xref: /netbsd-src/sys/uvm/uvm_fault.c (revision daf6c4152fcddc27c445489775ed1f66ab4ea9a9)
1 /*	$NetBSD: uvm_fault.c,v 1.182 2011/02/10 21:05:52 skrll 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  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
28  */
29 
30 /*
31  * uvm_fault.c: fault handler
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.182 2011/02/10 21:05:52 skrll Exp $");
36 
37 #include "opt_uvmhist.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/proc.h>
43 #include <sys/malloc.h>
44 #include <sys/mman.h>
45 
46 #include <uvm/uvm.h>
47 
48 /*
49  *
50  * a word on page faults:
51  *
52  * types of page faults we handle:
53  *
54  * CASE 1: upper layer faults                   CASE 2: lower layer faults
55  *
56  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
57  *    read/write1     write>1                  read/write   +-cow_write/zero
58  *         |             |                         |        |
59  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
60  * amap |  V  |       |  ---------> new |          |        | |  ^  |
61  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
62  *                                                 |        |    |
63  *      +-----+       +-----+                   +--|--+     | +--|--+
64  * uobj | d/c |       | d/c |                   |  V  |     +----+  |
65  *      +-----+       +-----+                   +-----+       +-----+
66  *
67  * d/c = don't care
68  *
69  *   case [0]: layerless fault
70  *	no amap or uobj is present.   this is an error.
71  *
72  *   case [1]: upper layer fault [anon active]
73  *     1A: [read] or [write with anon->an_ref == 1]
74  *		I/O takes place in upper level anon and uobj is not touched.
75  *     1B: [write with anon->an_ref > 1]
76  *		new anon is alloc'd and data is copied off ["COW"]
77  *
78  *   case [2]: lower layer fault [uobj]
79  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
80  *		I/O takes place directly in object.
81  *     2B: [write to copy_on_write] or [read on NULL uobj]
82  *		data is "promoted" from uobj to a new anon.
83  *		if uobj is null, then we zero fill.
84  *
85  * we follow the standard UVM locking protocol ordering:
86  *
87  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
88  * we hold a PG_BUSY page if we unlock for I/O
89  *
90  *
91  * the code is structured as follows:
92  *
93  *     - init the "IN" params in the ufi structure
94  *   ReFault: (ERESTART returned to the loop in uvm_fault_internal)
95  *     - do lookups [locks maps], check protection, handle needs_copy
96  *     - check for case 0 fault (error)
97  *     - establish "range" of fault
98  *     - if we have an amap lock it and extract the anons
99  *     - if sequential advice deactivate pages behind us
100  *     - at the same time check pmap for unmapped areas and anon for pages
101  *	 that we could map in (and do map it if found)
102  *     - check object for resident pages that we could map in
103  *     - if (case 2) goto Case2
104  *     - >>> handle case 1
105  *           - ensure source anon is resident in RAM
106  *           - if case 1B alloc new anon and copy from source
107  *           - map the correct page in
108  *   Case2:
109  *     - >>> handle case 2
110  *           - ensure source page is resident (if uobj)
111  *           - if case 2B alloc new anon and copy from source (could be zero
112  *		fill if uobj == NULL)
113  *           - map the correct page in
114  *     - done!
115  *
116  * note on paging:
117  *   if we have to do I/O we place a PG_BUSY page in the correct object,
118  * unlock everything, and do the I/O.   when I/O is done we must reverify
119  * the state of the world before assuming that our data structures are
120  * valid.   [because mappings could change while the map is unlocked]
121  *
122  *  alternative 1: unbusy the page in question and restart the page fault
123  *    from the top (ReFault).   this is easy but does not take advantage
124  *    of the information that we already have from our previous lookup,
125  *    although it is possible that the "hints" in the vm_map will help here.
126  *
127  * alternative 2: the system already keeps track of a "version" number of
128  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
129  *    mapping) you bump the version number up by one...]   so, we can save
130  *    the version number of the map before we release the lock and start I/O.
131  *    then when I/O is done we can relock and check the version numbers
132  *    to see if anything changed.    this might save us some over 1 because
133  *    we don't have to unbusy the page and may be less compares(?).
134  *
135  * alternative 3: put in backpointers or a way to "hold" part of a map
136  *    in place while I/O is in progress.   this could be complex to
137  *    implement (especially with structures like amap that can be referenced
138  *    by multiple map entries, and figuring out what should wait could be
139  *    complex as well...).
140  *
141  * we use alternative 2.  given that we are multi-threaded now we may want
142  * to reconsider the choice.
143  */
144 
145 /*
146  * local data structures
147  */
148 
149 struct uvm_advice {
150 	int advice;
151 	int nback;
152 	int nforw;
153 };
154 
155 /*
156  * page range array:
157  * note: index in array must match "advice" value
158  * XXX: borrowed numbers from freebsd.   do they work well for us?
159  */
160 
161 static const struct uvm_advice uvmadvice[] = {
162 	{ MADV_NORMAL, 3, 4 },
163 	{ MADV_RANDOM, 0, 0 },
164 	{ MADV_SEQUENTIAL, 8, 7},
165 };
166 
167 #define UVM_MAXRANGE 16	/* must be MAX() of nback+nforw+1 */
168 
169 /*
170  * private prototypes
171  */
172 
173 /*
174  * inline functions
175  */
176 
177 /*
178  * uvmfault_anonflush: try and deactivate pages in specified anons
179  *
180  * => does not have to deactivate page if it is busy
181  */
182 
183 static inline void
184 uvmfault_anonflush(struct vm_anon **anons, int n)
185 {
186 	int lcv;
187 	struct vm_page *pg;
188 
189 	for (lcv = 0; lcv < n; lcv++) {
190 		if (anons[lcv] == NULL)
191 			continue;
192 		mutex_enter(&anons[lcv]->an_lock);
193 		pg = anons[lcv]->an_page;
194 		if (pg && (pg->flags & PG_BUSY) == 0) {
195 			mutex_enter(&uvm_pageqlock);
196 			if (pg->wire_count == 0) {
197 				uvm_pagedeactivate(pg);
198 			}
199 			mutex_exit(&uvm_pageqlock);
200 		}
201 		mutex_exit(&anons[lcv]->an_lock);
202 	}
203 }
204 
205 /*
206  * normal functions
207  */
208 
209 /*
210  * uvmfault_amapcopy: clear "needs_copy" in a map.
211  *
212  * => called with VM data structures unlocked (usually, see below)
213  * => we get a write lock on the maps and clear needs_copy for a VA
214  * => if we are out of RAM we sleep (waiting for more)
215  */
216 
217 static void
218 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
219 {
220 	for (;;) {
221 
222 		/*
223 		 * no mapping?  give up.
224 		 */
225 
226 		if (uvmfault_lookup(ufi, true) == false)
227 			return;
228 
229 		/*
230 		 * copy if needed.
231 		 */
232 
233 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
234 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
235 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
236 
237 		/*
238 		 * didn't work?  must be out of RAM.   unlock and sleep.
239 		 */
240 
241 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
242 			uvmfault_unlockmaps(ufi, true);
243 			uvm_wait("fltamapcopy");
244 			continue;
245 		}
246 
247 		/*
248 		 * got it!   unlock and return.
249 		 */
250 
251 		uvmfault_unlockmaps(ufi, true);
252 		return;
253 	}
254 	/*NOTREACHED*/
255 }
256 
257 /*
258  * uvmfault_anonget: get data in an anon into a non-busy, non-released
259  * page in that anon.
260  *
261  * => maps, amap, and anon locked by caller.
262  * => if we fail (result != 0) we unlock everything.
263  * => if we are successful, we return with everything still locked.
264  * => we don't move the page on the queues [gets moved later]
265  * => if we allocate a new page [we_own], it gets put on the queues.
266  *    either way, the result is that the page is on the queues at return time
267  * => for pages which are on loan from a uvm_object (and thus are not
268  *    owned by the anon): if successful, we return with the owning object
269  *    locked.   the caller must unlock this object when it unlocks everything
270  *    else.
271  */
272 
273 int
274 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
275     struct vm_anon *anon)
276 {
277 	bool we_own;	/* we own anon's page? */
278 	bool locked;	/* did we relock? */
279 	struct vm_page *pg;
280 	int error;
281 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
282 
283 	KASSERT(mutex_owned(&anon->an_lock));
284 
285 	error = 0;
286 	uvmexp.fltanget++;
287         /* bump rusage counters */
288 	if (anon->an_page)
289 		curlwp->l_ru.ru_minflt++;
290 	else
291 		curlwp->l_ru.ru_majflt++;
292 
293 	/*
294 	 * loop until we get it, or fail.
295 	 */
296 
297 	for (;;) {
298 		we_own = false;		/* true if we set PG_BUSY on a page */
299 		pg = anon->an_page;
300 
301 		/*
302 		 * if there is a resident page and it is loaned, then anon
303 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
304 		 * the real owner of the page has been identified and locked.
