Home | History | Annotate | Line # | Download | only in uvm
uvm_pager.c revision 1.23
      1 /*	$NetBSD: uvm_pager.c,v 1.23 1999/09/12 01:17:41 chs Exp $	*/
      2 
      3 /*
      4  *
      5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
      6  * All rights reserved.
      7  *
      8  * Redistribution and use in source and binary forms, with or without
      9  * modification, are permitted provided that the following conditions
     10  * are met:
     11  * 1. Redistributions of source code must retain the above copyright
     12  *    notice, this list of conditions and the following disclaimer.
     13  * 2. Redistributions in binary form must reproduce the above copyright
     14  *    notice, this list of conditions and the following disclaimer in the
     15  *    documentation and/or other materials provided with the distribution.
     16  * 3. All advertising materials mentioning features or use of this software
     17  *    must display the following acknowledgement:
     18  *      This product includes software developed by Charles D. Cranor and
     19  *      Washington University.
     20  * 4. The name of the author may not be used to endorse or promote products
     21  *    derived from this software without specific prior written permission.
     22  *
     23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
     24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
     25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
     27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
     28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
     29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
     30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
     31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
     32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     33  *
     34  * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
     35  */
     36 
     37 #include "opt_uvmhist.h"
     38 
     39 /*
     40  * uvm_pager.c: generic functions used to assist the pagers.
     41  */
     42 
     43 #include <sys/param.h>
     44 #include <sys/systm.h>
     45 #include <sys/proc.h>
     46 #include <sys/malloc.h>
     47 
     48 #include <vm/vm.h>
     49 #include <vm/vm_page.h>
     50 #include <vm/vm_kern.h>
     51 
     52 #define UVM_PAGER
     53 #include <uvm/uvm.h>
     54 
     55 /*
     56  * list of uvm pagers in the system
     57  */
     58 
     59 extern struct uvm_pagerops aobj_pager;
     60 extern struct uvm_pagerops uvm_deviceops;
     61 extern struct uvm_pagerops uvm_vnodeops;
     62 
     63 struct uvm_pagerops *uvmpagerops[] = {
     64 	&aobj_pager,
     65 	&uvm_deviceops,
     66 	&uvm_vnodeops,
     67 };
     68 
     69 /*
     70  * the pager map: provides KVA for I/O
     71  */
     72 
     73 #define PAGER_MAP_SIZE       (4 * 1024 * 1024)
     74 vm_map_t pager_map;		/* XXX */
     75 simple_lock_data_t pager_map_wanted_lock;
     76 boolean_t pager_map_wanted;	/* locked by pager map */
     77 
     78 
     79 /*
     80  * uvm_pager_init: init pagers (at boot time)
     81  */
     82 
     83 void
     84 uvm_pager_init()
     85 {
     86 	int lcv;
     87 
     88 	/*
     89 	 * init pager map
     90 	 */
     91 
     92 	 pager_map = uvm_km_suballoc(kernel_map, &uvm.pager_sva, &uvm.pager_eva,
     93 	 			PAGER_MAP_SIZE, 0, FALSE, NULL);
     94 	 simple_lock_init(&pager_map_wanted_lock);
     95 	 pager_map_wanted = FALSE;
     96 
     97 	/*
     98 	 * init ASYNC I/O queue
     99 	 */
    100 
    101 	TAILQ_INIT(&uvm.aio_done);
    102 
    103 	/*
    104 	 * call pager init functions
    105 	 */
    106 	for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
    107 	    lcv++) {
    108 		if (uvmpagerops[lcv]->pgo_init)
    109 			uvmpagerops[lcv]->pgo_init();
    110 	}
    111 }
    112 
    113 /*
    114  * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
    115  *
    116  * we basically just map in a blank map entry to reserve the space in the
    117  * map and then use pmap_enter() to put the mappings in by hand.
