xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 7a6a7ae08ac6c612f0fbb0d4425825c6be2a9050)
1 /*	$NetBSD: uvm_fault.c,v 1.183 2011/04/08 10:42:51 yamt 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 1.183 2011/04/08 10:42:51 yamt 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  * => it's caller's responsibility to put the promoted nanon->an_page to the
540  *    page queue.
541  */
542 
543 static int
544 uvmfault_promote(struct uvm_faultinfo *ufi,
545     struct vm_anon *oanon,
546     struct vm_page *uobjpage,
547     struct vm_anon **nanon, /* OUT: allocated anon */
548     struct vm_anon **spare)
549 {
550 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
551 	struct uvm_object *uobj;
552 	struct vm_anon *anon;
553 	struct vm_page *pg;
554 	struct vm_page *opg;
555 	int error;
556 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
557 
558 	if (oanon) {
559 		/* anon COW */
560 		opg = oanon->an_page;
561 		KASSERT(opg != NULL);
562 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
563 	} else if (uobjpage != PGO_DONTCARE) {
564 		/* object-backed COW */
565 		opg = uobjpage;
566 	} else {
567 		/* ZFOD */
568 		opg = NULL;
569 	}
570 	if (opg != NULL) {
571 		uobj = opg->uobject;
572 	} else {
573 		uobj = NULL;
574 	}
575 
576 	KASSERT(amap != NULL);
577 	KASSERT(uobjpage != NULL);
578 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
579 	KASSERT(mutex_owned(&amap->am_l));
580 	KASSERT(oanon == NULL || mutex_owned(&oanon->an_lock));
581 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
582 #if 0
583 	KASSERT(*spare == NULL || !mutex_owned(&(*spare)->an_lock));
584 #endif
585 
586 	if (*spare != NULL) {
587 		anon = *spare;
588 		*spare = NULL;
589 		mutex_enter(&anon->an_lock);
590 	} else if (ufi->map != kernel_map) {
591 		anon = uvm_analloc();
592 	} else {
593 		UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0);
594 
595 		/*
596 		 * we can't allocate anons with kernel_map locked.
597 		 */
598 
599 		uvm_page_unbusy(&uobjpage, 1);
600 		uvmfault_unlockall(ufi, amap, uobj, oanon);
601 
602 		*spare = uvm_analloc();
603 		if (*spare == NULL) {
604 			goto nomem;
605 		}
606 		mutex_exit(&(*spare)->an_lock);
607 		error = ERESTART;
608 		goto done;
609 	}
610 	if (anon) {
611 
612 		/*
613 		 * The new anon is locked.
614 		 *
615 		 * if opg == NULL, we want a zero'd, dirty page,
616 		 * so have uvm_pagealloc() do that for us.
617 		 */
618 
619 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
620 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
621 	} else {
622 		pg = NULL;
623 	}
624 
625 	/*
626 	 * out of memory resources?
627 	 */
628 
629 	if (pg == NULL) {
630 		/* save anon for the next try. */
631 		if (anon != NULL) {
632 			mutex_exit(&anon->an_lock);
633 			*spare = anon;
634 		}
635 
636 		/* unlock and fail ... */
637 		uvm_page_unbusy(&uobjpage, 1);
638 		uvmfault_unlockall(ufi, amap, uobj, oanon);
639 nomem:
640 		if (!uvm_reclaimable()) {
641 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
642 			uvmexp.fltnoanon++;
643 			error = ENOMEM;
644 			goto done;
645 		}
646 
647 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
648 		uvmexp.fltnoram++;
649 		uvm_wait("flt_noram5");
650 		error = ERESTART;
651 		goto done;
652 	}
653 
654 	/* copy page [pg now dirty] */
655 	if (opg) {
656 		uvm_pagecopy(opg, pg);
657 	}
658 
659 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
660 	    oanon != NULL);
661 
662 	*nanon = anon;
663 	error = 0;
664 done:
665 	return error;
666 }
667 
668 
669 /*
670  *   F A U L T   -   m a i n   e n t r y   p o i n t
671  */
672 
673 /*
674  * uvm_fault: page fault handler
675  *
676  * => called from MD code to resolve a page fault
677  * => VM data structures usually should be unlocked.   however, it is
678  *	possible to call here with the main map locked if the caller
679  *	gets a write lock, sets it recusive, and then calls us (c.f.
