xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision 7c7c171d130af9949261bc7dce2150a03c3d239c)
1 /*	$NetBSD: uvm_vnode.c,v 1.9 1998/03/11 01:37:40 chuck 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/mount.h>
62 #include <sys/proc.h>
63 #include <sys/malloc.h>
64 #include <sys/vnode.h>
65 
66 #include <vm/vm.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_kern.h>
69 
70 #include <sys/syscallargs.h>
71 
72 #include <uvm/uvm.h>
73 #include <uvm/uvm_vnode.h>
74 
75 /*
76  * private global data structure
77  *
78  * we keep a list of writeable active vnode-backed VM objects for sync op.
79  * we keep a simpleq of vnodes that are currently being sync'd.
80  */
81 
82 LIST_HEAD(uvn_list_struct, uvm_vnode);
83 static struct uvn_list_struct uvn_wlist;	/* writeable uvns */
84 static simple_lock_data_t uvn_wl_lock;		/* locks uvn_wlist */
85 
86 SIMPLEQ_HEAD(uvn_sq_struct, uvm_vnode);
87 static struct uvn_sq_struct uvn_sync_q;		/* sync'ing uvns */
88 lock_data_t uvn_sync_lock;			/* locks sync operation */
89 
90 /*
91  * functions
92  */
93 
94 static int		   uvn_asyncget __P((struct uvm_object *, vm_offset_t,
95 					    int));
96 struct uvm_object 	  *uvn_attach __P((void *, vm_prot_t));
97 static void		   uvn_cluster __P((struct uvm_object *, vm_offset_t,
98 					   vm_offset_t *, vm_offset_t *));
99 static void                uvn_detach __P((struct uvm_object *));
100 static boolean_t           uvn_flush __P((struct uvm_object *, vm_offset_t,
101 					 vm_offset_t, int));
102 static int                 uvn_get __P((struct uvm_object *, vm_offset_t,
103 					vm_page_t *, int *, int,
104 					vm_prot_t, int, int));
105 static void		   uvn_init __P((void));
106 static int		   uvn_io __P((struct uvm_vnode *, vm_page_t *,
107 				      int, int, int));
108 static int		   uvn_put __P((struct uvm_object *, vm_page_t *,
109 					int, boolean_t));
110 static void                uvn_reference __P((struct uvm_object *));
111 static boolean_t	   uvn_releasepg __P((struct vm_page *,
112 					      struct vm_page **));
113 
114 /*
115  * master pager structure
116  */
117 
118 struct uvm_pagerops uvm_vnodeops = {
119 	uvn_init,
120 	uvn_attach,
121 	uvn_reference,
122 	uvn_detach,
123 	NULL,			/* no specialized fault routine required */
124 	uvn_flush,
125 	uvn_get,
126 	uvn_asyncget,
127 	uvn_put,
128 	uvn_cluster,
129 	uvm_mk_pcluster, /* use generic version of this: see uvm_pager.c */
130 	uvm_shareprot,	 /* !NULL: allow us in share maps */
131 	NULL,		 /* AIO-DONE function (not until we have asyncio) */
132 	uvn_releasepg,
133 };
134 
135 /*
136  * the ops!
137  */
138 
139 /*
140  * uvn_init
141  *
142  * init pager private data structures.
143  */
144 
145 static void
146 uvn_init()
147 {
148 
149 	LIST_INIT(&uvn_wlist);
150 	simple_lock_init(&uvn_wl_lock);
151 	/* note: uvn_sync_q init'd in uvm_vnp_sync() */
152 	lockinit(&uvn_sync_lock, PVM, "uvnsync", 0, 0);
153 }
154 
155 /*
156  * uvn_attach
157  *
158  * attach a vnode structure to a VM object.  if the vnode is already
159  * attached, then just bump the reference count by one and return the
160  * VM object.   if not already attached, attach and return the new VM obj.
161  * the "accessprot" tells the max access the attaching thread wants to
162  * our pages.
163  *
164  * => caller must _not_ already be holding the lock on the uvm_object.
165  * => in fact, nothing should be locked so that we can sleep here.
166  * => note that uvm_object is first thing in vnode structure, so their
167  *    pointers are equiv.
168  */
169 
170 struct uvm_object *
171 uvn_attach(arg, accessprot)
172 	void *arg;
173 	vm_prot_t accessprot;
174 {
175 	struct vnode *vp = arg;
176 	struct uvm_vnode *uvn = &vp->v_uvm;
177 	struct vattr vattr;
178 	int oldflags, result;
179 	u_quad_t used_vnode_size;
180 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
181 
182 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
183 
184 	/*
185 	 * first get a lock on the uvn.
186 	 */
187 	simple_lock(&uvn->u_obj.vmobjlock);
188 	while (uvn->u_flags & UVM_VNODE_BLOCKED) {
189 		uvn->u_flags |= UVM_VNODE_WANTED;
190 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
191 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
192 		    "uvn_attach", 0);
193 		simple_lock(&uvn->u_obj.vmobjlock);
194 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
195 	}
196 
197 	/*
198 	 * now we have lock and uvn must not be in a blocked state.
199 	 * first check to see if it is already active, in which case
200 	 * we can bump the reference count, check to see if we need to
201 	 * add it to the writeable list, and then return.
202 	 */
203 	if (uvn->u_flags & UVM_VNODE_VALID) {	/* already active? */
204 
205 		/* regain VREF if we were persisting */
206 		if (uvn->u_obj.uo_refs == 0) {
207 			VREF(vp);
208 			UVMHIST_LOG(maphist," VREF (reclaim persisting vnode)",
209 			    0,0,0,0);
210 		}
211 		uvn->u_obj.uo_refs++;		/* bump uvn ref! */
212 
213 		/* check for new writeable uvn */
214 		if ((accessprot & VM_PROT_WRITE) != 0 &&
215 		    (uvn->u_flags & UVM_VNODE_WRITEABLE) == 0) {
216 			simple_lock(&uvn_wl_lock);
217 			LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
218 			simple_unlock(&uvn_wl_lock);
219 			/* we are now on wlist! */
220 			uvn->u_flags |= UVM_VNODE_WRITEABLE;
221 		}
222 
223 		/* unlock and return */
224 		simple_unlock(&uvn->u_obj.vmobjlock);
225 		UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
226 		    0, 0, 0);
227 		return (&uvn->u_obj);
228 	}
229 
230 	/*
231 	 * need to call VOP_GETATTR() to get the attributes, but that could
232 	 * block (due to I/O), so we want to unlock the object before calling.
233 	 * however, we want to keep anyone else from playing with the object
234 	 * while it is unlocked.   to do this we set UVM_VNODE_ALOCK which
235 	 * prevents anyone from attaching to the vnode until we are done with
236 	 * it.
237 	 */
238 	uvn->u_flags = UVM_VNODE_ALOCK;
239 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
240 		/* XXX: curproc? */
241 	result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
242 
243 	/*
244 	 * make sure that the newsize fits within a vm_offset_t
245 	 * XXX: need to revise addressing data types
246 	 */
247 	used_vnode_size = vattr.va_size;
248 	if (used_vnode_size > (vm_offset_t) -PAGE_SIZE) {
249 #ifdef DEBUG
250 		printf("uvn_attach: vn %p size truncated %qx->%x\n", vp,
251 		    used_vnode_size, -PAGE_SIZE);
252 #endif
253 		used_vnode_size = (vm_offset_t) -PAGE_SIZE;
254 	}
255 
256 	/* relock object */
257 	simple_lock(&uvn->u_obj.vmobjlock);
258 
259 	if (result != 0) {
260 		if (uvn->u_flags & UVM_VNODE_WANTED)
261 			wakeup(uvn);
262 		uvn->u_flags = 0;
263 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
264 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
265 		return(NULL);
266 	}
267 
268 	/*
269 	 * now set up the uvn.
270 	 */
271 	uvn->u_obj.pgops = &uvm_vnodeops;
272 	TAILQ_INIT(&uvn->u_obj.memq);
273 	uvn->u_obj.uo_npages = 0;
274 	uvn->u_obj.uo_refs = 1;			/* just us... */
275 	oldflags = uvn->u_flags;
276 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
277 	uvn->u_nio = 0;
278 	uvn->u_size = used_vnode_size;
279 
280 	/* if write access, we need to add it to the wlist */
281 	if (accessprot & VM_PROT_WRITE) {
282 		simple_lock(&uvn_wl_lock);
283 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
284 		simple_unlock(&uvn_wl_lock);
285 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
286 	}
287 
288 	/*
289 	 * add a reference to the vnode.   this reference will stay as long
290 	 * as there is a valid mapping of the vnode.   dropped when the
291 	 * reference count goes to zero [and we either free or persist].
