Home | History | Annotate | Line # | Download | only in uvm
uvm_anon.c revision 1.59
      1 /*	$NetBSD: uvm_anon.c,v 1.59 2011/08/06 17:25:03 rmind Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
      5  * All rights reserved.
      6  *
      7  * Redistribution and use in source and binary forms, with or without
      8  * modification, are permitted provided that the following conditions
      9  * are met:
     10  * 1. Redistributions of source code must retain the above copyright
     11  *    notice, this list of conditions and the following disclaimer.
     12  * 2. Redistributions in binary form must reproduce the above copyright
     13  *    notice, this list of conditions and the following disclaimer in the
     14  *    documentation and/or other materials provided with the distribution.
     15  *
     16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
     17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
     18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
     20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
     21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
     22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
     23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
     24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
     25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     26  */
     27 
     28 /*
     29  * uvm_anon.c: uvm anon ops
     30  */
     31 
     32 #include <sys/cdefs.h>
     33 __KERNEL_RCSID(0, "$NetBSD: uvm_anon.c,v 1.59 2011/08/06 17:25:03 rmind Exp $");
     34 
     35 #include "opt_uvmhist.h"
     36 
     37 #include <sys/param.h>
     38 #include <sys/systm.h>
     39 #include <sys/pool.h>
     40 #include <sys/kernel.h>
     41 
     42 #include <uvm/uvm.h>
     43 #include <uvm/uvm_swap.h>
     44 #include <uvm/uvm_pdpolicy.h>
     45 
     46 static struct pool_cache	uvm_anon_cache;
     47 
     48 static int			uvm_anon_ctor(void *, void *, int);
     49 
     50 void
     51 uvm_anon_init(void)
     52 {
     53 
     54 	pool_cache_bootstrap(&uvm_anon_cache, sizeof(struct vm_anon), 0, 0,
     55 	    PR_LARGECACHE, "anonpl", NULL, IPL_NONE, uvm_anon_ctor,
     56 	    NULL, NULL);
     57 }
     58 
     59 static int
     60 uvm_anon_ctor(void *arg, void *object, int flags)
     61 {
     62 	struct vm_anon *anon = object;
     63 
     64 	anon->an_ref = 0;
     65 	anon->an_lock = NULL;
     66 	anon->an_page = NULL;
     67 #if defined(VMSWAP)
     68 	anon->an_swslot = 0;
     69 #endif
     70 	return 0;
     71 }
     72 
     73 /*
     74  * uvm_analloc: allocate a new anon.
     75  *
     76  * => anon will have no lock associated.
     77  */
     78 struct vm_anon *
     79 uvm_analloc(void)
     80 {
     81 	struct vm_anon *anon;
     82 
     83 	anon = pool_cache_get(&uvm_anon_cache, PR_NOWAIT);
     84 	if (anon) {
     85 		KASSERT(anon->an_ref == 0);
     86 		KASSERT(anon->an_lock == NULL);
     87 		KASSERT(anon->an_page == NULL);
     88 #if defined(VMSWAP)
     89 		KASSERT(anon->an_swslot == 0);
     90 #endif
     91 		anon->an_ref = 1;
     92 	}
     93 	return anon;
     94 }
     95 
     96 /*
     97  * uvm_anon_dispose: free any resident page or swap resources of anon.
     98  *
     99  * => anon must be removed from the amap (if anon was in an amap).
    100  * => amap must be locked; we may drop and re-acquire the lock here.
    101  */
    102 
    103 static void
    104 uvm_anon_dispose(struct vm_anon *anon)
    105 {
    106 	struct vm_page *pg = anon->an_page;
    107 
    108 	UVMHIST_FUNC("uvm_anon_dispose"); UVMHIST_CALLED(maphist);
    109 	UVMHIST_LOG(maphist,"(anon=0x%x)", anon, 0,0,0);
    110 
    111 	KASSERT(mutex_owned(anon->an_lock));
    112 
    113 	/*
    114 	 * If there is a resident page and it is loaned, then anon may not
    115 	 * own it.  Call out to uvm_anon_lockloanpg() to identify and lock
    116 	 * the real owner of the page.