680  *	uvm_map_pageable).   this should be avoided because it keeps
681  *	the map locked off during I/O.
682  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
683  */
684 
685 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
686 			 ~VM_PROT_WRITE : VM_PROT_ALL)
687 
688 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
689 #define UVM_FAULT_WIRE		(1 << 0)
690 #define UVM_FAULT_MAXPROT	(1 << 1)
691 
692 struct uvm_faultctx {
693 	vm_prot_t access_type;
694 	vm_prot_t enter_prot;
695 	vaddr_t startva;
696 	int npages;
697 	int centeridx;
698 	struct vm_anon *anon_spare;
699 	bool wire_mapping;
700 	bool narrow;
701 	bool wire_paging;
702 	bool cow_now;
703 	bool promote;
704 };
705 
706 static inline int	uvm_fault_check(
707 			    struct uvm_faultinfo *, struct uvm_faultctx *,
708 			    struct vm_anon ***, bool);
709 
710 static int		uvm_fault_upper(
711 			    struct uvm_faultinfo *, struct uvm_faultctx *,
712 			    struct vm_anon **);
713 static inline int	uvm_fault_upper_lookup(
714 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
715 			    struct vm_anon **, struct vm_page **);
716 static inline void	uvm_fault_upper_neighbor(
717 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
718 			    vaddr_t, struct vm_page *, bool);
719 static inline int	uvm_fault_upper_loan(
720 			    struct uvm_faultinfo *, struct uvm_faultctx *,
721 			    struct vm_anon *, struct uvm_object **);
722 static inline int	uvm_fault_upper_promote(
723 			    struct uvm_faultinfo *, struct uvm_faultctx *,
724 			    struct uvm_object *, struct vm_anon *);
725 static inline int	uvm_fault_upper_direct(
726 			    struct uvm_faultinfo *, struct uvm_faultctx *,
727 			    struct uvm_object *, struct vm_anon *);
728 static int		uvm_fault_upper_enter(
729 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
730 			    struct uvm_object *, struct vm_anon *,
731 			    struct vm_page *, struct vm_anon *);
732 static inline void	uvm_fault_upper_done(
733 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
734 			    struct vm_anon *, struct vm_page *);
735 
736 static int		uvm_fault_lower(
737 			    struct uvm_faultinfo *, struct uvm_faultctx *,
738 			    struct vm_page **);
739 static inline void	uvm_fault_lower_lookup(
740 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
741 			    struct vm_page **);
742 static inline void	uvm_fault_lower_neighbor(
743 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
744 			    vaddr_t, struct vm_page *, bool);
745 static inline int	uvm_fault_lower_io(
746 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
747 			    struct uvm_object **, struct vm_page **);
748 static inline int	uvm_fault_lower_direct(
749 			    struct uvm_faultinfo *, struct uvm_faultctx *,
750 			    struct uvm_object *, struct vm_page *);
751 static inline int	uvm_fault_lower_direct_loan(
752 			    struct uvm_faultinfo *, struct uvm_faultctx *,
753 			    struct uvm_object *, struct vm_page **,
754 			    struct vm_page **);
755 static inline int	uvm_fault_lower_promote(
756 			    struct uvm_faultinfo *, struct uvm_faultctx *,
757 			    struct uvm_object *, struct vm_page *);
758 static int		uvm_fault_lower_enter(
759 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
760 			    struct uvm_object *,
761 			    struct vm_anon *, struct vm_page *);
762 static inline void	uvm_fault_lower_done(
763 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
764 			    struct uvm_object *, struct vm_page *);
765 
766 int
767 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
768     vm_prot_t access_type, int fault_flag)
769 {
770 	struct uvm_faultinfo ufi;
771 	struct uvm_faultctx flt = {
772 		.access_type = access_type,
773 
774 		/* don't look for neighborhood * pages on "wire" fault */
