xref: /netbsd-src/sys/uvm/uvm_pager.c (revision bada23909e740596d0a3785a73bd3583a9807fb8)
1 /*	$NetBSD: uvm_pager.c,v 1.14 1999/01/22 08:00:35 chs Exp $	*/
2 
3 /*
4  * XXXCDC: "ROUGH DRAFT" QUALITY UVM PRE-RELEASE FILE!
5  *	   >>>USE AT YOUR OWN RISK, WORK IS NOT FINISHED<<<
6  */
7 /*
8  *
9  * Copyright (c) 1997 Charles D. Cranor and Washington University.
10  * All rights reserved.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *      This product includes software developed by Charles D. Cranor and
23  *      Washington University.
24  * 4. The name of the author may not be used to endorse or promote products
25  *    derived from this software without specific prior written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
28  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
29  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
30  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
31  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
32  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
33  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
34  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
35  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
36  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37  *
38  * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
39  */
40 
41 #include "opt_uvmhist.h"
42 #include "opt_pmap_new.h"
43 
44 /*
45  * uvm_pager.c: generic functions used to assist the pagers.
46  */
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/proc.h>
51 #include <sys/malloc.h>
52 
53 #include <vm/vm.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_kern.h>
56 
57 #define UVM_PAGER
58 #include <uvm/uvm.h>
59 
60 /*
61  * list of uvm pagers in the system
62  */
63 
64 extern struct uvm_pagerops aobj_pager;
65 extern struct uvm_pagerops uvm_deviceops;
66 extern struct uvm_pagerops uvm_vnodeops;
67 
68 struct uvm_pagerops *uvmpagerops[] = {
69 	&aobj_pager,
70 	&uvm_deviceops,
71 	&uvm_vnodeops,
72 };
73 
74 /*
75  * the pager map: provides KVA for I/O
76  */
77 
78 #define PAGER_MAP_SIZE       (4 * 1024 * 1024)
79 vm_map_t pager_map;		/* XXX */
80 simple_lock_data_t pager_map_wanted_lock;
81 boolean_t pager_map_wanted;	/* locked by pager map */
82 
83 
84 /*
85  * uvm_pager_init: init pagers (at boot time)
86  */
87 
88 void
89 uvm_pager_init()
90 {
91 	int lcv;
92 
93 	/*
94 	 * init pager map
95 	 */
96 
97 	 pager_map = uvm_km_suballoc(kernel_map, &uvm.pager_sva, &uvm.pager_eva,
98 	 			PAGER_MAP_SIZE, FALSE, FALSE, NULL);
99 	 simple_lock_init(&pager_map_wanted_lock);
100 	 pager_map_wanted = FALSE;
101 
102 	/*
103 	 * init ASYNC I/O queue
104 	 */
105 
106 	TAILQ_INIT(&uvm.aio_done);
107 
108 	/*
109 	 * call pager init functions
110 	 */
111 	for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
112 	    lcv++) {
113 		if (uvmpagerops[lcv]->pgo_init)
114 			uvmpagerops[lcv]->pgo_init();
115 	}
116 }
117 
118 /*
119  * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
120  *
121  * we basically just map in a blank map entry to reserve the space in the
122  * map and then use pmap_enter() to put the mappings in by hand.
