xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision bada23909e740596d0a3785a73bd3583a9807fb8)
1 /*	$NetBSD: uvm_vnode.c,v 1.19 1999/03/04 06:48:54 chs Exp $	*/
2 
3 /*
4  * XXXCDC: "ROUGH DRAFT" QUALITY UVM PRE-RELEASE FILE!
5  *         >>>USE AT YOUR OWN RISK, WORK IS NOT FINISHED<<<
6  */
7 /*
8  * Copyright (c) 1997 Charles D. Cranor and Washington University.
9  * Copyright (c) 1991, 1993
10  *      The Regents of the University of California.
11  * Copyright (c) 1990 University of Utah.
12  *
13  * All rights reserved.
14  *
15  * This code is derived from software contributed to Berkeley by
16  * the Systems Programming Group of the University of Utah Computer
17  * Science Department.
18  *
19  * Redistribution and use in source and binary forms, with or without
20  * modification, are permitted provided that the following conditions
21  * are met:
22  * 1. Redistributions of source code must retain the above copyright
23  *    notice, this list of conditions and the following disclaimer.
24  * 2. Redistributions in binary form must reproduce the above copyright
25  *    notice, this list of conditions and the following disclaimer in the
26  *    documentation and/or other materials provided with the distribution.
27  * 3. All advertising materials mentioning features or use of this software
28  *    must display the following acknowledgement:
29  *      This product includes software developed by Charles D. Cranor,
30  *	Washington University, the University of California, Berkeley and
31  *	its contributors.
32  * 4. Neither the name of the University nor the names of its contributors
33  *    may be used to endorse or promote products derived from this software
34  *    without specific prior written permission.
35  *
36  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
37  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
38  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
39  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
40  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
41  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
42  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
43  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
44  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
45  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
46  * SUCH DAMAGE.
47  *
48  *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
49  * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
50  */
51 
52 #include "fs_nfs.h"
53 #include "opt_uvmhist.h"
54 
55 /*
56  * uvm_vnode.c: the vnode pager.
57  */
58 
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/proc.h>
62 #include <sys/malloc.h>
63 #include <sys/vnode.h>
64 #include <sys/disklabel.h>
65 #include <sys/ioctl.h>
66 #include <sys/fcntl.h>
67 #include <sys/conf.h>
68 
69 #include <miscfs/specfs/specdev.h>
70 
71 #include <vm/vm.h>
72 #include <vm/vm_page.h>
73 #include <vm/vm_kern.h>
74 
75 #include <uvm/uvm.h>
76 #include <uvm/uvm_vnode.h>
77 
78 /*
79  * private global data structure
80  *
81  * we keep a list of writeable active vnode-backed VM objects for sync op.
82  * we keep a simpleq of vnodes that are currently being sync'd.
83  */
84 
85 LIST_HEAD(uvn_list_struct, uvm_vnode);
86 static struct uvn_list_struct uvn_wlist;	/* writeable uvns */
87 static simple_lock_data_t uvn_wl_lock;		/* locks uvn_wlist */
88 
89 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
90 static struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
91 lock_data_t uvn_sync_lock;			/* locks sync operation */
92 
93 /*
94  * functions
95  */
96 
97 static int		   uvn_asyncget __P((struct uvm_object *, vaddr_t,
98 					    int));
99 struct uvm_object 	  *uvn_attach __P((void *, vm_prot_t));
100 static void		   uvn_cluster __P((struct uvm_object *, vaddr_t,
101 					   vaddr_t *, vaddr_t *));
102 static void                uvn_detach __P((struct uvm_object *));
103 static boolean_t           uvn_flush __P((struct uvm_object *, vaddr_t,
104 					 vaddr_t, int));
105 static int                 uvn_get __P((struct uvm_object *, vaddr_t,
106 					vm_page_t *, int *, int,
107 					vm_prot_t, int, int));
108 static void		   uvn_init __P((void));
109 static int		   uvn_io __P((struct uvm_vnode *, vm_page_t *,
110 				      int, int, int));
111 static int		   uvn_put __P((struct uvm_object *, vm_page_t *,
112 					int, boolean_t));
113 static void                uvn_reference __P((struct uvm_object *));
114 static boolean_t	   uvn_releasepg __P((struct vm_page *,
115 					      struct vm_page **));
116 
117 /*
118  * master pager structure
119  */
120 
121 struct uvm_pagerops uvm_vnodeops = {
122 	uvn_init,
123 	uvn_attach,
124 	uvn_reference,
125 	uvn_detach,
126 	NULL,			/* no specialized fault routine required */
127 	uvn_flush,
128 	uvn_get,
129 	uvn_asyncget,
130 	uvn_put,
131 	uvn_cluster,
132 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
133 	uvm_shareprot,	 /* !NULL: allow us in share maps */
134 	NULL,		 /* AIO-DONE function (not until we have asyncio) */
135 	uvn_releasepg,
136 };
137 
138 /*
139  * the ops!
140  */
141 
142 /*
143  * uvn_init
144  *
145  * init pager private data structures.
146  */
147 
148 static void
149 uvn_init()
150 {
151 
152 	LIST_INIT(&uvn_wlist);
153 	simple_lock_init(&uvn_wl_lock);
154 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
155 	lockinit(&uvn_sync_lock, PVM, "uvnsync", 0, 0);
156 }
157 
158 /*
159  * uvn_attach
160  *
161  * attach a vnode structure to a VM object.  if the vnode is already
162  * attached, then just bump the reference count by one and return the
163  * VM object.   if not already attached, attach and return the new VM obj.
164  * the "accessprot" tells the max access the attaching thread wants to
165  * our pages.
166  *
167  * => caller must _not_ already be holding the lock on the uvm_object.
168  * => in fact, nothing should be locked so that we can sleep here.
169  * => note that uvm_object is first thing in vnode structure, so their
170  *    pointers are equiv.
171  */
172 
173 struct uvm_object *
174 uvn_attach(arg, accessprot)
175 	void *arg;
176 	vm_prot_t accessprot;
177 {
178 	struct vnode *vp = arg;
179 	struct uvm_vnode *uvn = &vp->v_uvm;
180 	struct vattr vattr;
181 	int oldflags, result;
182 	struct partinfo pi;
183 	u_quad_t used_vnode_size;
184 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
185 
186 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
187 
188 	used_vnode_size = (u_quad_t)0;	/* XXX gcc -Wuninitialized */
189 
190 	/*
191 	 * first get a lock on the uvn.
192 	 */
193 	simple_lock(&uvn->u_obj.vmobjlock);
194 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
195 		uvn->u_flags |= UVM_VNODE_WANTED;
196 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
197 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
198 		    "uvn_attach", 0);
199 		simple_lock(&uvn->u_obj.vmobjlock);
200 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
201 	}
202 
203 	/*
204 	 * if we're mapping a BLK device, make sure it is a disk.
205 	 */
206 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
207 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
208 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
209 		return(NULL);
210 	}
211 
212 	/*
213 	 * now we have lock and uvn must not be in a blocked state.
214 	 * first check to see if it is already active, in which case
215 	 * we can bump the reference count, check to see if we need to
216 	 * add it to the writeable list, and then return.
217 	 */
218 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
219 
220 		/* regain VREF if we were persisting */
221 		if (uvn->u_obj.uo_refs == 0) {
222 			VREF(vp);
223 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
224 			    0,0,0,0);
225 		}
226 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
227 
228 		/* check for new writeable uvn */
229 		if ((accessprot & VM_PROT_WRITE) != 0 &&
230 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
231 			simple_lock(&uvn_wl_lock);
232 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
233 			simple_unlock(&uvn_wl_lock);
234 			/* we are now on wlist! */
235 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
236 		}
237 
238 		/* unlock and return */
239 		simple_unlock(&uvn->u_obj.vmobjlock);
240 		UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
241 		    0, 0, 0);
242 		return (&uvn->u_obj);
243 	}
244 
245 	/*
246 	 * need to call VOP_GETATTR() to get the attributes, but that could
247 	 * block (due to I/O), so we want to unlock the object before calling.
