xref: /netbsd-src/sys/uvm/uvm_vnode.c (revision 17dd36da8292193180754d5047c0926dbb56818c)
1 /*	$NetBSD: uvm_vnode.c,v 1.48 2001/03/10 22:46:51 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  */
254 static void
255 uvn_detach(uobj)
256 	struct uvm_object *uobj;
257 {
258 	vrele((struct vnode *)uobj);
259 }
260 
261 /*
262  * uvn_releasepg: handled a released page in a uvn
263  *
264  * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
265  *	to dispose of.
266  * => caller must handled PG_WANTED case
267  * => called with page's object locked, pageq's unlocked
268  * => returns TRUE if page's object is still alive, FALSE if we
269  *	killed the page's object.    if we return TRUE, then we
270  *	return with the object locked.
271  * => if (nextpgp != NULL) => we return the next page on the queue, and return
272  *				with the page queues locked [for pagedaemon]
273  * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
274  * => we kill the uvn if it is not referenced and we are suppose to
275  *	kill it ("relkill").
276  */
277 
278 boolean_t
279 uvn_releasepg(pg, nextpgp)
280 	struct vm_page *pg;
281 	struct vm_page **nextpgp;	/* OUT */
282 {
283 	KASSERT(pg->flags & PG_RELEASED);
284 
285 	/*
286 	 * dispose of the page [caller handles PG_WANTED]
287 	 */
288 	pmap_page_protect(pg, VM_PROT_NONE);
289 	uvm_lock_pageq();
290 	if (nextpgp)
291 		*nextpgp = TAILQ_NEXT(pg, pageq);
292 	uvm_pagefree(pg);
293 	if (!nextpgp)
294 		uvm_unlock_pageq();
295 
296 	return (TRUE);
297 }
298 
299 /*
300  * issues to consider:
301  * there are two tailq's in the uvm. structure... one for pending async
302  * i/o and one for "done" async i/o.   to do an async i/o one puts
303  * a buf on the "pending" list (protected by splbio()), starts the
304  * i/o and returns 0.    when the i/o is done, we expect
305  * some sort of "i/o done" function to be called (at splbio(), interrupt
306  * time).   this function should remove the buf from the pending list
307  * and place it on the "done" list and wakeup the daemon.   the daemon
308  * will run at normal spl() and will remove all items from the "done"
309  * list and call the iodone hook for each done request (see uvm_pager.c).
310  *
311  * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
312  *	later collection.
313  * => called with pageq's locked by the daemon.
314  *
315  * general outline:
316  * - "try" to lock object.   if fail, just return (will try again later)
317  * - drop "u_nio" (this req is done!)
318  * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
319  * - get "page" structures (atop?).
320  * - handle "wanted" pages
321  * - handle "released" pages [using pgo_releasepg]
322  *   >>> pgo_releasepg may kill the object
323  * dont forget to look at "object" wanted flag in all cases.
324  */
325 
326 
327 /*
328  * uvn_flush: flush pages out of a uvm object.
329  *
330  * => "stop == 0" means flush all pages at or after "start".
331  * => object should be locked by caller.   we may _unlock_ the object
332  *	if (and only if) we need to clean a page (PGO_CLEANIT), or
333  *	if PGO_SYNCIO is set and there are pages busy.
334  *	we return with the object locked.
335  * => if PGO_CLEANIT or PGO_SYNCIO is set, we may block (due to I/O).
336  *	thus, a caller might want to unlock higher level resources
337  *	(e.g. vm_map) before calling flush.
338  * => if neither PGO_CLEANIT nor PGO_SYNCIO is set, then we will neither
339  *	unlock the object nor block.
340  * => if PGO_ALLPAGES is set, then all pages in the object are valid targets
341  *	for flushing.
342  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
343  *	that new pages are inserted on the tail end of the list.   thus,
344  *	we can make a complete pass through the object in one go by starting
345  *	at the head and working towards the tail (new pages are put in
346  *	front of us).
