uvm_anon.c revision 1.28 
1/*	$NetBSD: uvm_anon.c,v 1.28 2004/03/24 07:55:01 junyoung 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
35/*
36 * uvm_anon.c: uvm anon ops
37 */
38
39#include <sys/cdefs.h>
40__KERNEL_RCSID(0, "$NetBSD: uvm_anon.c,v 1.28 2004/03/24 07:55:01 junyoung Exp $");
41
42#include "opt_uvmhist.h"
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/proc.h>
47#include <sys/malloc.h>
48#include <sys/pool.h>
49#include <sys/kernel.h>
50
51#include <uvm/uvm.h>
52#include <uvm/uvm_swap.h>
53
54/*
55 * anonblock_list: global list of anon blocks,
56 * locked by swap_syscall_lock (since we never remove
57 * anything from this list and we only add to it via swapctl(2)).
58 */
59
60struct uvm_anonblock {
61	LIST_ENTRY(uvm_anonblock) list;
62	int count;
63	struct vm_anon *anons;
64};
65static LIST_HEAD(anonlist, uvm_anonblock) anonblock_list;
66
67
68static boolean_t anon_pagein(struct vm_anon *);
69
70
71/*
72 * allocate anons
73 */
74void
75uvm_anon_init()
76{
77	int nanon = uvmexp.free - (uvmexp.free / 16); /* XXXCDC ??? */
78
79	simple_lock_init(&uvm.afreelock);
80	LIST_INIT(&anonblock_list);
81
82	/*
83	 * Allocate the initial anons.
84	 */
85	uvm_anon_add(nanon);
86}
87
88/*
89 * add some more anons to the free pool.  called when we add
90 * more swap space.
91 *
92 * => swap_syscall_lock should be held (protects anonblock_list).
93 */
94int
95uvm_anon_add(count)
96	int	count;
97{
98	struct uvm_anonblock *anonblock;
99	struct vm_anon *anon;
100	int lcv, needed;
101
102	simple_lock(&uvm.afreelock);
103	uvmexp.nanonneeded += count;
104	needed = uvmexp.nanonneeded - uvmexp.nanon;
105	simple_unlock(&uvm.afreelock);
106
107	if (needed <= 0) {
108		return 0;
109	}
110	anon = (void *)uvm_km_alloc(kernel_map, sizeof(*anon) * needed);
111	if (anon == NULL) {
112		simple_lock(&uvm.afreelock);
113		uvmexp.nanonneeded -= count;
114		simple_unlock(&uvm.afreelock);
115		return ENOMEM;
116	}
117	MALLOC(anonblock, void *, sizeof(*anonblock), M_UVMAMAP, M_WAITOK);
118
119	anonblock->count = needed;
120	anonblock->anons = anon;
121	LIST_INSERT_HEAD(&anonblock_list, anonblock, list);
122	memset(anon, 0, sizeof(*anon) * needed);
123
124	simple_lock(&uvm.afreelock);
125	uvmexp.nanon += needed;
126	uvmexp.nfreeanon += needed;
127	for (lcv = 0; lcv < needed; lcv++) {
128		simple_lock_init(&anon[lcv].an_lock);
129		anon[lcv].u.an_nxt = uvm.afree;
130		uvm.afree = &anon[lcv];
131	}
132	simple_unlock(&uvm.afreelock);
133	return 0;
134}
135
136/*
137 * remove anons from the free pool.
138 */
139void
140uvm_anon_remove(count)
141	int count;
142{
143	/*
144	 * we never actually free any anons, to avoid allocation overhead.
145	 * XXX someday we might want to try to free anons.
146	 */
147
148	simple_lock(&uvm.afreelock);
149	uvmexp.nanonneeded -= count;
150	simple_unlock(&uvm.afreelock);
151}
152
153/*
154 * allocate an anon
155 *
156 * => new anon is returned locked!
157 */
158struct vm_anon *
159uvm_analloc()
160{
161	struct vm_anon *a;
162
163	simple_lock(&uvm.afreelock);
164	a = uvm.afree;
165	if (a) {
166		uvm.afree = a->u.an_nxt;
167		uvmexp.nfreeanon--;
168		a->an_ref = 1;
169		a->an_swslot = 0;
170		a->u.an_page = NULL;		/* so we can free quickly */
171		LOCK_ASSERT(simple_lock_held(&a->an_lock) == 0);
172		simple_lock(&a->an_lock);
173	}
174	simple_unlock(&uvm.afreelock);
175	return(a);
176}
177
178/*
179 * uvm_anfree: free a single anon structure
180 *
181 * => caller must remove anon from its amap before calling (if it was in
182 *	an amap).
183 * => anon must be unlocked and have a zero reference count.
184 * => we may lock the pageq's.
