uvm_anon.c revision 1.59 1 1.59 rmind /* $NetBSD: uvm_anon.c,v 1.59 2011/08/06 17:25:03 rmind Exp $ */
2 1.1 chuck
3 1.1 chuck /*
4 1.1 chuck * Copyright (c) 1997 Charles D. Cranor and Washington University.
5 1.1 chuck * All rights reserved.
6 1.1 chuck *
7 1.1 chuck * Redistribution and use in source and binary forms, with or without
8 1.1 chuck * modification, are permitted provided that the following conditions
9 1.1 chuck * are met:
10 1.1 chuck * 1. Redistributions of source code must retain the above copyright
11 1.1 chuck * notice, this list of conditions and the following disclaimer.
12 1.1 chuck * 2. Redistributions in binary form must reproduce the above copyright
13 1.1 chuck * notice, this list of conditions and the following disclaimer in the
14 1.1 chuck * documentation and/or other materials provided with the distribution.
15 1.1 chuck *
16 1.1 chuck * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17 1.1 chuck * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18 1.1 chuck * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19 1.1 chuck * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20 1.1 chuck * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21 1.1 chuck * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22 1.1 chuck * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23 1.1 chuck * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24 1.1 chuck * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25 1.1 chuck * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 1.1 chuck */
27 1.1 chuck
28 1.1 chuck /*
29 1.1 chuck * uvm_anon.c: uvm anon ops
30 1.1 chuck */
31 1.21 lukem
32 1.21 lukem #include <sys/cdefs.h>
33 1.59 rmind __KERNEL_RCSID(0, "$NetBSD: uvm_anon.c,v 1.59 2011/08/06 17:25:03 rmind Exp $");
34 1.1 chuck
35 1.1 chuck #include "opt_uvmhist.h"
36 1.1 chuck
37 1.1 chuck #include <sys/param.h>
38 1.1 chuck #include <sys/systm.h>
39 1.1 chuck #include <sys/pool.h>
40 1.5 chs #include <sys/kernel.h>
41 1.1 chuck
42 1.1 chuck #include <uvm/uvm.h>
43 1.1 chuck #include <uvm/uvm_swap.h>
44 1.39 yamt #include <uvm/uvm_pdpolicy.h>
45 1.1 chuck
46 1.55 rmind static struct pool_cache uvm_anon_cache;
47 1.5 chs
48 1.55 rmind static int uvm_anon_ctor(void *, void *, int);
49 1.5 chs
50 1.1 chuck void
51 1.34 thorpej uvm_anon_init(void)
52 1.1 chuck {
53 1.5 chs
54 1.49 ad pool_cache_bootstrap(&uvm_anon_cache, sizeof(struct vm_anon), 0, 0,
55 1.49 ad PR_LARGECACHE, "anonpl", NULL, IPL_NONE, uvm_anon_ctor,
56 1.55 rmind NULL, NULL);
57 1.1 chuck }
58 1.1 chuck
59 1.33 yamt static int
60 1.41 yamt uvm_anon_ctor(void *arg, void *object, int flags)
61 1.1 chuck {
62 1.33 yamt struct vm_anon *anon = object;
63 1.1 chuck
64 1.33 yamt anon->an_ref = 0;
65 1.59 rmind anon->an_lock = NULL;
66 1.33 yamt anon->an_page = NULL;
67 1.37 yamt #if defined(VMSWAP)
68 1.33 yamt anon->an_swslot = 0;
69 1.55 rmind #endif
70 1.11 chs return 0;
71 1.1 chuck }
72 1.1 chuck
73 1.1 chuck /*
74 1.55 rmind * uvm_analloc: allocate a new anon.
75 1.13 thorpej *
76 1.55 rmind * => anon will have no lock associated.
77 1.1 chuck */
78 1.1 chuck struct vm_anon *
79 1.34 thorpej uvm_analloc(void)
80 1.1 chuck {
81 1.33 yamt struct vm_anon *anon;
82 1.1 chuck
83 1.47 ad anon = pool_cache_get(&uvm_anon_cache, PR_NOWAIT);
84 1.33 yamt if (anon) {
85 1.33 yamt KASSERT(anon->an_ref == 0);
86 1.59 rmind KASSERT(anon->an_lock == NULL);
87 1.33 yamt KASSERT(anon->an_page == NULL);
88 1.37 yamt #if defined(VMSWAP)
89 1.33 yamt KASSERT(anon->an_swslot == 0);
90 1.55 rmind #endif
91 1.33 yamt anon->an_ref = 1;
92 1.1 chuck }
93 1.33 yamt return anon;
94 1.1 chuck }
95 1.1 chuck
96 1.1 chuck /*
97 1.59 rmind * uvm_anon_dispose: free any resident page or swap resources of anon.
