xref: /openbsd-src/sys/uvm/uvm_vnode.c (revision 8445c53715e7030056b779e8ab40efb7820981f2)
1 /*	$OpenBSD: uvm_vnode.c,v 1.20 2001/09/11 20:05:26 miod 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 
68 #include <uvm/uvm.h>
69 #include <uvm/uvm_vnode.h>
70 
71 /*
72  * private global data structure
73  *
74  * we keep a list of writeable active vnode-backed VM objects for sync op.
75  * we keep a simpleq of vnodes that are currently being sync'd.
76  */
77 
78 LIST_HEAD(uvn_list_struct, uvm_vnode);
79 static struct uvn_list_struct uvn_wlist;	/* writeable uvns */
80 static simple_lock_data_t uvn_wl_lock;		/* locks uvn_wlist */
81 
82 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
83 static struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
84 lock_data_t uvn_sync_lock;			/* locks sync operation */
85 
86 /*
87  * functions
88  */
89 
90 static int		   uvn_asyncget __P((struct uvm_object *, voff_t,
91 					    int));
92 static void		   uvn_cluster __P((struct uvm_object *, voff_t,
93 					   voff_t *, voff_t *));
94 static void                uvn_detach __P((struct uvm_object *));
95 static boolean_t           uvn_flush __P((struct uvm_object *, voff_t,
96 					 voff_t, int));
97 static int                 uvn_get __P((struct uvm_object *, voff_t,
98 					vm_page_t *, int *, int,
99 					vm_prot_t, int, int));
100 static void		   uvn_init __P((void));
101 static int		   uvn_io __P((struct uvm_vnode *, vm_page_t *,
102 				      int, int, int));
103 static int		   uvn_put __P((struct uvm_object *, vm_page_t *,
104 					int, boolean_t));
105 static void                uvn_reference __P((struct uvm_object *));
106 static boolean_t	   uvn_releasepg __P((struct vm_page *,
107 					      struct vm_page **));
108 
109 /*
110  * master pager structure
111  */
112 
113 struct uvm_pagerops uvm_vnodeops = {
114 	uvn_init,
115 	uvn_reference,
116 	uvn_detach,
117 	NULL,			/* no specialized fault routine required */
118 	uvn_flush,
119 	uvn_get,
120 	uvn_asyncget,
121 	uvn_put,
122 	uvn_cluster,
123 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
124 	NULL,		 /* AIO-DONE function (not until we have asyncio) */
125 	uvn_releasepg,
126 };
127 
128 /*
129  * the ops!
130  */
131 
132 /*
133  * uvn_init
134  *
135  * init pager private data structures.
136  */
137 
138 static void
139 uvn_init()
140 {
141 
142 	LIST_INIT(&uvn_wlist);
143 	simple_lock_init(&uvn_wl_lock);
144 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
145 	lockinit(&uvn_sync_lock, PVM, "uvnsync", 0, 0);
146 }
147 
148 /*
149  * uvn_attach
150  *
151  * attach a vnode structure to a VM object.  if the vnode is already
152  * attached, then just bump the reference count by one and return the
153  * VM object.   if not already attached, attach and return the new VM obj.
154  * the "accessprot" tells the max access the attaching thread wants to
155  * our pages.
156  *
157  * => caller must _not_ already be holding the lock on the uvm_object.
158  * => in fact, nothing should be locked so that we can sleep here.
159  * => note that uvm_object is first thing in vnode structure, so their
160  *    pointers are equiv.
161  */
162 
163 struct uvm_object *
164 uvn_attach(arg, accessprot)
165 	void *arg;
166 	vm_prot_t accessprot;
167 {
168 	struct vnode *vp = arg;
169 	struct uvm_vnode *uvn = &vp->v_uvm;
170 	struct vattr vattr;
171 	int oldflags, result;
172 	struct partinfo pi;
173 	u_quad_t used_vnode_size;
174 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
175 
176 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
177 
178 	used_vnode_size = (u_quad_t)0;	/* XXX gcc -Wuninitialized */
179 
180 	/*
181 	 * first get a lock on the uvn.
182 	 */
183 	simple_lock(&uvn->u_obj.vmobjlock);
184 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
185 		printf("uvn_attach: blocked at 0x%p flags 0x%x\n",
186 		    uvn, uvn->u_flags);
187 		uvn->u_flags |= UVM_VNODE_WANTED;
188 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
189 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
190 		    "uvn_attach", 0);
191 		simple_lock(&uvn->u_obj.vmobjlock);
192 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
193 	}
194 
195 	/*
196 	 * if we're mapping a BLK device, make sure it is a disk.
197 	 */
198 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
199 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
200 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
201 		return(NULL);
202 	}
203 
204 	/*
205 	 * now we have lock and uvn must not be in a blocked state.
206 	 * first check to see if it is already active, in which case
207 	 * we can bump the reference count, check to see if we need to
208 	 * add it to the writeable list, and then return.
209 	 */
210 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
211 
212 		/* regain VREF if we were persisting */
213 		if (uvn->u_obj.uo_refs == 0) {
214 			VREF(vp);
215 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
216 			    0,0,0,0);
217 		}
218 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
219 
220 		/* check for new writeable uvn */
221 		if ((accessprot & VM_PROT_WRITE) != 0 &&
222 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
223 			simple_lock(&uvn_wl_lock);
224 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
225 			simple_unlock(&uvn_wl_lock);
226 			/* we are now on wlist! */
227 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
228 		}
229 
230 		/* unlock and return */
231 		simple_unlock(&uvn->u_obj.vmobjlock);
232 		UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
233 		    0, 0, 0);
234 		return (&uvn->u_obj);
235 	}
236 
237 	/*
238 	 * need to call VOP_GETATTR() to get the attributes, but that could
239 	 * block (due to I/O), so we want to unlock the object before calling.
240 	 * however, we want to keep anyone else from playing with the object
241 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
242 	 * prevents anyone from attaching to the vnode until we are done with
243 	 * it.
244 	 */
245 	uvn->u_flags = UVM_VNODE_ALOCK;
246 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
247 		/* XXX: curproc? */
248 
249 	if (vp->v_type == VBLK) {
250 		/*
251 		 * We could implement this as a specfs getattr call, but:
252 		 *
253 		 *	(1) VOP_GETATTR() would get the file system
254 		 *	    vnode operation, not the specfs operation.
255 		 *
256 		 *	(2) All we want is the size, anyhow.
257 		 */
258 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
259 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
260 		if (result == 0) {
261 			/* XXX should remember blocksize */
262 			used_vnode_size = (u_quad_t)pi.disklab->d_secsize *
263 			    (u_quad_t)pi.part->p_size;
264 		}
265 	} else {
266 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
267 		if (result == 0)
268 			used_vnode_size = vattr.va_size;
269 	}
270 
271 	/* relock object */
272 	simple_lock(&uvn->u_obj.vmobjlock);
273 
274 	if (result != 0) {
275 		if (uvn->u_flags & UVM_VNODE_WANTED)
276 			wakeup(uvn);
277 		uvn->u_flags = 0;
278 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
279 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
280 		return(NULL);
281 	}
282 
283 	/*
284 	 * make sure that the newsize fits within a vaddr_t
285 	 * XXX: need to revise addressing data types
286 	 */
287 #ifdef DEBUG
288 	if (vp->v_type == VBLK)
289 		printf("used_vnode_size = %llu\n", (long long)used_vnode_size);
290 #endif
291 
292 	/*
293 	 * now set up the uvn.
294 	 */
295 	uvn->u_obj.pgops = &uvm_vnodeops;
296 	TAILQ_INIT(&uvn->u_obj.memq);
297 	uvn->u_obj.uo_npages = 0;
298 	uvn->u_obj.uo_refs = 1;			/* just us... */
299 	oldflags = uvn->u_flags;
300 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
301 	uvn->u_nio = 0;
302 	uvn->u_size = used_vnode_size;
303 
304 	/* if write access, we need to add it to the wlist */
305 	if (accessprot & VM_PROT_WRITE) {
306 		simple_lock(&uvn_wl_lock);
307 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
308 		simple_unlock(&uvn_wl_lock);
309 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
310 	}
311 
312 	/*
313 	 * add a reference to the vnode.   this reference will stay as long
314 	 * as there is a valid mapping of the vnode.   dropped when the
315 	 * reference count goes to zero [and we either free or persist].