292 	 */
293 	VREF(vp);
294 	simple_unlock(&uvn->u_obj.vmobjlock);
295 	if (oldflags & UVM_VNODE_WANTED)
296 		wakeup(uvn);
297 
298 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
299 	return(&uvn->u_obj);
300 }
301 
302 
303 /*
304  * uvn_reference
305  *
306  * duplicate a reference to a VM object.  Note that the reference
307  * count must already be at least one (the passed in reference) so
308  * there is no chance of the uvn being killed or locked out here.
309  *
310  * => caller must call with object unlocked.
311  * => caller must be using the same accessprot as was used at attach time
312  */
313 
314 
315 static void
316 uvn_reference(uobj)
317 	struct uvm_object *uobj;
318 {
319 #ifdef DIAGNOSTIC
320 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
321 #endif
322 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
323 
324 	simple_lock(&uobj->vmobjlock);
325 #ifdef DIAGNOSTIC
326 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
327 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
328 		    uobj->uo_refs);
329 		panic("uvn_reference: invalid state");
330 	}
331 #endif
332 	uobj->uo_refs++;
333 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
334 	uobj, uobj->uo_refs,0,0);
335 	simple_unlock(&uobj->vmobjlock);
336 }
337 
338 /*
339  * uvn_detach
340  *
341  * remove a reference to a VM object.
342  *
343  * => caller must call with object unlocked and map locked.
344  * => this starts the detach process, but doesn't have to finish it
345  *    (async i/o could still be pending).
346  */
347 static void
348 uvn_detach(uobj)
349 	struct uvm_object *uobj;
350 {
351 	struct uvm_vnode *uvn;
352 	struct vnode *vp;
353 	int oldflags;
354 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
355 
356 	simple_lock(&uobj->vmobjlock);
357 
358 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
359 	uobj->uo_refs--;			/* drop ref! */
360 	if (uobj->uo_refs) {			/* still more refs */
361 		simple_unlock(&uobj->vmobjlock);
362 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
363 		return;
364 	}
365 
366 	/*
367 	 * get other pointers ...
368 	 */
369 
370 	uvn = (struct uvm_vnode *) uobj;
371 	vp = (struct vnode *) uobj;
372 
373 	/*
374 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
375 	 * to keep an active text file from being overwritten).
376 	 */
377 	vp->v_flag &= ~VTEXT;
378 
379 	/*
380 	 * we just dropped the last reference to the uvn.   see if we can
381 	 * let it "stick around".
382 	 */
383 
384 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
385 		/* won't block */
386 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
387 		vrele(vp);			/* drop vnode reference */
388 		simple_unlock(&uobj->vmobjlock);
389 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
390 		return;
391 	}
392 
393 	/*
394 	 * its a goner!
395 	 */
396 
397 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
398 
399 	uvn->u_flags |= UVM_VNODE_DYING;
400 
401 	/*
402 	 * even though we may unlock in flush, no one can gain a reference
403 	 * to us until we clear the "dying" flag [because it blocks
404 	 * attaches].  we will not do that until after we've disposed of all
405 	 * the pages with uvn_flush().  note that before the flush the only
406 	 * pages that could be marked PG_BUSY are ones that are in async
407 	 * pageout by the daemon.  (there can't be any pending "get"'s
408 	 * because there are no references to the object).
409 	 */
410 
411 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
412 
413 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
414 
415 	/*
416 	 * given the structure of this pager, the above flush request will
417 	 * create the following state: all the pages that were in the object
418 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
419 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
420 	 * in either case, if we have pages left we can't kill the object
421 	 * yet because i/o is pending.  in this case we set the "relkill"
422 	 * flag which will cause pgo_releasepg to kill the object once all
423 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
424 	 * routine or from the page daemon].
425 	 */
426 
427 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
428 #ifdef DIAGNOSTIC
429 		/*
430 		 * XXXCDC: very unlikely to happen until we have async i/o
431 		 * so print a little info message in case it does.
432 		 */
433 		printf("uvn_detach: vn %p has pages left after flush - "
434 		    "relkill mode\n", uobj);
435 #endif
436 		uvn->u_flags |= UVM_VNODE_RELKILL;
437 		simple_unlock(&uobj->vmobjlock);
438 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
439 		    0, 0);
440 		return;
441 	}
442 
443 	/*
444 	 * kill object now.   note that we can't be on the sync q because
445 	 * all references are gone.
446 	 */
447 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
448 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
449 		LIST_REMOVE(uvn, u_wlist);
450 		simple_unlock(&uvn_wl_lock);
451 	}
452 #ifdef DIAGNOSTIC
453 	if (uobj->memq.tqh_first != NULL)
454 		panic("uvn_deref: vnode VM object still has pages afer "
455 		    "syncio/free flush");
456 #endif
457 	oldflags = uvn->u_flags;
458 	uvn->u_flags = 0;
459 	simple_unlock(&uobj->vmobjlock);
460 
461 	/* wake up any sleepers */
462 	if (oldflags & UVM_VNODE_WANTED)
463 		wakeup(uvn);
464 
465 	/*
466 	 * drop our reference to the vnode.
467 	 */
468 	vrele(vp);
469 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
470 
471 	return;
472 }
473 
474 /*
475  * uvm_vnp_terminate: external hook to clear out a vnode's VM
476  *
477  * called in two cases:
478  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
479  *      count) needs to be freed so that a vnode can be reused.  this
480  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
481  *      the free list is still attached (i.e. not VBAD) then vgone is
482  *	called.   as part of the vgone trace this should get called to
483  *	free the vm object.   this is the common case.
484  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
485  *	"umount -f") the vgone() function is called on active vnodes
486  *	on the mounted file systems to kill their data (the vnodes become
487  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
488  *	call here (even if the uvn is still in use -- i.e. has a non-zero
489  *	reference count).  this case happens at "umount -f" and during a
490  *	"reboot/halt" operation.
491  *
492  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
493  *	[protects us from getting a vnode that is already in the DYING
494  *	 state...]
495  * => unlike uvn_detach, this function must not return until all the
496  *	uvn's pages are disposed of.
497  * => in case [2] the uvn is still alive after this call, but all I/O
498  *	ops will fail (due to the backing vnode now being "dead").  this
499  *	will prob. kill any process using the uvn due to pgo_get failing.
500  */
501 
502 void
503 uvm_vnp_terminate(vp)
504 	struct vnode *vp;
505 {
506 	struct uvm_vnode *uvn = &vp->v_uvm;
507 	int oldflags;
508 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
509 
510 	/*
511 	 * lock object and check if it is valid
512 	 */
513 	simple_lock(&uvn->u_obj.vmobjlock);
514 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
515 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
516 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
517 		simple_unlock(&uvn->u_obj.vmobjlock);
518 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
519 		return;
520 	}
521 
522 	/*
523 	 * must be a valid uvn that is not already dying (because XLOCK
524 	 * protects us from that).   the uvn can't in the the ALOCK state
525 	 * because it is valid, and uvn's that are in the ALOCK state haven't
526 	 * been marked valid yet.
527 	 */
528 
529 #ifdef DEBUG
530 	/*
531 	 * debug check: are we yanking the vnode out from under our uvn?
532 	 */
533 	if (uvn->u_obj.uo_refs) {
534 		printf("uvm_vnp_terminate(%p): terminating active vnode "
535 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
536 	}
537 #endif
538 
539 	/*
540 	 * it is possible that the uvn was detached and is in the relkill
541 	 * state [i.e. waiting for async i/o to finish so that releasepg can
542 	 * kill object].  we take over the vnode now and cancel the relkill.
543 	 * we want to know when the i/o is done so we can recycle right
544 	 * away.   note that a uvn can only be in the RELKILL state if it
545 	 * has a zero reference count.
546 	 */
547 
548 	if (uvn->u_flags & UVM_VNODE_RELKILL)
549 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
550 
551 	/*
552 	 * block the uvn by setting the dying flag, and then flush the
553 	 * pages.  (note that flush may unlock object while doing I/O, but
554 	 * it will re-lock it before it returns control here).