347  * => NOTE: we are allowed to lock the page queues, so the caller
348  *	must not be holding the lock on them [e.g. pagedaemon had
349  *	better not call us with the queues locked]
350  * => we return TRUE unless we encountered some sort of I/O error
351  *
352  * comment on "cleaning" object and PG_BUSY pages:
353  *	this routine is holding the lock on the object.   the only time
354  *	that it can run into a PG_BUSY page that it does not own is if
355  *	some other process has started I/O on the page (e.g. either
356  *	a pagein, or a pageout).    if the PG_BUSY page is being paged
357  *	in, then it can not be dirty (!PG_CLEAN) because no one has
358  *	had a chance to modify it yet.    if the PG_BUSY page is being
359  *	paged out then it means that someone else has already started
360  *	cleaning the page for us (how nice!).    in this case, if we
361  *	have syncio specified, then after we make our pass through the
362  *	object we need to wait for the other PG_BUSY pages to clear
363  *	off (i.e. we need to do an iosync).   also note that once a
364  *	page is PG_BUSY it must stay in its object until it is un-busyed.
365  *
366  * note on page traversal:
367  *	we can traverse the pages in an object either by going down the
368  *	linked list in "uobj->memq", or we can go over the address range
369  *	by page doing hash table lookups for each address.    depending
370  *	on how many pages are in the object it may be cheaper to do one
371  *	or the other.   we set "by_list" to true if we are using memq.
372  *	if the cost of a hash lookup was equal to the cost of the list
373  *	traversal we could compare the number of pages in the start->stop
374  *	range to the total number of pages in the object.   however, it
375  *	seems that a hash table lookup is more expensive than the linked
376  *	list traversal, so we multiply the number of pages in the
377  *	start->stop range by a penalty which we define below.
378  */
379 
380 #define UVN_HASH_PENALTY 4	/* XXX: a guess */
381 
382 static boolean_t
383 uvn_flush(uobj, start, stop, flags)
384 	struct uvm_object *uobj;
385 	voff_t start, stop;
386 	int flags;
387 {
388 	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;
389 	struct vnode *vp = (struct vnode *)uobj;
390 	struct vm_page *pp, *ppnext, *ptmp;
391 	struct vm_page *pps[256], **ppsp;
392 	int s;
393 	int npages, result, lcv;
394 	boolean_t retval, need_iosync, by_list, needs_clean, all, wasclean;
395 	boolean_t async = (flags & PGO_SYNCIO) == 0;
396 	voff_t curoff;
397 	u_short pp_version;
398 	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
399 	UVMHIST_LOG(maphist, "uobj %p start 0x%x stop 0x%x flags 0x%x",
400 		    uobj, start, stop, flags);
401 	KASSERT(flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));
402 
403 	if (uobj->uo_npages == 0) {
404 		if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL &&
405 		    (vp->v_flag & VONWORKLST)) {
406 			vp->v_flag &= ~VONWORKLST;
407 			LIST_REMOVE(vp, v_synclist);
408 		}
409 		return TRUE;
410 	}
411 
412 #ifdef DEBUG
413 	if (uvn->u_size == VSIZENOTSET) {
414 		printf("uvn_flush: size not set vp %p\n", uvn);
415 		vprint("uvn_flush VSIZENOTSET", vp);
416 		flags |= PGO_ALLPAGES;
417 	}
418 #endif
419 
420 	/*
421 	 * get init vals and determine how we are going to traverse object
422 	 */
423 
424 	if (stop == 0) {
425 		stop = trunc_page(LLONG_MAX);
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 		all = FALSE;
438 		by_list = (uobj->uo_npages <=
439 		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
440 	}
441 
442 	UVMHIST_LOG(maphist,
443 	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
444 	    start, stop, by_list, flags);
445 
446 	/*
447 	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
448 	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
449 	 * is wrong it will only prevent us from clustering... it won't break
450 	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
451 	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
452 	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
453 	 * bit is always up to date since there are no mappings).