123  */
124 
125 vaddr_t
126 uvm_pagermapin(pps, npages, aiop, waitf)
127 	struct vm_page **pps;
128 	int npages;
129 	struct uvm_aiodesc **aiop;	/* OUT */
130 	int waitf;
131 {
132 	vsize_t size;
133 	vaddr_t kva;
134 	struct uvm_aiodesc *aio;
135 #if !defined(PMAP_NEW)
136 	vaddr_t cva;
137 	struct vm_page *pp;
138 #endif
139 	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
140 
141 	UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d, aiop=0x%x, waitf=%d)",
142 	      pps, npages, aiop, waitf);
143 
144 ReStart:
145 	if (aiop) {
146 		MALLOC(aio, struct uvm_aiodesc *, sizeof(*aio), M_TEMP, waitf);
147 		if (aio == NULL)
148 			return(0);
149 		*aiop = aio;
150 	} else {
151 		aio = NULL;
152 	}
153 
154 	size = npages << PAGE_SHIFT;
155 	kva = NULL;			/* let system choose VA */
156 
157 	if (uvm_map(pager_map, &kva, size, NULL,
158 	      UVM_UNKNOWN_OFFSET, UVM_FLAG_NOMERGE) != KERN_SUCCESS) {
159 		if (waitf == M_NOWAIT) {
160 			if (aio)
161 				FREE(aio, M_TEMP);
162 			UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
163 			return(NULL);
164 		}
165 		simple_lock(&pager_map_wanted_lock);
166 		pager_map_wanted = TRUE;
167 		UVMHIST_LOG(maphist, "  SLEEPING on pager_map",0,0,0,0);
168 		UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, FALSE,
169 		    "pager_map",0);
170 		goto ReStart;
171 	}
172 
173 #if defined(PMAP_NEW)
174 	/*
175 	 * XXX: (ab)using the pmap module to store state info for us.
176 	 * (pmap stores the PAs... we fetch them back later and convert back
177 	 * to pages with PHYS_TO_VM_PAGE).
178 	 */
179 	pmap_kenter_pgs(kva, pps, npages);
180 
181 #else /* PMAP_NEW */
182 
183 	/* got it */
184 	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
185 		pp = *pps++;
186 #ifdef DEBUG
187 		if ((pp->flags & PG_BUSY) == 0)
188 			panic("uvm_pagermapin: page not busy");
189 #endif
190 
191 		pmap_enter(vm_map_pmap(pager_map), cva, VM_PAGE_TO_PHYS(pp),
192 		    VM_PROT_DEFAULT, TRUE);
193 	}
194 
195 #endif /* PMAP_NEW */
196 
197 	UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
198 	return(kva);
199 }
200 
201 /*
202  * uvm_pagermapout: remove pager_map mapping
203  *
204  * we remove our mappings by hand and then remove the mapping (waking
205  * up anyone wanting space).
206  */
207 
208 void
209 uvm_pagermapout(kva, npages)
210 	vaddr_t kva;
211 	int npages;
212 {
213 	vsize_t size = npages << PAGE_SHIFT;
214 	vm_map_entry_t entries;
215 	UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
216 
217 	UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
218 
219 	/*
220 	 * duplicate uvm_unmap, but add in pager_map_wanted handling.
221 	 */
222 
223 	vm_map_lock(pager_map);
224 	(void) uvm_unmap_remove(pager_map, kva, kva + size, &entries);
225 	simple_lock(&pager_map_wanted_lock);
226 	if (pager_map_wanted) {
227 		pager_map_wanted = FALSE;
228 		wakeup(pager_map);
229 	}
230 	simple_unlock(&pager_map_wanted_lock);
231 	vm_map_unlock(pager_map);
232 	if (entries)
233 		uvm_unmap_detach(entries, 0);
234 
235 	UVMHIST_LOG(maphist,"<- done",0,0,0,0);
236 }
237 
238 /*
239  * uvm_mk_pcluster
240  *
241  * generic "make 'pager put' cluster" function.  a pager can either
242  * [1] set pgo_mk_pcluster to NULL (never cluster), [2] set it to this
243  * generic function, or [3] set it to a pager specific function.
244  *
245  * => caller must lock object _and_ pagequeues (since we need to look
246  *    at active vs. inactive bits, etc.)
247  * => caller must make center page busy and write-protect it
248  * => we mark all cluster pages busy for the caller
249  * => the caller must unbusy all pages (and check wanted/released
250  *    status if it drops the object lock)
251  * => flags:
252  *      PGO_ALLPAGES:  all pages in object are valid targets
253  *      !PGO_ALLPAGES: use "lo" and "hi" to limit range of cluster
254  *      PGO_DOACTCLUST: include active pages in cluster.
255  *        NOTE: the caller should clear PG_CLEANCHK bits if PGO_DOACTCLUST.
256  *              PG_CLEANCHK is only a hint, but clearing will help reduce
257  *		the number of calls we make to the pmap layer.