98 1.1 chuck *
99 1.55 rmind * => anon must be removed from the amap (if anon was in an amap).
100 1.59 rmind * => amap must be locked; we may drop and re-acquire the lock here.
101 1.1 chuck */
102 1.20 chs
103 1.54 rmind static void
104 1.59 rmind uvm_anon_dispose(struct vm_anon *anon)
105 1.1 chuck {
106 1.55 rmind struct vm_page *pg = anon->an_page;
107 1.55 rmind
108 1.59 rmind UVMHIST_FUNC("uvm_anon_dispose"); UVMHIST_CALLED(maphist);
109 1.1 chuck UVMHIST_LOG(maphist,"(anon=0x%x)", anon, 0,0,0);
110 1.1 chuck
111 1.59 rmind KASSERT(mutex_owned(anon->an_lock));
112 1.12 thorpej
113 1.1 chuck /*
114 1.55 rmind * If there is a resident page and it is loaned, then anon may not
115 1.55 rmind * own it. Call out to uvm_anon_lockloanpg() to identify and lock
116 1.55 rmind * the real owner of the page.
117 1.1 chuck */
118 1.1 chuck
119 1.22 chs if (pg && pg->loan_count) {
120 1.54 rmind KASSERT(anon->an_lock != NULL);
121 1.1 chuck pg = uvm_anon_lockloanpg(anon);
122 1.22 chs }
123 1.1 chuck
124 1.1 chuck /*
125 1.55 rmind * Dispose the page, if it is resident.
126 1.1 chuck */
127 1.1 chuck
128 1.1 chuck if (pg) {
129 1.54 rmind KASSERT(anon->an_lock != NULL);
130 1.1 chuck
131 1.1 chuck /*
132 1.55 rmind * If the page is owned by a UVM object (now locked),
133 1.55 rmind * then kill the loan on the page rather than free it,
134 1.55 rmind * and release the object lock.
135 1.1 chuck */
136 1.1 chuck
137 1.1 chuck if (pg->uobject) {
138 1.50 ad mutex_enter(&uvm_pageqlock);
139 1.10 chs KASSERT(pg->loan_count > 0);
140 1.1 chuck pg->loan_count--;
141 1.1 chuck pg->uanon = NULL;
142 1.50 ad mutex_exit(&uvm_pageqlock);
143 1.54 rmind mutex_exit(pg->uobject->vmobjlock);
144 1.1 chuck } else {
145 1.1 chuck
146 1.1 chuck /*
147 1.55 rmind * If page has no UVM object, then anon is the owner,
148 1.55 rmind * and it is already locked.
149 1.1 chuck */
150 1.1 chuck
151 1.18 chs KASSERT((pg->flags & PG_RELEASED) == 0);
152 1.18 chs pmap_page_protect(pg, VM_PROT_NONE);
153 1.29 yamt
154 1.29 yamt /*
155 1.55 rmind * If the page is busy, mark it as PG_RELEASED, so
156 1.55 rmind * that uvm_anon_release(9) would release it later.
157 1.29 yamt */
158 1.29 yamt
159 1.29 yamt if (pg->flags & PG_BUSY) {
160 1.29 yamt pg->flags |= PG_RELEASED;
161 1.58 yamt mutex_obj_hold(anon->an_lock);
162 1.29 yamt return;
163 1.17 chs }
164 1.50 ad mutex_enter(&uvm_pageqlock);
165 1.29 yamt uvm_pagefree(pg);
166 1.50 ad mutex_exit(&uvm_pageqlock);
167 1.18 chs UVMHIST_LOG(maphist, "anon 0x%x, page 0x%x: "
168 1.18 chs "freed now!", anon, pg, 0, 0);
169 1.1 chuck }
170 1.1 chuck }
171 1.55 rmind
172 1.37 yamt #if defined(VMSWAP)
173 1.26 pk if (pg == NULL && anon->an_swslot > 0) {
174 1.55 rmind /* This page is no longer only in swap. */
175 1.45 ad mutex_enter(&uvm_swap_data_lock);
176 1.20 chs KASSERT(uvmexp.swpgonly > 0);
177 1.20 chs uvmexp.swpgonly--;
178 1.45 ad mutex_exit(&uvm_swap_data_lock);
179 1.20 chs }
180 1.55 rmind #endif
181 1.1 chuck
182 1.1 chuck /*
183 1.59 rmind * Free any swap resources, leave a page replacement hint.
