xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision dc306354b0b29af51801a7632f1e95265a68cd81)
1 /*	$NetBSD: uvm_vnode.c,v 1.17 1998/11/04 06:21:40 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 maping 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 if (vp->v_type == VBLK) printf("used_vnode_size = %qu\n", used_vnode_size);
296 	if (used_vnode_size > (vaddr_t) -PAGE_SIZE) {
297 #ifdef DEBUG
298 		printf("uvn_attach: vn %p size truncated %qx->%x\n", vp,
299 		    used_vnode_size, -PAGE_SIZE);
300 #endif
301 		used_vnode_size = (vaddr_t) -PAGE_SIZE;
302 	}
303 
304 	/*
305 	 * now set up the uvn.
306 	 */
307 	uvn->u_obj.pgops = &uvm_vnodeops;
308 	TAILQ_INIT(&uvn->u_obj.memq);
309 	uvn->u_obj.uo_npages = 0;
310 	uvn->u_obj.uo_refs = 1;			/* just us... */
311 	oldflags = uvn->u_flags;
312 	uvn->u_flags = UVM_VNODE_VALID|UVM_VNODE_CANPERSIST;
313 	uvn->u_nio = 0;
314 	uvn->u_size = used_vnode_size;
315 
316 	/* if write access, we need to add it to the wlist */
317 	if (accessprot & VM_PROT_WRITE) {
318 		simple_lock(&uvn_wl_lock);
319 		LIST_INSERT_HEAD(&uvn_wlist, uvn, u_wlist);
320 		simple_unlock(&uvn_wl_lock);
321 		uvn->u_flags |= UVM_VNODE_WRITEABLE;	/* we are on wlist! */
322 	}
323 
324 	/*
325 	 * add a reference to the vnode.   this reference will stay as long
326 	 * as there is a valid mapping of the vnode.   dropped when the
327 	 * reference count goes to zero [and we either free or persist].
328 	 */
329 	VREF(vp);
330 	simple_unlock(&uvn->u_obj.vmobjlock);
331 	if (oldflags & UVM_VNODE_WANTED)
332 		wakeup(uvn);
333 
334 	UVMHIST_LOG(maphist,"<- done/VREF, ret 0x%x", &uvn->u_obj,0,0,0);
335 	return(&uvn->u_obj);
336 }
337 
338 
339 /*
340  * uvn_reference
341  *
342  * duplicate a reference to a VM object.  Note that the reference
343  * count must already be at least one (the passed in reference) so
344  * there is no chance of the uvn being killed or locked out here.
345  *
346  * => caller must call with object unlocked.
347  * => caller must be using the same accessprot as was used at attach time
348  */
349 
350 
351 static void
352 uvn_reference(uobj)
353 	struct uvm_object *uobj;
354 {
355 #ifdef DIAGNOSTIC
356 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
357 #endif
358 	UVMHIST_FUNC("uvn_reference"); UVMHIST_CALLED(maphist);
359 
360 	simple_lock(&uobj->vmobjlock);
361 #ifdef DIAGNOSTIC
362 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
363 		printf("uvn_reference: ref=%d, flags=0x%x\n", uvn->u_flags,
364 		    uobj->uo_refs);
365 		panic("uvn_reference: invalid state");
366 	}
367 #endif
368 	uobj->uo_refs++;
369 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
370 	uobj, uobj->uo_refs,0,0);
371 	simple_unlock(&uobj->vmobjlock);
372 }
373 
374 /*
375  * uvn_detach
376  *
377  * remove a reference to a VM object.
378  *
379  * => caller must call with object unlocked and map locked.
380  * => this starts the detach process, but doesn't have to finish it
381  *    (async i/o could still be pending).
382  */
383 static void
384 uvn_detach(uobj)
385 	struct uvm_object *uobj;
386 {
387 	struct uvm_vnode *uvn;
388 	struct vnode *vp;
389 	int oldflags;
390 	UVMHIST_FUNC("uvn_detach"); UVMHIST_CALLED(maphist);
391 
392 	simple_lock(&uobj->vmobjlock);
393 
394 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
395 	uobj->uo_refs--;			/* drop ref! */
396 	if (uobj->uo_refs) {			/* still more refs */
397 		simple_unlock(&uobj->vmobjlock);
398 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
399 		return;
400 	}
401 
402 	/*
403 	 * get other pointers ...
404 	 */
405 
406 	uvn = (struct uvm_vnode *) uobj;
407 	vp = (struct vnode *) uobj;
408 
409 	/*
410 	 * clear VTEXT flag now that there are no mappings left (VTEXT is used
411 	 * to keep an active text file from being overwritten).
412 	 */
413 	vp->v_flag &= ~VTEXT;
414 
415 	/*
416 	 * we just dropped the last reference to the uvn.   see if we can
417 	 * let it "stick around".
418 	 */
419 
420 	if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
421 		/* won't block */
422 		uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
423 		simple_unlock(&uobj->vmobjlock);
424 		vrele(vp);			/* drop vnode reference */
425 		UVMHIST_LOG(maphist,"<- done/vrele!  (persist)", 0,0,0,0);
426 		return;
427 	}
428 
429 	/*
430 	 * its a goner!
431 	 */
432 
433 	UVMHIST_LOG(maphist,"  its a goner (flushing)!", 0,0,0,0);
434 
435 	uvn->u_flags |= UVM_VNODE_DYING;
436 
437 	/*
438 	 * even though we may unlock in flush, no one can gain a reference
439 	 * to us until we clear the "dying" flag [because it blocks
440 	 * attaches].  we will not do that until after we've disposed of all
441 	 * the pages with uvn_flush().  note that before the flush the only
442 	 * pages that could be marked PG_BUSY are ones that are in async
443 	 * pageout by the daemon.  (there can't be any pending "get"'s
444 	 * because there are no references to the object).
445 	 */
446 
447 	(void) uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
448 
449 	UVMHIST_LOG(maphist,"  its a goner (done flush)!", 0,0,0,0);
450 
451 	/*
452 	 * given the structure of this pager, the above flush request will
453 	 * create the following state: all the pages that were in the object
454 	 * have either been free'd or they are marked PG_BUSY|PG_RELEASED.
455 	 * the PG_BUSY bit was set either by us or the daemon for async I/O.
456 	 * in either case, if we have pages left we can't kill the object
457 	 * yet because i/o is pending.  in this case we set the "relkill"
458 	 * flag which will cause pgo_releasepg to kill the object once all
459 	 * the I/O's are done [pgo_releasepg will be called from the aiodone
460 	 * routine or from the page daemon].
461 	 */
462 
463 	if (uobj->uo_npages) {		/* I/O pending.  iodone will free */
464 #ifdef DIAGNOSTIC
465 		/*
466 		 * XXXCDC: very unlikely to happen until we have async i/o
467 		 * so print a little info message in case it does.
468 		 */
469 		printf("uvn_detach: vn %p has pages left after flush - "
470 		    "relkill mode\n", uobj);
471 #endif
472 		uvn->u_flags |= UVM_VNODE_RELKILL;
473 		simple_unlock(&uobj->vmobjlock);
474 		UVMHIST_LOG(maphist,"<- done! (releasepg will kill obj)", 0, 0,
475 		    0, 0);
476 		return;
477 	}
478 
479 	/*
480 	 * kill object now.   note that we can't be on the sync q because
481 	 * all references are gone.
482 	 */
483 	if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
484 		simple_lock(&uvn_wl_lock);		/* protect uvn_wlist */
485 		LIST_REMOVE(uvn, u_wlist);
486 		simple_unlock(&uvn_wl_lock);
487 	}
488 #ifdef DIAGNOSTIC
489 	if (uobj->memq.tqh_first != NULL)
490 		panic("uvn_deref: vnode VM object still has pages afer "
491 		    "syncio/free flush");
492 #endif
493 	oldflags = uvn->u_flags;
494 	uvn->u_flags = 0;
495 	simple_unlock(&uobj->vmobjlock);
496 
497 	/* wake up any sleepers */
498 	if (oldflags & UVM_VNODE_WANTED)
499 		wakeup(uvn);
500 
501 	/*
502 	 * drop our reference to the vnode.
503 	 */
504 	vrele(vp);
505 	UVMHIST_LOG(maphist,"<- done (vrele) final", 0,0,0,0);
506 
507 	return;
508 }
509 
510 /*
511  * uvm_vnp_terminate: external hook to clear out a vnode's VM
512  *
513  * called in two cases:
514  *  [1] when a persisting vnode vm object (i.e. one with a zero reference
515  *      count) needs to be freed so that a vnode can be reused.  this
516  *      happens under "getnewvnode" in vfs_subr.c.   if the vnode from
517  *      the free list is still attached (i.e. not VBAD) then vgone is
518  *	called.   as part of the vgone trace this should get called to
519  *	free the vm object.   this is the common case.
