xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision 5aefcfdc06931dd97e76246d2fe0302f7b3fe094)
1 /*	$NetBSD: uvm_vnode.c,v 1.40 2000/12/16 06:17:09 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, vm_page_t *,
86 				     int *, int, vm_prot_t, int, int));
87 static int		uvn_put __P((struct uvm_object *, vm_page_t *, int,
88 				     boolean_t));
89 static void		uvn_reference __P((struct uvm_object *));
90 static boolean_t	uvn_releasepg __P((struct vm_page *,
91 					   struct vm_page **));
92 
93 /*
94  * master pager structure
95  */
96 
97 struct uvm_pagerops uvm_vnodeops = {
98 	NULL,
99 	uvn_reference,
100 	uvn_detach,
101 	NULL,
102 	uvn_flush,
103 	uvn_get,
104 	uvn_put,
105 	uvn_cluster,
106 	uvm_mk_pcluster,
107 	uvn_releasepg,
108 };
109 
110 /*
111  * the ops!
112  */
113 
114 /*
115  * uvn_attach
116  *
117  * attach a vnode structure to a VM object.  if the vnode is already
118  * attached, then just bump the reference count by one and return the
119  * VM object.   if not already attached, attach and return the new VM obj.
120  * the "accessprot" tells the max access the attaching thread wants to
121  * our pages.
122  *
123  * => caller must _not_ already be holding the lock on the uvm_object.
124  * => in fact, nothing should be locked so that we can sleep here.
125  * => note that uvm_object is first thing in vnode structure, so their
126  *    pointers are equiv.
127  */
128 
129 struct uvm_object *
130 uvn_attach(arg, accessprot)
131 	void *arg;
132 	vm_prot_t accessprot;
133 {
134 	struct vnode *vp = arg;
135 	struct uvm_vnode *uvn = &vp->v_uvm;
136 	struct vattr vattr;
137 	int result;
138 	struct partinfo pi;
139 	voff_t used_vnode_size;
140 	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);
141 
142 	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
143 	used_vnode_size = (voff_t)0;
144 
145 	/*
146 	 * first get a lock on the uvn.
147 	 */
148 	simple_lock(&uvn->u_obj.vmobjlock);
149 	while (uvn->u_flags & VXLOCK) {
150 		uvn->u_flags |= VXWANT;
151 		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
152 		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
153 		    "uvn_attach", 0);
154 		simple_lock(&uvn->u_obj.vmobjlock);
155 		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
156 	}
157 
158 	/*
159 	 * if we're mapping a BLK device, make sure it is a disk.
160 	 */
161 	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
162 		simple_unlock(&uvn->u_obj.vmobjlock);
163 		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
164 		return(NULL);
165 	}
166 
167 #ifdef DIAGNOSTIC
168 	if (vp->v_type != VREG) {
169 		panic("uvn_attach: vp %p not VREG", vp);
170 	}
171 #endif
172 
173 	/*
174 	 * set up our idea of the size
175 	 * if this hasn't been done already.
176 	 */
177 	if (uvn->u_size == VSIZENOTSET) {
178 
179 	uvn->u_flags |= VXLOCK;
180 	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
181 		/* XXX: curproc? */
182 	if (vp->v_type == VBLK) {
183 		/*
184 		 * We could implement this as a specfs getattr call, but:
185 		 *
186 		 *	(1) VOP_GETATTR() would get the file system
187 		 *	    vnode operation, not the specfs operation.
188 		 *
189 		 *	(2) All we want is the size, anyhow.
190 		 */
191 		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
192 		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
193 		if (result == 0) {
194 			/* XXX should remember blocksize */
195 			used_vnode_size = (voff_t)pi.disklab->d_secsize *
196 			    (voff_t)pi.part->p_size;
197 		}
198 	} else {
199 		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
200 		if (result == 0)
201 			used_vnode_size = vattr.va_size;
202 	}
203 
204 	/* relock object */
205 	simple_lock(&uvn->u_obj.vmobjlock);
206 
207 	if (uvn->u_flags & VXWANT)
208 		wakeup(uvn);
209 	uvn->u_flags &= ~(VXLOCK|VXWANT);
210 
211 	if (result != 0) {
212 		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
213 		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
214 		return(NULL);
215 	}
216 	uvn->u_size = used_vnode_size;
217 
218 	}
219 
220 	/* unlock and return */
221 	simple_unlock(&uvn->u_obj.vmobjlock);
222 	UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
223 	    0, 0, 0);
224 	return (&uvn->u_obj);
225 }
226 
227 
228 /*
229  * uvn_reference
230  *
231  * duplicate a reference to a VM object.  Note that the reference
232  * count must already be at least one (the passed in reference) so
233  * there is no chance of the uvn being killed or locked out here.