305 		 */
306 
307 		if (pg && pg->loan_count)
308 			pg = uvm_anon_lockloanpg(anon);
309 
310 		/*
311 		 * page there?   make sure it is not busy/released.
312 		 */
313 
314 		if (pg) {
315 
316 			/*
317 			 * at this point, if the page has a uobject [meaning
318 			 * we have it on loan], then that uobject is locked
319 			 * by us!   if the page is busy, we drop all the
320 			 * locks (including uobject) and try again.
321 			 */
322 
323 			if ((pg->flags & PG_BUSY) == 0) {
324 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
325 				return (0);
326 			}
327 			pg->flags |= PG_WANTED;
328 			uvmexp.fltpgwait++;
329 
330 			/*
331 			 * the last unlock must be an atomic unlock+wait on
332 			 * the owner of page
333 			 */
334 
335 			if (pg->uobject) {	/* owner is uobject ? */
336 				uvmfault_unlockall(ufi, amap, NULL, anon);
337 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
338 				    0,0,0);
339 				UVM_UNLOCK_AND_WAIT(pg,
340 				    &pg->uobject->vmobjlock,
341 				    false, "anonget1",0);
342 			} else {
343 				/* anon owns page */
344 				uvmfault_unlockall(ufi, amap, NULL, NULL);
345 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
346 				    0,0,0);
347 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
348 				    "anonget2",0);
349 			}
350 		} else {
351 #if defined(VMSWAP)
352 
353 			/*
354 			 * no page, we must try and bring it in.
355 			 */
356 
357 			pg = uvm_pagealloc(NULL,
358 			    ufi != NULL ? ufi->orig_rvaddr : 0,
359 			    anon, UVM_FLAG_COLORMATCH);
360 			if (pg == NULL) {		/* out of RAM.  */
361 				uvmfault_unlockall(ufi, amap, NULL, anon);
362 				uvmexp.fltnoram++;
363 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
364 				    0,0,0);
365 				if (!uvm_reclaimable()) {
366 					return ENOMEM;
367 				}
368 				uvm_wait("flt_noram1");
369 			} else {
370 				/* we set the PG_BUSY bit */
371 				we_own = true;
372 				uvmfault_unlockall(ufi, amap, NULL, anon);
373 
374 				/*
375 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
376 				 * page into the uvm_swap_get function with
377 				 * all data structures unlocked.  note that
378 				 * it is ok to read an_swslot here because
379 				 * we hold PG_BUSY on the page.
380 				 */
381 				uvmexp.pageins++;
382 				error = uvm_swap_get(pg, anon->an_swslot,
383 				    PGO_SYNCIO);
384 
385 				/*
386 				 * we clean up after the i/o below in the
387 				 * "we_own" case
388 				 */
389 			}
390 #else /* defined(VMSWAP) */
391 			panic("%s: no page", __func__);
392 #endif /* defined(VMSWAP) */
393 		}
394 
395 		/*
396 		 * now relock and try again
397 		 */
398 
399 		locked = uvmfault_relock(ufi);
400 		if (locked && amap != NULL) {
401 			amap_lock(amap);
402 		}
403 		if (locked || we_own)
404 			mutex_enter(&anon->an_lock);
405 
406 		/*
407 		 * if we own the page (i.e. we set PG_BUSY), then we need
408 		 * to clean up after the I/O. there are three cases to
409 		 * consider:
410 		 *   [1] page released during I/O: free anon and ReFault.
411 		 *   [2] I/O not OK.   free the page and cause the fault
412 		 *       to fail.
413 		 *   [3] I/O OK!   activate the page and sync with the
414 		 *       non-we_own case (i.e. drop anon lock if not locked).
415 		 */
416 
417 		if (we_own) {
418 #if defined(VMSWAP)
419 			if (pg->flags & PG_WANTED) {
420 				wakeup(pg);
421 			}
422 			if (error) {
423 
424 				/*
425 				 * remove the swap slot from the anon
426 				 * and mark the anon as having no real slot.
427 				 * don't free the swap slot, thus preventing
428 				 * it from being used again.
429 				 */
430 
431 				if (anon->an_swslot > 0)
432 					uvm_swap_markbad(anon->an_swslot, 1);
433 				anon->an_swslot = SWSLOT_BAD;
434 
435 				if ((pg->flags & PG_RELEASED) != 0)
436 					goto released;
437 
438 				/*
439 				 * note: page was never !PG_BUSY, so it
440 				 * can't be mapped and thus no need to
441 				 * pmap_page_protect it...
442 				 */
443 
444 				mutex_enter(&uvm_pageqlock);
445 				uvm_pagefree(pg);
446 				mutex_exit(&uvm_pageqlock);
447 
448 				if (locked)
449 					uvmfault_unlockall(ufi, amap, NULL,
450 					    anon);
451 				else
452 					mutex_exit(&anon->an_lock);
453 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
454 				return error;
455 			}
456 
457 			if ((pg->flags & PG_RELEASED) != 0) {
458 released:
459 				KASSERT(anon->an_ref == 0);
460 
461 				/*
462 				 * released while we unlocked amap.
463 				 */
464 
465 				if (locked)
466 					uvmfault_unlockall(ufi, amap, NULL,
467 					    NULL);
468 
469 				uvm_anon_release(anon);
470 
471 				if (error) {
472 					UVMHIST_LOG(maphist,
473 					    "<- ERROR/RELEASED", 0,0,0,0);
474 					return error;
475 				}
476 
477 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
478 				return ERESTART;
479 			}
480 
481 			/*
482 			 * we've successfully read the page, activate it.
483 			 */
484 
485 			mutex_enter(&uvm_pageqlock);
486 			uvm_pageactivate(pg);
487 			mutex_exit(&uvm_pageqlock);
488 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
489 			UVM_PAGE_OWN(pg, NULL);
490 			if (!locked)
491 				mutex_exit(&anon->an_lock);
492 #else /* defined(VMSWAP) */
493 			panic("%s: we_own", __func__);
494 #endif /* defined(VMSWAP) */
495 		}
496 
497 		/*
498 		 * we were not able to relock.   restart fault.
499 		 */
500 
501 		if (!locked) {
502 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
503 			return (ERESTART);
504 		}
505 
506 		/*
507 		 * verify no one has touched the amap and moved the anon on us.
508 		 */
509 
510 		if (ufi != NULL &&
511 		    amap_lookup(&ufi->entry->aref,
512 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
513 
514 			uvmfault_unlockall(ufi, amap, NULL, anon);
515 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
516 			return (ERESTART);
517 		}
518 
519 		/*
520 		 * try it again!
521 		 */
522 
523 		uvmexp.fltanretry++;
524 		continue;
525 	}
526 	/*NOTREACHED*/
527 }
528 
529 /*
530  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
531  *
532  *	1. allocate an anon and a page.
533  *	2. fill its contents.
534  *	3. put it into amap.
535  *
536  * => if we fail (result != 0) we unlock everything.
537  * => on success, return a new locked anon via 'nanon'.
538  *    (*nanon)->an_page will be a resident, locked, dirty page.
539  */
540 
541 static int
542 uvmfault_promote(struct uvm_faultinfo *ufi,
543     struct vm_anon *oanon,
544     struct vm_page *uobjpage,
545     struct vm_anon **nanon, /* OUT: allocated anon */
546     struct vm_anon **spare)
547 {
548 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
549 	struct uvm_object *uobj;
550 	struct vm_anon *anon;
551 	struct vm_page *pg;
552 	struct vm_page *opg;
553 	int error;
554 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
555 
556 	if (oanon) {
557 		/* anon COW */
558 		opg = oanon->an_page;
559 		KASSERT(opg != NULL);
560 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
561 	} else if (uobjpage != PGO_DONTCARE) {
562 		/* object-backed COW */
563 		opg = uobjpage;
564 	} else {
565 		/* ZFOD */
566 		opg = NULL;
567 	}
568 	if (opg != NULL) {
569 		uobj = opg->uobject;
570 	} else {
571 		uobj = NULL;
572 	}
573 
574 	KASSERT(amap != NULL);
575 	KASSERT(uobjpage != NULL);
576 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
577 	KASSERT(mutex_owned(&amap->am_l));
578 	KASSERT(oanon == NULL || mutex_owned(&oanon->an_lock));
579 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
580 #if 0
581 	KASSERT(*spare == NULL || !mutex_owned(&(*spare)->an_lock));
582 #endif
583 
584 	if (*spare != NULL) {
585 		anon = *spare;
586 		*spare = NULL;
587 		mutex_enter(&anon->an_lock);
588 	} else if (ufi->map != kernel_map) {
589 		anon = uvm_analloc();
590 	} else {
591 		UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0);
592 
593 		/*
594 		 * we can't allocate anons with kernel_map locked.
595 		 */
596 
597 		uvm_page_unbusy(&uobjpage, 1);
598 		uvmfault_unlockall(ufi, amap, uobj, oanon);
599 
600 		*spare = uvm_analloc();
601 		if (*spare == NULL) {
602 			goto nomem;
603 		}
604 		mutex_exit(&(*spare)->an_lock);
605 		error = ERESTART;
606 		goto done;
607 	}
608 	if (anon) {
609 
610 		/*
611 		 * The new anon is locked.