316 	 */
317 	VREF(vp);
318 	simple_unlock(&uvn->u_obj.vmobjlock);
319 	if (oldflags & UVM_VNODE_WANTED)
320 		wakeup(uvn);
321 
322 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
323 	return(&uvn->u_obj);
324 }
325 
326 
327 /*
328  * uvn_reference
329  *
330  * duplicate a reference to a VM object.  Note that the reference
331  * count must already be at least one (the passed in reference) so
332  * there is no chance of the uvn being killed or locked out here.
333  *
334  * => caller must call with object unlocked.
335  * => caller must be using the same accessprot as was used at attach time
336  */
337 
338 
339 static void
340 uvn_reference(uobj)
341 	struct uvm_object *uobj;
342 {
343 #ifdef DEBUG
344 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
345 #endif
346 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
347 
348 	simple_lock(&uobj->vmobjlock);
349 #ifdef DEBUG
350 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
351 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
352 		    uobj->uo_refs);
353 		panic("uvn_reference: invalid state");
354 	}
355 #endif
356 	uobj->uo_refs++;
357 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
358 	uobj, uobj->uo_refs,0,0);
359 	simple_unlock(&uobj->vmobjlock);
360 }
361 
362 /*
363  * uvn_detach
364  *
365  * remove a reference to a VM object.
366  *
367  * => caller must call with object unlocked and map locked.
368  * => this starts the detach process, but doesn't have to finish it
369  *    (async i/o could still be pending).
370  */
371 static void
372 uvn_detach(uobj)
373 	struct uvm_object *uobj;
374 {
375 	struct uvm_vnode *uvn;
376 	struct vnode *vp;
377 	int oldflags;
378 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
379 
380 	simple_lock(&uobj->vmobjlock);
381 
382 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
383 	uobj->uo_refs--;			/* drop ref! */
384 	if (uobj->uo_refs) {			/* still more refs */
385 		simple_unlock(&uobj->vmobjlock);
386 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
387 		return;
388 	}
389 
390 	/*
391 	 * get other pointers ...
392 	 */
393 
394 	uvn = (struct uvm_vnode *) uobj;
395 	vp = (struct vnode *) uobj;
396 
397 	/*
398 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
399 	 * to keep an active text file from being overwritten).
400 	 */
401 	vp->v_flag &= ~VTEXT;
402 
403 	/*
404 	 * we just dropped the last reference to the uvn.   see if we can
405 	 * let it "stick around".
406 	 */
407 
408 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
409 		/* won't block */
410 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
411 		simple_unlock(&uobj->vmobjlock);
412 		vrele(vp);			/* drop vnode reference */
413 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
414 		return;
415 	}
416 
417 	/*
418 	 * its a goner!
419 	 */
420 
421 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
422 
423 	uvn->u_flags |= UVM_VNODE_DYING;
424 
425 	/*
426 	 * even though we may unlock in flush, no one can gain a reference
427 	 * to us until we clear the "dying" flag [because it blocks
428 	 * attaches].  we will not do that until after we've disposed of all
429 	 * the pages with uvn_flush().  note that before the flush the only
430 	 * pages that could be marked PG_BUSY are ones that are in async
431 	 * pageout by the daemon.  (there can't be any pending "get"'s
432 	 * because there are no references to the object).
433 	 */
434 
435 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
436 
437 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
438 
439 	/*
440 	 * given the structure of this pager, the above flush request will
441 	 * create the following state: all the pages that were in the object
442 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
443 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
444 	 * in either case, if we have pages left we can't kill the object
445 	 * yet because i/o is pending.  in this case we set the "relkill"
446 	 * flag which will cause pgo_releasepg to kill the object once all
447 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
448 	 * routine or from the page daemon].
449 	 */
450 
451 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
452 #ifdef DEBUG
453 		/*
454 		 * XXXCDC: very unlikely to happen until we have async i/o
455 		 * so print a little info message in case it does.
456 		 */
457 		printf("uvn_detach: vn %p has pages left after flush - "
458 		    "relkill mode\n", uobj);
459 #endif
460 		uvn->u_flags |= UVM_VNODE_RELKILL;
461 		simple_unlock(&uobj->vmobjlock);
462 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
463 		    0, 0);
464 		return;
465 	}
466 
467 	/*
468 	 * kill object now.   note that we can't be on the sync q because
469 	 * all references are gone.
470 	 */
471 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
472 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
473 		LIST_REMOVE(uvn, u_wlist);
474 		simple_unlock(&uvn_wl_lock);
475 	}
476 #ifdef DIAGNOSTIC
477 	if (uobj->memq.tqh_first != NULL)
478 		panic("uvn_deref: vnode VM object still has pages afer "
479 		    "syncio/free flush");
480 #endif
481 	oldflags = uvn->u_flags;
482 	uvn->u_flags = 0;
483 	simple_unlock(&uobj->vmobjlock);
484 
485 	/* wake up any sleepers */
486 	if (oldflags & UVM_VNODE_WANTED)
487 		wakeup(uvn);
488 
489 	/*
490 	 * drop our reference to the vnode.
491 	 */
492 	vrele(vp);
493 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
494 
495 	return;
496 }
497 
498 /*
499  * uvm_vnp_terminate: external hook to clear out a vnode's VM
500  *
501  * called in two cases:
502  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
503  *      count) needs to be freed so that a vnode can be reused.  this
504  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
505  *      the free list is still attached (i.e. not VBAD) then vgone is
506  *	called.   as part of the vgone trace this should get called to
507  *	free the vm object.   this is the common case.
508  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
509  *	"umount -f") the vgone() function is called on active vnodes
510  *	on the mounted file systems to kill their data (the vnodes become
511  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
512  *	call here (even if the uvn is still in use -- i.e. has a non-zero
513  *	reference count).  this case happens at "umount -f" and during a
514  *	"reboot/halt" operation.
515  *
516  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
517  *	[protects us from getting a vnode that is already in the DYING
518  *	 state...]
519  * => unlike uvn_detach, this function must not return until all the
520  *	uvn's pages are disposed of.
521  * => in case [2] the uvn is still alive after this call, but all I/O
522  *	ops will fail (due to the backing vnode now being "dead").  this
523  *	will prob. kill any process using the uvn due to pgo_get failing.
524  */
525 
526 void
527 uvm_vnp_terminate(vp)
528 	struct vnode *vp;
529 {
530 	struct uvm_vnode *uvn = &vp->v_uvm;
531 	int oldflags;
532 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
533 
534 	/*
535 	 * lock object and check if it is valid
536 	 */
537 	simple_lock(&uvn->u_obj.vmobjlock);
538 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
539 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
540 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
541 		simple_unlock(&uvn->u_obj.vmobjlock);
542 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
543 		return;
544 	}
545 
546 	/*
547 	 * must be a valid uvn that is not already dying (because XLOCK
548 	 * protects us from that).   the uvn can't in the ALOCK state
549 	 * because it is valid, and uvn's that are in the ALOCK state haven't
550 	 * been marked valid yet.
551 	 */
552 
553 #ifdef DEBUG
554 	/*
555 	 * debug check: are we yanking the vnode out from under our uvn?
556 	 */
557 	if (uvn->u_obj.uo_refs) {
558 		printf("uvm_vnp_terminate(%p): terminating active vnode "
559 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
560 	}
561 #endif
562 
563 	/*
564 	 * it is possible that the uvn was detached and is in the relkill
565 	 * state [i.e. waiting for async i/o to finish so that releasepg can
566 	 * kill object].  we take over the vnode now and cancel the relkill.
567 	 * we want to know when the i/o is done so we can recycle right
568 	 * away.   note that a uvn can only be in the RELKILL state if it
569 	 * has a zero reference count.
570 	 */
571 
572 	if (uvn->u_flags & UVM_VNODE_RELKILL)
573 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
574 
575 	/*
576 	 * block the uvn by setting the dying flag, and then flush the
577 	 * pages.  (note that flush may unlock object while doing I/O, but
578 	 * it will re-lock it before it returns control here).
579 	 *
580 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
581 	 * that uvn_io doesn't attempt to VOP_LOCK again.