234  *
235  * => caller must call with object unlocked.
236  * => caller must be using the same accessprot as was used at attach time
237  */
238 
239 
240 static void
241 uvn_reference(uobj)
242 	struct uvm_object *uobj;
243 {
244 	VREF((struct vnode *)uobj);
245 }
246 
247 /*
248  * uvn_detach
249  *
250  * remove a reference to a VM object.
251  *
252  * => caller must call with object unlocked and map locked.
253  * => this starts the detach process, but doesn't have to finish it
254  *    (async i/o could still be pending).
255  */
256 static void
257 uvn_detach(uobj)
258 	struct uvm_object *uobj;
259 {
260 	vrele((struct vnode *)uobj);
261 }
262 
263 /*
264  * uvn_releasepg: handled a released page in a uvn
265  *
266  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
267  *	to dispose of.
268  * => caller must handled PG_WANTED case
269  * => called with page's object locked, pageq's unlocked
270  * => returns TRUE if page's object is still alive, FALSE if we
271  *	killed the page's object.    if we return TRUE, then we
272  *	return with the object locked.
273  * => if (nextpgp != NULL) => we return the next page on the queue, and return
274  *				with the page queues locked [for pagedaemon]
275  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
276  * => we kill the uvn if it is not referenced and we are suppose to
277  *	kill it ("relkill").
278  */
279 
280 boolean_t
281 uvn_releasepg(pg, nextpgp)
282 	struct vm_page *pg;
283 	struct vm_page **nextpgp;	/* OUT */
284 {
285 	KASSERT(pg->flags & PG_RELEASED);
286 
287 	/*
288 	 * dispose of the page [caller handles PG_WANTED]
289 	 */
290 	pmap_page_protect(pg, VM_PROT_NONE);
291 	uvm_lock_pageq();
292 	if (nextpgp)
293 		*nextpgp = TAILQ_NEXT(pg, pageq);
294 	uvm_pagefree(pg);
295 	if (!nextpgp)
296 		uvm_unlock_pageq();
297 
298 	return (TRUE);
299 }
300 
301 /*
302  * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
303  * through the buffer cache and allow I/O in any size.  These VOPs use
304  * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
305  * go through the buffer cache or allow I/O sizes larger than a
306  * block].  we will eventually want to change this.
307  *
308  * issues to consider:
309  *   uvm provides the uvm_aiodesc structure for async i/o management.
310  * there are two tailq's in the uvm. structure... one for pending async
311  * i/o and one for "done" async i/o.   to do an async i/o one puts
312  * an aiodesc on the "pending" list (protected by splbio()), starts the
313  * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
314  * some sort of "i/o done" function to be called (at splbio(), interrupt
315  * time).   this function should remove the aiodesc from the pending list
316  * and place it on the "done" list and wakeup the daemon.   the daemon
317  * will run at normal spl() and will remove all items from the "done"
318  * list and call the "aiodone" hook for each done request (see uvm_pager.c).
319  * [in the old vm code, this was done by calling the "put" routine with
320  * null arguments which made the code harder to read and understand because
321  * you had one function ("put") doing two things.]
322  *
323  * so the current pager needs:
324  *   int uvn_aiodone(struct uvm_aiodesc *)
325  *
326  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
327  *	later collection.
328  * => called with pageq's locked by the daemon.
329  *
330  * general outline:
331  * - "try" to lock object.   if fail, just return (will try again later)
332  * - drop "u_nio" (this req is done!)
333  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
334  * - get "page" structures (atop?).
335  * - handle "wanted" pages
336  * - handle "released" pages [using pgo_releasepg]
337  *   >>> pgo_releasepg may kill the object
338  * dont forget to look at "object" wanted flag in all cases.
339  */
340 
341 
342 /*
343  * uvn_flush: flush pages out of a uvm object.