185 */
186
187void
188uvm_anfree(anon)
189	struct vm_anon *anon;
190{
191	struct vm_page *pg;
192	UVMHIST_FUNC("uvm_anfree"); UVMHIST_CALLED(maphist);
193	UVMHIST_LOG(maphist,"(anon=0x%x)", anon, 0,0,0);
194
195	KASSERT(anon->an_ref == 0);
196	LOCK_ASSERT(!simple_lock_held(&anon->an_lock));
197
198	/*
199	 * get page
200	 */
201
202	pg = anon->u.an_page;
203
204	/*
205	 * if there is a resident page and it is loaned, then anon may not
206	 * own it.   call out to uvm_anon_lockpage() to ensure the real owner
207 	 * of the page has been identified and locked.
208	 */
209
210	if (pg && pg->loan_count) {
211		simple_lock(&anon->an_lock);
212		pg = uvm_anon_lockloanpg(anon);
213		simple_unlock(&anon->an_lock);
214	}
215
216	/*
217	 * if we have a resident page, we must dispose of it before freeing
218	 * the anon.
219	 */
220
221	if (pg) {
222
223		/*
224		 * if the page is owned by a uobject (now locked), then we must
225		 * kill the loan on the page rather than free it.
226		 */
227
228		if (pg->uobject) {
229			uvm_lock_pageq();
230			KASSERT(pg->loan_count > 0);
231			pg->loan_count--;
232			pg->uanon = NULL;
233			uvm_unlock_pageq();
234			simple_unlock(&pg->uobject->vmobjlock);
235		} else {
236
237			/*
238			 * page has no uobject, so we must be the owner of it.
239			 * if page is busy then we wait until it is not busy,
240			 * and then free it.
241			 */
242
243			KASSERT((pg->flags & PG_RELEASED) == 0);
244			simple_lock(&anon->an_lock);
245			pmap_page_protect(pg, VM_PROT_NONE);
246			while ((pg = anon->u.an_page) &&
247			       (pg->flags & PG_BUSY) != 0) {
248				pg->flags |= PG_WANTED;
249				UVM_UNLOCK_AND_WAIT(pg, &anon->an_lock, 0,
250				    "anfree", 0);
251				simple_lock(&anon->an_lock);
252			}
253			if (pg) {
254				uvm_lock_pageq();
255				uvm_pagefree(pg);
256				uvm_unlock_pageq();
257			}
258			simple_unlock(&anon->an_lock);
259			UVMHIST_LOG(maphist, "anon 0x%x, page 0x%x: "
260				    "freed now!", anon, pg, 0, 0);
261		}
262	}
263	if (pg == NULL && anon->an_swslot > 0) {
264		/* this page is no longer only in swap. */
265		simple_lock(&uvm.swap_data_lock);
266		KASSERT(uvmexp.swpgonly > 0);
267		uvmexp.swpgonly--;
268		simple_unlock(&uvm.swap_data_lock);
269	}
270
271	/*
272	 * free any swap resources.
273	 */
274
275	uvm_anon_dropswap(anon);
276
277	/*
278	 * now that we've stripped the data areas from the anon,
279	 * free the anon itself.
280	 */
281
282	simple_lock(&uvm.afreelock);
283	anon->u.an_nxt = uvm.afree;
284	uvm.afree = anon;
285	uvmexp.nfreeanon++;
286	simple_unlock(&uvm.afreelock);
287	UVMHIST_LOG(maphist,"<- done!",0,0,0,0);
288}
289
290/*
291 * uvm_anon_dropswap:  release any swap resources from this anon.
292 *
293 * => anon must be locked or have a reference count of 0.
294 */
295void
296uvm_anon_dropswap(anon)
297	struct vm_anon *anon;
298{
299	UVMHIST_FUNC("uvm_anon_dropswap"); UVMHIST_CALLED(maphist);
300
301	if (anon->an_swslot == 0)
302		return;
303
304	UVMHIST_LOG(maphist,"freeing swap for anon %p, paged to swslot 0x%x",
305		    anon, anon->an_swslot, 0, 0);
306	uvm_swap_free(anon->an_swslot, 1);
307	anon->an_swslot = 0;
308}
309
310/*
311 * uvm_anon_lockloanpg: given a locked anon, lock its resident page
312 *
313 * => anon is locked by caller
314 * => on return: anon is locked
315 *		 if there is a resident page:
316 *			if it has a uobject, it is locked by us
317 *			if it is ownerless, we take over as owner
318 *		 we return the resident page (it can change during
319 *		 this function)
320 * => note that the only time an anon has an ownerless resident page
321 *	is if the page was loaned from a uvm_object and the uvm_object
322 *	disowned it
323 * => this only needs to be called when you want to do an operation
324 *	on an anon's resident page and that page has a non-zero loan
325 *	count.
326 */
327struct vm_page *
328uvm_anon_lockloanpg(anon)
329	struct vm_anon *anon;
330{
331	struct vm_page *pg;
332	boolean_t locked = FALSE;
333
334	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
335
336	/*
337	 * loop while we have a resident page that has a non-zero loan count.
338	 * if we successfully get our lock, we will "break" the loop.
339	 * note that the test for pg->loan_count is not protected -- this
340	 * may produce false positive results.   note that a false positive
341	 * result may cause us to do more work than we need to, but it will
342	 * not produce an incorrect result.
