xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: uvm_vnode.c,v 1.50 2001/05/26 21:27:21 chs Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993
6  *      The Regents of the University of California.
7  * Copyright (c) 1990 University of Utah.
8  *
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * the Systems Programming Group of the University of Utah Computer
13  * Science Department.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. All advertising materials mentioning features or use of this software
24  *    must display the following acknowledgement:
25  *      This product includes software developed by Charles D. Cranor,
26  *	Washington University, the University of California, Berkeley and
27  *	its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  *
44  *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
45  * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
46  */
47 
48 #include "fs_nfs.h"
49 #include "opt_uvmhist.h"
50 #include "opt_ddb.h"
51 
52 /*
53  * uvm_vnode.c: the vnode pager.
54  */
55 
56 #include <sys/param.h>
57 #include <sys/systm.h>
58 #include <sys/kernel.h>
59 #include <sys/proc.h>
60 #include <sys/malloc.h>
61 #include <sys/vnode.h>
62 #include <sys/disklabel.h>
63 #include <sys/ioctl.h>
64 #include <sys/fcntl.h>
65 #include <sys/conf.h>
66 #include <sys/pool.h>
67 #include <sys/mount.h>
68 
69 #include <miscfs/specfs/specdev.h>
70 
71 #include <uvm/uvm.h>
72 #include <uvm/uvm_vnode.h>
73 
74 /*
75  * functions
76  */
77 
78 static void		uvn_cluster __P((struct uvm_object *, voff_t, voff_t *,
79 					 voff_t *));
80 static void		uvn_detach __P((struct uvm_object *));
81 static int		uvn_findpage __P((struct uvm_object *, voff_t,
82 					  struct vm_page **, int));
83 static boolean_t	uvn_flush __P((struct uvm_object *, voff_t, voff_t,
84 				       int));
85 static int		uvn_get __P((struct uvm_object *, voff_t,
86 				     struct vm_page **, int *, int, vm_prot_t,
87 				     int, int));
88 static int		uvn_put __P((struct uvm_object *, struct vm_page **,
89 				     int, boolean_t));
90 static void		uvn_reference __P((struct uvm_object *));
91 static boolean_t	uvn_releasepg __P((struct vm_page *,
92 					   struct vm_page **));
93 
94 /*
95  * master pager structure
96  */
97 
98 struct uvm_pagerops uvm_vnodeops = {
99 	NULL,
100 	uvn_reference,
101 	uvn_detach,
102 	NULL,
103 	uvn_flush,
104 	uvn_get,
105 	uvn_put,
106 	uvn_cluster,
107 	uvm_mk_pcluster,
108 	uvn_releasepg,
109 };
110 
111 /*
112  * the ops!
113  */
114 
115 /*
116  * uvn_attach
117  *
118  * attach a vnode structure to a VM object.  if the vnode is already
119  * attached, then just bump the reference count by one and return the
120  * VM object.   if not already attached, attach and return the new VM obj.
121  * the "accessprot" tells the max access the attaching thread wants to
122  * our pages.
123  *
124  * => caller must _not_ already be holding the lock on the uvm_object.
125  * => in fact, nothing should be locked so that we can sleep here.
126  * => note that uvm_object is first thing in vnode structure, so their
127  *    pointers are equiv.
128  */
129 
130 struct uvm_object *
131 uvn_attach(arg, accessprot)
132 	void *arg;
133 	vm_prot_t accessprot;
134 {
135 	struct vnode *vp = arg;
136 	struct uvm_vnode *uvn = &vp->v_uvm;
137 	struct vattr vattr;
138 	int result;
139 	struct partinfo pi;
140 	voff_t used_vnode_size;
141 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
142 
143 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
144 	used_vnode_size = (voff_t)0;
145 
146 	/*
147 	 * first get a lock on the uvn.
148 	 */
149 	simple_lock(&uvn->u_obj.vmobjlock);
150 	while (uvn->u_flags & VXLOCK) {
151 		uvn->u_flags |= VXWANT;
152 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
153 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
154 		    "uvn_attach", 0);
155 		simple_lock(&uvn->u_obj.vmobjlock);
156 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
157 	}
158 
159 	/*
160 	 * if we're mapping a BLK device, make sure it is a disk.