    118  *
    119  * XXX It would be nice to know the direction of the I/O, so that we can
    120  * XXX map only what is necessary.
    121  */
    122 
    123 vaddr_t
    124 uvm_pagermapin(pps, npages, aiop, waitf)
    125 	struct vm_page **pps;
    126 	int npages;
    127 	struct uvm_aiodesc **aiop;	/* OUT */
    128 	int waitf;
    129 {
    130 	vsize_t size;
    131 	vaddr_t kva;
    132 	struct uvm_aiodesc *aio;
    133 	vaddr_t cva;
    134 	struct vm_page *pp;
    135 	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
    136 
    137 	UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d, aiop=0x%x, waitf=%d)",
    138 	      pps, npages, aiop, waitf);
    139 
    140 ReStart:
    141 	if (aiop) {
    142 		MALLOC(aio, struct uvm_aiodesc *, sizeof(*aio), M_TEMP, waitf);
    143 		if (aio == NULL)
    144 			return(0);
    145 		*aiop = aio;
    146 	} else {
    147 		aio = NULL;
    148 	}
    149 
    150 	size = npages << PAGE_SHIFT;
    151 	kva = NULL;			/* let system choose VA */
    152 
    153 	if (uvm_map(pager_map, &kva, size, NULL,
    154 	      UVM_UNKNOWN_OFFSET, UVM_FLAG_NOMERGE) != KERN_SUCCESS) {
    155 		if (waitf == M_NOWAIT) {
    156 			if (aio)
    157 				FREE(aio, M_TEMP);
    158 			UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
    159 			return(NULL);
    160 		}
    161 		simple_lock(&pager_map_wanted_lock);
    162 		pager_map_wanted = TRUE;
    163 		UVMHIST_LOG(maphist, "  SLEEPING on pager_map",0,0,0,0);
    164 		UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, FALSE,
    165 		    "pager_map",0);
    166 		goto ReStart;
    167 	}
    168 
    169 	/* got it */
    170 	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
    171 		pp = *pps++;
    172 #ifdef DEBUG
    173 		if ((pp->flags & PG_BUSY) == 0)
    174 			panic("uvm_pagermapin: page not busy");
    175 #endif
    176 
    177 		/*
    178 		 * XXX VM_PROT_DEFAULT includes VM_PROT_EXEC; is that
    179 		 * XXX really necessary?  It could lead to unnecessary
    180 		 * XXX instruction cache flushes.
    181 		 */
    182 		pmap_enter(vm_map_pmap(pager_map), cva, VM_PAGE_TO_PHYS(pp),
    183 		    VM_PROT_DEFAULT, TRUE,
    184 		    VM_PROT_READ | VM_PROT_WRITE);
    185 	}
    186 
    187 	UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
    188 	return(kva);
    189 }
    190 
    191 /*
    192  * uvm_pagermapout: remove pager_map mapping
    193  *
    194  * we remove our mappings by hand and then remove the mapping (waking
    195  * up anyone wanting space).
    196  */
    197 
    198 void
    199 uvm_pagermapout(kva, npages)
    200 	vaddr_t kva;
    201 	int npages;
    202 {
    203 	vsize_t size = npages << PAGE_SHIFT;
    204 	vm_map_entry_t entries;
    205 	UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
    206 
    207 	UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
    208 
    209 	/*
    210 	 * duplicate uvm_unmap, but add in pager_map_wanted handling.
    211 	 */
    212 
    213 	vm_map_lock(pager_map);
    214 	(void) uvm_unmap_remove(pager_map, kva, kva + size, &entries);
    215 	simple_lock(&pager_map_wanted_lock);
    216 	if (pager_map_wanted) {
    217 		pager_map_wanted = FALSE;
    218 		wakeup(pager_map);
    219 	}
    220 	simple_unlock(&pager_map_wanted_lock);
    221 	vm_map_unlock(pager_map);
    222 	if (entries)
    223 		uvm_unmap_detach(entries, 0);
    224 
    225 	UVMHIST_LOG(maphist,"<- done",0,0,0,0);
    226 }
    227 
    228 /*
    229  * uvm_mk_pcluster
    230  *
    231  * generic "make 'pager put' cluster" function.  a pager can either
    232  * [1] set pgo_mk_pcluster to NULL (never cluster), [2] set it to this
    233  * generic function, or [3] set it to a pager specific function.
