xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision 4472dbe5e3bd91ef2540bada7a7ca7384627ff9b)
1 /*	$NetBSD: uvm_vnode.c,v 1.33 2000/05/19 03:45:05 thorpej Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993
6  *      The Regents of the University of California.
7  * Copyright (c) 1990 University of Utah.
8  *
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * the Systems Programming Group of the University of Utah Computer
13  * Science Department.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. All advertising materials mentioning features or use of this software
24  *    must display the following acknowledgement:
25  *      This product includes software developed by Charles D. Cranor,
26  *	Washington University, the University of California, Berkeley and
27  *	its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  *
44  *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
45  * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
46  */
47 
48 #include "fs_nfs.h"
49 #include "opt_uvmhist.h"
50 
51 /*
52  * uvm_vnode.c: the vnode pager.
53  */
54 
55 #include <sys/param.h>
56 #include <sys/systm.h>
57 #include <sys/proc.h>
58 #include <sys/malloc.h>
59 #include <sys/vnode.h>
60 #include <sys/disklabel.h>
61 #include <sys/ioctl.h>
62 #include <sys/fcntl.h>
63 #include <sys/conf.h>
64 
65 #include <miscfs/specfs/specdev.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_kern.h>
70 
71 #include <uvm/uvm.h>
72 #include <uvm/uvm_vnode.h>
73 
74 /*
75  * private global data structure
76  *
77  * we keep a list of writeable active vnode-backed VM objects for sync op.
78  * we keep a simpleq of vnodes that are currently being sync'd.
79  */
80 
81 LIST_HEAD(uvn_list_struct, uvm_vnode);
82 static struct uvn_list_struct uvn_wlist;	/* writeable uvns */
83 static simple_lock_data_t uvn_wl_lock;		/* locks uvn_wlist */
84 
85 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
86 static struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
87 lock_data_t uvn_sync_lock;			/* locks sync operation */
88 
89 /*
90  * functions
91  */
92 
93 static int		   uvn_asyncget __P((struct uvm_object *, voff_t,
94 					    int));
95 static void		   uvn_cluster __P((struct uvm_object *, voff_t,
96 					   voff_t *, voff_t *));
97 static void                uvn_detach __P((struct uvm_object *));
98 static boolean_t           uvn_flush __P((struct uvm_object *, voff_t,
99 					 voff_t, int));
100 static int                 uvn_get __P((struct uvm_object *, voff_t,
101 					vm_page_t *, int *, int,
102 					vm_prot_t, int, int));
103 static void		   uvn_init __P((void));
104 static int		   uvn_io __P((struct uvm_vnode *, vm_page_t *,
105 				      int, int, int));
106 static int		   uvn_put __P((struct uvm_object *, vm_page_t *,
107 					int, boolean_t));
108 static void                uvn_reference __P((struct uvm_object *));
109 static boolean_t	   uvn_releasepg __P((struct vm_page *,
110 					      struct vm_page **));
111 
112 /*
113  * master pager structure
114  */
115 
116 struct uvm_pagerops uvm_vnodeops = {
117 	uvn_init,
118 	uvn_reference,
119 	uvn_detach,
120 	NULL,			/* no specialized fault routine required */
121 	uvn_flush,
122 	uvn_get,
123 	uvn_asyncget,
124 	uvn_put,
125 	uvn_cluster,
126 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
127 	NULL,		 /* AIO-DONE function (not until we have asyncio) */
128 	uvn_releasepg,
129 };
130 
131 /*
132  * the ops!
133  */
134 
135 /*
136  * uvn_init
137  *
138  * init pager private data structures.
139  */
140 
141 static void
142 uvn_init()
143 {
144 
145 	LIST_INIT(&uvn_wlist);
146 	simple_lock_init(&uvn_wl_lock);
147 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
148 	lockinit(&uvn_sync_lock, PVM, "uvnsync", 0, 0);
149 }
150 
151 /*
152  * uvn_attach
153  *
154  * attach a vnode structure to a VM object.  if the vnode is already
155  * attached, then just bump the reference count by one and return the
156  * VM object.   if not already attached, attach and return the new VM obj.
157  * the "accessprot" tells the max access the attaching thread wants to
158  * our pages.
159  *
160  * => caller must _not_ already be holding the lock on the uvm_object.
161  * => in fact, nothing should be locked so that we can sleep here.
162  * => note that uvm_object is first thing in vnode structure, so their
163  *    pointers are equiv.
164  */
165 
166 struct uvm_object *
167 uvn_attach(arg, accessprot)
168 	void *arg;
169 	vm_prot_t accessprot;
170 {
171 	struct vnode *vp = arg;
172 	struct uvm_vnode *uvn = &vp->v_uvm;
173 	struct vattr vattr;
174 	int oldflags, result;
175 	struct partinfo pi;
176 	u_quad_t used_vnode_size;
177 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
178 
179 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
180 
181 	used_vnode_size = (u_quad_t)0;	/* XXX gcc -Wuninitialized */
182 
183 	/*
184 	 * first get a lock on the uvn.
185 	 */
186 	simple_lock(&uvn->u_obj.vmobjlock);
187 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
188 		printf("uvn_attach: blocked at 0x%p flags 0x%x\n",
189 		    uvn, uvn->u_flags);
190 		uvn->u_flags |= UVM_VNODE_WANTED;
191 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
192 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
193 		    "uvn_attach", 0);
194 		simple_lock(&uvn->u_obj.vmobjlock);
195 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
196 	}
197 
198 	/*
199 	 * if we're mapping a BLK device, make sure it is a disk.
200 	 */
201 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
202 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
203 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
204 		return(NULL);
205 	}
206 
207 	/*
208 	 * now we have lock and uvn must not be in a blocked state.
209 	 * first check to see if it is already active, in which case
210 	 * we can bump the reference count, check to see if we need to
211 	 * add it to the writeable list, and then return.
212 	 */
213 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
214 
215 		/* regain VREF if we were persisting */
216 		if (uvn->u_obj.uo_refs == 0) {
217 			VREF(vp);
218 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
219 			    0,0,0,0);
220 		}
221 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
222 
223 		/* check for new writeable uvn */
224 		if ((accessprot & VM_PROT_WRITE) != 0 &&
225 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
226 			simple_lock(&uvn_wl_lock);
227 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
228 			simple_unlock(&uvn_wl_lock);
229 			/* we are now on wlist! */
230 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
231 		}
232 
233 		/* unlock and return */
234 		simple_unlock(&uvn->u_obj.vmobjlock);
235 		UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
236 		    0, 0, 0);
237 		return (&uvn->u_obj);
238 	}
239 
240 	/*
241 	 * need to call VOP_GETATTR() to get the attributes, but that could
242 	 * block (due to I/O), so we want to unlock the object before calling.
243 	 * however, we want to keep anyone else from playing with the object
244 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
245 	 * prevents anyone from attaching to the vnode until we are done with
246 	 * it.
247 	 */
248 	uvn->u_flags = UVM_VNODE_ALOCK;
249 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
250 		/* XXX: curproc? */
251 
252 	if (vp->v_type == VBLK) {
253 		/*
254 		 * We could implement this as a specfs getattr call, but:
255 		 *
256 		 *	(1) VOP_GETATTR() would get the file system
257 		 *	    vnode operation, not the specfs operation.
258 		 *
259 		 *	(2) All we want is the size, anyhow.
260 		 */
261 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
262 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
263 		if (result == 0) {
264 			/* XXX should remember blocksize */
265 			used_vnode_size = (u_quad_t)pi.disklab->d_secsize *
266 			    (u_quad_t)pi.part->p_size;
267 		}
268 	} else {
269 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
270 		if (result == 0)
271 			used_vnode_size = vattr.va_size;
272 	}
273 
274 	/* relock object */
275 	simple_lock(&uvn->u_obj.vmobjlock);
276 
277 	if (result != 0) {
278 		if (uvn->u_flags & UVM_VNODE_WANTED)
279 			wakeup(uvn);
280 		uvn->u_flags = 0;
281 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
282 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
283 		return(NULL);
284 	}
285 
286 	/*
287 	 * make sure that the newsize fits within a vaddr_t
288 	 * XXX: need to revise addressing data types
289 	 */
290 #ifdef DEBUG
291 	if (vp->v_type == VBLK)
292 		printf("used_vnode_size = %qu\n", (long long)used_vnode_size);
293 #endif
294 
295 	/*
296 	 * now set up the uvn.