258  */
259 
260 struct vm_page **
261 uvm_mk_pcluster(uobj, pps, npages, center, flags, mlo, mhi)
262 	struct uvm_object *uobj;	/* IN */
263 	struct vm_page **pps, *center;  /* IN/OUT, IN */
264 	int *npages, flags;		/* IN/OUT, IN */
265 	vaddr_t mlo, mhi;		/* IN (if !PGO_ALLPAGES) */
266 {
267 	struct vm_page **ppsp, *pclust;
268 	vaddr_t lo, hi, curoff;
269 	int center_idx, forward;
270 	UVMHIST_FUNC("uvm_mk_pcluster"); UVMHIST_CALLED(maphist);
271 
272 	/*
273 	 * center page should already be busy and write protected.  XXX:
274 	 * suppose page is wired?  if we lock, then a process could
275 	 * fault/block on it.  if we don't lock, a process could write the
276 	 * pages in the middle of an I/O.  (consider an msync()).  let's
277 	 * lock it for now (better to delay than corrupt data?).
278 	 */
279 
280 	/*
281 	 * get cluster boundaries, check sanity, and apply our limits as well.
282 	 */
283 
284 	uobj->pgops->pgo_cluster(uobj, center->offset, &lo, &hi);
285 	if ((flags & PGO_ALLPAGES) == 0) {
286 		if (lo < mlo)
287 			lo = mlo;
288 		if (hi > mhi)
289 			hi = mhi;
290 	}
291 	if ((hi - lo) >> PAGE_SHIFT > *npages) {  /* pps too small, bail out! */
292 #ifdef DIAGNOSTIC
293 	    printf("uvm_mk_pcluster: provided page array too small (fixed)\n");
294 #endif
295 		pps[0] = center;
296 		*npages = 1;
297 		return(pps);
298 	}
299 
300 	/*
301 	 * now determine the center and attempt to cluster around the
302 	 * edges
303 	 */
304 
305 	center_idx = (center->offset - lo) >> PAGE_SHIFT;
306 	pps[center_idx] = center;	/* plug in the center page */
307 	ppsp = &pps[center_idx];
308 	*npages = 1;
309 
310 	/*
311 	 * attempt to cluster around the left [backward], and then
312 	 * the right side [forward].
313 	 *
314 	 * note that for inactive pages (pages that have been deactivated)
315 	 * there are no valid mappings and PG_CLEAN should be up to date.
316 	 * [i.e. there is no need to query the pmap with pmap_is_modified
317 	 * since there are no mappings].
318 	 */
319 
320 	for (forward  = 0 ; forward <= 1 ; forward++) {
321 
322 		curoff = center->offset + (forward ? PAGE_SIZE : -PAGE_SIZE);
323 		for ( ;(forward == 0 && curoff >= lo) ||
324 		       (forward && curoff < hi);
325 		      curoff += (forward ? 1 : -1) << PAGE_SHIFT) {
326 
327 			pclust = uvm_pagelookup(uobj, curoff); /* lookup page */
328 			if (pclust == NULL)
329 				break;			/* no page */
330 			/* handle active pages */
331 			/* NOTE: inactive pages don't have pmap mappings */
332 			if ((pclust->pqflags & PQ_INACTIVE) == 0) {
333 				if ((flags & PGO_DOACTCLUST) == 0)
334 					/* dont want mapped pages at all */
335 					break;
336 
337 				/* make sure "clean" bit is sync'd */
338 				if ((pclust->flags & PG_CLEANCHK) == 0) {
339 					if ((pclust->flags & (PG_CLEAN|PG_BUSY))
340 					   == PG_CLEAN &&
341 					   pmap_is_modified(PMAP_PGARG(pclust)))
342 					pclust->flags &= ~PG_CLEAN;
343 					/* now checked */
344 					pclust->flags |= PG_CLEANCHK;
345 				}
346 			}
347 			/* is page available for cleaning and does it need it */
348 			if ((pclust->flags & (PG_CLEAN|PG_BUSY)) != 0)
349 				break;	/* page is already clean or is busy */
350 
351 			/* yes!   enroll the page in our array */
352 			pclust->flags |= PG_BUSY;		/* busy! */
353 			UVM_PAGE_OWN(pclust, "uvm_mk_pcluster");
354 			/* XXX: protect wired page?   see above comment. */
355 			pmap_page_protect(PMAP_PGARG(pclust), VM_PROT_READ);
356 			if (!forward) {
357 				ppsp--;			/* back up one page */
358 				*ppsp = pclust;
359 			} else {
360 				/* move forward one page */
361 				ppsp[*npages] = pclust;
362 			}
363 			*npages = *npages + 1;
364 		}
365 	}
366 
367 	/*
368 	 * done!  return the cluster array to the caller!!!