    234  *
    235  * => caller must lock object _and_ pagequeues (since we need to look
    236  *    at active vs. inactive bits, etc.)
    237  * => caller must make center page busy and write-protect it
    238  * => we mark all cluster pages busy for the caller
    239  * => the caller must unbusy all pages (and check wanted/released
    240  *    status if it drops the object lock)
    241  * => flags:
    242  *      PGO_ALLPAGES:  all pages in object are valid targets
    243  *      !PGO_ALLPAGES: use "lo" and "hi" to limit range of cluster
    244  *      PGO_DOACTCLUST: include active pages in cluster.
    245  *        NOTE: the caller should clear PG_CLEANCHK bits if PGO_DOACTCLUST.
    246  *              PG_CLEANCHK is only a hint, but clearing will help reduce
    247  *		the number of calls we make to the pmap layer.
    248  */
    249 
    250 struct vm_page **
    251 uvm_mk_pcluster(uobj, pps, npages, center, flags, mlo, mhi)
    252 	struct uvm_object *uobj;	/* IN */
    253 	struct vm_page **pps, *center;  /* IN/OUT, IN */
    254 	int *npages, flags;		/* IN/OUT, IN */
    255 	vaddr_t mlo, mhi;		/* IN (if !PGO_ALLPAGES) */
    256 {
    257 	struct vm_page **ppsp, *pclust;
    258 	vaddr_t lo, hi, curoff;
    259 	int center_idx, forward;
    260 	UVMHIST_FUNC("uvm_mk_pcluster"); UVMHIST_CALLED(maphist);
    261 
    262 	/*
    263 	 * center page should already be busy and write protected.  XXX:
    264 	 * suppose page is wired?  if we lock, then a process could
    265 	 * fault/block on it.  if we don't lock, a process could write the
    266 	 * pages in the middle of an I/O.  (consider an msync()).  let's
    267 	 * lock it for now (better to delay than corrupt data?).
    268 	 */
    269 
    270 	/*
    271 	 * get cluster boundaries, check sanity, and apply our limits as well.
    272 	 */
    273 
    274 	uobj->pgops->pgo_cluster(uobj, center->offset, &lo, &hi);
    275 	if ((flags & PGO_ALLPAGES) == 0) {
    276 		if (lo < mlo)
    277 			lo = mlo;
    278 		if (hi > mhi)
    279 			hi = mhi;
    280 	}
    281 	if ((hi - lo) >> PAGE_SHIFT > *npages) {  /* pps too small, bail out! */
    282 #ifdef DIAGNOSTIC
    283 	    printf("uvm_mk_pcluster: provided page array too small (fixed)\n");
    284 #endif
    285 		pps[0] = center;
    286 		*npages = 1;
    287 		return(pps);
    288 	}
    289 
    290 	/*
    291 	 * now determine the center and attempt to cluster around the
    292 	 * edges
    293 	 */
    294 
    295 	center_idx = (center->offset - lo) >> PAGE_SHIFT;
    296 	pps[center_idx] = center;	/* plug in the center page */
    297 	ppsp = &pps[center_idx];
    298 	*npages = 1;
    299 
    300 	/*
    301 	 * attempt to cluster around the left [backward], and then
    302 	 * the right side [forward].
    303 	 *
    304 	 * note that for inactive pages (pages that have been deactivated)
    305 	 * there are no valid mappings and PG_CLEAN should be up to date.
    306 	 * [i.e. there is no need to query the pmap with pmap_is_modified
    307 	 * since there are no mappings].