555 	 *
556 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
557 	 * that uvn_io doesn't attempt to VOP_LOCK again.
558 	 *
559 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
560 	 *	due to a forceful unmount might not be a good idea.  maybe we
561 	 *	need a way to pass in this info to uvn_flush through a
562 	 *	pager-defined PGO_ constant [currently there are none].
563 	 */
564 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
565 
566 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
567 
568 	/*
569 	 * as we just did a flush we expect all the pages to be gone or in
570 	 * the process of going.  sleep to wait for the rest to go [via iosync].
571 	 */
572 
573 	while (uvn->u_obj.uo_npages) {
574 #ifdef DIAGNOSTIC
575 		struct vm_page *pp;
576 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
577 		     pp = pp->listq.tqe_next) {
578 			if ((pp->flags & PG_BUSY) == 0)
579 				panic("uvm_vnp_terminate: detected unbusy pg");
580 		}
581 		if (uvn->u_nio == 0)
582 			panic("uvm_vnp_terminate: no I/O to wait for?");
583 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
584 		/*
585 		 * XXXCDC: this is unlikely to happen without async i/o so we
586 		 * put a printf in just to keep an eye on it.
587 		 */
588 #endif
589 		uvn->u_flags |= UVM_VNODE_IOSYNC;
590 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
591 		    "uvn_term",0);
592 		simple_lock(&uvn->u_obj.vmobjlock);
593 	}
594 
595 	/*
596 	 * done.   now we free the uvn if its reference count is zero
597 	 * (true if we are zapping a persisting uvn).   however, if we are
598 	 * terminating a uvn with active mappings we let it live ... future
599 	 * calls down to the vnode layer will fail.
600 	 */
601 
602 	oldflags = uvn->u_flags;
603 	if (uvn->u_obj.uo_refs) {
604 
605 		/*
606 		 * uvn must live on it is dead-vnode state until all references
607 		 * are gone.   restore flags.    clear CANPERSIST state.
608 		 */
609 
610 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
611 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
612 
613 	} else {
614 
615 		/*
616 		 * free the uvn now.   note that the VREF reference is already
617 		 * gone [it is dropped when we enter the persist state].
618 		 */
619 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
620 			panic("uvm_vnp_terminate: io sync wanted bit set");
621 
622 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
623 			simple_lock(&uvn_wl_lock);
624 			LIST_REMOVE(uvn, u_wlist);
625 			simple_unlock(&uvn_wl_lock);
626 		}
627 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
628 	}
629 
630 	if (oldflags & UVM_VNODE_WANTED)
631 		wakeup(uvn);		/* object lock still held */
632 
633 	simple_unlock(&uvn->u_obj.vmobjlock);
634 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
635 
636 }
637 
638 /*
639  * uvn_releasepg: handled a released page in a uvn
640  *
641  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
642  *	to dispose of.
643  * => caller must handled PG_WANTED case
644  * => called with page's object locked, pageq's unlocked
645  * => returns TRUE if page's object is still alive, FALSE if we
646  *	killed the page's object.    if we return TRUE, then we
647  *	return with the object locked.
648  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
649  *				with the page queues locked [for pagedaemon]
650  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
651  * => we kill the uvn if it is not referenced and we are suppose to
652  *	kill it ("relkill").
653  */
654 
655 boolean_t
656 uvn_releasepg(pg, nextpgp)
657 	struct vm_page *pg;
658 	struct vm_page **nextpgp;	/* OUT */
659 {
660 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
661 #ifdef DIAGNOSTIC
662 	if ((pg->flags & PG_RELEASED) == 0)
663 		panic("uvn_releasepg: page not released!");
664 #endif
665 
666 	/*
667 	 * dispose of the page [caller handles PG_WANTED]
668 	 */
669 	pmap_page_protect(PMAP_PGARG(pg), VM_PROT_NONE);
670 	uvm_lock_pageq();
671 	if (nextpgp)
672 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
673 	uvm_pagefree(pg);
674 	if (!nextpgp)
675 		uvm_unlock_pageq();
676 
677 	/*
678 	 * now see if we need to kill the object
679 	 */
680 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
681 		if (uvn->u_obj.uo_refs)
682 			panic("uvn_releasepg: kill flag set on referenced "
683 			    "object!");
684 		if (uvn->u_obj.uo_npages == 0) {
685 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
686 				simple_lock(&uvn_wl_lock);
687 				LIST_REMOVE(uvn, u_wlist);
688 				simple_unlock(&uvn_wl_lock);
689 			}
690 #ifdef DIAGNOSTIC
691 			if (uvn->u_obj.memq.tqh_first)
692 	panic("uvn_releasepg: pages in object with npages == 0");
693 #endif
694 			if (uvn->u_flags & UVM_VNODE_WANTED)
695 				/* still holding object lock */
696 				wakeup(uvn);
697 
698 			uvn->u_flags = 0;		/* DEAD! */
699 			simple_unlock(&uvn->u_obj.vmobjlock);
700 			return (FALSE);
701 		}
702 	}
703 	return (TRUE);
704 }
705 
706 /*
707  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
708  * through the buffer cache and allow I/O in any size.  These VOPs use
709  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
710  * go through the buffer cache or allow I/O sizes larger than a
711  * block].  we will eventually want to change this.
712  *
713  * issues to consider:
714  *   uvm provides the uvm_aiodesc structure for async i/o management.
715  * there are two tailq's in the uvm. structure... one for pending async
716  * i/o and one for "done" async i/o.   to do an async i/o one puts
717  * an aiodesc on the "pending" list (protected by splbio()), starts the
718  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
719  * some sort of "i/o done" function to be called (at splbio(), interrupt
720  * time).   this function should remove the aiodesc from the pending list
721  * and place it on the "done" list and wakeup the daemon.   the daemon
722  * will run at normal spl() and will remove all items from the "done"
723  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
724  * [in the old vm code, this was done by calling the "put" routine with
725  * null arguments which made the code harder to read and understand because
726  * you had one function ("put") doing two things.]
727  *
728  * so the current pager needs:
729  *   int uvn_aiodone(struct uvm_aiodesc *)
730  *
731  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
732  *	later collection.
733  * => called with pageq's locked by the daemon.
734  *
735  * general outline:
736  * - "try" to lock object.   if fail, just return (will try again later)
737  * - drop "u_nio" (this req is done!)
738  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
739  * - get "page" structures (atop?).
740  * - handle "wanted" pages
741  * - handle "released" pages [using pgo_releasepg]
742  *   >>> pgo_releasepg may kill the object
743  * dont forget to look at "object" wanted flag in all cases.
744  */
745 
746 
747 /*
748  * uvn_flush: flush pages out of a uvm object.
749  *
750  * => object should be locked by caller.   we may _unlock_ the object
751  *	if (and only if) we need to clean a page (PGO_CLEANIT).
752  *	we return with the object locked.
753  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
754  *	might want to unlock higher level resources (e.g. vm_map)
755  *	before calling flush.
756  * => if PGO_CLEANIT is not set, then we will neither unlock the object
757  *	or block.
758  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
759  *	for flushing.
760  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
761  *	that new pages are inserted on the tail end of the list.   thus,
762  *	we can make a complete pass through the object in one go by starting
763  *	at the head and working towards the tail (new pages are put in
764  *	front of us).
765  * => NOTE: we are allowed to lock the page queues, so the caller
766  *	must not be holding the lock on them [e.g. pagedaemon had
767  *	better not call us with the queues locked]
768  * => we return TRUE unless we encountered some sort of I/O error
769  *
770  * comment on "cleaning" object and PG_BUSY pages:
771  *	this routine is holding the lock on the object.   the only time
772  *	that it can run into a PG_BUSY page that it does not own is if
773  *	some other process has started I/O on the page (e.g. either
774  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
775  *	in, then it can not be dirty (!PG_CLEAN) because no one has
776  *	had a chance to modify it yet.    if the PG_BUSY page is being
777  *	paged out then it means that someone else has already started
778  *	cleaning the page for us (how nice!).    in this case, if we
779  *	have syncio specified, then after we make our pass through the
780  *	object we need to wait for the other PG_BUSY pages to clear
781  *	off (i.e. we need to do an iosync).   also note that once a
782  *	page is PG_BUSY it must stay in its object until it is un-busyed.