297 	 */
298 	uvn->u_obj.pgops = &uvm_vnodeops;
299 	TAILQ_INIT(&uvn->u_obj.memq);
300 	uvn->u_obj.uo_npages = 0;
301 	uvn->u_obj.uo_refs = 1;			/* just us... */
302 	oldflags = uvn->u_flags;
303 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
304 	uvn->u_nio = 0;
305 	uvn->u_size = used_vnode_size;
306 
307 	/* if write access, we need to add it to the wlist */
308 	if (accessprot & VM_PROT_WRITE) {
309 		simple_lock(&uvn_wl_lock);
310 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
311 		simple_unlock(&uvn_wl_lock);
312 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
313 	}
314 
315 	/*
316 	 * add a reference to the vnode.   this reference will stay as long
317 	 * as there is a valid mapping of the vnode.   dropped when the
318 	 * reference count goes to zero [and we either free or persist].
319 	 */
320 	VREF(vp);
321 	simple_unlock(&uvn->u_obj.vmobjlock);
322 	if (oldflags & UVM_VNODE_WANTED)
323 		wakeup(uvn);
324 
325 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
326 	return(&uvn->u_obj);
327 }
328 
329 
330 /*
331  * uvn_reference
332  *
333  * duplicate a reference to a VM object.  Note that the reference
334  * count must already be at least one (the passed in reference) so
335  * there is no chance of the uvn being killed or locked out here.
336  *
337  * => caller must call with object unlocked.
338  * => caller must be using the same accessprot as was used at attach time
339  */
340 
341 
342 static void
343 uvn_reference(uobj)
344 	struct uvm_object *uobj;
345 {
346 #ifdef DEBUG
347 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
348 #endif
349 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
350 
351 	simple_lock(&uobj->vmobjlock);
352 #ifdef DEBUG
353 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
354 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
355 		    uobj->uo_refs);
356 		panic("uvn_reference: invalid state");
357 	}
358 #endif
359 	uobj->uo_refs++;
360 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
361 	uobj, uobj->uo_refs,0,0);
362 	simple_unlock(&uobj->vmobjlock);
363 }
364 
365 /*
366  * uvn_detach
367  *
368  * remove a reference to a VM object.
369  *
370  * => caller must call with object unlocked and map locked.
371  * => this starts the detach process, but doesn't have to finish it
372  *    (async i/o could still be pending).
373  */
374 static void
375 uvn_detach(uobj)
376 	struct uvm_object *uobj;
377 {
378 	struct uvm_vnode *uvn;
379 	struct vnode *vp;
380 	int oldflags;
381 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
382 
383 	simple_lock(&uobj->vmobjlock);
384 
385 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
386 	uobj->uo_refs--;			/* drop ref! */
387 	if (uobj->uo_refs) {			/* still more refs */
388 		simple_unlock(&uobj->vmobjlock);
389 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
390 		return;
391 	}
392 
393 	/*
394 	 * get other pointers ...
395 	 */
396 
397 	uvn = (struct uvm_vnode *) uobj;
398 	vp = (struct vnode *) uobj;
399 
400 	/*
401 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
402 	 * to keep an active text file from being overwritten).
403 	 */
404 	vp->v_flag &= ~VTEXT;
405 
406 	/*
407 	 * we just dropped the last reference to the uvn.   see if we can
408 	 * let it "stick around".
409 	 */
410 
411 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
412 		/* won't block */
413 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
414 		simple_unlock(&uobj->vmobjlock);
415 		vrele(vp);			/* drop vnode reference */
416 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
417 		return;
418 	}
419 
420 	/*
421 	 * its a goner!
422 	 */
423 
424 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
425 
426 	uvn->u_flags |= UVM_VNODE_DYING;
427 
428 	/*
429 	 * even though we may unlock in flush, no one can gain a reference
430 	 * to us until we clear the "dying" flag [because it blocks
431 	 * attaches].  we will not do that until after we've disposed of all
432 	 * the pages with uvn_flush().  note that before the flush the only
433 	 * pages that could be marked PG_BUSY are ones that are in async
434 	 * pageout by the daemon.  (there can't be any pending "get"'s
435 	 * because there are no references to the object).
436 	 */
437 
438 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
439 
440 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
441 
442 	/*
443 	 * given the structure of this pager, the above flush request will
444 	 * create the following state: all the pages that were in the object
445 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
446 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
447 	 * in either case, if we have pages left we can't kill the object
448 	 * yet because i/o is pending.  in this case we set the "relkill"
449 	 * flag which will cause pgo_releasepg to kill the object once all
450 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
451 	 * routine or from the page daemon].
452 	 */
453 
454 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
455 #ifdef DEBUG
456 		/*
457 		 * XXXCDC: very unlikely to happen until we have async i/o
458 		 * so print a little info message in case it does.
459 		 */
460 		printf("uvn_detach: vn %p has pages left after flush - "
461 		    "relkill mode\n", uobj);
462 #endif
463 		uvn->u_flags |= UVM_VNODE_RELKILL;
464 		simple_unlock(&uobj->vmobjlock);
465 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
466 		    0, 0);
467 		return;
468 	}
469 
470 	/*
471 	 * kill object now.   note that we can't be on the sync q because
472 	 * all references are gone.
473 	 */
474 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
475 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
476 		LIST_REMOVE(uvn, u_wlist);
477 		simple_unlock(&uvn_wl_lock);
478 	}
479 #ifdef DIAGNOSTIC
480 	if (uobj->memq.tqh_first != NULL)
481 		panic("uvn_deref: vnode VM object still has pages afer "
482 		    "syncio/free flush");
483 #endif
484 	oldflags = uvn->u_flags;
485 	uvn->u_flags = 0;
486 	simple_unlock(&uobj->vmobjlock);
487 
488 	/* wake up any sleepers */
489 	if (oldflags & UVM_VNODE_WANTED)
490 		wakeup(uvn);
491 
492 	/*
493 	 * drop our reference to the vnode.
494 	 */
495 	vrele(vp);
496 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
497 
498 	return;
499 }
500 
501 /*
502  * uvm_vnp_terminate: external hook to clear out a vnode's VM
503  *
504  * called in two cases:
505  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
506  *      count) needs to be freed so that a vnode can be reused.  this
507  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
508  *      the free list is still attached (i.e. not VBAD) then vgone is
509  *	called.   as part of the vgone trace this should get called to
510  *	free the vm object.   this is the common case.
511  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
512  *	"umount -f") the vgone() function is called on active vnodes
513  *	on the mounted file systems to kill their data (the vnodes become
514  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
515  *	call here (even if the uvn is still in use -- i.e. has a non-zero
516  *	reference count).  this case happens at "umount -f" and during a
517  *	"reboot/halt" operation.
518  *
519  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
520  *	[protects us from getting a vnode that is already in the DYING
521  *	 state...]
522  * => unlike uvn_detach, this function must not return until all the
523  *	uvn's pages are disposed of.
524  * => in case [2] the uvn is still alive after this call, but all I/O
525  *	ops will fail (due to the backing vnode now being "dead").  this
526  *	will prob. kill any process using the uvn due to pgo_get failing.
527  */
528 
529 void
530 uvm_vnp_terminate(vp)
531 	struct vnode *vp;
532 {
533 	struct uvm_vnode *uvn = &vp->v_uvm;
534 	int oldflags;
535 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
536 
537 	/*
538 	 * lock object and check if it is valid
539 	 */
540 	simple_lock(&uvn->u_obj.vmobjlock);
541 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
542 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
543 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
544 		simple_unlock(&uvn->u_obj.vmobjlock);
545 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
546 		return;
547 	}
548 
549 	/*
550 	 * must be a valid uvn that is not already dying (because XLOCK
551 	 * protects us from that).   the uvn can't in the ALOCK state
552 	 * because it is valid, and uvn's that are in the ALOCK state haven't
553 	 * been marked valid yet.
554 	 */
555 
556 #ifdef DEBUG
557 	/*
558 	 * debug check: are we yanking the vnode out from under our uvn?
559 	 */
560 	if (uvn->u_obj.uo_refs) {
561 		printf("uvm_vnp_terminate(%p): terminating active vnode "
562 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
563 	}
564 #endif
565 
566 	/*
567 	 * it is possible that the uvn was detached and is in the relkill
568 	 * state [i.e. waiting for async i/o to finish so that releasepg can
569 	 * kill object].  we take over the vnode now and cancel the relkill.