    308 	 */
    309 
    310 	for (forward  = 0 ; forward <= 1 ; forward++) {
    311 
    312 		curoff = center->offset + (forward ? PAGE_SIZE : -PAGE_SIZE);
    313 		for ( ;(forward == 0 && curoff >= lo) ||
    314 		       (forward && curoff < hi);
    315 		      curoff += (forward ? 1 : -1) << PAGE_SHIFT) {
    316 
    317 			pclust = uvm_pagelookup(uobj, curoff); /* lookup page */
    318 			if (pclust == NULL)
    319 				break;			/* no page */
    320 			/* handle active pages */
    321 			/* NOTE: inactive pages don't have pmap mappings */
    322 			if ((pclust->pqflags & PQ_INACTIVE) == 0) {
    323 				if ((flags & PGO_DOACTCLUST) == 0)
    324 					/* dont want mapped pages at all */
    325 					break;
    326 
    327 				/* make sure "clean" bit is sync'd */
    328 				if ((pclust->flags & PG_CLEANCHK) == 0) {
    329 					if ((pclust->flags & (PG_CLEAN|PG_BUSY))
    330 					   == PG_CLEAN &&
    331 					   pmap_is_modified(pclust))
    332 						pclust->flags &= ~PG_CLEAN;
    333 
    334 					/* now checked */
    335 					pclust->flags |= PG_CLEANCHK;
    336 				}
    337 			}
    338 			/* is page available for cleaning and does it need it */
    339 			if ((pclust->flags & (PG_CLEAN|PG_BUSY)) != 0)
    340 				break;	/* page is already clean or is busy */
    341 
    342 			/* yes!   enroll the page in our array */
    343 			pclust->flags |= PG_BUSY;		/* busy! */
    344 			UVM_PAGE_OWN(pclust, "uvm_mk_pcluster");
    345 			/* XXX: protect wired page?   see above comment. */
    346 			pmap_page_protect(pclust, VM_PROT_READ);
    347 			if (!forward) {
    348 				ppsp--;			/* back up one page */
    349 				*ppsp = pclust;
    350 			} else {
    351 				/* move forward one page */
    352 				ppsp[*npages] = pclust;
    353 			}
    354 			*npages = *npages + 1;
    355 		}
    356 	}
    357 
    358 	/*
    359 	 * done!  return the cluster array to the caller!!!
    360 	 */
    361 
    362 	UVMHIST_LOG(maphist, "<- done",0,0,0,0);
    363 	return(ppsp);
    364 }
    365 
    366 
    367 /*
    368  * uvm_shareprot: generic share protect routine
    369  *
    370  * => caller must lock map entry's map
    371  * => caller must lock object pointed to by map entry
    372  */
    373 
    374 void
    375 uvm_shareprot(entry, prot)
    376 	vm_map_entry_t entry;
    377 	vm_prot_t prot;
    378 {
    379 	struct uvm_object *uobj = entry->object.uvm_obj;
    380 	struct vm_page *pp;
    381 	vaddr_t start, stop;
    382 	UVMHIST_FUNC("uvm_shareprot"); UVMHIST_CALLED(maphist);
    383 
    384 	if (UVM_ET_ISSUBMAP(entry))
    385 		panic("uvm_shareprot: non-object attached");
    386 
    387 	start = entry->offset;
    388 	stop = start + (entry->end - entry->start);
    389 
    390 	/*
    391 	 * traverse list of pages in object.   if page in range, pmap_prot it
    392 	 */
    393 
    394 	for (pp = uobj->memq.tqh_first ; pp != NULL ; pp = pp->listq.tqe_next) {
    395 		if (pp->offset >= start && pp->offset < stop)
    396 			pmap_page_protect(pp, prot);
    397 	}
    398 	UVMHIST_LOG(maphist, "<- done",0,0,0,0);
    399 }
    400 
    401 /*
    402  * uvm_pager_put: high level pageout routine
    403  *
    404  * we want to pageout page "pg" to backing store, clustering if
    405  * possible.