783  *
784  * note on page traversal:
785  *	we can traverse the pages in an object either by going down the
786  *	linked list in "uobj->memq", or we can go over the address range
787  *	by page doing hash table lookups for each address.    depending
788  *	on how many pages are in the object it may be cheaper to do one
789  *	or the other.   we set "by_list" to true if we are using memq.
790  *	if the cost of a hash lookup was equal to the cost of the list
791  *	traversal we could compare the number of pages in the start->stop
792  *	range to the total number of pages in the object.   however, it
793  *	seems that a hash table lookup is more expensive than the linked
794  *	list traversal, so we multiply the number of pages in the
795  *	start->stop range by a penalty which we define below.
796  */
797 
798 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
799 
800 static boolean_t
801 uvn_flush(uobj, start, stop, flags)
802 	struct uvm_object *uobj;
803 	vm_offset_t start, stop;
804 	int flags;
805 {
806 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
807 	struct vm_page *pp, *ppnext, *ptmp;
808 	struct vm_page *pps[MAXBSIZE/PAGE_SIZE], **ppsp;
809 	int npages, result, lcv;
810 	boolean_t retval, need_iosync, by_list, needs_clean;
811 	vm_offset_t curoff;
812 	u_short pp_version;
813 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
814 
815 	curoff = 0;	/* XXX: shut up gcc */
816 	/*
817 	 * get init vals and determine how we are going to traverse object
818 	 */
819 
820 	need_iosync = FALSE;
821 	retval = TRUE;		/* return value */
822 	if (flags & PGO_ALLPAGES) {
823 		start = 0;
824 		stop = round_page(uvn->u_size);
825 		by_list = TRUE;		/* always go by the list */
826 	} else {
827 		start = trunc_page(start);
828 		stop = round_page(stop);
829 		if (stop > round_page(uvn->u_size))
830 			printf("uvn_flush: strange, got an out of range "
831 			    "flush (fixed)\n");
832 
833 		by_list = (uobj->uo_npages <=
834 		    ((stop - start) / PAGE_SIZE) * UVN_HASH_PENALTY);
835 	}
836 
837 	UVMHIST_LOG(maphist,
838 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
839 	    start, stop, by_list, flags);
840 
841 	/*
842 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
843 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
844 	 * is wrong it will only prevent us from clustering... it won't break
845 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
846 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
847 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
848 	 * bit is always up to date since there are no mappings).
849 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
850 	 */
851 
852 	if ((flags & PGO_CLEANIT) != 0 &&
853 	    uobj->pgops->pgo_mk_pcluster != NULL) {
854 		if (by_list) {
855 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
856 			    pp = pp->listq.tqe_next) {
857 				if (pp->offset < start || pp->offset >= stop)
858 					continue;
859 				pp->flags &= ~PG_CLEANCHK;
860 			}
861 
862 		} else {   /* by hash */
863 			for (curoff = start ; curoff < stop;
864 			    curoff += PAGE_SIZE) {
865 				pp = uvm_pagelookup(uobj, curoff);
866 				if (pp)
867 					pp->flags &= ~PG_CLEANCHK;
868 			}
869 		}
870 	}
871 
872 	/*
873 	 * now do it.   note: we must update ppnext in body of loop or we
874 	 * will get stuck.  we need to use ppnext because we may free "pp"
875 	 * before doing the next loop.
876 	 */
877 
878 	if (by_list) {
879 		pp = uobj->memq.tqh_first;
880 	} else {
881 		curoff = start;
882 		pp = uvm_pagelookup(uobj, curoff);
883 	}
884 
885 	ppnext = NULL;	/* XXX: shut up gcc */
886 	ppsp = NULL;		/* XXX: shut up gcc */
887 	uvm_lock_pageq();	/* page queues locked */
888 
889 	/* locked: both page queues and uobj */
890 	for ( ; (by_list && pp != NULL) ||
891 	  (!by_list && curoff < stop) ; pp = ppnext) {
892 
893 		if (by_list) {
894 
895 			/*
896 			 * range check
897 			 */
898 
899 			if (pp->offset < start || pp->offset >= stop) {
900 				ppnext = pp->listq.tqe_next;
901 				continue;
902 			}
903 
904 		} else {
905 
906 			/*
907 			 * null check
908 			 */
909 
910 			curoff += PAGE_SIZE;
911 			if (pp == NULL) {
912 				if (curoff < stop)
913 					ppnext = uvm_pagelookup(uobj, curoff);
914 				continue;
915 			}
916 
917 		}
918 
919 		/*
920 		 * handle case where we do not need to clean page (either
921 		 * because we are not clean or because page is not dirty or
922 		 * is busy):
923 		 *
924 		 * NOTE: we are allowed to deactivate a non-wired active
925 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
926 		 * queue it must stay put until it is !PG_BUSY (so as not to
927 		 * confuse pagedaemon).
928 		 */
929 
930 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
931 			needs_clean = FALSE;
932 			if ((pp->flags & PG_BUSY) != 0 &&
933 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
934 			             (PGO_CLEANIT|PGO_SYNCIO))
935 				need_iosync = TRUE;
936 		} else {
937 			/*
938 			 * freeing: nuke all mappings so we can sync
939 			 * PG_CLEAN bit with no race
940 			 */
941 			if ((pp->flags & PG_CLEAN) != 0 &&
942 			    (flags & PGO_FREE) != 0 &&
943 			    (pp->pqflags & PQ_ACTIVE) != 0)
944 				pmap_page_protect(PMAP_PGARG(pp), VM_PROT_NONE);
945 			if ((pp->flags & PG_CLEAN) != 0 &&
946 			    pmap_is_modified(PMAP_PGARG(pp)))
947 				pp->flags &= ~(PG_CLEAN);
948 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
949 
950 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
951 		}
952 
953 		/*
954 		 * if we don't need a clean... load ppnext and dispose of pp
955 		 */
956 		if (!needs_clean) {
957 			/* load ppnext */
958 			if (by_list)
959 				ppnext = pp->listq.tqe_next;
960 			else {
961 				if (curoff < stop)
962 					ppnext = uvm_pagelookup(uobj, curoff);
963 			}
964 
965 			/* now dispose of pp */
966 			if (flags & PGO_DEACTIVATE) {
967 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
968 				    pp->wire_count == 0) {
969 					pmap_page_protect(PMAP_PGARG(pp),
970 					    VM_PROT_NONE);
971 					uvm_pagedeactivate(pp);
972 				}
973 
974 			} else if (flags & PGO_FREE) {
975 				if (pp->flags & PG_BUSY) {
976 					/* release busy pages */
977 					pp->flags |= PG_RELEASED;
978 				} else {
979 					pmap_page_protect(PMAP_PGARG(pp),
980 					    VM_PROT_NONE);
981 					/* removed page from object */
982 					uvm_pagefree(pp);
983 				}
984 			}
985 			/* ppnext is valid so we can continue... */
986 			continue;
987 		}
988 
989 		/*
990 		 * pp points to a page in the locked object that we are
991 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
992 		 * for cleaning (PGO_CLEANIT).  we clean it now.
993 		 *
994 		 * let uvm_pager_put attempted a clustered page out.
995 		 * note: locked: uobj and page queues.
996 		 */
997 
998 		pp->flags |= PG_BUSY;	/* we 'own' page now */
999 		UVM_PAGE_OWN(pp, "uvn_flush");
1000 		pmap_page_protect(PMAP_PGARG(pp), VM_PROT_READ);
1001 		pp_version = pp->version;
1002 ReTry:
1003 		ppsp = pps;
1004 		npages = sizeof(pps) / sizeof(struct vm_page *);
1005 
1006 		/* locked: page queues, uobj */
1007 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1008 			   flags | PGO_DOACTCLUST, start, stop);
1009 		/* unlocked: page queues, uobj */
1010 
1011 		/*
1012 		 * at this point nothing is locked.   if we did an async I/O
1013 		 * it is remotely possible for the async i/o to complete and
1014 		 * the page "pp" be freed or what not before we get a chance
1015 		 * to relock the object.   in order to detect this, we have
1016 		 * saved the version number of the page in "pp_version".