    117 	 */
    118 
    119 	if (pg && pg->loan_count) {
    120 		KASSERT(anon->an_lock != NULL);
    121 		pg = uvm_anon_lockloanpg(anon);
    122 	}
    123 
    124 	/*
    125 	 * Dispose the page, if it is resident.
    126 	 */
    127 
    128 	if (pg) {
    129 		KASSERT(anon->an_lock != NULL);
    130 
    131 		/*
    132 		 * If the page is owned by a UVM object (now locked),
    133 		 * then kill the loan on the page rather than free it,
    134 		 * and release the object lock.
    135 		 */
    136 
    137 		if (pg->uobject) {
    138 			mutex_enter(&uvm_pageqlock);
    139 			KASSERT(pg->loan_count > 0);
    140 			pg->loan_count--;
    141 			pg->uanon = NULL;
    142 			mutex_exit(&uvm_pageqlock);
    143 			mutex_exit(pg->uobject->vmobjlock);
    144 		} else {
    145 
    146 			/*
    147 			 * If page has no UVM object, then anon is the owner,
    148 			 * and it is already locked.
    149 			 */
    150 
    151 			KASSERT((pg->flags & PG_RELEASED) == 0);
    152 			pmap_page_protect(pg, VM_PROT_NONE);
    153 
    154 			/*
    155 			 * If the page is busy, mark it as PG_RELEASED, so
    156 			 * that uvm_anon_release(9) would release it later.
    157 			 */
    158 
    159 			if (pg->flags & PG_BUSY) {
    160 				pg->flags |= PG_RELEASED;
    161 				mutex_obj_hold(anon->an_lock);
    162 				return;
    163 			}
    164 			mutex_enter(&uvm_pageqlock);
    165 			uvm_pagefree(pg);
    166 			mutex_exit(&uvm_pageqlock);
    167 			UVMHIST_LOG(maphist, "anon 0x%x, page 0x%x: "
    168 				    "freed now!", anon, pg, 0, 0);
    169 		}
    170 	}
    171 
    172 #if defined(VMSWAP)
    173 	if (pg == NULL && anon->an_swslot > 0) {
    174 		/* This page is no longer only in swap. */
    175 		mutex_enter(&uvm_swap_data_lock);
    176 		KASSERT(uvmexp.swpgonly > 0);
    177 		uvmexp.swpgonly--;
    178 		mutex_exit(&uvm_swap_data_lock);
    179 	}
    180 #endif
    181 
    182 	/*
    183 	 * Free any swap resources, leave a page replacement hint.
    184 	 */
    185 
    186 	uvm_anon_dropswap(anon);
    187 	uvmpdpol_anfree(anon);
    188 	UVMHIST_LOG(maphist,"<- done!",0,0,0,0);
    189 }
    190 
    191 /*
    192  * uvm_anon_free: free a single anon.
    193  *
    194  * => anon must be already disposed.
    195  */
    196 void
    197 uvm_anon_free(struct vm_anon *anon)
    198 {
    199 
    200 	KASSERT(anon->an_ref == 0);
    201 	KASSERT(anon->an_lock == NULL);
    202 	KASSERT(anon->an_page == NULL);
    203 #if defined(VMSWAP)
    204 	KASSERT(anon->an_swslot == 0);
    205 #endif
    206 	pool_cache_put(&uvm_anon_cache, anon);
    207 }
    208 
    209 /*
    210  * uvm_anon_freelst: free a linked list of anon structures.
    211  *
    212  * => anon must be locked, we will unlock it.