248 	 * however, we want to keep anyone else from playing with the object
249 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
250 	 * prevents anyone from attaching to the vnode until we are done with
251 	 * it.
252 	 */
253 	uvn->u_flags = UVM_VNODE_ALOCK;
254 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
255 		/* XXX: curproc? */
256 
257 	if (vp->v_type == VBLK) {
258 		/*
259 		 * We could implement this as a specfs getattr call, but:
260 		 *
261 		 *	(1) VOP_GETATTR() would get the file system
262 		 *	    vnode operation, not the specfs operation.
263 		 *
264 		 *	(2) All we want is the size, anyhow.
265 		 */
266 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
267 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
268 		if (result == 0) {
269 			/* XXX should remember blocksize */
270 			used_vnode_size = (u_quad_t)pi.disklab->d_secsize *
271 			    (u_quad_t)pi.part->p_size;
272 		}
273 	} else {
274 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
275 		if (result == 0)
276 			used_vnode_size = vattr.va_size;
277 	}
278 
279 	/* relock object */
280 	simple_lock(&uvn->u_obj.vmobjlock);
281 
282 	if (result != 0) {
283 		if (uvn->u_flags & UVM_VNODE_WANTED)
284 			wakeup(uvn);
285 		uvn->u_flags = 0;
286 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
287 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
288 		return(NULL);
289 	}
290 
291 	/*
292 	 * make sure that the newsize fits within a vaddr_t
293 	 * XXX: need to revise addressing data types
294 	 */
295 #ifdef DEBUG
296 	if (vp->v_type == VBLK)
297 		printf("used_vnode_size = %qu\n", (long long)used_vnode_size);
298 #endif
299 	if (used_vnode_size > (vaddr_t) -PAGE_SIZE) {
300 #ifdef DEBUG
301 		printf("uvn_attach: vn %p size truncated %qx->%x\n", vp,
302 		    (long long)used_vnode_size, -PAGE_SIZE);
303 #endif
304 		used_vnode_size = (vaddr_t) -PAGE_SIZE;
305 	}
306 
307 	/*
308 	 * now set up the uvn.
309 	 */
310 	uvn->u_obj.pgops = &uvm_vnodeops;
311 	TAILQ_INIT(&uvn->u_obj.memq);
312 	uvn->u_obj.uo_npages = 0;
313 	uvn->u_obj.uo_refs = 1;			/* just us... */
314 	oldflags = uvn->u_flags;
315 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
316 	uvn->u_nio = 0;
317 	uvn->u_size = used_vnode_size;
318 
319 	/* if write access, we need to add it to the wlist */
320 	if (accessprot & VM_PROT_WRITE) {
321 		simple_lock(&uvn_wl_lock);
322 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
323 		simple_unlock(&uvn_wl_lock);
324 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
325 	}
326 
327 	/*
328 	 * add a reference to the vnode.   this reference will stay as long
329 	 * as there is a valid mapping of the vnode.   dropped when the
330 	 * reference count goes to zero [and we either free or persist].
331 	 */
332 	VREF(vp);
333 	simple_unlock(&uvn->u_obj.vmobjlock);
334 	if (oldflags & UVM_VNODE_WANTED)
335 		wakeup(uvn);
336 
337 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
338 	return(&uvn->u_obj);
339 }
340 
341 
342 /*
343  * uvn_reference
344  *
345  * duplicate a reference to a VM object.  Note that the reference
346  * count must already be at least one (the passed in reference) so
347  * there is no chance of the uvn being killed or locked out here.
348  *
349  * => caller must call with object unlocked.
350  * => caller must be using the same accessprot as was used at attach time
351  */
352 
353 
354 static void
355 uvn_reference(uobj)
356 	struct uvm_object *uobj;
357 {
358 #ifdef DIAGNOSTIC
359 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
360 #endif
361 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
362 
363 	simple_lock(&uobj->vmobjlock);
364 #ifdef DIAGNOSTIC
365 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
366 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
367 		    uobj->uo_refs);
368 		panic("uvn_reference: invalid state");
369 	}
370 #endif
371 	uobj->uo_refs++;
372 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
373 	uobj, uobj->uo_refs,0,0);
374 	simple_unlock(&uobj->vmobjlock);
375 }
376 
377 /*
378  * uvn_detach
379  *
380  * remove a reference to a VM object.
381  *
382  * => caller must call with object unlocked and map locked.
383  * => this starts the detach process, but doesn't have to finish it
384  *    (async i/o could still be pending).
385  */
386 static void
387 uvn_detach(uobj)
388 	struct uvm_object *uobj;
389 {
390 	struct uvm_vnode *uvn;
391 	struct vnode *vp;
392 	int oldflags;
393 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
394 
395 	simple_lock(&uobj->vmobjlock);
396 
397 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
398 	uobj->uo_refs--;			/* drop ref! */
399 	if (uobj->uo_refs) {			/* still more refs */
400 		simple_unlock(&uobj->vmobjlock);
401 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
402 		return;
403 	}
404 
405 	/*
406 	 * get other pointers ...
407 	 */
408 
409 	uvn = (struct uvm_vnode *) uobj;
410 	vp = (struct vnode *) uobj;
411 
412 	/*
413 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
414 	 * to keep an active text file from being overwritten).
415 	 */
416 	vp->v_flag &= ~VTEXT;
417 
418 	/*
419 	 * we just dropped the last reference to the uvn.   see if we can
420 	 * let it "stick around".
421 	 */
422 
423 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
424 		/* won't block */
425 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
426 		simple_unlock(&uobj->vmobjlock);
427 		vrele(vp);			/* drop vnode reference */
428 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
429 		return;
430 	}
431 
432 	/*
433 	 * its a goner!
434 	 */
435 
436 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
437 
438 	uvn->u_flags |= UVM_VNODE_DYING;
439 
440 	/*
441 	 * even though we may unlock in flush, no one can gain a reference
442 	 * to us until we clear the "dying" flag [because it blocks
443 	 * attaches].  we will not do that until after we've disposed of all
444 	 * the pages with uvn_flush().  note that before the flush the only
445 	 * pages that could be marked PG_BUSY are ones that are in async
446 	 * pageout by the daemon.  (there can't be any pending "get"'s
447 	 * because there are no references to the object).
448 	 */
449 
450 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
451 
452 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
453 
454 	/*
455 	 * given the structure of this pager, the above flush request will
456 	 * create the following state: all the pages that were in the object
457 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
458 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
459 	 * in either case, if we have pages left we can't kill the object
460 	 * yet because i/o is pending.  in this case we set the "relkill"
461 	 * flag which will cause pgo_releasepg to kill the object once all
462 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
463 	 * routine or from the page daemon].
464 	 */
465 
466 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
467 #ifdef DIAGNOSTIC
468 		/*
469 		 * XXXCDC: very unlikely to happen until we have async i/o
470 		 * so print a little info message in case it does.
471 		 */
472 		printf("uvn_detach: vn %p has pages left after flush - "
473 		    "relkill mode\n", uobj);
474 #endif
475 		uvn->u_flags |= UVM_VNODE_RELKILL;
476 		simple_unlock(&uobj->vmobjlock);
477 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
478 		    0, 0);
479 		return;
480 	}
481 
482 	/*
483 	 * kill object now.   note that we can't be on the sync q because
484 	 * all references are gone.