1017 		 */
1018 
1019 		/* relock! */
1020 		simple_lock(&uobj->vmobjlock);
1021 		uvm_lock_pageq();
1022 
1023 		/*
1024 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1025 		 * can only happen when  we are doing async I/O and can't
1026 		 * map the pages into kernel memory (pager_map) due to lack
1027 		 * of vm space.   if this happens we drop back to sync I/O.
1028 		 */
1029 
1030 		if (result == VM_PAGER_AGAIN) {
1031 			/*
1032 			 * it is unlikely, but page could have been released
1033 			 * while we had the object lock dropped.   we ignore
1034 			 * this now and retry the I/O.  we will detect and
1035 			 * handle the released page after the syncio I/O
1036 			 * completes.
1037 			 */
1038 #ifdef DIAGNOSTIC
1039 			if (flags & PGO_SYNCIO)
1040 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1041 #endif
1042 			flags |= PGO_SYNCIO;
1043 			goto ReTry;
1044 		}
1045 
1046 		/*
1047 		 * the cleaning operation is now done.   finish up.  note that
1048 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1049 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1050 		 * to us in ppsp/npages.
1051 		 */
1052 
1053 		/*
1054 		 * for pending async i/o if we are not deactivating/freeing
1055 		 * we can move on to the next page.
1056 		 */
1057 
1058 		if (result == VM_PAGER_PEND) {
1059 
1060 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1061 				/*
1062 				 * no per-page ops: refresh ppnext and continue
1063 				 */
1064 				if (by_list) {
1065 					if (pp->version == pp_version)
1066 						ppnext = pp->listq.tqe_next;
1067 					else
1068 						/* reset */
1069 						ppnext = uobj->memq.tqh_first;
1070 				} else {
1071 					if (curoff < stop)
1072 						ppnext = uvm_pagelookup(uobj,
1073 						    curoff);
1074 				}
1075 				continue;
1076 			}
1077 
1078 			/* need to do anything here? */
1079 		}
1080 
1081 		/*
1082 		 * need to look at each page of the I/O operation.  we defer
1083 		 * processing "pp" until the last trip through this "for" loop
1084 		 * so that we can load "ppnext" for the main loop after we
1085 		 * play with the cluster pages [thus the "npages + 1" in the
1086 		 * loop below].
1087 		 */
1088 
1089 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1090 
1091 			/*
1092 			 * handle ppnext for outside loop, and saving pp
1093 			 * until the end.
1094 			 */
1095 			if (lcv < npages) {
1096 				if (ppsp[lcv] == pp)
1097 					continue; /* skip pp until the end */
1098 				ptmp = ppsp[lcv];
1099 			} else {
1100 				ptmp = pp;
1101 
1102 				/* set up next page for outer loop */
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, curoff);
1112 				}
1113 			}
1114 
1115 			/*
1116 			 * verify the page didn't get moved while obj was
1117 			 * unlocked
1118 			 */
1119 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1120 				continue;
1121 
1122 			/*
1123 			 * unbusy the page if I/O is done.   note that for
1124 			 * pending I/O it is possible that the I/O op
1125 			 * finished before we relocked the object (in
1126 			 * which case the page is no longer busy).
1127 			 */
1128 
1129 			if (result != VM_PAGER_PEND) {
1130 				if (ptmp->flags & PG_WANTED)
1131 					/* still holding object lock */
1132 					thread_wakeup(ptmp);
1133 
1134 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1135 				UVM_PAGE_OWN(ptmp, NULL);
1136 				if (ptmp->flags & PG_RELEASED) {
1137 
1138 					/* pgo_releasepg wants this */
1139 					uvm_unlock_pageq();
1140 					if (!uvn_releasepg(ptmp, NULL))
1141 						return (TRUE);
1142 
1143 					uvm_lock_pageq();	/* relock */
1144 					continue;		/* next page */
1145 
1146 				} else {
1147 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1148 					if ((flags & PGO_FREE) == 0)
1149 						pmap_clear_modify(
1150 						    PMAP_PGARG(ptmp));
1151 				}
1152 			}
1153 
1154 			/*
1155 			 * dispose of page
1156 			 */
1157 
1158 			if (flags & PGO_DEACTIVATE) {
1159 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1160 				    pp->wire_count == 0) {
1161 					pmap_page_protect(PMAP_PGARG(ptmp),
1162 					    VM_PROT_NONE);
1163 					uvm_pagedeactivate(ptmp);
1164 				}
1165 
1166 			} else if (flags & PGO_FREE) {
1167 				if (result == VM_PAGER_PEND) {
1168 					if ((ptmp->flags & PG_BUSY) != 0)
1169 						/* signal for i/o done */
1170 						ptmp->flags |= PG_RELEASED;
1171 				} else {
1172 					if (result != VM_PAGER_OK) {
1173 						printf("uvn_flush: obj=%p, "
1174 						   "offset=0x%lx.  error "
1175 						   "during pageout.\n",
1176 						    pp->uobject, pp->offset);
1177 						printf("uvn_flush: WARNING: "
1178 						    "changes to page may be "
1179 						    "lost!\n");
1180 						retval = FALSE;
1181 					}
1182 					pmap_page_protect(PMAP_PGARG(ptmp),
1183 					    VM_PROT_NONE);
1184 					uvm_pagefree(ptmp);
1185 				}
1186 			}
1187 
1188 		}		/* end of "lcv" for loop */
1189 
1190 	}		/* end of "pp" for loop */
1191 
1192 	/*
1193 	 * done with pagequeues: unlock
1194 	 */
1195 	uvm_unlock_pageq();
1196 
1197 	/*
1198 	 * now wait for all I/O if required.
1199 	 */
1200 	if (need_iosync) {
1201 
1202 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1203 		while (uvn->u_nio != 0) {
1204 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1205 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1206 			  FALSE, "uvn_flush",0);
1207 			simple_lock(&uvn->u_obj.vmobjlock);
1208 		}
1209 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1210 			wakeup(&uvn->u_flags);
1211 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1212 	}
1213 
1214 	/* return, with object locked! */
1215 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1216 	return(retval);
1217 }
1218 
1219 /*
1220  * uvn_cluster
1221  *
1222  * we are about to do I/O in an object at offset.   this function is called
1223  * to establish a range of offsets around "offset" in which we can cluster
1224  * I/O.
1225  *
1226  * - currently doesn't matter if obj locked or not.
1227  */
1228 
1229 static void
1230 uvn_cluster(uobj, offset, loffset, hoffset)
1231 	struct uvm_object *uobj;
1232 	vm_offset_t offset;
1233 	vm_offset_t *loffset, *hoffset; /* OUT */
1234 {
1235 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1236 	*loffset = offset;
1237 
1238 	if (*loffset >= uvn->u_size)
1239 		panic("uvn_cluster: offset out of range");
1240 
1241 	/*
1242 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1243 	 */
1244 	*hoffset = *loffset + MAXBSIZE;
1245 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1246 		*hoffset = round_page(uvn->u_size);
1247 
1248 	return;
1249 }
1250 
1251 /*
1252  * uvn_put: flush page data to backing store.
1253  *
1254  * => prefer map unlocked (not required)
1255  * => object must be locked!   we will _unlock_ it before starting I/O.
1256  * => flags: PGO_SYNCIO -- use sync. I/O
1257  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1258  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1259  *	[thus we never do async i/o!  see iodone comment]
1260  */
1261 
1262 static int
1263 uvn_put(uobj, pps, npages, flags)
1264 	struct uvm_object *uobj;
1265 	struct vm_page **pps;
1266 	int npages, flags;
1267 {
1268 	int retval;
1269 
1270 	/* note: object locked */
1271 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1272 	/* note: object unlocked */
1273 
1274 	return(retval);
1275 }
1276 
1277 
1278 /*
1279  * uvn_get: get pages (synchronously) from backing store
1280  *
1281  * => prefer map unlocked (not required)
1282  * => object must be locked!  we will _unlock_ it before starting any I/O.
1283  * => flags: PGO_ALLPAGES: get all of the pages
1284  *           PGO_LOCKED: fault data structures are locked
1285  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1286  * => NOTE: caller must check for released pages!!