    406  *
    407  * => page queues must be locked by caller
    408  * => if page is not swap-backed, then "uobj" points to the object
    409  *	backing it.   this object should be locked by the caller.
    410  * => if page is swap-backed, then "uobj" should be NULL.
    411  * => "pg" should be PG_BUSY (by caller), and !PG_CLEAN
    412  *    for swap-backed memory, "pg" can be NULL if there is no page
    413  *    of interest [sometimes the case for the pagedaemon]
    414  * => "ppsp_ptr" should point to an array of npages vm_page pointers
    415  *	for possible cluster building
    416  * => flags (first two for non-swap-backed pages)
    417  *	PGO_ALLPAGES: all pages in uobj are valid targets
    418  *	PGO_DOACTCLUST: include "PQ_ACTIVE" pages as valid targets
    419  *	PGO_SYNCIO: do SYNC I/O (no async)
    420  *	PGO_PDFREECLUST: pagedaemon: drop cluster on successful I/O
    421  * => start/stop: if (uobj && !PGO_ALLPAGES) limit targets to this range
    422  *		  if (!uobj) start is the (daddr_t) of the starting swapblk
    423  * => return state:
    424  *	1. we return the VM_PAGER status code of the pageout
    425  *	2. we return with the page queues unlocked
    426  *	3. if (uobj != NULL) [!swap_backed] we return with
    427  *		uobj locked _only_ if PGO_PDFREECLUST is set
    428  *		AND result != VM_PAGER_PEND.   in all other cases
    429  *		we return with uobj unlocked.   [this is a hack
    430  *		that allows the pagedaemon to save one lock/unlock
    431  *		pair in the !swap_backed case since we have to
    432  *		lock the uobj to drop the cluster anyway]
    433  *	4. on errors we always drop the cluster.   thus, if we return
    434  *		!PEND, !OK, then the caller only has to worry about
    435  *		un-busying the main page (not the cluster pages).
    436  *	5. on success, if !PGO_PDFREECLUST, we return the cluster
    437  *		with all pages busy (caller must un-busy and check
    438  *		wanted/released flags).
    439  */
    440 
    441 int
    442 uvm_pager_put(uobj, pg, ppsp_ptr, npages, flags, start, stop)
    443 	struct uvm_object *uobj;	/* IN */
    444 	struct vm_page *pg, ***ppsp_ptr;/* IN, IN/OUT */
    445 	int *npages;			/* IN/OUT */
    446 	int flags;			/* IN */
    447 	vaddr_t start, stop;	/* IN, IN */
    448 {
    449 	int result;
    450 	daddr_t swblk;
    451 	struct vm_page **ppsp = *ppsp_ptr;
    452 
    453 	/*
    454 	 * note that uobj is null  if we are doing a swap-backed pageout.
    455 	 * note that uobj is !null if we are doing normal object pageout.
    456 	 * note that the page queues must be locked to cluster.
    457 	 */
    458 
    459 	if (uobj) {	/* if !swap-backed */
    460 
    461 		/*
    462 		 * attempt to build a cluster for pageout using its
    463 		 * make-put-cluster function (if it has one).
    464 		 */
    465 
    466 		if (uobj->pgops->pgo_mk_pcluster) {
    467 			ppsp = uobj->pgops->pgo_mk_pcluster(uobj, ppsp,
    468 			    npages, pg, flags, start, stop);
    469 			*ppsp_ptr = ppsp;  /* update caller's pointer */
    470 		} else {
    471 			ppsp[0] = pg;
    472 			*npages = 1;
    473 		 }
    474 
    475 		swblk = 0;		/* XXX: keep gcc happy */
    476 
    477 	} else {
    478 
    479 		/*
    480 		 * for swap-backed pageout, the caller (the pagedaemon) has
    481 		 * already built the cluster for us.   the starting swap
    482 		 * block we are writing to has been passed in as "start."