485 	 */
486 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
487 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
488 		LIST_REMOVE(uvn, u_wlist);
489 		simple_unlock(&uvn_wl_lock);
490 	}
491 #ifdef DIAGNOSTIC
492 	if (uobj->memq.tqh_first != NULL)
493 		panic("uvn_deref: vnode VM object still has pages afer "
494 		    "syncio/free flush");
495 #endif
496 	oldflags = uvn->u_flags;
497 	uvn->u_flags = 0;
498 	simple_unlock(&uobj->vmobjlock);
499 
500 	/* wake up any sleepers */
501 	if (oldflags & UVM_VNODE_WANTED)
502 		wakeup(uvn);
503 
504 	/*
505 	 * drop our reference to the vnode.
506 	 */
507 	vrele(vp);
508 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
509 
510 	return;
511 }
512 
513 /*
514  * uvm_vnp_terminate: external hook to clear out a vnode's VM
515  *
516  * called in two cases:
517  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
518  *      count) needs to be freed so that a vnode can be reused.  this
519  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
520  *      the free list is still attached (i.e. not VBAD) then vgone is
521  *	called.   as part of the vgone trace this should get called to
522  *	free the vm object.   this is the common case.
523  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
524  *	"umount -f") the vgone() function is called on active vnodes
525  *	on the mounted file systems to kill their data (the vnodes become
526  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
527  *	call here (even if the uvn is still in use -- i.e. has a non-zero
528  *	reference count).  this case happens at "umount -f" and during a
529  *	"reboot/halt" operation.
530  *
531  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
532  *	[protects us from getting a vnode that is already in the DYING
533  *	 state...]
534  * => unlike uvn_detach, this function must not return until all the
535  *	uvn's pages are disposed of.
536  * => in case [2] the uvn is still alive after this call, but all I/O
537  *	ops will fail (due to the backing vnode now being "dead").  this
538  *	will prob. kill any process using the uvn due to pgo_get failing.
539  */
540 
541 void
542 uvm_vnp_terminate(vp)
543 	struct vnode *vp;
544 {
545 	struct uvm_vnode *uvn = &vp->v_uvm;
546 	int oldflags;
547 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
548 
549 	/*
550 	 * lock object and check if it is valid
551 	 */
552 	simple_lock(&uvn->u_obj.vmobjlock);
553 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
554 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
555 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
556 		simple_unlock(&uvn->u_obj.vmobjlock);
557 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
558 		return;
559 	}
560 
561 	/*
562 	 * must be a valid uvn that is not already dying (because XLOCK
563 	 * protects us from that).   the uvn can't in the the ALOCK state
564 	 * because it is valid, and uvn's that are in the ALOCK state haven't
565 	 * been marked valid yet.
566 	 */
567 
568 #ifdef DEBUG
569 	/*
570 	 * debug check: are we yanking the vnode out from under our uvn?
571 	 */
572 	if (uvn->u_obj.uo_refs) {
573 		printf("uvm_vnp_terminate(%p): terminating active vnode "
574 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
575 	}
576 #endif
577 
578 	/*
579 	 * it is possible that the uvn was detached and is in the relkill
580 	 * state [i.e. waiting for async i/o to finish so that releasepg can
581 	 * kill object].  we take over the vnode now and cancel the relkill.
582 	 * we want to know when the i/o is done so we can recycle right
583 	 * away.   note that a uvn can only be in the RELKILL state if it
584 	 * has a zero reference count.
585 	 */
586 
587 	if (uvn->u_flags & UVM_VNODE_RELKILL)
588 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
589 
590 	/*
591 	 * block the uvn by setting the dying flag, and then flush the
592 	 * pages.  (note that flush may unlock object while doing I/O, but
593 	 * it will re-lock it before it returns control here).
594 	 *
595 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
596 	 * that uvn_io doesn't attempt to VOP_LOCK again.
597 	 *
598 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
599 	 *	due to a forceful unmount might not be a good idea.  maybe we
600 	 *	need a way to pass in this info to uvn_flush through a
601 	 *	pager-defined PGO_ constant [currently there are none].
602 	 */
603 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
604 
605 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
606 
607 	/*
608 	 * as we just did a flush we expect all the pages to be gone or in
609 	 * the process of going.  sleep to wait for the rest to go [via iosync].
610 	 */
611 
612 	while (uvn->u_obj.uo_npages) {
613 #ifdef DIAGNOSTIC
614 		struct vm_page *pp;
615 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
616 		     pp = pp->listq.tqe_next) {
617 			if ((pp->flags & PG_BUSY) == 0)
618 				panic("uvm_vnp_terminate: detected unbusy pg");
619 		}
620 		if (uvn->u_nio == 0)
621 			panic("uvm_vnp_terminate: no I/O to wait for?");
622 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
623 		/*
624 		 * XXXCDC: this is unlikely to happen without async i/o so we
625 		 * put a printf in just to keep an eye on it.
626 		 */
627 #endif
628 		uvn->u_flags |= UVM_VNODE_IOSYNC;
629 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
630 		    "uvn_term",0);
631 		simple_lock(&uvn->u_obj.vmobjlock);
632 	}
633 
634 	/*
635 	 * done.   now we free the uvn if its reference count is zero
636 	 * (true if we are zapping a persisting uvn).   however, if we are
637 	 * terminating a uvn with active mappings we let it live ... future
638 	 * calls down to the vnode layer will fail.
639 	 */
640 
641 	oldflags = uvn->u_flags;
642 	if (uvn->u_obj.uo_refs) {
643 
644 		/*
645 		 * uvn must live on it is dead-vnode state until all references
646 		 * are gone.   restore flags.    clear CANPERSIST state.
647 		 */
648 
649 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
650 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
651 
652 	} else {
653 
654 		/*
655 		 * free the uvn now.   note that the VREF reference is already
656 		 * gone [it is dropped when we enter the persist state].
657 		 */
658 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
659 			panic("uvm_vnp_terminate: io sync wanted bit set");
660 
661 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
662 			simple_lock(&uvn_wl_lock);
663 			LIST_REMOVE(uvn, u_wlist);
664 			simple_unlock(&uvn_wl_lock);
665 		}
666 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
667 	}
668 
669 	if (oldflags & UVM_VNODE_WANTED)
670 		wakeup(uvn);		/* object lock still held */
671 
672 	simple_unlock(&uvn->u_obj.vmobjlock);
673 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
674 
675 }
676 
677 /*
678  * uvn_releasepg: handled a released page in a uvn
679  *
680  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
681  *	to dispose of.
682  * => caller must handled PG_WANTED case
683  * => called with page's object locked, pageq's unlocked
684  * => returns TRUE if page's object is still alive, FALSE if we
685  *	killed the page's object.    if we return TRUE, then we
686  *	return with the object locked.
687  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
688  *				with the page queues locked [for pagedaemon]
689  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
690  * => we kill the uvn if it is not referenced and we are suppose to
691  *	kill it ("relkill").
