uvm_anon.c revision 1.20 
1/*	$NetBSD: uvm_anon.c,v 1.20 2001/11/06 08:07:49 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
35/*
36 * uvm_anon.c: uvm anon ops
37 */
38
39#include "opt_uvmhist.h"
40
41#include <sys/param.h>
42#include <sys/systm.h>
43#include <sys/proc.h>
44#include <sys/malloc.h>
45#include <sys/pool.h>
46#include <sys/kernel.h>
47
48#include <uvm/uvm.h>
49#include <uvm/uvm_swap.h>
50
51/*
52 * anonblock_list: global list of anon blocks,
53 * locked by swap_syscall_lock (since we never remove
54 * anything from this list and we only add to it via swapctl(2)).
55 */
56
57struct uvm_anonblock {
58	LIST_ENTRY(uvm_anonblock) list;
59	int count;
60	struct vm_anon *anons;
61};
62static LIST_HEAD(anonlist, uvm_anonblock) anonblock_list;
63
64
65static boolean_t anon_pagein __P((struct vm_anon *));
66
67
68/*
69 * allocate anons
70 */
71void
72uvm_anon_init()
73{
74	int nanon = uvmexp.free - (uvmexp.free / 16); /* XXXCDC ??? */
75
76	simple_lock_init(&uvm.afreelock);
77	LIST_INIT(&anonblock_list);
78
79	/*
80	 * Allocate the initial anons.
81	 */
82	uvm_anon_add(nanon);
83}
84
85/*
86 * add some more anons to the free pool.  called when we add
87 * more swap space.
88 *
89 * => swap_syscall_lock should be held (protects anonblock_list).
90 */
91int
92uvm_anon_add(count)
93	int	count;
94{
95	struct uvm_anonblock *anonblock;
96	struct vm_anon *anon;
97	int lcv, needed;
98
99	simple_lock(&uvm.afreelock);
100	uvmexp.nanonneeded += count;
101	needed = uvmexp.nanonneeded - uvmexp.nanon;
102	simple_unlock(&uvm.afreelock);
103
104	if (needed <= 0) {
105		return 0;
106	}
107	anon = (void *)uvm_km_alloc(kernel_map, sizeof(*anon) * needed);
108	if (anon == NULL) {
109		simple_lock(&uvm.afreelock);
110		uvmexp.nanonneeded -= count;
111		simple_unlock(&uvm.afreelock);
112		return ENOMEM;
113	}
114	MALLOC(anonblock, void *, sizeof(*anonblock), M_UVMAMAP, M_WAITOK);
115
116	anonblock->count = needed;
117	anonblock->anons = anon;
118	LIST_INSERT_HEAD(&anonblock_list, anonblock, list);
119	memset(anon, 0, sizeof(*anon) * needed);
120
121	simple_lock(&uvm.afreelock);
122	uvmexp.nanon += needed;
123	uvmexp.nfreeanon += needed;
124	for (lcv = 0; lcv < needed; lcv++) {
125		simple_lock_init(&anon->an_lock);
126		anon[lcv].u.an_nxt = uvm.afree;
127		uvm.afree = &anon[lcv];
128		simple_lock_init(&uvm.afree->an_lock);
129	}
130	simple_unlock(&uvm.afreelock);
131	return 0;
132}
133
134/*
135 * remove anons from the free pool.
136 */
137void
138uvm_anon_remove(count)
139	int count;
140{
141	/*
142	 * we never actually free any anons, to avoid allocation overhead.
143	 * XXX someday we might want to try to free anons.
144	 */
145
146	simple_lock(&uvm.afreelock);
147	uvmexp.nanonneeded -= count;
148	simple_unlock(&uvm.afreelock);
149}
150
151/*
152 * allocate an anon
153 *
154 * => new anon is returned locked!
155 */
156struct vm_anon *
157uvm_analloc()
158{
159	struct vm_anon *a;
160
161	simple_lock(&uvm.afreelock);
162	a = uvm.afree;
163	if (a) {
164		uvm.afree = a->u.an_nxt;
165		uvmexp.nfreeanon--;
166		a->an_ref = 1;
167		a->an_swslot = 0;
168		a->u.an_page = NULL;		/* so we can free quickly */
169		LOCK_ASSERT(simple_lock_held(&a->an_lock) == 0);
170		simple_lock(&a->an_lock);
171	}
172	simple_unlock(&uvm.afreelock);
173	return(a);
174}
175
176/*
177 * uvm_anfree: free a single anon structure
178 *
179 * => caller must remove anon from its amap before calling (if it was in
180 *	an amap).
181 * => anon must be unlocked and have a zero reference count.
182 * => we may lock the pageq's.