184 1.1 chuck */
185 1.18 chs
186 1.2 chs uvm_anon_dropswap(anon);
187 1.39 yamt uvmpdpol_anfree(anon);
188 1.59 rmind UVMHIST_LOG(maphist,"<- done!",0,0,0,0);
189 1.59 rmind }
190 1.59 rmind
191 1.59 rmind /*
192 1.59 rmind * uvm_anon_free: free a single anon.
193 1.59 rmind *
194 1.59 rmind * => anon must be already disposed.
195 1.59 rmind */
196 1.59 rmind void
197 1.59 rmind uvm_anon_free(struct vm_anon *anon)
198 1.59 rmind {
199 1.39 yamt
200 1.59 rmind KASSERT(anon->an_ref == 0);
201 1.59 rmind KASSERT(anon->an_lock == NULL);
202 1.33 yamt KASSERT(anon->an_page == NULL);
203 1.37 yamt #if defined(VMSWAP)
204 1.31 yamt KASSERT(anon->an_swslot == 0);
205 1.55 rmind #endif
206 1.47 ad pool_cache_put(&uvm_anon_cache, anon);
207 1.2 chs }
208 1.2 chs
209 1.55 rmind /*
210 1.59 rmind * uvm_anon_freelst: free a linked list of anon structures.
211 1.59 rmind *
212 1.59 rmind * => anon must be locked, we will unlock it.
213 1.55 rmind */
214 1.54 rmind void
215 1.59 rmind uvm_anon_freelst(struct vm_amap *amap, struct vm_anon *anonlst)
216 1.54 rmind {
217 1.59 rmind struct vm_anon *anon = anonlst;
218 1.59 rmind
219 1.59 rmind KASSERT(mutex_owned(amap->am_lock));
220 1.54 rmind
221 1.59 rmind while (anon) {
222 1.59 rmind uvm_anon_dispose(anon);
223 1.59 rmind anon = anon->an_link;
224 1.59 rmind }
225 1.59 rmind amap_unlock(amap);
226 1.59 rmind
227 1.59 rmind while (anonlst) {
228 1.59 rmind anon = anonlst->an_link;
229 1.59 rmind /* Note: clears an_ref as well. */
230 1.59 rmind anonlst->an_link = NULL;
231 1.59 rmind anonlst->an_lock = NULL;
232 1.59 rmind uvm_anon_free(anonlst);
233 1.59 rmind anonlst = anon;
234 1.54 rmind }
235 1.54 rmind }
236 1.54 rmind
237 1.2 chs /*
238 1.55 rmind * uvm_anon_lockloanpg: given a locked anon, lock its resident page owner.
239 1.1 chuck *
240 1.1 chuck * => anon is locked by caller
241 1.1 chuck * => on return: anon is locked
242 1.1 chuck * if there is a resident page:
243 1.1 chuck * if it has a uobject, it is locked by us
244 1.1 chuck * if it is ownerless, we take over as owner
245 1.1 chuck * we return the resident page (it can change during
246 1.1 chuck * this function)
247 1.1 chuck * => note that the only time an anon has an ownerless resident page
248 1.1 chuck * is if the page was loaned from a uvm_object and the uvm_object
249 1.1 chuck * disowned it
250 1.1 chuck * => this only needs to be called when you want to do an operation
251 1.1 chuck * on an anon's resident page and that page has a non-zero loan
252 1.1 chuck * count.
253 1.1 chuck */
254 1.1 chuck struct vm_page *
255 1.34 thorpej uvm_anon_lockloanpg(struct vm_anon *anon)
256 1.1 chuck {
257 1.1 chuck struct vm_page *pg;
258 1.43 thorpej bool locked = false;
259 1.1 chuck
260 1.54 rmind KASSERT(mutex_owned(anon->an_lock));
261 1.12 thorpej
262 1.1 chuck /*
263 1.1 chuck * loop while we have a resident page that has a non-zero loan count.
264 1.1 chuck * if we successfully get our lock, we will "break" the loop.