343	 */
344
345	while (((pg = anon->u.an_page) != NULL) && pg->loan_count != 0) {
346
347		/*
348		 * quickly check to see if the page has an object before
349		 * bothering to lock the page queues.   this may also produce
350		 * a false positive result, but that's ok because we do a real
351		 * check after that.
352		 */
353
354		if (pg->uobject) {
355			uvm_lock_pageq();
356			if (pg->uobject) {
357				locked =
358				    simple_lock_try(&pg->uobject->vmobjlock);
359			} else {
360				/* object disowned before we got PQ lock */
361				locked = TRUE;
362			}
363			uvm_unlock_pageq();
364
365			/*
366			 * if we didn't get a lock (try lock failed), then we
367			 * toggle our anon lock and try again
368			 */
369
370			if (!locked) {
371				simple_unlock(&anon->an_lock);
372
373				/*
374				 * someone locking the object has a chance to
375				 * lock us right now
376				 */
377
378				simple_lock(&anon->an_lock);
379				continue;
380			}
381		}
382
383		/*
384		 * if page is un-owned [i.e. the object dropped its ownership],
385		 * then we can take over as owner!
386		 */
387
388		if (pg->uobject == NULL && (pg->pqflags & PQ_ANON) == 0) {
389			uvm_lock_pageq();
390			pg->pqflags |= PQ_ANON;
391			pg->loan_count--;
392			uvm_unlock_pageq();
393		}
394		break;
395	}
396	return(pg);
397}
398
399
400
401/*
402 * page in every anon that is paged out to a range of swslots.
403 *
404 * swap_syscall_lock should be held (protects anonblock_list).
405 */
406
407boolean_t
408anon_swap_off(startslot, endslot)
409	int startslot, endslot;
410{
411	struct uvm_anonblock *anonblock;
412
413	LIST_FOREACH(anonblock, &anonblock_list, list) {
414		int i;
415
416		/*
417		 * loop thru all the anons in the anonblock,
418		 * paging in where needed.
419		 */
420
421		for (i = 0; i < anonblock->count; i++) {
422			struct vm_anon *anon = &anonblock->anons[i];
423			int slot;
424
425			/*
426			 * lock anon to work on it.
427			 */
428
429			simple_lock(&anon->an_lock);
430
431			/*
432			 * is this anon's swap slot in range?
433			 */
434
435			slot = anon->an_swslot;
436			if (slot >= startslot && slot < endslot) {
437				boolean_t rv;
438
439				/*
440				 * yup, page it in.
441				 */
442
443				/* locked: anon */
444				rv = anon_pagein(anon);
445				/* unlocked: anon */
446
447				if (rv) {
448					return rv;
449				}
450			} else {
451
452				/*
453				 * nope, unlock and proceed.
454				 */
455
456				simple_unlock(&anon->an_lock);
457			}
458		}
459	}
460	return FALSE;
461}
462
463
464/*
465 * fetch an anon's page.
466 *
467 * => anon must be locked, and is unlocked upon return.
468 * => returns TRUE if pagein was aborted due to lack of memory.
469 */
470
471static boolean_t
472anon_pagein(anon)
473	struct vm_anon *anon;
474{
475	struct vm_page *pg;
476	struct uvm_object *uobj;
477	int rv;
478
479	/* locked: anon */
480	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
481
482	rv = uvmfault_anonget(NULL, NULL, anon);
483
484	/*
485	 * if rv == 0, anon is still locked, else anon
486	 * is unlocked
487	 */
488
489	switch (rv) {
490	case 0:
491		break;
492
493	case EIO:
494	case ERESTART:
495
496		/*
497		 * nothing more to do on errors.
498		 * ERESTART can only mean that the anon was freed,
499		 * so again there's nothing to do.
500		 */
501
502		return FALSE;
503
504	default:
505		return TRUE;
506	}
507
508	/*
509	 * ok, we've got the page now.
510	 * mark it as dirty, clear its swslot and un-busy it.
511	 */
512
513	pg = anon->u.an_page;
514	uobj = pg->uobject;
515	if (anon->an_swslot > 0)
516		uvm_swap_free(anon->an_swslot, 1);
517	anon->an_swslot = 0;
518	pg->flags &= ~(PG_CLEAN);
519
520	/*
521	 * deactivate the page (to put it on a page queue)
522	 */
523
524	pmap_clear_reference(pg);
525	uvm_lock_pageq();
526	if (pg->wire_count == 0)
527		uvm_pagedeactivate(pg);
528	uvm_unlock_pageq();
529
530	if (pg->flags & PG_WANTED) {
531		wakeup(pg);
532		pg->flags &= ~(PG_WANTED);
533	}
534
535	/*
536	 * unlock the anon and we're done.
537	 */
538
539	simple_unlock(&anon->an_lock);
540	if (uobj) {
541		simple_unlock(&uobj->vmobjlock);
542	}
543	return FALSE;
544}
545