    483 		 * "pg" could be NULL if there is no page we are especially
    484 		 * interested in (in which case the whole cluster gets dropped
    485 		 * in the event of an error or a sync "done").
    486 		 */
    487 		swblk = (daddr_t) start;
    488 		/* ppsp and npages should be ok */
    489 	}
    490 
    491 	/* now that we've clustered we can unlock the page queues */
    492 	uvm_unlock_pageq();
    493 
    494 	/*
    495 	 * now attempt the I/O.   if we have a failure and we are
    496 	 * clustered, we will drop the cluster and try again.
    497 	 */
    498 
    499 ReTry:
    500 	if (uobj) {
    501 		/* object is locked */
    502 		result = uobj->pgops->pgo_put(uobj, ppsp, *npages,
    503 		    flags & PGO_SYNCIO);
    504 		/* object is now unlocked */
    505 	} else {
    506 		/* nothing locked */
    507 		result = uvm_swap_put(swblk, ppsp, *npages, flags & PGO_SYNCIO);
    508 		/* nothing locked */
    509 	}
    510 
    511 	/*
    512 	 * we have attempted the I/O.
    513 	 *
    514 	 * if the I/O was a success then:
    515 	 * 	if !PGO_PDFREECLUST, we return the cluster to the
    516 	 *		caller (who must un-busy all pages)
    517 	 *	else we un-busy cluster pages for the pagedaemon
    518 	 *
    519 	 * if I/O is pending (async i/o) then we return the pending code.
    520 	 * [in this case the async i/o done function must clean up when
    521 	 *  i/o is done...]
    522 	 */
    523 
    524 	if (result == VM_PAGER_PEND || result == VM_PAGER_OK) {
    525 		if (result == VM_PAGER_OK && (flags & PGO_PDFREECLUST)) {
    526 			/*
    527 			 * drop cluster and relock object (only if I/O is
    528 			 * not pending)
    529 			 */
    530 			if (uobj)
    531 				/* required for dropcluster */
    532 				simple_lock(&uobj->vmobjlock);
    533 			if (*npages > 1 || pg == NULL)
    534 				uvm_pager_dropcluster(uobj, pg, ppsp, npages,
    535 				    PGO_PDFREECLUST, 0);
    536 			/* if (uobj): object still locked, as per
    537 			 * return-state item #3 */
    538 		}
    539 		return (result);
    540 	}
    541 
    542 	/*
    543 	 * a pager error occured.    if we have clustered, we drop the
    544 	 * cluster and try again.
    545 	 */
    546 
    547 	if (*npages > 1 || pg == NULL) {
    548 		if (uobj)
    549 			simple_lock(&uobj->vmobjlock);
    550 		uvm_pager_dropcluster(uobj, pg, ppsp, npages, PGO_REALLOCSWAP,
    551 		    swblk);
    552 		if (pg != NULL)
    553 			goto ReTry;
    554 	}
    555 
    556 	/*
    557 	 * a pager error occured (even after dropping the cluster, if there
    558 	 * was one).    give up!   the caller only has one page ("pg")
    559 	 * to worry about.
    560 	 */
    561 
    562 	if (uobj && (flags & PGO_PDFREECLUST) != 0)
    563 		simple_lock(&uobj->vmobjlock);
    564 	return(result);
    565 }
    566 
    567 /*
    568  * uvm_pager_dropcluster: drop a cluster we have built (because we
    569  * got an error, or, if PGO_PDFREECLUST we are un-busying the
    570  * cluster pages on behalf of the pagedaemon).
    571  *
    572  * => uobj, if non-null, is a non-swap-backed object that is
    573  *	locked by the caller.   we return with this object still
    574  *	locked.