692  */
693 
694 boolean_t
695 uvn_releasepg(pg, nextpgp)
696 	struct vm_page *pg;
697 	struct vm_page **nextpgp;	/* OUT */
698 {
699 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
700 #ifdef DIAGNOSTIC
701 	if ((pg->flags & PG_RELEASED) == 0)
702 		panic("uvn_releasepg: page not released!");
703 #endif
704 
705 	/*
706 	 * dispose of the page [caller handles PG_WANTED]
707 	 */
708 	pmap_page_protect(PMAP_PGARG(pg), VM_PROT_NONE);
709 	uvm_lock_pageq();
710 	if (nextpgp)
711 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
712 	uvm_pagefree(pg);
713 	if (!nextpgp)
714 		uvm_unlock_pageq();
715 
716 	/*
717 	 * now see if we need to kill the object
718 	 */
719 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
720 		if (uvn->u_obj.uo_refs)
721 			panic("uvn_releasepg: kill flag set on referenced "
722 			    "object!");
723 		if (uvn->u_obj.uo_npages == 0) {
724 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
725 				simple_lock(&uvn_wl_lock);
726 				LIST_REMOVE(uvn, u_wlist);
727 				simple_unlock(&uvn_wl_lock);
728 			}
729 #ifdef DIAGNOSTIC
730 			if (uvn->u_obj.memq.tqh_first)
731 	panic("uvn_releasepg: pages in object with npages == 0");
732 #endif
733 			if (uvn->u_flags & UVM_VNODE_WANTED)
734 				/* still holding object lock */
735 				wakeup(uvn);
736 
737 			uvn->u_flags = 0;		/* DEAD! */
738 			simple_unlock(&uvn->u_obj.vmobjlock);
739 			return (FALSE);
740 		}
741 	}
742 	return (TRUE);
743 }
744 
745 /*
746  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
747  * through the buffer cache and allow I/O in any size.  These VOPs use
748  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
749  * go through the buffer cache or allow I/O sizes larger than a
750  * block].  we will eventually want to change this.
751  *
752  * issues to consider:
753  *   uvm provides the uvm_aiodesc structure for async i/o management.
754  * there are two tailq's in the uvm. structure... one for pending async
755  * i/o and one for "done" async i/o.   to do an async i/o one puts
756  * an aiodesc on the "pending" list (protected by splbio()), starts the
757  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
758  * some sort of "i/o done" function to be called (at splbio(), interrupt
759  * time).   this function should remove the aiodesc from the pending list
760  * and place it on the "done" list and wakeup the daemon.   the daemon
761  * will run at normal spl() and will remove all items from the "done"
762  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
763  * [in the old vm code, this was done by calling the "put" routine with
764  * null arguments which made the code harder to read and understand because
765  * you had one function ("put") doing two things.]
766  *
767  * so the current pager needs:
768  *   int uvn_aiodone(struct uvm_aiodesc *)
769  *
770  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
771  *	later collection.
772  * => called with pageq's locked by the daemon.
773  *
774  * general outline:
775  * - "try" to lock object.   if fail, just return (will try again later)
776  * - drop "u_nio" (this req is done!)
777  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
778  * - get "page" structures (atop?).
779  * - handle "wanted" pages
780  * - handle "released" pages [using pgo_releasepg]
781  *   >>> pgo_releasepg may kill the object
782  * dont forget to look at "object" wanted flag in all cases.
783  */
784 
785 
786 /*
787  * uvn_flush: flush pages out of a uvm object.
788  *
789  * => object should be locked by caller.   we may _unlock_ the object
790  *	if (and only if) we need to clean a page (PGO_CLEANIT).
791  *	we return with the object locked.
792  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
793  *	might want to unlock higher level resources (e.g. vm_map)
794  *	before calling flush.
795  * => if PGO_CLEANIT is not set, then we will neither unlock the object
796  *	or block.
797  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
798  *	for flushing.
799  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
800  *	that new pages are inserted on the tail end of the list.   thus,
801  *	we can make a complete pass through the object in one go by starting
802  *	at the head and working towards the tail (new pages are put in
803  *	front of us).
804  * => NOTE: we are allowed to lock the page queues, so the caller
805  *	must not be holding the lock on them [e.g. pagedaemon had
806  *	better not call us with the queues locked]
807  * => we return TRUE unless we encountered some sort of I/O error
808  *
809  * comment on "cleaning" object and PG_BUSY pages:
810  *	this routine is holding the lock on the object.   the only time
811  *	that it can run into a PG_BUSY page that it does not own is if
812  *	some other process has started I/O on the page (e.g. either
813  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
814  *	in, then it can not be dirty (!PG_CLEAN) because no one has
815  *	had a chance to modify it yet.    if the PG_BUSY page is being
816  *	paged out then it means that someone else has already started
817  *	cleaning the page for us (how nice!).    in this case, if we
818  *	have syncio specified, then after we make our pass through the
819  *	object we need to wait for the other PG_BUSY pages to clear
820  *	off (i.e. we need to do an iosync).   also note that once a
821  *	page is PG_BUSY it must stay in its object until it is un-busyed.
822  *
823  * note on page traversal:
824  *	we can traverse the pages in an object either by going down the
825  *	linked list in "uobj->memq", or we can go over the address range
826  *	by page doing hash table lookups for each address.    depending
827  *	on how many pages are in the object it may be cheaper to do one
828  *	or the other.   we set "by_list" to true if we are using memq.
829  *	if the cost of a hash lookup was equal to the cost of the list
830  *	traversal we could compare the number of pages in the start->stop
831  *	range to the total number of pages in the object.   however, it
832  *	seems that a hash table lookup is more expensive than the linked
833  *	list traversal, so we multiply the number of pages in the
834  *	start->stop range by a penalty which we define below.
835  */
836 
837 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
838 
839 static boolean_t
840 uvn_flush(uobj, start, stop, flags)
841 	struct uvm_object *uobj;
842 	vaddr_t start, stop;
843 	int flags;
844 {
845 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
846 	struct vm_page *pp, *ppnext, *ptmp;
847 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
848 	int npages, result, lcv;
849 	boolean_t retval, need_iosync, by_list, needs_clean;
850 	vaddr_t curoff;
851 	u_short pp_version;
852 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
853 
854 	curoff = 0;	/* XXX: shut up gcc */
855 	/*
856 	 * get init vals and determine how we are going to traverse object
857 	 */
858 
859 	need_iosync = FALSE;
860 	retval = TRUE;		/* return value */
861 	if (flags & PGO_ALLPAGES) {
862 		start = 0;
863 		stop = round_page(uvn->u_size);
864 		by_list = TRUE;		/* always go by the list */
865 	} else {
866 		start = trunc_page(start);
867 		stop = round_page(stop);
868 		if (stop > round_page(uvn->u_size))
869 			printf("uvn_flush: strange, got an out of range "
870 			    "flush (fixed)\n");
871 
872 		by_list = (uobj->uo_npages <=
873 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
874 	}
875 
876 	UVMHIST_LOG(maphist,
877 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
878 	    start, stop, by_list, flags);
879 
880 	/*
881 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
882 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
883 	 * is wrong it will only prevent us from clustering... it won't break
884 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
885 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
886 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
887 	 * bit is always up to date since there are no mappings).
888 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
889 	 */
890 
891 	if ((flags & PGO_CLEANIT) != 0 &&
892 	    uobj->pgops->pgo_mk_pcluster != NULL) {
893 		if (by_list) {
894 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
895 			    pp = pp->listq.tqe_next) {
896 				if (pp->offset < start || pp->offset >= stop)
897 					continue;
898 				pp->flags &= ~PG_CLEANCHK;
899 			}
900 
901 		} else {   /* by hash */
902 			for (curoff = start ; curoff < stop;
903 			    curoff += PAGE_SIZE) {
904 				pp = uvm_pagelookup(uobj, curoff);
905 				if (pp)
906 					pp->flags &= ~PG_CLEANCHK;
907 			}
908 		}
909 	}
910 
911 	/*
912 	 * now do it.   note: we must update ppnext in body of loop or we
913 	 * will get stuck.  we need to use ppnext because we may free "pp"
914 	 * before doing the next loop.