775 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
776 
777 		/* "wire" fault causes wiring of both mapping and paging */
778 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
779 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
780 	};
781 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
782 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
783 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
784 	int error;
785 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
786 
787 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)",
788 	      orig_map, vaddr, access_type, fault_flag);
789 
790 	curcpu()->ci_data.cpu_nfault++;
791 
792 	/*
793 	 * init the IN parameters in the ufi
794 	 */
795 
796 	ufi.orig_map = orig_map;
797 	ufi.orig_rvaddr = trunc_page(vaddr);
798 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
799 
800 	error = ERESTART;
801 	while (error == ERESTART) { /* ReFault: */
802 		anons = anons_store;
803 		pages = pages_store;
804 
805 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
806 		if (error != 0)
807 			continue;
808 
809 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
810 		if (error != 0)
811 			continue;
812 
813 		if (pages[flt.centeridx] == PGO_DONTCARE)
814 			error = uvm_fault_upper(&ufi, &flt, anons);
815 		else {
816 			struct uvm_object * const uobj =
817 			    ufi.entry->object.uvm_obj;
818 
819 			if (uobj && uobj->pgops->pgo_fault != NULL) {
820 				/*
821 				 * invoke "special" fault routine.
822 				 */
823 				mutex_enter(&uobj->vmobjlock);
824 				/* locked: maps(read), amap(if there), uobj */
825 				error = uobj->pgops->pgo_fault(&ufi,
826 				    flt.startva, pages, flt.npages,
827 				    flt.centeridx, flt.access_type,
828 				    PGO_LOCKED|PGO_SYNCIO);
829 
830 				/*
831 				 * locked: nothing, pgo_fault has unlocked
832 				 * everything
833 				 */
834 
835 				/*
836 				 * object fault routine responsible for
837 				 * pmap_update().
838 				 */
839 			} else {
840 				error = uvm_fault_lower(&ufi, &flt, pages);
841 			}
842 		}
843 	}
844 
845 	if (flt.anon_spare != NULL) {
846 		flt.anon_spare->an_ref--;
847 		uvm_anfree(flt.anon_spare);
848 	}
849 	return error;
850 }
851 
852 /*
853  * uvm_fault_check: check prot, handle needs-copy, etc.
854  *
855  *	1. lookup entry.
856  *	2. check protection.
857  *	3. adjust fault condition (mainly for simulated fault).
858  *	4. handle needs-copy (lazy amap copy).
859  *	5. establish range of interest for neighbor fault (aka pre-fault).
860  *	6. look up anons (if amap exists).
861  *	7. flush pages (if MADV_SEQUENTIAL)
862  *
863  * => called with nothing locked.
864  * => if we fail (result != 0) we unlock everything.
865  * => initialize/adjust many members of flt.
866  */
867 
868 static int
869 uvm_fault_check(
870 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
871 	struct vm_anon ***ranons, bool maxprot)
872 {
873 	struct vm_amap *amap;
874 	struct uvm_object *uobj;
875 	vm_prot_t check_prot;
876 	int nback, nforw;
877 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
878 
879 	/*
880 	 * lookup and lock the maps
881 	 */
882 
883 	if (uvmfault_lookup(ufi, false) == false) {
884 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", ufi->orig_rvaddr,
885 		    0,0,0);
886 		return EFAULT;
887 	}
888 	/* locked: maps(read) */
889 
890 #ifdef DIAGNOSTIC
891 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
892 		printf("Page fault on non-pageable map:\n");
893 		printf("ufi->map = %p\n", ufi->map);
894 		printf("ufi->orig_map = %p\n", ufi->orig_map);
895 		printf("ufi->orig_rvaddr = 0x%lx\n", (u_long) ufi->orig_rvaddr);
896 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
897 	}
898 #endif
899 
900 	/*
901 	 * check protection
902 	 */
903 
904 	check_prot = maxprot ?