183 */
184
185void
186uvm_anfree(anon)
187	struct vm_anon *anon;
188{
189	struct vm_page *pg;
190	UVMHIST_FUNC("uvm_anfree"); UVMHIST_CALLED(maphist);
191	UVMHIST_LOG(maphist,"(anon=0x%x)", anon, 0,0,0);
192
193	KASSERT(anon->an_ref == 0);
194	LOCK_ASSERT(!simple_lock_held(&anon->an_lock));
195
196	/*
197	 * get page
198	 */
199
200	pg = anon->u.an_page;
201
202	/*
203	 * if there is a resident page and it is loaned, then anon may not
204	 * own it.   call out to uvm_anon_lockpage() to ensure the real owner
205 	 * of the page has been identified and locked.
206	 */
207
208	if (pg && pg->loan_count)
209		pg = uvm_anon_lockloanpg(anon);
210
211	/*
212	 * if we have a resident page, we must dispose of it before freeing
213	 * the anon.
214	 */
215
216	if (pg) {
217
218		/*
219		 * if the page is owned by a uobject (now locked), then we must
220		 * kill the loan on the page rather than free it.
221		 */
222
223		if (pg->uobject) {
224			uvm_lock_pageq();
225			KASSERT(pg->loan_count > 0);
226			pg->loan_count--;
227			pg->uanon = NULL;
228			uvm_unlock_pageq();
229			simple_unlock(&pg->uobject->vmobjlock);
230		} else {
231
232			/*
233			 * page has no uobject, so we must be the owner of it.
234			 * if page is busy then we wait until it is not busy,
235			 * and then free it.
236			 */
237
238			KASSERT((pg->flags & PG_RELEASED) == 0);
239			simple_lock(&anon->an_lock);
240			pmap_page_protect(pg, VM_PROT_NONE);
241			while ((pg = anon->u.an_page) &&
242			       (pg->flags & PG_BUSY) != 0) {
243				pg->flags |= PG_WANTED;
244				UVM_UNLOCK_AND_WAIT(pg, &anon->an_lock, 0,
245				    "anfree", 0);
246				simple_lock(&anon->an_lock);
247			}
248			if (pg) {
249				uvm_lock_pageq();
250				uvm_pagefree(pg);
251				uvm_unlock_pageq();
252			}
253			simple_unlock(&anon->an_lock);
254			UVMHIST_LOG(maphist, "anon 0x%x, page 0x%x: "
255				    "freed now!", anon, pg, 0, 0);
256		}
257	}
258	if (pg == NULL && anon->an_swslot != 0) {
259		/* this page is no longer only in swap. */
260		simple_lock(&uvm.swap_data_lock);
261		KASSERT(uvmexp.swpgonly > 0);
262		uvmexp.swpgonly--;
263		simple_unlock(&uvm.swap_data_lock);
264	}
265
266	/*
267	 * free any swap resources.
268	 */
269
270	uvm_anon_dropswap(anon);
271
272	/*
273	 * now that we've stripped the data areas from the anon,
274	 * free the anon itself.
275	 */
276
277	simple_lock(&uvm.afreelock);
278	anon->u.an_nxt = uvm.afree;
279	uvm.afree = anon;
280	uvmexp.nfreeanon++;
281	simple_unlock(&uvm.afreelock);
282	UVMHIST_LOG(maphist,"<- done!",0,0,0,0);
283}
284
285/*
286 * uvm_anon_dropswap:  release any swap resources from this anon.
287 *
288 * => anon must be locked or have a reference count of 0.
289 */
290void
291uvm_anon_dropswap(anon)
292	struct vm_anon *anon;
293{
294	UVMHIST_FUNC("uvm_anon_dropswap"); UVMHIST_CALLED(maphist);
295
296	if (anon->an_swslot == 0)
297		return;
298
299	UVMHIST_LOG(maphist,"freeing swap for anon %p, paged to swslot 0x%x",
300		    anon, anon->an_swslot, 0, 0);
301	uvm_swap_free(anon->an_swslot, 1);
302	anon->an_swslot = 0;
303}
304
305/*
306 * uvm_anon_lockloanpg: given a locked anon, lock its resident page
307 *
308 * => anon is locked by caller
309 * => on return: anon is locked
310 *		 if there is a resident page:
311 *			if it has a uobject, it is locked by us
312 *			if it is ownerless, we take over as owner
313 *		 we return the resident page (it can change during
314 *		 this function)
315 * => note that the only time an anon has an ownerless resident page
316 *	is if the page was loaned from a uvm_object and the uvm_object
317 *	disowned it
318 * => this only needs to be called when you want to do an operation
319 *	on an anon's resident page and that page has a non-zero loan
320 *	count.
321 */
322struct vm_page *
323uvm_anon_lockloanpg(anon)
324	struct vm_anon *anon;
325{
326	struct vm_page *pg;
327	boolean_t locked = FALSE;
328
329	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
330
331	/*
332	 * loop while we have a resident page that has a non-zero loan count.
333	 * if we successfully get our lock, we will "break" the loop.