369 	 */
370 
371 	UVMHIST_LOG(maphist, "<- done",0,0,0,0);
372 	return(ppsp);
373 }
374 
375 
376 /*
377  * uvm_shareprot: generic share protect routine
378  *
379  * => caller must lock map entry's map
380  * => caller must lock object pointed to by map entry
381  */
382 
383 void
384 uvm_shareprot(entry, prot)
385 	vm_map_entry_t entry;
386 	vm_prot_t prot;
387 {
388 	struct uvm_object *uobj = entry->object.uvm_obj;
389 	struct vm_page *pp;
390 	vaddr_t start, stop;
391 	UVMHIST_FUNC("uvm_shareprot"); UVMHIST_CALLED(maphist);
392 
393 	if (UVM_ET_ISSUBMAP(entry))
394 		panic("uvm_shareprot: non-object attached");
395 
396 	start = entry->offset;
397 	stop = start + (entry->end - entry->start);
398 
399 	/*
400 	 * traverse list of pages in object.   if page in range, pmap_prot it
401 	 */
402 
403 	for (pp = uobj->memq.tqh_first ; pp != NULL ; pp = pp->listq.tqe_next) {
404 		if (pp->offset >= start && pp->offset < stop)
405 			pmap_page_protect(PMAP_PGARG(pp), prot);
406 	}
407 	UVMHIST_LOG(maphist, "<- done",0,0,0,0);
408 }
409 
410 /*
411  * uvm_pager_put: high level pageout routine
412  *
413  * we want to pageout page "pg" to backing store, clustering if
414  * possible.
415  *
416  * => page queues must be locked by caller
417  * => if page is not swap-backed, then "uobj" points to the object
418  *	backing it.   this object should be locked by the caller.
419  * => if page is swap-backed, then "uobj" should be NULL.
420  * => "pg" should be PG_BUSY (by caller), and !PG_CLEAN
421  *    for swap-backed memory, "pg" can be NULL if there is no page
422  *    of interest [sometimes the case for the pagedaemon]
423  * => "ppsp_ptr" should point to an array of npages vm_page pointers
424  *	for possible cluster building
425  * => flags (first two for non-swap-backed pages)
426  *	PGO_ALLPAGES: all pages in uobj are valid targets
427  *	PGO_DOACTCLUST: include "PQ_ACTIVE" pages as valid targets
428  *	PGO_SYNCIO: do SYNC I/O (no async)
429  *	PGO_PDFREECLUST: pagedaemon: drop cluster on successful I/O
430  * => start/stop: if (uobj && !PGO_ALLPAGES) limit targets to this range
431  *		  if (!uobj) start is the (daddr_t) of the starting swapblk
432  * => return state:
433  *	1. we return the VM_PAGER status code of the pageout
434  *	2. we return with the page queues unlocked
435  *	3. if (uobj != NULL) [!swap_backed] we return with
436  *		uobj locked _only_ if PGO_PDFREECLUST is set
437  *		AND result != VM_PAGER_PEND.   in all other cases
438  *		we return with uobj unlocked.   [this is a hack
439  *		that allows the pagedaemon to save one lock/unlock
440  *		pair in the !swap_backed case since we have to
441  *		lock the uobj to drop the cluster anyway]
442  *	4. on errors we always drop the cluster.   thus, if we return
443  *		!PEND, !OK, then the caller only has to worry about
444  *		un-busying the main page (not the cluster pages).
445  *	5. on success, if !PGO_PDFREECLUST, we return the cluster
446  *		with all pages busy (caller must un-busy and check
447  *		wanted/released flags).