1287  */
1288 
1289 static int
1290 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1291 	struct uvm_object *uobj;
1292 	vm_offset_t offset;
1293 	struct vm_page **pps;		/* IN/OUT */
1294 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1295 	int centeridx, advice, flags;
1296 	vm_prot_t access_type;
1297 {
1298 	vm_offset_t current_offset;
1299 	struct vm_page *ptmp;
1300 	int lcv, result, gotpages;
1301 	boolean_t done;
1302 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1303 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1304 
1305 	/*
1306 	 * step 1: handled the case where fault data structures are locked.
1307 	 */
1308 
1309 	if (flags & PGO_LOCKED) {
1310 
1311 		/*
1312 		 * gotpages is the current number of pages we've gotten (which
1313 		 * we pass back up to caller via *npagesp.
1314 		 */
1315 
1316 		gotpages = 0;
1317 
1318 		/*
1319 		 * step 1a: get pages that are already resident.   only do this
1320 		 * if the data structures are locked (i.e. the first time
1321 		 * through).
1322 		 */
1323 
1324 		done = TRUE;	/* be optimistic */
1325 
1326 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1327 		    lcv++, current_offset += PAGE_SIZE) {
1328 
1329 			/* do we care about this page?  if not, skip it */
1330 			if (pps[lcv] == PGO_DONTCARE)
1331 				continue;
1332 
1333 			/* lookup page */
1334 			ptmp = uvm_pagelookup(uobj, current_offset);
1335 
1336 			/* to be useful must get a non-busy, non-released pg */
1337 			if (ptmp == NULL ||
1338 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1339 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1340 				    != 0)
1341 				done = FALSE;	/* need to do a wait or I/O! */
1342 				continue;
1343 			}
1344 
1345 			/*
1346 			 * useful page: busy/lock it and plug it in our
1347 			 * result array
1348 			 */
1349 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1350 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1351 			pps[lcv] = ptmp;
1352 			gotpages++;
1353 
1354 		}	/* "for" lcv loop */
1355 
1356 		/*
1357 		 * XXX: given the "advice", should we consider async read-ahead?
1358 		 * XXX: fault current does deactive of pages behind us.  is
1359 		 * this good (other callers might now).
1360 		 */
1361 		/*
1362 		 * XXX: read-ahead currently handled by buffer cache (bread)
1363 		 * level.
1364 		 * XXX: no async i/o available.
1365 		 * XXX: so we don't do anything now.
1366 		 */
1367 
1368 		/*
1369 		 * step 1c: now we've either done everything needed or we to
1370 		 * unlock and do some waiting or I/O.
1371 		 */
1372 
1373 		*npagesp = gotpages;		/* let caller know */
1374 		if (done)
1375 			return(VM_PAGER_OK);		/* bingo! */
1376 		else
1377 			/* EEK!   Need to unlock and I/O */
1378 			return(VM_PAGER_UNLOCK);
1379 	}
1380 
1381 	/*
1382 	 * step 2: get non-resident or busy pages.
1383 	 * object is locked.   data structures are unlocked.
1384 	 *
1385 	 * XXX: because we can't do async I/O at this level we get things
1386 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1387 	 * async-read-ahead for us at a lower level.
1388 	 */
1389 
1390 	for (lcv = 0, current_offset = offset ;
1391 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1392 
1393 		/* skip over pages we've already gotten or don't want */
1394 		/* skip over pages we don't _have_ to get */
1395 		if (pps[lcv] != NULL || (lcv != centeridx &&
1396 		    (flags & PGO_ALLPAGES) == 0))
1397 			continue;
1398 
1399 		/*
1400 		 * we have yet to locate the current page (pps[lcv]).   we first
1401 		 * look for a page that is already at the current offset.   if
1402 		 * we fine a page, we check to see if it is busy or released.
1403 		 * if that is the case, then we sleep on the page until it is
1404 		 * no longer busy or released and repeat the lookup.    if the
1405 		 * page we found is neither busy nor released, then we busy it
1406 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1407 		 * following while loop and indicates we are ready to move on
1408 		 * to the next page in the "lcv" loop above.
1409 		 *
1410 		 * if we exit the while loop with pps[lcv] still set to NULL,
1411 		 * then it means that we allocated a new busy/fake/clean page
1412 		 * ptmp in the object and we need to do I/O to fill in the data.
1413 		 */
1414 
1415 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1416 
1417 			/* look for a current page */
1418 			ptmp = uvm_pagelookup(uobj, current_offset);
1419 
1420 			/* nope?   allocate one now (if we can) */
1421 			if (ptmp == NULL) {
1422 
1423 				ptmp = uvm_pagealloc(uobj, current_offset,
1424 				    NULL);	/* alloc */
1425 
1426 				/* out of RAM? */
1427 				if (ptmp == NULL) {
1428 					simple_unlock(&uobj->vmobjlock);
1429 					uvm_wait("uvn_getpage");
1430 					simple_lock(&uobj->vmobjlock);
1431 
1432 					/* goto top of pps while loop */
1433 					continue;
1434 				}
1435 
1436 				/*
1437 				 * got new page ready for I/O.  break pps
1438 				 * while loop.  pps[lcv] is still NULL.
1439 				 */
1440 				break;
1441 			}
1442 
1443 			/* page is there, see if we need to wait on it */
1444 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1445 				ptmp->flags |= PG_WANTED;
1446 				UVM_UNLOCK_AND_WAIT(ptmp,
1447 				    &uobj->vmobjlock, 0, "uvn_get",0);
1448 				simple_lock(&uobj->vmobjlock);
1449 				continue;	/* goto top of pps while loop */
1450 			}
1451 
1452 			/*
1453 			 * if we get here then the page has become resident
1454 			 * and unbusy between steps 1 and 2.  we busy it
1455 			 * now (so we own it) and set pps[lcv] (so that we
1456 			 * exit the while loop).
1457 			 */
1458 			ptmp->flags |= PG_BUSY;
1459 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1460 			pps[lcv] = ptmp;
1461 		}
1462 
1463 		/*
1464 		 * if we own the a valid page at the correct offset, pps[lcv]
1465 		 * will point to it.   nothing more to do except go to the
1466 		 * next page.
1467 		 */
1468 
1469 		if (pps[lcv])
1470 			continue;			/* next lcv */
1471 
1472 		/*
1473 		 * we have a "fake/busy/clean" page that we just allocated.  do
1474 		 * I/O to fill it with valid data.  note that object must be
1475 		 * locked going into uvn_io, but will be unlocked afterwards.
1476 		 */
1477 
1478 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1479 		    PGO_SYNCIO, UIO_READ);
1480 
1481 		/*
1482 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1483 		 * syncio the result can not be PEND or AGAIN.   we must relock
1484 		 * and check for errors.
1485 		 */
1486 
1487 		/* lock object.   check for errors.   */
1488 		simple_lock(&uobj->vmobjlock);
1489 		if (result != VM_PAGER_OK) {
1490 			if (ptmp->flags & PG_WANTED)
1491 				/* object lock still held */
1492 				thread_wakeup(ptmp);
1493 
1494 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1495 			UVM_PAGE_OWN(ptmp, NULL);
1496 			uvm_lock_pageq();
1497 			uvm_pagefree(ptmp);
1498 			uvm_unlock_pageq();
1499 			simple_unlock(&uobj->vmobjlock);
1500 			return(result);
1501 		}
1502 
1503 		/*
1504 		 * we got the page!   clear the fake flag (indicates valid
1505 		 * data now in page) and plug into our result array.   note
1506 		 * that page is still busy.
1507 		 *
1508 		 * it is the callers job to:
1509 		 * => check if the page is released
1510 		 * => unbusy the page
1511 		 * => activate the page
1512 		 */
1513 
1514 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1515 		pmap_clear_modify(PMAP_PGARG(ptmp));	/* ... and clean */
1516 		pps[lcv] = ptmp;
1517 
1518 	}	/* lcv loop */
1519 
1520 	/*
1521 	 * finally, unlock object and return.
1522 	 */
1523 
1524 	simple_unlock(&uobj->vmobjlock);
1525 	return (VM_PAGER_OK);
1526 }
1527 
1528 /*
1529  * uvn_asyncget: start async I/O to bring pages into ram
1530  *
1531  * => caller must lock object(???XXX: see if this is best)
1532  * => could be called from uvn_get or a madvise() fault-ahead.
1533  * => if it fails, it doesn't matter.
