xref: /openbsd-src/sys/uvm/uvm_vnode.c (revision 62a742911104f98b9185b2c6b6007d9b1c36396c)
1 /*	$OpenBSD: uvm_vnode.c,v 1.2 1999/02/26 05:32:08 art Exp $	*/
2 /*	$NetBSD: uvm_vnode.c,v 1.18 1999/01/29 12:56:17 bouyer Exp $	*/
3 
4 /*
5  * XXXCDC: "ROUGH DRAFT" QUALITY UVM PRE-RELEASE FILE!
6  *         >>>USE AT YOUR OWN RISK, WORK IS NOT FINISHED<<<
7  */
8 /*
9  * Copyright (c) 1997 Charles D. Cranor and Washington University.
10  * Copyright (c) 1991, 1993
11  *      The Regents of the University of California.
12  * Copyright (c) 1990 University of Utah.
13  *
14  * All rights reserved.
15  *
16  * This code is derived from software contributed to Berkeley by
17  * the Systems Programming Group of the University of Utah Computer
18  * Science Department.
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  * 1. Redistributions of source code must retain the above copyright
24  *    notice, this list of conditions and the following disclaimer.
25  * 2. Redistributions in binary form must reproduce the above copyright
26  *    notice, this list of conditions and the following disclaimer in the
27  *    documentation and/or other materials provided with the distribution.
28  * 3. All advertising materials mentioning features or use of this software
29  *    must display the following acknowledgement:
30  *      This product includes software developed by Charles D. Cranor,
31  *	Washington University, the University of California, Berkeley and
32  *	its contributors.
33  * 4. Neither the name of the University nor the names of its contributors
34  *    may be used to endorse or promote products derived from this software
35  *    without specific prior written permission.
36  *
37  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
38  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
39  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
40  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
41  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
42  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
43  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
44  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
45  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
46  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
47  * SUCH DAMAGE.
48  *
49  *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
50  * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
51  */
52 
53 /*
54  * uvm_vnode.c: the vnode pager.
55  */
56 
57 #include <sys/param.h>
58 #include <sys/systm.h>
59 #include <sys/proc.h>
60 #include <sys/malloc.h>
61 #include <sys/vnode.h>
62 #include <sys/disklabel.h>
63 #include <sys/ioctl.h>
64 #include <sys/fcntl.h>
65 #include <sys/conf.h>
66 
67 #include <miscfs/specfs/specdev.h>
68 
69 #include <vm/vm.h>
70 #include <vm/vm_page.h>
71 #include <vm/vm_kern.h>
72 
73 #include <uvm/uvm.h>
74 #include <uvm/uvm_vnode.h>
75 
76 /*
77  * private global data structure
78  *
79  * we keep a list of writeable active vnode-backed VM objects for sync op.
80  * we keep a simpleq of vnodes that are currently being sync'd.
81  */
82 
83 LIST_HEAD(uvn_list_struct, uvm_vnode);
84 static struct uvn_list_struct uvn_wlist;	/* writeable uvns */
85 static simple_lock_data_t uvn_wl_lock;		/* locks uvn_wlist */
86 
87 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
88 static struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
89 lock_data_t uvn_sync_lock;			/* locks sync operation */
90 
91 /*
92  * functions
93  */
94 
95 static int		   uvn_asyncget __P((struct uvm_object *, vaddr_t,
96 					    int));
97 struct uvm_object 	  *uvn_attach __P((void *, vm_prot_t));
98 static void		   uvn_cluster __P((struct uvm_object *, vaddr_t,
99 					   vaddr_t *, vaddr_t *));
100 static void                uvn_detach __P((struct uvm_object *));
101 static boolean_t           uvn_flush __P((struct uvm_object *, vaddr_t,
102 					 vaddr_t, int));
103 static int                 uvn_get __P((struct uvm_object *, vaddr_t,
104 					vm_page_t *, int *, int,
105 					vm_prot_t, int, int));
106 static void		   uvn_init __P((void));
107 static int		   uvn_io __P((struct uvm_vnode *, vm_page_t *,
108 				      int, int, int));
109 static int		   uvn_put __P((struct uvm_object *, vm_page_t *,
110 					int, boolean_t));
111 static void                uvn_reference __P((struct uvm_object *));
112 static boolean_t	   uvn_releasepg __P((struct vm_page *,
113 					      struct vm_page **));
114 
115 /*
116  * master pager structure
117  */
118 
119 struct uvm_pagerops uvm_vnodeops = {
120 	uvn_init,
121 	uvn_attach,
122 	uvn_reference,
123 	uvn_detach,
124 	NULL,			/* no specialized fault routine required */
125 	uvn_flush,
126 	uvn_get,
127 	uvn_asyncget,
128 	uvn_put,
129 	uvn_cluster,
130 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
131 	uvm_shareprot,	 /* !NULL: allow us in share maps */
132 	NULL,		 /* AIO-DONE function (not until we have asyncio) */
133 	uvn_releasepg,
134 };
135 
136 /*
137  * the ops!
138  */
139 
140 /*
141  * uvn_init
142  *
143  * init pager private data structures.
144  */
145 
146 static void
147 uvn_init()
148 {
149 
150 	LIST_INIT(&uvn_wlist);
151 	simple_lock_init(&uvn_wl_lock);
152 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
153 	lockinit(&uvn_sync_lock, PVM, "uvnsync", 0, 0);
154 }
155 
156 /*
157  * uvn_attach
158  *
159  * attach a vnode structure to a VM object.  if the vnode is already
160  * attached, then just bump the reference count by one and return the
161  * VM object.   if not already attached, attach and return the new VM obj.
162  * the "accessprot" tells the max access the attaching thread wants to
163  * our pages.
164  *
165  * => caller must _not_ already be holding the lock on the uvm_object.
166  * => in fact, nothing should be locked so that we can sleep here.
167  * => note that uvm_object is first thing in vnode structure, so their
168  *    pointers are equiv.
169  */
170 
171 struct uvm_object *
172 uvn_attach(arg, accessprot)
173 	void *arg;
174 	vm_prot_t accessprot;
175 {
176 	struct vnode *vp = arg;
177 	struct uvm_vnode *uvn = &vp->v_uvm;
178 	struct vattr vattr;
179 	int oldflags, result;
180 	struct partinfo pi;
181 	u_quad_t used_vnode_size;
182 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
183 
184 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
185 
186 	used_vnode_size = (u_quad_t)0;	/* XXX gcc -Wuninitialized */
187 
188 	/*
189 	 * first get a lock on the uvn.
190 	 */
191 	simple_lock(&uvn->u_obj.vmobjlock);
192 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
193 		uvn->u_flags |= UVM_VNODE_WANTED;
194 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
195 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
196 		    "uvn_attach", 0);
197 		simple_lock(&uvn->u_obj.vmobjlock);
198 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
199 	}
200 
201 	/*
202 	 * if we're mapping a BLK device, make sure it is a disk.
203 	 */
204 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
205 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
206 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
207 		return(NULL);
208 	}
209 
210 	/*
211 	 * now we have lock and uvn must not be in a blocked state.
212 	 * first check to see if it is already active, in which case
213 	 * we can bump the reference count, check to see if we need to
214 	 * add it to the writeable list, and then return.
215 	 */
216 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
217 
218 		/* regain VREF if we were persisting */
219 		if (uvn->u_obj.uo_refs == 0) {
220 			VREF(vp);
221 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
222 			    0,0,0,0);
223 		}
224 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
225 
226 		/* check for new writeable uvn */
227 		if ((accessprot & VM_PROT_WRITE) != 0 &&
228 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
229 			simple_lock(&uvn_wl_lock);
230 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
231 			simple_unlock(&uvn_wl_lock);
232 			/* we are now on wlist! */
233 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
234 		}
235 
236 		/* unlock and return */
237 		simple_unlock(&uvn->u_obj.vmobjlock);
238 		UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
239 		    0, 0, 0);
240 		return (&uvn->u_obj);
241 	}
242 
243 	/*
244 	 * need to call VOP_GETATTR() to get the attributes, but that could
245 	 * block (due to I/O), so we want to unlock the object before calling.
246 	 * however, we want to keep anyone else from playing with the object
247 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
248 	 * prevents anyone from attaching to the vnode until we are done with
249 	 * it.
250 	 */
251 	uvn->u_flags = UVM_VNODE_ALOCK;
252 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
253 		/* XXX: curproc? */
254 
255 	if (vp->v_type == VBLK) {
256 		/*
257 		 * We could implement this as a specfs getattr call, but:
258 		 *
259 		 *	(1) VOP_GETATTR() would get the file system
260 		 *	    vnode operation, not the specfs operation.