448  */
449 
450 int
451 uvm_pager_put(uobj, pg, ppsp_ptr, npages, flags, start, stop)
452 	struct uvm_object *uobj;	/* IN */
453 	struct vm_page *pg, ***ppsp_ptr;/* IN, IN/OUT */
454 	int *npages;			/* IN/OUT */
455 	int flags;			/* IN */
456 	vaddr_t start, stop;	/* IN, IN */
457 {
458 	int result;
459 	daddr_t swblk;
460 	struct vm_page **ppsp = *ppsp_ptr;
461 
462 	/*
463 	 * note that uobj is null  if we are doing a swap-backed pageout.
464 	 * note that uobj is !null if we are doing normal object pageout.
465 	 * note that the page queues must be locked to cluster.
466 	 */
467 
468 	if (uobj) {	/* if !swap-backed */
469 
470 		/*
471 		 * attempt to build a cluster for pageout using its
472 		 * make-put-cluster function (if it has one).
473 		 */
474 
475 		if (uobj->pgops->pgo_mk_pcluster) {
476 			ppsp = uobj->pgops->pgo_mk_pcluster(uobj, ppsp,
477 			    npages, pg, flags, start, stop);
478 			*ppsp_ptr = ppsp;  /* update caller's pointer */
479 		} else {
480 			ppsp[0] = pg;
481 			*npages = 1;
482 		 }
483 
484 		swblk = 0;		/* XXX: keep gcc happy */
485 
486 	} else {
487 
488 		/*
489 		 * for swap-backed pageout, the caller (the pagedaemon) has
490 		 * already built the cluster for us.   the starting swap
491 		 * block we are writing to has been passed in as "start."
492 		 * "pg" could be NULL if there is no page we are especially
493 		 * interested in (in which case the whole cluster gets dropped
494 		 * in the event of an error or a sync "done").
495 		 */
496 		swblk = (daddr_t) start;
497 		/* ppsp and npages should be ok */
498 	}
499 
500 	/* now that we've clustered we can unlock the page queues */
501 	uvm_unlock_pageq();
502 
503 	/*
504 	 * now attempt the I/O.   if we have a failure and we are
505 	 * clustered, we will drop the cluster and try again.
506 	 */
507 
508 ReTry:
509 	if (uobj) {
510 		/* object is locked */
511 		result = uobj->pgops->pgo_put(uobj, ppsp, *npages,
512 		    flags & PGO_SYNCIO);
513 		/* object is now unlocked */
514 	} else {
515 		/* nothing locked */
516 		result = uvm_swap_put(swblk, ppsp, *npages, flags & PGO_SYNCIO);
517 		/* nothing locked */
518 	}
519 
520 	/*
521 	 * we have attempted the I/O.
522 	 *
523 	 * if the I/O was a success then:
524 	 * 	if !PGO_PDFREECLUST, we return the cluster to the
525 	 *		caller (who must un-busy all pages)
526 	 *	else we un-busy cluster pages for the pagedaemon
527 	 *
528 	 * if I/O is pending (async i/o) then we return the pending code.
529 	 * [in this case the async i/o done function must clean up when
530 	 *  i/o is done...]
531 	 */
532 
533 	if (result == VM_PAGER_PEND || result == VM_PAGER_OK) {
534 		if (result == VM_PAGER_OK && (flags & PGO_PDFREECLUST)) {
535 			/*
536 			 * drop cluster and relock object (only if I/O is
537 			 * not pending)
538 			 */
539 			if (uobj)
540 				/* required for dropcluster */
541 				simple_lock(&uobj->vmobjlock);
542 			if (*npages > 1 || pg == NULL)
543 				uvm_pager_dropcluster(uobj, pg, ppsp, npages,
544 				    PGO_PDFREECLUST, 0);
545 			/* if (uobj): object still locked, as per
546 			 * return-state item #3 */
547 		}
548 		return (result);
549 	}
550 
551 	/*
552 	 * a pager error occured.    if we have clustered, we drop the
553 	 * cluster and try again.
554 	 */
555 
556 	if (*npages > 1 || pg == NULL) {
557 		if (uobj)
558 			simple_lock(&uobj->vmobjlock);
559 		uvm_pager_dropcluster(uobj, pg, ppsp, npages, PGO_REALLOCSWAP,
560 		    swblk);
561 		if (pg != NULL)
562 			goto ReTry;
563 	}
564 
565 	/*
566 	 * a pager error occured (even after dropping the cluster, if there
567 	 * was one).    give up!   the caller only has one page ("pg")
568 	 * to worry about.