344  *
345  * => object should be locked by caller.   we may _unlock_ the object
346  *	if (and only if) we need to clean a page (PGO_CLEANIT).
347  *	we return with the object locked.
348  * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
349  *	might want to unlock higher level resources (e.g. vm_map)
350  *	before calling flush.
351  * => if PGO_CLEANIT is not set, then we will neither unlock the object
352  *	or block.
353  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
354  *	for flushing.
355  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
356  *	that new pages are inserted on the tail end of the list.   thus,
357  *	we can make a complete pass through the object in one go by starting
358  *	at the head and working towards the tail (new pages are put in
359  *	front of us).
360  * => NOTE: we are allowed to lock the page queues, so the caller
361  *	must not be holding the lock on them [e.g. pagedaemon had
362  *	better not call us with the queues locked]
363  * => we return TRUE unless we encountered some sort of I/O error
364  *
365  * comment on "cleaning" object and PG_BUSY pages:
366  *	this routine is holding the lock on the object.   the only time
367  *	that it can run into a PG_BUSY page that it does not own is if
368  *	some other process has started I/O on the page (e.g. either
369  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
370  *	in, then it can not be dirty (!PG_CLEAN) because no one has
371  *	had a chance to modify it yet.    if the PG_BUSY page is being
372  *	paged out then it means that someone else has already started
373  *	cleaning the page for us (how nice!).    in this case, if we
374  *	have syncio specified, then after we make our pass through the
375  *	object we need to wait for the other PG_BUSY pages to clear
376  *	off (i.e. we need to do an iosync).   also note that once a
377  *	page is PG_BUSY it must stay in its object until it is un-busyed.
378  *
379  * note on page traversal:
380  *	we can traverse the pages in an object either by going down the
381  *	linked list in "uobj->memq", or we can go over the address range
382  *	by page doing hash table lookups for each address.    depending
383  *	on how many pages are in the object it may be cheaper to do one
384  *	or the other.   we set "by_list" to true if we are using memq.
385  *	if the cost of a hash lookup was equal to the cost of the list
386  *	traversal we could compare the number of pages in the start->stop
387  *	range to the total number of pages in the object.   however, it
388  *	seems that a hash table lookup is more expensive than the linked
389  *	list traversal, so we multiply the number of pages in the
390  *	start->stop range by a penalty which we define below.
391  */
392 
393 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
394 
395 static boolean_t
396 uvn_flush(uobj, start, stop, flags)
397 	struct uvm_object *uobj;
398 	voff_t start, stop;
399 	int flags;
400 {
401 	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;
402 	struct vnode *vp = (struct vnode *)uobj;
403 	struct vm_page *pp, *ppnext, *ptmp;
404 	struct vm_page *pps[256], **ppsp;
405 	int s;
406 	int npages, result, lcv;
407 	boolean_t retval, need_iosync, by_list, needs_clean, all, wasclean;
408 	voff_t curoff;
409 	u_short pp_version;
410 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
411 	UVMHIST_LOG(maphist, "uobj %p start 0x%x stop 0x%x flags 0x%x",
412 		    uobj, start, stop, flags);
413 	KASSERT(flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
414 
415 #ifdef DEBUG
416 	if (uvn->u_size == VSIZENOTSET) {
417 		printf("uvn_flush: size not set vp %p\n", uvn);
418 		vprint("uvn_flush VSIZENOTSET", vp);
419 		flags |= PGO_ALLPAGES;
420 	}
421 #endif
422 
423 	/*
424 	 * get init vals and determine how we are going to traverse object
425 	 */
426 
427 	curoff = 0;
428 	need_iosync = FALSE;
429 	retval = TRUE;
430 	wasclean = TRUE;
431 	if (flags & PGO_ALLPAGES) {
432 		all = TRUE;
433 		by_list = TRUE;
434 	} else {
435 		start = trunc_page(start);
436 		stop = round_page(stop);
437 #ifdef DEBUG
438 		if (stop > round_page(uvn->u_size)) {
439 			printf("uvn_flush: oor vp %p start 0x%x stop 0x%x "
440 			       "size 0x%x\n", uvn, (int)start, (int)stop,
441 			       (int)round_page(uvn->u_size));
442 		}
443 #endif
444 		all = FALSE;
445 		by_list = (uobj->uo_npages <=
446 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
447 	}
448 
449 	UVMHIST_LOG(maphist,
450 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
451 	    start, stop, by_list, flags);
452 
453 	/*
454 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
455 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
456 	 * is wrong it will only prevent us from clustering... it won't break
457 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
458 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
459 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
460 	 * bit is always up to date since there are no mappings).