570 	 * we want to know when the i/o is done so we can recycle right
571 	 * away.   note that a uvn can only be in the RELKILL state if it
572 	 * has a zero reference count.
573 	 */
574 
575 	if (uvn->u_flags & UVM_VNODE_RELKILL)
576 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
577 
578 	/*
579 	 * block the uvn by setting the dying flag, and then flush the
580 	 * pages.  (note that flush may unlock object while doing I/O, but
581 	 * it will re-lock it before it returns control here).
582 	 *
583 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
584 	 * that uvn_io doesn't attempt to VOP_LOCK again.
585 	 *
586 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
587 	 *	due to a forceful unmount might not be a good idea.  maybe we
588 	 *	need a way to pass in this info to uvn_flush through a
589 	 *	pager-defined PGO_ constant [currently there are none].
590 	 */
591 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
592 
593 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
594 
595 	/*
596 	 * as we just did a flush we expect all the pages to be gone or in
597 	 * the process of going.  sleep to wait for the rest to go [via iosync].
598 	 */
599 
600 	while (uvn->u_obj.uo_npages) {
601 #ifdef DEBUG
602 		struct vm_page *pp;
603 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
604 		     pp = pp->listq.tqe_next) {
605 			if ((pp->flags & PG_BUSY) == 0)
606 				panic("uvm_vnp_terminate: detected unbusy pg");
607 		}
608 		if (uvn->u_nio == 0)
609 			panic("uvm_vnp_terminate: no I/O to wait for?");
610 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
611 		/*
612 		 * XXXCDC: this is unlikely to happen without async i/o so we
613 		 * put a printf in just to keep an eye on it.
614 		 */
615 #endif
616 		uvn->u_flags |= UVM_VNODE_IOSYNC;
617 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
618 		    "uvn_term",0);
619 		simple_lock(&uvn->u_obj.vmobjlock);
620 	}
621 
622 	/*
623 	 * done.   now we free the uvn if its reference count is zero
624 	 * (true if we are zapping a persisting uvn).   however, if we are
625 	 * terminating a uvn with active mappings we let it live ... future
626 	 * calls down to the vnode layer will fail.
627 	 */
628 
629 	oldflags = uvn->u_flags;
630 	if (uvn->u_obj.uo_refs) {
631 
632 		/*
633 		 * uvn must live on it is dead-vnode state until all references
634 		 * are gone.   restore flags.    clear CANPERSIST state.
635 		 */
636 
637 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
638 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
639 
640 	} else {
641 
642 		/*
643 		 * free the uvn now.   note that the VREF reference is already
644 		 * gone [it is dropped when we enter the persist state].
645 		 */
646 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
647 			panic("uvm_vnp_terminate: io sync wanted bit set");
648 
649 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
650 			simple_lock(&uvn_wl_lock);
651 			LIST_REMOVE(uvn, u_wlist);
652 			simple_unlock(&uvn_wl_lock);
653 		}
654 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
655 	}
656 
657 	if (oldflags & UVM_VNODE_WANTED)
658 		wakeup(uvn);		/* object lock still held */
659 
660 	simple_unlock(&uvn->u_obj.vmobjlock);
661 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
662 
663 }
664 
665 /*
666  * uvn_releasepg: handled a released page in a uvn
667  *
668  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
669  *	to dispose of.
670  * => caller must handled PG_WANTED case
671  * => called with page's object locked, pageq's unlocked
672  * => returns TRUE if page's object is still alive, FALSE if we
673  *	killed the page's object.    if we return TRUE, then we
674  *	return with the object locked.
675  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
676  *				with the page queues locked [for pagedaemon]
677  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
678  * => we kill the uvn if it is not referenced and we are suppose to
679  *	kill it ("relkill").
680  */
681 
682 boolean_t
683 uvn_releasepg(pg, nextpgp)
684 	struct vm_page *pg;
685 	struct vm_page **nextpgp;	/* OUT */
686 {
687 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
688 #ifdef DIAGNOSTIC
689 	if ((pg->flags & PG_RELEASED) == 0)
690 		panic("uvn_releasepg: page not released!");
691 #endif
692 
693 	/*
694 	 * dispose of the page [caller handles PG_WANTED]
695 	 */
696 	pmap_page_protect(pg, VM_PROT_NONE);
697 	uvm_lock_pageq();
698 	if (nextpgp)
699 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
700 	uvm_pagefree(pg);
701 	if (!nextpgp)
702 		uvm_unlock_pageq();
703 
704 	/*
705 	 * now see if we need to kill the object
706 	 */
707 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
708 		if (uvn->u_obj.uo_refs)
709 			panic("uvn_releasepg: kill flag set on referenced "
710 			    "object!");
711 		if (uvn->u_obj.uo_npages == 0) {
712 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
713 				simple_lock(&uvn_wl_lock);
714 				LIST_REMOVE(uvn, u_wlist);
715 				simple_unlock(&uvn_wl_lock);
716 			}
717 #ifdef DIAGNOSTIC
718 			if (uvn->u_obj.memq.tqh_first)
719 	panic("uvn_releasepg: pages in object with npages == 0");
720 #endif
721 			if (uvn->u_flags & UVM_VNODE_WANTED)
722 				/* still holding object lock */
723 				wakeup(uvn);
724 
725 			uvn->u_flags = 0;		/* DEAD! */
726 			simple_unlock(&uvn->u_obj.vmobjlock);
727 			return (FALSE);
728 		}
729 	}
730 	return (TRUE);
731 }
732 
733 /*
734  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
735  * through the buffer cache and allow I/O in any size.  These VOPs use
736  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
737  * go through the buffer cache or allow I/O sizes larger than a
738  * block].  we will eventually want to change this.
739  *
740  * issues to consider:
741  *   uvm provides the uvm_aiodesc structure for async i/o management.
742  * there are two tailq's in the uvm. structure... one for pending async
743  * i/o and one for "done" async i/o.   to do an async i/o one puts
744  * an aiodesc on the "pending" list (protected by splbio()), starts the
745  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
746  * some sort of "i/o done" function to be called (at splbio(), interrupt
747  * time).   this function should remove the aiodesc from the pending list
748  * and place it on the "done" list and wakeup the daemon.   the daemon
749  * will run at normal spl() and will remove all items from the "done"
750  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
751  * [in the old vm code, this was done by calling the "put" routine with
752  * null arguments which made the code harder to read and understand because
753  * you had one function ("put") doing two things.]
754  *
755  * so the current pager needs:
756  *   int uvn_aiodone(struct uvm_aiodesc *)
757  *
758  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
759  *	later collection.
760  * => called with pageq's locked by the daemon.
761  *
762  * general outline:
763  * - "try" to lock object.   if fail, just return (will try again later)
764  * - drop "u_nio" (this req is done!)
765  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
766  * - get "page" structures (atop?).
767  * - handle "wanted" pages
768  * - handle "released" pages [using pgo_releasepg]
769  *   >>> pgo_releasepg may kill the object
770  * dont forget to look at "object" wanted flag in all cases.
771  */
772 
773 
774 /*
775  * uvn_flush: flush pages out of a uvm object.
776  *
777  * => object should be locked by caller.   we may _unlock_ the object
778  *	if (and only if) we need to clean a page (PGO_CLEANIT).
779  *	we return with the object locked.
780  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
781  *	might want to unlock higher level resources (e.g. vm_map)
782  *	before calling flush.
783  * => if PGO_CLEANIT is not set, then we will neither unlock the object
784  *	or block.
785  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
786  *	for flushing.
787  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
788  *	that new pages are inserted on the tail end of the list.   thus,
789  *	we can make a complete pass through the object in one go by starting
790  *	at the head and working towards the tail (new pages are put in
791  *	front of us).