    575  * => page queues are not locked
    576  * => pg is our page of interest (the one we clustered around, can be null)
    577  * => ppsp/npages is our current cluster
    578  * => flags: PGO_PDFREECLUST: pageout was a success: un-busy cluster
    579  *	pages on behalf of the pagedaemon.
    580  *           PGO_REALLOCSWAP: drop previously allocated swap slots for
    581  *		clustered swap-backed pages (except for "pg" if !NULL)
    582  *		"swblk" is the start of swap alloc (e.g. for ppsp[0])
    583  *		[only meaningful if swap-backed (uobj == NULL)]
    584  */
    585 
    586 void
    587 uvm_pager_dropcluster(uobj, pg, ppsp, npages, flags, swblk)
    588 	struct uvm_object *uobj;	/* IN */
    589 	struct vm_page *pg, **ppsp;	/* IN, IN/OUT */
    590 	int *npages;			/* IN/OUT */
    591 	int flags;
    592 	int swblk;			/* valid if
    593 					   (uobj == NULL && PGO_REALLOCSWAP) */
    594 {
    595 	int lcv;
    596 	boolean_t obj_is_alive;
    597 	struct uvm_object *saved_uobj;
    598 
    599 	/*
    600 	 * if we need to reallocate swap space for the cluster we are dropping
    601 	 * (true if swap-backed and PGO_REALLOCSWAP) then free the old
    602 	 * allocation now.   save a block for "pg" if it is non-NULL.
    603 	 *
    604 	 * note that we will zap the object's pointer to swap in the "for" loop
    605 	 * below...
    606 	 */
    607 
    608 	if (uobj == NULL && (flags & PGO_REALLOCSWAP)) {
    609 		if (pg)
    610 			uvm_swap_free(swblk + 1, *npages - 1);
    611 		else
    612 			uvm_swap_free(swblk, *npages);
    613 	}
    614 
    615 	/*
    616 	 * drop all pages but "pg"
    617 	 */
    618 
    619 	for (lcv = 0 ; lcv < *npages ; lcv++) {
    620 
    621 		if (ppsp[lcv] == pg)		/* skip "pg" */
    622 			continue;
    623 
    624 		/*
    625 		 * if swap-backed, gain lock on object that owns page.  note
    626 		 * that PQ_ANON bit can't change as long as we are holding
    627 		 * the PG_BUSY bit (so there is no need to lock the page
    628 		 * queues to test it).
    629 		 *
    630 		 * once we have the lock, dispose of the pointer to swap, if
    631 		 * requested
    632 		 */
    633 		if (!uobj) {
    634 			if (ppsp[lcv]->pqflags & PQ_ANON) {
    635 				simple_lock(&ppsp[lcv]->uanon->an_lock);
    636 				if (flags & PGO_REALLOCSWAP)
    637 					  /* zap swap block */
    638 					  ppsp[lcv]->uanon->an_swslot = 0;
    639 			} else {
    640 				simple_lock(&ppsp[lcv]->uobject->vmobjlock);
    641 				if (flags & PGO_REALLOCSWAP)
    642 					uao_set_swslot(ppsp[lcv]->uobject,
    643 					    ppsp[lcv]->offset >> PAGE_SHIFT, 0);
    644 			}
    645 		}
    646 
    647 		/* did someone want the page while we had it busy-locked? */
    648 		if (ppsp[lcv]->flags & PG_WANTED)
    649 			/* still holding obj lock */
    650 			wakeup(ppsp[lcv]);
    651 
    652 		/* if page was released, release it.  otherwise un-busy it */
    653 		if (ppsp[lcv]->flags & PG_RELEASED) {
    654 
    655 			if (ppsp[lcv]->pqflags & PQ_ANON) {
    656 				/* so that anfree will free */
    657 				ppsp[lcv]->flags &= ~(PG_BUSY);