265 1.1 chuck * note that the test for pg->loan_count is not protected -- this
266 1.1 chuck * may produce false positive results. note that a false positive
267 1.1 chuck * result may cause us to do more work than we need to, but it will
268 1.1 chuck * not produce an incorrect result.
269 1.1 chuck */
270 1.1 chuck
271 1.33 yamt while (((pg = anon->an_page) != NULL) && pg->loan_count != 0) {
272 1.1 chuck
273 1.1 chuck /*
274 1.1 chuck * quickly check to see if the page has an object before
275 1.1 chuck * bothering to lock the page queues. this may also produce
276 1.1 chuck * a false positive result, but that's ok because we do a real
277 1.1 chuck * check after that.
278 1.1 chuck */
279 1.1 chuck
280 1.1 chuck if (pg->uobject) {
281 1.50 ad mutex_enter(&uvm_pageqlock);
282 1.18 chs if (pg->uobject) {
283 1.1 chuck locked =
284 1.54 rmind mutex_tryenter(pg->uobject->vmobjlock);
285 1.1 chuck } else {
286 1.1 chuck /* object disowned before we got PQ lock */
287 1.43 thorpej locked = true;
288 1.1 chuck }
289 1.50 ad mutex_exit(&uvm_pageqlock);
290 1.1 chuck
291 1.1 chuck /*
292 1.1 chuck * if we didn't get a lock (try lock failed), then we
293 1.1 chuck * toggle our anon lock and try again
294 1.1 chuck */
295 1.1 chuck
296 1.1 chuck if (!locked) {
297 1.1 chuck /*
298 1.1 chuck * someone locking the object has a chance to
299 1.1 chuck * lock us right now
300 1.54 rmind *
301 1.54 rmind * XXX Better than yielding but inadequate.
302 1.1 chuck */
303 1.54 rmind kpause("livelock", false, 1, anon->an_lock);
304 1.10 chs continue;
305 1.1 chuck }
306 1.1 chuck }
307 1.1 chuck
308 1.1 chuck /*
309 1.55 rmind * If page is un-owned i.e. the object dropped its ownership,
310 1.55 rmind * then we have to take the ownership.
311 1.1 chuck */
312 1.1 chuck
313 1.1 chuck if (pg->uobject == NULL && (pg->pqflags & PQ_ANON) == 0) {
314 1.50 ad mutex_enter(&uvm_pageqlock);
315 1.18 chs pg->pqflags |= PQ_ANON;
316 1.18 chs pg->loan_count--;
317 1.50 ad mutex_exit(&uvm_pageqlock);
318 1.1 chuck }
319 1.1 chuck break;
320 1.1 chuck }
321 1.55 rmind return pg;
322 1.5 chs }
323 1.5 chs
324 1.37 yamt #if defined(VMSWAP)
325 1.37 yamt
326 1.5 chs /*
327 1.55 rmind * uvm_anon_pagein: fetch an anon's page.
328 1.5 chs *
329 1.5 chs * => anon must be locked, and is unlocked upon return.
330 1.43 thorpej * => returns true if pagein was aborted due to lack of memory.
331 1.5 chs */
332 1.5 chs
333 1.42 thorpej bool
334 1.57 rmind uvm_anon_pagein(struct vm_amap *amap, struct vm_anon *anon)
335 1.5 chs {
336 1.5 chs struct vm_page *pg;
337 1.5 chs struct uvm_object *uobj;
338 1.8 thorpej
339 1.54 rmind KASSERT(mutex_owned(anon->an_lock));
340 1.57 rmind KASSERT(anon->an_lock == amap->am_lock);
341 1.12 thorpej
342 1.8 thorpej /*
343 1.55 rmind * Get the page of the anon.
344 1.8 thorpej */
345 1.5 chs
346 1.57 rmind switch (uvmfault_anonget(NULL, amap, anon)) {
347 1.16 chs case 0:
348 1.55 rmind /* Success - we have the page. */
349 1.55 rmind KASSERT(mutex_owned(anon->an_lock));
350 1.5 chs break;
351 1.16 chs case EIO:
352 1.16 chs case ERESTART:
353 1.5 chs /*
354 1.55 rmind * Nothing more to do on errors. ERESTART means that the
355 1.55 rmind * anon was freed.
356 1.5 chs */
357 1.43 thorpej return false;
358 1.25 pk default:
359 1.43 thorpej return true;
360 1.5 chs }
361 1.5 chs
362 1.5 chs /*
363 1.55 rmind * Mark the page as dirty, clear its swslot and un-busy it.