261 		 *
262 		 *	(2) All we want is the size, anyhow.
263 		 */
264 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
265 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
266 		if (result == 0) {
267 			/* XXX should remember blocksize */
268 			used_vnode_size = (u_quad_t)pi.disklab->d_secsize *
269 			    (u_quad_t)pi.part->p_size;
270 		}
271 	} else {
272 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
273 		if (result == 0)
274 			used_vnode_size = vattr.va_size;
275 	}
276 
277 	/* relock object */
278 	simple_lock(&uvn->u_obj.vmobjlock);
279 
280 	if (result != 0) {
281 		if (uvn->u_flags & UVM_VNODE_WANTED)
282 			wakeup(uvn);
283 		uvn->u_flags = 0;
284 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
285 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
286 		return(NULL);
287 	}
288 
289 	/*
290 	 * make sure that the newsize fits within a vaddr_t
291 	 * XXX: need to revise addressing data types
292 	 */
293 #ifdef DEBUG
294 	if (vp->v_type == VBLK)
295 		printf("used_vnode_size = %qu\n", used_vnode_size);
296 #endif
297 	if (used_vnode_size > (vaddr_t) -PAGE_SIZE) {
298 #ifdef DEBUG
299 		printf("uvn_attach: vn %p size truncated %qx->%x\n", vp,
300 		    used_vnode_size, -PAGE_SIZE);
301 #endif
302 		used_vnode_size = (vaddr_t) -PAGE_SIZE;
303 	}
304 
305 	/*
306 	 * now set up the uvn.
307 	 */
308 	uvn->u_obj.pgops = &uvm_vnodeops;
309 	TAILQ_INIT(&uvn->u_obj.memq);
310 	uvn->u_obj.uo_npages = 0;
311 	uvn->u_obj.uo_refs = 1;			/* just us... */
312 	oldflags = uvn->u_flags;
313 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
314 	uvn->u_nio = 0;
315 	uvn->u_size = used_vnode_size;
316 
317 	/* if write access, we need to add it to the wlist */
318 	if (accessprot & VM_PROT_WRITE) {
319 		simple_lock(&uvn_wl_lock);
320 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
321 		simple_unlock(&uvn_wl_lock);
322 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
323 	}
324 
325 	/*
326 	 * add a reference to the vnode.   this reference will stay as long
327 	 * as there is a valid mapping of the vnode.   dropped when the
328 	 * reference count goes to zero [and we either free or persist].
329 	 */
330 	VREF(vp);
331 	simple_unlock(&uvn->u_obj.vmobjlock);
332 	if (oldflags & UVM_VNODE_WANTED)
333 		wakeup(uvn);
334 
335 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
336 	return(&uvn->u_obj);
337 }
338 
339 
340 /*
341  * uvn_reference
342  *
343  * duplicate a reference to a VM object.  Note that the reference
344  * count must already be at least one (the passed in reference) so
345  * there is no chance of the uvn being killed or locked out here.
346  *
347  * => caller must call with object unlocked.
348  * => caller must be using the same accessprot as was used at attach time
349  */
350 
351 
352 static void
353 uvn_reference(uobj)
354 	struct uvm_object *uobj;
355 {
356 #ifdef DIAGNOSTIC
357 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
358 #endif
359 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
360 
361 	simple_lock(&uobj->vmobjlock);
362 #ifdef DIAGNOSTIC
363 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
364 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
365 		    uobj->uo_refs);
366 		panic("uvn_reference: invalid state");
367 	}
368 #endif
369 	uobj->uo_refs++;
370 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
371 	uobj, uobj->uo_refs,0,0);
372 	simple_unlock(&uobj->vmobjlock);
373 }
374 
375 /*
376  * uvn_detach
377  *
378  * remove a reference to a VM object.
379  *
380  * => caller must call with object unlocked and map locked.
381  * => this starts the detach process, but doesn't have to finish it
382  *    (async i/o could still be pending).
383  */
384 static void
385 uvn_detach(uobj)
386 	struct uvm_object *uobj;
387 {
388 	struct uvm_vnode *uvn;
389 	struct vnode *vp;
390 	int oldflags;
391 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
392 
393 	simple_lock(&uobj->vmobjlock);
394 
395 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
396 	uobj->uo_refs--;			/* drop ref! */
397 	if (uobj->uo_refs) {			/* still more refs */
398 		simple_unlock(&uobj->vmobjlock);
399 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
400 		return;
401 	}
402 
403 	/*
404 	 * get other pointers ...
405 	 */
406 
407 	uvn = (struct uvm_vnode *) uobj;
408 	vp = (struct vnode *) uobj;
409 
410 	/*
411 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
412 	 * to keep an active text file from being overwritten).
413 	 */
414 	vp->v_flag &= ~VTEXT;
415 
416 	/*
417 	 * we just dropped the last reference to the uvn.   see if we can
418 	 * let it "stick around".
419 	 */
420 
421 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
422 		/* won't block */
423 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
424 		simple_unlock(&uobj->vmobjlock);
425 		vrele(vp);			/* drop vnode reference */
426 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
427 		return;
428 	}
429 
430 	/*
431 	 * its a goner!
432 	 */
433 
434 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
435 
436 	uvn->u_flags |= UVM_VNODE_DYING;
437 
438 	/*
439 	 * even though we may unlock in flush, no one can gain a reference
440 	 * to us until we clear the "dying" flag [because it blocks
441 	 * attaches].  we will not do that until after we've disposed of all
442 	 * the pages with uvn_flush().  note that before the flush the only
443 	 * pages that could be marked PG_BUSY are ones that are in async
444 	 * pageout by the daemon.  (there can't be any pending "get"'s
445 	 * because there are no references to the object).
446 	 */
447 
448 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
449 
450 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
451 
452 	/*
453 	 * given the structure of this pager, the above flush request will
454 	 * create the following state: all the pages that were in the object
455 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
456 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
457 	 * in either case, if we have pages left we can't kill the object
458 	 * yet because i/o is pending.  in this case we set the "relkill"
459 	 * flag which will cause pgo_releasepg to kill the object once all
460 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
461 	 * routine or from the page daemon].
462 	 */
463 
464 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
465 #ifdef DIAGNOSTIC
466 		/*
467 		 * XXXCDC: very unlikely to happen until we have async i/o
468 		 * so print a little info message in case it does.
469 		 */
470 		printf("uvn_detach: vn %p has pages left after flush - "
471 		    "relkill mode\n", uobj);
472 #endif
473 		uvn->u_flags |= UVM_VNODE_RELKILL;
474 		simple_unlock(&uobj->vmobjlock);
475 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
476 		    0, 0);
477 		return;
478 	}
479 
480 	/*
481 	 * kill object now.   note that we can't be on the sync q because
482 	 * all references are gone.
483 	 */
484 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
485 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
486 		LIST_REMOVE(uvn, u_wlist);
487 		simple_unlock(&uvn_wl_lock);
488 	}
489 #ifdef DIAGNOSTIC
490 	if (uobj->memq.tqh_first != NULL)
491 		panic("uvn_deref: vnode VM object still has pages afer "
492 		    "syncio/free flush");
493 #endif
494 	oldflags = uvn->u_flags;
495 	uvn->u_flags = 0;
496 	simple_unlock(&uobj->vmobjlock);
497 
498 	/* wake up any sleepers */
499 	if (oldflags & UVM_VNODE_WANTED)
500 		wakeup(uvn);
501 
502 	/*
503 	 * drop our reference to the vnode.
504 	 */
505 	vrele(vp);
506 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
507 
508 	return;
509 }
510 
511 /*
512  * uvm_vnp_terminate: external hook to clear out a vnode's VM
513  *
514  * called in two cases:
515  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
516  *      count) needs to be freed so that a vnode can be reused.  this
517  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
518  *      the free list is still attached (i.e. not VBAD) then vgone is
519  *	called.   as part of the vgone trace this should get called to
520  *	free the vm object.   this is the common case.
521  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
522  *	"umount -f") the vgone() function is called on active vnodes
523  *	on the mounted file systems to kill their data (the vnodes become
524  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
525  *	call here (even if the uvn is still in use -- i.e. has a non-zero
526  *	reference count).  this case happens at "umount -f" and during a
527  *	"reboot/halt" operation.
528  *
529  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
530  *	[protects us from getting a vnode that is already in the DYING
531  *	 state...]
532  * => unlike uvn_detach, this function must not return until all the
533  *	uvn's pages are disposed of.