905 	    ufi->entry->max_protection : ufi->entry->protection;
906 	if ((check_prot & flt->access_type) != flt->access_type) {
907 		UVMHIST_LOG(maphist,
908 		    "<- protection failure (prot=0x%x, access=0x%x)",
909 		    ufi->entry->protection, flt->access_type, 0, 0);
910 		uvmfault_unlockmaps(ufi, false);
911 		return EACCES;
912 	}
913 
914 	/*
915 	 * "enter_prot" is the protection we want to enter the page in at.
916 	 * for certain pages (e.g. copy-on-write pages) this protection can
917 	 * be more strict than ufi->entry->protection.  "wired" means either
918 	 * the entry is wired or we are fault-wiring the pg.
919 	 */
920 
921 	flt->enter_prot = ufi->entry->protection;
922 	if (VM_MAPENT_ISWIRED(ufi->entry))
923 		flt->wire_mapping = true;
924 
925 	if (flt->wire_mapping) {
926 		flt->access_type = flt->enter_prot; /* full access for wired */
927 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
928 	} else {
929 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
930 	}
931 
932 	flt->promote = false;
933 
934 	/*
935 	 * handle "needs_copy" case.   if we need to copy the amap we will
936 	 * have to drop our readlock and relock it with a write lock.  (we
937 	 * need a write lock to change anything in a map entry [e.g.
938 	 * needs_copy]).
939 	 */
940 
941 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
942 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
943 			KASSERT(!maxprot);
944 			/* need to clear */
945 			UVMHIST_LOG(maphist,
946 			    "  need to clear needs_copy and refault",0,0,0,0);
947 			uvmfault_unlockmaps(ufi, false);
948 			uvmfault_amapcopy(ufi);
949 			uvmexp.fltamcopy++;
950 			return ERESTART;
951 
952 		} else {
953 
954 			/*
955 			 * ensure that we pmap_enter page R/O since
956 			 * needs_copy is still true
957 			 */
958 
959 			flt->enter_prot &= ~VM_PROT_WRITE;
960 		}
961 	}
962 
963 	/*
964 	 * identify the players
965 	 */
966 
967 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
968 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
969 
970 	/*
971 	 * check for a case 0 fault.  if nothing backing the entry then
972 	 * error now.
973 	 */
974 
975 	if (amap == NULL && uobj == NULL) {
976 		uvmfault_unlockmaps(ufi, false);
977 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
978 		return EFAULT;
979 	}
980 
981 	/*
982 	 * establish range of interest based on advice from mapper
983 	 * and then clip to fit map entry.   note that we only want
984 	 * to do this the first time through the fault.   if we
985 	 * ReFault we will disable this by setting "narrow" to true.
986 	 */
987 
988 	if (flt->narrow == false) {
989 
990 		/* wide fault (!narrow) */
991 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
992 			 ufi->entry->advice);
993 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
994 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
995 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
996 		/*
997 		 * note: "-1" because we don't want to count the
998 		 * faulting page as forw
999 		 */
1000 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1001 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1002 			     PAGE_SHIFT) - 1);
1003 		flt->npages = nback + nforw + 1;
1004 		flt->centeridx = nback;
1005 
1006 		flt->narrow = true;	/* ensure only once per-fault */
1007 
1008 	} else {
1009 
1010 		/* narrow fault! */
1011 		nback = nforw = 0;
1012 		flt->startva = ufi->orig_rvaddr;
1013 		flt->npages = 1;
1014 		flt->centeridx = 0;
1015 
1016 	}
1017 	/* offset from entry's start to pgs' start */
1018 	const voff_t eoff = flt->startva - ufi->entry->start;
1019 
1020 	/* locked: maps(read) */
1021 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
1022 		    flt->narrow, nback, nforw, flt->startva);
1023 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi->entry,
1024 		    amap, uobj, 0);
1025 
1026 	/*
1027 	 * if we've got an amap, lock it and extract current anons.