161 	 */
162 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
163 		simple_unlock(&uvn->u_obj.vmobjlock);
164 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
165 		return(NULL);
166 	}
167 
168 #ifdef DIAGNOSTIC
169 	if (vp->v_type != VREG) {
170 		panic("uvn_attach: vp %p not VREG", vp);
171 	}
172 #endif
173 
174 	/*
175 	 * set up our idea of the size
176 	 * if this hasn't been done already.
177 	 */
178 	if (uvn->u_size == VSIZENOTSET) {
179 
180 	uvn->u_flags |= VXLOCK;
181 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
182 		/* XXX: curproc? */
183 	if (vp->v_type == VBLK) {
184 		/*
185 		 * We could implement this as a specfs getattr call, but:
186 		 *
187 		 *	(1) VOP_GETATTR() would get the file system
188 		 *	    vnode operation, not the specfs operation.
189 		 *
190 		 *	(2) All we want is the size, anyhow.
191 		 */
192 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
193 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
194 		if (result == 0) {
195 			/* XXX should remember blocksize */
196 			used_vnode_size = (voff_t)pi.disklab->d_secsize *
197 			    (voff_t)pi.part->p_size;
198 		}
199 	} else {
200 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
201 		if (result == 0)
202 			used_vnode_size = vattr.va_size;
203 	}
204 
205 	/* relock object */
206 	simple_lock(&uvn->u_obj.vmobjlock);
207 
208 	if (uvn->u_flags & VXWANT)
209 		wakeup(uvn);
210 	uvn->u_flags &= ~(VXLOCK|VXWANT);
211 
212 	if (result != 0) {
213 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
214 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
215 		return(NULL);
216 	}
217 	uvn->u_size = used_vnode_size;
218 
219 	}
220 
221 	/* unlock and return */
222 	simple_unlock(&uvn->u_obj.vmobjlock);
223 	UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
224 	    0, 0, 0);
225 	return (&uvn->u_obj);
226 }
227 
228 
229 /*
230  * uvn_reference
231  *
232  * duplicate a reference to a VM object.  Note that the reference
233  * count must already be at least one (the passed in reference) so
234  * there is no chance of the uvn being killed or locked out here.
235  *
236  * => caller must call with object unlocked.
237  * => caller must be using the same accessprot as was used at attach time
238  */
239 
240 
241 static void
242 uvn_reference(uobj)
243 	struct uvm_object *uobj;
244 {
245 	VREF((struct vnode *)uobj);
246 }
247 
248 /*
249  * uvn_detach
250  *
251  * remove a reference to a VM object.
252  *
253  * => caller must call with object unlocked and map locked.
254  */
255 static void
256 uvn_detach(uobj)
257 	struct uvm_object *uobj;
258 {
259 	vrele((struct vnode *)uobj);
260 }
261 
262 /*
263  * uvn_releasepg: handled a released page in a uvn
264  *
265  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
266  *	to dispose of.
267  * => caller must handled PG_WANTED case
268  * => called with page's object locked, pageq's unlocked
269  * => returns TRUE if page's object is still alive, FALSE if we
270  *	killed the page's object.    if we return TRUE, then we
271  *	return with the object locked.
272  * => if (nextpgp != NULL) => we return the next page on the queue, and return
273  *				with the page queues locked [for pagedaemon]
274  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
275  * => we kill the uvn if it is not referenced and we are suppose to
276  *	kill it ("relkill").
277  */
278 
279 boolean_t
280 uvn_releasepg(pg, nextpgp)
281 	struct vm_page *pg;
282 	struct vm_page **nextpgp;	/* OUT */
283 {
284 	KASSERT(pg->flags & PG_RELEASED);
285 
286 	/*
287 	 * dispose of the page [caller handles PG_WANTED]
288 	 */
289 	pmap_page_protect(pg, VM_PROT_NONE);
290 	uvm_lock_pageq();
291 	if (nextpgp)
292 		*nextpgp = TAILQ_NEXT(pg, pageq);
293 	uvm_pagefree(pg);
294 	if (!nextpgp)
295 		uvm_unlock_pageq();
296 
297 	return (TRUE);
298 }
299 
300 /*
301  * issues to consider:
302  * there are two tailq's in the uvm. structure... one for pending async
303  * i/o and one for "done" async i/o.   to do an async i/o one puts
304  * a buf on the "pending" list (protected by splbio()), starts the
305  * i/o and returns 0.    when the i/o is done, we expect
306  * some sort of "i/o done" function to be called (at splbio(), interrupt
307  * time).   this function should remove the buf from the pending list
308  * and place it on the "done" list and wakeup the daemon.   the daemon
309  * will run at normal spl() and will remove all items from the "done"
310  * list and call the iodone hook for each done request (see uvm_pager.c).