582 	 *
583 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
584 	 *	due to a forceful unmount might not be a good idea.  maybe we
585 	 *	need a way to pass in this info to uvn_flush through a
586 	 *	pager-defined PGO_ constant [currently there are none].
587 	 */
588 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
589 
590 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
591 
592 	/*
593 	 * as we just did a flush we expect all the pages to be gone or in
594 	 * the process of going.  sleep to wait for the rest to go [via iosync].
595 	 */
596 
597 	while (uvn->u_obj.uo_npages) {
598 #ifdef DEBUG
599 		struct vm_page *pp;
600 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
601 		     pp = pp->listq.tqe_next) {
602 			if ((pp->flags & PG_BUSY) == 0)
603 				panic("uvm_vnp_terminate: detected unbusy pg");
604 		}
605 		if (uvn->u_nio == 0)
606 			panic("uvm_vnp_terminate: no I/O to wait for?");
607 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
608 		/*
609 		 * XXXCDC: this is unlikely to happen without async i/o so we
610 		 * put a printf in just to keep an eye on it.
611 		 */
612 #endif
613 		uvn->u_flags |= UVM_VNODE_IOSYNC;
614 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
615 		    "uvn_term",0);
616 		simple_lock(&uvn->u_obj.vmobjlock);
617 	}
618 
619 	/*
620 	 * done.   now we free the uvn if its reference count is zero
621 	 * (true if we are zapping a persisting uvn).   however, if we are
622 	 * terminating a uvn with active mappings we let it live ... future
623 	 * calls down to the vnode layer will fail.
624 	 */
625 
626 	oldflags = uvn->u_flags;
627 	if (uvn->u_obj.uo_refs) {
628 
629 		/*
630 		 * uvn must live on it is dead-vnode state until all references
631 		 * are gone.   restore flags.    clear CANPERSIST state.
632 		 */
633 
634 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
635 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
636 
637 	} else {
638 
639 		/*
640 		 * free the uvn now.   note that the VREF reference is already
641 		 * gone [it is dropped when we enter the persist state].
642 		 */
643 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
644 			panic("uvm_vnp_terminate: io sync wanted bit set");
645 
646 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
647 			simple_lock(&uvn_wl_lock);
648 			LIST_REMOVE(uvn, u_wlist);
649 			simple_unlock(&uvn_wl_lock);
650 		}
651 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
652 	}
653 
654 	if (oldflags & UVM_VNODE_WANTED)
655 		wakeup(uvn);		/* object lock still held */
656 
657 	simple_unlock(&uvn->u_obj.vmobjlock);
658 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
659 
660 }
661 
662 /*
663  * uvn_releasepg: handled a released page in a uvn
664  *
665  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
666  *	to dispose of.
667  * => caller must handled PG_WANTED case
668  * => called with page's object locked, pageq's unlocked
669  * => returns TRUE if page's object is still alive, FALSE if we
670  *	killed the page's object.    if we return TRUE, then we
671  *	return with the object locked.
672  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
673  *				with the page queues locked [for pagedaemon]
674  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
675  * => we kill the uvn if it is not referenced and we are suppose to
676  *	kill it ("relkill").
677  */
678 
679 boolean_t
680 uvn_releasepg(pg, nextpgp)
681 	struct vm_page *pg;
682 	struct vm_page **nextpgp;	/* OUT */
683 {
684 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
685 #ifdef DIAGNOSTIC
686 	if ((pg->flags & PG_RELEASED) == 0)
687 		panic("uvn_releasepg: page not released!");
688 #endif
689 
690 	/*
691 	 * dispose of the page [caller handles PG_WANTED]
692 	 */
693 	pmap_page_protect(pg, VM_PROT_NONE);
694 	uvm_lock_pageq();
695 	if (nextpgp)
696 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
697 	uvm_pagefree(pg);
698 	if (!nextpgp)
699 		uvm_unlock_pageq();
700 
701 	/*
702 	 * now see if we need to kill the object
703 	 */
704 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
705 		if (uvn->u_obj.uo_refs)
706 			panic("uvn_releasepg: kill flag set on referenced "
707 			    "object!");
708 		if (uvn->u_obj.uo_npages == 0) {
709 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
710 				simple_lock(&uvn_wl_lock);
711 				LIST_REMOVE(uvn, u_wlist);
712 				simple_unlock(&uvn_wl_lock);
713 			}
714 #ifdef DIAGNOSTIC
715 			if (uvn->u_obj.memq.tqh_first)
716 	panic("uvn_releasepg: pages in object with npages == 0");
717 #endif
718 			if (uvn->u_flags & UVM_VNODE_WANTED)
719 				/* still holding object lock */
720 				wakeup(uvn);
721 
722 			uvn->u_flags = 0;		/* DEAD! */
723 			simple_unlock(&uvn->u_obj.vmobjlock);
724 			return (FALSE);
725 		}
726 	}
727 	return (TRUE);
728 }
729 
730 /*
731  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
732  * through the buffer cache and allow I/O in any size.  These VOPs use
733  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
734  * go through the buffer cache or allow I/O sizes larger than a
735  * block].  we will eventually want to change this.
736  *
737  * issues to consider:
738  *   uvm provides the uvm_aiodesc structure for async i/o management.
739  * there are two tailq's in the uvm. structure... one for pending async
740  * i/o and one for "done" async i/o.   to do an async i/o one puts
741  * an aiodesc on the "pending" list (protected by splbio()), starts the
742  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
743  * some sort of "i/o done" function to be called (at splbio(), interrupt
744  * time).   this function should remove the aiodesc from the pending list
745  * and place it on the "done" list and wakeup the daemon.   the daemon
746  * will run at normal spl() and will remove all items from the "done"
747  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
748  * [in the old vm code, this was done by calling the "put" routine with
749  * null arguments which made the code harder to read and understand because
750  * you had one function ("put") doing two things.]
751  *
752  * so the current pager needs:
753  *   int uvn_aiodone(struct uvm_aiodesc *)
754  *
755  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
756  *	later collection.
757  * => called with pageq's locked by the daemon.
758  *
759  * general outline:
760  * - "try" to lock object.   if fail, just return (will try again later)
761  * - drop "u_nio" (this req is done!)
762  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
763  * - get "page" structures (atop?).
764  * - handle "wanted" pages
765  * - handle "released" pages [using pgo_releasepg]
766  *   >>> pgo_releasepg may kill the object
767  * dont forget to look at "object" wanted flag in all cases.
768  */
769 
770 
771 /*
772  * uvn_flush: flush pages out of a uvm object.
773  *
774  * => object should be locked by caller.   we may _unlock_ the object
775  *	if (and only if) we need to clean a page (PGO_CLEANIT).
776  *	we return with the object locked.
777  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
778  *	might want to unlock higher level resources (e.g. vm_map)
779  *	before calling flush.
780  * => if PGO_CLEANIT is not set, then we will neither unlock the object
781  *	or block.
782  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
783  *	for flushing.
784  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
785  *	that new pages are inserted on the tail end of the list.   thus,
786  *	we can make a complete pass through the object in one go by starting
787  *	at the head and working towards the tail (new pages are put in
788  *	front of us).
789  * => NOTE: we are allowed to lock the page queues, so the caller
790  *	must not be holding the lock on them [e.g. pagedaemon had
791  *	better not call us with the queues locked]
792  * => we return TRUE unless we encountered some sort of I/O error
793  *
794  * comment on "cleaning" object and PG_BUSY pages:
795  *	this routine is holding the lock on the object.   the only time
796  *	that it can run into a PG_BUSY page that it does not own is if
797  *	some other process has started I/O on the page (e.g. either
798  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
799  *	in, then it can not be dirty (!PG_CLEAN) because no one has
800  *	had a chance to modify it yet.    if the PG_BUSY page is being
801  *	paged out then it means that someone else has already started
802  *	cleaning the page for us (how nice!).    in this case, if we
803  *	have syncio specified, then after we make our pass through the
804  *	object we need to wait for the other PG_BUSY pages to clear
805  *	off (i.e. we need to do an iosync).   also note that once a
806  *	page is PG_BUSY it must stay in its object until it is un-busyed.
807  *
808  * note on page traversal:
809  *	we can traverse the pages in an object either by going down the
810  *	linked list in "uobj->memq", or we can go over the address range
811  *	by page doing hash table lookups for each address.    depending
812  *	on how many pages are in the object it may be cheaper to do one
813  *	or the other.   we set "by_list" to true if we are using memq.