520  *  [2] when a filesystem is being unmounted by force (MNT_FORCE,
521  *	"umount -f") the vgone() function is called on active vnodes
522  *	on the mounted file systems to kill their data (the vnodes become
523  *	"dead" ones [see src/sys/miscfs/deadfs/...]).  that results in a
524  *	call here (even if the uvn is still in use -- i.e. has a non-zero
525  *	reference count).  this case happens at "umount -f" and during a
526  *	"reboot/halt" operation.
527  *
528  * => the caller must XLOCK and VOP_LOCK the vnode before calling us
529  *	[protects us from getting a vnode that is already in the DYING
530  *	 state...]
531  * => unlike uvn_detach, this function must not return until all the
532  *	uvn's pages are disposed of.
533  * => in case [2] the uvn is still alive after this call, but all I/O
534  *	ops will fail (due to the backing vnode now being "dead").  this
535  *	will prob. kill any process using the uvn due to pgo_get failing.
536  */
537 
538 void
539 uvm_vnp_terminate(vp)
540 	struct vnode *vp;
541 {
542 	struct uvm_vnode *uvn = &vp->v_uvm;
543 	int oldflags;
544 	UVMHIST_FUNC("uvm_vnp_terminate"); UVMHIST_CALLED(maphist);
545 
546 	/*
547 	 * lock object and check if it is valid
548 	 */
549 	simple_lock(&uvn->u_obj.vmobjlock);
550 	UVMHIST_LOG(maphist, "  vp=0x%x, ref=%d, flag=0x%x", vp,
551 	    uvn->u_obj.uo_refs, uvn->u_flags, 0);
552 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0) {
553 		simple_unlock(&uvn->u_obj.vmobjlock);
554 		UVMHIST_LOG(maphist, "<- done (not active)", 0, 0, 0, 0);
555 		return;
556 	}
557 
558 	/*
559 	 * must be a valid uvn that is not already dying (because XLOCK
560 	 * protects us from that).   the uvn can't in the the ALOCK state
561 	 * because it is valid, and uvn's that are in the ALOCK state haven't
562 	 * been marked valid yet.
563 	 */
564 
565 #ifdef DEBUG
566 	/*
567 	 * debug check: are we yanking the vnode out from under our uvn?
568 	 */
569 	if (uvn->u_obj.uo_refs) {
570 		printf("uvm_vnp_terminate(%p): terminating active vnode "
571 		    "(refs=%d)\n", uvn, uvn->u_obj.uo_refs);
572 	}
573 #endif
574 
575 	/*
576 	 * it is possible that the uvn was detached and is in the relkill
577 	 * state [i.e. waiting for async i/o to finish so that releasepg can
578 	 * kill object].  we take over the vnode now and cancel the relkill.
579 	 * we want to know when the i/o is done so we can recycle right
580 	 * away.   note that a uvn can only be in the RELKILL state if it
581 	 * has a zero reference count.
582 	 */
583 
584 	if (uvn->u_flags & UVM_VNODE_RELKILL)
585 		uvn->u_flags &= ~UVM_VNODE_RELKILL;	/* cancel RELKILL */
586 
587 	/*
588 	 * block the uvn by setting the dying flag, and then flush the
589 	 * pages.  (note that flush may unlock object while doing I/O, but
590 	 * it will re-lock it before it returns control here).
591 	 *
592 	 * also, note that we tell I/O that we are already VOP_LOCK'd so
593 	 * that uvn_io doesn't attempt to VOP_LOCK again.
594 	 *
595 	 * XXXCDC: setting VNISLOCKED on an active uvn which is being terminated
596 	 *	due to a forceful unmount might not be a good idea.  maybe we
597 	 *	need a way to pass in this info to uvn_flush through a
598 	 *	pager-defined PGO_ constant [currently there are none].
599 	 */
600 	uvn->u_flags |= UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED;
601 
602 	(void) uvn_flush(&uvn->u_obj, 0, 0, PGO_CLEANIT|PGO_FREE|PGO_ALLPAGES);
603 
604 	/*
605 	 * as we just did a flush we expect all the pages to be gone or in
606 	 * the process of going.  sleep to wait for the rest to go [via iosync].
607 	 */
608 
609 	while (uvn->u_obj.uo_npages) {
610 #ifdef DIAGNOSTIC
611 		struct vm_page *pp;
612 		for (pp = uvn->u_obj.memq.tqh_first ; pp != NULL ;
613 		     pp = pp->listq.tqe_next) {
614 			if ((pp->flags & PG_BUSY) == 0)
615 				panic("uvm_vnp_terminate: detected unbusy pg");
616 		}
617 		if (uvn->u_nio == 0)
618 			panic("uvm_vnp_terminate: no I/O to wait for?");
619 		printf("uvm_vnp_terminate: waiting for I/O to fin.\n");
620 		/*
621 		 * XXXCDC: this is unlikely to happen without async i/o so we
622 		 * put a printf in just to keep an eye on it.
623 		 */
624 #endif
625 		uvn->u_flags |= UVM_VNODE_IOSYNC;
626 		UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock, FALSE,
627 		    "uvn_term",0);
628 		simple_lock(&uvn->u_obj.vmobjlock);
629 	}
630 
631 	/*
632 	 * done.   now we free the uvn if its reference count is zero
633 	 * (true if we are zapping a persisting uvn).   however, if we are
634 	 * terminating a uvn with active mappings we let it live ... future
635 	 * calls down to the vnode layer will fail.
636 	 */
637 
638 	oldflags = uvn->u_flags;
639 	if (uvn->u_obj.uo_refs) {
640 
641 		/*
642 		 * uvn must live on it is dead-vnode state until all references
643 		 * are gone.   restore flags.    clear CANPERSIST state.
644 		 */
645 
646 		uvn->u_flags &= ~(UVM_VNODE_DYING|UVM_VNODE_VNISLOCKED|
647 		      UVM_VNODE_WANTED|UVM_VNODE_CANPERSIST);
648 
649 	} else {
650 
651 		/*
652 		 * free the uvn now.   note that the VREF reference is already
653 		 * gone [it is dropped when we enter the persist state].
654 		 */
655 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
656 			panic("uvm_vnp_terminate: io sync wanted bit set");
657 
658 		if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
659 			simple_lock(&uvn_wl_lock);
660 			LIST_REMOVE(uvn, u_wlist);
661 			simple_unlock(&uvn_wl_lock);
662 		}
663 		uvn->u_flags = 0;	/* uvn is history, clear all bits */
664 	}
665 
666 	if (oldflags & UVM_VNODE_WANTED)
667 		wakeup(uvn);		/* object lock still held */
668 
669 	simple_unlock(&uvn->u_obj.vmobjlock);
670 	UVMHIST_LOG(maphist, "<- done", 0, 0, 0, 0);
671 
672 }
673 
674 /*
675  * uvn_releasepg: handled a released page in a uvn
676  *
677  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
678  *	to dispose of.
679  * => caller must handled PG_WANTED case
680  * => called with page's object locked, pageq's unlocked
681  * => returns TRUE if page's object is still alive, FALSE if we
682  *	killed the page's object.    if we return TRUE, then we
683  *	return with the object locked.
684  * => if (nextpgp != NULL) => we return pageq.tqe_next here, and return
685  *				with the page queues locked [for pagedaemon]
686  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
687  * => we kill the uvn if it is not referenced and we are suppose to
688  *	kill it ("relkill").