311  *
312  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
313  *	later collection.
314  * => called with pageq's locked by the daemon.
315  *
316  * general outline:
317  * - "try" to lock object.   if fail, just return (will try again later)
318  * - drop "u_nio" (this req is done!)
319  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
320  * - get "page" structures (atop?).
321  * - handle "wanted" pages
322  * - handle "released" pages [using pgo_releasepg]
323  *   >>> pgo_releasepg may kill the object
324  * dont forget to look at "object" wanted flag in all cases.
325  */
326 
327 
328 /*
329  * uvn_flush: flush pages out of a uvm object.
330  *
331  * => "stop == 0" means flush all pages at or after "start".
332  * => object should be locked by caller.   we may _unlock_ the object
333  *	if (and only if) we need to clean a page (PGO_CLEANIT), or
334  *	if PGO_SYNCIO is set and there are pages busy.
335  *	we return with the object locked.
336  * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
337  *	thus, a caller might want to unlock higher level resources
338  *	(e.g. vm_map) before calling flush.
339  * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, then we will neither
340  *	unlock the object nor block.
341  * => if PGO_ALLPAGES is set, then all pages in the object are valid targets
342  *	for flushing.
343  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
344  *	that new pages are inserted on the tail end of the list.   thus,
345  *	we can make a complete pass through the object in one go by starting
346  *	at the head and working towards the tail (new pages are put in
347  *	front of us).
348  * => NOTE: we are allowed to lock the page queues, so the caller
349  *	must not be holding the lock on them [e.g. pagedaemon had
350  *	better not call us with the queues locked]
351  * => we return TRUE unless we encountered some sort of I/O error
352  *
353  * comment on "cleaning" object and PG_BUSY pages:
354  *	this routine is holding the lock on the object.   the only time
355  *	that it can run into a PG_BUSY page that it does not own is if
356  *	some other process has started I/O on the page (e.g. either
357  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
358  *	in, then it can not be dirty (!PG_CLEAN) because no one has
359  *	had a chance to modify it yet.    if the PG_BUSY page is being
360  *	paged out then it means that someone else has already started
361  *	cleaning the page for us (how nice!).    in this case, if we
362  *	have syncio specified, then after we make our pass through the
363  *	object we need to wait for the other PG_BUSY pages to clear
364  *	off (i.e. we need to do an iosync).   also note that once a
365  *	page is PG_BUSY it must stay in its object until it is un-busyed.
366  *
367  * note on page traversal:
368  *	we can traverse the pages in an object either by going down the
369  *	linked list in "uobj->memq", or we can go over the address range
370  *	by page doing hash table lookups for each address.    depending
371  *	on how many pages are in the object it may be cheaper to do one
372  *	or the other.   we set "by_list" to true if we are using memq.
373  *	if the cost of a hash lookup was equal to the cost of the list
374  *	traversal we could compare the number of pages in the start->stop
375  *	range to the total number of pages in the object.   however, it
376  *	seems that a hash table lookup is more expensive than the linked
377  *	list traversal, so we multiply the number of pages in the
378  *	start->stop range by a penalty which we define below.