915 	 */
916 
917 	if (by_list) {
918 		pp = uobj->memq.tqh_first;
919 	} else {
920 		curoff = start;
921 		pp = uvm_pagelookup(uobj, curoff);
922 	}
923 
924 	ppnext = NULL;	/* XXX: shut up gcc */
925 	ppsp = NULL;		/* XXX: shut up gcc */
926 	uvm_lock_pageq();	/* page queues locked */
927 
928 	/* locked: both page queues and uobj */
929 	for ( ; (by_list && pp != NULL) ||
930 	  (!by_list && curoff < stop) ; pp = ppnext) {
931 
932 		if (by_list) {
933 
934 			/*
935 			 * range check
936 			 */
937 
938 			if (pp->offset < start || pp->offset >= stop) {
939 				ppnext = pp->listq.tqe_next;
940 				continue;
941 			}
942 
943 		} else {
944 
945 			/*
946 			 * null check
947 			 */
948 
949 			curoff += PAGE_SIZE;
950 			if (pp == NULL) {
951 				if (curoff < stop)
952 					ppnext = uvm_pagelookup(uobj, curoff);
953 				continue;
954 			}
955 
956 		}
957 
958 		/*
959 		 * handle case where we do not need to clean page (either
960 		 * because we are not clean or because page is not dirty or
961 		 * is busy):
962 		 *
963 		 * NOTE: we are allowed to deactivate a non-wired active
964 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
965 		 * queue it must stay put until it is !PG_BUSY (so as not to
966 		 * confuse pagedaemon).
967 		 */
968 
969 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
970 			needs_clean = FALSE;
971 			if ((pp->flags & PG_BUSY) != 0 &&
972 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
973 			             (PGO_CLEANIT|PGO_SYNCIO))
974 				need_iosync = TRUE;
975 		} else {
976 			/*
977 			 * freeing: nuke all mappings so we can sync
978 			 * PG_CLEAN bit with no race
979 			 */
980 			if ((pp->flags & PG_CLEAN) != 0 &&
981 			    (flags & PGO_FREE) != 0 &&
982 			    (pp->pqflags & PQ_ACTIVE) != 0)
983 				pmap_page_protect(PMAP_PGARG(pp), VM_PROT_NONE);
984 			if ((pp->flags & PG_CLEAN) != 0 &&
985 			    pmap_is_modified(PMAP_PGARG(pp)))
986 				pp->flags &= ~(PG_CLEAN);
987 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
988 
989 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
990 		}
991 
992 		/*
993 		 * if we don't need a clean... load ppnext and dispose of pp
994 		 */
995 		if (!needs_clean) {
996 			/* load ppnext */
997 			if (by_list)
998 				ppnext = pp->listq.tqe_next;
999 			else {
1000 				if (curoff < stop)
1001 					ppnext = uvm_pagelookup(uobj, curoff);
1002 			}
1003 
1004 			/* now dispose of pp */
1005 			if (flags & PGO_DEACTIVATE) {
1006 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1007 				    pp->wire_count == 0) {
1008 					pmap_page_protect(PMAP_PGARG(pp),
1009 					    VM_PROT_NONE);
1010 					uvm_pagedeactivate(pp);
1011 				}
1012 
1013 			} else if (flags & PGO_FREE) {
1014 				if (pp->flags & PG_BUSY) {
1015 					/* release busy pages */
1016 					pp->flags |= PG_RELEASED;
1017 				} else {
1018 					pmap_page_protect(PMAP_PGARG(pp),
1019 					    VM_PROT_NONE);
1020 					/* removed page from object */
1021 					uvm_pagefree(pp);
1022 				}
1023 			}
1024 			/* ppnext is valid so we can continue... */
1025 			continue;
1026 		}
1027 
1028 		/*
1029 		 * pp points to a page in the locked object that we are
1030 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
1031 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1032 		 *
1033 		 * let uvm_pager_put attempted a clustered page out.
1034 		 * note: locked: uobj and page queues.
1035 		 */
1036 
1037 		pp->flags |= PG_BUSY;	/* we 'own' page now */
1038 		UVM_PAGE_OWN(pp, "uvn_flush");
1039 		pmap_page_protect(PMAP_PGARG(pp), VM_PROT_READ);
1040 		pp_version = pp->version;
1041 ReTry:
1042 		ppsp = pps;
1043 		npages = sizeof(pps) / sizeof(struct vm_page *);
1044 
1045 		/* locked: page queues, uobj */
1046 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1047 			   flags | PGO_DOACTCLUST, start, stop);
1048 		/* unlocked: page queues, uobj */
1049 
1050 		/*
1051 		 * at this point nothing is locked.   if we did an async I/O
1052 		 * it is remotely possible for the async i/o to complete and
1053 		 * the page "pp" be freed or what not before we get a chance
1054 		 * to relock the object.   in order to detect this, we have
1055 		 * saved the version number of the page in "pp_version".
1056 		 */
1057 
1058 		/* relock! */
1059 		simple_lock(&uobj->vmobjlock);
1060 		uvm_lock_pageq();
1061 
1062 		/*
1063 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1064 		 * can only happen when  we are doing async I/O and can't
1065 		 * map the pages into kernel memory (pager_map) due to lack
1066 		 * of vm space.   if this happens we drop back to sync I/O.
1067 		 */
1068 
1069 		if (result == VM_PAGER_AGAIN) {
1070 			/*
1071 			 * it is unlikely, but page could have been released
1072 			 * while we had the object lock dropped.   we ignore
1073 			 * this now and retry the I/O.  we will detect and
1074 			 * handle the released page after the syncio I/O
1075 			 * completes.
1076 			 */
1077 #ifdef DIAGNOSTIC
1078 			if (flags & PGO_SYNCIO)
1079 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1080 #endif
1081 			flags |= PGO_SYNCIO;
1082 			goto ReTry;
1083 		}
1084 
1085 		/*
1086 		 * the cleaning operation is now done.   finish up.  note that
1087 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1088 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1089 		 * to us in ppsp/npages.
1090 		 */
1091 
1092 		/*
1093 		 * for pending async i/o if we are not deactivating/freeing
1094 		 * we can move on to the next page.
1095 		 */
1096 
1097 		if (result == VM_PAGER_PEND) {
1098 
1099 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1100 				/*
1101 				 * no per-page ops: refresh ppnext and continue
1102 				 */
1103 				if (by_list) {
1104 					if (pp->version == pp_version)
1105 						ppnext = pp->listq.tqe_next;
1106 					else
1107 						/* reset */
1108 						ppnext = uobj->memq.tqh_first;
1109 				} else {
1110 					if (curoff < stop)
1111 						ppnext = uvm_pagelookup(uobj,
1112 						    curoff);
1113 				}
1114 				continue;
1115 			}
1116 
1117 			/* need to do anything here? */
1118 		}
1119 
1120 		/*
1121 		 * need to look at each page of the I/O operation.  we defer
1122 		 * processing "pp" until the last trip through this "for" loop
1123 		 * so that we can load "ppnext" for the main loop after we
1124 		 * play with the cluster pages [thus the "npages + 1" in the
1125 		 * loop below].
1126 		 */
1127 
1128 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1129 
1130 			/*
1131 			 * handle ppnext for outside loop, and saving pp
1132 			 * until the end.
1133 			 */
1134 			if (lcv < npages) {
1135 				if (ppsp[lcv] == pp)
1136 					continue; /* skip pp until the end */
1137 				ptmp = ppsp[lcv];
1138 			} else {
1139 				ptmp = pp;
1140 
1141 				/* set up next page for outer loop */
1142 				if (by_list) {
1143 					if (pp->version == pp_version)
1144 						ppnext = pp->listq.tqe_next;
1145 					else
1146 						/* reset */
1147 						ppnext = uobj->memq.tqh_first;
1148 				} else {
1149 					if (curoff < stop)
1150 					ppnext = uvm_pagelookup(uobj, curoff);
1151 				}
1152 			}
1153 
1154 			/*
1155 			 * verify the page didn't get moved while obj was
1156 			 * unlocked
1157 			 */
1158 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1159 				continue;
1160 
1161 			/*
1162 			 * unbusy the page if I/O is done.   note that for
1163 			 * pending I/O it is possible that the I/O op
1164 			 * finished before we relocked the object (in
1165 			 * which case the page is no longer busy).