689  */
690 
691 boolean_t
692 uvn_releasepg(pg, nextpgp)
693 	struct vm_page *pg;
694 	struct vm_page **nextpgp;	/* OUT */
695 {
696 	struct uvm_vnode *uvn = (struct uvm_vnode *) pg->uobject;
697 #ifdef DIAGNOSTIC
698 	if ((pg->flags & PG_RELEASED) == 0)
699 		panic("uvn_releasepg: page not released!");
700 #endif
701 
702 	/*
703 	 * dispose of the page [caller handles PG_WANTED]
704 	 */
705 	pmap_page_protect(PMAP_PGARG(pg), VM_PROT_NONE);
706 	uvm_lock_pageq();
707 	if (nextpgp)
708 		*nextpgp = pg->pageq.tqe_next;	/* next page for daemon */
709 	uvm_pagefree(pg);
710 	if (!nextpgp)
711 		uvm_unlock_pageq();
712 
713 	/*
714 	 * now see if we need to kill the object
715 	 */
716 	if (uvn->u_flags & UVM_VNODE_RELKILL) {
717 		if (uvn->u_obj.uo_refs)
718 			panic("uvn_releasepg: kill flag set on referenced "
719 			    "object!");
720 		if (uvn->u_obj.uo_npages == 0) {
721 			if (uvn->u_flags & UVM_VNODE_WRITEABLE) {
722 				simple_lock(&uvn_wl_lock);
723 				LIST_REMOVE(uvn, u_wlist);
724 				simple_unlock(&uvn_wl_lock);
725 			}
726 #ifdef DIAGNOSTIC
727 			if (uvn->u_obj.memq.tqh_first)
728 	panic("uvn_releasepg: pages in object with npages == 0");
729 #endif
730 			if (uvn->u_flags & UVM_VNODE_WANTED)
731 				/* still holding object lock */
732 				wakeup(uvn);
733 
734 			uvn->u_flags = 0;		/* DEAD! */
735 			simple_unlock(&uvn->u_obj.vmobjlock);
736 			return (FALSE);
737 		}
738 	}
739 	return (TRUE);
740 }
741 
742 /*
743  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
744  * through the buffer cache and allow I/O in any size.  These VOPs use
745  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
746  * go through the buffer cache or allow I/O sizes larger than a
747  * block].  we will eventually want to change this.
748  *
749  * issues to consider:
750  *   uvm provides the uvm_aiodesc structure for async i/o management.
751  * there are two tailq's in the uvm. structure... one for pending async
752  * i/o and one for "done" async i/o.   to do an async i/o one puts
753  * an aiodesc on the "pending" list (protected by splbio()), starts the
754  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
755  * some sort of "i/o done" function to be called (at splbio(), interrupt
756  * time).   this function should remove the aiodesc from the pending list
757  * and place it on the "done" list and wakeup the daemon.   the daemon
758  * will run at normal spl() and will remove all items from the "done"
759  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
760  * [in the old vm code, this was done by calling the "put" routine with
761  * null arguments which made the code harder to read and understand because
762  * you had one function ("put") doing two things.]
763  *
764  * so the current pager needs:
765  *   int uvn_aiodone(struct uvm_aiodesc *)
766  *
767  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
768  *	later collection.
769  * => called with pageq's locked by the daemon.
770  *
771  * general outline:
772  * - "try" to lock object.   if fail, just return (will try again later)
773  * - drop "u_nio" (this req is done!)
774  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
775  * - get "page" structures (atop?).
776  * - handle "wanted" pages
777  * - handle "released" pages [using pgo_releasepg]
778  *   >>> pgo_releasepg may kill the object
779  * dont forget to look at "object" wanted flag in all cases.
780  */
781 
782 
783 /*
784  * uvn_flush: flush pages out of a uvm object.
785  *
786  * => object should be locked by caller.   we may _unlock_ the object
787  *	if (and only if) we need to clean a page (PGO_CLEANIT).
788  *	we return with the object locked.
789  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
790  *	might want to unlock higher level resources (e.g. vm_map)
791  *	before calling flush.
792  * => if PGO_CLEANIT is not set, then we will neither unlock the object
793  *	or block.
794  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
795  *	for flushing.
796  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
797  *	that new pages are inserted on the tail end of the list.   thus,
798  *	we can make a complete pass through the object in one go by starting
799  *	at the head and working towards the tail (new pages are put in
800  *	front of us).
801  * => NOTE: we are allowed to lock the page queues, so the caller
802  *	must not be holding the lock on them [e.g. pagedaemon had
803  *	better not call us with the queues locked]
804  * => we return TRUE unless we encountered some sort of I/O error
805  *
806  * comment on "cleaning" object and PG_BUSY pages:
807  *	this routine is holding the lock on the object.   the only time
808  *	that it can run into a PG_BUSY page that it does not own is if
809  *	some other process has started I/O on the page (e.g. either
810  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
811  *	in, then it can not be dirty (!PG_CLEAN) because no one has
812  *	had a chance to modify it yet.    if the PG_BUSY page is being
813  *	paged out then it means that someone else has already started
814  *	cleaning the page for us (how nice!).    in this case, if we
815  *	have syncio specified, then after we make our pass through the
816  *	object we need to wait for the other PG_BUSY pages to clear
817  *	off (i.e. we need to do an iosync).   also note that once a
818  *	page is PG_BUSY it must stay in its object until it is un-busyed.
819  *
820  * note on page traversal:
821  *	we can traverse the pages in an object either by going down the
822  *	linked list in "uobj->memq", or we can go over the address range
823  *	by page doing hash table lookups for each address.    depending
824  *	on how many pages are in the object it may be cheaper to do one
825  *	or the other.   we set "by_list" to true if we are using memq.
826  *	if the cost of a hash lookup was equal to the cost of the list
827  *	traversal we could compare the number of pages in the start->stop
828  *	range to the total number of pages in the object.   however, it
829  *	seems that a hash table lookup is more expensive than the linked
830  *	list traversal, so we multiply the number of pages in the
831  *	start->stop range by a penalty which we define below.
832  */
833 
834 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
835 
836 static boolean_t
837 uvn_flush(uobj, start, stop, flags)
838 	struct uvm_object *uobj;
839 	vaddr_t start, stop;
840 	int flags;
841 {
842 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
843 	struct vm_page *pp, *ppnext, *ptmp;
844 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
845 	int npages, result, lcv;
846 	boolean_t retval, need_iosync, by_list, needs_clean;
847 	vaddr_t curoff;
848 	u_short pp_version;
849 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
850 
851 	curoff = 0;	/* XXX: shut up gcc */
852 	/*
853 	 * get init vals and determine how we are going to traverse object
854 	 */
855 
856 	need_iosync = FALSE;
857 	retval = TRUE;		/* return value */
858 	if (flags & PGO_ALLPAGES) {
859 		start = 0;
860 		stop = round_page(uvn->u_size);
861 		by_list = TRUE;		/* always go by the list */
862 	} else {
863 		start = trunc_page(start);
864 		stop = round_page(stop);
865 		if (stop > round_page(uvn->u_size))
866 			printf("uvn_flush: strange, got an out of range "
867 			    "flush (fixed)\n");
868 
869 		by_list = (uobj->uo_npages <=
870 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
871 	}
872 
873 	UVMHIST_LOG(maphist,
874 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
875 	    start, stop, by_list, flags);
876 
877 	/*
878 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
879 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
880 	 * is wrong it will only prevent us from clustering... it won't break
881 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
882 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
883 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
884 	 * bit is always up to date since there are no mappings).
885 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
886 	 */
887 
888 	if ((flags & PGO_CLEANIT) != 0 &&
889 	    uobj->pgops->pgo_mk_pcluster != NULL) {
890 		if (by_list) {
891 			for (pp = uobj->memq.tqh_first ; pp != NULL ;
892 			    pp = pp->listq.tqe_next) {
893 				if (pp->offset < start || pp->offset >= stop)
894 					continue;
895 				pp->flags &= ~PG_CLEANCHK;
896 			}
897 
898 		} else {   /* by hash */
899 			for (curoff = start ; curoff < stop;
900 			    curoff += PAGE_SIZE) {
901 				pp = uvm_pagelookup(uobj, curoff);
902 				if (pp)
903 					pp->flags &= ~PG_CLEANCHK;
904 			}
905 		}
906 	}
907 
908 	/*
909 	 * now do it.   note: we must update ppnext in body of loop or we
910 	 * will get stuck.  we need to use ppnext because we may free "pp"
911 	 * before doing the next loop.
912 	 */
913 
914 	if (by_list) {
915 		pp = uobj->memq.tqh_first;
916 	} else {
917 		curoff = start;
918 		pp = uvm_pagelookup(uobj, curoff);
919 	}
920 
921 	ppnext = NULL;	/* XXX: shut up gcc */
922 	ppsp = NULL;		/* XXX: shut up gcc */
923 	uvm_lock_pageq();	/* page queues locked */
924 
925 	/* locked: both page queues and uobj */
926 	for ( ; (by_list && pp != NULL) ||
927 	  (!by_list && curoff < stop) ; pp = ppnext) {
928 
929 		if (by_list) {
930 
931 			/*
932 			 * range check
933 			 */
934 
935 			if (pp->offset < start || pp->offset >= stop) {
936 				ppnext = pp->listq.tqe_next;
937 				continue;
938 			}
939 
940 		} else {
941 
942 			/*
943 			 * null check
944 			 */
945 
946 			curoff += PAGE_SIZE;
947 			if (pp == NULL) {
948 				if (curoff < stop)
949 					ppnext = uvm_pagelookup(uobj, curoff);
950 				continue;
951 			}
952 
953 		}
954 
955 		/*
956 		 * handle case where we do not need to clean page (either
957 		 * because we are not clean or because page is not dirty or
958 		 * is busy):
959 		 *
960 		 * NOTE: we are allowed to deactivate a non-wired active
961 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
962 		 * queue it must stay put until it is !PG_BUSY (so as not to
963 		 * confuse pagedaemon).