379  */
380 
381 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
382 
383 static boolean_t
384 uvn_flush(uobj, start, stop, flags)
385 	struct uvm_object *uobj;
386 	voff_t start, stop;
387 	int flags;
388 {
389 	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;
390 	struct vnode *vp = (struct vnode *)uobj;
391 	struct vm_page *pp, *ppnext, *ptmp;
392 	struct vm_page *pps[256], **ppsp;
393 	int s;
394 	int npages, result, lcv;
395 	boolean_t retval, need_iosync, by_list, needs_clean, all, wasclean;
396 	boolean_t async = (flags & PGO_SYNCIO) == 0;
397 	voff_t curoff;
398 	u_short pp_version;
399 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
400 	UVMHIST_LOG(maphist, "uobj %p start 0x%x stop 0x%x flags 0x%x",
401 		    uobj, start, stop, flags);
402 	KASSERT(flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
403 
404 	if (uobj->uo_npages == 0) {
405 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL &&
406 		    (vp->v_flag & VONWORKLST)) {
407 			vp->v_flag &= ~VONWORKLST;
408 			LIST_REMOVE(vp, v_synclist);
409 		}
410 		return TRUE;
411 	}
412 
413 #ifdef DEBUG
414 	if (uvn->u_size == VSIZENOTSET) {
415 		printf("uvn_flush: size not set vp %p\n", uvn);
416 		vprint("uvn_flush VSIZENOTSET", vp);
417 		flags |= PGO_ALLPAGES;
418 	}
419 #endif
420 
421 	/*
422 	 * get init vals and determine how we are going to traverse object
423 	 */
424 
425 	if (stop == 0) {
426 		stop = trunc_page(LLONG_MAX);
427 	}
428 	curoff = 0;
429 	need_iosync = FALSE;
430 	retval = TRUE;
431 	wasclean = TRUE;
432 	if (flags & PGO_ALLPAGES) {
433 		all = TRUE;
434 		by_list = TRUE;
435 	} else {
436 		start = trunc_page(start);
437 		stop = round_page(stop);
438 		all = FALSE;
439 		by_list = (uobj->uo_npages <=
440 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
441 	}
442 
443 	UVMHIST_LOG(maphist,
444 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
445 	    start, stop, by_list, flags);
446 
447 	/*
448 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
449 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
450 	 * is wrong it will only prevent us from clustering... it won't break
451 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
452 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
453 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
454 	 * bit is always up to date since there are no mappings).
455 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
456 	 */
457 
458 	if ((flags & PGO_CLEANIT) != 0 &&
459 	    uobj->pgops->pgo_mk_pcluster != NULL) {
460 		if (by_list) {
461 			TAILQ_FOREACH(pp, &uobj->memq, listq) {
462 				if (!all &&
463 				    (pp->offset < start || pp->offset >= stop))
464 					continue;
465 				pp->flags &= ~PG_CLEANCHK;
466 			}
467 
468 		} else {   /* by hash */
469 			for (curoff = start ; curoff < stop;
470 			    curoff += PAGE_SIZE) {
471 				pp = uvm_pagelookup(uobj, curoff);
472 				if (pp)
473 					pp->flags &= ~PG_CLEANCHK;
474 			}
475 		}
476 	}
477 
478 	/*
479 	 * now do it.   note: we must update ppnext in body of loop or we
480 	 * will get stuck.  we need to use ppnext because we may free "pp"
481 	 * before doing the next loop.
482 	 */
483 
484 	if (by_list) {
485 		pp = TAILQ_FIRST(&uobj->memq);
486 	} else {
487 		curoff = start;
488 		pp = uvm_pagelookup(uobj, curoff);
489 	}
490 
491 	ppnext = NULL;
492 	ppsp = NULL;
493 	uvm_lock_pageq();
494 
495 	/* locked: both page queues and uobj */
496 	for ( ; (by_list && pp != NULL) ||
497 		      (!by_list && curoff < stop) ; pp = ppnext) {
498 		if (by_list) {
499 			if (!all &&
500 			    (pp->offset < start || pp->offset >= stop)) {
501 				ppnext = TAILQ_NEXT(pp, listq);
502 				continue;
503 			}
504 		} else {
505 			curoff += PAGE_SIZE;
506 			if (pp == NULL) {
507 				if (curoff < stop)
508 					ppnext = uvm_pagelookup(uobj, curoff);
509 				continue;
510 			}
511 		}
512 
513 		/*
514 		 * handle case where we do not need to clean page (either
515 		 * because we are not clean or because page is not dirty or
516 		 * is busy):
517 		 *
518 		 * NOTE: we are allowed to deactivate a non-wired active
519 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
520 		 * queue it must stay put until it is !PG_BUSY (so as not to
521 		 * confuse pagedaemon).
