  *    until we are done.
1947  */
1948 
1949 void
1950 uvm_vnp_sync(mp)
1951 	struct mount *mp;
1952 {
1953 	struct uvm_vnode *uvn;
1954 	struct vnode *vp;
1955 	boolean_t got_lock;
1956 
1957 	/*
1958 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1959 	 * our lock...
1960 	 */
1961 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0, curproc /*XXX*/);
1962 
1963 	/*
1964 	 * step 2: build up a simpleq of uvns of interest based on the
1965 	 * write list.   we gain a reference to uvns of interest.  must
1966 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1967 	 * in the body of the loop.
1968 	 */
1969 	SIMPLEQ_INIT(&uvn_sync_q);
1970 	simple_lock(&uvn_wl_lock);
1971 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1972 	    uvn = uvn->u_wlist.le_next) {
1973 
1974 		vp = (struct vnode *) uvn;
1975 		if (mp && vp->v_mount != mp)
1976 			continue;
1977 
1978 		/* attempt to gain reference */
1979 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1980 		    						FALSE &&
1981 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1982 			/* spin */ ;
1983 
1984 		/*
1985 		 * we will exit the loop if either if the following are true:
1986 		 *  - we got the lock [always true if NCPU == 1]
1987 		 *  - we failed to get the lock but noticed the vnode was
1988 		 * 	"blocked" -- in this case the vnode must be a dying
1989 		 *	vnode, and since dying vnodes are in the process of
1990 		 *	being flushed out, we can safely skip this one
1991 		 *
1992 		 * we want to skip over the vnode if we did not get the lock,
1993 		 * or if the vnode is already dying (due to the above logic).
1994 		 *
1995 		 * note that uvn must already be valid because we found it on
1996 		 * the wlist (this also means it can't be ALOCK'd).
1997 		 */
1998 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1999 			if (got_lock)
2000 				simple_unlock(&uvn->u_obj.vmobjlock);
2001 			continue;		/* skip it */
2002 		}
2003 
2004 		/*
2005 		 * gain reference.   watch out for persisting uvns (need to
2006 		 * regain vnode REF).
2007 		 */
2008 		if (uvn->u_obj.uo_refs == 0)
2009 			VREF(vp);
2010 		uvn->u_obj.uo_refs++;
2011 		simple_unlock(&uvn->u_obj.vmobjlock);
2012 
2013 		/*
2014 		 * got it!
2015 		 */
2016 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
2017 	}
2018 	simple_unlock(&uvn_wl_lock);
2019 
2020 	/*
2021 	 * step 3: we now have a list of uvn's that may need cleaning.
2022 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
2023 	 * (so we can now safely lock uvn's again).
2024 	 */
2025 
2026 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2027 		simple_lock(&uvn->u_obj.vmobjlock);
2028 #ifdef DEBUG
2029 		if (uvn->u_flags & UVM_VNODE_DYING) {
2030 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2031 		}
2032 #endif
2033 		uvn_flush(&uvn->u_obj, 0, 0,
2034 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2035 
2036 		/*
2037 		 * if we have the only reference and we just cleaned the uvn,
2038 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2039 		 * thus allowing us to avoid thinking about flushing it again
2040 		 * on later sync ops.
2041 		 */
2042 		if (uvn->u_obj.uo_refs == 1 &&
2043 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2044 			LIST_REMOVE(uvn, u_wlist);
2045 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2046 		}
2047 
2048 		simple_unlock(&uvn->u_obj.vmobjlock);
2049 
2050 		/* now drop our reference to the uvn */
2051 		uvn_detach(&uvn->u_obj);
2052 	}
2053 
2054 	/*
2055 	 * done!  release sync lock
2056 	 */
2057 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0, curproc /*XXX*/);
2058 }
2059