534  * => in case [2] the uvn is still alive after this call, but all I/O
535  *	ops will fail (due to the backing vnode now being "dead").  this
536  *	will prob. kill any process using the uvn due to pgo_get failing.
537  */
538 
539 void
540 uvm_vnp_terminate(vp)
541 	struct vnode *vp;
542 {
543 	struct uvm_vnode *uvn = &vp->v_uvm;
544 	int oldflags;
545 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
546 
547 	/*
548 	 * lock object and check if it is valid
549 	 */
550 	simple_lock(&uvn->u_obj.vmobjlock);
551 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
552 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
553 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
554 		simple_unlock(&uvn->u_obj.vmobjlock);
555 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
556 		return;
557 	}
558 
559 	/*
560 	 * must be a valid uvn that is not already dying (because XLOCK
561 	 * protects us from that).   the uvn can't in the the ALOCK state
562 	 * because it is valid, and uvn's that are in the ALOCK state haven't
563 	 * been marked valid yet.
564 	 */
565 
566 #ifdef DEBUG
567 	/*
568 	 * debug check: are we yanking the vnode out from under our uvn?
569 	 */
570 	if (uvn->u_obj.uo_refs) {
571 		printf("uvm_vnp_terminate(%p): terminating active vnode "
572 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
573 	}
574 #endif
575 
576 	/*
577 	 * it is possible that the uvn was detached and is in the relkill
578 	 * state [i.e. waiting for async i/o to finish so that releasepg can
579 	 * kill object].  we take over the vnode now and cancel the relkill.
580 	 * we want to know when the i/o is done so we can recycle right
581 	 * away.   note that a uvn can only be in the RELKILL state if it
582 	 * has a zero reference count.
583 	 */
584 
585 	if (uvn->u_flags & UVM_VNODE_RELKILL)
586 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
587 
588 	/*
589 	 * block the uvn by setting the dying flag, and then flush the
590 	 * pages.  (note that flush may unlock object while doing I/O, but
591 	 * it will re-lock it before it returns control here).
592 	 *
593 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
594 	 * that uvn_io doesn't attempt to VOP_LOCK again.
595 	 *
596 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
597 	 *	due to a forceful unmount might not be a good idea.  maybe we
598 	 *	need a way to pass in this info to uvn_flush through a
599 	 *	pager-defined PGO_ constant [currently there are none].
600 	 */
601 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
602 
603 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
604 
605 	/*
606 	 * as we just did a flush we expect all the pages to be gone or in
607 	 * the process of going.  sleep to wait for the rest to go [via iosync].
608 	 */
609 
610 	while (uvn->u_obj.uo_npages) {
611 #ifdef DIAGNOSTIC
612 		struct vm_page *pp;
613 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
614 		     pp = pp->listq.tqe_next) {
615 			if ((pp->flags & PG_BUSY) == 0)
616 				panic("uvm_vnp_terminate: detected unbusy pg");
617 		}
618 		if (uvn->u_nio == 0)
619 			panic("uvm_vnp_terminate: no I/O to wait for?");
620 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
621 		/*
622 		 * XXXCDC: this is unlikely to happen without async i/o so we
623 		 * put a printf in just to keep an eye on it.
624 		 */
625 #endif
626 		uvn->u_flags |= UVM_VNODE_IOSYNC;
627 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
628 		    "uvn_term",0);
629 		simple_lock(&uvn->u_obj.vmobjlock);
630 	}
631 
632 	/*
633 	 * done.   now we free the uvn if its reference count is zero
634 	 * (true if we are zapping a persisting uvn).   however, if we are
635 	 * terminating a uvn with active mappings we let it live ... future
636 	 * calls down to the vnode layer will fail.
637 	 */
638 
639 	oldflags = uvn->u_flags;
640 	if (uvn->u_obj.uo_refs) {
641 
642 		/*
643 		 * uvn must live on it is dead-vnode state until all references
644 		 * are gone.   restore flags.    clear CANPERSIST state.
645 		 */
646 
647 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
648 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
649 
650 	} else {
651 
652 		/*
653 		 * free the uvn now.   note that the VREF reference is already
654 		 * gone [it is dropped when we enter the persist state].
655 		 */
656 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
657 			panic("uvm_vnp_terminate: io sync wanted bit set");
658 
659 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
660 			simple_lock(&uvn_wl_lock);
661 			LIST_REMOVE(uvn, u_wlist);
662 			simple_unlock(&uvn_wl_lock);
663 		}
664 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
665 	}
666 
667 	if (oldflags & UVM_VNODE_WANTED)
668 		wakeup(uvn);		/* object lock still held */
669 
670 	simple_unlock(&uvn->u_obj.vmobjlock);
671 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
672 
673 }
674 
675 /*
676  * uvn_releasepg: handled a released page in a uvn
677  *
678  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
679  *	to dispose of.
680  * => caller must handled PG_WANTED case
681  * => called with page's object locked, pageq's unlocked
682  * => returns TRUE if page's object is still alive, FALSE if we
683  *	killed the page's object.    if we return TRUE, then we
684  *	return with the object locked.
685  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
686  *				with the page queues locked [for pagedaemon]
687  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
688  * => we kill the uvn if it is not referenced and we are suppose to
689  *	kill it ("relkill").
690  */
691 
692 boolean_t
693 uvn_releasepg(pg, nextpgp)
694 	struct vm_page *pg;
695 	struct vm_page **nextpgp;	/* OUT */
696 {
697 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
698 #ifdef DIAGNOSTIC
699 	if ((pg->flags & PG_RELEASED) == 0)
700 		panic("uvn_releasepg: page not released!");
701 #endif
702 
703 	/*
704 	 * dispose of the page [caller handles PG_WANTED]
705 	 */
706 	pmap_page_protect(PMAP_PGARG(pg), VM_PROT_NONE);
707 	uvm_lock_pageq();
708 	if (nextpgp)
709 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
710 	uvm_pagefree(pg);
711 	if (!nextpgp)
712 		uvm_unlock_pageq();
713 
714 	/*
715 	 * now see if we need to kill the object
716 	 */
717 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
718 		if (uvn->u_obj.uo_refs)
719 			panic("uvn_releasepg: kill flag set on referenced "
720 			    "object!");
721 		if (uvn->u_obj.uo_npages == 0) {
722 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
723 				simple_lock(&uvn_wl_lock);
724 				LIST_REMOVE(uvn, u_wlist);
725 				simple_unlock(&uvn_wl_lock);
726 			}
727 #ifdef DIAGNOSTIC
728 			if (uvn->u_obj.memq.tqh_first)
729 	panic("uvn_releasepg: pages in object with npages == 0");
730 #endif
731 			if (uvn->u_flags & UVM_VNODE_WANTED)
732 				/* still holding object lock */
733 				wakeup(uvn);
734 
735 			uvn->u_flags = 0;		/* DEAD! */
736 			simple_unlock(&uvn->u_obj.vmobjlock);
737 			return (FALSE);
738 		}
739 	}
740 	return (TRUE);
741 }
742 
743 /*
744  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
745  * through the buffer cache and allow I/O in any size.  These VOPs use
746  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
747  * go through the buffer cache or allow I/O sizes larger than a
748  * block].  we will eventually want to change this.
749  *
750  * issues to consider:
751  *   uvm provides the uvm_aiodesc structure for async i/o management.
752  * there are two tailq's in the uvm. structure... one for pending async
753  * i/o and one for "done" async i/o.   to do an async i/o one puts
754  * an aiodesc on the "pending" list (protected by splbio()), starts the
755  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
756  * some sort of "i/o done" function to be called (at splbio(), interrupt
757  * time).   this function should remove the aiodesc from the pending list
758  * and place it on the "done" list and wakeup the daemon.   the daemon
759  * will run at normal spl() and will remove all items from the "done"
760  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
761  * [in the old vm code, this was done by calling the "put" routine with
762  * null arguments which made the code harder to read and understand because
763  * you had one function ("put") doing two things.]
764  *
765  * so the current pager needs:
766  *   int uvn_aiodone(struct uvm_aiodesc *)
767  *
768  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
769  *	later collection.
770  * => called with pageq's locked by the daemon.
771  *
772  * general outline:
773  * - "try" to lock object.   if fail, just return (will try again later)
774  * - drop "u_nio" (this req is done!)
775  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
776  * - get "page" structures (atop?).
777  * - handle "wanted" pages
778  * - handle "released" pages [using pgo_releasepg]
779  *   >>> pgo_releasepg may kill the object
780  * dont forget to look at "object" wanted flag in all cases.
781  */
782 
783 
784 /*
785  * uvn_flush: flush pages out of a uvm object.