461 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
462 	 */
463 
464 	if ((flags & PGO_CLEANIT) != 0 &&
465 	    uobj->pgops->pgo_mk_pcluster != NULL) {
466 		if (by_list) {
467 			TAILQ_FOREACH(pp, &uobj->memq, listq) {
468 				if (!all &&
469 				    (pp->offset < start || pp->offset >= stop))
470 					continue;
471 				pp->flags &= ~PG_CLEANCHK;
472 			}
473 
474 		} else {   /* by hash */
475 			for (curoff = start ; curoff < stop;
476 			    curoff += PAGE_SIZE) {
477 				pp = uvm_pagelookup(uobj, curoff);
478 				if (pp)
479 					pp->flags &= ~PG_CLEANCHK;
480 			}
481 		}
482 	}
483 
484 	/*
485 	 * now do it.   note: we must update ppnext in body of loop or we
486 	 * will get stuck.  we need to use ppnext because we may free "pp"
487 	 * before doing the next loop.
488 	 */
489 
490 	if (by_list) {
491 		pp = TAILQ_FIRST(&uobj->memq);
492 	} else {
493 		curoff = start;
494 		pp = uvm_pagelookup(uobj, curoff);
495 	}
496 
497 	ppnext = NULL;
498 	ppsp = NULL;
499 	uvm_lock_pageq();
500 
501 	/* locked: both page queues and uobj */
502 	for ( ; (by_list && pp != NULL) ||
503 		      (!by_list && curoff < stop) ; pp = ppnext) {
504 		if (by_list) {
505 			if (!all &&
506 			    (pp->offset < start || pp->offset >= stop)) {
507 				ppnext = TAILQ_NEXT(pp, listq);
508 				continue;
509 			}
510 		} else {
511 			curoff += PAGE_SIZE;
512 			if (pp == NULL) {
513 				if (curoff < stop)
514 					ppnext = uvm_pagelookup(uobj, curoff);
515 				continue;
516 			}
517 		}
518 
519 		/*
520 		 * handle case where we do not need to clean page (either
521 		 * because we are not clean or because page is not dirty or
522 		 * is busy):
523 		 *
524 		 * NOTE: we are allowed to deactivate a non-wired active
525 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
526 		 * queue it must stay put until it is !PG_BUSY (so as not to
527 		 * confuse pagedaemon).
528 		 */
529 
530 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
531 			needs_clean = FALSE;
532 			if ((flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
533 			             (PGO_CLEANIT|PGO_SYNCIO))
534 				need_iosync = TRUE;
535 		} else {
536 
537 			/*
538 			 * freeing: nuke all mappings so we can sync
539 			 * PG_CLEAN bit with no race
540 			 */
541 			if ((pp->flags & PG_CLEAN) != 0 &&
542 			    (flags & PGO_FREE) != 0 &&
543 			    (pp->pqflags & PQ_ACTIVE) != 0)
544 				pmap_page_protect(pp, VM_PROT_NONE);
545 			if ((pp->flags & PG_CLEAN) != 0 &&
546 			    pmap_is_modified(pp))
547 				pp->flags &= ~(PG_CLEAN);
548 			pp->flags |= PG_CLEANCHK;
549 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
550 		}
551 
552 		/*
553 		 * if we don't need a clean... load ppnext and dispose of pp
554 		 */
555 		if (!needs_clean) {
556 			if (by_list)
557 				ppnext = TAILQ_NEXT(pp, listq);
558 			else {
559 				if (curoff < stop)
560 					ppnext = uvm_pagelookup(uobj, curoff);
561 			}
562 
563 			if (flags & PGO_DEACTIVATE) {
564 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
565 				    (pp->flags & PG_BUSY) == 0 &&
566 				    pp->wire_count == 0) {
567 					pmap_page_protect(pp, VM_PROT_NONE);
568 					uvm_pagedeactivate(pp);
569 				}
570 
571 			} else if (flags & PGO_FREE) {
572 				if (pp->flags & PG_BUSY) {
573 					pp->flags |= PG_RELEASED;
574 				} else {
575 					pmap_page_protect(pp, VM_PROT_NONE);
576 					uvm_pagefree(pp);
577 				}
578 			}
579 			/* ppnext is valid so we can continue... */
580 			continue;
581 		}
582 
583 		/*
584 		 * pp points to a page in the locked object that we are
585 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
586 		 * for cleaning (PGO_CLEANIT).  we clean it now.