612 		 *
613 		 * if opg == NULL, we want a zero'd, dirty page,
614 		 * so have uvm_pagealloc() do that for us.
615 		 */
616 
617 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
618 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
619 	} else {
620 		pg = NULL;
621 	}
622 
623 	/*
624 	 * out of memory resources?
625 	 */
626 
627 	if (pg == NULL) {
628 		/* save anon for the next try. */
629 		if (anon != NULL) {
630 			mutex_exit(&anon->an_lock);
631 			*spare = anon;
632 		}
633 
634 		/* unlock and fail ... */
635 		uvm_page_unbusy(&uobjpage, 1);
636 		uvmfault_unlockall(ufi, amap, uobj, oanon);
637 nomem:
638 		if (!uvm_reclaimable()) {
639 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
640 			uvmexp.fltnoanon++;
641 			error = ENOMEM;
642 			goto done;
643 		}
644 
645 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
646 		uvmexp.fltnoram++;
647 		uvm_wait("flt_noram5");
648 		error = ERESTART;
649 		goto done;
650 	}
651 
652 	/* copy page [pg now dirty] */
653 	if (opg) {
654 		uvm_pagecopy(opg, pg);
655 	}
656 
657 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
658 	    oanon != NULL);
659 
660 	*nanon = anon;
661 	error = 0;
662 done:
663 	return error;
664 }
665 
666 
667 /*
668  *   F A U L T   -   m a i n   e n t r y   p o i n t
669  */
670 
671 /*
672  * uvm_fault: page fault handler
673  *
674  * => called from MD code to resolve a page fault
675  * => VM data structures usually should be unlocked.   however, it is
676  *	possible to call here with the main map locked if the caller
677  *	gets a write lock, sets it recusive, and then calls us (c.f.
678  *	uvm_map_pageable).   this should be avoided because it keeps
679  *	the map locked off during I/O.
680  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
681  */
682 
683 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
684 			 ~VM_PROT_WRITE : VM_PROT_ALL)
685 
686 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
687 #define UVM_FAULT_WIRE		(1 << 0)
688 #define UVM_FAULT_MAXPROT	(1 << 1)
689 
690 struct uvm_faultctx {
691 	vm_prot_t access_type;
692 	vm_prot_t enter_prot;
693 	vaddr_t startva;
694 	int npages;
695 	int centeridx;
696 	struct vm_anon *anon_spare;
697 	bool wire_mapping;
698 	bool narrow;
699 	bool wire_paging;
700 	bool cow_now;
701 	bool promote;
702 };
703 
704 static inline int	uvm_fault_check(
705 			    struct uvm_faultinfo *, struct uvm_faultctx *,
706 			    struct vm_anon ***, bool);
707 
708 static int		uvm_fault_upper(
709 			    struct uvm_faultinfo *, struct uvm_faultctx *,
710 			    struct vm_anon **);
711 static inline int	uvm_fault_upper_lookup(
712 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
713 			    struct vm_anon **, struct vm_page **);
714 static inline void	uvm_fault_upper_neighbor(
715 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
716 			    vaddr_t, struct vm_page *, bool);
717 static inline int	uvm_fault_upper_loan(
718 			    struct uvm_faultinfo *, struct uvm_faultctx *,
719 			    struct vm_anon *, struct uvm_object **);
720 static inline int	uvm_fault_upper_promote(
721 			    struct uvm_faultinfo *, struct uvm_faultctx *,
722 			    struct uvm_object *, struct vm_anon *);
723 static inline int	uvm_fault_upper_direct(
724 			    struct uvm_faultinfo *, struct uvm_faultctx *,
725 			    struct uvm_object *, struct vm_anon *);
726 static int		uvm_fault_upper_enter(
727 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
728 			    struct uvm_object *, struct vm_anon *,
729 			    struct vm_page *, struct vm_anon *);
730 static inline void	uvm_fault_upper_done(
731 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
732 			    struct vm_anon *, struct vm_page *);
733 
734 static int		uvm_fault_lower(
735 			    struct uvm_faultinfo *, struct uvm_faultctx *,
736 			    struct vm_page **);
737 static inline void	uvm_fault_lower_lookup(
738 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
739 			    struct vm_page **);
740 static inline void	uvm_fault_lower_neighbor(
741 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
742 			    vaddr_t, struct vm_page *, bool);
743 static inline int	uvm_fault_lower_io(
744 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
745 			    struct uvm_object **, struct vm_page **);
746 static inline int	uvm_fault_lower_direct(
747 			    struct uvm_faultinfo *, struct uvm_faultctx *,
748 			    struct uvm_object *, struct vm_page *);
749 static inline int	uvm_fault_lower_direct_loan(
750 			    struct uvm_faultinfo *, struct uvm_faultctx *,
751 			    struct uvm_object *, struct vm_page **,
752 			    struct vm_page **);
753 static inline int	uvm_fault_lower_promote(
754 			    struct uvm_faultinfo *, struct uvm_faultctx *,
755 			    struct uvm_object *, struct vm_page *);
756 static int		uvm_fault_lower_enter(
757 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
758 			    struct uvm_object *,
759 			    struct vm_anon *, struct vm_page *,
760 			    struct vm_page *);
761 static inline void	uvm_fault_lower_done(
762 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
763 			    struct uvm_object *, struct vm_page *);
764 
765 int
766 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
767     vm_prot_t access_type, int fault_flag)
768 {
769 	struct uvm_faultinfo ufi;
770 	struct uvm_faultctx flt = {
771 		.access_type = access_type,
772 
773 		/* don't look for neighborhood * pages on "wire" fault */
774 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
775 
776 		/* "wire" fault causes wiring of both mapping and paging */
777 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
778 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
779 	};
780 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
781 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
782 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
783 	int error;
784 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
785 
786 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)",
787 	      orig_map, vaddr, access_type, fault_flag);
788 
789 	curcpu()->ci_data.cpu_nfault++;
790 
791 	/*
792 	 * init the IN parameters in the ufi
793 	 */
794 
795 	ufi.orig_map = orig_map;
796 	ufi.orig_rvaddr = trunc_page(vaddr);
797 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
798 
799 	error = ERESTART;
800 	while (error == ERESTART) {
801 		anons = anons_store;
802 		pages = pages_store;
803 
804 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
805 		if (error != 0)
806 			continue;
807 
808 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
809 		if (error != 0)
810 			continue;
811 
812 		if (pages[flt.centeridx] == PGO_DONTCARE)
813 			error = uvm_fault_upper(&ufi, &flt, anons);
814 		else {
815 			struct uvm_object * const uobj =
816 			    ufi.entry->object.uvm_obj;
817 
818 			if (uobj && uobj->pgops->pgo_fault != NULL) {
819 				/*
820 				 * invoke "special" fault routine.
821 				 */
822 				mutex_enter(&uobj->vmobjlock);
823 				/* locked: maps(read), amap(if there), uobj */
824 				error = uobj->pgops->pgo_fault(&ufi,
825 				    flt.startva, pages, flt.npages,
826 				    flt.centeridx, flt.access_type,
827 				    PGO_LOCKED|PGO_SYNCIO);
828 
829 				/*
830 				 * locked: nothing, pgo_fault has unlocked
831 				 * everything
832 				 */
833 
834 				/*
835 				 * object fault routine responsible for
836 				 * pmap_update().
837 				 */
838 			} else {
839 				error = uvm_fault_lower(&ufi, &flt, pages);
840 			}
841 		}
842 	}
843 
844 	if (flt.anon_spare != NULL) {
845 		flt.anon_spare->an_ref--;
846 		uvm_anfree(flt.anon_spare);
847 	}
848 	return error;
849 }
850 
851 /*
852  * uvm_fault_check: check prot, handle needs-copy, etc.
853  *
854  *	1. lookup entry.
855  *	2. check protection.
856  *	3. adjust fault condition (mainly for simulated fault).
857  *	4. handle needs-copy (lazy amap copy).
858  *	5. establish range of interest for neighbor fault (aka pre-fault).
859  *	6. look up anons (if amap exists).
860  *	7. flush pages (if MADV_SEQUENTIAL)
861  *
862  * => called with nothing locked.
863  * => if we fail (result != 0) we unlock everything.
864  * => initialize/adjust many members of flt.
865  */
866 
867 static int
868 uvm_fault_check(
869 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
870 	struct vm_anon ***ranons, bool maxprot)
871 {
872 	struct vm_amap *amap;
873 	struct uvm_object *uobj;
874 	vm_prot_t check_prot;
875 	int nback, nforw;
876 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
877 
878 	/*
879 	 * lookup and lock the maps
880 	 */
881 
882 	if (uvmfault_lookup(ufi, false) == false) {
883 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", ufi->orig_rvaddr,
884 		    0,0,0);
885 		return EFAULT;
886 	}
887 	/* locked: maps(read) */
888 
889 #ifdef DIAGNOSTIC
890 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
891 		printf("Page fault on non-pageable map:\n");
892 		printf("ufi->map = %p\n", ufi->map);
893 		printf("ufi->orig_map = %p\n", ufi->orig_map);
894 		printf("ufi->orig_rvaddr = 0x%lx\n", (u_long) ufi->orig_rvaddr);
895 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
896 	}
897 #endif
898 
899 	/*
900 	 * check protection
901 	 */
902 
903 	check_prot = maxprot ?