964 		 */
965 
966 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
967 			needs_clean = FALSE;
968 			if ((pp->flags & PG_BUSY) != 0 &&
969 			    (flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
970 			             (PGO_CLEANIT|PGO_SYNCIO))
971 				need_iosync = TRUE;
972 		} else {
973 			/*
974 			 * freeing: nuke all mappings so we can sync
975 			 * PG_CLEAN bit with no race
976 			 */
977 			if ((pp->flags & PG_CLEAN) != 0 &&
978 			    (flags & PGO_FREE) != 0 &&
979 			    (pp->pqflags & PQ_ACTIVE) != 0)
980 				pmap_page_protect(PMAP_PGARG(pp), VM_PROT_NONE);
981 			if ((pp->flags & PG_CLEAN) != 0 &&
982 			    pmap_is_modified(PMAP_PGARG(pp)))
983 				pp->flags &= ~(PG_CLEAN);
984 			pp->flags |= PG_CLEANCHK;	/* update "hint" */
985 
986 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
987 		}
988 
989 		/*
990 		 * if we don't need a clean... load ppnext and dispose of pp
991 		 */
992 		if (!needs_clean) {
993 			/* load ppnext */
994 			if (by_list)
995 				ppnext = pp->listq.tqe_next;
996 			else {
997 				if (curoff < stop)
998 					ppnext = uvm_pagelookup(uobj, curoff);
999 			}
1000 
1001 			/* now dispose of pp */
1002 			if (flags & PGO_DEACTIVATE) {
1003 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1004 				    pp->wire_count == 0) {
1005 					pmap_page_protect(PMAP_PGARG(pp),
1006 					    VM_PROT_NONE);
1007 					uvm_pagedeactivate(pp);
1008 				}
1009 
1010 			} else if (flags & PGO_FREE) {
1011 				if (pp->flags & PG_BUSY) {
1012 					/* release busy pages */
1013 					pp->flags |= PG_RELEASED;
1014 				} else {
1015 					pmap_page_protect(PMAP_PGARG(pp),
1016 					    VM_PROT_NONE);
1017 					/* removed page from object */
1018 					uvm_pagefree(pp);
1019 				}
1020 			}
1021 			/* ppnext is valid so we can continue... */
1022 			continue;
1023 		}
1024 
1025 		/*
1026 		 * pp points to a page in the locked object that we are
1027 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
1028 		 * for cleaning (PGO_CLEANIT).  we clean it now.
1029 		 *
1030 		 * let uvm_pager_put attempted a clustered page out.
1031 		 * note: locked: uobj and page queues.
1032 		 */
1033 
1034 		pp->flags |= PG_BUSY;	/* we 'own' page now */
1035 		UVM_PAGE_OWN(pp, "uvn_flush");
1036 		pmap_page_protect(PMAP_PGARG(pp), VM_PROT_READ);
1037 		pp_version = pp->version;
1038 ReTry:
1039 		ppsp = pps;
1040 		npages = sizeof(pps) / sizeof(struct vm_page *);
1041 
1042 		/* locked: page queues, uobj */
1043 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
1044 			   flags | PGO_DOACTCLUST, start, stop);
1045 		/* unlocked: page queues, uobj */
1046 
1047 		/*
1048 		 * at this point nothing is locked.   if we did an async I/O
1049 		 * it is remotely possible for the async i/o to complete and
1050 		 * the page "pp" be freed or what not before we get a chance
1051 		 * to relock the object.   in order to detect this, we have
1052 		 * saved the version number of the page in "pp_version".
1053 		 */
1054 
1055 		/* relock! */
1056 		simple_lock(&uobj->vmobjlock);
1057 		uvm_lock_pageq();
1058 
1059 		/*
1060 		 * VM_PAGER_AGAIN: given the structure of this pager, this
1061 		 * can only happen when  we are doing async I/O and can't
1062 		 * map the pages into kernel memory (pager_map) due to lack
1063 		 * of vm space.   if this happens we drop back to sync I/O.
1064 		 */
1065 
1066 		if (result == VM_PAGER_AGAIN) {
1067 			/*
1068 			 * it is unlikely, but page could have been released
1069 			 * while we had the object lock dropped.   we ignore
1070 			 * this now and retry the I/O.  we will detect and
1071 			 * handle the released page after the syncio I/O
1072 			 * completes.
1073 			 */
1074 #ifdef DIAGNOSTIC
1075 			if (flags & PGO_SYNCIO)
1076 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
1077 #endif
1078 			flags |= PGO_SYNCIO;
1079 			goto ReTry;
1080 		}
1081 
1082 		/*
1083 		 * the cleaning operation is now done.   finish up.  note that
1084 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
1085 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
1086 		 * to us in ppsp/npages.
1087 		 */
1088 
1089 		/*
1090 		 * for pending async i/o if we are not deactivating/freeing
1091 		 * we can move on to the next page.
1092 		 */
1093 
1094 		if (result == VM_PAGER_PEND) {
1095 
1096 			if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
1097 				/*
1098 				 * no per-page ops: refresh ppnext and continue
1099 				 */
1100 				if (by_list) {
1101 					if (pp->version == pp_version)
1102 						ppnext = pp->listq.tqe_next;
1103 					else
1104 						/* reset */
1105 						ppnext = uobj->memq.tqh_first;
1106 				} else {
1107 					if (curoff < stop)
1108 						ppnext = uvm_pagelookup(uobj,
1109 						    curoff);
1110 				}
1111 				continue;
1112 			}
1113 
1114 			/* need to do anything here? */
1115 		}
1116 
1117 		/*
1118 		 * need to look at each page of the I/O operation.  we defer
1119 		 * processing "pp" until the last trip through this "for" loop
1120 		 * so that we can load "ppnext" for the main loop after we
1121 		 * play with the cluster pages [thus the "npages + 1" in the
1122 		 * loop below].
1123 		 */
1124 
1125 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
1126 
1127 			/*
1128 			 * handle ppnext for outside loop, and saving pp
1129 			 * until the end.
1130 			 */
1131 			if (lcv < npages) {
1132 				if (ppsp[lcv] == pp)
1133 					continue; /* skip pp until the end */
1134 				ptmp = ppsp[lcv];
1135 			} else {
1136 				ptmp = pp;
1137 
1138 				/* set up next page for outer loop */
1139 				if (by_list) {
1140 					if (pp->version == pp_version)
1141 						ppnext = pp->listq.tqe_next;
1142 					else
1143 						/* reset */
1144 						ppnext = uobj->memq.tqh_first;
1145 				} else {
1146 					if (curoff < stop)
1147 					ppnext = uvm_pagelookup(uobj, curoff);
1148 				}
1149 			}
1150 
1151 			/*
1152 			 * verify the page didn't get moved while obj was
1153 			 * unlocked
1154 			 */
1155 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
1156 				continue;
1157 
1158 			/*
1159 			 * unbusy the page if I/O is done.   note that for
1160 			 * pending I/O it is possible that the I/O op
1161 			 * finished before we relocked the object (in
1162 			 * which case the page is no longer busy).
