xref: /netbsd-src/sys/uvm/uvm_anon.c (revision 481fca6e59249d8ffcf24fef7cfbe7b131bfb080)
1 /*	$NetBSD: uvm_anon.c,v 1.7 2000/06/27 17:29:18 mrg 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 
57 struct uvm_anonblock {
58 	LIST_ENTRY(uvm_anonblock) list;
59 	int count;
60 	struct vm_anon *anons;
61 };
62 static LIST_HEAD(anonlist, uvm_anonblock) anonblock_list;
63 
64 
65 static boolean_t anon_pagein __P((struct vm_anon *));
66 
67 
68 /*
69  * allocate anons
70  */
71 void
72 uvm_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  */
91 void
92 uvm_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;
106 	}
107 
108 	MALLOC(anonblock, void *, sizeof(*anonblock), M_UVMAMAP, M_WAITOK);
109 	anon = (void *)uvm_km_alloc(kernel_map, sizeof(*anon) * needed);
110 
111 	/* XXX Should wait for VM to free up. */
112 	if (anonblock == NULL || anon == NULL) {
113 		printf("uvm_anon_add: can not allocate %d anons\n", needed);
114 		panic("uvm_anon_add");
115 	}
116 
117 	anonblock->count = needed;
118 	anonblock->anons = anon;
119 	LIST_INSERT_HEAD(&anonblock_list, anonblock, list);
120 	memset(anon, 0, sizeof(*anon) * needed);
121 
122 	simple_lock(&uvm.afreelock);
123 	uvmexp.nanon += needed;
124 	uvmexp.nfreeanon += needed;
125 	for (lcv = 0; lcv < needed; lcv++) {
126 		simple_lock_init(&anon->an_lock);
127 		anon[lcv].u.an_nxt = uvm.afree;
128 		uvm.afree = &anon[lcv];
129 		simple_lock_init(&uvm.afree->an_lock);
130 	}
131 	simple_unlock(&uvm.afreelock);
132 }
133 
134 /*
135  * remove anons from the free pool.
136  */
137 void
138 uvm_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 struct vm_anon *
155 uvm_analloc()
156 {
157 	struct vm_anon *a;
158 
159 	simple_lock(&uvm.afreelock);
160 	a = uvm.afree;
161 	if (a) {
162 		uvm.afree = a->u.an_nxt;
163 		uvmexp.nfreeanon--;
164 		a->an_ref = 1;
165 		a->an_swslot = 0;
166 		a->u.an_page = NULL;		/* so we can free quickly */
167 	}
168 	simple_unlock(&uvm.afreelock);
169 	return(a);
170 }
171 
172 /*
173  * uvm_anfree: free a single anon structure
174  *
175  * => caller must remove anon from its amap before calling (if it was in
176  *	an amap).
177  * => anon must be unlocked and have a zero reference count.
178  * => we may lock the pageq's.
179  */
180 void
181 uvm_anfree(anon)
182 	struct vm_anon *anon;
183 {
184 	struct vm_page *pg;
185 	UVMHIST_FUNC("uvm_anfree"); UVMHIST_CALLED(maphist);
186 	UVMHIST_LOG(maphist,"(anon=0x%x)", anon, 0,0,0);
187 
188 	/*
189 	 * get page
190 	 */
191 
192 	pg = anon->u.an_page;
193 
194 	/*
195 	 * if there is a resident page and it is loaned, then anon may not
196 	 * own it.   call out to uvm_anon_lockpage() to ensure the real owner
197  	 * of the page has been identified and locked.
198 	 */
199 
200 	if (pg && pg->loan_count)
201 		pg = uvm_anon_lockloanpg(anon);
202 
203 	/*
204 	 * if we have a resident page, we must dispose of it before freeing
205 	 * the anon.
206 	 */
207 
208 	if (pg) {
209 
210 		/*
211 		 * if the page is owned by a uobject (now locked), then we must
212 		 * kill the loan on the page rather than free it.
213 		 */
214 
215 		if (pg->uobject) {
216 
217 			/* kill loan */
218 			uvm_lock_pageq();
219 #ifdef DIAGNOSTIC
220 			if (pg->loan_count < 1)
221 				panic("uvm_anfree: obj owned page "
222 				      "with no loan count");
223 #endif
224 			pg->loan_count--;
225 			pg->uanon = NULL;
226 			uvm_unlock_pageq();
227 			simple_unlock(&pg->uobject->vmobjlock);
228 
229 		} else {
230 
231 			/*
232 			 * page has no uobject, so we must be the owner of it.
233 			 *
234 			 * if page is busy then we just mark it as released
235 			 * (who ever has it busy must check for this when they
236 			 * wake up).    if the page is not busy then we can
237 			 * free it now.