904 	    ufi->entry->max_protection : ufi->entry->protection;
905 	if ((check_prot & flt->access_type) != flt->access_type) {
906 		UVMHIST_LOG(maphist,
907 		    "<- protection failure (prot=0x%x, access=0x%x)",
908 		    ufi->entry->protection, flt->access_type, 0, 0);
909 		uvmfault_unlockmaps(ufi, false);
910 		return EACCES;
911 	}
912 
913 	/*
914 	 * "enter_prot" is the protection we want to enter the page in at.
915 	 * for certain pages (e.g. copy-on-write pages) this protection can
916 	 * be more strict than ufi->entry->protection.  "wired" means either
917 	 * the entry is wired or we are fault-wiring the pg.
918 	 */
919 
920 	flt->enter_prot = ufi->entry->protection;
921 	if (VM_MAPENT_ISWIRED(ufi->entry))
922 		flt->wire_mapping = true;
923 
924 	if (flt->wire_mapping) {
925 		flt->access_type = flt->enter_prot; /* full access for wired */
926 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
927 	} else {
928 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
929 	}
930 
931 	flt->promote = false;
932 
933 	/*
934 	 * handle "needs_copy" case.   if we need to copy the amap we will
935 	 * have to drop our readlock and relock it with a write lock.  (we
936 	 * need a write lock to change anything in a map entry [e.g.
937 	 * needs_copy]).
938 	 */
939 
940 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
941 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
942 			KASSERT(!maxprot);
943 			/* need to clear */
944 			UVMHIST_LOG(maphist,
945 			    "  need to clear needs_copy and refault",0,0,0,0);
946 			uvmfault_unlockmaps(ufi, false);
947 			uvmfault_amapcopy(ufi);
948 			uvmexp.fltamcopy++;
949 			return ERESTART;
950 
951 		} else {
952 
953 			/*
954 			 * ensure that we pmap_enter page R/O since
955 			 * needs_copy is still true
956 			 */
957 
958 			flt->enter_prot &= ~VM_PROT_WRITE;
959 		}
960 	}
961 
962 	/*
963 	 * identify the players
964 	 */
965 
966 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
967 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
968 
969 	/*
970 	 * check for a case 0 fault.  if nothing backing the entry then
971 	 * error now.
972 	 */
973 
974 	if (amap == NULL && uobj == NULL) {
975 		uvmfault_unlockmaps(ufi, false);
976 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
977 		return EFAULT;
978 	}
979 
980 	/*
981 	 * establish range of interest based on advice from mapper
982 	 * and then clip to fit map entry.   note that we only want
983 	 * to do this the first time through the fault.   if we
984 	 * ReFault we will disable this by setting "narrow" to true.
985 	 */
986 
987 	if (flt->narrow == false) {
988 
989 		/* wide fault (!narrow) */
990 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
991 			 ufi->entry->advice);
992 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
993 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
994 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
995 		/*
996 		 * note: "-1" because we don't want to count the
997 		 * faulting page as forw
998 		 */
999 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1000 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1001 			     PAGE_SHIFT) - 1);
1002 		flt->npages = nback + nforw + 1;
1003 		flt->centeridx = nback;
1004 
1005 		flt->narrow = true;	/* ensure only once per-fault */
1006 
1007 	} else {
1008 
1009 		/* narrow fault! */
1010 		nback = nforw = 0;
1011 		flt->startva = ufi->orig_rvaddr;
1012 		flt->npages = 1;
1013 		flt->centeridx = 0;
1014 
1015 	}
1016 	/* offset from entry's start to pgs' start */
1017 	const voff_t eoff = flt->startva - ufi->entry->start;
1018 
1019 	/* locked: maps(read) */
1020 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
1021 		    flt->narrow, nback, nforw, flt->startva);
1022 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi->entry,
1023 		    amap, uobj, 0);
1024 
1025 	/*
1026 	 * if we've got an amap, lock it and extract current anons.
1027 	 */
1028 
1029 	if (amap) {
1030 		amap_lock(amap);
1031 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1032 	} else {
1033 		*ranons = NULL;	/* to be safe */
1034 	}
1035 
1036 	/* locked: maps(read), amap(if there) */
1037 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1038 
1039 	/*
1040 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1041 	 * now and then forget about them (for the rest of the fault).
1042 	 */
1043 
1044 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1045 
1046 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1047 		    0,0,0,0);
1048 		/* flush back-page anons? */
1049 		if (amap)
1050 			uvmfault_anonflush(*ranons, nback);
1051 
1052 		/* flush object? */
1053 		if (uobj) {
1054 			voff_t uoff;
1055 
1056 			uoff = ufi->entry->offset + eoff;
1057 			mutex_enter(&uobj->vmobjlock);
1058 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1059 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1060 		}
1061 
1062 		/* now forget about the backpages */
1063 		if (amap)
1064 			*ranons += nback;
1065 		flt->startva += (nback << PAGE_SHIFT);
1066 		flt->npages -= nback;
1067 		flt->centeridx = 0;
1068 	}
1069 	/*
1070 	 * => startva is fixed
1071 	 * => npages is fixed
1072 	 */
1073 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1074 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1075 	    flt->startva + (flt->npages << PAGE_SHIFT));
1076 	return 0;
1077 }
1078 
1079 /*
1080  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1081  *
1082  * iterate range of interest:
1083  *	1. check if h/w mapping exists.  if yes, we don't care
1084  *	2. check if anon exists.  if not, page is lower.
1085  *	3. if anon exists, enter h/w mapping for neighbors.
1086  *
1087  * => called with amap locked (if exists).
1088  */
1089 
1090 static int
1091 uvm_fault_upper_lookup(
1092 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1093 	struct vm_anon **anons, struct vm_page **pages)
1094 {
1095 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1096 	int lcv;
1097 	vaddr_t currva;
1098 	bool shadowed;
1099 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1100 
1101 	/* locked: maps(read), amap(if there) */
1102 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1103 
1104 	/*
1105 	 * map in the backpages and frontpages we found in the amap in hopes
1106 	 * of preventing future faults.    we also init the pages[] array as
1107 	 * we go.
1108 	 */
1109 
1110 	currva = flt->startva;
1111 	shadowed = false;
1112 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1113 		/*
1114 		 * don't play with VAs that are already mapped
1115 		 * (except for center)
1116 		 */
1117 		if (lcv != flt->centeridx &&
1118 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1119 			pages[lcv] = PGO_DONTCARE;
1120 			continue;
1121 		}
1122 
1123 		/*
1124 		 * unmapped or center page.   check if any anon at this level.
1125 		 */
1126 		if (amap == NULL || anons[lcv] == NULL) {
1127 			pages[lcv] = NULL;
1128 			continue;
1129 		}
1130 
1131 		/*
1132 		 * check for present page and map if possible.   re-activate it.
1133 		 */
1134 
1135 		pages[lcv] = PGO_DONTCARE;
1136 		if (lcv == flt->centeridx) {	/* save center for later! */
1137 			shadowed = true;
1138 		} else {
1139 			struct vm_anon *anon = anons[lcv];
1140 
1141 			mutex_enter(&anon->an_lock);
1142 			struct vm_page *pg = anon->an_page;
1143 
1144 			/* ignore loaned and busy pages */
1145 			if (pg != NULL && pg->loan_count == 0 &&
1146 			    (pg->flags & PG_BUSY) == 0)
1147 				uvm_fault_upper_neighbor(ufi, flt, currva,
1148 				    pg, anon->an_ref > 1);
1149 			mutex_exit(&anon->an_lock);
1150 		}
1151 	}
1152 
1153 	/* locked: maps(read), amap(if there) */
1154 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1155 	/* (shadowed == true) if there is an anon at the faulting address */
1156 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
1157 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1158 
1159 	/*
1160 	 * note that if we are really short of RAM we could sleep in the above
1161 	 * call to pmap_enter with everything locked.   bad?
1162 	 *
1163 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1164 	 * XXX case.  --thorpej
1165 	 */
1166 
1167 	return 0;
1168 }
1169 
1170 /*
1171  * uvm_fault_upper_neighbor: enter single lower neighbor page.
1172  *
1173  * => called with amap and anon locked.