587 		 *
588 		 * let uvm_pager_put attempted a clustered page out.
589 		 * note: locked: uobj and page queues.
590 		 */
591 
592 		wasclean = FALSE;
593 		pp->flags |= PG_BUSY;	/* we 'own' page now */
594 		UVM_PAGE_OWN(pp, "uvn_flush");
595 		pmap_page_protect(pp, VM_PROT_READ);
596 		pp_version = pp->version;
597 ReTry:
598 		ppsp = pps;
599 		npages = sizeof(pps) / sizeof(struct vm_page *);
600 
601 		/* locked: page queues, uobj */
602 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
603 				       flags | PGO_DOACTCLUST, start, stop);
604 		/* unlocked: page queues, uobj */
605 
606 		/*
607 		 * at this point nothing is locked.   if we did an async I/O
608 		 * it is remotely possible for the async i/o to complete and
609 		 * the page "pp" be freed or what not before we get a chance
610 		 * to relock the object.   in order to detect this, we have
611 		 * saved the version number of the page in "pp_version".
612 		 */
613 
614 		/* relock! */
615 		simple_lock(&uobj->vmobjlock);
616 		uvm_lock_pageq();
617 
618 		/*
619 		 * VM_PAGER_AGAIN: given the structure of this pager, this
620 		 * can only happen when  we are doing async I/O and can't
621 		 * map the pages into kernel memory (pager_map) due to lack
622 		 * of vm space.   if this happens we drop back to sync I/O.
623 		 */
624 
625 		if (result == VM_PAGER_AGAIN) {
626 
627 			/*
628 			 * it is unlikely, but page could have been released
629 			 * while we had the object lock dropped.   we ignore
630 			 * this now and retry the I/O.  we will detect and
631 			 * handle the released page after the syncio I/O
632 			 * completes.
633 			 */
634 #ifdef DIAGNOSTIC
635 			if (flags & PGO_SYNCIO)
636 	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
637 #endif
638 			flags |= PGO_SYNCIO;
639 			goto ReTry;
640 		}
641 
642 		/*
643 		 * the cleaning operation is now done.   finish up.  note that
644 		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
645 		 * if success (OK, PEND) then uvm_pager_put returns the cluster
646 		 * to us in ppsp/npages.
647 		 */
648 
649 		/*
650 		 * for pending async i/o if we are not deactivating/freeing
651 		 * we can move on to the next page.
652 		 */
653 
654 		if (result == VM_PAGER_PEND &&
655 		    (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
656 
657 			/*
658 			 * no per-page ops: refresh ppnext and continue
659 			 */
660 			if (by_list) {
661 				if (pp->version == pp_version)
662 					ppnext = TAILQ_NEXT(pp, listq);
663 				else
664 					ppnext = TAILQ_FIRST(&uobj->memq);
665 			} else {
666 				if (curoff < stop)
667 					ppnext = uvm_pagelookup(uobj, curoff);
668 			}
669 			continue;
670 		}
671 
672 		/*
673 		 * need to look at each page of the I/O operation.  we defer
674 		 * processing "pp" until the last trip through this "for" loop
675 		 * so that we can load "ppnext" for the main loop after we
676 		 * play with the cluster pages [thus the "npages + 1" in the
677 		 * loop below].
678 		 */
679 
680 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
681 
682 			/*
683 			 * handle ppnext for outside loop, and saving pp
684 			 * until the end.