    213  */
    214 void
    215 uvm_anon_freelst(struct vm_amap *amap, struct vm_anon *anonlst)
    216 {
    217 	struct vm_anon *anon = anonlst;
    218 
    219 	KASSERT(mutex_owned(amap->am_lock));
    220 
    221 	while (anon) {
    222 		uvm_anon_dispose(anon);
    223 		anon = anon->an_link;
    224 	}
    225 	amap_unlock(amap);
    226 
    227 	while (anonlst) {
    228 		anon = anonlst->an_link;
    229 		/* Note: clears an_ref as well. */
    230 		anonlst->an_link = NULL;
    231 		anonlst->an_lock = NULL;
    232 		uvm_anon_free(anonlst);
    233 		anonlst = anon;
    234 	}
    235 }
    236 
    237 /*
    238  * uvm_anon_lockloanpg: given a locked anon, lock its resident page owner.
    239  *
    240  * => anon is locked by caller
    241  * => on return: anon is locked
    242  *		 if there is a resident page:
    243  *			if it has a uobject, it is locked by us
    244  *			if it is ownerless, we take over as owner
    245  *		 we return the resident page (it can change during
    246  *		 this function)
    247  * => note that the only time an anon has an ownerless resident page
    248  *	is if the page was loaned from a uvm_object and the uvm_object
    249  *	disowned it
    250  * => this only needs to be called when you want to do an operation
    251  *	on an anon's resident page and that page has a non-zero loan
    252  *	count.
    253  */
    254 struct vm_page *
    255 uvm_anon_lockloanpg(struct vm_anon *anon)
    256 {
    257 	struct vm_page *pg;
    258 	bool locked = false;
    259 
    260 	KASSERT(mutex_owned(anon->an_lock));
    261 
    262 	/*
    263 	 * loop while we have a resident page that has a non-zero loan count.
    264 	 * if we successfully get our lock, we will "break" the loop.
    265 	 * note that the test for pg->loan_count is not protected -- this
    266 	 * may produce false positive results.   note that a false positive
    267 	 * result may cause us to do more work than we need to, but it will
    268 	 * not produce an incorrect result.
    269 	 */
    270 
    271 	while (((pg = anon->an_page) != NULL) && pg->loan_count != 0) {
    272 
    273 		/*
    274 		 * quickly check to see if the page has an object before
    275 		 * bothering to lock the page queues.   this may also produce
    276 		 * a false positive result, but that's ok because we do a real
    277 		 * check after that.
    278 		 */
    279 
    280 		if (pg->uobject) {
    281 			mutex_enter(&uvm_pageqlock);
    282 			if (pg->uobject) {
    283 				locked =
    284 				    mutex_tryenter(pg->uobject->vmobjlock);
    285 			} else {
    286 				/* object disowned before we got PQ lock */
    287 				locked = true;
    288 			}
    289 			mutex_exit(&uvm_pageqlock);
    290 
    291 			/*
    292 			 * if we didn't get a lock (try lock failed), then we
    293 			 * toggle our anon lock and try again
    294 			 */
    295 
    296 			if (!locked) {
    297 				/*
    298 				 * someone locking the object has a chance to
    299 				 * lock us right now
    300 				 *
    301 				 * XXX Better than yielding but inadequate.
    302 				 */
    303 				kpause("livelock", false, 1, anon->an_lock);
    304 				continue;
    305 			}
    306 		}
    307 
    308 		/*
    309 		 * If page is un-owned i.e. the object dropped its ownership,
    310 		 * then we have to take the ownership.
    311 		 */
    312 
    313 		if (pg->uobject == NULL && (pg->pqflags & PQ_ANON) == 0) {
    314 			mutex_enter(&uvm_pageqlock);
    315 			pg->pqflags |= PQ_ANON;
    316 			pg->loan_count--;
    317 			mutex_exit(&uvm_pageqlock);
    318 		}
    319 		break;
    320 	}
    321 	return pg;
    322 }
    323 
    324 #if defined(VMSWAP)
    325 
    326 /*
    327  * uvm_anon_pagein: fetch an anon's page.
    328  *
    329  * => anon must be locked, and is unlocked upon return.
    330  * => returns true if pagein was aborted due to lack of memory.
    331  */
    332 
    333 bool
    334 uvm_anon_pagein(struct vm_amap *amap, struct vm_anon *anon)
    335 {
    336 	struct vm_page *pg;
    337 	struct uvm_object *uobj;
    338 
    339 	KASSERT(mutex_owned(anon->an_lock));
    340 	KASSERT(anon->an_lock == amap->am_lock);
    341 
    342 	/*
    343 	 * Get the page of the anon.