454 	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
455 	 */
456 
457 	if ((flags & PGO_CLEANIT) != 0 &&
458 	    uobj->pgops->pgo_mk_pcluster != NULL) {
459 		if (by_list) {
460 			TAILQ_FOREACH(pp, &uobj->memq, listq) {
461 				if (!all &&
462 				    (pp->offset < start || pp->offset >= stop))
463 					continue;
464 				pp->flags &= ~PG_CLEANCHK;
465 			}
466 
467 		} else {   /* by hash */
468 			for (curoff = start ; curoff < stop;
469 			    curoff += PAGE_SIZE) {
470 				pp = uvm_pagelookup(uobj, curoff);
471 				if (pp)
472 					pp->flags &= ~PG_CLEANCHK;
473 			}
474 		}
475 	}
476 
477 	/*
478 	 * now do it.   note: we must update ppnext in body of loop or we
479 	 * will get stuck.  we need to use ppnext because we may free "pp"
480 	 * before doing the next loop.
481 	 */
482 
483 	if (by_list) {
484 		pp = TAILQ_FIRST(&uobj->memq);
485 	} else {
486 		curoff = start;
487 		pp = uvm_pagelookup(uobj, curoff);
488 	}
489 
490 	ppnext = NULL;
491 	ppsp = NULL;
492 	uvm_lock_pageq();
493 
494 	/* locked: both page queues and uobj */
495 	for ( ; (by_list && pp != NULL) ||
496 		      (!by_list && curoff < stop) ; pp = ppnext) {
497 		if (by_list) {
498 			if (!all &&
499 			    (pp->offset < start || pp->offset >= stop)) {
500 				ppnext = TAILQ_NEXT(pp, listq);
501 				continue;
502 			}
503 		} else {
504 			curoff += PAGE_SIZE;
505 			if (pp == NULL) {
506 				if (curoff < stop)
507 					ppnext = uvm_pagelookup(uobj, curoff);
508 				continue;
509 			}
510 		}
511 
512 		/*
513 		 * handle case where we do not need to clean page (either
514 		 * because we are not clean or because page is not dirty or
515 		 * is busy):
516 		 *
517 		 * NOTE: we are allowed to deactivate a non-wired active
518 		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
519 		 * queue it must stay put until it is !PG_BUSY (so as not to
520 		 * confuse pagedaemon).
521 		 */
522 
523 		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
524 			needs_clean = FALSE;
525 			if (!async)
526 				need_iosync = TRUE;
527 		} else {
528 
529 			/*
530 			 * freeing: nuke all mappings so we can sync
531 			 * PG_CLEAN bit with no race
532 			 */
533 			if ((pp->flags & PG_CLEAN) != 0 &&
534 			    (flags & PGO_FREE) != 0 &&
535 			    /* XXX ACTIVE|INACTIVE test unnecessary? */
536 			    (pp->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) != 0)
537 				pmap_page_protect(pp, VM_PROT_NONE);
538 			if ((pp->flags & PG_CLEAN) != 0 &&
539 			    pmap_is_modified(pp))
540 				pp->flags &= ~(PG_CLEAN);
541 			pp->flags |= PG_CLEANCHK;
542 			needs_clean = ((pp->flags & PG_CLEAN) == 0);
543 		}
544 
545 		/*
546 		 * if we don't need a clean... load ppnext and dispose of pp
547 		 */
548 		if (!needs_clean) {
549 			if (by_list)
550 				ppnext = TAILQ_NEXT(pp, listq);
551 			else {
552 				if (curoff < stop)
553 					ppnext = uvm_pagelookup(uobj, curoff);
554 			}
555 
556 			if (flags & PGO_DEACTIVATE) {
557 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
558 				    (pp->flags & PG_BUSY) == 0 &&
559 				    pp->wire_count == 0) {
560 					pmap_clear_reference(pp);
561 					uvm_pagedeactivate(pp);
562 				}
563 
564 			} else if (flags & PGO_FREE) {
565 				if (pp->flags & PG_BUSY) {
566 					pp->flags |= PG_RELEASED;
567 				} else {
568 					pmap_page_protect(pp, VM_PROT_NONE);
569 					uvm_pagefree(pp);
570 				}
571 			}
572 			/* ppnext is valid so we can continue... */
573 			continue;
574 		}
575 
576 		/*
577 		 * pp points to a page in the locked object that we are
578 		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
579 		 * for cleaning (PGO_CLEANIT).  we clean it now.