238 			 */
239 
240 			if ((pg->flags & PG_BUSY) != 0) {
241 				/* tell them to dump it when done */
242 				pg->flags |= PG_RELEASED;
243 				UVMHIST_LOG(maphist,
244 				    "  anon 0x%x, page 0x%x: BUSY (released!)",
245 				    anon, pg, 0, 0);
246 				return;
247 			}
248 
249 			pmap_page_protect(pg, VM_PROT_NONE);
250 			uvm_lock_pageq();	/* lock out pagedaemon */
251 			uvm_pagefree(pg);	/* bye bye */
252 			uvm_unlock_pageq();	/* free the daemon */
253 
254 			UVMHIST_LOG(maphist,"  anon 0x%x, page 0x%x: freed now!",
255 			    anon, pg, 0, 0);
256 		}
257 	}
258 
259 	/*
260 	 * free any swap resources.
261 	 */
262 	uvm_anon_dropswap(anon);
263 
264 	/*
265 	 * now that we've stripped the data areas from the anon, free the anon
266 	 * itself!
267 	 */
268 	simple_lock(&uvm.afreelock);
269 	anon->u.an_nxt = uvm.afree;
270 	uvm.afree = anon;
271 	uvmexp.nfreeanon++;
272 	simple_unlock(&uvm.afreelock);
273 	UVMHIST_LOG(maphist,"<- done!",0,0,0,0);
274 }
275 
276 /*
277  * uvm_anon_dropswap:  release any swap resources from this anon.
278  *
279  * => anon must be locked or have a reference count of 0.
280  */
281 void
282 uvm_anon_dropswap(anon)
283 	struct vm_anon *anon;
284 {
285 	UVMHIST_FUNC("uvm_anon_dropswap"); UVMHIST_CALLED(maphist);
286 	if (anon->an_swslot == 0) {
287 		return;
288 	}
289 
290 	UVMHIST_LOG(maphist,"freeing swap for anon %p, paged to swslot 0x%x",
291 		    anon, anon->an_swslot, 0, 0);
292 	uvm_swap_free(anon->an_swslot, 1);
293 	anon->an_swslot = 0;
294 
295 	if (anon->u.an_page == NULL) {
296 		/* this page is no longer only in swap. */
297 		simple_lock(&uvm.swap_data_lock);
298 		uvmexp.swpgonly--;
299 		simple_unlock(&uvm.swap_data_lock);
300 	}
301 }
302 
303 /*
304  * uvm_anon_lockloanpg: given a locked anon, lock its resident page
305  *
306  * => anon is locked by caller
307  * => on return: anon is locked
308  *		 if there is a resident page:
309  *			if it has a uobject, it is locked by us
310  *			if it is ownerless, we take over as owner
311  *		 we return the resident page (it can change during
312  *		 this function)
313  * => note that the only time an anon has an ownerless resident page
314  *	is if the page was loaned from a uvm_object and the uvm_object
315  *	disowned it
316  * => this only needs to be called when you want to do an operation
317  *	on an anon's resident page and that page has a non-zero loan
318  *	count.
319  */
320 struct vm_page *
321 uvm_anon_lockloanpg(anon)
322 	struct vm_anon *anon;
323 {
324 	struct vm_page *pg;
325 	boolean_t locked = FALSE;
326 
327 	/*
328 	 * loop while we have a resident page that has a non-zero loan count.
329 	 * if we successfully get our lock, we will "break" the loop.
330 	 * note that the test for pg->loan_count is not protected -- this
331 	 * may produce false positive results.   note that a false positive
332 	 * result may cause us to do more work than we need to, but it will
333 	 * not produce an incorrect result.
334 	 */
335 
336 	while (((pg = anon->u.an_page) != NULL) && pg->loan_count != 0) {
337 
338 		/*
339 		 * quickly check to see if the page has an object before
340 		 * bothering to lock the page queues.   this may also produce
341 		 * a false positive result, but that's ok because we do a real
342 		 * check after that.
343 		 *
344 		 * XXX: quick check -- worth it?   need volatile?
345 		 */
346 
347 		if (pg->uobject) {
348 
349 			uvm_lock_pageq();
350 			if (pg->uobject) {	/* the "real" check */
351 				locked =
352 				    simple_lock_try(&pg->uobject->vmobjlock);
353 			} else {
354 				/* object disowned before we got PQ lock */
355 				locked = TRUE;
356 			}
357 			uvm_unlock_pageq();
358 
359 			/*
360 			 * if we didn't get a lock (try lock failed), then we
361 			 * toggle our anon lock and try again
362 			 */
363 
364 			if (!locked) {
365 				simple_unlock(&anon->an_lock);
366 				/*
367 				 * someone locking the object has a chance to
368 				 * lock us right now
369 				 */
370 				simple_lock(&anon->an_lock);
371 				continue;		/* start over */
372 			}
373 		}
374 
375 		/*
376 		 * if page is un-owned [i.e. the object dropped its ownership],
377 		 * then we can take over as owner!