786  *
787  * => object should be locked by caller.   we may _unlock_ the object
788  *	if (and only if) we need to clean a page (PGO_CLEANIT).
789  *	we return with the object locked.
790  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
791  *	might want to unlock higher level resources (e.g. vm_map)
792  *	before calling flush.
793  * => if PGO_CLEANIT is not set, then we will neither unlock the object
794  *	or block.
795  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
796  *	for flushing.
797  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
798  *	that new pages are inserted on the tail end of the list.   thus,
799  *	we can make a complete pass through the object in one go by starting
800  *	at the head and working towards the tail (new pages are put in
801  *	front of us).
802  * => NOTE: we are allowed to lock the page queues, so the caller
803  *	must not be holding the lock on them [e.g. pagedaemon had
804  *	better not call us with the queues locked]
805  * => we return TRUE unless we encountered some sort of I/O error
806  *
807  * comment on "cleaning" object and PG_BUSY pages:
808  *	this routine is holding the lock on the object.   the only time
809  *	that it can run into a PG_BUSY page that it does not own is if
810  *	some other process has started I/O on the page (e.g. either
811  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
812  *	in, then it can not be dirty (!PG_CLEAN) because no one has
813  *	had a chance to modify it yet.    if the PG_BUSY page is being
814  *	paged out then it means that someone else has already started
815  *	cleaning the page for us (how nice!).    in this case, if we
816  *	have syncio specified, then after we make our pass through the
817  *	object we need to wait for the other PG_BUSY pages to clear
818  *	off (i.e. we need to do an iosync).   also note that once a
819  *	page is PG_BUSY it must stay in its object until it is un-busyed.
820  *
821  * note on page traversal:
822  *	we can traverse the pages in an object either by going down the
823  *	linked list in "uobj->memq", or we can go over the address range
824  *	by page doing hash table lookups for each address.    depending
825  *	on how many pages are in the object it may be cheaper to do one
826  *	or the other.   we set "by_list" to true if we are using memq.
827  *	if the cost of a hash lookup was equal to the cost of the list
828  *	traversal we could compare the number of pages in the start->stop
829  *	range to the total number of pages in the object.   however, it
830  *	seems that a hash table lookup is more expensive than the linked
831  *	list traversal, so we multiply the number of pages in the
832  *	start->stop range by a penalty which we define below.
833  */
834 
835 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
836 
837 static boolean_t
838 uvn_flush(uobj, start, stop, flags)
839 	struct uvm_object *uobj;
840 	vaddr_t start, stop;
841 	int flags;
842 {
843 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
844 	struct vm_page *pp, *ppnext, *ptmp;
845 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
846 	int npages, result, lcv;
847 	boolean_t retval, need_iosync, by_list, needs_clean;
848 	vaddr_t curoff;
849 	u_short pp_version;
850 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
851 
852 	curoff = 0;	/* XXX: shut up gcc */
853 	/*
854 	 * get init vals and determine how we are going to traverse object
855 	 */
856 
857 	need_iosync = FALSE;
858 	retval = TRUE;		/* return value */
859 	if (flags & PGO_ALLPAGES) {
860 		start = 0;
861 		stop = round_page(uvn->u_size);
862 		by_list = TRUE;		/* always go by the list */
863 	} else {
864 		start = trunc_page(start);
865 		stop = round_page(stop);
866 		if (stop > round_page(uvn->u_size))
867 			printf("uvn_flush: strange, got an out of range "
868 			    "flush (fixed)\n");
869 
870 		by_list = (uobj->uo_npages <=
871 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
872 	}
873 
874 	UVMHIST_LOG(maphist,
875 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
876 	    start, stop, by_list, flags);
877 
878 	/*
879 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
880 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
881 	 * is wrong it will only prevent us from clustering... it won't break
882 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
883 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
884 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
885 	 * bit is always up to date since there are no mappings).
886 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
887 	 */
888 
889 	if ((flags & PGO_CLEANIT) != 0 &&
890 	    uobj->pgops->pgo_mk_pcluster != NULL) {
891 		if (by_list) {
892 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
893 			    pp = pp->listq.tqe_next) {
894 				if (pp->offset < start || pp->offset >= stop)
895 					continue;
896 				pp->flags &= ~PG_CLEANCHK;
897 			}
898 
899 		} else {   /* by hash */
900 			for (curoff = start ; curoff < stop;
901 			    curoff += PAGE_SIZE) {
902 				pp = uvm_pagelookup(uobj, curoff);
903 				if (pp)
904 					pp->flags &= ~PG_CLEANCHK;
905 			}
906 		}
907 	}
908 
909 	/*
910 	 * now do it.   note: we must update ppnext in body of loop or we
911 	 * will get stuck.  we need to use ppnext because we may free "pp"
912 	 * before doing the next loop.
913 	 */
914 
915 	if (by_list) {
916 		pp = uobj->memq.tqh_first;
917 	} else {
918 		curoff = start;
919 		pp = uvm_pagelookup(uobj, curoff);
920 	}
921 
922 	ppnext = NULL;	/* XXX: shut up gcc */
923 	ppsp = NULL;		/* XXX: shut up gcc */
924 	uvm_lock_pageq();	/* page queues locked */
925 
926 	/* locked: both page queues and uobj */
927 	for ( ; (by_list && pp != NULL) ||
928 	  (!by_list && curoff < stop) ; pp = ppnext) {
929 
930 		if (by_list) {
931 
932 			/*
933 			 * range check
934 			 */
935 
936 			if (pp->offset < start || pp->offset >= stop) {
937 				ppnext = pp->listq.tqe_next;
938 				continue;
939 			}
940 
941 		} else {
942 
943 			/*
944 			 * null check
945 			 */
946 
947 			curoff += PAGE_SIZE;
948 			if (pp == NULL) {
949 				if (curoff < stop)
950 					ppnext = uvm_pagelookup(uobj, curoff);
951 				continue;
952 			}
953 
954 		}
955 
956 		/*
957 		 * handle case where we do not need to clean page (either
958 		 * because we are not clean or because page is not dirty or
959 		 * is busy):
960 		 *
961 		 * NOTE: we are allowed to deactivate a non-wired active
962 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
963 		 * queue it must stay put until it is !PG_BUSY (so as not to
964 		 * confuse pagedaemon).
965 		 */
966 
967 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
968 			needs_clean = FALSE;
969 			if ((pp->flags & PG_BUSY) != 0 &&
970 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
971 			             (PGO_CLEANIT|PGO_SYNCIO))
972 				need_iosync = TRUE;
973 		} else {
974 			/*
975 			 * freeing: nuke all mappings so we can sync
976 			 * PG_CLEAN bit with no race
977 			 */
978 			if ((pp->flags & PG_CLEAN) != 0 &&
979 			    (flags & PGO_FREE) != 0 &&
980 			    (pp->pqflags & PQ_ACTIVE) != 0)
981 				pmap_page_protect(PMAP_PGARG(pp), VM_PROT_NONE);
982 			if ((pp->flags & PG_CLEAN) != 0 &&
983 			    pmap_is_modified(PMAP_PGARG(pp)))
984 				pp->flags &= ~(PG_CLEAN);
985 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
986 
987 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
988 		}
989 
990 		/*
991 		 * if we don't need a clean... load ppnext and dispose of pp
992 		 */
993 		if (!needs_clean) {
994 			/* load ppnext */
995 			if (by_list)
996 				ppnext = pp->listq.tqe_next;
997 			else {
998 				if (curoff < stop)
999 					ppnext = uvm_pagelookup(uobj, curoff);
1000 			}
1001 
1002 			/* now dispose of pp */
1003 			if (flags & PGO_DEACTIVATE) {
1004 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1005 				    pp->wire_count == 0) {
1006 					pmap_page_protect(PMAP_PGARG(pp),
1007 					    VM_PROT_NONE);
1008 					uvm_pagedeactivate(pp);
1009 				}
1010 
1011 			} else if (flags & PGO_FREE) {
1012 				if (pp->flags & PG_BUSY) {
1013 					/* release busy pages */
1014 					pp->flags |= PG_RELEASED;
1015 				} else {
1016 					pmap_page_protect(PMAP_PGARG(pp),
1017 					    VM_PROT_NONE);
1018 					/* removed page from object */
1019 					uvm_pagefree(pp);
1020 				}
1021 			}
1022 			/* ppnext is valid so we can continue... */
1023 			continue;
1024 		}
1025 
1026 		/*
1027 		 * pp points to a page in the locked object that we are
1028 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
1029 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1030 		 *
1031 		 * let uvm_pager_put attempted a clustered page out.