1174  */
1175 
1176 static void
1177 uvm_fault_upper_neighbor(
1178 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1179 	vaddr_t currva, struct vm_page *pg, bool readonly)
1180 {
1181 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1182 
1183 	/* locked: amap, anon */
1184 
1185 	mutex_enter(&uvm_pageqlock);
1186 	uvm_pageenqueue(pg);
1187 	mutex_exit(&uvm_pageqlock);
1188 	UVMHIST_LOG(maphist,
1189 	    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
1190 	    ufi->orig_map->pmap, currva, pg, 0);
1191 	uvmexp.fltnamap++;
1192 
1193 	/*
1194 	 * Since this page isn't the page that's actually faulting,
1195 	 * ignore pmap_enter() failures; it's not critical that we
1196 	 * enter these right now.
1197 	 */
1198 
1199 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1200 	    VM_PAGE_TO_PHYS(pg),
1201 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1202 	    flt->enter_prot,
1203 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1204 
1205 	pmap_update(ufi->orig_map->pmap);
1206 }
1207 
1208 /*
1209  * uvm_fault_upper: handle upper fault.
1210  *
1211  *	1. acquire anon lock.
1212  *	2. get anon.  let uvmfault_anonget do the dirty work.
1213  *	3. handle loan.
1214  *	4. dispatch direct or promote handlers.
1215  */
1216 
1217 static int
1218 uvm_fault_upper(
1219 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1220 	struct vm_anon **anons)
1221 {
1222 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1223 	struct vm_anon * const anon = anons[flt->centeridx];
1224 	struct uvm_object *uobj;
1225 	int error;
1226 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1227 
1228 	/* locked: maps(read), amap */
1229 	KASSERT(mutex_owned(&amap->am_l));
1230 
1231 	/*
1232 	 * handle case 1: fault on an anon in our amap
1233 	 */
1234 
1235 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1236 	mutex_enter(&anon->an_lock);
1237 
1238 	/* locked: maps(read), amap, anon */
1239 	KASSERT(mutex_owned(&amap->am_l));
1240 	KASSERT(mutex_owned(&anon->an_lock));
1241 
1242 	/*
1243 	 * no matter if we have case 1A or case 1B we are going to need to
1244 	 * have the anon's memory resident.   ensure that now.
1245 	 */
1246 
1247 	/*
1248 	 * let uvmfault_anonget do the dirty work.
1249 	 * if it fails (!OK) it will unlock everything for us.
1250 	 * if it succeeds, locks are still valid and locked.
1251 	 * also, if it is OK, then the anon's page is on the queues.
1252 	 * if the page is on loan from a uvm_object, then anonget will
1253 	 * lock that object for us if it does not fail.
1254 	 */
1255 
1256 	error = uvmfault_anonget(ufi, amap, anon);
1257 	switch (error) {
1258 	case 0:
1259 		break;
1260 
1261 	case ERESTART:
1262 		return ERESTART;
1263 
1264 	case EAGAIN:
1265 		kpause("fltagain1", false, hz/2, NULL);
1266 		return ERESTART;
1267 
1268 	default:
1269 		return error;
1270 	}
1271 
1272 	/*
1273 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1274 	 */
1275 
1276 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1277 
1278 	/* locked: maps(read), amap, anon, uobj(if one) */
1279 	KASSERT(mutex_owned(&amap->am_l));
1280 	KASSERT(mutex_owned(&anon->an_lock));
1281 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1282 
1283 	/*
1284 	 * special handling for loaned pages
1285 	 */
1286 
1287 	if (anon->an_page->loan_count) {
1288 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1289 		if (error != 0)
1290 			return error;
1291 	}
1292 
1293 	/*
1294 	 * if we are case 1B then we will need to allocate a new blank
1295 	 * anon to transfer the data into.   note that we have a lock
1296 	 * on anon, so no one can busy or release the page until we are done.
1297 	 * also note that the ref count can't drop to zero here because
1298 	 * it is > 1 and we are only dropping one ref.
1299 	 *
1300 	 * in the (hopefully very rare) case that we are out of RAM we
1301 	 * will unlock, wait for more RAM, and refault.
1302 	 *
1303 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1304 	 */
1305 
1306 	if (flt->cow_now && anon->an_ref > 1) {
1307 		flt->promote = true;
1308 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1309 	} else {
1310 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1311 	}
1312 	return error;
1313 }
1314 
1315 /*
1316  * uvm_fault_upper_loan: handle loaned upper page.
1317  *
1318  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1319  *	2. if cow'ing now, and if ref count is 1, break loan.
1320  */
1321 
1322 static int
1323 uvm_fault_upper_loan(
1324 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1325 	struct vm_anon *anon, struct uvm_object **ruobj)
1326 {
1327 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1328 	int error = 0;
1329 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1330 
1331 	if (!flt->cow_now) {
1332 
1333 		/*
1334 		 * for read faults on loaned pages we just cap the
1335 		 * protection at read-only.
1336 		 */
1337 
1338 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1339 
1340 	} else {
1341 		/*
1342 		 * note that we can't allow writes into a loaned page!
1343 		 *
1344 		 * if we have a write fault on a loaned page in an
1345 		 * anon then we need to look at the anon's ref count.
1346 		 * if it is greater than one then we are going to do
1347 		 * a normal copy-on-write fault into a new anon (this
1348 		 * is not a problem).  however, if the reference count
1349 		 * is one (a case where we would normally allow a
1350 		 * write directly to the page) then we need to kill
1351 		 * the loan before we continue.
1352 		 */
1353 
1354 		/* >1 case is already ok */
1355 		if (anon->an_ref == 1) {
1356 			error = uvm_loanbreak_anon(anon, *ruobj);
1357 			if (error != 0) {
1358 				uvmfault_unlockall(ufi, amap, *ruobj, anon);
1359 				uvm_wait("flt_noram2");
1360 				return ERESTART;
1361 			}
1362 			/* if we were a loan reciever uobj is gone */
1363 			if (*ruobj)
1364 				*ruobj = NULL;
1365 		}
1366 	}
1367 	return error;
1368 }
1369 
1370 /*
1371  * uvm_fault_upper_promote: promote upper page.
1372  *
1373  *	1. call uvmfault_promote.
1374  *	2. enqueue page.
1375  *	3. deref.
1376  *	4. pass page to uvm_fault_upper_enter.
1377  */
1378 
1379 static int
1380 uvm_fault_upper_promote(
1381 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1382 	struct uvm_object *uobj, struct vm_anon *anon)
1383 {
1384 	struct vm_anon * const oanon = anon;
1385 	struct vm_page *pg;
1386 	int error;
1387 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1388 
1389 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1390 	uvmexp.flt_acow++;
1391 
1392 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1393 	    &flt->anon_spare);
1394 	switch (error) {
1395 	case 0:
1396 		break;
1397 	case ERESTART:
1398 		return ERESTART;
1399 	default:
1400 		return error;
1401 	}
1402 
1403 	pg = anon->an_page;
1404 	mutex_enter(&uvm_pageqlock);
1405 	uvm_pageactivate(pg);
1406 	mutex_exit(&uvm_pageqlock);
1407 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1408 	UVM_PAGE_OWN(pg, NULL);
1409 
1410 	/* deref: can not drop to zero here by defn! */
1411 	oanon->an_ref--;
1412 
1413 	/*
1414 	 * note: oanon is still locked, as is the new anon.  we
1415 	 * need to check for this later when we unlock oanon; if
1416 	 * oanon != anon, we'll have to unlock anon, too.
1417 	 */
1418 
1419 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1420 }
1421 
1422 /*
1423  * uvm_fault_upper_direct: handle direct fault.
1424  */
1425 
1426 static int
1427 uvm_fault_upper_direct(
1428 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1429 	struct uvm_object *uobj, struct vm_anon *anon)
1430 {
1431 	struct vm_anon * const oanon = anon;
1432 	struct vm_page *pg;
1433 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1434 
1435 	uvmexp.flt_anon++;
1436 	pg = anon->an_page;
1437 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1438 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1439 
1440 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1441 }
1442 
1443 /*
1444  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1445  */
1446 
1447 static int
1448 uvm_fault_upper_enter(
1449 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1450 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1451 	struct vm_anon *oanon)
1452 {
1453 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1454 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1455 
1456 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1457 	KASSERT(mutex_owned(&amap->am_l));
1458 	KASSERT(mutex_owned(&anon->an_lock));
1459 	KASSERT(mutex_owned(&oanon->an_lock));
1460 
1461 	/*
1462 	 * now map the page in.
1463 	 */
1464 
1465 	UVMHIST_LOG(maphist,
1466 	    "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1467 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
1468 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1469 	    VM_PAGE_TO_PHYS(pg),
1470 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1471 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1472 
1473 		/*
1474 		 * No need to undo what we did; we can simply think of
1475 		 * this as the pmap throwing away the mapping information.
1476 		 *
1477 		 * We do, however, have to go through the ReFault path,
1478 		 * as the map may change while we're asleep.