1166 			 */
1167 
1168 			if (result != VM_PAGER_PEND) {
1169 				if (ptmp->flags & PG_WANTED)
1170 					/* still holding object lock */
1171 					thread_wakeup(ptmp);
1172 
1173 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1174 				UVM_PAGE_OWN(ptmp, NULL);
1175 				if (ptmp->flags & PG_RELEASED) {
1176 
1177 					/* pgo_releasepg wants this */
1178 					uvm_unlock_pageq();
1179 					if (!uvn_releasepg(ptmp, NULL))
1180 						return (TRUE);
1181 
1182 					uvm_lock_pageq();	/* relock */
1183 					continue;		/* next page */
1184 
1185 				} else {
1186 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1187 					if ((flags & PGO_FREE) == 0)
1188 						pmap_clear_modify(
1189 						    PMAP_PGARG(ptmp));
1190 				}
1191 			}
1192 
1193 			/*
1194 			 * dispose of page
1195 			 */
1196 
1197 			if (flags & PGO_DEACTIVATE) {
1198 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1199 				    pp->wire_count == 0) {
1200 					pmap_page_protect(PMAP_PGARG(ptmp),
1201 					    VM_PROT_NONE);
1202 					uvm_pagedeactivate(ptmp);
1203 				}
1204 
1205 			} else if (flags & PGO_FREE) {
1206 				if (result == VM_PAGER_PEND) {
1207 					if ((ptmp->flags & PG_BUSY) != 0)
1208 						/* signal for i/o done */
1209 						ptmp->flags |= PG_RELEASED;
1210 				} else {
1211 					if (result != VM_PAGER_OK) {
1212 						printf("uvn_flush: obj=%p, "
1213 						   "offset=0x%lx.  error "
1214 						   "during pageout.\n",
1215 						    pp->uobject, pp->offset);
1216 						printf("uvn_flush: WARNING: "
1217 						    "changes to page may be "
1218 						    "lost!\n");
1219 						retval = FALSE;
1220 					}
1221 					pmap_page_protect(PMAP_PGARG(ptmp),
1222 					    VM_PROT_NONE);
1223 					uvm_pagefree(ptmp);
1224 				}
1225 			}
1226 
1227 		}		/* end of "lcv" for loop */
1228 
1229 	}		/* end of "pp" for loop */
1230 
1231 	/*
1232 	 * done with pagequeues: unlock
1233 	 */
1234 	uvm_unlock_pageq();
1235 
1236 	/*
1237 	 * now wait for all I/O if required.
1238 	 */
1239 	if (need_iosync) {
1240 
1241 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1242 		while (uvn->u_nio != 0) {
1243 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1244 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1245 			  FALSE, "uvn_flush",0);
1246 			simple_lock(&uvn->u_obj.vmobjlock);
1247 		}
1248 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1249 			wakeup(&uvn->u_flags);
1250 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1251 	}
1252 
1253 	/* return, with object locked! */
1254 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1255 	return(retval);
1256 }
1257 
1258 /*
1259  * uvn_cluster
1260  *
1261  * we are about to do I/O in an object at offset.   this function is called
1262  * to establish a range of offsets around "offset" in which we can cluster
1263  * I/O.
1264  *
1265  * - currently doesn't matter if obj locked or not.
1266  */
1267 
1268 static void
1269 uvn_cluster(uobj, offset, loffset, hoffset)
1270 	struct uvm_object *uobj;
1271 	vaddr_t offset;
1272 	vaddr_t *loffset, *hoffset; /* OUT */
1273 {
1274 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1275 	*loffset = offset;
1276 
1277 	if (*loffset >= uvn->u_size)
1278 		panic("uvn_cluster: offset out of range");
1279 
1280 	/*
1281 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1282 	 */
1283 	*hoffset = *loffset + MAXBSIZE;
1284 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1285 		*hoffset = round_page(uvn->u_size);
1286 
1287 	return;
1288 }
1289 
1290 /*
1291  * uvn_put: flush page data to backing store.
1292  *
1293  * => prefer map unlocked (not required)
1294  * => object must be locked!   we will _unlock_ it before starting I/O.
1295  * => flags: PGO_SYNCIO -- use sync. I/O
1296  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1297  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1298  *	[thus we never do async i/o!  see iodone comment]
1299  */
1300 
1301 static int
1302 uvn_put(uobj, pps, npages, flags)
1303 	struct uvm_object *uobj;
1304 	struct vm_page **pps;
1305 	int npages, flags;
1306 {
1307 	int retval;
1308 
1309 	/* note: object locked */
1310 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1311 	/* note: object unlocked */
1312 
1313 	return(retval);
1314 }
1315 
1316 
1317 /*
1318  * uvn_get: get pages (synchronously) from backing store
1319  *
1320  * => prefer map unlocked (not required)
1321  * => object must be locked!  we will _unlock_ it before starting any I/O.
1322  * => flags: PGO_ALLPAGES: get all of the pages
1323  *           PGO_LOCKED: fault data structures are locked
1324  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1325  * => NOTE: caller must check for released pages!!
1326  */
1327 
1328 static int
1329 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1330 	struct uvm_object *uobj;
1331 	vaddr_t offset;
1332 	struct vm_page **pps;		/* IN/OUT */
1333 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1334 	int centeridx, advice, flags;
1335 	vm_prot_t access_type;
1336 {
1337 	vaddr_t current_offset;
1338 	struct vm_page *ptmp;
1339 	int lcv, result, gotpages;
1340 	boolean_t done;
1341 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1342 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1343 
1344 	/*
1345 	 * step 1: handled the case where fault data structures are locked.
1346 	 */
1347 
1348 	if (flags & PGO_LOCKED) {
1349 
1350 		/*
1351 		 * gotpages is the current number of pages we've gotten (which
1352 		 * we pass back up to caller via *npagesp.
1353 		 */
1354 
1355 		gotpages = 0;
1356 
1357 		/*
1358 		 * step 1a: get pages that are already resident.   only do this
1359 		 * if the data structures are locked (i.e. the first time
1360 		 * through).
1361 		 */
1362 
1363 		done = TRUE;	/* be optimistic */
1364 
1365 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1366 		    lcv++, current_offset += PAGE_SIZE) {
1367 
1368 			/* do we care about this page?  if not, skip it */
1369 			if (pps[lcv] == PGO_DONTCARE)
1370 				continue;
1371 
1372 			/* lookup page */
1373 			ptmp = uvm_pagelookup(uobj, current_offset);
1374 
1375 			/* to be useful must get a non-busy, non-released pg */
1376 			if (ptmp == NULL ||
1377 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1378 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1379 				    != 0)
1380 				done = FALSE;	/* need to do a wait or I/O! */
1381 				continue;
1382 			}
1383 
1384 			/*
1385 			 * useful page: busy/lock it and plug it in our
1386 			 * result array
1387 			 */
1388 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1389 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1390 			pps[lcv] = ptmp;
1391 			gotpages++;
1392 
1393 		}	/* "for" lcv loop */
1394 
1395 		/*
1396 		 * XXX: given the "advice", should we consider async read-ahead?
1397 		 * XXX: fault current does deactive of pages behind us.  is
1398 		 * this good (other callers might now).
1399 		 */
1400 		/*
1401 		 * XXX: read-ahead currently handled by buffer cache (bread)
1402 		 * level.
1403 		 * XXX: no async i/o available.
1404 		 * XXX: so we don't do anything now.
1405 		 */
1406 
1407 		/*
1408 		 * step 1c: now we've either done everything needed or we to
1409 		 * unlock and do some waiting or I/O.