814  *	if the cost of a hash lookup was equal to the cost of the list
815  *	traversal we could compare the number of pages in the start->stop
816  *	range to the total number of pages in the object.   however, it
817  *	seems that a hash table lookup is more expensive than the linked
818  *	list traversal, so we multiply the number of pages in the
819  *	start->stop range by a penalty which we define below.
820  */
821 
822 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
823 
824 static boolean_t
825 uvn_flush(uobj, start, stop, flags)
826 	struct uvm_object *uobj;
827 	voff_t start, stop;
828 	int flags;
829 {
830 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
831 	struct vm_page *pp, *ppnext, *ptmp;
832 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
833 	int npages, result, lcv;
834 	boolean_t retval, need_iosync, by_list, needs_clean, all;
835 	voff_t curoff;
836 	u_short pp_version;
837 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
838 
839 	curoff = 0;	/* XXX: shut up gcc */
840 	/*
841 	 * get init vals and determine how we are going to traverse object
842 	 */
843 
844 	need_iosync = FALSE;
845 	retval = TRUE;		/* return value */
846 	if (flags & PGO_ALLPAGES) {
847 		all = TRUE;
848 		by_list = TRUE;		/* always go by the list */
849 	} else {
850 		start = trunc_page(start);
851 		stop = round_page(stop);
852 #ifdef DEBUG
853 		if (stop > round_page(uvn->u_size))
854 			printf("uvn_flush: strange, got an out of range "
855 			    "flush (fixed)\n");
856 #endif
857 		all = FALSE;
858 		by_list = (uobj->uo_npages <=
859 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
860 	}
861 
862 	UVMHIST_LOG(maphist,
863 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
864 	    start, stop, by_list, flags);
865 
866 	/*
867 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
868 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
869 	 * is wrong it will only prevent us from clustering... it won't break
870 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
871 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
872 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
873 	 * bit is always up to date since there are no mappings).
874 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
875 	 */
876 
877 	if ((flags & PGO_CLEANIT) != 0 &&
878 	    uobj->pgops->pgo_mk_pcluster != NULL) {
879 		if (by_list) {
880 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
881 			    pp = pp->listq.tqe_next) {
882 				if (!all &&
883 				    (pp->offset < start || pp->offset >= stop))
884 					continue;
885 				pp->flags &= ~PG_CLEANCHK;
886 			}
887 
888 		} else {   /* by hash */
889 			for (curoff = start ; curoff < stop;
890 			    curoff += PAGE_SIZE) {
891 				pp = uvm_pagelookup(uobj, curoff);
892 				if (pp)
893 					pp->flags &= ~PG_CLEANCHK;
894 			}
895 		}
896 	}
897 
898 	/*
899 	 * now do it.   note: we must update ppnext in body of loop or we
900 	 * will get stuck.  we need to use ppnext because we may free "pp"
901 	 * before doing the next loop.
902 	 */
903 
904 	if (by_list) {
905 		pp = uobj->memq.tqh_first;
906 	} else {
907 		curoff = start;
908 		pp = uvm_pagelookup(uobj, curoff);
909 	}
910 
911 	ppnext = NULL;	/* XXX: shut up gcc */
912 	ppsp = NULL;		/* XXX: shut up gcc */
913 	uvm_lock_pageq();	/* page queues locked */
914 
915 	/* locked: both page queues and uobj */
916 	for ( ; (by_list && pp != NULL) ||
917 	  (!by_list && curoff < stop) ; pp = ppnext) {
918 
919 		if (by_list) {
920 
921 			/*
922 			 * range check
923 			 */
924 
925 			if (!all &&
926 			    (pp->offset < start || pp->offset >= stop)) {
927 				ppnext = pp->listq.tqe_next;
928 				continue;
929 			}
930 
931 		} else {
932 
933 			/*
934 			 * null check
935 			 */
936 
937 			curoff += PAGE_SIZE;
938 			if (pp == NULL) {
939 				if (curoff < stop)
940 					ppnext = uvm_pagelookup(uobj, curoff);
941 				continue;
942 			}
943 
944 		}
945 
946 		/*
947 		 * handle case where we do not need to clean page (either
948 		 * because we are not clean or because page is not dirty or
949 		 * is busy):
950 		 *
951 		 * NOTE: we are allowed to deactivate a non-wired active
952 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
953 		 * queue it must stay put until it is !PG_BUSY (so as not to
954 		 * confuse pagedaemon).
955 		 */
956 
957 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
958 			needs_clean = FALSE;
959 			if ((pp->flags & PG_BUSY) != 0 &&
960 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
961 			             (PGO_CLEANIT|PGO_SYNCIO))
962 				need_iosync = TRUE;
963 		} else {
964 			/*
965 			 * freeing: nuke all mappings so we can sync
966 			 * PG_CLEAN bit with no race
967 			 */
968 			if ((pp->flags & PG_CLEAN) != 0 &&
969 			    (flags & PGO_FREE) != 0 &&
970 			    (pp->pqflags & PQ_ACTIVE) != 0)
971 				pmap_page_protect(pp, VM_PROT_NONE);
972 			if ((pp->flags & PG_CLEAN) != 0 &&
973 			    pmap_is_modified(pp))
974 				pp->flags &= ~(PG_CLEAN);
975 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
976 
977 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
978 		}
979 
980 		/*
981 		 * if we don't need a clean... load ppnext and dispose of pp
982 		 */
983 		if (!needs_clean) {
984 			/* load ppnext */
985 			if (by_list)
986 				ppnext = pp->listq.tqe_next;
987 			else {
988 				if (curoff < stop)
989 					ppnext = uvm_pagelookup(uobj, curoff);
990 			}
991 
992 			/* now dispose of pp */
993 			if (flags & PGO_DEACTIVATE) {
994 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
995 				    pp->wire_count == 0) {
996 					pmap_page_protect(pp, VM_PROT_NONE);
997 					uvm_pagedeactivate(pp);
998 				}
999 
1000 			} else if (flags & PGO_FREE) {
1001 				if (pp->flags & PG_BUSY) {
1002 					/* release busy pages */
1003 					pp->flags |= PG_RELEASED;
1004 				} else {
1005 					pmap_page_protect(pp, VM_PROT_NONE);
1006 					/* removed page from object */
1007 					uvm_pagefree(pp);
1008 				}
1009 			}
1010 			/* ppnext is valid so we can continue... */
1011 			continue;
1012 		}
1013 
1014 		/*
1015 		 * pp points to a page in the locked object that we are
1016 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
1017 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1018 		 *
1019 		 * let uvm_pager_put attempted a clustered page out.
1020 		 * note: locked: uobj and page queues.
1021 		 */
1022 
1023 		pp->flags |= PG_BUSY;	/* we 'own' page now */
1024 		UVM_PAGE_OWN(pp, "uvn_flush");
1025 		pmap_page_protect(pp, VM_PROT_READ);
1026 		pp_version = pp->version;
1027 ReTry:
1028 		ppsp = pps;
1029 		npages = sizeof(pps) / sizeof(struct vm_page *);
1030 
1031 		/* locked: page queues, uobj */
1032 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1033 			   flags | PGO_DOACTCLUST, start, stop);
1034 		/* unlocked: page queues, uobj */
1035 
1036 		/*
1037 		 * at this point nothing is locked.   if we did an async I/O
1038 		 * it is remotely possible for the async i/o to complete and
1039 		 * the page "pp" be freed or what not before we get a chance
1040 		 * to relock the object.   in order to detect this, we have
1041 		 * saved the version number of the page in "pp_version".
1042 		 */
1043 
1044 		/* relock! */
1045 		simple_lock(&uobj->vmobjlock);
1046 		uvm_lock_pageq();
1047 
1048 		/*
1049 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1050 		 * can only happen when  we are doing async I/O and can't
1051 		 * map the pages into kernel memory (pager_map) due to lack
1052 		 * of vm space.   if this happens we drop back to sync I/O.
1053 		 */
1054 
1055 		if (result == VM_PAGER_AGAIN) {
1056 			/*
1057 			 * it is unlikely, but page could have been released
1058 			 * while we had the object lock dropped.   we ignore
1059 			 * this now and retry the I/O.  we will detect and
1060 			 * handle the released page after the syncio I/O
1061 			 * completes.