1479 		 */
1480 
1481 		if (anon != oanon)
1482 			mutex_exit(&anon->an_lock);
1483 		uvmfault_unlockall(ufi, amap, uobj, oanon);
1484 		if (!uvm_reclaimable()) {
1485 			UVMHIST_LOG(maphist,
1486 			    "<- failed.  out of VM",0,0,0,0);
1487 			/* XXX instrumentation */
1488 			return ENOMEM;
1489 		}
1490 		/* XXX instrumentation */
1491 		uvm_wait("flt_pmfail1");
1492 		return ERESTART;
1493 	}
1494 
1495 	uvm_fault_upper_done(ufi, flt, anon, pg);
1496 
1497 	/*
1498 	 * done case 1!  finish up by unlocking everything and returning success
1499 	 */
1500 
1501 	if (anon != oanon) {
1502 		mutex_exit(&anon->an_lock);
1503 	}
1504 	pmap_update(ufi->orig_map->pmap);
1505 	uvmfault_unlockall(ufi, amap, uobj, oanon);
1506 	return 0;
1507 }
1508 
1509 /*
1510  * uvm_fault_upper_done: queue upper center page.
1511  */
1512 
1513 static void
1514 uvm_fault_upper_done(
1515 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1516 	struct vm_anon *anon, struct vm_page *pg)
1517 {
1518 	const bool wire_paging = flt->wire_paging;
1519 
1520 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1521 
1522 	/*
1523 	 * ... update the page queues.
1524 	 */
1525 
1526 	mutex_enter(&uvm_pageqlock);
1527 	if (wire_paging) {
1528 		uvm_pagewire(pg);
1529 
1530 		/*
1531 		 * since the now-wired page cannot be paged out,
1532 		 * release its swap resources for others to use.
1533 		 * since an anon with no swap cannot be PG_CLEAN,
1534 		 * clear its clean flag now.
1535 		 */
1536 
1537 		pg->flags &= ~(PG_CLEAN);
1538 
1539 	} else {
1540 		uvm_pageactivate(pg);
1541 	}
1542 	mutex_exit(&uvm_pageqlock);
1543 
1544 	if (wire_paging) {
1545 		uvm_anon_dropswap(anon);
1546 	}
1547 }
1548 
1549 /*
1550  * uvm_fault_lower: handle lower fault.
1551  *
1552  *	1. check uobj
1553  *	1.1. if null, ZFOD.
1554  *	1.2. if not null, look up unnmapped neighbor pages.
1555  *	2. for center page, check if promote.
1556  *	2.1. ZFOD always needs promotion.
1557  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1558  *	3. if uobj is not ZFOD and page is not found, do i/o.
1559  *	4. dispatch either direct / promote fault.
1560  */
1561 
1562 static int
1563 uvm_fault_lower(
1564 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1565 	struct vm_page **pages)
1566 {
1567 #ifdef DIAGNOSTIC
1568 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1569 #endif
1570 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1571 	struct vm_page *uobjpage;
1572 	int error;
1573 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1574 
1575 	/* locked: maps(read), amap(if there), uobj(if !null) */
1576 
1577 	/*
1578 	 * now, if the desired page is not shadowed by the amap and we have
1579 	 * a backing object that does not have a special fault routine, then
1580 	 * we ask (with pgo_get) the object for resident pages that we care
1581 	 * about and attempt to map them in.  we do not let pgo_get block
1582 	 * (PGO_LOCKED).
1583 	 */
1584 
1585 	if (uobj == NULL) {
1586 		/* zero fill; don't care neighbor pages */
1587 		uobjpage = NULL;
1588 	} else {
1589 		uvm_fault_lower_lookup(ufi, flt, pages);
1590 		uobjpage = pages[flt->centeridx];
1591 	}
1592 
1593 	/*
1594 	 * note that at this point we are done with any front or back pages.
1595 	 * we are now going to focus on the center page (i.e. the one we've
1596 	 * faulted on).  if we have faulted on the upper (anon) layer
1597 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1598 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1599 	 * layer [i.e. case 2] and the page was both present and available,
1600 	 * then we've got a pointer to it as "uobjpage" and we've already
1601 	 * made it BUSY.
1602 	 */
1603 
1604 	/*
1605 	 * locked:
1606 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1607 	 */
1608 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1609 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1610 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1611 
1612 	/*
1613 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1614 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1615 	 * have a backing object, check and see if we are going to promote
1616 	 * the data up to an anon during the fault.
1617 	 */
1618 
1619 	if (uobj == NULL) {
1620 		uobjpage = PGO_DONTCARE;
1621 		flt->promote = true;		/* always need anon here */
1622 	} else {
1623 		KASSERT(uobjpage != PGO_DONTCARE);
1624 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1625 	}
1626 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1627 	    flt->promote, (uobj == NULL), 0,0);
1628 
1629 	/*
1630 	 * if uobjpage is not null then we do not need to do I/O to get the
1631 	 * uobjpage.
1632 	 *
1633 	 * if uobjpage is null, then we need to unlock and ask the pager to
1634 	 * get the data for us.   once we have the data, we need to reverify
1635 	 * the state the world.   we are currently not holding any resources.
1636 	 */
1637 
1638 	if (uobjpage) {
1639 		/* update rusage counters */
1640 		curlwp->l_ru.ru_minflt++;
1641 	} else {
1642 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1643 		if (error != 0)
1644 			return error;
1645 	}
1646 
1647 	/*
1648 	 * locked:
1649 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1650 	 */
1651 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1652 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1653 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1654 
1655 	/*
1656 	 * notes:
1657 	 *  - at this point uobjpage can not be NULL
1658 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1659 	 *  for it above)
1660 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1661 	 */
1662 
1663 	KASSERT(uobjpage != NULL);
1664 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1665 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1666 	    (uobjpage->flags & PG_CLEAN) != 0);
1667 
1668 	if (!flt->promote) {
1669 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1670 	} else {
1671 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1672 	}
1673 	return error;
1674 }
1675 
1676 /*
1677  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1678  *
1679  *	1. get on-memory pages.
1680  *	2. if failed, give up (get only center page later).
1681  *	3. if succeeded, enter h/w mapping of neighbor pages.
1682  */
1683 
1684 static void
1685 uvm_fault_lower_lookup(
1686 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1687 	struct vm_page **pages)
1688 {
1689 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1690 	int lcv, gotpages;
1691 	vaddr_t currva;
1692 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1693 
1694 	mutex_enter(&uobj->vmobjlock);
1695 	/* locked: maps(read), amap(if there), uobj */
1696 	/*
1697 	 * the following call to pgo_get does _not_ change locking state
1698 	 */
1699 
1700 	uvmexp.fltlget++;
1701 	gotpages = flt->npages;
1702 	(void) uobj->pgops->pgo_get(uobj,
1703 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1704 	    pages, &gotpages, flt->centeridx,
1705 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1706 
1707 	/*
1708 	 * check for pages to map, if we got any
1709 	 */
1710 
1711 	if (gotpages == 0) {
1712 		pages[flt->centeridx] = NULL;
1713 		return;
1714 	}
1715 
1716 	currva = flt->startva;
1717 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1718 		struct vm_page *curpg;
1719 
1720 		curpg = pages[lcv];
1721 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1722 			continue;
1723 		}
1724 		KASSERT(curpg->uobject == uobj);
1725 
1726 		/*
1727 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1728 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1729 		 * to it.
1730 		 */
1731 
1732 		if (lcv == flt->centeridx) {
1733 			UVMHIST_LOG(maphist, "  got uobjpage "
1734 			    "(0x%x) with locked get",
1735 			    curpg, 0,0,0);
1736 		} else {
1737 			bool readonly = (curpg->flags & PG_RDONLY)
1738 			    || (curpg->loan_count > 0)
1739 			    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
1740 
1741 			uvm_fault_lower_neighbor(ufi, flt,
1742 			    currva, curpg, readonly);
1743 		}
1744 	}
1745 	pmap_update(ufi->orig_map->pmap);
1746 }
1747 
1748 /*
1749  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1750  */
1751 
1752 static void
1753 uvm_fault_lower_neighbor(
1754 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1755 	vaddr_t currva, struct vm_page *pg, bool readonly)
1756 {
1757 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1758 
1759 	/* locked: maps(read), amap(if there), uobj */
1760 
1761 	/*
1762 	 * calling pgo_get with PGO_LOCKED returns us pages which
1763 	 * are neither busy nor released, so we don't need to check
1764 	 * for this.  we can just directly enter the pages.
1765 	 */
1766 
1767 	mutex_enter(&uvm_pageqlock);
1768 	uvm_pageenqueue(pg);
1769 	mutex_exit(&uvm_pageqlock);
1770 	UVMHIST_LOG(maphist,
1771 	    "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
1772 	    ufi->orig_map->pmap, currva, pg, 0);
1773 	uvmexp.fltnomap++;
1774 
1775 	/*
1776 	 * Since this page isn't the page that's actually faulting,
1777 	 * ignore pmap_enter() failures; it's not critical that we
1778 	 * enter these right now.
1779 	 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've
1780 	 * held the lock the whole time we've had the handle.