792  * => NOTE: we are allowed to lock the page queues, so the caller
793  *	must not be holding the lock on them [e.g. pagedaemon had
794  *	better not call us with the queues locked]
795  * => we return TRUE unless we encountered some sort of I/O error
796  *
797  * comment on "cleaning" object and PG_BUSY pages:
798  *	this routine is holding the lock on the object.   the only time
799  *	that it can run into a PG_BUSY page that it does not own is if
800  *	some other process has started I/O on the page (e.g. either
801  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
802  *	in, then it can not be dirty (!PG_CLEAN) because no one has
803  *	had a chance to modify it yet.    if the PG_BUSY page is being
804  *	paged out then it means that someone else has already started
805  *	cleaning the page for us (how nice!).    in this case, if we
806  *	have syncio specified, then after we make our pass through the
807  *	object we need to wait for the other PG_BUSY pages to clear
808  *	off (i.e. we need to do an iosync).   also note that once a
809  *	page is PG_BUSY it must stay in its object until it is un-busyed.
810  *
811  * note on page traversal:
812  *	we can traverse the pages in an object either by going down the
813  *	linked list in "uobj->memq", or we can go over the address range
814  *	by page doing hash table lookups for each address.    depending
815  *	on how many pages are in the object it may be cheaper to do one
816  *	or the other.   we set "by_list" to true if we are using memq.
817  *	if the cost of a hash lookup was equal to the cost of the list
818  *	traversal we could compare the number of pages in the start->stop
819  *	range to the total number of pages in the object.   however, it
820  *	seems that a hash table lookup is more expensive than the linked
821  *	list traversal, so we multiply the number of pages in the
822  *	start->stop range by a penalty which we define below.
823  */
824 
825 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
826 
827 static boolean_t
828 uvn_flush(uobj, start, stop, flags)
829 	struct uvm_object *uobj;
830 	voff_t start, stop;
831 	int flags;
832 {
833 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
834 	struct vm_page *pp, *ppnext, *ptmp;
835 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
836 	int npages, result, lcv;
837 	boolean_t retval, need_iosync, by_list, needs_clean, all;
838 	voff_t curoff;
839 	u_short pp_version;
840 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
841 
842 	curoff = 0;	/* XXX: shut up gcc */
843 	/*
844 	 * get init vals and determine how we are going to traverse object
845 	 */
846 
847 	need_iosync = FALSE;
848 	retval = TRUE;		/* return value */
849 	if (flags & PGO_ALLPAGES) {
850 		all = TRUE;
851 		by_list = TRUE;		/* always go by the list */
852 	} else {
853 		start = trunc_page(start);
854 		stop = round_page(stop);
855 #ifdef DEBUG
856 		if (stop > round_page(uvn->u_size))
857 			printf("uvn_flush: strange, got an out of range "
858 			    "flush (fixed)\n");
859 #endif
860 		all = FALSE;
861 		by_list = (uobj->uo_npages <=
862 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
863 	}
864 
865 	UVMHIST_LOG(maphist,
866 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
867 	    start, stop, by_list, flags);
868 
869 	/*
870 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
871 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
872 	 * is wrong it will only prevent us from clustering... it won't break
873 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
874 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
875 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
876 	 * bit is always up to date since there are no mappings).
877 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
878 	 */
879 
880 	if ((flags & PGO_CLEANIT) != 0 &&
881 	    uobj->pgops->pgo_mk_pcluster != NULL) {
882 		if (by_list) {
883 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
884 			    pp = pp->listq.tqe_next) {
885 				if (!all &&
886 				    (pp->offset < start || pp->offset >= stop))
887 					continue;
888 				pp->flags &= ~PG_CLEANCHK;
889 			}
890 
891 		} else {   /* by hash */
892 			for (curoff = start ; curoff < stop;
893 			    curoff += PAGE_SIZE) {
894 				pp = uvm_pagelookup(uobj, curoff);
895 				if (pp)
896 					pp->flags &= ~PG_CLEANCHK;
897 			}
898 		}
899 	}
900 
901 	/*
902 	 * now do it.   note: we must update ppnext in body of loop or we
903 	 * will get stuck.  we need to use ppnext because we may free "pp"
904 	 * before doing the next loop.
905 	 */
906 
907 	if (by_list) {
908 		pp = uobj->memq.tqh_first;
909 	} else {
910 		curoff = start;
911 		pp = uvm_pagelookup(uobj, curoff);
912 	}
913 
914 	ppnext = NULL;	/* XXX: shut up gcc */
915 	ppsp = NULL;		/* XXX: shut up gcc */
916 	uvm_lock_pageq();	/* page queues locked */
917 
918 	/* locked: both page queues and uobj */
919 	for ( ; (by_list && pp != NULL) ||
920 	  (!by_list && curoff < stop) ; pp = ppnext) {
921 
922 		if (by_list) {
923 
924 			/*
925 			 * range check
926 			 */
927 
928 			if (!all &&
929 			    (pp->offset < start || pp->offset >= stop)) {
930 				ppnext = pp->listq.tqe_next;
931 				continue;
932 			}
933 
934 		} else {
935 
936 			/*
937 			 * null check
938 			 */
939 
940 			curoff += PAGE_SIZE;
941 			if (pp == NULL) {
942 				if (curoff < stop)
943 					ppnext = uvm_pagelookup(uobj, curoff);
944 				continue;
945 			}
946 
947 		}
948 
949 		/*
950 		 * handle case where we do not need to clean page (either
951 		 * because we are not clean or because page is not dirty or
952 		 * is busy):
953 		 *
954 		 * NOTE: we are allowed to deactivate a non-wired active
955 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
956 		 * queue it must stay put until it is !PG_BUSY (so as not to
957 		 * confuse pagedaemon).
958 		 */
959 
960 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
961 			needs_clean = FALSE;
962 			if ((pp->flags & PG_BUSY) != 0 &&
963 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
964 			             (PGO_CLEANIT|PGO_SYNCIO))
965 				need_iosync = TRUE;
966 		} else {
967 			/*
968 			 * freeing: nuke all mappings so we can sync
969 			 * PG_CLEAN bit with no race
970 			 */
971 			if ((pp->flags & PG_CLEAN) != 0 &&
972 			    (flags & PGO_FREE) != 0 &&
973 			    (pp->pqflags & PQ_ACTIVE) != 0)
974 				pmap_page_protect(pp, VM_PROT_NONE);
975 			if ((pp->flags & PG_CLEAN) != 0 &&
976 			    pmap_is_modified(pp))
977 				pp->flags &= ~(PG_CLEAN);
978 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
979 
980 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
981 		}
982 
983 		/*
984 		 * if we don't need a clean... load ppnext and dispose of pp
985 		 */
986 		if (!needs_clean) {
987 			/* load ppnext */
988 			if (by_list)
989 				ppnext = pp->listq.tqe_next;
990 			else {
991 				if (curoff < stop)
992 					ppnext = uvm_pagelookup(uobj, curoff);
993 			}
994 
995 			/* now dispose of pp */
996 			if (flags & PGO_DEACTIVATE) {
997 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
998 				    pp->wire_count == 0) {
999 					pmap_page_protect(pp, VM_PROT_NONE);
1000 					uvm_pagedeactivate(pp);
1001 				}
1002 
1003 			} else if (flags & PGO_FREE) {
1004 				if (pp->flags & PG_BUSY) {
1005 					/* release busy pages */
1006 					pp->flags |= PG_RELEASED;
1007 				} else {
1008 					pmap_page_protect(pp, VM_PROT_NONE);
1009 					/* removed page from object */
1010 					uvm_pagefree(pp);
1011 				}
1012 			}
1013 			/* ppnext is valid so we can continue... */
1014 			continue;
1015 		}
1016 
1017 		/*
1018 		 * pp points to a page in the locked object that we are
1019 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
1020 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1021 		 *
1022 		 * let uvm_pager_put attempted a clustered page out.
1023 		 * note: locked: uobj and page queues.
1024 		 */
1025 
1026 		pp->flags |= PG_BUSY;	/* we 'own' page now */
1027 		UVM_PAGE_OWN(pp, "uvn_flush");
1028 		pmap_page_protect(pp, VM_PROT_READ);
1029 		pp_version = pp->version;
1030 ReTry:
1031 		ppsp = pps;
1032 		npages = sizeof(pps) / sizeof(struct vm_page *);
1033 
1034 		/* locked: page queues, uobj */
1035 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1036 			   flags | PGO_DOACTCLUST, start, stop);
1037 		/* unlocked: page queues, uobj */
1038 
1039 		/*
1040 		 * at this point nothing is locked.   if we did an async I/O
1041 		 * it is remotely possible for the async i/o to complete and
1042 		 * the page "pp" be freed or what not before we get a chance
1043 		 * to relock the object.   in order to detect this, we have
1044 		 * saved the version number of the page in "pp_version".