1032 		 * note: locked: uobj and page queues.
1033 		 */
1034 
1035 		pp->flags |= PG_BUSY;	/* we 'own' page now */
1036 		UVM_PAGE_OWN(pp, "uvn_flush");
1037 		pmap_page_protect(PMAP_PGARG(pp), VM_PROT_READ);
1038 		pp_version = pp->version;
1039 ReTry:
1040 		ppsp = pps;
1041 		npages = sizeof(pps) / sizeof(struct vm_page *);
1042 
1043 		/* locked: page queues, uobj */
1044 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1045 			   flags | PGO_DOACTCLUST, start, stop);
1046 		/* unlocked: page queues, uobj */
1047 
1048 		/*
1049 		 * at this point nothing is locked.   if we did an async I/O
1050 		 * it is remotely possible for the async i/o to complete and
1051 		 * the page "pp" be freed or what not before we get a chance
1052 		 * to relock the object.   in order to detect this, we have
1053 		 * saved the version number of the page in "pp_version".
1054 		 */
1055 
1056 		/* relock! */
1057 		simple_lock(&uobj->vmobjlock);
1058 		uvm_lock_pageq();
1059 
1060 		/*
1061 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1062 		 * can only happen when  we are doing async I/O and can't
1063 		 * map the pages into kernel memory (pager_map) due to lack
1064 		 * of vm space.   if this happens we drop back to sync I/O.
1065 		 */
1066 
1067 		if (result == VM_PAGER_AGAIN) {
1068 			/*
1069 			 * it is unlikely, but page could have been released
1070 			 * while we had the object lock dropped.   we ignore
1071 			 * this now and retry the I/O.  we will detect and
1072 			 * handle the released page after the syncio I/O
1073 			 * completes.
1074 			 */
1075 #ifdef DIAGNOSTIC
1076 			if (flags & PGO_SYNCIO)
1077 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1078 #endif
1079 			flags |= PGO_SYNCIO;
1080 			goto ReTry;
1081 		}
1082 
1083 		/*
1084 		 * the cleaning operation is now done.   finish up.  note that
1085 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1086 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1087 		 * to us in ppsp/npages.
1088 		 */
1089 
1090 		/*
1091 		 * for pending async i/o if we are not deactivating/freeing
1092 		 * we can move on to the next page.
1093 		 */
1094 
1095 		if (result == VM_PAGER_PEND) {
1096 
1097 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1098 				/*
1099 				 * no per-page ops: refresh ppnext and continue
1100 				 */
1101 				if (by_list) {
1102 					if (pp->version == pp_version)
1103 						ppnext = pp->listq.tqe_next;
1104 					else
1105 						/* reset */
1106 						ppnext = uobj->memq.tqh_first;
1107 				} else {
1108 					if (curoff < stop)
1109 						ppnext = uvm_pagelookup(uobj,
1110 						    curoff);
1111 				}
1112 				continue;
1113 			}
1114 
1115 			/* need to do anything here? */
1116 		}
1117 
1118 		/*
1119 		 * need to look at each page of the I/O operation.  we defer
1120 		 * processing "pp" until the last trip through this "for" loop
1121 		 * so that we can load "ppnext" for the main loop after we
1122 		 * play with the cluster pages [thus the "npages + 1" in the
1123 		 * loop below].
1124 		 */
1125 
1126 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1127 
1128 			/*
1129 			 * handle ppnext for outside loop, and saving pp
1130 			 * until the end.
1131 			 */
1132 			if (lcv < npages) {
1133 				if (ppsp[lcv] == pp)
1134 					continue; /* skip pp until the end */
1135 				ptmp = ppsp[lcv];
1136 			} else {
1137 				ptmp = pp;
1138 
1139 				/* set up next page for outer loop */
1140 				if (by_list) {
1141 					if (pp->version == pp_version)
1142 						ppnext = pp->listq.tqe_next;
1143 					else
1144 						/* reset */
1145 						ppnext = uobj->memq.tqh_first;
1146 				} else {
1147 					if (curoff < stop)
1148 					ppnext = uvm_pagelookup(uobj, curoff);
1149 				}
1150 			}
1151 
1152 			/*
1153 			 * verify the page didn't get moved while obj was
1154 			 * unlocked
1155 			 */
1156 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1157 				continue;
1158 
1159 			/*
1160 			 * unbusy the page if I/O is done.   note that for
1161 			 * pending I/O it is possible that the I/O op
1162 			 * finished before we relocked the object (in
1163 			 * which case the page is no longer busy).
1164 			 */
1165 
1166 			if (result != VM_PAGER_PEND) {
1167 				if (ptmp->flags & PG_WANTED)
1168 					/* still holding object lock */
1169 					thread_wakeup(ptmp);
1170 
1171 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1172 				UVM_PAGE_OWN(ptmp, NULL);
1173 				if (ptmp->flags & PG_RELEASED) {
1174 
1175 					/* pgo_releasepg wants this */
1176 					uvm_unlock_pageq();
1177 					if (!uvn_releasepg(ptmp, NULL))
1178 						return (TRUE);
1179 
1180 					uvm_lock_pageq();	/* relock */
1181 					continue;		/* next page */
1182 
1183 				} else {
1184 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1185 					if ((flags & PGO_FREE) == 0)
1186 						pmap_clear_modify(
1187 						    PMAP_PGARG(ptmp));
1188 				}
1189 			}
1190 
1191 			/*
1192 			 * dispose of page
1193 			 */
1194 
1195 			if (flags & PGO_DEACTIVATE) {
1196 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1197 				    pp->wire_count == 0) {
1198 					pmap_page_protect(PMAP_PGARG(ptmp),
1199 					    VM_PROT_NONE);
1200 					uvm_pagedeactivate(ptmp);
1201 				}
1202 
1203 			} else if (flags & PGO_FREE) {
1204 				if (result == VM_PAGER_PEND) {
1205 					if ((ptmp->flags & PG_BUSY) != 0)
1206 						/* signal for i/o done */
1207 						ptmp->flags |= PG_RELEASED;
1208 				} else {
1209 					if (result != VM_PAGER_OK) {
1210 						printf("uvn_flush: obj=%p, "
1211 						   "offset=0x%lx.  error "
1212 						   "during pageout.\n",
1213 						    pp->uobject, pp->offset);
1214 						printf("uvn_flush: WARNING: "
1215 						    "changes to page may be "
1216 						    "lost!\n");
1217 						retval = FALSE;
1218 					}
1219 					pmap_page_protect(PMAP_PGARG(ptmp),
1220 					    VM_PROT_NONE);
1221 					uvm_pagefree(ptmp);
1222 				}
1223 			}
1224 
1225 		}		/* end of "lcv" for loop */
1226 
1227 	}		/* end of "pp" for loop */
1228 
1229 	/*
1230 	 * done with pagequeues: unlock
1231 	 */
1232 	uvm_unlock_pageq();
1233 
1234 	/*
1235 	 * now wait for all I/O if required.
1236 	 */
1237 	if (need_iosync) {
1238 
1239 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1240 		while (uvn->u_nio != 0) {
1241 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1242 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1243 			  FALSE, "uvn_flush",0);
1244 			simple_lock(&uvn->u_obj.vmobjlock);
1245 		}
1246 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1247 			wakeup(&uvn->u_flags);
1248 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1249 	}
1250 
1251 	/* return, with object locked! */
1252 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1253 	return(retval);
1254 }
1255 
1256 /*
1257  * uvn_cluster
1258  *
1259  * we are about to do I/O in an object at offset.   this function is called
1260  * to establish a range of offsets around "offset" in which we can cluster
1261  * I/O.
1262  *
1263  * - currently doesn't matter if obj locked or not.
1264  */
1265 
1266 static void
1267 uvn_cluster(uobj, offset, loffset, hoffset)
1268 	struct uvm_object *uobj;
1269 	vaddr_t offset;
1270 	vaddr_t *loffset, *hoffset; /* OUT */
1271 {
1272 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1273 	*loffset = offset;
1274 
1275 	if (*loffset >= uvn->u_size)
1276 		panic("uvn_cluster: offset out of range");
1277 
1278 	/*
1279 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1280 	 */
1281 	*hoffset = *loffset + MAXBSIZE;
1282 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1283 		*hoffset = round_page(uvn->u_size);
1284 
1285 	return;
1286 }
1287 
1288 /*
1289  * uvn_put: flush page data to backing store.
1290  *
1291  * => prefer map unlocked (not required)
1292  * => object must be locked!   we will _unlock_ it before starting I/O.