1163 			 */
1164 
1165 			if (result != VM_PAGER_PEND) {
1166 				if (ptmp->flags & PG_WANTED)
1167 					/* still holding object lock */
1168 					thread_wakeup(ptmp);
1169 
1170 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1171 				UVM_PAGE_OWN(ptmp, NULL);
1172 				if (ptmp->flags & PG_RELEASED) {
1173 
1174 					/* pgo_releasepg wants this */
1175 					uvm_unlock_pageq();
1176 					if (!uvn_releasepg(ptmp, NULL))
1177 						return (TRUE);
1178 
1179 					uvm_lock_pageq();	/* relock */
1180 					continue;		/* next page */
1181 
1182 				} else {
1183 					ptmp->flags |= (PG_CLEAN|PG_CLEANCHK);
1184 					if ((flags & PGO_FREE) == 0)
1185 						pmap_clear_modify(
1186 						    PMAP_PGARG(ptmp));
1187 				}
1188 			}
1189 
1190 			/*
1191 			 * dispose of page
1192 			 */
1193 
1194 			if (flags & PGO_DEACTIVATE) {
1195 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
1196 				    pp->wire_count == 0) {
1197 					pmap_page_protect(PMAP_PGARG(ptmp),
1198 					    VM_PROT_NONE);
1199 					uvm_pagedeactivate(ptmp);
1200 				}
1201 
1202 			} else if (flags & PGO_FREE) {
1203 				if (result == VM_PAGER_PEND) {
1204 					if ((ptmp->flags & PG_BUSY) != 0)
1205 						/* signal for i/o done */
1206 						ptmp->flags |= PG_RELEASED;
1207 				} else {
1208 					if (result != VM_PAGER_OK) {
1209 						printf("uvn_flush: obj=%p, "
1210 						   "offset=0x%lx.  error "
1211 						   "during pageout.\n",
1212 						    pp->uobject, pp->offset);
1213 						printf("uvn_flush: WARNING: "
1214 						    "changes to page may be "
1215 						    "lost!\n");
1216 						retval = FALSE;
1217 					}
1218 					pmap_page_protect(PMAP_PGARG(ptmp),
1219 					    VM_PROT_NONE);
1220 					uvm_pagefree(ptmp);
1221 				}
1222 			}
1223 
1224 		}		/* end of "lcv" for loop */
1225 
1226 	}		/* end of "pp" for loop */
1227 
1228 	/*
1229 	 * done with pagequeues: unlock
1230 	 */
1231 	uvm_unlock_pageq();
1232 
1233 	/*
1234 	 * now wait for all I/O if required.
1235 	 */
1236 	if (need_iosync) {
1237 
1238 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
1239 		while (uvn->u_nio != 0) {
1240 			uvn->u_flags |= UVM_VNODE_IOSYNC;
1241 			UVM_UNLOCK_AND_WAIT(&uvn->u_nio, &uvn->u_obj.vmobjlock,
1242 			  FALSE, "uvn_flush",0);
1243 			simple_lock(&uvn->u_obj.vmobjlock);
1244 		}
1245 		if (uvn->u_flags & UVM_VNODE_IOSYNCWANTED)
1246 			wakeup(&uvn->u_flags);
1247 		uvn->u_flags &= ~(UVM_VNODE_IOSYNC|UVM_VNODE_IOSYNCWANTED);
1248 	}
1249 
1250 	/* return, with object locked! */
1251 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
1252 	return(retval);
1253 }
1254 
1255 /*
1256  * uvn_cluster
1257  *
1258  * we are about to do I/O in an object at offset.   this function is called
1259  * to establish a range of offsets around "offset" in which we can cluster
1260  * I/O.
1261  *
1262  * - currently doesn't matter if obj locked or not.
1263  */
1264 
1265 static void
1266 uvn_cluster(uobj, offset, loffset, hoffset)
1267 	struct uvm_object *uobj;
1268 	vaddr_t offset;
1269 	vaddr_t *loffset, *hoffset; /* OUT */
1270 {
1271 	struct uvm_vnode *uvn = (struct uvm_vnode *) uobj;
1272 	*loffset = offset;
1273 
1274 	if (*loffset >= uvn->u_size)
1275 		panic("uvn_cluster: offset out of range");
1276 
1277 	/*
1278 	 * XXX: old pager claims we could use VOP_BMAP to get maxcontig value.
1279 	 */
1280 	*hoffset = *loffset + MAXBSIZE;
1281 	if (*hoffset > round_page(uvn->u_size))	/* past end? */
1282 		*hoffset = round_page(uvn->u_size);
1283 
1284 	return;
1285 }
1286 
1287 /*
1288  * uvn_put: flush page data to backing store.
1289  *
1290  * => prefer map unlocked (not required)
1291  * => object must be locked!   we will _unlock_ it before starting I/O.
1292  * => flags: PGO_SYNCIO -- use sync. I/O
1293  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
1294  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1295  *	[thus we never do async i/o!  see iodone comment]
1296  */
1297 
1298 static int
1299 uvn_put(uobj, pps, npages, flags)
1300 	struct uvm_object *uobj;
1301 	struct vm_page **pps;
1302 	int npages, flags;
1303 {
1304 	int retval;
1305 
1306 	/* note: object locked */
1307 	retval = uvn_io((struct uvm_vnode*)uobj, pps, npages, flags, UIO_WRITE);
1308 	/* note: object unlocked */
1309 
1310 	return(retval);
1311 }
1312 
1313 
1314 /*
1315  * uvn_get: get pages (synchronously) from backing store
1316  *
1317  * => prefer map unlocked (not required)
1318  * => object must be locked!  we will _unlock_ it before starting any I/O.
1319  * => flags: PGO_ALLPAGES: get all of the pages
1320  *           PGO_LOCKED: fault data structures are locked
1321  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1322  * => NOTE: caller must check for released pages!!
1323  */
1324 
1325 static int
1326 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
1327 	struct uvm_object *uobj;
1328 	vaddr_t offset;
1329 	struct vm_page **pps;		/* IN/OUT */
1330 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
1331 	int centeridx, advice, flags;
1332 	vm_prot_t access_type;
1333 {
1334 	vaddr_t current_offset;
1335 	struct vm_page *ptmp;
1336 	int lcv, result, gotpages;
1337 	boolean_t done;
1338 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(maphist);
1339 	UVMHIST_LOG(maphist, "flags=%d", flags,0,0,0);
1340 
1341 	/*
1342 	 * step 1: handled the case where fault data structures are locked.
1343 	 */
1344 
1345 	if (flags & PGO_LOCKED) {
1346 
1347 		/*
1348 		 * gotpages is the current number of pages we've gotten (which
1349 		 * we pass back up to caller via *npagesp.
1350 		 */
1351 
1352 		gotpages = 0;
1353 
1354 		/*
1355 		 * step 1a: get pages that are already resident.   only do this
1356 		 * if the data structures are locked (i.e. the first time
1357 		 * through).
1358 		 */
1359 
1360 		done = TRUE;	/* be optimistic */
1361 
1362 		for (lcv = 0, current_offset = offset ; lcv < *npagesp ;
1363 		    lcv++, current_offset += PAGE_SIZE) {
1364 
1365 			/* do we care about this page?  if not, skip it */
1366 			if (pps[lcv] == PGO_DONTCARE)
1367 				continue;
1368 
1369 			/* lookup page */
1370 			ptmp = uvm_pagelookup(uobj, current_offset);
1371 
1372 			/* to be useful must get a non-busy, non-released pg */
1373 			if (ptmp == NULL ||
1374 			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1375 				if (lcv == centeridx || (flags & PGO_ALLPAGES)
1376 				    != 0)
1377 				done = FALSE;	/* need to do a wait or I/O! */
1378 				continue;
1379 			}
1380 
1381 			/*
1382 			 * useful page: busy/lock it and plug it in our
1383 			 * result array
1384 			 */
1385 			ptmp->flags |= PG_BUSY;		/* loan up to caller */
1386 			UVM_PAGE_OWN(ptmp, "uvn_get1");
1387 			pps[lcv] = ptmp;
1388 			gotpages++;
1389 
1390 		}	/* "for" lcv loop */
1391 
1392 		/*
1393 		 * XXX: given the "advice", should we consider async read-ahead?
1394 		 * XXX: fault current does deactive of pages behind us.  is
1395 		 * this good (other callers might now).
1396 		 */
1397 		/*
1398 		 * XXX: read-ahead currently handled by buffer cache (bread)
1399 		 * level.
1400 		 * XXX: no async i/o available.
1401 		 * XXX: so we don't do anything now.
1402 		 */
1403 
1404 		/*
1405 		 * step 1c: now we've either done everything needed or we to
1406 		 * unlock and do some waiting or I/O.
1407 		 */
1408 
1409 		*npagesp = gotpages;		/* let caller know */
1410 		if (done)
1411 			return(VM_PAGER_OK);		/* bingo! */
1412 		else
1413 			/* EEK!   Need to unlock and I/O */
1414 			return(VM_PAGER_UNLOCK);
1415 	}
1416 
1417 	/*
1418 	 * step 2: get non-resident or busy pages.
1419 	 * object is locked.   data structures are unlocked.
1420 	 *
1421 	 * XXX: because we can't do async I/O at this level we get things
1422 	 * page at a time (otherwise we'd chunk).   the VOP_READ() will do
1423 	 * async-read-ahead for us at a lower level.
1424 	 */
1425 
1426 	for (lcv = 0, current_offset = offset ;
1427 			 lcv < *npagesp ; lcv++, current_offset += PAGE_SIZE) {
1428 
1429 		/* skip over pages we've already gotten or don't want */
1430 		/* skip over pages we don't _have_ to get */
1431 		if (pps[lcv] != NULL || (lcv != centeridx &&
1432 		    (flags & PGO_ALLPAGES) == 0))
1433 			continue;
1434 
1435 		/*
1436 		 * we have yet to locate the current page (pps[lcv]).   we first
1437 		 * look for a page that is already at the current offset.   if
1438 		 * we fine a page, we check to see if it is busy or released.