685 			 */
686 			if (lcv < npages) {
687 				if (ppsp[lcv] == pp)
688 					continue; /* skip pp until the end */
689 				ptmp = ppsp[lcv];
690 			} else {
691 				ptmp = pp;
692 
693 				/* set up next page for outer loop */
694 				if (by_list) {
695 					if (pp->version == pp_version)
696 						ppnext = TAILQ_NEXT(pp, listq);
697 					else
698 						ppnext = TAILQ_FIRST(
699 						    &uobj->memq);
700 				} else {
701 					if (curoff < stop)
702 						ppnext = uvm_pagelookup(uobj,
703 						    curoff);
704 				}
705 			}
706 
707 			/*
708 			 * verify the page wasn't moved while obj was
709 			 * unlocked
710 			 */
711 			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
712 				continue;
713 
714 			/*
715 			 * unbusy the page if I/O is done.   note that for
716 			 * pending I/O it is possible that the I/O op
717 			 * finished before we relocked the object (in
718 			 * which case the page is no longer busy).
719 			 */
720 
721 			if (result != VM_PAGER_PEND) {
722 				if (ptmp->flags & PG_WANTED) {
723 					/* still holding object lock */
724 					wakeup(ptmp);
725 				}
726 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
727 				UVM_PAGE_OWN(ptmp, NULL);
728 				if (ptmp->flags & PG_RELEASED) {
729 					uvm_unlock_pageq();
730 					if (!uvn_releasepg(ptmp, NULL)) {
731 						UVMHIST_LOG(maphist,
732 							    "released %p",
733 							    ptmp, 0,0,0);
734 						return (TRUE);
735 					}
736 					uvm_lock_pageq();
737 					continue;
738 				} else {
739 					if ((flags & PGO_WEAK) == 0 &&
740 					    !(result == VM_PAGER_ERROR &&
741 					      curproc == uvm.pagedaemon_proc)) {
742 						ptmp->flags |=
743 							(PG_CLEAN|PG_CLEANCHK);
744 						if ((flags & PGO_FREE) == 0) {
745 							pmap_clear_modify(ptmp);
746 						}
747 					}
748 				}
749 			}
750 
751 			/*
752 			 * dispose of page
753 			 */
754 
755 			if (flags & PGO_DEACTIVATE) {
756 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
757 				    (pp->flags & PG_BUSY) == 0 &&
758 				    pp->wire_count == 0) {
759 					pmap_page_protect(ptmp, VM_PROT_NONE);
760 					uvm_pagedeactivate(ptmp);
761 				}
762 			} else if (flags & PGO_FREE) {
763 				if (result == VM_PAGER_PEND) {
764 					if ((ptmp->flags & PG_BUSY) != 0)
765 						/* signal for i/o done */
766 						ptmp->flags |= PG_RELEASED;
767 				} else {
768 					if (result != VM_PAGER_OK) {
769 						printf("uvn_flush: obj=%p, "
770 						   "offset=0x%llx.  error %d\n",
771 						    pp->uobject,
772 						    (long long)pp->offset,
773 						    result);
774 						printf("uvn_flush: WARNING: "
775 						    "changes to page may be "
776 						    "lost!\n");
777 						retval = FALSE;
778 					}
779 					pmap_page_protect(ptmp, VM_PROT_NONE);
780 					uvm_pagefree(ptmp);
781 				}
782 			}
783 		}		/* end of "lcv" for loop */
784 	}		/* end of "pp" for loop */
785 
786 	uvm_unlock_pageq();
787 	if ((flags & PGO_CLEANIT) && all && wasclean &&
788 	    LIST_FIRST(&vp->v_dirtyblkhd) == NULL &&
789 	    (vp->v_flag & VONWORKLST)) {
790 		vp->v_flag &= ~VONWORKLST;
791 		LIST_REMOVE(vp, v_synclist);
792 	}
793 	if (need_iosync) {
794 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
795 
796 		/*
797 		 * XXX this doesn't use the new two-flag scheme,
798 		 * but to use that, all i/o initiators will have to change.
799 		 */
800 
801 		s = splbio();
802 		while (vp->v_numoutput != 0) {
803 			UVMHIST_LOG(ubchist, "waiting for vp %p num %d",
804 				    vp, vp->v_numoutput,0,0);
805 
806 			vp->v_flag |= VBWAIT;
807 			UVM_UNLOCK_AND_WAIT(&vp->v_numoutput,
808 					    &uvn->u_obj.vmobjlock,
809 					    FALSE, "uvn_flush",0);
810 			simple_lock(&uvn->u_obj.vmobjlock);
811 		}
812 		splx(s);
813 	}
814 
815 	/* return, with object locked! */
816 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
817 	return(retval);
818 }
819 
820 /*
821  * uvn_cluster
822  *
823  * we are about to do I/O in an object at offset.   this function is called
824  * to establish a range of offsets around "offset" in which we can cluster
825  * I/O.