1045 		 */
1046 
1047 		/* relock! */
1048 		simple_lock(&uobj->vmobjlock);
1049 		uvm_lock_pageq();
1050 
1051 		/*
1052 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1053 		 * can only happen when  we are doing async I/O and can't
1054 		 * map the pages into kernel memory (pager_map) due to lack
1055 		 * of vm space.   if this happens we drop back to sync I/O.
1056 		 */
1057 
1058 		if (result == VM_PAGER_AGAIN) {
1059 			/*
1060 			 * it is unlikely, but page could have been released
1061 			 * while we had the object lock dropped.   we ignore
1062 			 * this now and retry the I/O.  we will detect and
1063 			 * handle the released page after the syncio I/O
1064 			 * completes.
1065 			 */
1066 #ifdef DIAGNOSTIC
1067 			if (flags & PGO_SYNCIO)
1068 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1069 #endif
1070 			flags |= PGO_SYNCIO;
1071 			goto ReTry;
1072 		}
1073 
1074 		/*
1075 		 * the cleaning operation is now done.   finish up.  note that
1076 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1077 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1078 		 * to us in ppsp/npages.
1079 		 */
1080 
1081 		/*
1082 		 * for pending async i/o if we are not deactivating/freeing
1083 		 * we can move on to the next page.
1084 		 */
1085 
1086 		if (result == VM_PAGER_PEND) {
1087 
1088 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1089 				/*
1090 				 * no per-page ops: refresh ppnext and continue
1091 				 */
1092 				if (by_list) {
1093 					if (pp->version == pp_version)
1094 						ppnext = pp->listq.tqe_next;
1095 					else
1096 						/* reset */
1097 						ppnext = uobj->memq.tqh_first;
1098 				} else {
1099 					if (curoff < stop)
1100 						ppnext = uvm_pagelookup(uobj,
1101 						    curoff);
1102 				}
1103 				continue;
1104 			}
1105 
1106 			/* need to do anything here? */
1107 		}
1108 
1109 		/*
1110 		 * need to look at each page of the I/O operation.  we defer
1111 		 * processing "pp" until the last trip through this "for" loop
1112 		 * so that we can load "ppnext" for the main loop after we
1113 		 * play with the cluster pages [thus the "npages + 1" in the
1114 		 * loop below].
1115 		 */
1116 
1117 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1118 
1119 			/*
1120 			 * handle ppnext for outside loop, and saving pp
1121 			 * until the end.
1122 			 */
1123 			if (lcv < npages) {
1124 				if (ppsp[lcv] == pp)
1125 					continue; /* skip pp until the end */
1126 				ptmp = ppsp[lcv];
1127 			} else {
1128 				ptmp = pp;
1129 
1130 				/* set up next page for outer loop */
1131 				if (by_list) {
1132 					if (pp->version == pp_version)
1133 						ppnext = pp->listq.tqe_next;
1134 					else
1135 						/* reset */
1136 						ppnext = uobj->memq.tqh_first;
1137 				} else {
1138 					if (curoff < stop)
1139 					ppnext = uvm_pagelookup(uobj, curoff);
1140 				}
1141 			}
1142 
1143 			/*
1144 			 * verify the page didn't get moved while obj was
1145 			 * unlocked
1146 			 */
1147 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1148 				continue;
1149 
1150 			/*
1151 			 * unbusy the page if I/O is done.   note that for
1152 			 * pending I/O it is possible that the I/O op
1153 			 * finished before we relocked the object (in
1154 			 * which case the page is no longer busy).
1155 			 */
1156 
1157 			if (result != VM_PAGER_PEND) {
1158 				if (ptmp->flags & PG_WANTED)
1159 					/* still holding object lock */
1160 					wakeup(ptmp);
1161 
1162 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1163 				UVM_PAGE_OWN(ptmp, NULL);
1164 				if (ptmp->flags & PG_RELEASED) {
1165 
1166 					/* pgo_releasepg wants this */
1167 					uvm_unlock_pageq();
1168 					if (!uvn_releasepg(ptmp, NULL))
1169 						return (TRUE);
1170 
1171 					uvm_lock_pageq();	/* relock */
1172 					continue;		/* next page */
1173 
1174 				} else {
1175 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1176 					if ((flags & PGO_FREE) == 0)
1177 						pmap_clear_modify(ptmp);
1178 				}
1179 			}
1180 
1181 			/*
1182 			 * dispose of page
1183 			 */
1184 
1185 			if (flags & PGO_DEACTIVATE) {
1186 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1187 				    pp->wire_count == 0) {
1188 					pmap_page_protect(ptmp, VM_PROT_NONE);
1189 					uvm_pagedeactivate(ptmp);
1190 				}
1191 
1192 			} else if (flags & PGO_FREE) {
1193 				if (result == VM_PAGER_PEND) {
1194 					if ((ptmp->flags & PG_BUSY) != 0)
1195 						/* signal for i/o done */
1196 						ptmp->flags |= PG_RELEASED;
1197 				} else {
1198 					if (result != VM_PAGER_OK) {
1199 						printf("uvn_flush: obj=%p, "
1200 						   "offset=0x%llx.  error "
1201 						   "during pageout.\n",
1202 						    pp->uobject,
1203 						    (long long)pp->offset);
1204 						printf("uvn_flush: WARNING: "
1205 						    "changes to page may be "
1206 						    "lost!\n");
1207 						retval = FALSE;
1208 					}
1209 					pmap_page_protect(ptmp, VM_PROT_NONE);
1210 					uvm_pagefree(ptmp);
1211 				}
1212 			}
1213 
1214 		}		/* end of "lcv" for loop */
1215 
1216 	}		/* end of "pp" for loop */
1217 
1218 	/*
1219 	 * done with pagequeues: unlock
1220 	 */
1221 	uvm_unlock_pageq();
1222 
1223 	/*
1224 	 * now wait for all I/O if required.
1225 	 */
1226 	if (need_iosync) {
1227 
1228 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1229 		while (uvn->u_nio != 0) {
1230 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1231 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1232 			  FALSE, "uvn_flush",0);
1233 			simple_lock(&uvn->u_obj.vmobjlock);
1234 		}
1235 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1236 			wakeup(&uvn->u_flags);
1237 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1238 	}
1239 
1240 	/* return, with object locked! */
1241 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1242 	return(retval);
1243 }
1244 
1245 /*
1246  * uvn_cluster
1247  *
1248  * we are about to do I/O in an object at offset.   this function is called
1249  * to establish a range of offsets around "offset" in which we can cluster
1250  * I/O.
1251  *
1252  * - currently doesn't matter if obj locked or not.
1253  */
1254 
1255 static void
1256 uvn_cluster(uobj, offset, loffset, hoffset)
1257 	struct uvm_object *uobj;
1258 	voff_t offset;
1259 	voff_t *loffset, *hoffset; /* OUT */
1260 {
1261 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1262 	*loffset = offset;
1263 
1264 	if (*loffset >= uvn->u_size)
1265 		panic("uvn_cluster: offset out of range");
1266 
1267 	/*
1268 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1269 	 */
1270 	*hoffset = *loffset + MAXBSIZE;
1271 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1272 		*hoffset = round_page(uvn->u_size);
1273 
1274 	return;
1275 }
1276 
1277 /*
1278  * uvn_put: flush page data to backing store.
1279  *
1280  * => prefer map unlocked (not required)
1281  * => object must be locked!   we will _unlock_ it before starting I/O.
1282  * => flags: PGO_SYNCIO -- use sync. I/O
1283  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1284  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1285  *	[thus we never do async i/o!  see iodone comment]
1286  */
1287 
1288 static int
1289 uvn_put(uobj, pps, npages, flags)
1290 	struct uvm_object *uobj;
1291 	struct vm_page **pps;
1292 	int npages, flags;
1293 {
1294 	int retval;
1295 
1296 	/* note: object locked */
1297 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1298 	/* note: object unlocked */
1299 
1300 	return(retval);
1301 }
1302 
1303 
1304 /*
1305  * uvn_get: get pages (synchronously) from backing store
1306  *
1307  * => prefer map unlocked (not required)
1308  * => object must be locked!  we will _unlock_ it before starting any I/O.
1309  * => flags: PGO_ALLPAGES: get all of the pages
1310  *           PGO_LOCKED: fault data structures are locked
1311  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1312  * => NOTE: caller must check for released pages!!