1062 			 */
1063 #ifdef DIAGNOSTIC
1064 			if (flags & PGO_SYNCIO)
1065 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1066 #endif
1067 			flags |= PGO_SYNCIO;
1068 			goto ReTry;
1069 		}
1070 
1071 		/*
1072 		 * the cleaning operation is now done.   finish up.  note that
1073 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1074 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1075 		 * to us in ppsp/npages.
1076 		 */
1077 
1078 		/*
1079 		 * for pending async i/o if we are not deactivating/freeing
1080 		 * we can move on to the next page.
1081 		 */
1082 
1083 		if (result == VM_PAGER_PEND) {
1084 
1085 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1086 				/*
1087 				 * no per-page ops: refresh ppnext and continue
1088 				 */
1089 				if (by_list) {
1090 					if (pp->version == pp_version)
1091 						ppnext = pp->listq.tqe_next;
1092 					else
1093 						/* reset */
1094 						ppnext = uobj->memq.tqh_first;
1095 				} else {
1096 					if (curoff < stop)
1097 						ppnext = uvm_pagelookup(uobj,
1098 						    curoff);
1099 				}
1100 				continue;
1101 			}
1102 
1103 			/* need to do anything here? */
1104 		}
1105 
1106 		/*
1107 		 * need to look at each page of the I/O operation.  we defer
1108 		 * processing "pp" until the last trip through this "for" loop
1109 		 * so that we can load "ppnext" for the main loop after we
1110 		 * play with the cluster pages [thus the "npages + 1" in the
1111 		 * loop below].
1112 		 */
1113 
1114 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1115 
1116 			/*
1117 			 * handle ppnext for outside loop, and saving pp
1118 			 * until the end.
1119 			 */
1120 			if (lcv < npages) {
1121 				if (ppsp[lcv] == pp)
1122 					continue; /* skip pp until the end */
1123 				ptmp = ppsp[lcv];
1124 			} else {
1125 				ptmp = pp;
1126 
1127 				/* set up next page for outer loop */
1128 				if (by_list) {
1129 					if (pp->version == pp_version)
1130 						ppnext = pp->listq.tqe_next;
1131 					else
1132 						/* reset */
1133 						ppnext = uobj->memq.tqh_first;
1134 				} else {
1135 					if (curoff < stop)
1136 					ppnext = uvm_pagelookup(uobj, curoff);
1137 				}
1138 			}
1139 
1140 			/*
1141 			 * verify the page didn't get moved while obj was
1142 			 * unlocked
1143 			 */
1144 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1145 				continue;
1146 
1147 			/*
1148 			 * unbusy the page if I/O is done.   note that for
1149 			 * pending I/O it is possible that the I/O op
1150 			 * finished before we relocked the object (in
1151 			 * which case the page is no longer busy).
1152 			 */
1153 
1154 			if (result != VM_PAGER_PEND) {
1155 				if (ptmp->flags & PG_WANTED)
1156 					/* still holding object lock */
1157 					wakeup(ptmp);
1158 
1159 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1160 				UVM_PAGE_OWN(ptmp, NULL);
1161 				if (ptmp->flags & PG_RELEASED) {
1162 
1163 					/* pgo_releasepg wants this */
1164 					uvm_unlock_pageq();
1165 					if (!uvn_releasepg(ptmp, NULL))
1166 						return (TRUE);
1167 
1168 					uvm_lock_pageq();	/* relock */
1169 					continue;		/* next page */
1170 
1171 				} else {
1172 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1173 					if ((flags & PGO_FREE) == 0)
1174 						pmap_clear_modify(ptmp);
1175 				}
1176 			}
1177 
1178 			/*
1179 			 * dispose of page
1180 			 */
1181 
1182 			if (flags & PGO_DEACTIVATE) {
1183 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1184 				    pp->wire_count == 0) {
1185 					pmap_page_protect(ptmp, VM_PROT_NONE);
1186 					uvm_pagedeactivate(ptmp);
1187 				}
1188 
1189 			} else if (flags & PGO_FREE) {
1190 				if (result == VM_PAGER_PEND) {
1191 					if ((ptmp->flags & PG_BUSY) != 0)
1192 						/* signal for i/o done */
1193 						ptmp->flags |= PG_RELEASED;
1194 				} else {
1195 					if (result != VM_PAGER_OK) {
1196 						printf("uvn_flush: obj=%p, "
1197 						   "offset=0x%llx.  error "
1198 						   "during pageout.\n",
1199 						    pp->uobject,
1200 						    (long long)pp->offset);
1201 						printf("uvn_flush: WARNING: "
1202 						    "changes to page may be "
1203 						    "lost!\n");
1204 						retval = FALSE;
1205 					}
1206 					pmap_page_protect(ptmp, VM_PROT_NONE);
1207 					uvm_pagefree(ptmp);
1208 				}
1209 			}
1210 
1211 		}		/* end of "lcv" for loop */
1212 
1213 	}		/* end of "pp" for loop */
1214 
1215 	/*
1216 	 * done with pagequeues: unlock
1217 	 */
1218 	uvm_unlock_pageq();
1219 
1220 	/*
1221 	 * now wait for all I/O if required.
1222 	 */
1223 	if (need_iosync) {
1224 
1225 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1226 		while (uvn->u_nio != 0) {
1227 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1228 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1229 			  FALSE, "uvn_flush",0);
1230 			simple_lock(&uvn->u_obj.vmobjlock);
1231 		}
1232 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1233 			wakeup(&uvn->u_flags);
1234 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1235 	}
1236 
1237 	/* return, with object locked! */
1238 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1239 	return(retval);
1240 }
1241 
1242 /*
1243  * uvn_cluster
1244  *
1245  * we are about to do I/O in an object at offset.   this function is called
1246  * to establish a range of offsets around "offset" in which we can cluster
1247  * I/O.
1248  *
1249  * - currently doesn't matter if obj locked or not.
1250  */
1251 
1252 static void
1253 uvn_cluster(uobj, offset, loffset, hoffset)
1254 	struct uvm_object *uobj;
1255 	voff_t offset;
1256 	voff_t *loffset, *hoffset; /* OUT */
1257 {
1258 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1259 	*loffset = offset;
1260 
1261 	if (*loffset >= uvn->u_size)
1262 		panic("uvn_cluster: offset out of range");
1263 
1264 	/*
1265 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1266 	 */
1267 	*hoffset = *loffset + MAXBSIZE;
1268 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1269 		*hoffset = round_page(uvn->u_size);
1270 
1271 	return;
1272 }
1273 
1274 /*
1275  * uvn_put: flush page data to backing store.
1276  *
1277  * => prefer map unlocked (not required)
1278  * => object must be locked!   we will _unlock_ it before starting I/O.
1279  * => flags: PGO_SYNCIO -- use sync. I/O
1280  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1281  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1282  *	[thus we never do async i/o!  see iodone comment]
1283  */
1284 
1285 static int
1286 uvn_put(uobj, pps, npages, flags)
1287 	struct uvm_object *uobj;
1288 	struct vm_page **pps;
1289 	int npages, flags;
1290 {
1291 	int retval;
1292 
1293 	/* note: object locked */
1294 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1295 	/* note: object unlocked */
1296 
1297 	return(retval);
1298 }
1299 
1300 
1301 /*
1302  * uvn_get: get pages (synchronously) from backing store
1303  *
1304  * => prefer map unlocked (not required)
1305  * => object must be locked!  we will _unlock_ it before starting any I/O.
1306  * => flags: PGO_ALLPAGES: get all of the pages
1307  *           PGO_LOCKED: fault data structures are locked
1308  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1309  * => NOTE: caller must check for released pages!!
1310  */
1311 
1312 static int
1313 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1314 	struct uvm_object *uobj;
1315 	voff_t offset;
1316 	struct vm_page **pps;		/* IN/OUT */
1317 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1318 	int centeridx, advice, flags;
1319 	vm_prot_t access_type;
1320 {
1321 	voff_t current_offset;
1322 	struct vm_page *ptmp;
1323 	int lcv, result, gotpages;
1324 	boolean_t done;
1325 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1326 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1327 
1328 	/*
1329 	 * step 1: handled the case where fault data structures are locked.