1439 		 * if that is the case, then we sleep on the page until it is
1440 		 * no longer busy or released and repeat the lookup.    if the
1441 		 * page we found is neither busy nor released, then we busy it
1442 		 * (so we own it) and plug it into pps[lcv].   this breaks the
1443 		 * following while loop and indicates we are ready to move on
1444 		 * to the next page in the "lcv" loop above.
1445 		 *
1446 		 * if we exit the while loop with pps[lcv] still set to NULL,
1447 		 * then it means that we allocated a new busy/fake/clean page
1448 		 * ptmp in the object and we need to do I/O to fill in the data.
1449 		 */
1450 
1451 		while (pps[lcv] == NULL) {	/* top of "pps" while loop */
1452 
1453 			/* look for a current page */
1454 			ptmp = uvm_pagelookup(uobj, current_offset);
1455 
1456 			/* nope?   allocate one now (if we can) */
1457 			if (ptmp == NULL) {
1458 
1459 				ptmp = uvm_pagealloc(uobj, current_offset,
1460 				    NULL);	/* alloc */
1461 
1462 				/* out of RAM? */
1463 				if (ptmp == NULL) {
1464 					simple_unlock(&uobj->vmobjlock);
1465 					uvm_wait("uvn_getpage");
1466 					simple_lock(&uobj->vmobjlock);
1467 
1468 					/* goto top of pps while loop */
1469 					continue;
1470 				}
1471 
1472 				/*
1473 				 * got new page ready for I/O.  break pps
1474 				 * while loop.  pps[lcv] is still NULL.
1475 				 */
1476 				break;
1477 			}
1478 
1479 			/* page is there, see if we need to wait on it */
1480 			if ((ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
1481 				ptmp->flags |= PG_WANTED;
1482 				UVM_UNLOCK_AND_WAIT(ptmp,
1483 				    &uobj->vmobjlock, 0, "uvn_get",0);
1484 				simple_lock(&uobj->vmobjlock);
1485 				continue;	/* goto top of pps while loop */
1486 			}
1487 
1488 			/*
1489 			 * if we get here then the page has become resident
1490 			 * and unbusy between steps 1 and 2.  we busy it
1491 			 * now (so we own it) and set pps[lcv] (so that we
1492 			 * exit the while loop).
1493 			 */
1494 			ptmp->flags |= PG_BUSY;
1495 			UVM_PAGE_OWN(ptmp, "uvn_get2");
1496 			pps[lcv] = ptmp;
1497 		}
1498 
1499 		/*
1500 		 * if we own the a valid page at the correct offset, pps[lcv]
1501 		 * will point to it.   nothing more to do except go to the
1502 		 * next page.
1503 		 */
1504 
1505 		if (pps[lcv])
1506 			continue;			/* next lcv */
1507 
1508 		/*
1509 		 * we have a "fake/busy/clean" page that we just allocated.  do
1510 		 * I/O to fill it with valid data.  note that object must be
1511 		 * locked going into uvn_io, but will be unlocked afterwards.
1512 		 */
1513 
1514 		result = uvn_io((struct uvm_vnode *) uobj, &ptmp, 1,
1515 		    PGO_SYNCIO, UIO_READ);
1516 
1517 		/*
1518 		 * I/O done.   object is unlocked (by uvn_io).   because we used
1519 		 * syncio the result can not be PEND or AGAIN.   we must relock
1520 		 * and check for errors.
1521 		 */
1522 
1523 		/* lock object.   check for errors.   */
1524 		simple_lock(&uobj->vmobjlock);
1525 		if (result != VM_PAGER_OK) {
1526 			if (ptmp->flags & PG_WANTED)
1527 				/* object lock still held */
1528 				thread_wakeup(ptmp);
1529 
1530 			ptmp->flags &= ~(PG_WANTED|PG_BUSY);
1531 			UVM_PAGE_OWN(ptmp, NULL);
1532 			uvm_lock_pageq();
1533 			uvm_pagefree(ptmp);
1534 			uvm_unlock_pageq();
1535 			simple_unlock(&uobj->vmobjlock);
1536 			return(result);
1537 		}
1538 
1539 		/*
1540 		 * we got the page!   clear the fake flag (indicates valid
1541 		 * data now in page) and plug into our result array.   note
1542 		 * that page is still busy.
1543 		 *
1544 		 * it is the callers job to:
1545 		 * => check if the page is released
1546 		 * => unbusy the page
1547 		 * => activate the page
1548 		 */
1549 
1550 		ptmp->flags &= ~PG_FAKE;		/* data is valid ... */
1551 		pmap_clear_modify(PMAP_PGARG(ptmp));	/* ... and clean */
1552 		pps[lcv] = ptmp;
1553 
1554 	}	/* lcv loop */
1555 
1556 	/*
1557 	 * finally, unlock object and return.
1558 	 */
1559 
1560 	simple_unlock(&uobj->vmobjlock);
1561 	return (VM_PAGER_OK);
1562 }
1563 
1564 /*
1565  * uvn_asyncget: start async I/O to bring pages into ram
1566  *
1567  * => caller must lock object(???XXX: see if this is best)
1568  * => could be called from uvn_get or a madvise() fault-ahead.
1569  * => if it fails, it doesn't matter.
1570  */
1571 
1572 static int
1573 uvn_asyncget(uobj, offset, npages)
1574 	struct uvm_object *uobj;
1575 	vaddr_t offset;
1576 	int npages;
1577 {
1578 
1579 	/*
1580 	 * XXXCDC: we can't do async I/O yet
1581 	 */
1582 	printf("uvn_asyncget called\n");
1583 	return (KERN_SUCCESS);
1584 }
1585 
1586 /*
1587  * uvn_io: do I/O to a vnode
1588  *
1589  * => prefer map unlocked (not required)
1590  * => object must be locked!   we will _unlock_ it before starting I/O.
1591  * => flags: PGO_SYNCIO -- use sync. I/O
1592  * => XXX: currently we use VOP_READ/VOP_WRITE which are only sync.
1593  *	[thus we never do async i/o!  see iodone comment]
1594  */
1595 
1596 static int
1597 uvn_io(uvn, pps, npages, flags, rw)
1598 	struct uvm_vnode *uvn;
1599 	vm_page_t *pps;
1600 	int npages, flags, rw;
1601 {
1602 	struct vnode *vn;
1603 	struct uio uio;
1604 	struct iovec iov;
1605 	vaddr_t kva, file_offset;
1606 	int waitf, result, got, wanted;
1607 	UVMHIST_FUNC("uvn_io"); UVMHIST_CALLED(maphist);
1608 
1609 	UVMHIST_LOG(maphist, "rw=%d", rw,0,0,0);
1610 
1611 	/*
1612 	 * init values
1613 	 */
1614 
1615 	waitf = (flags & PGO_SYNCIO) ? M_WAITOK : M_NOWAIT;
1616 	vn = (struct vnode *) uvn;
1617 	file_offset = pps[0]->offset;
1618 
1619 	/*
1620 	 * check for sync'ing I/O.
1621 	 */
1622 
1623 	while (uvn->u_flags & UVM_VNODE_IOSYNC) {
1624 		if (waitf == M_NOWAIT) {
1625 			simple_unlock(&uvn->u_obj.vmobjlock);
1626 			UVMHIST_LOG(maphist,"<- try again (iosync)",0,0,0,0);
1627 			return(VM_PAGER_AGAIN);
1628 		}
1629 		uvn->u_flags |= UVM_VNODE_IOSYNCWANTED;
1630 		UVM_UNLOCK_AND_WAIT(&uvn->u_flags, &uvn->u_obj.vmobjlock,
1631 			FALSE, "uvn_iosync",0);
1632 		simple_lock(&uvn->u_obj.vmobjlock);
1633 	}
1634 
1635 	/*
1636 	 * check size
1637 	 */
1638 
1639 	if (file_offset >= uvn->u_size) {
1640 			simple_unlock(&uvn->u_obj.vmobjlock);
1641 			UVMHIST_LOG(maphist,"<- BAD (size check)",0,0,0,0);
1642 #ifdef DIAGNOSTIC
1643 			printf("uvn_io: note: size check fired\n");
1644 #endif
1645 			return(VM_PAGER_BAD);
1646 	}
1647 
1648 	/*
1649 	 * first try and map the pages in (without waiting)
1650 	 */
1651 
1652 	kva = uvm_pagermapin(pps, npages, NULL, M_NOWAIT);
1653 	if (kva == NULL && waitf == M_NOWAIT) {
1654 		simple_unlock(&uvn->u_obj.vmobjlock);
1655 		UVMHIST_LOG(maphist,"<- mapin failed (try again)",0,0,0,0);
1656 		return(VM_PAGER_AGAIN);
1657 	}
1658 
1659 	/*
1660 	 * ok, now bump u_nio up.   at this point we are done with uvn
1661 	 * and can unlock it.   if we still don't have a kva, try again
1662 	 * (this time with sleep ok).