1293  * => flags: PGO_SYNCIO -- use sync. I/O
1294  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1295  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1296  *	[thus we never do async i/o!  see iodone comment]
1297  */
1298 
1299 static int
1300 uvn_put(uobj, pps, npages, flags)
1301 	struct uvm_object *uobj;
1302 	struct vm_page **pps;
1303 	int npages, flags;
1304 {
1305 	int retval;
1306 
1307 	/* note: object locked */
1308 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1309 	/* note: object unlocked */
1310 
1311 	return(retval);
1312 }
1313 
1314 
1315 /*
1316  * uvn_get: get pages (synchronously) from backing store
1317  *
1318  * => prefer map unlocked (not required)
1319  * => object must be locked!  we will _unlock_ it before starting any I/O.
1320  * => flags: PGO_ALLPAGES: get all of the pages
1321  *           PGO_LOCKED: fault data structures are locked
1322  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1323  * => NOTE: caller must check for released pages!!
1324  */
1325 
1326 static int
1327 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1328 	struct uvm_object *uobj;
1329 	vaddr_t offset;
1330 	struct vm_page **pps;		/* IN/OUT */
1331 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1332 	int centeridx, advice, flags;
1333 	vm_prot_t access_type;
1334 {
1335 	vaddr_t current_offset;
1336 	struct vm_page *ptmp;
1337 	int lcv, result, gotpages;
1338 	boolean_t done;
1339 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1340 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1341 
1342 	/*
1343 	 * step 1: handled the case where fault data structures are locked.
1344 	 */
1345 
1346 	if (flags & PGO_LOCKED) {
1347 
1348 		/*
1349 		 * gotpages is the current number of pages we've gotten (which
1350 		 * we pass back up to caller via *npagesp.
1351 		 */
1352 
1353 		gotpages = 0;
1354 
1355 		/*
1356 		 * step 1a: get pages that are already resident.   only do this
1357 		 * if the data structures are locked (i.e. the first time
1358 		 * through).
1359 		 */
1360 
1361 		done = TRUE;	/* be optimistic */
1362 
1363 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1364 		    lcv++, current_offset += PAGE_SIZE) {
1365 
1366 			/* do we care about this page?  if not, skip it */
1367 			if (pps[lcv] == PGO_DONTCARE)
1368 				continue;
1369 
1370 			/* lookup page */
1371 			ptmp = uvm_pagelookup(uobj, current_offset);
1372 
1373 			/* to be useful must get a non-busy, non-released pg */
1374 			if (ptmp == NULL ||
1375 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1376 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1377 				    != 0)
1378 				done = FALSE;	/* need to do a wait or I/O! */
1379 				continue;
1380 			}
1381 
1382 			/*
1383 			 * useful page: busy/lock it and plug it in our
1384 			 * result array
1385 			 */
1386 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1387 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1388 			pps[lcv] = ptmp;
1389 			gotpages++;
1390 
1391 		}	/* "for" lcv loop */
1392 
1393 		/*
1394 		 * XXX: given the "advice", should we consider async read-ahead?
1395 		 * XXX: fault current does deactive of pages behind us.  is
1396 		 * this good (other callers might now).
1397 		 */
1398 		/*
1399 		 * XXX: read-ahead currently handled by buffer cache (bread)
1400 		 * level.
1401 		 * XXX: no async i/o available.
1402 		 * XXX: so we don't do anything now.
1403 		 */
1404 
1405 		/*
1406 		 * step 1c: now we've either done everything needed or we to
1407 		 * unlock and do some waiting or I/O.
1408 		 */
1409 
1410 		*npagesp = gotpages;		/* let caller know */
1411 		if (done)
1412 			return(VM_PAGER_OK);		/* bingo! */
1413 		else
1414 			/* EEK!   Need to unlock and I/O */
1415 			return(VM_PAGER_UNLOCK);
1416 	}
1417 
1418 	/*
1419 	 * step 2: get non-resident or busy pages.
1420 	 * object is locked.   data structures are unlocked.
1421 	 *
1422 	 * XXX: because we can't do async I/O at this level we get things
1423 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1424 	 * async-read-ahead for us at a lower level.
1425 	 */
1426 
1427 	for (lcv = 0, current_offset = offset ;
1428 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1429 
1430 		/* skip over pages we've already gotten or don't want */
1431 		/* skip over pages we don't _have_ to get */
1432 		if (pps[lcv] != NULL || (lcv != centeridx &&
1433 		    (flags & PGO_ALLPAGES) == 0))
1434 			continue;
1435 
1436 		/*
1437 		 * we have yet to locate the current page (pps[lcv]).   we first
1438 		 * look for a page that is already at the current offset.   if
1439 		 * we fine a page, we check to see if it is busy or released.
1440 		 * if that is the case, then we sleep on the page until it is
1441 		 * no longer busy or released and repeat the lookup.    if the
1442 		 * page we found is neither busy nor released, then we busy it
1443 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1444 		 * following while loop and indicates we are ready to move on
1445 		 * to the next page in the "lcv" loop above.
1446 		 *
1447 		 * if we exit the while loop with pps[lcv] still set to NULL,
1448 		 * then it means that we allocated a new busy/fake/clean page
1449 		 * ptmp in the object and we need to do I/O to fill in the data.
1450 		 */
1451 
1452 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1453 
1454 			/* look for a current page */
1455 			ptmp = uvm_pagelookup(uobj, current_offset);
1456 
1457 			/* nope?   allocate one now (if we can) */
1458 			if (ptmp == NULL) {
1459 
1460 				ptmp = uvm_pagealloc(uobj, current_offset,
1461 				    NULL);	/* alloc */
1462 
1463 				/* out of RAM? */
1464 				if (ptmp == NULL) {
1465 					simple_unlock(&uobj->vmobjlock);
1466 					uvm_wait("uvn_getpage");
1467 					simple_lock(&uobj->vmobjlock);
1468 
1469 					/* goto top of pps while loop */
1470 					continue;
1471 				}
1472 
1473 				/*
1474 				 * got new page ready for I/O.  break pps
1475 				 * while loop.  pps[lcv] is still NULL.
1476 				 */
1477 				break;
1478 			}
1479 
1480 			/* page is there, see if we need to wait on it */
1481 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1482 				ptmp->flags |= PG_WANTED;
1483 				UVM_UNLOCK_AND_WAIT(ptmp,
1484 				    &uobj->vmobjlock, 0, "uvn_get",0);
1485 				simple_lock(&uobj->vmobjlock);
1486 				continue;	/* goto top of pps while loop */
1487 			}
1488 
1489 			/*
1490 			 * if we get here then the page has become resident
1491 			 * and unbusy between steps 1 and 2.  we busy it
1492 			 * now (so we own it) and set pps[lcv] (so that we
1493 			 * exit the while loop).
1494 			 */
1495 			ptmp->flags |= PG_BUSY;
1496 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1497 			pps[lcv] = ptmp;
1498 		}
1499 
1500 		/*
1501 		 * if we own the a valid page at the correct offset, pps[lcv]
1502 		 * will point to it.   nothing more to do except go to the
1503 		 * next page.
1504 		 */
1505 
1506 		if (pps[lcv])
1507 			continue;			/* next lcv */
1508 
1509 		/*
1510 		 * we have a "fake/busy/clean" page that we just allocated.  do
1511 		 * I/O to fill it with valid data.  note that object must be
1512 		 * locked going into uvn_io, but will be unlocked afterwards.
1513 		 */
1514 
1515 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1516 		    PGO_SYNCIO, UIO_READ);
1517 
1518 		/*
1519 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1520 		 * syncio the result can not be PEND or AGAIN.   we must relock
1521 		 * and check for errors.
1522 		 */
1523 
1524 		/* lock object.   check for errors.   */
1525 		simple_lock(&uobj->vmobjlock);
1526 		if (result != VM_PAGER_OK) {
1527 			if (ptmp->flags & PG_WANTED)
1528 				/* object lock still held */
1529 				thread_wakeup(ptmp);
1530 
1531 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1532 			UVM_PAGE_OWN(ptmp, NULL);
1533 			uvm_lock_pageq();
1534 			uvm_pagefree(ptmp);
1535 			uvm_unlock_pageq();
1536 			simple_unlock(&uobj->vmobjlock);
1537 			return(result);
1538 		}
1539 
1540 		/*
1541 		 * we got the page!   clear the fake flag (indicates valid
1542 		 * data now in page) and plug into our result array.   note
1543 		 * that page is still busy.