826  *
827  * - currently doesn't matter if obj locked or not.
828  */
829 
830 static void
831 uvn_cluster(uobj, offset, loffset, hoffset)
832 	struct uvm_object *uobj;
833 	voff_t offset;
834 	voff_t *loffset, *hoffset; /* OUT */
835 {
836 	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;
837 
838 	*loffset = offset;
839 	*hoffset = min(offset + MAXBSIZE, round_page(uvn->u_size));
840 }
841 
842 /*
843  * uvn_put: flush page data to backing store.
844  *
845  * => object must be locked!   we will _unlock_ it before starting I/O.
846  * => flags: PGO_SYNCIO -- use sync. I/O
847  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
848  */
849 
850 static int
851 uvn_put(uobj, pps, npages, flags)
852 	struct uvm_object *uobj;
853 	struct vm_page **pps;
854 	int npages, flags;
855 {
856 	struct vnode *vp = (struct vnode *)uobj;
857 	int error;
858 
859 	error = VOP_PUTPAGES(vp, pps, npages, flags, NULL);
860 	return uvm_errno2vmerror(error);
861 }
862 
863 
864 /*
865  * uvn_get: get pages (synchronously) from backing store
866  *
867  * => prefer map unlocked (not required)
868  * => object must be locked!  we will _unlock_ it before starting any I/O.
869  * => flags: PGO_ALLPAGES: get all of the pages
870  *           PGO_LOCKED: fault data structures are locked
871  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
872  * => NOTE: caller must check for released pages!!
873  */
874 
875 static int
876 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
877 	struct uvm_object *uobj;
878 	voff_t offset;
879 	struct vm_page **pps;		/* IN/OUT */
880 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
881 	int centeridx;
882 	vm_prot_t access_type;
883 	int advice, flags;
884 {
885 	struct vnode *vp = (struct vnode *)uobj;
886 	int error;
887 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(ubchist);
888 
889 	UVMHIST_LOG(ubchist, "vp %p off 0x%x", vp, (int)offset, 0,0);
890 	error = VOP_GETPAGES(vp, offset, pps, npagesp, centeridx,
891 			     access_type, advice, flags);
892 	return uvm_errno2vmerror(error);
893 }
894 
895 
896 /*
897  * uvn_findpages:
898  * return the page for the uobj and offset requested, allocating if needed.
899  * => uobj must be locked.
900  * => returned page will be BUSY.
901  */
902 
903 void
904 uvn_findpages(uobj, offset, npagesp, pps, flags)
905 	struct uvm_object *uobj;
906 	voff_t offset;
907 	int *npagesp;
908 	struct vm_page **pps;
909 	int flags;
910 {
911 	int i, rv, npages;
912 
913 	rv = 0;
914 	npages = *npagesp;
915 	for (i = 0; i < npages; i++, offset += PAGE_SIZE) {
916 		rv += uvn_findpage(uobj, offset, &pps[i], flags);
917 	}
918 	*npagesp = rv;
919 }
920 
921 static int
922 uvn_findpage(uobj, offset, pgp, flags)
923 	struct uvm_object *uobj;
924 	voff_t offset;
925 	struct vm_page **pgp;
926 	int flags;
927 {
928 	struct vm_page *pg;
929 	UVMHIST_FUNC("uvn_findpage"); UVMHIST_CALLED(ubchist);
930 	UVMHIST_LOG(ubchist, "vp %p off 0x%lx", uobj, offset,0,0);
931 
932 	if (*pgp != NULL) {
933 		UVMHIST_LOG(ubchist, "dontcare", 0,0,0,0);
934 		return 0;
935 	}
936 	for (;;) {
937 		/* look for an existing page */
938 		pg = uvm_pagelookup(uobj, offset);
939 
940 		/* nope?   allocate one now */
941 		if (pg == NULL) {
942 			if (flags & UFP_NOALLOC) {
943 				UVMHIST_LOG(ubchist, "noalloc", 0,0,0,0);
944 				return 0;
945 			}
946 			if (uvmexp.vnodepages >
947 			    (uvmexp.active + uvmexp.inactive + uvmexp.wired +