1410 		 */
1411 
1412 		*npagesp = gotpages;		/* let caller know */
1413 		if (done)
1414 			return(VM_PAGER_OK);		/* bingo! */
1415 		else
1416 			/* EEK!   Need to unlock and I/O */
1417 			return(VM_PAGER_UNLOCK);
1418 	}
1419 
1420 	/*
1421 	 * step 2: get non-resident or busy pages.
1422 	 * object is locked.   data structures are unlocked.
1423 	 *
1424 	 * XXX: because we can't do async I/O at this level we get things
1425 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1426 	 * async-read-ahead for us at a lower level.
1427 	 */
1428 
1429 	for (lcv = 0, current_offset = offset ;
1430 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1431 
1432 		/* skip over pages we've already gotten or don't want */
1433 		/* skip over pages we don't _have_ to get */
1434 		if (pps[lcv] != NULL || (lcv != centeridx &&
1435 		    (flags & PGO_ALLPAGES) == 0))
1436 			continue;
1437 
1438 		/*
1439 		 * we have yet to locate the current page (pps[lcv]).   we first
1440 		 * look for a page that is already at the current offset.   if
1441 		 * we fine a page, we check to see if it is busy or released.
1442 		 * if that is the case, then we sleep on the page until it is
1443 		 * no longer busy or released and repeat the lookup.    if the
1444 		 * page we found is neither busy nor released, then we busy it
1445 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1446 		 * following while loop and indicates we are ready to move on
1447 		 * to the next page in the "lcv" loop above.
1448 		 *
1449 		 * if we exit the while loop with pps[lcv] still set to NULL,
1450 		 * then it means that we allocated a new busy/fake/clean page
1451 		 * ptmp in the object and we need to do I/O to fill in the data.
1452 		 */
1453 
1454 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1455 
1456 			/* look for a current page */
1457 			ptmp = uvm_pagelookup(uobj, current_offset);
1458 
1459 			/* nope?   allocate one now (if we can) */
1460 			if (ptmp == NULL) {
1461 
1462 				ptmp = uvm_pagealloc(uobj, current_offset,
1463 				    NULL);	/* alloc */
1464 
1465 				/* out of RAM? */
1466 				if (ptmp == NULL) {
1467 					simple_unlock(&uobj->vmobjlock);
1468 					uvm_wait("uvn_getpage");
1469 					simple_lock(&uobj->vmobjlock);
1470 
1471 					/* goto top of pps while loop */
1472 					continue;
1473 				}
1474 
1475 				/*
1476 				 * got new page ready for I/O.  break pps
1477 				 * while loop.  pps[lcv] is still NULL.
1478 				 */
1479 				break;
1480 			}
1481 
1482 			/* page is there, see if we need to wait on it */
1483 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1484 				ptmp->flags |= PG_WANTED;
1485 				UVM_UNLOCK_AND_WAIT(ptmp,
1486 				    &uobj->vmobjlock, 0, "uvn_get",0);
1487 				simple_lock(&uobj->vmobjlock);
1488 				continue;	/* goto top of pps while loop */
1489 			}
1490 
1491 			/*
1492 			 * if we get here then the page has become resident
1493 			 * and unbusy between steps 1 and 2.  we busy it
1494 			 * now (so we own it) and set pps[lcv] (so that we
1495 			 * exit the while loop).
1496 			 */
1497 			ptmp->flags |= PG_BUSY;
1498 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1499 			pps[lcv] = ptmp;
1500 		}
1501 
1502 		/*
1503 		 * if we own the a valid page at the correct offset, pps[lcv]
1504 		 * will point to it.   nothing more to do except go to the
1505 		 * next page.
1506 		 */
1507 
1508 		if (pps[lcv])
1509 			continue;			/* next lcv */
1510 
1511 		/*
1512 		 * we have a "fake/busy/clean" page that we just allocated.  do
1513 		 * I/O to fill it with valid data.  note that object must be
1514 		 * locked going into uvn_io, but will be unlocked afterwards.
1515 		 */
1516 
1517 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1518 		    PGO_SYNCIO, UIO_READ);
1519 
1520 		/*
1521 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1522 		 * syncio the result can not be PEND or AGAIN.   we must relock
1523 		 * and check for errors.
1524 		 */
1525 
1526 		/* lock object.   check for errors.   */
1527 		simple_lock(&uobj->vmobjlock);
1528 		if (result != VM_PAGER_OK) {
1529 			if (ptmp->flags & PG_WANTED)
1530 				/* object lock still held */
1531 				thread_wakeup(ptmp);
1532 
1533 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1534 			UVM_PAGE_OWN(ptmp, NULL);
1535 			uvm_lock_pageq();
1536 			uvm_pagefree(ptmp);
1537 			uvm_unlock_pageq();
1538 			simple_unlock(&uobj->vmobjlock);
1539 			return(result);
1540 		}
1541 
1542 		/*
1543 		 * we got the page!   clear the fake flag (indicates valid
1544 		 * data now in page) and plug into our result array.   note
1545 		 * that page is still busy.
1546 		 *
1547 		 * it is the callers job to:
1548 		 * => check if the page is released
1549 		 * => unbusy the page
1550 		 * => activate the page
1551 		 */
1552 
1553 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1554 		pmap_clear_modify(PMAP_PGARG(ptmp));	/* ... and clean */
1555 		pps[lcv] = ptmp;
1556 
1557 	}	/* lcv loop */
1558 
1559 	/*
1560 	 * finally, unlock object and return.
1561 	 */
1562 
1563 	simple_unlock(&uobj->vmobjlock);
1564 	return (VM_PAGER_OK);
1565 }
1566 
1567 /*
1568  * uvn_asyncget: start async I/O to bring pages into ram
1569  *
1570  * => caller must lock object(???XXX: see if this is best)
1571  * => could be called from uvn_get or a madvise() fault-ahead.
1572  * => if it fails, it doesn't matter.
1573  */
1574 
1575 static int
1576 uvn_asyncget(uobj, offset, npages)
1577 	struct uvm_object *uobj;
1578 	vaddr_t offset;
1579 	int npages;
1580 {
1581 
1582 	/*
1583 	 * XXXCDC: we can't do async I/O yet
1584 	 */
1585 	printf("uvn_asyncget called\n");
1586 	return (KERN_SUCCESS);
1587 }
1588 
1589 /*
1590  * uvn_io: do I/O to a vnode
1591  *
1592  * => prefer map unlocked (not required)
1593  * => object must be locked!   we will _unlock_ it before starting I/O.
1594  * => flags: PGO_SYNCIO -- use sync. I/O
1595  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1596  *	[thus we never do async i/o!  see iodone comment]
1597  */
1598 
1599 static int
1600 uvn_io(uvn, pps, npages, flags, rw)
1601 	struct uvm_vnode *uvn;
1602 	vm_page_t *pps;
1603 	int npages, flags, rw;
1604 {
1605 	struct vnode *vn;
1606 	struct uio uio;
1607 	struct iovec iov;
1608 	vaddr_t kva, file_offset;
1609 	int waitf, result, got, wanted;
1610 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1611 
1612 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1613 
1614 	/*
1615 	 * init values
1616 	 */
1617 
1618 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1619 	vn = (struct vnode *) uvn;
1620 	file_offset = pps[0]->offset;
1621 
1622 	/*
1623 	 * check for sync'ing I/O.