522 		 */
523 
524 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
525 			needs_clean = FALSE;
526 			if (!async)
527 				need_iosync = TRUE;
528 		} else {
529 
530 			/*
531 			 * freeing: nuke all mappings so we can sync
532 			 * PG_CLEAN bit with no race
533 			 */
534 			if ((pp->flags & PG_CLEAN) != 0 &&
535 			    (flags & PGO_FREE) != 0 &&
536 			    /* XXX ACTIVE|INACTIVE test unnecessary? */
537 			    (pp->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) != 0)
538 				pmap_page_protect(pp, VM_PROT_NONE);
539 			if ((pp->flags & PG_CLEAN) != 0 &&
540 			    pmap_is_modified(pp))
541 				pp->flags &= ~(PG_CLEAN);
542 			pp->flags |= PG_CLEANCHK;
543 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
544 		}
545 
546 		/*
547 		 * if we don't need a clean... load ppnext and dispose of pp
548 		 */
549 		if (!needs_clean) {
550 			if (by_list)
551 				ppnext = TAILQ_NEXT(pp, listq);
552 			else {
553 				if (curoff < stop)
554 					ppnext = uvm_pagelookup(uobj, curoff);
555 			}
556 
557 			if (flags & PGO_DEACTIVATE) {
558 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
559 				    (pp->flags & PG_BUSY) == 0 &&
560 				    pp->wire_count == 0) {
561 					pmap_clear_reference(pp);
562 					uvm_pagedeactivate(pp);
563 				}
564 
565 			} else if (flags & PGO_FREE) {
566 				if (pp->flags & PG_BUSY) {
567 					pp->flags |= PG_RELEASED;
568 				} else {
569 					pmap_page_protect(pp, VM_PROT_NONE);
570 					uvm_pagefree(pp);
571 				}
572 			}
573 			/* ppnext is valid so we can continue... */
574 			continue;
575 		}
576 
577 		/*
578 		 * pp points to a page in the locked object that we are
579 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
580 		 * for cleaning (PGO_CLEANIT).  we clean it now.
581 		 *
582 		 * let uvm_pager_put attempted a clustered page out.
583 		 * note: locked: uobj and page queues.
584 		 */
585 
586 		wasclean = FALSE;
587 		pp->flags |= PG_BUSY;	/* we 'own' page now */
588 		UVM_PAGE_OWN(pp, "uvn_flush");
589 		pmap_page_protect(pp, VM_PROT_READ);
590 		pp_version = pp->version;
591 		ppsp = pps;
592 		npages = sizeof(pps) / sizeof(struct vm_page *);
593 
594 		/* locked: page queues, uobj */
595 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
596 				       flags | PGO_DOACTCLUST, start, stop);
597 		/* unlocked: page queues, uobj */
598 
599 		/*
600 		 * at this point nothing is locked.   if we did an async I/O
601 		 * it is remotely possible for the async i/o to complete and
602 		 * the page "pp" be freed or what not before we get a chance
603 		 * to relock the object.   in order to detect this, we have
604 		 * saved the version number of the page in "pp_version".
605 		 */
606 
607 		/* relock! */
608 		simple_lock(&uobj->vmobjlock);
609 		uvm_lock_pageq();
610 
611 		/*
612 		 * the cleaning operation is now done.  finish up.  note that
613 		 * on error uvm_pager_put drops the cluster for us.
614 		 * on success uvm_pager_put returns the cluster to us in
615 		 * ppsp/npages.
616 		 */
617 
618 		/*
619 		 * for pending async i/o if we are not deactivating/freeing
620 		 * we can move on to the next page.
621 		 */
622 
623 		if (result == 0 && async &&
624 		    (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
625 
626 			/*
627 			 * no per-page ops: refresh ppnext and continue
628 			 */
629 			if (by_list) {
630 				if (pp->version == pp_version)
631 					ppnext = TAILQ_NEXT(pp, listq);
632 				else
633 					ppnext = TAILQ_FIRST(&uobj->memq);
634 			} else {
635 				if (curoff < stop)
636 					ppnext = uvm_pagelookup(uobj, curoff);
637 			}
638 			continue;
639 		}
640 
641 		/*
642 		 * need to look at each page of the I/O operation.  we defer
643 		 * processing "pp" until the last trip through this "for" loop
644 		 * so that we can load "ppnext" for the main loop after we
645 		 * play with the cluster pages [thus the "npages + 1" in the
646 		 * loop below].
647 		 */
648 
649 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
650 
651 			/*
652 			 * handle ppnext for outside loop, and saving pp
653 			 * until the end.