1313  */
1314 
1315 static int
1316 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1317 	struct uvm_object *uobj;
1318 	voff_t offset;
1319 	struct vm_page **pps;		/* IN/OUT */
1320 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1321 	int centeridx, advice, flags;
1322 	vm_prot_t access_type;
1323 {
1324 	voff_t current_offset;
1325 	struct vm_page *ptmp;
1326 	int lcv, result, gotpages;
1327 	boolean_t done;
1328 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1329 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1330 
1331 	/*
1332 	 * step 1: handled the case where fault data structures are locked.
1333 	 */
1334 
1335 	if (flags & PGO_LOCKED) {
1336 
1337 		/*
1338 		 * gotpages is the current number of pages we've gotten (which
1339 		 * we pass back up to caller via *npagesp.
1340 		 */
1341 
1342 		gotpages = 0;
1343 
1344 		/*
1345 		 * step 1a: get pages that are already resident.   only do this
1346 		 * if the data structures are locked (i.e. the first time
1347 		 * through).
1348 		 */
1349 
1350 		done = TRUE;	/* be optimistic */
1351 
1352 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1353 		    lcv++, current_offset += PAGE_SIZE) {
1354 
1355 			/* do we care about this page?  if not, skip it */
1356 			if (pps[lcv] == PGO_DONTCARE)
1357 				continue;
1358 
1359 			/* lookup page */
1360 			ptmp = uvm_pagelookup(uobj, current_offset);
1361 
1362 			/* to be useful must get a non-busy, non-released pg */
1363 			if (ptmp == NULL ||
1364 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1365 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1366 				    != 0)
1367 				done = FALSE;	/* need to do a wait or I/O! */
1368 				continue;
1369 			}
1370 
1371 			/*
1372 			 * useful page: busy/lock it and plug it in our
1373 			 * result array
1374 			 */
1375 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1376 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1377 			pps[lcv] = ptmp;
1378 			gotpages++;
1379 
1380 		}	/* "for" lcv loop */
1381 
1382 		/*
1383 		 * XXX: given the "advice", should we consider async read-ahead?
1384 		 * XXX: fault current does deactive of pages behind us.  is
1385 		 * this good (other callers might now).
1386 		 */
1387 		/*
1388 		 * XXX: read-ahead currently handled by buffer cache (bread)
1389 		 * level.
1390 		 * XXX: no async i/o available.
1391 		 * XXX: so we don't do anything now.
1392 		 */
1393 
1394 		/*
1395 		 * step 1c: now we've either done everything needed or we to
1396 		 * unlock and do some waiting or I/O.
1397 		 */
1398 
1399 		*npagesp = gotpages;		/* let caller know */
1400 		if (done)
1401 			return(VM_PAGER_OK);		/* bingo! */
1402 		else
1403 			/* EEK!   Need to unlock and I/O */
1404 			return(VM_PAGER_UNLOCK);
1405 	}
1406 
1407 	/*
1408 	 * step 2: get non-resident or busy pages.
1409 	 * object is locked.   data structures are unlocked.
1410 	 *
1411 	 * XXX: because we can't do async I/O at this level we get things
1412 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1413 	 * async-read-ahead for us at a lower level.
1414 	 */
1415 
1416 	for (lcv = 0, current_offset = offset ;
1417 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1418 
1419 		/* skip over pages we've already gotten or don't want */
1420 		/* skip over pages we don't _have_ to get */
1421 		if (pps[lcv] != NULL || (lcv != centeridx &&
1422 		    (flags & PGO_ALLPAGES) == 0))
1423 			continue;
1424 
1425 		/*
1426 		 * we have yet to locate the current page (pps[lcv]).   we first
1427 		 * look for a page that is already at the current offset.   if
1428 		 * we fine a page, we check to see if it is busy or released.
1429 		 * if that is the case, then we sleep on the page until it is
1430 		 * no longer busy or released and repeat the lookup.    if the
1431 		 * page we found is neither busy nor released, then we busy it
1432 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1433 		 * following while loop and indicates we are ready to move on
1434 		 * to the next page in the "lcv" loop above.
1435 		 *
1436 		 * if we exit the while loop with pps[lcv] still set to NULL,
1437 		 * then it means that we allocated a new busy/fake/clean page
1438 		 * ptmp in the object and we need to do I/O to fill in the data.
1439 		 */
1440 
1441 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1442 
1443 			/* look for a current page */
1444 			ptmp = uvm_pagelookup(uobj, current_offset);
1445 
1446 			/* nope?   allocate one now (if we can) */
1447 			if (ptmp == NULL) {
1448 
1449 				ptmp = uvm_pagealloc(uobj, current_offset,
1450 				    NULL, 0);
1451 
1452 				/* out of RAM? */
1453 				if (ptmp == NULL) {
1454 					simple_unlock(&uobj->vmobjlock);
1455 					uvm_wait("uvn_getpage");
1456 					simple_lock(&uobj->vmobjlock);
1457 
1458 					/* goto top of pps while loop */
1459 					continue;
1460 				}
1461 
1462 				/*
1463 				 * got new page ready for I/O.  break pps
1464 				 * while loop.  pps[lcv] is still NULL.
1465 				 */
1466 				break;
1467 			}
1468 
1469 			/* page is there, see if we need to wait on it */
1470 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1471 				ptmp->flags |= PG_WANTED;
1472 				UVM_UNLOCK_AND_WAIT(ptmp,
1473 				    &uobj->vmobjlock, FALSE, "uvn_get",0);
1474 				simple_lock(&uobj->vmobjlock);
1475 				continue;	/* goto top of pps while loop */
1476 			}
1477 
1478 			/*
1479 			 * if we get here then the page has become resident
1480 			 * and unbusy between steps 1 and 2.  we busy it
1481 			 * now (so we own it) and set pps[lcv] (so that we
1482 			 * exit the while loop).
1483 			 */
1484 			ptmp->flags |= PG_BUSY;
1485 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1486 			pps[lcv] = ptmp;
1487 		}
1488 
1489 		/*
1490 		 * if we own the a valid page at the correct offset, pps[lcv]
1491 		 * will point to it.   nothing more to do except go to the
1492 		 * next page.
1493 		 */
1494 
1495 		if (pps[lcv])
1496 			continue;			/* next lcv */
1497 
1498 		/*
1499 		 * we have a "fake/busy/clean" page that we just allocated.  do
1500 		 * I/O to fill it with valid data.  note that object must be
1501 		 * locked going into uvn_io, but will be unlocked afterwards.
1502 		 */
1503 
1504 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1505 		    PGO_SYNCIO, UIO_READ);
1506 
1507 		/*
1508 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1509 		 * syncio the result can not be PEND or AGAIN.   we must relock
1510 		 * and check for errors.
1511 		 */
1512 
1513 		/* lock object.   check for errors.   */
1514 		simple_lock(&uobj->vmobjlock);
1515 		if (result != VM_PAGER_OK) {
1516 			if (ptmp->flags & PG_WANTED)
1517 				/* object lock still held */
1518 				wakeup(ptmp);
1519 
1520 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1521 			UVM_PAGE_OWN(ptmp, NULL);
1522 			uvm_lock_pageq();
1523 			uvm_pagefree(ptmp);
1524 			uvm_unlock_pageq();
1525 			simple_unlock(&uobj->vmobjlock);
1526 			return(result);
1527 		}
1528 
1529 		/*
1530 		 * we got the page!   clear the fake flag (indicates valid
1531 		 * data now in page) and plug into our result array.   note
1532 		 * that page is still busy.
1533 		 *
1534 		 * it is the callers job to:
1535 		 * => check if the page is released
1536 		 * => unbusy the page
1537 		 * => activate the page
1538 		 */
1539 
1540 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1541 		pmap_clear_modify(ptmp);		/* ... and clean */
1542 		pps[lcv] = ptmp;
1543 
1544 	}	/* lcv loop */
1545 
1546 	/*
1547 	 * finally, unlock object and return.
1548 	 */
1549 
1550 	simple_unlock(&uobj->vmobjlock);
1551 	return (VM_PAGER_OK);
1552 }
1553 
1554 /*
1555  * uvn_asyncget: start async I/O to bring pages into ram
1556  *
1557  * => caller must lock object(???XXX: see if this is best)
1558  * => could be called from uvn_get or a madvise() fault-ahead.