569 	 */
570 
571 	if (uobj && (flags & PGO_PDFREECLUST) != 0)
572 		simple_lock(&uobj->vmobjlock);
573 	return(result);
574 }
575 
576 /*
577  * uvm_pager_dropcluster: drop a cluster we have built (because we
578  * got an error, or, if PGO_PDFREECLUST we are un-busying the
579  * cluster pages on behalf of the pagedaemon).
580  *
581  * => uobj, if non-null, is a non-swap-backed object that is
582  *	locked by the caller.   we return with this object still
583  *	locked.
584  * => page queues are not locked
585  * => pg is our page of interest (the one we clustered around, can be null)
586  * => ppsp/npages is our current cluster
587  * => flags: PGO_PDFREECLUST: pageout was a success: un-busy cluster
588  *	pages on behalf of the pagedaemon.
589  *           PGO_REALLOCSWAP: drop previously allocated swap slots for
590  *		clustered swap-backed pages (except for "pg" if !NULL)
591  *		"swblk" is the start of swap alloc (e.g. for ppsp[0])
592  *		[only meaningful if swap-backed (uobj == NULL)]
593  */
594 
595 
596 void uvm_pager_dropcluster(uobj, pg, ppsp, npages, flags, swblk)
597 
598 struct uvm_object *uobj;	/* IN */
599 struct vm_page *pg, **ppsp;	/* IN, IN/OUT */
600 int *npages;			/* IN/OUT */
601 int flags;
602 int swblk;			/* valid if (uobj == NULL && PGO_REALLOCSWAP) */
603 
604 {
605 	int lcv;
606 	boolean_t obj_is_alive;
607 	struct uvm_object *saved_uobj;
608 
609 	/*
610 	 * if we need to reallocate swap space for the cluster we are dropping
611 	 * (true if swap-backed and PGO_REALLOCSWAP) then free the old
612 	 * allocation now.   save a block for "pg" if it is non-NULL.
613 	 *
614 	 * note that we will zap the object's pointer to swap in the "for" loop
615 	 * below...
616 	 */
617 
618 	if (uobj == NULL && (flags & PGO_REALLOCSWAP)) {
619 		if (pg)
620 			uvm_swap_free(swblk + 1, *npages - 1);
621 		else
622 			uvm_swap_free(swblk, *npages);
623 	}
624 
625 	/*
626 	 * drop all pages but "pg"
627 	 */
628 
629 	for (lcv = 0 ; lcv < *npages ; lcv++) {
630 
631 		if (ppsp[lcv] == pg)		/* skip "pg" */
632 			continue;
633 
634 		/*
635 		 * if swap-backed, gain lock on object that owns page.  note
636 		 * that PQ_ANON bit can't change as long as we are holding
637 		 * the PG_BUSY bit (so there is no need to lock the page
638 		 * queues to test it).