654 			 */
655 			if (lcv < npages) {
656 				if (ppsp[lcv] == pp)
657 					continue; /* skip pp until the end */
658 				ptmp = ppsp[lcv];
659 			} else {
660 				ptmp = pp;
661 
662 				/* set up next page for outer loop */
663 				if (by_list) {
664 					if (pp->version == pp_version)
665 						ppnext = TAILQ_NEXT(pp, listq);
666 					else
667 						ppnext = TAILQ_FIRST(
668 						    &uobj->memq);
669 				} else {
670 					if (curoff < stop)
671 						ppnext = uvm_pagelookup(uobj,
672 						    curoff);
673 				}
674 			}
675 
676 			/*
677 			 * verify the page wasn't moved while obj was
678 			 * unlocked
679 			 */
680 			if (result == 0 && async && ptmp->uobject != uobj)
681 				continue;
682 
683 			/*
684 			 * unbusy the page if I/O is done.   note that for
685 			 * async I/O it is possible that the I/O op
686 			 * finished before we relocked the object (in
687 			 * which case the page is no longer busy).
688 			 */
689 
690 			if (result != 0 || !async) {
691 				if (ptmp->flags & PG_WANTED) {
692 					/* still holding object lock */
693 					wakeup(ptmp);
694 				}
695 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
696 				UVM_PAGE_OWN(ptmp, NULL);
697 				if (ptmp->flags & PG_RELEASED) {
698 					uvm_unlock_pageq();
699 					if (!uvn_releasepg(ptmp, NULL)) {
700 						UVMHIST_LOG(maphist,
701 							    "released %p",
702 							    ptmp, 0,0,0);
703 						return (TRUE);
704 					}
705 					uvm_lock_pageq();
706 					continue;
707 				} else {
708 					if ((flags & PGO_WEAK) == 0 &&
709 					    !(result == EIO &&
710 					      curproc == uvm.pagedaemon_proc)) {
711 						ptmp->flags |=
712 							(PG_CLEAN|PG_CLEANCHK);
713 						if ((flags & PGO_FREE) == 0) {
714 							pmap_clear_modify(ptmp);
715 						}
716 					}
717 				}
718 			}
719 
720 			/*
721 			 * dispose of page
722 			 */
723 
724 			if (flags & PGO_DEACTIVATE) {
725 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
726 				    (pp->flags & PG_BUSY) == 0 &&
727 				    pp->wire_count == 0) {
728 					pmap_clear_reference(ptmp);
729 					uvm_pagedeactivate(ptmp);
730 				}
731 			} else if (flags & PGO_FREE) {
732 				if (result == 0 && async) {
733 					if ((ptmp->flags & PG_BUSY) != 0)
734 						/* signal for i/o done */
735 						ptmp->flags |= PG_RELEASED;
736 				} else {
737 					if (result != 0) {
738 						printf("uvn_flush: obj=%p, "
739 						   "offset=0x%llx.  error %d\n",
740 						    pp->uobject,
741 						    (long long)pp->offset,
742 						    result);
743 						printf("uvn_flush: WARNING: "
744 						    "changes to page may be "
745 						    "lost!\n");
746 						retval = FALSE;
747 					}
748 					pmap_page_protect(ptmp, VM_PROT_NONE);
749 					uvm_pagefree(ptmp);
750 				}
751 			}
752 		}		/* end of "lcv" for loop */
753 	}		/* end of "pp" for loop */
754 
755 	uvm_unlock_pageq();
756 	if ((flags & PGO_CLEANIT) && all && wasclean &&
757 	    LIST_FIRST(&vp->v_dirtyblkhd) == NULL &&
758 	    (vp->v_flag & VONWORKLST)) {
759 		vp->v_flag &= ~VONWORKLST;
760 		LIST_REMOVE(vp, v_synclist);
761 	}
762 	if (need_iosync) {
763 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
764 
765 		/*
766 		 * XXX this doesn't use the new two-flag scheme,
767 		 * but to use that, all i/o initiators will have to change.