1624 	 */
1625 
1626 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1627 		if (waitf == M_NOWAIT) {
1628 			simple_unlock(&uvn->u_obj.vmobjlock);
1629 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1630 			return(VM_PAGER_AGAIN);
1631 		}
1632 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1633 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1634 			FALSE, "uvn_iosync",0);
1635 		simple_lock(&uvn->u_obj.vmobjlock);
1636 	}
1637 
1638 	/*
1639 	 * check size
1640 	 */
1641 
1642 	if (file_offset >= uvn->u_size) {
1643 			simple_unlock(&uvn->u_obj.vmobjlock);
1644 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1645 #ifdef DIAGNOSTIC
1646 			printf("uvn_io: note: size check fired\n");
1647 #endif
1648 			return(VM_PAGER_BAD);
1649 	}
1650 
1651 	/*
1652 	 * first try and map the pages in (without waiting)
1653 	 */
1654 
1655 	kva = uvm_pagermapin(pps, npages, NULL, M_NOWAIT);
1656 	if (kva == NULL && waitf == M_NOWAIT) {
1657 		simple_unlock(&uvn->u_obj.vmobjlock);
1658 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1659 		return(VM_PAGER_AGAIN);
1660 	}
1661 
1662 	/*
1663 	 * ok, now bump u_nio up.   at this point we are done with uvn
1664 	 * and can unlock it.   if we still don't have a kva, try again
1665 	 * (this time with sleep ok).
1666 	 */
1667 
1668 	uvn->u_nio++;			/* we have an I/O in progress! */
1669 	simple_unlock(&uvn->u_obj.vmobjlock);
1670 	/* NOTE: object now unlocked */
1671 	if (kva == NULL) {
1672 		kva = uvm_pagermapin(pps, npages, NULL, M_WAITOK);
1673 	}
1674 
1675 	/*
1676 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1677 	 * get touched (so we can look at "offset" without having to lock
1678 	 * the object).  set up for I/O.
1679 	 */
1680 
1681 	/*
1682 	 * fill out uio/iov
1683 	 */
1684 
1685 	iov.iov_base = (caddr_t) kva;
1686 	wanted = npages << PAGE_SHIFT;
1687 	if (file_offset + wanted > uvn->u_size)
1688 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1689 	iov.iov_len = wanted;
1690 	uio.uio_iov = &iov;
1691 	uio.uio_iovcnt = 1;
1692 	uio.uio_offset = file_offset;
1693 	uio.uio_segflg = UIO_SYSSPACE;
1694 	uio.uio_rw = rw;
1695 	uio.uio_resid = wanted;
1696 	uio.uio_procp = NULL;
1697 
1698 	/*
1699 	 * do the I/O!  (XXX: curproc?)
1700 	 */
1701 
1702 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1703 
1704 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1705 		vn_lock(vn, LK_EXCLUSIVE | LK_RETRY);
1706 	/* NOTE: vnode now locked! */
1707 
1708 	if (rw == UIO_READ)
1709 		result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1710 	else
1711 		result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1712 
1713 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1714 		VOP_UNLOCK(vn, 0);
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 #ifdef NFS
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	/* NFS */
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);
1877 	uvn_detach(&uvn->u_obj);
1878 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
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 	u_quad_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 		 * make sure that the newsize fits within a vaddr_t
1920 		 * XXX: need to revise addressing data types
1921 		 */
1922 
1923 		if (newsize > (vaddr_t) -PAGE_SIZE) {
1924 #ifdef DEBUG
1925 			printf("uvm_vnp_setsize: vn %p size truncated "
1926 			       "%qx->%lx\n", vp, (long long)newsize,
1927 			       (vaddr_t)-PAGE_SIZE);
1928 #endif
1929 			newsize = (vaddr_t)-PAGE_SIZE;
1930 		}
1931 
1932 		/*
1933 		 * now check if the size has changed: if we shrink we had better
1934 		 * toss some pages...
1935 		 */
1936 
1937 		if (uvn->u_size > newsize) {
1938 			(void)uvn_flush(&uvn->u_obj, (vaddr_t) newsize,
1939 			    uvn->u_size, PGO_FREE);
1940 		}
1941 		uvn->u_size = (vaddr_t)newsize;
1942 	}
1943 	simple_unlock(&uvn->u_obj.vmobjlock);
1944 
1945 	/*
1946 	 * done
1947 	 */
1948 	return;
1949 }
1950 
1951 /*
1952  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1953  *
1954  * => called from sys_sync with no VM structures locked
1955  * => only one process can do a sync at a time (because the uvn
1956  *    structure only has one queue for sync'ing).  we ensure this
1957  *    by holding the uvn_sync_lock while the sync is in progress.
1958  *    other processes attempting a sync will sleep on this lock
1959  *    until we are done.
1960  */
1961 
1962 void
1963 uvm_vnp_sync(mp)
1964 	struct mount *mp;
1965 {
1966 	struct uvm_vnode *uvn;
1967 	struct vnode *vp;
1968 	boolean_t got_lock;
1969 
1970 	/*
1971 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1972 	 * our lock...
1973 	 */
1974 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0);
1975 
1976 	/*
1977 	 * step 2: build up a simpleq of uvns of interest based on the
1978 	 * write list.   we gain a reference to uvns of interest.  must
1979 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1980 	 * in the body of the loop.
1981 	 */
1982 	SIMPLEQ_INIT(&uvn_sync_q);
1983 	simple_lock(&uvn_wl_lock);
1984 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1985 	    uvn = uvn->u_wlist.le_next) {
1986 
1987 		vp = (struct vnode *) uvn;
1988 		if (mp && vp->v_mount != mp)
1989 			continue;
1990 
1991 		/* attempt to gain reference */
1992 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1993 		    						FALSE &&
1994 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1995 			/* spin */ ;
1996 
1997 		/*
1998 		 * we will exit the loop if either if the following are true:
1999 		 *  - we got the lock [always true if NCPU == 1]
2000 		 *  - we failed to get the lock but noticed the vnode was
2001 		 * 	"blocked" -- in this case the vnode must be a dying
2002 		 *	vnode, and since dying vnodes are in the process of
2003 		 *	being flushed out, we can safely skip this one
2004 		 *
2005 		 * we want to skip over the vnode if we did not get the lock,
2006 		 * or if the vnode is already dying (due to the above logic).
2007 		 *
2008 		 * note that uvn must already be valid because we found it on
2009 		 * the wlist (this also means it can't be ALOCK'd).
2010 		 */
2011 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
2012 			if (got_lock)
2013 				simple_unlock(&uvn->u_obj.vmobjlock);
2014 			continue;		/* skip it */
2015 		}
2016 
2017 		/*
2018 		 * gain reference.   watch out for persisting uvns (need to
2019 		 * regain vnode REF).
2020 		 */
2021 		if (uvn->u_obj.uo_refs == 0)
2022 			VREF(vp);
2023 		uvn->u_obj.uo_refs++;
2024 		simple_unlock(&uvn->u_obj.vmobjlock);
2025 
2026 		/*
2027 		 * got it!
2028 		 */
2029 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
2030 	}
2031 	simple_unlock(&uvn_wl_lock);
2032 
2033 	/*
2034 	 * step 3: we now have a list of uvn's that may need cleaning.
2035 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
2036 	 * (so we can now safely lock uvn's again).
2037 	 */
2038 
2039 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2040 		simple_lock(&uvn->u_obj.vmobjlock);
2041 #ifdef DIAGNOSTIC
2042 		if (uvn->u_flags & UVM_VNODE_DYING) {
2043 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2044 		}
2045 #endif
2046 		uvn_flush(&uvn->u_obj, 0, 0,
2047 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2048 
2049 		/*
2050 		 * if we have the only reference and we just cleaned the uvn,
2051 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2052 		 * thus allowing us to avoid thinking about flushing it again
2053 		 * on later sync ops.
2054 		 */
2055 		if (uvn->u_obj.uo_refs == 1 &&
2056 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2057 			LIST_REMOVE(uvn, u_wlist);
2058 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2059 		}
2060 
2061 		simple_unlock(&uvn->u_obj.vmobjlock);
2062 
2063 		/* now drop our reference to the uvn */
2064 		uvn_detach(&uvn->u_obj);
2065 	}
2066 
2067 	/*
2068 	 * done!  release sync lock
2069 	 */
2070 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0);
2071 }
2072