1559  * => if it fails, it doesn't matter.
1560  */
1561 
1562 static int
1563 uvn_asyncget(uobj, offset, npages)
1564 	struct uvm_object *uobj;
1565 	voff_t offset;
1566 	int npages;
1567 {
1568 
1569 	/*
1570 	 * XXXCDC: we can't do async I/O yet
1571 	 */
1572 	printf("uvn_asyncget called\n");
1573 	return (KERN_SUCCESS);
1574 }
1575 
1576 /*
1577  * uvn_io: do I/O to a vnode
1578  *
1579  * => prefer map unlocked (not required)
1580  * => object must be locked!   we will _unlock_ it before starting I/O.
1581  * => flags: PGO_SYNCIO -- use sync. I/O
1582  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1583  *	[thus we never do async i/o!  see iodone comment]
1584  */
1585 
1586 static int
1587 uvn_io(uvn, pps, npages, flags, rw)
1588 	struct uvm_vnode *uvn;
1589 	vm_page_t *pps;
1590 	int npages, flags, rw;
1591 {
1592 	struct vnode *vn;
1593 	struct uio uio;
1594 	struct iovec iov;
1595 	vaddr_t kva;
1596 	off_t file_offset;
1597 	int waitf, result, mapinflags;
1598 	size_t got, wanted;
1599 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1600 
1601 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1602 
1603 	/*
1604 	 * init values
1605 	 */
1606 
1607 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1608 	vn = (struct vnode *) uvn;
1609 	file_offset = pps[0]->offset;
1610 
1611 	/*
1612 	 * check for sync'ing I/O.
1613 	 */
1614 
1615 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1616 		if (waitf == M_NOWAIT) {
1617 			simple_unlock(&uvn->u_obj.vmobjlock);
1618 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1619 			return(VM_PAGER_AGAIN);
1620 		}
1621 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1622 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1623 			FALSE, "uvn_iosync",0);
1624 		simple_lock(&uvn->u_obj.vmobjlock);
1625 	}
1626 
1627 	/*
1628 	 * check size
1629 	 */
1630 
1631 	if (file_offset >= uvn->u_size) {
1632 			simple_unlock(&uvn->u_obj.vmobjlock);
1633 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1634 			return(VM_PAGER_BAD);
1635 	}
1636 
1637 	/*
1638 	 * first try and map the pages in (without waiting)
1639 	 */
1640 
1641 	mapinflags = (rw == UIO_READ) ?
1642 	    UVMPAGER_MAPIN_READ : UVMPAGER_MAPIN_WRITE;
1643 
1644 	kva = uvm_pagermapin(pps, npages, NULL, mapinflags);
1645 	if (kva == 0 && waitf == M_NOWAIT) {
1646 		simple_unlock(&uvn->u_obj.vmobjlock);
1647 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1648 		return(VM_PAGER_AGAIN);
1649 	}
1650 
1651 	/*
1652 	 * ok, now bump u_nio up.   at this point we are done with uvn
1653 	 * and can unlock it.   if we still don't have a kva, try again
1654 	 * (this time with sleep ok).
1655 	 */
1656 
1657 	uvn->u_nio++;			/* we have an I/O in progress! */
1658 	simple_unlock(&uvn->u_obj.vmobjlock);
1659 	/* NOTE: object now unlocked */
1660 	if (kva == 0)
1661 		kva = uvm_pagermapin(pps, npages, NULL,
1662 		    mapinflags | UVMPAGER_MAPIN_WAITOK);
1663 
1664 	/*
1665 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1666 	 * get touched (so we can look at "offset" without having to lock
1667 	 * the object).  set up for I/O.
1668 	 */
1669 
1670 	/*
1671 	 * fill out uio/iov
1672 	 */
1673 
1674 	iov.iov_base = (caddr_t) kva;
1675 	wanted = npages << PAGE_SHIFT;
1676 	if (file_offset + wanted > uvn->u_size)
1677 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1678 	iov.iov_len = wanted;
1679 	uio.uio_iov = &iov;
1680 	uio.uio_iovcnt = 1;
1681 	uio.uio_offset = file_offset;
1682 	uio.uio_segflg = UIO_SYSSPACE;
1683 	uio.uio_rw = rw;
1684 	uio.uio_resid = wanted;
1685 	uio.uio_procp = NULL;
1686 
1687 	/*
1688 	 * do the I/O!  (XXX: curproc?)
1689 	 */
1690 
1691 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1692 
1693 	/*
1694 	 * This process may already have this vnode locked, if we faulted in
1695 	 * copyin() or copyout() on a region backed by this vnode
1696 	 * while doing I/O to the vnode.  If this is the case, don't
1697 	 * panic.. instead, return the error to the user.
1698 	 *
1699 	 * XXX this is a stopgap to prevent a panic.
1700 	 * Ideally, this kind of operation *should* work.
1701 	 */
1702 	result = 0;
1703 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1704 		result = vn_lock(vn, LK_EXCLUSIVE | LK_RETRY | LK_RECURSEFAIL);
1705 
1706 	if (result == 0) {
1707 		/* NOTE: vnode now locked! */
1708 
1709 		if (rw == UIO_READ)
1710 			result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1711 		else
1712 			result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1713 
1714 		if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1715 			VOP_UNLOCK(vn, 0);
1716 	}
1717 
1718 	/* NOTE: vnode now unlocked (unless vnislocked) */
1719 
1720 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1721 
1722 	/*
1723 	 * result == unix style errno (0 == OK!)
1724 	 *
1725 	 * zero out rest of buffer (if needed)
1726 	 */
1727 
1728 	if (result == 0) {
1729 		got = wanted - uio.uio_resid;
1730 
1731 		if (wanted && got == 0) {
1732 			result = EIO;		/* XXX: error? */
1733 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1734 			memset((void *) (kva + got), 0,
1735 			       (npages << PAGE_SHIFT) - got);
1736 		}
1737 	}
1738 
1739 	/*
1740 	 * now remove pager mapping
1741 	 */
1742 	uvm_pagermapout(kva, npages);
1743 
1744 	/*
1745 	 * now clean up the object (i.e. drop I/O count)
1746 	 */
1747 
1748 	simple_lock(&uvn->u_obj.vmobjlock);
1749 	/* NOTE: object now locked! */
1750 
1751 	uvn->u_nio--;			/* I/O DONE! */
1752 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1753 		wakeup(&uvn->u_nio);
1754 	}
1755 	simple_unlock(&uvn->u_obj.vmobjlock);
1756 	/* NOTE: object now unlocked! */
1757 
1758 	/*
1759 	 * done!
1760 	 */
1761 
1762 	UVMHIST_LOG(maphist, "<- done (result %d)", result,0,0,0);
1763 	if (result == 0)
1764 		return(VM_PAGER_OK);
1765 	else
1766 		return(VM_PAGER_ERROR);
1767 }
1768 
1769 /*
1770  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1771  * is gone we will kill the object (flushing dirty pages back to the vnode
1772  * if needed).
1773  *
1774  * => returns TRUE if there was no uvm_object attached or if there was
1775  *	one and we killed it [i.e. if there is no active uvn]
1776  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1777  *	needed]
1778  *
1779  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1780  *	having to hold a reference on the vnode) and given a working
1781  *	uvm_vnp_sync(), how does that effect the need for this function?
1782  *      [XXXCDC: seems like it can die?]
1783  *
1784  * => XXX: this function should DIE once we merge the VM and buffer
1785  *	cache.
1786  *
1787  * research shows that this is called in the following places:
1788  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1789  *	changes sizes
1790  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1791  *	are written to
1792  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1793  *	is off
1794  * ffs_realloccg: when we can't extend the current block and have
1795  *	to allocate a new one we call this [XXX: why?]
1796  * nfsrv_rename, rename_files: called when the target filename is there
1797  *	and we want to remove it
1798  * nfsrv_remove, sys_unlink: called on file we are removing
1799  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1800  *	then return "text busy"
1801  * nfs_open: seems to uncache any file opened with nfs
1802  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1803  */
1804 
1805 boolean_t
1806 uvm_vnp_uncache(vp)
1807 	struct vnode *vp;
1808 {
1809 	struct uvm_vnode *uvn = &vp->v_uvm;
1810 
1811 	/*
1812 	 * lock uvn part of the vnode and check to see if we need to do anything
1813 	 */
1814 
1815 	simple_lock(&uvn->u_obj.vmobjlock);
1816 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1817 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1818 		simple_unlock(&uvn->u_obj.vmobjlock);
1819 		return(TRUE);
1820 	}
1821 
1822 	/*
1823 	 * we have a valid, non-blocked uvn.   clear persist flag.