580 		 *
581 		 * let uvm_pager_put attempted a clustered page out.
582 		 * note: locked: uobj and page queues.
583 		 */
584 
585 		wasclean = FALSE;
586 		pp->flags |= PG_BUSY;	/* we 'own' page now */
587 		UVM_PAGE_OWN(pp, "uvn_flush");
588 		pmap_page_protect(pp, VM_PROT_READ);
589 		pp_version = pp->version;
590 		ppsp = pps;
591 		npages = sizeof(pps) / sizeof(struct vm_page *);
592 
593 		/* locked: page queues, uobj */
594 		result = uvm_pager_put(uobj, pp, &ppsp, &npages,
595 				       flags | PGO_DOACTCLUST, start, stop);
596 		/* unlocked: page queues, uobj */
597 
598 		/*
599 		 * at this point nothing is locked.   if we did an async I/O
600 		 * it is remotely possible for the async i/o to complete and
601 		 * the page "pp" be freed or what not before we get a chance
602 		 * to relock the object.   in order to detect this, we have
603 		 * saved the version number of the page in "pp_version".
604 		 */
605 
606 		/* relock! */
607 		simple_lock(&uobj->vmobjlock);
608 		uvm_lock_pageq();
609 
610 		/*
611 		 * the cleaning operation is now done.  finish up.  note that
612 		 * on error uvm_pager_put drops the cluster for us.
613 		 * on success uvm_pager_put returns the cluster to us in
614 		 * ppsp/npages.
615 		 */
616 
617 		/*
618 		 * for pending async i/o if we are not deactivating/freeing
619 		 * we can move on to the next page.
620 		 */
621 
622 		if (result == 0 && async &&
623 		    (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
624 
625 			/*
626 			 * no per-page ops: refresh ppnext and continue
627 			 */
628 			if (by_list) {
629 				if (pp->version == pp_version)
630 					ppnext = TAILQ_NEXT(pp, listq);
631 				else
632 					ppnext = TAILQ_FIRST(&uobj->memq);
633 			} else {
634 				if (curoff < stop)
635 					ppnext = uvm_pagelookup(uobj, curoff);
636 			}
637 			continue;
638 		}
639 
640 		/*
641 		 * need to look at each page of the I/O operation.  we defer
642 		 * processing "pp" until the last trip through this "for" loop
643 		 * so that we can load "ppnext" for the main loop after we
644 		 * play with the cluster pages [thus the "npages + 1" in the
645 		 * loop below].
646 		 */
647 
648 		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {
649 
650 			/*
651 			 * handle ppnext for outside loop, and saving pp
652 			 * until the end.
653 			 */
654 			if (lcv < npages) {
655 				if (ppsp[lcv] == pp)
656 					continue; /* skip pp until the end */
657 				ptmp = ppsp[lcv];
658 			} else {
659 				ptmp = pp;
660 
661 				/* set up next page for outer loop */
662 				if (by_list) {
663 					if (pp->version == pp_version)
664 						ppnext = TAILQ_NEXT(pp, listq);
665 					else
666 						ppnext = TAILQ_FIRST(
667 						    &uobj->memq);
668 				} else {
669 					if (curoff < stop)
670 						ppnext = uvm_pagelookup(uobj,
671 						    curoff);
672 				}
673 			}
674 
675 			/*
676 			 * verify the page wasn't moved while obj was
677 			 * unlocked
678 			 */
679 			if (result == 0 && async && ptmp->uobject != uobj)
680 				continue;
681 
682 			/*
683 			 * unbusy the page if I/O is done.   note that for
684 			 * async I/O it is possible that the I/O op
685 			 * finished before we relocked the object (in
686 			 * which case the page is no longer busy).