768 		 */
769 
770 		s = splbio();
771 		while (vp->v_numoutput != 0) {
772 			UVMHIST_LOG(ubchist, "waiting for vp %p num %d",
773 				    vp, vp->v_numoutput,0,0);
774 
775 			vp->v_flag |= VBWAIT;
776 			UVM_UNLOCK_AND_WAIT(&vp->v_numoutput,
777 					    &uvn->u_obj.vmobjlock,
778 					    FALSE, "uvn_flush",0);
779 			simple_lock(&uvn->u_obj.vmobjlock);
780 		}
781 		splx(s);
782 	}
783 
784 	/* return, with object locked! */
785 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
786 	return(retval);
787 }
788 
789 /*
790  * uvn_cluster
791  *
792  * we are about to do I/O in an object at offset.   this function is called
793  * to establish a range of offsets around "offset" in which we can cluster
794  * I/O.
795  *
796  * - currently doesn't matter if obj locked or not.
797  */
798 
799 static void
800 uvn_cluster(uobj, offset, loffset, hoffset)
801 	struct uvm_object *uobj;
802 	voff_t offset;
803 	voff_t *loffset, *hoffset; /* OUT */
804 {
805 	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;
806 
807 	*loffset = offset;
808 	*hoffset = MIN(offset + MAXBSIZE, round_page(uvn->u_size));
809 }
810 
811 /*
812  * uvn_put: flush page data to backing store.
813  *
814  * => object must be locked!   we will _unlock_ it before starting I/O.
815  * => flags: PGO_SYNCIO -- use sync. I/O
816  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
817  */
818 
819 static int
820 uvn_put(uobj, pps, npages, flags)
821 	struct uvm_object *uobj;
822 	struct vm_page **pps;
823 	int npages, flags;
824 {
825 	struct vnode *vp = (struct vnode *)uobj;
826 	int error;
827 
828 	error = VOP_PUTPAGES(vp, pps, npages, flags, NULL);
829 	return error;
830 }
831 
832 
833 /*
834  * uvn_get: get pages (synchronously) from backing store
835  *
836  * => prefer map unlocked (not required)
837  * => object must be locked!  we will _unlock_ it before starting any I/O.
838  * => flags: PGO_ALLPAGES: get all of the pages
839  *           PGO_LOCKED: fault data structures are locked
840  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
841  * => NOTE: caller must check for released pages!!
842  */
843 
844 static int
845 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
846 	struct uvm_object *uobj;
847 	voff_t offset;
848 	struct vm_page **pps;		/* IN/OUT */
849 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
850 	int centeridx;
851 	vm_prot_t access_type;
852 	int advice, flags;
853 {
854 	struct vnode *vp = (struct vnode *)uobj;
855 	int error;
856 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(ubchist);
857 
858 	UVMHIST_LOG(ubchist, "vp %p off 0x%x", vp, (int)offset, 0,0);
859 	error = VOP_GETPAGES(vp, offset, pps, npagesp, centeridx,
860 			     access_type, advice, flags);
861 	return error;
862 }
863 
864 
865 /*
866  * uvn_findpages:
867  * return the page for the uobj and offset requested, allocating if needed.
868  * => uobj must be locked.
869  * => returned page will be BUSY.
870  */
871 
872 void
873 uvn_findpages(uobj, offset, npagesp, pps, flags)
874 	struct uvm_object *uobj;
875 	voff_t offset;
876 	int *npagesp;
877 	struct vm_page **pps;
878 	int flags;
879 {
880 	int i, rv, npages;
881 
882 	rv = 0;
883 	npages = *npagesp;
884 	for (i = 0; i < npages; i++, offset += PAGE_SIZE) {
885 		rv += uvn_findpage(uobj, offset, &pps[i], flags);
886 	}
887 	*npagesp = rv;
888 }
889 
890 static int
891 uvn_findpage(uobj, offset, pgp, flags)
892 	struct uvm_object *uobj;
893 	voff_t offset;
894 	struct vm_page **pgp;
895 	int flags;
896 {
897 	struct vm_page *pg;
898 	UVMHIST_FUNC("uvn_findpage"); UVMHIST_CALLED(ubchist);
899 	UVMHIST_LOG(ubchist, "vp %p off 0x%lx", uobj, offset,0,0);
900 
901 	if (*pgp != NULL) {
902 		UVMHIST_LOG(ubchist, "dontcare", 0,0,0,0);
903 		return 0;