364 1.5 chs */
365 1.5 chs
366 1.33 yamt pg = anon->an_page;
367 1.5 chs uobj = pg->uobject;
368 1.55 rmind if (anon->an_swslot > 0) {
369 1.23 pk uvm_swap_free(anon->an_swslot, 1);
370 1.55 rmind }
371 1.5 chs anon->an_swslot = 0;
372 1.55 rmind pg->flags &= ~PG_CLEAN;
373 1.5 chs
374 1.5 chs /*
375 1.55 rmind * Deactivate the page (to put it on a page queue).
376 1.5 chs */
377 1.5 chs
378 1.50 ad mutex_enter(&uvm_pageqlock);
379 1.55 rmind if (pg->wire_count == 0) {
380 1.24 pk uvm_pagedeactivate(pg);
381 1.55 rmind }
382 1.50 ad mutex_exit(&uvm_pageqlock);
383 1.25 pk
384 1.25 pk if (pg->flags & PG_WANTED) {
385 1.55 rmind pg->flags &= ~PG_WANTED;
386 1.25 pk wakeup(pg);
387 1.25 pk }
388 1.5 chs
389 1.54 rmind mutex_exit(anon->an_lock);
390 1.5 chs if (uobj) {
391 1.54 rmind mutex_exit(uobj->vmobjlock);
392 1.5 chs }
393 1.43 thorpej return false;
394 1.1 chuck }
395 1.29 yamt
396 1.55 rmind /*
397 1.55 rmind * uvm_anon_dropswap: release any swap resources from this anon.
398 1.55 rmind *
399 1.55 rmind * => anon must be locked or have a reference count of 0.
400 1.55 rmind */
401 1.55 rmind void
402 1.55 rmind uvm_anon_dropswap(struct vm_anon *anon)
403 1.55 rmind {
404 1.55 rmind UVMHIST_FUNC("uvm_anon_dropswap"); UVMHIST_CALLED(maphist);
405 1.55 rmind
406 1.55 rmind if (anon->an_swslot == 0)
407 1.55 rmind return;
408 1.55 rmind
409 1.55 rmind UVMHIST_LOG(maphist,"freeing swap for anon %p, paged to swslot 0x%x",
410 1.55 rmind anon, anon->an_swslot, 0, 0);
411 1.55 rmind uvm_swap_free(anon->an_swslot, 1);
412 1.55 rmind anon->an_swslot = 0;
413 1.55 rmind }
414 1.55 rmind
415 1.55 rmind #endif
416 1.37 yamt
417 1.29 yamt /*
418 1.29 yamt * uvm_anon_release: release an anon and its page.
419 1.29 yamt *
420 1.55 rmind * => anon should not have any references.
421 1.55 rmind * => anon must be locked.
422 1.29 yamt */
423 1.29 yamt
424 1.29 yamt void
425 1.34 thorpej uvm_anon_release(struct vm_anon *anon)
426 1.29 yamt {
427 1.33 yamt struct vm_page *pg = anon->an_page;
428 1.59 rmind kmutex_t *lock;
429 1.29 yamt
430 1.59 rmind KASSERT(mutex_owned(anon->an_lock));
431 1.29 yamt KASSERT(pg != NULL);
432 1.29 yamt KASSERT((pg->flags & PG_RELEASED) != 0);
433 1.29 yamt KASSERT((pg->flags & PG_BUSY) != 0);
434 1.29 yamt KASSERT(pg->uobject == NULL);
435 1.29 yamt KASSERT(pg->uanon == anon);
436 1.29 yamt KASSERT(pg->loan_count == 0);
437 1.29 yamt KASSERT(anon->an_ref == 0);
438 1.29 yamt
439 1.50 ad mutex_enter(&uvm_pageqlock);
440 1.29 yamt uvm_pagefree(pg);
441 1.50 ad mutex_exit(&uvm_pageqlock);
442 1.59 rmind mutex_exit(anon->an_lock);
443 1.29 yamt
444 1.33 yamt KASSERT(anon->an_page == NULL);
445 1.29 yamt
446 1.59 rmind /* Note: extra reference is held for PG_RELEASED case. */
447 1.59 rmind lock = anon->an_lock;
448 1.59 rmind anon->an_lock = NULL;
449 1.59 rmind uvm_anon_free(anon);
450 1.56 yamt mutex_obj_free(lock);
451 1.29 yamt }
452