687 			 */
688 
689 			if (result != 0 || !async) {
690 				if (ptmp->flags & PG_WANTED) {
691 					/* still holding object lock */
692 					wakeup(ptmp);
693 				}
694 				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
695 				UVM_PAGE_OWN(ptmp, NULL);
696 				if (ptmp->flags & PG_RELEASED) {
697 					uvm_unlock_pageq();
698 					if (!uvn_releasepg(ptmp, NULL)) {
699 						UVMHIST_LOG(maphist,
700 							    "released %p",
701 							    ptmp, 0,0,0);
702 						return (TRUE);
703 					}
704 					uvm_lock_pageq();
705 					continue;
706 				} else {
707 					if ((flags & PGO_WEAK) == 0 &&
708 					    !(result == EIO &&
709 					      curproc == uvm.pagedaemon_proc)) {
710 						ptmp->flags |=
711 							(PG_CLEAN|PG_CLEANCHK);
712 						if ((flags & PGO_FREE) == 0) {
713 							pmap_clear_modify(ptmp);
714 						}
715 					}
716 				}
717 			}
718 
719 			/*
720 			 * dispose of page
721 			 */
722 
723 			if (flags & PGO_DEACTIVATE) {
724 				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
725 				    (pp->flags & PG_BUSY) == 0 &&
726 				    pp->wire_count == 0) {
727 					pmap_clear_reference(ptmp);
728 					uvm_pagedeactivate(ptmp);
729 				}
730 			} else if (flags & PGO_FREE) {
731 				if (result == 0 && async) {
732 					if ((ptmp->flags & PG_BUSY) != 0)
733 						/* signal for i/o done */
734 						ptmp->flags |= PG_RELEASED;
735 				} else {
736 					if (result != 0) {
737 						printf("uvn_flush: obj=%p, "
738 						   "offset=0x%llx.  error %d\n",
739 						    pp->uobject,
740 						    (long long)pp->offset,
741 						    result);
742 						printf("uvn_flush: WARNING: "
743 						    "changes to page may be "
744 						    "lost!\n");
745 						retval = FALSE;
746 					}
747 					pmap_page_protect(ptmp, VM_PROT_NONE);
748 					uvm_pagefree(ptmp);
749 				}
750 			}
751 		}		/* end of "lcv" for loop */
752 	}		/* end of "pp" for loop */
753 
754 	uvm_unlock_pageq();
755 	if ((flags & PGO_CLEANIT) && all && wasclean &&
756 	    LIST_FIRST(&vp->v_dirtyblkhd) == NULL &&
757 	    (vp->v_flag & VONWORKLST)) {
758 		vp->v_flag &= ~VONWORKLST;
759 		LIST_REMOVE(vp, v_synclist);
760 	}
761 	if (need_iosync) {
762 		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);
763 
764 		/*
765 		 * XXX this doesn't use the new two-flag scheme,
766 		 * but to use that, all i/o initiators will have to change.
767 		 */
768 
769 		s = splbio();
770 		while (vp->v_numoutput != 0) {
771 			UVMHIST_LOG(ubchist, "waiting for vp %p num %d",
772 				    vp, vp->v_numoutput,0,0);
773 
774 			vp->v_flag |= VBWAIT;
775 			UVM_UNLOCK_AND_WAIT(&vp->v_numoutput,
776 					    &uvn->u_obj.vmobjlock,
777 					    FALSE, "uvn_flush",0);
778 			simple_lock(&uvn->u_obj.vmobjlock);
779 		}
780 		splx(s);
781 	}
782 
783 	/* return, with object locked! */
784 	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
785 	return(retval);
786 }
787 
788 /*
789  * uvn_cluster
790  *
791  * we are about to do I/O in an object at offset.   this function is called
792  * to establish a range of offsets around "offset" in which we can cluster
793  * I/O.
794  *
795  * - currently doesn't matter if obj locked or not.