334	 * note that the test for pg->loan_count is not protected -- this
335	 * may produce false positive results.   note that a false positive
336	 * result may cause us to do more work than we need to, but it will
337	 * not produce an incorrect result.
338	 */
339
340	while (((pg = anon->u.an_page) != NULL) && pg->loan_count != 0) {
341
342		/*
343		 * quickly check to see if the page has an object before
344		 * bothering to lock the page queues.   this may also produce
345		 * a false positive result, but that's ok because we do a real
346		 * check after that.
347		 */
348
349		if (pg->uobject) {
350			uvm_lock_pageq();
351			if (pg->uobject) {
352				locked =
353				    simple_lock_try(&pg->uobject->vmobjlock);
354			} else {
355				/* object disowned before we got PQ lock */
356				locked = TRUE;
357			}
358			uvm_unlock_pageq();
359
360			/*
361			 * if we didn't get a lock (try lock failed), then we
362			 * toggle our anon lock and try again
363			 */
364
365			if (!locked) {
366				simple_unlock(&anon->an_lock);
367
368				/*
369				 * someone locking the object has a chance to
370				 * lock us right now
371				 */
372
373				simple_lock(&anon->an_lock);
374				continue;
375			}
376		}
377
378		/*
379		 * if page is un-owned [i.e. the object dropped its ownership],
380		 * then we can take over as owner!
381		 */
382
383		if (pg->uobject == NULL && (pg->pqflags & PQ_ANON) == 0) {
384			uvm_lock_pageq();
385			pg->pqflags |= PQ_ANON;
386			pg->loan_count--;
387			uvm_unlock_pageq();
388		}
389		break;
390	}
391	return(pg);
392}
393
394
395
396/*
397 * page in every anon that is paged out to a range of swslots.
398 *
399 * swap_syscall_lock should be held (protects anonblock_list).
400 */
401
402boolean_t
403anon_swap_off(startslot, endslot)
404	int startslot, endslot;
405{
406	struct uvm_anonblock *anonblock;
407
408	LIST_FOREACH(anonblock, &anonblock_list, list) {
409		int i;
410
411		/*
412		 * loop thru all the anons in the anonblock,
413		 * paging in where needed.
414		 */
415
416		for (i = 0; i < anonblock->count; i++) {
417			struct vm_anon *anon = &anonblock->anons[i];
418			int slot;
419
420			/*
421			 * lock anon to work on it.
422			 */
423
424			simple_lock(&anon->an_lock);
425
426			/*
427			 * is this anon's swap slot in range?
428			 */
429
430			slot = anon->an_swslot;
431			if (slot >= startslot && slot < endslot) {
432				boolean_t rv;
433
434				/*
435				 * yup, page it in.
436				 */
437
438				/* locked: anon */
439				rv = anon_pagein(anon);
440				/* unlocked: anon */
441
442				if (rv) {
443					return rv;
444				}
445			} else {
446
447				/*
448				 * nope, unlock and proceed.
449				 */
450
451				simple_unlock(&anon->an_lock);
452			}
453		}
454	}
455	return FALSE;
456}
457
458
459/*
460 * fetch an anon's page.
461 *
462 * => anon must be locked, and is unlocked upon return.
463 * => returns TRUE if pagein was aborted due to lack of memory.
464 */
465
466static boolean_t
467anon_pagein(anon)
468	struct vm_anon *anon;
469{
470	struct vm_page *pg;
471	struct uvm_object *uobj;
472	int rv;
473
474	/* locked: anon */
475	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
476
477	rv = uvmfault_anonget(NULL, NULL, anon);
478
479	/*
480	 * if rv == 0, anon is still locked, else anon
481	 * is unlocked
482	 */
483
484	switch (rv) {
485	case 0:
486		break;
487
488	case EIO:
489	case ERESTART:
490
491		/*
492		 * nothing more to do on errors.
493		 * ERESTART can only mean that the anon was freed,
494		 * so again there's nothing to do.
495		 */
496
497		return FALSE;
498	}
499
500	/*
501	 * ok, we've got the page now.
502	 * mark it as dirty, clear its swslot and un-busy it.
503	 */
504
505	pg = anon->u.an_page;
506	uobj = pg->uobject;
507	uvm_swap_free(anon->an_swslot, 1);
508	anon->an_swslot = 0;
509	pg->flags &= ~(PG_CLEAN);
510
511	/*
512	 * deactivate the page (to put it on a page queue)
513	 */
514
515	pmap_clear_reference(pg);
516	uvm_lock_pageq();
517	uvm_pagedeactivate(pg);
518	uvm_unlock_pageq();
519
520	/*
521	 * unlock the anon and we're done.
522	 */
523
524	simple_unlock(&anon->an_lock);
525	if (uobj) {
526		simple_unlock(&uobj->vmobjlock);
527	}
528	return FALSE;
529}
530