    658 				UVM_PAGE_OWN(ppsp[lcv], NULL);
    659 
    660 				pmap_page_protect(ppsp[lcv], VM_PROT_NONE);
    661 				simple_unlock(&ppsp[lcv]->uanon->an_lock);
    662 				/* kills anon and frees pg */
    663 				uvm_anfree(ppsp[lcv]->uanon);
    664 
    665 				continue;
    666 			}
    667 
    668 			/*
    669 			 * pgo_releasepg will dump the page for us
    670 			 */
    671 
    672 #ifdef DIAGNOSTIC
    673 			if (ppsp[lcv]->uobject->pgops->pgo_releasepg == NULL)
    674 				panic("uvm_pager_dropcluster: no releasepg "
    675 				    "function");
    676 #endif
    677 			saved_uobj = ppsp[lcv]->uobject;
    678 			obj_is_alive =
    679 			    saved_uobj->pgops->pgo_releasepg(ppsp[lcv], NULL);
    680 
    681 #ifdef DIAGNOSTIC
    682 			/* for normal objects, "pg" is still PG_BUSY by us,
    683 			 * so obj can't die */
    684 			if (uobj && !obj_is_alive)
    685 				panic("uvm_pager_dropcluster: object died "
    686 				    "with active page");
    687 #endif
    688 			/* only unlock the object if it is still alive...  */
    689 			if (obj_is_alive && saved_uobj != uobj)
    690 				simple_unlock(&saved_uobj->vmobjlock);
    691 
    692 			/*
    693 			 * XXXCDC: suppose uobj died in the pgo_releasepg?
    694 			 * how pass that
    695 			 * info up to caller.  we are currently ignoring it...
    696 			 */
    697 
    698 			continue;		/* next page */
    699 
    700 		} else {
    701 			ppsp[lcv]->flags &= ~(PG_BUSY|PG_WANTED);
    702 			UVM_PAGE_OWN(ppsp[lcv], NULL);
    703 		}
    704 
    705 		/*
    706 		 * if we are operating on behalf of the pagedaemon and we
    707 		 * had a successful pageout update the page!
    708 		 */
    709 		if (flags & PGO_PDFREECLUST) {
    710 			pmap_clear_reference(ppsp[lcv]);
    711 			pmap_clear_modify(ppsp[lcv]);
    712 			ppsp[lcv]->flags |= PG_CLEAN;
    713 		}
    714 
    715 		/* if anonymous cluster, unlock object and move on */
    716 		if (!uobj) {
    717 			if (ppsp[lcv]->pqflags & PQ_ANON)
    718 				simple_unlock(&ppsp[lcv]->uanon->an_lock);
    719 			else
    720 				simple_unlock(&ppsp[lcv]->uobject->vmobjlock);
    721 		}
    722 
    723 	}
    724 
    725 	/*
    726 	 * drop to a cluster of 1 page ("pg") if requested
    727 	 */
    728 
    729 	if (pg && (flags & PGO_PDFREECLUST) == 0) {
    730 		/*
    731 		 * if we are not a successful pageout, we make a 1 page cluster.
    732 		 */
    733 		ppsp[0] = pg;
    734 		*npages = 1;
    735 
    736 		/*
    737 		 * assign new swap block to new cluster, if anon backed
    738 		 */
    739 		if (uobj == NULL && (flags & PGO_REALLOCSWAP)) {
    740 			if (pg->pqflags & PQ_ANON) {
    741 				simple_lock(&pg->uanon->an_lock);
    742 				pg->uanon->an_swslot = swblk;	/* reassign */
    743 				simple_unlock(&pg->uanon->an_lock);
    744 			} else {
    745 				simple_lock(&pg->uobject->vmobjlock);
    746 				uao_set_swslot(pg->uobject,
    747 				    pg->offset >> PAGE_SHIFT, swblk);
    748 				simple_unlock(&pg->uobject->vmobjlock);
    749 			}
    750 		}
    751 	}
    752 }
    753