796  */
797 
798 static void
799 uvn_cluster(uobj, offset, loffset, hoffset)
800 	struct uvm_object *uobj;
801 	voff_t offset;
802 	voff_t *loffset, *hoffset; /* OUT */
803 {
804 	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;
805 
806 	*loffset = offset;
807 	*hoffset = MIN(offset + MAXBSIZE, round_page(uvn->u_size));
808 }
809 
810 /*
811  * uvn_put: flush page data to backing store.
812  *
813  * => object must be locked!   we will _unlock_ it before starting I/O.
814  * => flags: PGO_SYNCIO -- use sync. I/O
815  * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
816  */
817 
818 static int
819 uvn_put(uobj, pps, npages, flags)
820 	struct uvm_object *uobj;
821 	struct vm_page **pps;
822 	int npages, flags;
823 {
824 	struct vnode *vp = (struct vnode *)uobj;
825 	int error;
826 
827 	error = VOP_PUTPAGES(vp, pps, npages, flags, NULL);
828 	return error;
829 }
830 
831 
832 /*
833  * uvn_get: get pages (synchronously) from backing store
834  *
835  * => prefer map unlocked (not required)
836  * => object must be locked!  we will _unlock_ it before starting any I/O.
837  * => flags: PGO_ALLPAGES: get all of the pages
838  *           PGO_LOCKED: fault data structures are locked
839  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
840  * => NOTE: caller must check for released pages!!
841  */
842 
843 static int
844 uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
845 	struct uvm_object *uobj;
846 	voff_t offset;
847 	struct vm_page **pps;		/* IN/OUT */
848 	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
849 	int centeridx;
850 	vm_prot_t access_type;
851 	int advice, flags;
852 {
853 	struct vnode *vp = (struct vnode *)uobj;
854 	int error;
855 	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(ubchist);
856 
857 	UVMHIST_LOG(ubchist, "vp %p off 0x%x", vp, (int)offset, 0,0);
858 	error = VOP_GETPAGES(vp, offset, pps, npagesp, centeridx,
859 			     access_type, advice, flags);
860 	return error;
861 }
862 
863 
864 /*
865  * uvn_findpages:
866  * return the page for the uobj and offset requested, allocating if needed.
867  * => uobj must be locked.
868  * => returned page will be BUSY.
869  */
870 
871 void
872 uvn_findpages(uobj, offset, npagesp, pps, flags)
873 	struct uvm_object *uobj;
874 	voff_t offset;
875 	int *npagesp;
876 	struct vm_page **pps;
877 	int flags;
878 {
879 	int i, rv, npages;
880 
881 	rv = 0;
882 	npages = *npagesp;
883 	for (i = 0; i < npages; i++, offset += PAGE_SIZE) {
884 		rv += uvn_findpage(uobj, offset, &pps[i], flags);
885 	}
886 	*npagesp = rv;
887 }
888 
889 static int
890 uvn_findpage(uobj, offset, pgp, flags)
891 	struct uvm_object *uobj;
892 	voff_t offset;
893 	struct vm_page **pgp;
894 	int flags;
895 {
896 	struct vm_page *pg;
897 	UVMHIST_FUNC("uvn_findpage"); UVMHIST_CALLED(ubchist);
898 	UVMHIST_LOG(ubchist, "vp %p off 0x%lx", uobj, offset,0,0);
899 
900 	if (*pgp != NULL) {
901 		UVMHIST_LOG(ubchist, "dontcare", 0,0,0,0);
902 		return 0;
903 	}
904 	for (;;) {
905 		/* look for an existing page */
906 		pg = uvm_pagelookup(uobj, offset);
907 
908 		/* nope?   allocate one now */
909 		if (pg == NULL) {
910 			if (flags & UFP_NOALLOC) {
911 				UVMHIST_LOG(ubchist, "noalloc", 0,0,0,0);
912 				return 0;
913 			}
914 			pg = uvm_pagealloc(uobj, offset, NULL, 0);
915 			if (pg == NULL) {
916 				if (flags & UFP_NOWAIT) {