378 		 */
379 
380 		if (pg->uobject == NULL && (pg->pqflags & PQ_ANON) == 0) {
381 			uvm_lock_pageq();
382 			pg->pqflags |= PQ_ANON;		/* take ownership... */
383 			pg->loan_count--;	/* ... and drop our loan */
384 			uvm_unlock_pageq();
385 		}
386 
387 		/*
388 		 * we did it!   break the loop
389 		 */
390 		break;
391 	}
392 
393 	/*
394 	 * done!
395 	 */
396 
397 	return(pg);
398 }
399 
400 
401 
402 /*
403  * page in every anon that is paged out to a range of swslots.
404  *
405  * swap_syscall_lock should be held (protects anonblock_list).
406  */
407 
408 boolean_t
409 anon_swap_off(startslot, endslot)
410 	int startslot, endslot;
411 {
412 	struct uvm_anonblock *anonblock;
413 
414 	for (anonblock = LIST_FIRST(&anonblock_list);
415 	     anonblock != NULL;
416 	     anonblock = LIST_NEXT(anonblock, list)) {
417 		int i;
418 
419 		/*
420 		 * loop thru all the anons in the anonblock,
421 		 * paging in where needed.
422 		 */
423 
424 		for (i = 0; i < anonblock->count; i++) {
425 			struct vm_anon *anon = &anonblock->anons[i];
426 			int slot;
427 
428 			/*
429 			 * lock anon to work on it.
430 			 */
431 
432 			simple_lock(&anon->an_lock);
433 
434 			/*
435 			 * is this anon's swap slot in range?
436 			 */
437 
438 			slot = anon->an_swslot;
439 			if (slot >= startslot && slot < endslot) {
440 				boolean_t rv;
441 
442 				/*
443 				 * yup, page it in.
444 				 */
445 
446 				/* locked: anon */
447 				rv = anon_pagein(anon);
448 				/* unlocked: anon */
449 
450 				if (rv) {
451 					return rv;
452 				}
453 			} else {
454 
455 				/*
456 				 * nope, unlock and proceed.
457 				 */
458 
459 				simple_unlock(&anon->an_lock);
460 			}
461 		}
462 	}
463 	return FALSE;
464 }
465 
466 
467 /*
468  * fetch an anon's page.
469  *
470  * => anon must be locked, and is unlocked upon return.
471  * => returns TRUE if pagein was aborted due to lack of memory.
472  */
473 
474 static boolean_t
475 anon_pagein(anon)
476 	struct vm_anon *anon;
477 {
478 	struct vm_page *pg;
479 	struct uvm_object *uobj;
480 	int rv;
481 	UVMHIST_FUNC("anon_pagein"); UVMHIST_CALLED(pdhist);
482 
483 	/* locked: anon */
484 	rv = uvmfault_anonget(NULL, NULL, anon);
485 	/* unlocked: anon */
486 
487 	switch (rv) {
488 	case VM_PAGER_OK:
489 		break;
490 
491 	case VM_PAGER_ERROR:
492 	case VM_PAGER_REFAULT:
493 
494 		/*
495 		 * nothing more to do on errors.
496 		 * VM_PAGER_REFAULT can only mean that the anon was freed,
497 		 * so again there's nothing to do.
498 		 */
499 
500 		return FALSE;
501 
502 #ifdef DIAGNOSTIC
503 	default:
504 		panic("anon_pagein: uvmfault_anonget -> %d", rv);
505 #endif
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 	uvm_swap_free(anon->an_swslot, 1);
516 	anon->an_swslot = 0;
517 	pg->flags &= ~(PG_CLEAN);
518 
519 	/*
520 	 * deactivate the page (to put it on a page queue)
521 	 */
522 
523 	pmap_clear_reference(pg);
524 	pmap_page_protect(pg, VM_PROT_NONE);
525 	uvm_lock_pageq();
526 	uvm_pagedeactivate(pg);
527 	uvm_unlock_pageq();
528 
529 	/*
530 	 * unlock the anon and we're done.
531 	 */
532 
533 	simple_unlock(&anon->an_lock);
534 	if (uobj) {
535 		simple_unlock(&uobj->vmobjlock);
536 	}
537 	return FALSE;
538 }
539