639 		 *
640 		 * once we have the lock, dispose of the pointer to swap, if
641 		 * requested
642 		 */
643 		if (!uobj) {
644 			if (ppsp[lcv]->pqflags & PQ_ANON) {
645 				simple_lock(&ppsp[lcv]->uanon->an_lock);
646 				if (flags & PGO_REALLOCSWAP)
647 					  /* zap swap block */
648 					  ppsp[lcv]->uanon->an_swslot = 0;
649 			} else {
650 				simple_lock(&ppsp[lcv]->uobject->vmobjlock);
651 				if (flags & PGO_REALLOCSWAP)
652 					uao_set_swslot(ppsp[lcv]->uobject,
653 					    ppsp[lcv]->offset >> PAGE_SHIFT, 0);
654 			}
655 		}
656 
657 		/* did someone want the page while we had it busy-locked? */
658 		if (ppsp[lcv]->flags & PG_WANTED)
659 			/* still holding obj lock */
660 			thread_wakeup(ppsp[lcv]);
661 
662 		/* if page was released, release it.  otherwise un-busy it */
663 		if (ppsp[lcv]->flags & PG_RELEASED) {
664 
665 			if (ppsp[lcv]->pqflags & PQ_ANON) {
666 				/* so that anfree will free */
667 				ppsp[lcv]->flags &= ~(PG_BUSY);
668 				UVM_PAGE_OWN(ppsp[lcv], NULL);
669 
670 				pmap_page_protect(PMAP_PGARG(ppsp[lcv]),
671 				    VM_PROT_NONE); /* be safe */
672 				simple_unlock(&ppsp[lcv]->uanon->an_lock);
673 				/* kills anon and frees pg */
674 				uvm_anfree(ppsp[lcv]->uanon);
675 
676 				continue;
677 			}
678 
679 			/*
680 			 * pgo_releasepg will dump the page for us
681 			 */
682 
683 #ifdef DIAGNOSTIC
684 			if (ppsp[lcv]->uobject->pgops->pgo_releasepg == NULL)
685 				panic("uvm_pager_dropcluster: no releasepg "
686 				    "function");
687 #endif
688 			saved_uobj = ppsp[lcv]->uobject;
689 			obj_is_alive =
690 			    saved_uobj->pgops->pgo_releasepg(ppsp[lcv], NULL);
691 
692 #ifdef DIAGNOSTIC
693 			/* for normal objects, "pg" is still PG_BUSY by us,
694 			 * so obj can't die */
695 			if (uobj && !obj_is_alive)
696 				panic("uvm_pager_dropcluster: object died "
697 				    "with active page");
698 #endif
699 			/* only unlock the object if it is still alive...  */
700 			if (obj_is_alive && saved_uobj != uobj)
701 				simple_unlock(&saved_uobj->vmobjlock);
702 
703 			/*
704 			 * XXXCDC: suppose uobj died in the pgo_releasepg?
705 			 * how pass that
706 			 * info up to caller.  we are currently ignoring it...
707 			 */
708 
709 			continue;		/* next page */
710 
711 		} else {
712 			ppsp[lcv]->flags &= ~(PG_BUSY|PG_WANTED);
713 			UVM_PAGE_OWN(ppsp[lcv], NULL);
714 		}
715 
716 		/*
717 		 * if we are operating on behalf of the pagedaemon and we
718 		 * had a successful pageout update the page!
719 		 */
720 		if (flags & PGO_PDFREECLUST) {
721 			/* XXX: with PMAP_NEW ref should already be clear,
722 			 * but don't trust! */
723 			pmap_clear_reference(PMAP_PGARG(ppsp[lcv]));
724 			pmap_clear_modify(PMAP_PGARG(ppsp[lcv]));
725 			ppsp[lcv]->flags |= PG_CLEAN;
726 		}
727 
728 		/* if anonymous cluster, unlock object and move on */
729 		if (!uobj) {
730 			if (ppsp[lcv]->pqflags & PQ_ANON)
731 				simple_unlock(&ppsp[lcv]->uanon->an_lock);
732 			else
733 				simple_unlock(&ppsp[lcv]->uobject->vmobjlock);
734 		}
735 
736 	}
737 
738 	/*
739 	 * drop to a cluster of 1 page ("pg") if requested
740 	 */
741 
742 	if (pg && (flags & PGO_PDFREECLUST) == 0) {
743 		/*
744 		 * if we are not a successful pageout, we make a 1 page cluster.
745 		 */
746 		ppsp[0] = pg;
747 		*npages = 1;
748 
749 		/*
750 		 * assign new swap block to new cluster, if anon backed
751 		 */
752 		if (uobj == NULL && (flags & PGO_REALLOCSWAP)) {
753 			if (pg->pqflags & PQ_ANON) {
754 				simple_lock(&pg->uanon->an_lock);
755 				pg->uanon->an_swslot = swblk;	/* reassign */
756 				simple_unlock(&pg->uanon->an_lock);
757 			} else {
758 				simple_lock(&pg->uobject->vmobjlock);
759 				uao_set_swslot(pg->uobject,
760 				    pg->offset >> PAGE_SHIFT, swblk);
761 				simple_unlock(&pg->uobject->vmobjlock);
762 			}
763 		}
764 	}
765 }
766