1534  */
1535 
1536 static int
1537 uvn_asyncget(uobj, offset, npages)
1538 	struct uvm_object *uobj;
1539 	vm_offset_t offset;
1540 	int npages;
1541 {
1542 
1543 	/*
1544 	 * XXXCDC: we can't do async I/O yet
1545 	 */
1546 	printf("uvn_asyncget called\n");
1547 	return (KERN_SUCCESS);
1548 }
1549 
1550 /*
1551  * uvn_io: do I/O to a vnode
1552  *
1553  * => prefer map unlocked (not required)
1554  * => object must be locked!   we will _unlock_ it before starting I/O.
1555  * => flags: PGO_SYNCIO -- use sync. I/O
1556  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1557  *	[thus we never do async i/o!  see iodone comment]
1558  */
1559 
1560 static int
1561 uvn_io(uvn, pps, npages, flags, rw)
1562 	struct uvm_vnode *uvn;
1563 	vm_page_t *pps;
1564 	int npages, flags, rw;
1565 {
1566 	struct vnode *vn;
1567 	struct uio uio;
1568 	struct iovec iov;
1569 	vm_offset_t kva, file_offset;
1570 	int waitf, result, got, wanted;
1571 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1572 
1573 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1574 
1575 	/*
1576 	 * init values
1577 	 */
1578 
1579 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1580 	vn = (struct vnode *) uvn;
1581 	file_offset = pps[0]->offset;
1582 
1583 	/*
1584 	 * check for sync'ing I/O.
1585 	 */
1586 
1587 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1588 		if (waitf == M_NOWAIT) {
1589 			simple_unlock(&uvn->u_obj.vmobjlock);
1590 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1591 			return(VM_PAGER_AGAIN);
1592 		}
1593 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1594 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1595 			FALSE, "uvn_iosync",0);
1596 		simple_lock(&uvn->u_obj.vmobjlock);
1597 	}
1598 
1599 	/*
1600 	 * check size
1601 	 */
1602 
1603 	if (file_offset >= uvn->u_size) {
1604 			simple_unlock(&uvn->u_obj.vmobjlock);
1605 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1606 #ifdef DIAGNOSTIC
1607 			printf("uvn_io: note: size check fired\n");
1608 #endif
1609 			return(VM_PAGER_BAD);
1610 	}
1611 
1612 	/*
1613 	 * first try and map the pages in (without waiting)
1614 	 */
1615 
1616 	kva = uvm_pagermapin(pps, npages, NULL, M_NOWAIT);
1617 	if (kva == NULL && waitf == M_NOWAIT) {
1618 		simple_unlock(&uvn->u_obj.vmobjlock);
1619 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1620 		return(VM_PAGER_AGAIN);
1621 	}
1622 
1623 	/*
1624 	 * ok, now bump u_nio up.   at this point we are done with uvn
1625 	 * and can unlock it.   if we still don't have a kva, try again
1626 	 * (this time with sleep ok).
1627 	 */
1628 
1629 	uvn->u_nio++;			/* we have an I/O in progress! */
1630 	simple_unlock(&uvn->u_obj.vmobjlock);
1631 	/* NOTE: object now unlocked */
1632 	if (kva == NULL) {
1633 		kva = uvm_pagermapin(pps, npages, NULL, M_WAITOK);
1634 	}
1635 
1636 	/*
1637 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1638 	 * get touched (so we can look at "offset" without having to lock
1639 	 * the object).  set up for I/O.
1640 	 */
1641 
1642 	/*
1643 	 * fill out uio/iov
1644 	 */
1645 
1646 	iov.iov_base = (caddr_t) kva;
1647 	wanted = npages * PAGE_SIZE;
1648 	if (file_offset + wanted > uvn->u_size)
1649 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1650 	iov.iov_len = wanted;
1651 	uio.uio_iov = &iov;
1652 	uio.uio_iovcnt = 1;
1653 	uio.uio_offset = file_offset;
1654 	uio.uio_segflg = UIO_SYSSPACE;
1655 	uio.uio_rw = rw;
1656 	uio.uio_resid = wanted;
1657 	uio.uio_procp = NULL;
1658 
1659 	/*
1660 	 * do the I/O!  (XXX: curproc?)
1661 	 */
1662 
1663 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1664 
1665 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1666 		vn_lock(vn, LK_EXCLUSIVE | LK_RETRY);
1667 	/* NOTE: vnode now locked! */
1668 
1669 	if (rw == UIO_READ)
1670 		result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1671 	else
1672 		result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1673 
1674 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1675 		VOP_UNLOCK(vn, 0);
1676 	/* NOTE: vnode now unlocked (unless vnislocked) */
1677 
1678 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1679 
1680 	/*
1681 	 * result == unix style errno (0 == OK!)
1682 	 *
1683 	 * zero out rest of buffer (if needed)
1684 	 */
1685 
1686 	if (result == 0) {
1687 		got = wanted - uio.uio_resid;
1688 
1689 		if (wanted && got == 0) {
1690 			result = EIO;		/* XXX: error? */
1691 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1692 			bzero((void *) (kva + got), (PAGE_SIZE * npages) - got);
1693 		}
1694 	}
1695 
1696 	/*
1697 	 * now remove pager mapping
1698 	 */
1699 	uvm_pagermapout(kva, npages);
1700 
1701 	/*
1702 	 * now clean up the object (i.e. drop I/O count)
1703 	 */
1704 
1705 	simple_lock(&uvn->u_obj.vmobjlock);
1706 	/* NOTE: object now locked! */
1707 
1708 	uvn->u_nio--;			/* I/O DONE! */
1709 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1710 		wakeup(&uvn->u_nio);
1711 	}
1712 	simple_unlock(&uvn->u_obj.vmobjlock);
1713 	/* NOTE: object now unlocked! */
1714 
1715 	/*
1716 	 * done!
1717 	 */
1718 
1719 	UVMHIST_LOG(maphist, "<- done (result %d)", result,0,0,0);
1720 	if (result == 0)
1721 		return(VM_PAGER_OK);
1722 	else
1723 		return(VM_PAGER_ERROR);
1724 }
1725 
1726 /*
1727  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1728  * is gone we will kill the object (flushing dirty pages back to the vnode
1729  * if needed).
1730  *
1731  * => returns TRUE if there was no uvm_object attached or if there was
1732  *	one and we killed it [i.e. if there is no active uvn]
1733  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1734  *	needed]
1735  *
1736  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1737  *	having to hold a reference on the vnode) and given a working
1738  *	uvm_vnp_sync(), how does that effect the need for this function?
1739  *      [XXXCDC: seems like it can die?]
1740  *
1741  * => XXX: this function should DIE once we merge the VM and buffer
1742  *	cache.
1743  *
1744  * research shows that this is called in the following places:
1745  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1746  *	changes sizes
1747  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1748  *	are written to
1749  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1750  *	is off
1751  * ffs_realloccg: when we can't extend the current block and have
1752  *	to allocate a new one we call this [XXX: why?]
1753  * nfsrv_rename, rename_files: called when the target filename is there
1754  *	and we want to remove it
1755  * nfsrv_remove, sys_unlink: called on file we are removing
1756  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1757  *	then return "text busy"
1758  * nfs_open: seems to uncache any file opened with nfs
1759  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1760  */
1761 
1762 boolean_t
1763 uvm_vnp_uncache(vp)
1764 	struct vnode *vp;
1765 {
1766 	struct uvm_vnode *uvn = &vp->v_uvm;
1767 
1768 	/*
1769 	 * lock uvn part of the vnode and check to see if we need to do anything
1770 	 */
1771 
1772 	simple_lock(&uvn->u_obj.vmobjlock);
1773 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1774 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1775 		simple_unlock(&uvn->u_obj.vmobjlock);
1776 		return(TRUE);
1777 	}
1778 
1779 	/*
1780 	 * we have a valid, non-blocked uvn.   clear persist flag.
1781 	 * if uvn is currently active we can return now.
1782 	 */
1783 
1784 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1785 	if (uvn->u_obj.uo_refs) {
1786 		simple_unlock(&uvn->u_obj.vmobjlock);
1787 		return(FALSE);
1788 	}
1789 
1790 	/*
1791 	 * uvn is currently persisting!   we have to gain a reference to
1792 	 * it so that we can call uvn_detach to kill the uvn.