948 			     uvmexp.free) * 7 / 8) {
949 				pg = NULL;
950 			} else {
951 				pg = uvm_pagealloc(uobj, offset, NULL, 0);
952 			}
953 			if (pg == NULL) {
954 				if (flags & UFP_NOWAIT) {
955 					UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
956 					return 0;
957 				}
958 				simple_unlock(&uobj->vmobjlock);
959 				uvm_wait("uvn_fp1");
960 				simple_lock(&uobj->vmobjlock);
961 				continue;
962 			}
963 			uvmexp.vnodepages++;
964 			UVMHIST_LOG(ubchist, "alloced",0,0,0,0);
965 			break;
966 		} else if (flags & UFP_NOCACHE) {
967 			UVMHIST_LOG(ubchist, "nocache",0,0,0,0);
968 			return 0;
969 		}
970 
971 		/* page is there, see if we need to wait on it */
972 		if ((pg->flags & (PG_BUSY|PG_RELEASED)) != 0) {
973 			if (flags & UFP_NOWAIT) {
974 				UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
975 				return 0;
976 			}
977 			pg->flags |= PG_WANTED;
978 			UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0,
979 					    "uvn_fp2", 0);
980 			simple_lock(&uobj->vmobjlock);
981 			continue;
982 		}
983 
984 		/* skip PG_RDONLY pages if requested */
985 		if ((flags & UFP_NORDONLY) && (pg->flags & PG_RDONLY)) {
986 			UVMHIST_LOG(ubchist, "nordonly",0,0,0,0);
987 			return 0;
988 		}
989 
990 		/* mark the page BUSY and we're done. */
991 		pg->flags |= PG_BUSY;
992 		UVM_PAGE_OWN(pg, "uvn_findpage");
993 		UVMHIST_LOG(ubchist, "found",0,0,0,0);
994 		break;
995 	}
996 	*pgp = pg;
997 	return 1;
998 }
999 
1000 /*
1001  * uvm_vnp_setsize: grow or shrink a vnode uvn
1002  *
1003  * grow   => just update size value
1004  * shrink => toss un-needed pages
1005  *
1006  * => we assume that the caller has a reference of some sort to the
1007  *	vnode in question so that it will not be yanked out from under
1008  *	us.
1009  *
1010  * called from:
1011  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
1012  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
1013  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
1014  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
1015  *  => union fs: union_newsize
1016  */
1017 
1018 void
1019 uvm_vnp_setsize(vp, newsize)
1020 	struct vnode *vp;
1021 	voff_t newsize;
1022 {
1023 	struct uvm_vnode *uvn = &vp->v_uvm;
1024 	UVMHIST_FUNC("uvm_vnp_setsize"); UVMHIST_CALLED(ubchist);
1025 
1026 	simple_lock(&uvn->u_obj.vmobjlock);
1027 
1028 	UVMHIST_LOG(ubchist, "old 0x%x new 0x%x", uvn->u_size, newsize, 0,0);
1029 
1030 	/*
1031 	 * now check if the size has changed: if we shrink we had better
1032 	 * toss some pages...
1033 	 */
1034 
1035 	if (uvn->u_size > newsize && uvn->u_size != VSIZENOTSET) {
1036 		(void) uvn_flush(&uvn->u_obj, newsize, uvn->u_size, PGO_FREE);
1037 	}
1038 	uvn->u_size = newsize;
1039 	simple_unlock(&uvn->u_obj.vmobjlock);
1040 }
1041 
1042 /*
1043  * uvm_vnp_zerorange:  set a range of bytes in a file to zero.
1044  */
1045 
1046 void
1047 uvm_vnp_zerorange(vp, off, len)
1048 	struct vnode *vp;
1049 	off_t off;
1050 	size_t len;
1051 {
1052         void *win;
1053 
1054         /*
1055          * XXXUBC invent kzero() and use it
1056          */
1057 
1058         while (len) {
1059                 vsize_t bytelen = len;
1060 
1061                 win = ubc_alloc(&vp->v_uvm.u_obj, off, &bytelen, UBC_WRITE);
1062                 memset(win, 0, bytelen);
1063                 ubc_release(win, 0);
1064 
1065                 off += bytelen;
1066                 len -= bytelen;
1067         }
1068 }
1069