1330 	 */
1331 
1332 	if (flags & PGO_LOCKED) {
1333 
1334 		/*
1335 		 * gotpages is the current number of pages we've gotten (which
1336 		 * we pass back up to caller via *npagesp.
1337 		 */
1338 
1339 		gotpages = 0;
1340 
1341 		/*
1342 		 * step 1a: get pages that are already resident.   only do this
1343 		 * if the data structures are locked (i.e. the first time
1344 		 * through).
1345 		 */
1346 
1347 		done = TRUE;	/* be optimistic */
1348 
1349 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1350 		    lcv++, current_offset += PAGE_SIZE) {
1351 
1352 			/* do we care about this page?  if not, skip it */
1353 			if (pps[lcv] == PGO_DONTCARE)
1354 				continue;
1355 
1356 			/* lookup page */
1357 			ptmp = uvm_pagelookup(uobj, current_offset);
1358 
1359 			/* to be useful must get a non-busy, non-released pg */
1360 			if (ptmp == NULL ||
1361 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1362 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1363 				    != 0)
1364 				done = FALSE;	/* need to do a wait or I/O! */
1365 				continue;
1366 			}
1367 
1368 			/*
1369 			 * useful page: busy/lock it and plug it in our
1370 			 * result array
1371 			 */
1372 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1373 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1374 			pps[lcv] = ptmp;
1375 			gotpages++;
1376 
1377 		}	/* "for" lcv loop */
1378 
1379 		/*
1380 		 * XXX: given the "advice", should we consider async read-ahead?
1381 		 * XXX: fault current does deactive of pages behind us.  is
1382 		 * this good (other callers might now).
1383 		 */
1384 		/*
1385 		 * XXX: read-ahead currently handled by buffer cache (bread)
1386 		 * level.
1387 		 * XXX: no async i/o available.
1388 		 * XXX: so we don't do anything now.
1389 		 */
1390 
1391 		/*
1392 		 * step 1c: now we've either done everything needed or we to
1393 		 * unlock and do some waiting or I/O.
1394 		 */
1395 
1396 		*npagesp = gotpages;		/* let caller know */
1397 		if (done)
1398 			return(VM_PAGER_OK);		/* bingo! */
1399 		else
1400 			/* EEK!   Need to unlock and I/O */
1401 			return(VM_PAGER_UNLOCK);
1402 	}
1403 
1404 	/*
1405 	 * step 2: get non-resident or busy pages.
1406 	 * object is locked.   data structures are unlocked.
1407 	 *
1408 	 * XXX: because we can't do async I/O at this level we get things
1409 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1410 	 * async-read-ahead for us at a lower level.
1411 	 */
1412 
1413 	for (lcv = 0, current_offset = offset ;
1414 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1415 
1416 		/* skip over pages we've already gotten or don't want */
1417 		/* skip over pages we don't _have_ to get */
1418 		if (pps[lcv] != NULL || (lcv != centeridx &&
1419 		    (flags & PGO_ALLPAGES) == 0))
1420 			continue;
1421 
1422 		/*
1423 		 * we have yet to locate the current page (pps[lcv]).   we first
1424 		 * look for a page that is already at the current offset.   if
1425 		 * we fine a page, we check to see if it is busy or released.
1426 		 * if that is the case, then we sleep on the page until it is
1427 		 * no longer busy or released and repeat the lookup.    if the
1428 		 * page we found is neither busy nor released, then we busy it
1429 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1430 		 * following while loop and indicates we are ready to move on
1431 		 * to the next page in the "lcv" loop above.
1432 		 *
1433 		 * if we exit the while loop with pps[lcv] still set to NULL,
1434 		 * then it means that we allocated a new busy/fake/clean page
1435 		 * ptmp in the object and we need to do I/O to fill in the data.
1436 		 */
1437 
1438 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1439 
1440 			/* look for a current page */
1441 			ptmp = uvm_pagelookup(uobj, current_offset);
1442 
1443 			/* nope?   allocate one now (if we can) */
1444 			if (ptmp == NULL) {
1445 
1446 				ptmp = uvm_pagealloc(uobj, current_offset,
1447 				    NULL, 0);
1448 
1449 				/* out of RAM? */
1450 				if (ptmp == NULL) {
1451 					simple_unlock(&uobj->vmobjlock);
1452 					uvm_wait("uvn_getpage");
1453 					simple_lock(&uobj->vmobjlock);
1454 
1455 					/* goto top of pps while loop */
1456 					continue;
1457 				}
1458 
1459 				/*
1460 				 * got new page ready for I/O.  break pps
1461 				 * while loop.  pps[lcv] is still NULL.
1462 				 */
1463 				break;
1464 			}
1465 
1466 			/* page is there, see if we need to wait on it */
1467 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1468 				ptmp->flags |= PG_WANTED;
1469 				UVM_UNLOCK_AND_WAIT(ptmp,
1470 				    &uobj->vmobjlock, FALSE, "uvn_get",0);
1471 				simple_lock(&uobj->vmobjlock);
1472 				continue;	/* goto top of pps while loop */
1473 			}
1474 
1475 			/*
1476 			 * if we get here then the page has become resident
1477 			 * and unbusy between steps 1 and 2.  we busy it
1478 			 * now (so we own it) and set pps[lcv] (so that we
1479 			 * exit the while loop).
1480 			 */
1481 			ptmp->flags |= PG_BUSY;
1482 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1483 			pps[lcv] = ptmp;
1484 		}
1485 
1486 		/*
1487 		 * if we own the a valid page at the correct offset, pps[lcv]
1488 		 * will point to it.   nothing more to do except go to the
1489 		 * next page.
1490 		 */
1491 
1492 		if (pps[lcv])
1493 			continue;			/* next lcv */
1494 
1495 		/*
1496 		 * we have a "fake/busy/clean" page that we just allocated.  do
1497 		 * I/O to fill it with valid data.  note that object must be
1498 		 * locked going into uvn_io, but will be unlocked afterwards.
1499 		 */
1500 
1501 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1502 		    PGO_SYNCIO, UIO_READ);
1503 
1504 		/*
1505 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1506 		 * syncio the result can not be PEND or AGAIN.   we must relock
1507 		 * and check for errors.
1508 		 */
1509 
1510 		/* lock object.   check for errors.   */
1511 		simple_lock(&uobj->vmobjlock);
1512 		if (result != VM_PAGER_OK) {
1513 			if (ptmp->flags & PG_WANTED)
1514 				/* object lock still held */
1515 				wakeup(ptmp);
1516 
1517 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1518 			UVM_PAGE_OWN(ptmp, NULL);
1519 			uvm_lock_pageq();
1520 			uvm_pagefree(ptmp);
1521 			uvm_unlock_pageq();
1522 			simple_unlock(&uobj->vmobjlock);
1523 			return(result);
1524 		}
1525 
1526 		/*
1527 		 * we got the page!   clear the fake flag (indicates valid
1528 		 * data now in page) and plug into our result array.   note
1529 		 * that page is still busy.
1530 		 *
1531 		 * it is the callers job to:
1532 		 * => check if the page is released
1533 		 * => unbusy the page
1534 		 * => activate the page
1535 		 */
1536 
1537 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1538 		pmap_clear_modify(ptmp);		/* ... and clean */
1539 		pps[lcv] = ptmp;
1540 
1541 	}	/* lcv loop */
1542 
1543 	/*
1544 	 * finally, unlock object and return.
1545 	 */
1546 
1547 	simple_unlock(&uobj->vmobjlock);
1548 	return (VM_PAGER_OK);
1549 }
1550 
1551 /*
1552  * uvn_asyncget: start async I/O to bring pages into ram
1553  *
1554  * => caller must lock object(???XXX: see if this is best)
1555  * => could be called from uvn_get or a madvise() fault-ahead.
1556  * => if it fails, it doesn't matter.
1557  */
1558 
1559 static int
1560 uvn_asyncget(uobj, offset, npages)
1561 	struct uvm_object *uobj;
1562 	voff_t offset;
1563 	int npages;
1564 {
1565 
1566 	/*
1567 	 * XXXCDC: we can't do async I/O yet
1568 	 */
1569 	printf("uvn_asyncget called\n");
1570 	return (KERN_SUCCESS);
1571 }
1572 
1573 /*
1574  * uvn_io: do I/O to a vnode
1575  *
1576  * => prefer map unlocked (not required)
1577  * => object must be locked!   we will _unlock_ it before starting I/O.