1544 		 *
1545 		 * it is the callers job to:
1546 		 * => check if the page is released
1547 		 * => unbusy the page
1548 		 * => activate the page
1549 		 */
1550 
1551 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1552 		pmap_clear_modify(PMAP_PGARG(ptmp));	/* ... and clean */
1553 		pps[lcv] = ptmp;
1554 
1555 	}	/* lcv loop */
1556 
1557 	/*
1558 	 * finally, unlock object and return.
1559 	 */
1560 
1561 	simple_unlock(&uobj->vmobjlock);
1562 	return (VM_PAGER_OK);
1563 }
1564 
1565 /*
1566  * uvn_asyncget: start async I/O to bring pages into ram
1567  *
1568  * => caller must lock object(???XXX: see if this is best)
1569  * => could be called from uvn_get or a madvise() fault-ahead.
1570  * => if it fails, it doesn't matter.
1571  */
1572 
1573 static int
1574 uvn_asyncget(uobj, offset, npages)
1575 	struct uvm_object *uobj;
1576 	vaddr_t offset;
1577 	int npages;
1578 {
1579 
1580 	/*
1581 	 * XXXCDC: we can't do async I/O yet
1582 	 */
1583 	printf("uvn_asyncget called\n");
1584 	return (KERN_SUCCESS);
1585 }
1586 
1587 /*
1588  * uvn_io: do I/O to a vnode
1589  *
1590  * => prefer map unlocked (not required)
1591  * => object must be locked!   we will _unlock_ it before starting I/O.
1592  * => flags: PGO_SYNCIO -- use sync. I/O
1593  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1594  *	[thus we never do async i/o!  see iodone comment]
1595  */
1596 
1597 static int
1598 uvn_io(uvn, pps, npages, flags, rw)
1599 	struct uvm_vnode *uvn;
1600 	vm_page_t *pps;
1601 	int npages, flags, rw;
1602 {
1603 	struct vnode *vn;
1604 	struct uio uio;
1605 	struct iovec iov;
1606 	vaddr_t kva, file_offset;
1607 	int waitf, result, got, wanted;
1608 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1609 
1610 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1611 
1612 	/*
1613 	 * init values
1614 	 */
1615 
1616 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1617 	vn = (struct vnode *) uvn;
1618 	file_offset = pps[0]->offset;
1619 
1620 	/*
1621 	 * check for sync'ing I/O.
1622 	 */
1623 
1624 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1625 		if (waitf == M_NOWAIT) {
1626 			simple_unlock(&uvn->u_obj.vmobjlock);
1627 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1628 			return(VM_PAGER_AGAIN);
1629 		}
1630 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1631 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1632 			FALSE, "uvn_iosync",0);
1633 		simple_lock(&uvn->u_obj.vmobjlock);
1634 	}
1635 
1636 	/*
1637 	 * check size
1638 	 */
1639 
1640 	if (file_offset >= uvn->u_size) {
1641 			simple_unlock(&uvn->u_obj.vmobjlock);
1642 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1643 #ifdef DIAGNOSTIC
1644 			printf("uvn_io: note: size check fired\n");
1645 #endif
1646 			return(VM_PAGER_BAD);
1647 	}
1648 
1649 	/*
1650 	 * first try and map the pages in (without waiting)
1651 	 */
1652 
1653 	kva = uvm_pagermapin(pps, npages, NULL, M_NOWAIT);
1654 	if (kva == NULL && waitf == M_NOWAIT) {
1655 		simple_unlock(&uvn->u_obj.vmobjlock);
1656 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1657 		return(VM_PAGER_AGAIN);
1658 	}
1659 
1660 	/*
1661 	 * ok, now bump u_nio up.   at this point we are done with uvn
1662 	 * and can unlock it.   if we still don't have a kva, try again
1663 	 * (this time with sleep ok).
1664 	 */
1665 
1666 	uvn->u_nio++;			/* we have an I/O in progress! */
1667 	simple_unlock(&uvn->u_obj.vmobjlock);
1668 	/* NOTE: object now unlocked */
1669 	if (kva == NULL) {
1670 		kva = uvm_pagermapin(pps, npages, NULL, M_WAITOK);
1671 	}
1672 
1673 	/*
1674 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1675 	 * get touched (so we can look at "offset" without having to lock
1676 	 * the object).  set up for I/O.
1677 	 */
1678 
1679 	/*
1680 	 * fill out uio/iov
1681 	 */
1682 
1683 	iov.iov_base = (caddr_t) kva;
1684 	wanted = npages << PAGE_SHIFT;
1685 	if (file_offset + wanted > uvn->u_size)
1686 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1687 	iov.iov_len = wanted;
1688 	uio.uio_iov = &iov;
1689 	uio.uio_iovcnt = 1;
1690 	uio.uio_offset = file_offset;
1691 	uio.uio_segflg = UIO_SYSSPACE;
1692 	uio.uio_rw = rw;
1693 	uio.uio_resid = wanted;
1694 	uio.uio_procp = NULL;
1695 
1696 	/*
1697 	 * do the I/O!  (XXX: curproc?)
1698 	 */
1699 
1700 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1701 
1702 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1703 		vn_lock(vn, LK_EXCLUSIVE | LK_RETRY, curproc /*XXX*/);
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 	/* NOTE: vnode now unlocked (unless vnislocked) */
1714 
1715 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1716 
1717 	/*
1718 	 * result == unix style errno (0 == OK!)
1719 	 *
1720 	 * zero out rest of buffer (if needed)
1721 	 */
1722 
1723 	if (result == 0) {
1724 		got = wanted - uio.uio_resid;
1725 
1726 		if (wanted && got == 0) {
1727 			result = EIO;		/* XXX: error? */
1728 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1729 			bzero((void *) (kva + got),
1730 			      (npages << PAGE_SHIFT) - got);
1731 		}
1732 	}
1733 
1734 	/*
1735 	 * now remove pager mapping
1736 	 */
1737 	uvm_pagermapout(kva, npages);
1738 
1739 	/*
1740 	 * now clean up the object (i.e. drop I/O count)
1741 	 */
1742 
1743 	simple_lock(&uvn->u_obj.vmobjlock);
1744 	/* NOTE: object now locked! */
1745 
1746 	uvn->u_nio--;			/* I/O DONE! */
1747 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1748 		wakeup(&uvn->u_nio);
1749 	}
1750 	simple_unlock(&uvn->u_obj.vmobjlock);
1751 	/* NOTE: object now unlocked! */
1752 
1753 	/*
1754 	 * done!
1755 	 */
1756 
1757 	UVMHIST_LOG(maphist, "<- done (result %d)", result,0,0,0);
1758 	if (result == 0)
1759 		return(VM_PAGER_OK);
1760 	else
1761 		return(VM_PAGER_ERROR);
1762 }
1763 
1764 /*
1765  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1766  * is gone we will kill the object (flushing dirty pages back to the vnode
1767  * if needed).
1768  *
1769  * => returns TRUE if there was no uvm_object attached or if there was
1770  *	one and we killed it [i.e. if there is no active uvn]
1771  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1772  *	needed]
1773  *
1774  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1775  *	having to hold a reference on the vnode) and given a working
1776  *	uvm_vnp_sync(), how does that effect the need for this function?
1777  *      [XXXCDC: seems like it can die?]
1778  *
1779  * => XXX: this function should DIE once we merge the VM and buffer
1780  *	cache.
1781  *
1782  * research shows that this is called in the following places:
1783  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1784  *	changes sizes
1785  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1786  *	are written to
1787  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1788  *	is off
1789  * ffs_realloccg: when we can't extend the current block and have
1790  *	to allocate a new one we call this [XXX: why?]
1791  * nfsrv_rename, rename_files: called when the target filename is there
1792  *	and we want to remove it
1793  * nfsrv_remove, sys_unlink: called on file we are removing
1794  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1795  *	then return "text busy"
1796  * nfs_open: seems to uncache any file opened with nfs
1797  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1798  */
1799 
1800 boolean_t
1801 uvm_vnp_uncache(vp)
1802 	struct vnode *vp;
1803 {
1804 	struct uvm_vnode *uvn = &vp->v_uvm;
1805 
1806 	/*
1807 	 * lock uvn part of the vnode and check to see if we need to do anything
1808 	 */
1809 
1810 	simple_lock(&uvn->u_obj.vmobjlock);
1811 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1812 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1813 		simple_unlock(&uvn->u_obj.vmobjlock);
1814 		return(TRUE);
1815 	}
1816 
1817 	/*
1818 	 * we have a valid, non-blocked uvn.   clear persist flag.
1819 	 * if uvn is currently active we can return now.