1793 	 */
1794 
1795 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1796 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1797 	simple_unlock(&uvn->u_obj.vmobjlock);
1798 
1799 
1800 #ifdef DEBUG
1801 	/*
1802 	 * carry over sanity check from old vnode pager: the vnode should
1803 	 * be VOP_LOCK'd, and we confirm it here.
1804 	 */
1805 	if (!VOP_ISLOCKED(vp)) {
1806 		boolean_t is_ok_anyway = FALSE;
1807 #ifdef NFS
1808 		extern int (**nfsv2_vnodeop_p) __P((void *));
1809 		extern int (**spec_nfsv2nodeop_p) __P((void *));
1810 #ifdef FIFO
1811 		extern int (**fifo_nfsv2nodeop_p) __P((void *));
1812 #endif	/* FIFO */
1813 
1814 		/* vnode is NOT VOP_LOCKed: some vnode types _never_ lock */
1815 		if (vp->v_op == nfsv2_vnodeop_p ||
1816 		    vp->v_op == spec_nfsv2nodeop_p) {
1817 			is_ok_anyway = TRUE;
1818 		}
1819 #ifdef FIFO
1820 		if (vp->v_op == fifo_nfsv2nodeop_p) {
1821 			is_ok_anyway = TRUE;
1822 		}
1823 #endif	/* FIFO */
1824 #endif	/* NFS */
1825 		if (!is_ok_anyway)
1826 			panic("uvm_vnp_uncache: vnode not locked!");
1827 	}
1828 #endif	/* DEBUG */
1829 
1830 	/*
1831 	 * now drop our reference to the vnode.   if we have the sole
1832 	 * reference to the vnode then this will cause it to die [as we
1833 	 * just cleared the persist flag].   we have to unlock the vnode
1834 	 * while we are doing this as it may trigger I/O.
1835 	 *
1836 	 * XXX: it might be possible for uvn to get reclaimed while we are
1837 	 * unlocked causing us to return TRUE when we should not.   we ignore
1838 	 * this as a false-positive return value doesn't hurt us.
1839 	 */
1840 	VOP_UNLOCK(vp, 0);
1841 	uvn_detach(&uvn->u_obj);
1842 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1843 
1844 	/*
1845 	 * and return...
1846 	 */
1847 
1848 	return(TRUE);
1849 }
1850 
1851 /*
1852  * uvm_vnp_setsize: grow or shrink a vnode uvn
1853  *
1854  * grow   => just update size value
1855  * shrink => toss un-needed pages
1856  *
1857  * => we assume that the caller has a reference of some sort to the
1858  *	vnode in question so that it will not be yanked out from under
1859  *	us.
1860  *
1861  * called from:
1862  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1863  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1864  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1865  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1866  *  => union fs: union_newsize
1867  */
1868 
1869 void
1870 uvm_vnp_setsize(vp, newsize)
1871 	struct vnode *vp;
1872 	u_quad_t newsize;
1873 {
1874 	struct uvm_vnode *uvn = &vp->v_uvm;
1875 
1876 	/*
1877 	 * lock uvn and check for valid object, and if valid: do it!
1878 	 */
1879 	simple_lock(&uvn->u_obj.vmobjlock);
1880 	if (uvn->u_flags & UVM_VNODE_VALID) {
1881 
1882 		/*
1883 		 * make sure that the newsize fits within a vm_offset_t
1884 		 * XXX: need to revise addressing data types
1885 		 */
1886 
1887 		if (newsize > (vm_offset_t) -PAGE_SIZE) {
1888 #ifdef DEBUG
1889 			printf("uvm_vnp_setsize: vn %p size truncated "
1890 			    "%qx->%lx\n", vp, newsize, (vm_offset_t)-PAGE_SIZE);
1891 #endif
1892 			newsize = (vm_offset_t)-PAGE_SIZE;
1893 		}
1894 
1895 		/*
1896 		 * now check if the size has changed: if we shrink we had better
1897 		 * toss some pages...
1898 		 */
1899 
1900 		if (uvn->u_size > newsize) {
1901 			(void)uvn_flush(&uvn->u_obj, (vm_offset_t) newsize,
1902 			    uvn->u_size, PGO_FREE);
1903 		}
1904 		uvn->u_size = (vm_offset_t)newsize;
1905 	}
1906 	simple_unlock(&uvn->u_obj.vmobjlock);
1907 
1908 	/*
1909 	 * done
1910 	 */
1911 	return;
1912 }
1913 
1914 /*
1915  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1916  *
1917  * => called from sys_sync with no VM structures locked
1918  * => only one process can do a sync at a time (because the uvn
1919  *    structure only has one queue for sync'ing).  we ensure this
1920  *    by holding the uvn_sync_lock while the sync is in progress.
1921  *    other processes attempting a sync will sleep on this lock
1922  *    until we are done.
1923  */
1924 
1925 void
1926 uvm_vnp_sync(mp)
1927 	struct mount *mp;
1928 {
1929 	struct uvm_vnode *uvn;
1930 	struct vnode *vp;
1931 	boolean_t got_lock;
1932 
1933 	/*
1934 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1935 	 * our lock...
1936 	 */
1937 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0);
1938 
1939 	/*
1940 	 * step 2: build up a simpleq of uvns of interest based on the
1941 	 * write list.   we gain a reference to uvns of interest.  must
1942 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1943 	 * in the body of the loop.
1944 	 */
1945 	SIMPLEQ_INIT(&uvn_sync_q);
1946 	simple_lock(&uvn_wl_lock);
1947 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1948 	    uvn = uvn->u_wlist.le_next) {
1949 
1950 		vp = (struct vnode *) uvn;
1951 		if (mp && vp->v_mount != mp)
1952 			continue;
1953 
1954 		/* attempt to gain reference */
1955 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1956 		    						FALSE &&
1957 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1958 			/* spin */ ;
1959 
1960 		/*
1961 		 * we will exit the loop if either if the following are true:
1962 		 *  - we got the lock [always true if NCPU == 1]
1963 		 *  - we failed to get the lock but noticed the vnode was
1964 		 * 	"blocked" -- in this case the vnode must be a dying
1965 		 *	vnode, and since dying vnodes are in the process of
1966 		 *	being flushed out, we can safely skip this one
1967 		 *
1968 		 * we want to skip over the vnode if we did not get the lock,
1969 		 * or if the vnode is already dying (due to the above logic).
1970 		 *
1971 		 * note that uvn must already be valid because we found it on
1972 		 * the wlist (this also means it can't be ALOCK'd).
1973 		 */
1974 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1975 			if (got_lock)
1976 				simple_unlock(&uvn->u_obj.vmobjlock);
1977 			continue;		/* skip it */
1978 		}
1979 
1980 		/*
1981 		 * gain reference.   watch out for persisting uvns (need to
1982 		 * regain vnode REF).
1983 		 */
1984 		if (uvn->u_obj.uo_refs == 0)
1985 			VREF(vp);
1986 		uvn->u_obj.uo_refs++;
1987 		simple_unlock(&uvn->u_obj.vmobjlock);
1988 
1989 		/*
1990 		 * got it!
1991 		 */
1992 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
1993 	}
1994 	simple_unlock(&uvn_wl_lock);
1995 
1996 	/*
1997 	 * step 3: we now have a list of uvn's that may need cleaning.
1998 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
1999 	 * (so we can now safely lock uvn's again).
2000 	 */
2001 
2002 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2003 		simple_lock(&uvn->u_obj.vmobjlock);
2004 #ifdef DIAGNOSTIC
2005 		if (uvn->u_flags & UVM_VNODE_DYING) {
2006 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2007 		}
2008 #endif
2009 		uvn_flush(&uvn->u_obj, 0, 0,
2010 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2011 
2012 		/*
2013 		 * if we have the only reference and we just cleaned the uvn,
2014 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2015 		 * thus allowing us to avoid thinking about flushing it again
2016 		 * on later sync ops.
2017 		 */
2018 		if (uvn->u_obj.uo_refs == 1 &&
2019 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2020 			LIST_REMOVE(uvn, u_wlist);
2021 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2022 		}
2023 
2024 		simple_unlock(&uvn->u_obj.vmobjlock);
2025 
2026 		/* now drop our reference to the uvn */
2027 		uvn_detach(&uvn->u_obj);
2028 	}
2029 
2030 	/*
2031 	 * done!  release sync lock
2032 	 */
2033 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0);
2034 }
2035