1781 	 */
1782 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1783 	KASSERT((pg->flags & PG_RELEASED) == 0);
1784 	KASSERT((pg->flags & PG_WANTED) == 0);
1785 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) ||
1786 	    (pg->flags & PG_CLEAN) != 0);
1787 	pg->flags &= ~(PG_BUSY);
1788 	UVM_PAGE_OWN(pg, NULL);
1789 
1790 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1791 	    VM_PAGE_TO_PHYS(pg),
1792 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1793 	    flt->enter_prot & MASK(ufi->entry),
1794 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1795 }
1796 
1797 /*
1798  * uvm_fault_lower_io: get lower page from backing store.
1799  *
1800  *	1. unlock everything, because i/o will block.
1801  *	2. call pgo_get.
1802  *	3. if failed, recover.
1803  *	4. if succeeded, relock everything and verify things.
1804  */
1805 
1806 static int
1807 uvm_fault_lower_io(
1808 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1809 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1810 {
1811 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1812 	struct uvm_object *uobj = *ruobj;
1813 	struct vm_page *pg;
1814 	bool locked;
1815 	int gotpages;
1816 	int error;
1817 	voff_t uoff;
1818 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1819 
1820 	/* update rusage counters */
1821 	curlwp->l_ru.ru_majflt++;
1822 
1823 	/* locked: maps(read), amap(if there), uobj */
1824 	uvmfault_unlockall(ufi, amap, NULL, NULL);
1825 	/* locked: uobj */
1826 
1827 	uvmexp.fltget++;
1828 	gotpages = 1;
1829 	pg = NULL;
1830 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1831 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1832 	    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1833 	    PGO_SYNCIO);
1834 	/* locked: pg(if no error) */
1835 
1836 	/*
1837 	 * recover from I/O
1838 	 */
1839 
1840 	if (error) {
1841 		if (error == EAGAIN) {
1842 			UVMHIST_LOG(maphist,
1843 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1844 			kpause("fltagain2", false, hz/2, NULL);
1845 			return ERESTART;
1846 		}
1847 
1848 #if 0
1849 		KASSERT(error != ERESTART);
1850 #else
1851 		/* XXXUEBS don't re-fault? */
1852 		if (error == ERESTART)
1853 			error = EIO;
1854 #endif
1855 
1856 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1857 		    error, 0,0,0);
1858 		return error;
1859 	}
1860 
1861 	/* locked: pg */
1862 
1863 	KASSERT((pg->flags & PG_BUSY) != 0);
1864 
1865 	mutex_enter(&uvm_pageqlock);
1866 	uvm_pageactivate(pg);
1867 	mutex_exit(&uvm_pageqlock);
1868 
1869 	/*
1870 	 * re-verify the state of the world by first trying to relock
1871 	 * the maps.  always relock the object.
1872 	 */
1873 
1874 	locked = uvmfault_relock(ufi);
1875 	if (locked && amap)
1876 		amap_lock(amap);
1877 
1878 	/* might be changed */
1879 	uobj = pg->uobject;
1880 
1881 	mutex_enter(&uobj->vmobjlock);
1882 
1883 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1884 	/* locked(!locked): uobj, pg */
1885 
1886 	/*
1887 	 * verify that the page has not be released and re-verify
1888 	 * that amap slot is still free.   if there is a problem,
1889 	 * we unlock and clean up.
1890 	 */
1891 
1892 	if ((pg->flags & PG_RELEASED) != 0 ||
1893 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1894 	      ufi->orig_rvaddr - ufi->entry->start))) {
1895 		if (locked)
1896 			uvmfault_unlockall(ufi, amap, NULL, NULL);
1897 		locked = false;
1898 	}
1899 
1900 	/*
1901 	 * didn't get the lock?   release the page and retry.
1902 	 */
1903 
1904 	if (locked == false) {
1905 		UVMHIST_LOG(maphist,
1906 		    "  wasn't able to relock after fault: retry",
1907 		    0,0,0,0);
1908 		if (pg->flags & PG_WANTED) {
1909 			wakeup(pg);
1910 		}
1911 		if (pg->flags & PG_RELEASED) {
1912 			uvmexp.fltpgrele++;
1913 			uvm_pagefree(pg);
1914 			mutex_exit(&uobj->vmobjlock);
1915 			return ERESTART;
1916 		}
1917 		pg->flags &= ~(PG_BUSY|PG_WANTED);
1918 		UVM_PAGE_OWN(pg, NULL);
1919 		mutex_exit(&uobj->vmobjlock);
1920 		return ERESTART;
1921 	}
1922 
1923 	/*
1924 	 * we have the data in pg which is busy and
1925 	 * not released.  we are holding object lock (so the page
1926 	 * can't be released on us).
1927 	 */
1928 
1929 	/* locked: maps(read), amap(if !null), uobj, pg */
1930 
1931 	*ruobj = uobj;
1932 	*ruobjpage = pg;
1933 	return 0;
1934 }
1935 
1936 /*
1937  * uvm_fault_lower_direct: fault lower center page
1938  *
1939  *	1. adjust flt->enter_prot.
1940  *	2. if page is loaned, resolve.
1941  */
1942 
1943 int
1944 uvm_fault_lower_direct(
1945 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1946 	struct uvm_object *uobj, struct vm_page *uobjpage)
1947 {
1948 	struct vm_page *pg;
1949 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
1950 
1951 	/*
1952 	 * we are not promoting.   if the mapping is COW ensure that we
1953 	 * don't give more access than we should (e.g. when doing a read
1954 	 * fault on a COPYONWRITE mapping we want to map the COW page in
1955 	 * R/O even though the entry protection could be R/W).
1956 	 *
1957 	 * set "pg" to the page we want to map in (uobjpage, usually)
1958 	 */
1959 
1960 	uvmexp.flt_obj++;
1961 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
1962 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
1963 		flt->enter_prot &= ~VM_PROT_WRITE;
1964 	pg = uobjpage;		/* map in the actual object */
1965 
1966 	KASSERT(uobjpage != PGO_DONTCARE);
1967 
1968 	/*
1969 	 * we are faulting directly on the page.   be careful
1970 	 * about writing to loaned pages...
1971 	 */
1972 
1973 	if (uobjpage->loan_count) {
1974 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
1975 	}
1976 	KASSERT(pg == uobjpage);
1977 
1978 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg, uobjpage);
1979 }
1980 
1981 /*
1982  * uvm_fault_lower_direct_loan: resolve loaned page.
1983  *
1984  *	1. if not cow'ing, adjust flt->enter_prot.
1985  *	2. if cow'ing, break loan.
1986  */
1987 
1988 static int
1989 uvm_fault_lower_direct_loan(
1990 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1991 	struct uvm_object *uobj, struct vm_page **rpg,
1992 	struct vm_page **ruobjpage)
1993 {
1994 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1995 	struct vm_page *pg;
1996 	struct vm_page *uobjpage = *ruobjpage;
1997 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
1998 
1999 	if (!flt->cow_now) {
2000 		/* read fault: cap the protection at readonly */
2001 		/* cap! */
2002 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2003 	} else {
2004 		/* write fault: must break the loan here */
2005 
2006 		pg = uvm_loanbreak(uobjpage);
2007 		if (pg == NULL) {
2008 
2009 			/*
2010 			 * drop ownership of page, it can't be released
2011 			 */
2012 
2013 			if (uobjpage->flags & PG_WANTED)
2014 				wakeup(uobjpage);
2015 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2016 			UVM_PAGE_OWN(uobjpage, NULL);
2017 
2018 			uvmfault_unlockall(ufi, amap, uobj, NULL);
2019 			UVMHIST_LOG(maphist,
2020 			  "  out of RAM breaking loan, waiting",
2021 			  0,0,0,0);
2022 			uvmexp.fltnoram++;
2023 			uvm_wait("flt_noram4");
2024 			return ERESTART;
2025 		}
2026 		*rpg = pg;
2027 		*ruobjpage = pg;
2028 	}
2029 	return 0;
2030 }
2031 
2032 /*
2033  * uvm_fault_lower_promote: promote lower page.
2034  *
2035  *	1. call uvmfault_promote.
2036  *	2. fill in data.
2037  *	3. if not ZFOD, dispose old page.
2038  */
2039 
2040 int
2041 uvm_fault_lower_promote(
2042 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2043 	struct uvm_object *uobj, struct vm_page *uobjpage)
2044 {
2045 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2046 	struct vm_anon *anon;
2047 	struct vm_page *pg;
2048 	int error;
2049 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2050 
2051 	/*
2052 	 * if we are going to promote the data to an anon we
2053 	 * allocate a blank anon here and plug it into our amap.
2054 	 */
2055 #if DIAGNOSTIC
2056 	if (amap == NULL)
2057 		panic("uvm_fault: want to promote data, but no anon");
2058 #endif
2059 	error = uvmfault_promote(ufi, NULL, uobjpage,
2060 	    &anon, &flt->anon_spare);
2061 	switch (error) {
2062 	case 0:
2063 		break;
2064 	case ERESTART:
2065 		return ERESTART;
2066 	default:
2067 		return error;
2068 	}
2069 
2070 	pg = anon->an_page;
2071 
2072 	/*
2073 	 * fill in the data
2074 	 */
2075 
2076 	if (uobjpage != PGO_DONTCARE) {
2077 		uvmexp.flt_prcopy++;
2078 
2079 		/*
2080 		 * promote to shared amap?  make sure all sharing
2081 		 * procs see it
2082 		 */
2083 
2084 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2085 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2086 			/*
2087 			 * XXX: PAGE MIGHT BE WIRED!