1820 	 */
1821 
1822 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1823 	if (uvn->u_obj.uo_refs) {
1824 		simple_unlock(&uvn->u_obj.vmobjlock);
1825 		return(FALSE);
1826 	}
1827 
1828 	/*
1829 	 * uvn is currently persisting!   we have to gain a reference to
1830 	 * it so that we can call uvn_detach to kill the uvn.
1831 	 */
1832 
1833 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1834 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1835 	simple_unlock(&uvn->u_obj.vmobjlock);
1836 
1837 
1838 #ifdef DEBUG
1839 	/*
1840 	 * carry over sanity check from old vnode pager: the vnode should
1841 	 * be VOP_LOCK'd, and we confirm it here.
1842 	 */
1843 	if (!VOP_ISLOCKED(vp)) {
1844 		boolean_t is_ok_anyway = FALSE;
1845 #ifdef NFS
1846 		extern int (**nfsv2_vnodeop_p) __P((void *));
1847 		extern int (**spec_nfsv2nodeop_p) __P((void *));
1848 		extern int (**fifo_nfsv2nodeop_p) __P((void *));
1849 
1850 		/* vnode is NOT VOP_LOCKed: some vnode types _never_ lock */
1851 		if (vp->v_op == nfsv2_vnodeop_p ||
1852 		    vp->v_op == spec_nfsv2nodeop_p) {
1853 			is_ok_anyway = TRUE;
1854 		}
1855 		if (vp->v_op == fifo_nfsv2nodeop_p) {
1856 			is_ok_anyway = TRUE;
1857 		}
1858 #endif	/* NFS */
1859 		if (!is_ok_anyway)
1860 			panic("uvm_vnp_uncache: vnode not locked!");
1861 	}
1862 #endif	/* DEBUG */
1863 
1864 	/*
1865 	 * now drop our reference to the vnode.   if we have the sole
1866 	 * reference to the vnode then this will cause it to die [as we
1867 	 * just cleared the persist flag].   we have to unlock the vnode
1868 	 * while we are doing this as it may trigger I/O.
1869 	 *
1870 	 * XXX: it might be possible for uvn to get reclaimed while we are
1871 	 * unlocked causing us to return TRUE when we should not.   we ignore
1872 	 * this as a false-positive return value doesn't hurt us.
1873 	 */
1874 	VOP_UNLOCK(vp, 0, curproc /*XXX*/);
1875 	uvn_detach(&uvn->u_obj);
1876 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curproc/*XXX*/);
1877 
1878 	/*
1879 	 * and return...
1880 	 */
1881 
1882 	return(TRUE);
1883 }
1884 
1885 /*
1886  * uvm_vnp_setsize: grow or shrink a vnode uvn
1887  *
1888  * grow   => just update size value
1889  * shrink => toss un-needed pages
1890  *
1891  * => we assume that the caller has a reference of some sort to the
1892  *	vnode in question so that it will not be yanked out from under
1893  *	us.
1894  *
1895  * called from:
1896  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1897  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1898  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1899  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1900  *  => union fs: union_newsize
1901  */
1902 
1903 void
1904 uvm_vnp_setsize(vp, newsize)
1905 	struct vnode *vp;
1906 	u_quad_t newsize;
1907 {
1908 	struct uvm_vnode *uvn = &vp->v_uvm;
1909 
1910 	/*
1911 	 * lock uvn and check for valid object, and if valid: do it!
1912 	 */
1913 	simple_lock(&uvn->u_obj.vmobjlock);
1914 	if (uvn->u_flags & UVM_VNODE_VALID) {
1915 
1916 		/*
1917 		 * make sure that the newsize fits within a vaddr_t
1918 		 * XXX: need to revise addressing data types
1919 		 */
1920 
1921 		if (newsize > (vaddr_t) -PAGE_SIZE) {
1922 #ifdef DEBUG
1923 			printf("uvm_vnp_setsize: vn %p size truncated "
1924 			    "%qx->%lx\n", vp, newsize, (vaddr_t)-PAGE_SIZE);
1925 #endif
1926 			newsize = (vaddr_t)-PAGE_SIZE;
1927 		}
1928 
1929 		/*
1930 		 * now check if the size has changed: if we shrink we had better
1931 		 * toss some pages...
1932 		 */
1933 
1934 		if (uvn->u_size > newsize) {
1935 			(void)uvn_flush(&uvn->u_obj, (vaddr_t) newsize,
1936 			    uvn->u_size, PGO_FREE);
1937 		}
1938 		uvn->u_size = (vaddr_t)newsize;
1939 	}
1940 	simple_unlock(&uvn->u_obj.vmobjlock);
1941 
1942 	/*
1943 	 * done
1944 	 */
1945 	return;
1946 }
1947 
1948 /*
1949  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1950  *
1951  * => called from sys_sync with no VM structures locked
1952  * => only one process can do a sync at a time (because the uvn
1953  *    structure only has one queue for sync'ing).  we ensure this
1954  *    by holding the uvn_sync_lock while the sync is in progress.
1955  *    other processes attempting a sync will sleep on this lock
1956  *    until we are done.
1957  */
1958 
1959 void
1960 uvm_vnp_sync(mp)
1961 	struct mount *mp;
1962 {
1963 	struct uvm_vnode *uvn;
1964 	struct vnode *vp;
1965 	boolean_t got_lock;
1966 
1967 	/*
1968 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1969 	 * our lock...
1970 	 */
1971 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0, curproc /*XXX*/);
1972 
1973 	/*
1974 	 * step 2: build up a simpleq of uvns of interest based on the
1975 	 * write list.   we gain a reference to uvns of interest.  must
1976 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1977 	 * in the body of the loop.
1978 	 */
1979 	SIMPLEQ_INIT(&uvn_sync_q);
1980 	simple_lock(&uvn_wl_lock);
1981 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1982 	    uvn = uvn->u_wlist.le_next) {
1983 
1984 		vp = (struct vnode *) uvn;
1985 		if (mp && vp->v_mount != mp)
1986 			continue;
1987 
1988 		/* attempt to gain reference */
1989 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1990 		    						FALSE &&
1991 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1992 			/* spin */ ;
1993 
1994 		/*
1995 		 * we will exit the loop if either if the following are true:
1996 		 *  - we got the lock [always true if NCPU == 1]
1997 		 *  - we failed to get the lock but noticed the vnode was
1998 		 * 	"blocked" -- in this case the vnode must be a dying
1999 		 *	vnode, and since dying vnodes are in the process of
2000 		 *	being flushed out, we can safely skip this one
2001 		 *
2002 		 * we want to skip over the vnode if we did not get the lock,
2003 		 * or if the vnode is already dying (due to the above logic).
2004 		 *
2005 		 * note that uvn must already be valid because we found it on
2006 		 * the wlist (this also means it can't be ALOCK'd).
2007 		 */
2008 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
2009 			if (got_lock)
2010 				simple_unlock(&uvn->u_obj.vmobjlock);
2011 			continue;		/* skip it */
2012 		}
2013 
2014 		/*
2015 		 * gain reference.   watch out for persisting uvns (need to
2016 		 * regain vnode REF).
2017 		 */
2018 		if (uvn->u_obj.uo_refs == 0)
2019 			VREF(vp);
2020 		uvn->u_obj.uo_refs++;
2021 		simple_unlock(&uvn->u_obj.vmobjlock);
2022 
2023 		/*
2024 		 * got it!
2025 		 */
2026 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
2027 	}
2028 	simple_unlock(&uvn_wl_lock);
2029 
2030 	/*
2031 	 * step 3: we now have a list of uvn's that may need cleaning.
2032 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
2033 	 * (so we can now safely lock uvn's again).
2034 	 */
2035 
2036 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2037 		simple_lock(&uvn->u_obj.vmobjlock);
2038 #ifdef DIAGNOSTIC
2039 		if (uvn->u_flags & UVM_VNODE_DYING) {
2040 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2041 		}
2042 #endif
2043 		uvn_flush(&uvn->u_obj, 0, 0,
2044 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2045 
2046 		/*
2047 		 * if we have the only reference and we just cleaned the uvn,
2048 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2049 		 * thus allowing us to avoid thinking about flushing it again
2050 		 * on later sync ops.
2051 		 */
2052 		if (uvn->u_obj.uo_refs == 1 &&
2053 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2054 			LIST_REMOVE(uvn, u_wlist);
2055 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2056 		}
2057 
2058 		simple_unlock(&uvn->u_obj.vmobjlock);
2059 
2060 		/* now drop our reference to the uvn */
2061 		uvn_detach(&uvn->u_obj);
2062 	}
2063 
2064 	/*
2065 	 * done!  release sync lock
2066 	 */
2067 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0, curproc /*XXX*/);
2068 }
2069