    344 	 */
    345 
    346 	switch (uvmfault_anonget(NULL, amap, anon)) {
    347 	case 0:
    348 		/* Success - we have the page. */
    349 		KASSERT(mutex_owned(anon->an_lock));
    350 		break;
    351 	case EIO:
    352 	case ERESTART:
    353 		/*
    354 		 * Nothing more to do on errors.  ERESTART means that the
    355 		 * anon was freed.
    356 		 */
    357 		return false;
    358 	default:
    359 		return true;
    360 	}
    361 
    362 	/*
    363 	 * Mark the page as dirty, clear its swslot and un-busy it.
    364 	 */
    365 
    366 	pg = anon->an_page;
    367 	uobj = pg->uobject;
    368 	if (anon->an_swslot > 0) {
    369 		uvm_swap_free(anon->an_swslot, 1);
    370 	}
    371 	anon->an_swslot = 0;
    372 	pg->flags &= ~PG_CLEAN;
    373 
    374 	/*
    375 	 * Deactivate the page (to put it on a page queue).
    376 	 */
    377 
    378 	mutex_enter(&uvm_pageqlock);
    379 	if (pg->wire_count == 0) {
    380 		uvm_pagedeactivate(pg);
    381 	}
    382 	mutex_exit(&uvm_pageqlock);
    383 
    384 	if (pg->flags & PG_WANTED) {
    385 		pg->flags &= ~PG_WANTED;
    386 		wakeup(pg);
    387 	}
    388 
    389 	mutex_exit(anon->an_lock);
    390 	if (uobj) {
    391 		mutex_exit(uobj->vmobjlock);
    392 	}
    393 	return false;
    394 }
    395 
    396 /*
    397  * uvm_anon_dropswap: release any swap resources from this anon.
    398  *
    399  * => anon must be locked or have a reference count of 0.
    400  */
    401 void
    402 uvm_anon_dropswap(struct vm_anon *anon)
    403 {
    404 	UVMHIST_FUNC("uvm_anon_dropswap"); UVMHIST_CALLED(maphist);
    405 
    406 	if (anon->an_swslot == 0)
    407 		return;
    408 
    409 	UVMHIST_LOG(maphist,"freeing swap for anon %p, paged to swslot 0x%x",
    410 		    anon, anon->an_swslot, 0, 0);
    411 	uvm_swap_free(anon->an_swslot, 1);
    412 	anon->an_swslot = 0;
    413 }
    414 
    415 #endif
    416 
    417 /*
    418  * uvm_anon_release: release an anon and its page.
    419  *
    420  * => anon should not have any references.
    421  * => anon must be locked.
    422  */
    423 
    424 void
    425 uvm_anon_release(struct vm_anon *anon)
    426 {
    427 	struct vm_page *pg = anon->an_page;
    428 	kmutex_t *lock;
    429 
    430 	KASSERT(mutex_owned(anon->an_lock));
    431 	KASSERT(pg != NULL);
    432 	KASSERT((pg->flags & PG_RELEASED) != 0);
    433 	KASSERT((pg->flags & PG_BUSY) != 0);
    434 	KASSERT(pg->uobject == NULL);
    435 	KASSERT(pg->uanon == anon);
    436 	KASSERT(pg->loan_count == 0);
    437 	KASSERT(anon->an_ref == 0);
    438 
    439 	mutex_enter(&uvm_pageqlock);
    440 	uvm_pagefree(pg);
    441 	mutex_exit(&uvm_pageqlock);
    442 	mutex_exit(anon->an_lock);
    443 
    444 	KASSERT(anon->an_page == NULL);
    445 
    446 	/* Note: extra reference is held for PG_RELEASED case. */
    447 	lock = anon->an_lock;
    448 	anon->an_lock = NULL;
    449 	uvm_anon_free(anon);
    450 	mutex_obj_free(lock);
    451 }
    452