1578  * => flags: PGO_SYNCIO -- use sync. I/O
1579  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1580  *	[thus we never do async i/o!  see iodone comment]
1581  */
1582 
1583 static int
1584 uvn_io(uvn, pps, npages, flags, rw)
1585 	struct uvm_vnode *uvn;
1586 	vm_page_t *pps;
1587 	int npages, flags, rw;
1588 {
1589 	struct vnode *vn;
1590 	struct uio uio;
1591 	struct iovec iov;
1592 	vaddr_t kva;
1593 	off_t file_offset;
1594 	int waitf, result, mapinflags;
1595 	size_t got, wanted;
1596 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1597 
1598 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1599 
1600 	/*
1601 	 * init values
1602 	 */
1603 
1604 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1605 	vn = (struct vnode *) uvn;
1606 	file_offset = pps[0]->offset;
1607 
1608 	/*
1609 	 * check for sync'ing I/O.
1610 	 */
1611 
1612 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1613 		if (waitf == M_NOWAIT) {
1614 			simple_unlock(&uvn->u_obj.vmobjlock);
1615 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1616 			return(VM_PAGER_AGAIN);
1617 		}
1618 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1619 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1620 			FALSE, "uvn_iosync",0);
1621 		simple_lock(&uvn->u_obj.vmobjlock);
1622 	}
1623 
1624 	/*
1625 	 * check size
1626 	 */
1627 
1628 	if (file_offset >= uvn->u_size) {
1629 			simple_unlock(&uvn->u_obj.vmobjlock);
1630 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1631 			return(VM_PAGER_BAD);
1632 	}
1633 
1634 	/*
1635 	 * first try and map the pages in (without waiting)
1636 	 */
1637 
1638 	mapinflags = (rw == UIO_READ) ?
1639 	    UVMPAGER_MAPIN_READ : UVMPAGER_MAPIN_WRITE;
1640 
1641 	kva = uvm_pagermapin(pps, npages, NULL, mapinflags);
1642 	if (kva == 0 && waitf == M_NOWAIT) {
1643 		simple_unlock(&uvn->u_obj.vmobjlock);
1644 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1645 		return(VM_PAGER_AGAIN);
1646 	}
1647 
1648 	/*
1649 	 * ok, now bump u_nio up.   at this point we are done with uvn
1650 	 * and can unlock it.   if we still don't have a kva, try again
1651 	 * (this time with sleep ok).
1652 	 */
1653 
1654 	uvn->u_nio++;			/* we have an I/O in progress! */
1655 	simple_unlock(&uvn->u_obj.vmobjlock);
1656 	/* NOTE: object now unlocked */
1657 	if (kva == 0)
1658 		kva = uvm_pagermapin(pps, npages, NULL,
1659 		    mapinflags | UVMPAGER_MAPIN_WAITOK);
1660 
1661 	/*
1662 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1663 	 * get touched (so we can look at "offset" without having to lock
1664 	 * the object).  set up for I/O.
1665 	 */
1666 
1667 	/*
1668 	 * fill out uio/iov
1669 	 */
1670 
1671 	iov.iov_base = (caddr_t) kva;
1672 	wanted = npages << PAGE_SHIFT;
1673 	if (file_offset + wanted > uvn->u_size)
1674 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1675 	iov.iov_len = wanted;
1676 	uio.uio_iov = &iov;
1677 	uio.uio_iovcnt = 1;
1678 	uio.uio_offset = file_offset;
1679 	uio.uio_segflg = UIO_SYSSPACE;
1680 	uio.uio_rw = rw;
1681 	uio.uio_resid = wanted;
1682 	uio.uio_procp = curproc;
1683 
1684 	/*
1685 	 * do the I/O!  (XXX: curproc?)
1686 	 */
1687 
1688 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1689 
1690 	/*
1691 	 * This process may already have this vnode locked, if we faulted in
1692 	 * copyin() or copyout() on a region backed by this vnode
1693 	 * while doing I/O to the vnode.  If this is the case, don't
1694 	 * panic.. instead, return the error to the user.
1695 	 *
1696 	 * XXX this is a stopgap to prevent a panic.
1697 	 * Ideally, this kind of operation *should* work.
1698 	 */
1699 	result = 0;
1700 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1701 		result = vn_lock(vn, LK_EXCLUSIVE | LK_RETRY, curproc /*XXX*/);
1702 
1703 	if (result == 0) {
1704 		/* NOTE: vnode now locked! */
1705 
1706 		if (rw == UIO_READ)
1707 			result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1708 		else
1709 			result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1710 
1711 		if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1712 			VOP_UNLOCK(vn, 0, curproc /*XXX*/);
1713 	}
1714 
1715 	/* NOTE: vnode now unlocked (unless vnislocked) */
1716 
1717 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1718 
1719 	/*
1720 	 * result == unix style errno (0 == OK!)
1721 	 *
1722 	 * zero out rest of buffer (if needed)
1723 	 */
1724 
1725 	if (result == 0) {
1726 		got = wanted - uio.uio_resid;
1727 
1728 		if (wanted && got == 0) {
1729 			result = EIO;		/* XXX: error? */
1730 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1731 			memset((void *) (kva + got), 0,
1732 			       (npages << PAGE_SHIFT) - got);
1733 		}
1734 	}
1735 
1736 	/*
1737 	 * now remove pager mapping
1738 	 */
1739 	uvm_pagermapout(kva, npages);
1740 
1741 	/*
1742 	 * now clean up the object (i.e. drop I/O count)
1743 	 */
1744 
1745 	simple_lock(&uvn->u_obj.vmobjlock);
1746 	/* NOTE: object now locked! */
1747 
1748 	uvn->u_nio--;			/* I/O DONE! */
1749 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1750 		wakeup(&uvn->u_nio);
1751 	}
1752 	simple_unlock(&uvn->u_obj.vmobjlock);
1753 	/* NOTE: object now unlocked! */
1754 
1755 	/*
1756 	 * done!
1757 	 */
1758 
1759 	UVMHIST_LOG(maphist, "<- done (result %d)", result,0,0,0);
1760 	if (result == 0)
1761 		return(VM_PAGER_OK);
1762 	else
1763 		return(VM_PAGER_ERROR);
1764 }
1765 
1766 /*
1767  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1768  * is gone we will kill the object (flushing dirty pages back to the vnode
1769  * if needed).
1770  *
1771  * => returns TRUE if there was no uvm_object attached or if there was
1772  *	one and we killed it [i.e. if there is no active uvn]
1773  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1774  *	needed]
1775  *
1776  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1777  *	having to hold a reference on the vnode) and given a working
1778  *	uvm_vnp_sync(), how does that effect the need for this function?
1779  *      [XXXCDC: seems like it can die?]
1780  *
1781  * => XXX: this function should DIE once we merge the VM and buffer
1782  *	cache.
1783  *
1784  * research shows that this is called in the following places:
1785  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1786  *	changes sizes
1787  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1788  *	are written to
1789  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1790  *	is off
1791  * ffs_realloccg: when we can't extend the current block and have
1792  *	to allocate a new one we call this [XXX: why?]
1793  * nfsrv_rename, rename_files: called when the target filename is there
1794  *	and we want to remove it
1795  * nfsrv_remove, sys_unlink: called on file we are removing
1796  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1797  *	then return "text busy"
1798  * nfs_open: seems to uncache any file opened with nfs
1799  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1800  */
1801 
1802 boolean_t
1803 uvm_vnp_uncache(vp)
1804 	struct vnode *vp;
1805 {
1806 	struct uvm_vnode *uvn = &vp->v_uvm;
1807 
1808 	/*
1809 	 * lock uvn part of the vnode and check to see if we need to do anything
1810 	 */
1811 
1812 	simple_lock(&uvn->u_obj.vmobjlock);
1813 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1814 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1815 		simple_unlock(&uvn->u_obj.vmobjlock);
1816 		return(TRUE);
1817 	}
1818 
1819 	/*
1820 	 * we have a valid, non-blocked uvn.   clear persist flag.