2088 			 */
2089 		}
2090 
2091 		/*
2092 		 * dispose of uobjpage.  it can't be PG_RELEASED
2093 		 * since we still hold the object lock.
2094 		 * drop handle to uobj as well.
2095 		 */
2096 
2097 		if (uobjpage->flags & PG_WANTED)
2098 			/* still have the obj lock */
2099 			wakeup(uobjpage);
2100 		uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2101 		UVM_PAGE_OWN(uobjpage, NULL);
2102 		mutex_exit(&uobj->vmobjlock);
2103 		uobj = NULL;
2104 
2105 		UVMHIST_LOG(maphist,
2106 		    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
2107 		    uobjpage, anon, pg, 0);
2108 
2109 	} else {
2110 		uvmexp.flt_przero++;
2111 
2112 		/*
2113 		 * Page is zero'd and marked dirty by
2114 		 * uvmfault_promote().
2115 		 */
2116 
2117 		UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
2118 		    anon, pg, 0, 0);
2119 	}
2120 
2121 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg, uobjpage);
2122 }
2123 
2124 /*
2125  * uvm_fault_lower_enter: enter h/w mapping of lower page.
2126  */
2127 
2128 int
2129 uvm_fault_lower_enter(
2130 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2131 	struct uvm_object *uobj,
2132 	struct vm_anon *anon, struct vm_page *pg, struct vm_page *uobjpage)
2133 {
2134 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2135 	int error;
2136 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2137 
2138 	/*
2139 	 * locked:
2140 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
2141 	 *   anon(if !null), pg(if anon)
2142 	 *
2143 	 * note: pg is either the uobjpage or the new page in the new anon
2144 	 */
2145 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
2146 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
2147 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2148 	KASSERT(anon == NULL || mutex_owned(&anon->an_lock));
2149 	KASSERT((pg->flags & PG_BUSY) != 0);
2150 
2151 	/*
2152 	 * all resources are present.   we can now map it in and free our
2153 	 * resources.
2154 	 */
2155 
2156 	UVMHIST_LOG(maphist,
2157 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
2158 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
2159 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 ||
2160 		(pg->flags & PG_RDONLY) == 0);
2161 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2162 	    VM_PAGE_TO_PHYS(pg),
2163 	    (pg->flags & PG_RDONLY) != 0 ?
2164 	    flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2165 	    flt->access_type | PMAP_CANFAIL |
2166 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2167 
2168 		/*
2169 		 * No need to undo what we did; we can simply think of
2170 		 * this as the pmap throwing away the mapping information.
2171 		 *
2172 		 * We do, however, have to go through the ReFault path,
2173 		 * as the map may change while we're asleep.
2174 		 */
2175 
2176 		if (pg->flags & PG_WANTED)
2177 			wakeup(pg);
2178 
2179 		/*
2180 		 * note that pg can't be PG_RELEASED since we did not drop
2181 		 * the object lock since the last time we checked.
2182 		 */
2183 		KASSERT((pg->flags & PG_RELEASED) == 0);
2184 
2185 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2186 		UVM_PAGE_OWN(pg, NULL);
2187 
2188 		uvmfault_unlockall(ufi, amap, uobj, anon);
2189 		if (!uvm_reclaimable()) {
2190 			UVMHIST_LOG(maphist,
2191 			    "<- failed.  out of VM",0,0,0,0);
2192 			/* XXX instrumentation */
2193 			error = ENOMEM;
2194 			return error;
2195 		}
2196 		/* XXX instrumentation */
2197 		uvm_wait("flt_pmfail2");
2198 		return ERESTART;
2199 	}
2200 
2201 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2202 
2203 	/*
2204 	 * note that pg can't be PG_RELEASED since we did not drop the object
2205 	 * lock since the last time we checked.
2206 	 */
2207 	KASSERT((pg->flags & PG_RELEASED) == 0);
2208 	if (pg->flags & PG_WANTED)
2209 		wakeup(pg);
2210 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2211 	UVM_PAGE_OWN(pg, NULL);
2212 
2213 	pmap_update(ufi->orig_map->pmap);
2214 	uvmfault_unlockall(ufi, amap, uobj, anon);
2215 
2216 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2217 	return 0;
2218 }
2219 
2220 /*
2221  * uvm_fault_lower_done: queue lower center page.
2222  */
2223 
2224 void
2225 uvm_fault_lower_done(
2226 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2227 	struct uvm_object *uobj, struct vm_page *pg)
2228 {
2229 	bool dropswap = false;
2230 
2231 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2232 
2233 	mutex_enter(&uvm_pageqlock);
2234 	if (flt->wire_paging) {
2235 		uvm_pagewire(pg);
2236 		if (pg->pqflags & PQ_AOBJ) {
2237 
2238 			/*
2239 			 * since the now-wired page cannot be paged out,
2240 			 * release its swap resources for others to use.
2241 			 * since an aobj page with no swap cannot be PG_CLEAN,
2242 			 * clear its clean flag now.
2243 			 */
2244 
2245 			KASSERT(uobj != NULL);
2246 			pg->flags &= ~(PG_CLEAN);
2247 			dropswap = true;
2248 		}
2249 	} else {
2250 		uvm_pageactivate(pg);
2251 	}
2252 	mutex_exit(&uvm_pageqlock);
2253 
2254 	if (dropswap) {
2255 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2256 	}
2257 }
2258 
2259 
2260 /*
2261  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2262  *
2263  * => map may be read-locked by caller, but MUST NOT be write-locked.
2264  * => if map is read-locked, any operations which may cause map to
2265  *	be write-locked in uvm_fault() must be taken care of by
2266  *	the caller.  See uvm_map_pageable().
2267  */
2268 
2269 int
2270 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2271     vm_prot_t access_type, int maxprot)
2272 {
2273 	vaddr_t va;
2274 	int error;
2275 
2276 	/*
2277 	 * now fault it in a page at a time.   if the fault fails then we have
2278 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2279 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2280 	 */
2281 
2282 	/*
2283 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2284 	 * wiring the last page in the address space, though.
2285 	 */
2286 	if (start > end) {
2287 		return EFAULT;
2288 	}
2289 
2290 	for (va = start; va < end; va += PAGE_SIZE) {
2291 		error = uvm_fault_internal(map, va, access_type,
2292 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2293 		if (error) {
2294 			if (va != start) {
2295 				uvm_fault_unwire(map, start, va);
2296 			}
2297 			return error;
2298 		}
2299 	}
2300 	return 0;
2301 }
2302 
2303 /*
2304  * uvm_fault_unwire(): unwire range of virtual space.
2305  */
2306 
2307 void
2308 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2309 {
2310 	vm_map_lock_read(map);
2311 	uvm_fault_unwire_locked(map, start, end);
2312 	vm_map_unlock_read(map);
2313 }
2314 
2315 /*
2316  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2317  *
2318  * => map must be at least read-locked.
2319  */
2320 
2321 void
2322 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2323 {
2324 	struct vm_map_entry *entry;
2325 	pmap_t pmap = vm_map_pmap(map);
2326 	vaddr_t va;
2327 	paddr_t pa;
2328 	struct vm_page *pg;
2329 
2330 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
2331 
2332 	/*
2333 	 * we assume that the area we are unwiring has actually been wired
2334 	 * in the first place.   this means that we should be able to extract
2335 	 * the PAs from the pmap.   we also lock out the page daemon so that
2336 	 * we can call uvm_pageunwire.
2337 	 */
2338 
2339 	mutex_enter(&uvm_pageqlock);
2340 
2341 	/*
2342 	 * find the beginning map entry for the region.
2343 	 */
2344 
2345 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2346 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2347 		panic("uvm_fault_unwire_locked: address not in map");
2348 
2349 	for (va = start; va < end; va += PAGE_SIZE) {
2350 		if (pmap_extract(pmap, va, &pa) == false)
2351 			continue;
2352 
2353 		/*
2354 		 * find the map entry for the current address.
2355 		 */
2356 
2357 		KASSERT(va >= entry->start);
2358 		while (va >= entry->end) {
2359 			KASSERT(entry->next != &map->header &&
2360 				entry->next->start <= entry->end);
2361 			entry = entry->next;
2362 		}
2363 
2364 		/*
2365 		 * if the entry is no longer wired, tell the pmap.
2366 		 */
2367 
2368 		if (VM_MAPENT_ISWIRED(entry) == 0)
2369 			pmap_unwire(pmap, va);
2370 
2371 		pg = PHYS_TO_VM_PAGE(pa);
2372 		if (pg)
2373 			uvm_pageunwire(pg);
2374 	}
2375 
2376 	mutex_exit(&uvm_pageqlock);
2377 }
2378