1663 	 */
1664 
1665 	uvn->u_nio++;			/* we have an I/O in progress! */
1666 	simple_unlock(&uvn->u_obj.vmobjlock);
1667 	/* NOTE: object now unlocked */
1668 	if (kva == NULL) {
1669 		kva = uvm_pagermapin(pps, npages, NULL, M_WAITOK);
1670 	}
1671 
1672 	/*
1673 	 * ok, mapped in.  our pages are PG_BUSY so they are not going to
1674 	 * get touched (so we can look at "offset" without having to lock
1675 	 * the object).  set up for I/O.
1676 	 */
1677 
1678 	/*
1679 	 * fill out uio/iov
1680 	 */
1681 
1682 	iov.iov_base = (caddr_t) kva;
1683 	wanted = npages << PAGE_SHIFT;
1684 	if (file_offset + wanted > uvn->u_size)
1685 		wanted = uvn->u_size - file_offset;	/* XXX: needed? */
1686 	iov.iov_len = wanted;
1687 	uio.uio_iov = &iov;
1688 	uio.uio_iovcnt = 1;
1689 	uio.uio_offset = file_offset;
1690 	uio.uio_segflg = UIO_SYSSPACE;
1691 	uio.uio_rw = rw;
1692 	uio.uio_resid = wanted;
1693 	uio.uio_procp = NULL;
1694 
1695 	/*
1696 	 * do the I/O!  (XXX: curproc?)
1697 	 */
1698 
1699 	UVMHIST_LOG(maphist, "calling VOP",0,0,0,0);
1700 
1701 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1702 		vn_lock(vn, LK_EXCLUSIVE | LK_RETRY);
1703 	/* NOTE: vnode now locked! */
1704 
1705 	if (rw == UIO_READ)
1706 		result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
1707 	else
1708 		result = VOP_WRITE(vn, &uio, 0, curproc->p_ucred);
1709 
1710 	if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
1711 		VOP_UNLOCK(vn, 0);
1712 	/* NOTE: vnode now unlocked (unless vnislocked) */
1713 
1714 	UVMHIST_LOG(maphist, "done calling VOP",0,0,0,0);
1715 
1716 	/*
1717 	 * result == unix style errno (0 == OK!)
1718 	 *
1719 	 * zero out rest of buffer (if needed)
1720 	 */
1721 
1722 	if (result == 0) {
1723 		got = wanted - uio.uio_resid;
1724 
1725 		if (wanted && got == 0) {
1726 			result = EIO;		/* XXX: error? */
1727 		} else if (got < PAGE_SIZE * npages && rw == UIO_READ) {
1728 			memset((void *) (kva + got), 0,
1729 			       (npages << PAGE_SHIFT) - got);
1730 		}
1731 	}
1732 
1733 	/*
1734 	 * now remove pager mapping
1735 	 */
1736 	uvm_pagermapout(kva, npages);
1737 
1738 	/*
1739 	 * now clean up the object (i.e. drop I/O count)
1740 	 */
1741 
1742 	simple_lock(&uvn->u_obj.vmobjlock);
1743 	/* NOTE: object now locked! */
1744 
1745 	uvn->u_nio--;			/* I/O DONE! */
1746 	if ((uvn->u_flags & UVM_VNODE_IOSYNC) != 0 && uvn->u_nio == 0) {
1747 		wakeup(&uvn->u_nio);
1748 	}
1749 	simple_unlock(&uvn->u_obj.vmobjlock);
1750 	/* NOTE: object now unlocked! */
1751 
1752 	/*
1753 	 * done!
1754 	 */
1755 
1756 	UVMHIST_LOG(maphist, "<- done (result %d)", result,0,0,0);
1757 	if (result == 0)
1758 		return(VM_PAGER_OK);
1759 	else
1760 		return(VM_PAGER_ERROR);
1761 }
1762 
1763 /*
1764  * uvm_vnp_uncache: disable "persisting" in a vnode... when last reference
1765  * is gone we will kill the object (flushing dirty pages back to the vnode
1766  * if needed).
1767  *
1768  * => returns TRUE if there was no uvm_object attached or if there was
1769  *	one and we killed it [i.e. if there is no active uvn]
1770  * => called with the vnode VOP_LOCK'd [we will unlock it for I/O, if
1771  *	needed]
1772  *
1773  * => XXX: given that we now kill uvn's when a vnode is recycled (without
1774  *	having to hold a reference on the vnode) and given a working
1775  *	uvm_vnp_sync(), how does that effect the need for this function?
1776  *      [XXXCDC: seems like it can die?]
1777  *
1778  * => XXX: this function should DIE once we merge the VM and buffer
1779  *	cache.
1780  *
1781  * research shows that this is called in the following places:
1782  * ext2fs_truncate, ffs_truncate, detrunc[msdosfs]: called when vnode
1783  *	changes sizes
1784  * ext2fs_write, WRITE [ufs_readwrite], msdosfs_write: called when we
1785  *	are written to
1786  * ex2fs_chmod, ufs_chmod: called if VTEXT vnode and the sticky bit
1787  *	is off
1788  * ffs_realloccg: when we can't extend the current block and have
1789  *	to allocate a new one we call this [XXX: why?]
1790  * nfsrv_rename, rename_files: called when the target filename is there
1791  *	and we want to remove it
1792  * nfsrv_remove, sys_unlink: called on file we are removing
1793  * nfsrv_access: if VTEXT and we want WRITE access and we don't uncache
1794  *	then return "text busy"
1795  * nfs_open: seems to uncache any file opened with nfs
1796  * vn_writechk: if VTEXT vnode and can't uncache return "text busy"
1797  */
1798 
1799 boolean_t
1800 uvm_vnp_uncache(vp)
1801 	struct vnode *vp;
1802 {
1803 	struct uvm_vnode *uvn = &vp->v_uvm;
1804 
1805 	/*
1806 	 * lock uvn part of the vnode and check to see if we need to do anything
1807 	 */
1808 
1809 	simple_lock(&uvn->u_obj.vmobjlock);
1810 	if ((uvn->u_flags & UVM_VNODE_VALID) == 0 ||
1811 			(uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
1812 		simple_unlock(&uvn->u_obj.vmobjlock);
1813 		return(TRUE);
1814 	}
1815 
1816 	/*
1817 	 * we have a valid, non-blocked uvn.   clear persist flag.
1818 	 * if uvn is currently active we can return now.
1819 	 */
1820 
1821 	uvn->u_flags &= ~UVM_VNODE_CANPERSIST;
1822 	if (uvn->u_obj.uo_refs) {
1823 		simple_unlock(&uvn->u_obj.vmobjlock);
1824 		return(FALSE);
1825 	}
1826 
1827 	/*
1828 	 * uvn is currently persisting!   we have to gain a reference to
1829 	 * it so that we can call uvn_detach to kill the uvn.
1830 	 */
1831 
1832 	VREF(vp);			/* seems ok, even with VOP_LOCK */
1833 	uvn->u_obj.uo_refs++;		/* value is now 1 */
1834 	simple_unlock(&uvn->u_obj.vmobjlock);
1835 
1836 
1837 #ifdef DEBUG
1838 	/*
1839 	 * carry over sanity check from old vnode pager: the vnode should
1840 	 * be VOP_LOCK'd, and we confirm it here.
1841 	 */
1842 	if (!VOP_ISLOCKED(vp)) {
1843 		boolean_t is_ok_anyway = FALSE;
1844 #ifdef NFS
1845 		extern int (**nfsv2_vnodeop_p) __P((void *));
1846 		extern int (**spec_nfsv2nodeop_p) __P((void *));
1847 		extern int (**fifo_nfsv2nodeop_p) __P((void *));
1848 
1849 		/* vnode is NOT VOP_LOCKed: some vnode types _never_ lock */
1850 		if (vp->v_op == nfsv2_vnodeop_p ||
1851 		    vp->v_op == spec_nfsv2nodeop_p) {
1852 			is_ok_anyway = TRUE;
1853 		}
1854 		if (vp->v_op == fifo_nfsv2nodeop_p) {
1855 			is_ok_anyway = TRUE;
1856 		}
1857 #endif	/* NFS */
1858 		if (!is_ok_anyway)
1859 			panic("uvm_vnp_uncache: vnode not locked!");
1860 	}
1861 #endif	/* DEBUG */
1862 
1863 	/*
1864 	 * now drop our reference to the vnode.   if we have the sole
1865 	 * reference to the vnode then this will cause it to die [as we
1866 	 * just cleared the persist flag].   we have to unlock the vnode
1867 	 * while we are doing this as it may trigger I/O.