917 					UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
918 					return 0;
919 				}
920 				simple_unlock(&uobj->vmobjlock);
921 				uvm_wait("uvn_fp1");
922 				simple_lock(&uobj->vmobjlock);
923 				continue;
924 			}
925 			if (UVM_OBJ_IS_VTEXT(uobj)) {
926 				uvmexp.vtextpages++;
927 			} else {
928 				uvmexp.vnodepages++;
929 			}
930 			UVMHIST_LOG(ubchist, "alloced",0,0,0,0);
931 			break;
932 		} else if (flags & UFP_NOCACHE) {
933 			UVMHIST_LOG(ubchist, "nocache",0,0,0,0);
934 			return 0;
935 		}
936 
937 		/* page is there, see if we need to wait on it */
938 		if ((pg->flags & (PG_BUSY|PG_RELEASED)) != 0) {
939 			if (flags & UFP_NOWAIT) {
940 				UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
941 				return 0;
942 			}
943 			pg->flags |= PG_WANTED;
944 			UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0,
945 					    "uvn_fp2", 0);
946 			simple_lock(&uobj->vmobjlock);
947 			continue;
948 		}
949 
950 		/* skip PG_RDONLY pages if requested */
951 		if ((flags & UFP_NORDONLY) && (pg->flags & PG_RDONLY)) {
952 			UVMHIST_LOG(ubchist, "nordonly",0,0,0,0);
953 			return 0;
954 		}
955 
956 		/* mark the page BUSY and we're done. */
957 		pg->flags |= PG_BUSY;
958 		UVM_PAGE_OWN(pg, "uvn_findpage");
959 		UVMHIST_LOG(ubchist, "found",0,0,0,0);
960 		break;
961 	}
962 	*pgp = pg;
963 	return 1;
964 }
965 
966 /*
967  * uvm_vnp_setsize: grow or shrink a vnode uvn
968  *
969  * grow   => just update size value
970  * shrink => toss un-needed pages
971  *
972  * => we assume that the caller has a reference of some sort to the
973  *	vnode in question so that it will not be yanked out from under
974  *	us.
975  *
976  * called from:
977  *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
978  *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
979  *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
980  *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
981  *  => union fs: union_newsize
982  */
983 
984 void
985 uvm_vnp_setsize(vp, newsize)
986 	struct vnode *vp;
987 	voff_t newsize;
988 {
989 	struct uvm_vnode *uvn = &vp->v_uvm;
990 	voff_t pgend = round_page(newsize);
991 	UVMHIST_FUNC("uvm_vnp_setsize"); UVMHIST_CALLED(ubchist);
992 
993 	simple_lock(&uvn->u_obj.vmobjlock);
994 
995 	UVMHIST_LOG(ubchist, "old 0x%x new 0x%x", uvn->u_size, newsize, 0,0);
996 
997 	/*
998 	 * now check if the size has changed: if we shrink we had better
999 	 * toss some pages...
1000 	 */
1001 
1002 	if (uvn->u_size > pgend && uvn->u_size != VSIZENOTSET) {
1003 		(void) uvn_flush(&uvn->u_obj, pgend, 0, PGO_FREE);
1004 	}
1005 	uvn->u_size = newsize;
1006 	simple_unlock(&uvn->u_obj.vmobjlock);
1007 }
1008 
1009 /*
1010  * uvm_vnp_zerorange:  set a range of bytes in a file to zero.
1011  */
1012 
1013 void
1014 uvm_vnp_zerorange(vp, off, len)
1015 	struct vnode *vp;
1016 	off_t off;
1017 	size_t len;
1018 {
1019         void *win;
1020 
1021         /*
1022          * XXXUBC invent kzero() and use it
1023          */
1024 
1025         while (len) {
1026                 vsize_t bytelen = len;
1027 
1028                 win = ubc_alloc(&vp->v_uvm.u_obj, off, &bytelen, UBC_WRITE);
1029                 memset(win, 0, bytelen);
1030                 ubc_release(win, 0);
1031 
1032                 off += bytelen;
1033                 len -= bytelen;
1034         }
1035 }
1036