1821 	 * if uvn is currently active we can return now.
1822 	 */
1823 
1824 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1825 	if (uvn->u_obj.uo_refs) {
1826 		simple_unlock(&uvn->u_obj.vmobjlock);
1827 		return(FALSE);
1828 	}
1829 
1830 	/*
1831 	 * uvn is currently persisting!   we have to gain a reference to
1832 	 * it so that we can call uvn_detach to kill the uvn.
1833 	 */
1834 
1835 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1836 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1837 	simple_unlock(&uvn->u_obj.vmobjlock);
1838 
1839 
1840 #ifdef DEBUG
1841 	/*
1842 	 * carry over sanity check from old vnode pager: the vnode should
1843 	 * be VOP_LOCK'd, and we confirm it here.
1844 	 */
1845 	if (!VOP_ISLOCKED(vp)) {
1846 		boolean_t is_ok_anyway = FALSE;
1847 #if defined(NFSCLIENT)
1848 		extern int (**nfsv2_vnodeop_p) __P((void *));
1849 		extern int (**spec_nfsv2nodeop_p) __P((void *));
1850 		extern int (**fifo_nfsv2nodeop_p) __P((void *));
1851 
1852 		/* vnode is NOT VOP_LOCKed: some vnode types _never_ lock */
1853 		if (vp->v_op == nfsv2_vnodeop_p ||
1854 		    vp->v_op == spec_nfsv2nodeop_p) {
1855 			is_ok_anyway = TRUE;
1856 		}
1857 		if (vp->v_op == fifo_nfsv2nodeop_p) {
1858 			is_ok_anyway = TRUE;
1859 		}
1860 #endif	/* defined(NFSSERVER) || defined(NFSCLIENT) */
1861 		if (!is_ok_anyway)
1862 			panic("uvm_vnp_uncache: vnode not locked!");
1863 	}
1864 #endif	/* DEBUG */
1865 
1866 	/*
1867 	 * now drop our reference to the vnode.   if we have the sole
1868 	 * reference to the vnode then this will cause it to die [as we
1869 	 * just cleared the persist flag].   we have to unlock the vnode
1870 	 * while we are doing this as it may trigger I/O.
1871 	 *
1872 	 * XXX: it might be possible for uvn to get reclaimed while we are
1873 	 * unlocked causing us to return TRUE when we should not.   we ignore
1874 	 * this as a false-positive return value doesn't hurt us.
1875 	 */
1876 	VOP_UNLOCK(vp, 0, curproc /*XXX*/);
1877 	uvn_detach(&uvn->u_obj);
1878 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curproc/*XXX*/);
1879 
1880 	/*
1881 	 * and return...
1882 	 */
1883 
1884 	return(TRUE);
1885 }
1886 
1887 /*
1888  * uvm_vnp_setsize: grow or shrink a vnode uvn
1889  *
1890  * grow   => just update size value
1891  * shrink => toss un-needed pages
1892  *
1893  * => we assume that the caller has a reference of some sort to the
1894  *	vnode in question so that it will not be yanked out from under
1895  *	us.
1896  *
1897  * called from:
1898  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1899  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1900  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1901  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1902  *  => union fs: union_newsize
1903  */
1904 
1905 void
1906 uvm_vnp_setsize(vp, newsize)
1907 	struct vnode *vp;
1908 	voff_t newsize;
1909 {
1910 	struct uvm_vnode *uvn = &vp->v_uvm;
1911 
1912 	/*
1913 	 * lock uvn and check for valid object, and if valid: do it!
1914 	 */
1915 	simple_lock(&uvn->u_obj.vmobjlock);
1916 	if (uvn->u_flags & UVM_VNODE_VALID) {
1917 
1918 		/*
1919 		 * now check if the size has changed: if we shrink we had better
1920 		 * toss some pages...
1921 		 */
1922 
1923 		if (uvn->u_size > newsize) {
1924 			(void)uvn_flush(&uvn->u_obj, newsize,
1925 			    uvn->u_size, PGO_FREE);
1926 		}
1927 		uvn->u_size = newsize;
1928 	}
1929 	simple_unlock(&uvn->u_obj.vmobjlock);
1930 
1931 	/*
1932 	 * done
1933 	 */
1934 	return;
1935 }
1936 
1937 /*
1938  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1939  *
1940  * => called from sys_sync with no VM structures locked
1941  * => only one process can do a sync at a time (because the uvn
1942  *    structure only has one queue for sync'ing).  we ensure this
1943  *    by holding the uvn_sync_lock while the sync is in progress.
1944  *    other processes attempting a sync will sleep on this lock
1945  *    until we are done.
1946  */
1947 
1948 void
1949 uvm_vnp_sync(mp)
1950 	struct mount *mp;
1951 {
1952 	struct uvm_vnode *uvn;
1953 	struct vnode *vp;
1954 	boolean_t got_lock;
1955 
1956 	/*
1957 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1958 	 * our lock...
1959 	 */
1960 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0, curproc /*XXX*/);
1961 
1962 	/*
1963 	 * step 2: build up a simpleq of uvns of interest based on the
1964 	 * write list.   we gain a reference to uvns of interest.  must
1965 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1966 	 * in the body of the loop.
1967 	 */
1968 	SIMPLEQ_INIT(&uvn_sync_q);
1969 	simple_lock(&uvn_wl_lock);
1970 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1971 	    uvn = uvn->u_wlist.le_next) {
1972 
1973 		vp = (struct vnode *) uvn;
1974 		if (mp && vp->v_mount != mp)
1975 			continue;
1976 
1977 		/* attempt to gain reference */
1978 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1979 		    						FALSE &&
1980 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1981 			/* spin */ ;
1982 
1983 		/*
1984 		 * we will exit the loop if either if the following are true:
1985 		 *  - we got the lock [always true if NCPU == 1]
1986 		 *  - we failed to get the lock but noticed the vnode was
1987 		 * 	"blocked" -- in this case the vnode must be a dying
1988 		 *	vnode, and since dying vnodes are in the process of
1989 		 *	being flushed out, we can safely skip this one
1990 		 *
1991 		 * we want to skip over the vnode if we did not get the lock,
1992 		 * or if the vnode is already dying (due to the above logic).
1993 		 *
1994 		 * note that uvn must already be valid because we found it on
1995 		 * the wlist (this also means it can't be ALOCK'd).
1996 		 */
1997 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1998 			if (got_lock)
1999 				simple_unlock(&uvn->u_obj.vmobjlock);
2000 			continue;		/* skip it */
2001 		}
2002 
2003 		/*
2004 		 * gain reference.   watch out for persisting uvns (need to
2005 		 * regain vnode REF).
2006 		 */
2007 		if (uvn->u_obj.uo_refs == 0)
2008 			VREF(vp);
2009 		uvn->u_obj.uo_refs++;
2010 		simple_unlock(&uvn->u_obj.vmobjlock);
2011 
2012 		/*
2013 		 * got it!
2014 		 */
2015 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
2016 	}
2017 	simple_unlock(&uvn_wl_lock);
2018 
2019 	/*
2020 	 * step 3: we now have a list of uvn's that may need cleaning.
2021 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
2022 	 * (so we can now safely lock uvn's again).
2023 	 */
2024 
2025 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2026 		simple_lock(&uvn->u_obj.vmobjlock);
2027 #ifdef DEBUG
2028 		if (uvn->u_flags & UVM_VNODE_DYING) {
2029 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2030 		}
2031 #endif
2032 		uvn_flush(&uvn->u_obj, 0, 0,
2033 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2034 
2035 		/*
2036 		 * if we have the only reference and we just cleaned the uvn,
2037 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2038 		 * thus allowing us to avoid thinking about flushing it again
2039 		 * on later sync ops.
2040 		 */
2041 		if (uvn->u_obj.uo_refs == 1 &&
2042 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2043 			LIST_REMOVE(uvn, u_wlist);
2044 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2045 		}
2046 
2047 		simple_unlock(&uvn->u_obj.vmobjlock);
2048 
2049 		/* now drop our reference to the uvn */
2050 		uvn_detach(&uvn->u_obj);
2051 	}
2052 
2053 	/*
2054 	 * done!  release sync lock
2055 	 */
2056 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0, curproc /*XXX*/);
2057 }
2058