1824 	 * if uvn is currently active we can return now.
1825 	 */
1826 
1827 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1828 	if (uvn->u_obj.uo_refs) {
1829 		simple_unlock(&uvn->u_obj.vmobjlock);
1830 		return(FALSE);
1831 	}
1832 
1833 	/*
1834 	 * uvn is currently persisting!   we have to gain a reference to
1835 	 * it so that we can call uvn_detach to kill the uvn.
1836 	 */
1837 
1838 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1839 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1840 	simple_unlock(&uvn->u_obj.vmobjlock);
1841 
1842 
1843 #ifdef DEBUG
1844 	/*
1845 	 * carry over sanity check from old vnode pager: the vnode should
1846 	 * be VOP_LOCK'd, and we confirm it here.
1847 	 */
1848 	if (!VOP_ISLOCKED(vp)) {
1849 		boolean_t is_ok_anyway = FALSE;
1850 #ifdef NFS
1851 		extern int (**nfsv2_vnodeop_p) __P((void *));
1852 		extern int (**spec_nfsv2nodeop_p) __P((void *));
1853 		extern int (**fifo_nfsv2nodeop_p) __P((void *));
1854 
1855 		/* vnode is NOT VOP_LOCKed: some vnode types _never_ lock */
1856 		if (vp->v_op == nfsv2_vnodeop_p ||
1857 		    vp->v_op == spec_nfsv2nodeop_p) {
1858 			is_ok_anyway = TRUE;
1859 		}
1860 		if (vp->v_op == fifo_nfsv2nodeop_p) {
1861 			is_ok_anyway = TRUE;
1862 		}
1863 #endif	/* NFS */
1864 		if (!is_ok_anyway)
1865 			panic("uvm_vnp_uncache: vnode not locked!");
1866 	}
1867 #endif	/* DEBUG */
1868 
1869 	/*
1870 	 * now drop our reference to the vnode.   if we have the sole
1871 	 * reference to the vnode then this will cause it to die [as we
1872 	 * just cleared the persist flag].   we have to unlock the vnode
1873 	 * while we are doing this as it may trigger I/O.
1874 	 *
1875 	 * XXX: it might be possible for uvn to get reclaimed while we are
1876 	 * unlocked causing us to return TRUE when we should not.   we ignore
1877 	 * this as a false-positive return value doesn't hurt us.
1878 	 */
1879 	VOP_UNLOCK(vp, 0);
1880 	uvn_detach(&uvn->u_obj);
1881 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1882 
1883 	/*
1884 	 * and return...
1885 	 */
1886 
1887 	return(TRUE);
1888 }
1889 
1890 /*
1891  * uvm_vnp_setsize: grow or shrink a vnode uvn
1892  *
1893  * grow   => just update size value
1894  * shrink => toss un-needed pages
1895  *
1896  * => we assume that the caller has a reference of some sort to the
1897  *	vnode in question so that it will not be yanked out from under
1898  *	us.
1899  *
1900  * called from:
1901  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1902  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1903  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1904  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1905  *  => union fs: union_newsize
1906  */
1907 
1908 void
1909 uvm_vnp_setsize(vp, newsize)
1910 	struct vnode *vp;
1911 	voff_t newsize;
1912 {
1913 	struct uvm_vnode *uvn = &vp->v_uvm;
1914 
1915 	/*
1916 	 * lock uvn and check for valid object, and if valid: do it!
1917 	 */
1918 	simple_lock(&uvn->u_obj.vmobjlock);
1919 	if (uvn->u_flags & UVM_VNODE_VALID) {
1920 
1921 		/*
1922 		 * now check if the size has changed: if we shrink we had better
1923 		 * toss some pages...
1924 		 */
1925 
1926 		if (uvn->u_size > newsize) {
1927 			(void)uvn_flush(&uvn->u_obj, newsize,
1928 			    uvn->u_size, PGO_FREE);
1929 		}
1930 		uvn->u_size = newsize;
1931 	}
1932 	simple_unlock(&uvn->u_obj.vmobjlock);
1933 
1934 	/*
1935 	 * done
1936 	 */
1937 	return;
1938 }
1939 
1940 /*
1941  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1942  *
1943  * => called from sys_sync with no VM structures locked
1944  * => only one process can do a sync at a time (because the uvn
1945  *    structure only has one queue for sync'ing).  we ensure this
1946  *    by holding the uvn_sync_lock while the sync is in progress.
1947  *    other processes attempting a sync will sleep on this lock
1948  *    until we are done.
1949  */
1950 
1951 void
1952 uvm_vnp_sync(mp)
1953 	struct mount *mp;
1954 {
1955 	struct uvm_vnode *uvn;
1956 	struct vnode *vp;
1957 	boolean_t got_lock;
1958 
1959 	/*
1960 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1961 	 * our lock...
1962 	 */
1963 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0);
1964 
1965 	/*
1966 	 * step 2: build up a simpleq of uvns of interest based on the
1967 	 * write list.   we gain a reference to uvns of interest.  must
1968 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1969 	 * in the body of the loop.
1970 	 */
1971 	SIMPLEQ_INIT(&uvn_sync_q);
1972 	simple_lock(&uvn_wl_lock);
1973 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1974 	    uvn = uvn->u_wlist.le_next) {
1975 
1976 		vp = (struct vnode *) uvn;
1977 		if (mp && vp->v_mount != mp)
1978 			continue;
1979 
1980 		/* attempt to gain reference */
1981 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1982 		    						FALSE &&
1983 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1984 			/* spin */ ;
1985 
1986 		/*
1987 		 * we will exit the loop if either if the following are true:
1988 		 *  - we got the lock [always true if NCPU == 1]
1989 		 *  - we failed to get the lock but noticed the vnode was
1990 		 * 	"blocked" -- in this case the vnode must be a dying
1991 		 *	vnode, and since dying vnodes are in the process of
1992 		 *	being flushed out, we can safely skip this one
1993 		 *
1994 		 * we want to skip over the vnode if we did not get the lock,
1995 		 * or if the vnode is already dying (due to the above logic).
1996 		 *
1997 		 * note that uvn must already be valid because we found it on
1998 		 * the wlist (this also means it can't be ALOCK'd).
1999 		 */
2000 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
2001 			if (got_lock)
2002 				simple_unlock(&uvn->u_obj.vmobjlock);
2003 			continue;		/* skip it */
2004 		}
2005 
2006 		/*
2007 		 * gain reference.   watch out for persisting uvns (need to
2008 		 * regain vnode REF).
2009 		 */
2010 		if (uvn->u_obj.uo_refs == 0)
2011 			VREF(vp);
2012 		uvn->u_obj.uo_refs++;
2013 		simple_unlock(&uvn->u_obj.vmobjlock);
2014 
2015 		/*
2016 		 * got it!
2017 		 */
2018 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
2019 	}
2020 	simple_unlock(&uvn_wl_lock);
2021 
2022 	/*
2023 	 * step 3: we now have a list of uvn's that may need cleaning.
2024 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
2025 	 * (so we can now safely lock uvn's again).
2026 	 */
2027 
2028 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2029 		simple_lock(&uvn->u_obj.vmobjlock);
2030 #ifdef DEBUG
2031 		if (uvn->u_flags & UVM_VNODE_DYING) {
2032 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2033 		}
2034 #endif
2035 		uvn_flush(&uvn->u_obj, 0, 0,
2036 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2037 
2038 		/*
2039 		 * if we have the only reference and we just cleaned the uvn,
2040 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2041 		 * thus allowing us to avoid thinking about flushing it again
2042 		 * on later sync ops.
2043 		 */
2044 		if (uvn->u_obj.uo_refs == 1 &&
2045 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2046 			LIST_REMOVE(uvn, u_wlist);
2047 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2048 		}
2049 
2050 		simple_unlock(&uvn->u_obj.vmobjlock);
2051 
2052 		/* now drop our reference to the uvn */
2053 		uvn_detach(&uvn->u_obj);
2054 	}
2055 
2056 	/*
2057 	 * done!  release sync lock
2058 	 */
2059 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0);
2060 }
2061