904 	}
905 	for (;;) {
906 		/* look for an existing page */
907 		pg = uvm_pagelookup(uobj, offset);
908 
909 		/* nope?   allocate one now */
910 		if (pg == NULL) {
911 			if (flags & UFP_NOALLOC) {
912 				UVMHIST_LOG(ubchist, "noalloc", 0,0,0,0);
913 				return 0;
914 			}
915 			pg = uvm_pagealloc(uobj, offset, NULL, 0);
916 			if (pg == NULL) {
917 				if (flags & UFP_NOWAIT) {
918 					UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
919 					return 0;
920 				}
921 				simple_unlock(&uobj->vmobjlock);
922 				uvm_wait("uvn_fp1");
923 				simple_lock(&uobj->vmobjlock);
924 				continue;
925 			}
926 			if (UVM_OBJ_IS_VTEXT(uobj)) {
927 				uvmexp.vtextpages++;
928 			} else {
929 				uvmexp.vnodepages++;
930 			}
931 			UVMHIST_LOG(ubchist, "alloced",0,0,0,0);
932 			break;
933 		} else if (flags & UFP_NOCACHE) {
934 			UVMHIST_LOG(ubchist, "nocache",0,0,0,0);
935 			return 0;
936 		}
937 
938 		/* page is there, see if we need to wait on it */
939 		if ((pg->flags & (PG_BUSY|PG_RELEASED)) != 0) {
940 			if (flags & UFP_NOWAIT) {
941 				UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
942 				return 0;
943 			}
944 			pg->flags |= PG_WANTED;
945 			UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0,
946 					    "uvn_fp2", 0);
947 			simple_lock(&uobj->vmobjlock);
948 			continue;
949 		}
950 
951 		/* skip PG_RDONLY pages if requested */
952 		if ((flags & UFP_NORDONLY) && (pg->flags & PG_RDONLY)) {
953 			UVMHIST_LOG(ubchist, "nordonly",0,0,0,0);
954 			return 0;
955 		}
956 
957 		/* mark the page BUSY and we're done. */
958 		pg->flags |= PG_BUSY;
959 		UVM_PAGE_OWN(pg, "uvn_findpage");
960 		UVMHIST_LOG(ubchist, "found",0,0,0,0);
961 		break;
962 	}
963 	*pgp = pg;
964 	return 1;
965 }
966 
967 /*
968  * uvm_vnp_setsize: grow or shrink a vnode uvn
969  *
970  * grow   => just update size value
971  * shrink => toss un-needed pages
972  *
973  * => we assume that the caller has a reference of some sort to the
974  *	vnode in question so that it will not be yanked out from under
975  *	us.
976  *
977  * called from:
978  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
979  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
980  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
981  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
982  *  => union fs: union_newsize
983  */
984 
985 void
986 uvm_vnp_setsize(vp, newsize)
987 	struct vnode *vp;
988 	voff_t newsize;
989 {
990 	struct uvm_vnode *uvn = &vp->v_uvm;
991 	voff_t pgend = round_page(newsize);
992 	UVMHIST_FUNC("uvm_vnp_setsize"); UVMHIST_CALLED(ubchist);
993 
994 	simple_lock(&uvn->u_obj.vmobjlock);
995 
996 	UVMHIST_LOG(ubchist, "old 0x%x new 0x%x", uvn->u_size, newsize, 0,0);
997 
998 	/*
999 	 * now check if the size has changed: if we shrink we had better
1000 	 * toss some pages...
1001 	 */
1002 
1003 	if (uvn->u_size > pgend && uvn->u_size != VSIZENOTSET) {
1004 		(void) uvn_flush(&uvn->u_obj, pgend, 0, PGO_FREE);
1005 	}
1006 	uvn->u_size = newsize;
1007 	simple_unlock(&uvn->u_obj.vmobjlock);
1008 }
1009 
1010 /*
1011  * uvm_vnp_zerorange:  set a range of bytes in a file to zero.
1012  */
1013 
1014 void
1015 uvm_vnp_zerorange(vp, off, len)
1016 	struct vnode *vp;
1017 	off_t off;
1018 	size_t len;
1019 {
1020         void *win;
1021 
1022         /*
1023          * XXXUBC invent kzero() and use it
1024          */
1025 
1026         while (len) {
1027                 vsize_t bytelen = len;
1028 
1029                 win = ubc_alloc(&vp->v_uvm.u_obj, off, &bytelen, UBC_WRITE);
1030                 memset(win, 0, bytelen);
1031                 ubc_release(win, 0);
1032 
1033                 off += bytelen;
1034                 len -= bytelen;
1035         }
1036 }
1037