1868 	 *
1869 	 * XXX: it might be possible for uvn to get reclaimed while we are
1870 	 * unlocked causing us to return TRUE when we should not.   we ignore
1871 	 * this as a false-positive return value doesn't hurt us.
1872 	 */
1873 	VOP_UNLOCK(vp, 0);
1874 	uvn_detach(&uvn->u_obj);
1875 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1876 
1877 	/*
1878 	 * and return...
1879 	 */
1880 
1881 	return(TRUE);
1882 }
1883 
1884 /*
1885  * uvm_vnp_setsize: grow or shrink a vnode uvn
1886  *
1887  * grow   => just update size value
1888  * shrink => toss un-needed pages
1889  *
1890  * => we assume that the caller has a reference of some sort to the
1891  *	vnode in question so that it will not be yanked out from under
1892  *	us.
1893  *
1894  * called from:
1895  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1896  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1897  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1898  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1899  *  => union fs: union_newsize
1900  */
1901 
1902 void
1903 uvm_vnp_setsize(vp, newsize)
1904 	struct vnode *vp;
1905 	u_quad_t newsize;
1906 {
1907 	struct uvm_vnode *uvn = &vp->v_uvm;
1908 
1909 	/*
1910 	 * lock uvn and check for valid object, and if valid: do it!
1911 	 */
1912 	simple_lock(&uvn->u_obj.vmobjlock);
1913 	if (uvn->u_flags & UVM_VNODE_VALID) {
1914 
1915 		/*
1916 		 * make sure that the newsize fits within a vaddr_t
1917 		 * XXX: need to revise addressing data types
1918 		 */
1919 
1920 		if (newsize > (vaddr_t) -PAGE_SIZE) {
1921 #ifdef DEBUG
1922 			printf("uvm_vnp_setsize: vn %p size truncated "
1923 			    "%qx->%lx\n", vp, newsize, (vaddr_t)-PAGE_SIZE);
1924 #endif
1925 			newsize = (vaddr_t)-PAGE_SIZE;
1926 		}
1927 
1928 		/*
1929 		 * now check if the size has changed: if we shrink we had better
1930 		 * toss some pages...
1931 		 */
1932 
1933 		if (uvn->u_size > newsize) {
1934 			(void)uvn_flush(&uvn->u_obj, (vaddr_t) newsize,
1935 			    uvn->u_size, PGO_FREE);
1936 		}
1937 		uvn->u_size = (vaddr_t)newsize;
1938 	}
1939 	simple_unlock(&uvn->u_obj.vmobjlock);
1940 
1941 	/*
1942 	 * done
1943 	 */
1944 	return;
1945 }
1946 
1947 /*
1948  * uvm_vnp_sync: flush all dirty VM pages back to their backing vnodes.
1949  *
1950  * => called from sys_sync with no VM structures locked
1951  * => only one process can do a sync at a time (because the uvn
1952  *    structure only has one queue for sync'ing).  we ensure this
1953  *    by holding the uvn_sync_lock while the sync is in progress.
1954  *    other processes attempting a sync will sleep on this lock
1955  *    until we are done.
1956  */
1957 
1958 void
1959 uvm_vnp_sync(mp)
1960 	struct mount *mp;
1961 {
1962 	struct uvm_vnode *uvn;
1963 	struct vnode *vp;
1964 	boolean_t got_lock;
1965 
1966 	/*
1967 	 * step 1: ensure we are only ones using the uvn_sync_q by locking
1968 	 * our lock...
1969 	 */
1970 	lockmgr(&uvn_sync_lock, LK_EXCLUSIVE, (void *)0);
1971 
1972 	/*
1973 	 * step 2: build up a simpleq of uvns of interest based on the
1974 	 * write list.   we gain a reference to uvns of interest.  must
1975 	 * be careful about locking uvn's since we will be holding uvn_wl_lock
1976 	 * in the body of the loop.
1977 	 */
1978 	SIMPLEQ_INIT(&uvn_sync_q);
1979 	simple_lock(&uvn_wl_lock);
1980 	for (uvn = uvn_wlist.lh_first ; uvn != NULL ;
1981 	    uvn = uvn->u_wlist.le_next) {
1982 
1983 		vp = (struct vnode *) uvn;
1984 		if (mp && vp->v_mount != mp)
1985 			continue;
1986 
1987 		/* attempt to gain reference */
1988 		while ((got_lock = simple_lock_try(&uvn->u_obj.vmobjlock)) ==
1989 		    						FALSE &&
1990 				(uvn->u_flags & UVM_VNODE_BLOCKED) == 0)
1991 			/* spin */ ;
1992 
1993 		/*
1994 		 * we will exit the loop if either if the following are true:
1995 		 *  - we got the lock [always true if NCPU == 1]
1996 		 *  - we failed to get the lock but noticed the vnode was
1997 		 * 	"blocked" -- in this case the vnode must be a dying
1998 		 *	vnode, and since dying vnodes are in the process of
1999 		 *	being flushed out, we can safely skip this one
2000 		 *
2001 		 * we want to skip over the vnode if we did not get the lock,
2002 		 * or if the vnode is already dying (due to the above logic).
2003 		 *
2004 		 * note that uvn must already be valid because we found it on
2005 		 * the wlist (this also means it can't be ALOCK'd).
2006 		 */
2007 		if (!got_lock || (uvn->u_flags & UVM_VNODE_BLOCKED) != 0) {
2008 			if (got_lock)
2009 				simple_unlock(&uvn->u_obj.vmobjlock);
2010 			continue;		/* skip it */
2011 		}
2012 
2013 		/*
2014 		 * gain reference.   watch out for persisting uvns (need to
2015 		 * regain vnode REF).
2016 		 */
2017 		if (uvn->u_obj.uo_refs == 0)
2018 			VREF(vp);
2019 		uvn->u_obj.uo_refs++;
2020 		simple_unlock(&uvn->u_obj.vmobjlock);
2021 
2022 		/*
2023 		 * got it!
2024 		 */
2025 		SIMPLEQ_INSERT_HEAD(&uvn_sync_q, uvn, u_syncq);
2026 	}
2027 	simple_unlock(&uvn_wl_lock);
2028 
2029 	/*
2030 	 * step 3: we now have a list of uvn's that may need cleaning.
2031 	 * we are holding the uvn_sync_lock, but have dropped the uvn_wl_lock
2032 	 * (so we can now safely lock uvn's again).
2033 	 */
2034 
2035 	for (uvn = uvn_sync_q.sqh_first ; uvn ; uvn = uvn->u_syncq.sqe_next) {
2036 		simple_lock(&uvn->u_obj.vmobjlock);
2037 #ifdef DIAGNOSTIC
2038 		if (uvn->u_flags & UVM_VNODE_DYING) {
2039 			printf("uvm_vnp_sync: dying vnode on sync list\n");
2040 		}
2041 #endif
2042 		uvn_flush(&uvn->u_obj, 0, 0,
2043 		    PGO_CLEANIT|PGO_ALLPAGES|PGO_DOACTCLUST);
2044 
2045 		/*
2046 		 * if we have the only reference and we just cleaned the uvn,
2047 		 * then we can pull it out of the UVM_VNODE_WRITEABLE state
2048 		 * thus allowing us to avoid thinking about flushing it again
2049 		 * on later sync ops.
2050 		 */
2051 		if (uvn->u_obj.uo_refs == 1 &&
2052 		    (uvn->u_flags & UVM_VNODE_WRITEABLE)) {
2053 			LIST_REMOVE(uvn, u_wlist);
2054 			uvn->u_flags &= ~UVM_VNODE_WRITEABLE;
2055 		}
2056 
2057 		simple_unlock(&uvn->u_obj.vmobjlock);
2058 
2059 		/* now drop our reference to the uvn */
2060 		uvn_detach(&uvn->u_obj);
2061 	}
2062 
2063 	/*
2064 	 * done!  release sync lock
2065 	 */
2066 	lockmgr(&uvn_sync_lock, LK_RELEASE, (void *)0);
2067 }
2068