1028 	 */
1029 
1030 	if (amap) {
1031 		amap_lock(amap);
1032 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1033 	} else {
1034 		*ranons = NULL;	/* to be safe */
1035 	}
1036 
1037 	/* locked: maps(read), amap(if there) */
1038 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1039 
1040 	/*
1041 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1042 	 * now and then forget about them (for the rest of the fault).
1043 	 */
1044 
1045 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1046 
1047 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1048 		    0,0,0,0);
1049 		/* flush back-page anons? */
1050 		if (amap)
1051 			uvmfault_anonflush(*ranons, nback);
1052 
1053 		/* flush object? */
1054 		if (uobj) {
1055 			voff_t uoff;
1056 
1057 			uoff = ufi->entry->offset + eoff;
1058 			mutex_enter(&uobj->vmobjlock);
1059 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1060 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1061 		}
1062 
1063 		/* now forget about the backpages */
1064 		if (amap)
1065 			*ranons += nback;
1066 		flt->startva += (nback << PAGE_SHIFT);
1067 		flt->npages -= nback;
1068 		flt->centeridx = 0;
1069 	}
1070 	/*
1071 	 * => startva is fixed
1072 	 * => npages is fixed
1073 	 */
1074 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1075 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1076 	    flt->startva + (flt->npages << PAGE_SHIFT));
1077 	return 0;
1078 }
1079 
1080 /*
1081  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1082  *
1083  * iterate range of interest:
1084  *	1. check if h/w mapping exists.  if yes, we don't care
1085  *	2. check if anon exists.  if not, page is lower.
1086  *	3. if anon exists, enter h/w mapping for neighbors.
1087  *
1088  * => called with amap locked (if exists).
1089  */
1090 
1091 static int
1092 uvm_fault_upper_lookup(
1093 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1094 	struct vm_anon **anons, struct vm_page **pages)
1095 {
1096 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1097 	int lcv;
1098 	vaddr_t currva;
1099 	bool shadowed;
1100 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1101 
1102 	/* locked: maps(read), amap(if there) */
1103 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1104 
1105 	/*
1106 	 * map in the backpages and frontpages we found in the amap in hopes
1107 	 * of preventing future faults.    we also init the pages[] array as
1108 	 * we go.
1109 	 */
1110 
1111 	currva = flt->startva;
1112 	shadowed = false;
1113 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1114 		/*
1115 		 * don't play with VAs that are already mapped
1116 		 * (except for center)
1117 		 */
1118 		if (lcv != flt->centeridx &&
1119 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1120 			pages[lcv] = PGO_DONTCARE;
1121 			continue;
1122 		}
1123 
1124 		/*
1125 		 * unmapped or center page.   check if any anon at this level.
1126 		 */
1127 		if (amap == NULL || anons[lcv] == NULL) {
1128 			pages[lcv] = NULL;
1129 			continue;
1130 		}
1131 
1132 		/*
1133 		 * check for present page and map if possible.   re-activate it.
1134 		 */
1135 
1136 		pages[lcv] = PGO_DONTCARE;
1137 		if (lcv == flt->centeridx) {	/* save center for later! */
1138 			shadowed = true;
1139 		} else {
1140 			struct vm_anon *anon = anons[lcv];
1141 
1142 			mutex_enter(&anon->an_lock);
1143 			struct vm_page *pg = anon->an_page;
1144 
1145 			/* ignore loaned and busy pages */
1146 			if (pg != NULL && pg->loan_count == 0 &&
1147 			    (pg->flags & PG_BUSY) == 0)
1148 				uvm_fault_upper_neighbor(ufi, flt, currva,
1149 				    pg, anon->an_ref > 1);
1150 			mutex_exit(&anon->an_lock);
1151 		}
1152 	}
1153 
1154 	/* locked: maps(read), amap(if there) */
1155 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1156 	/* (shadowed == true) if there is an anon at the faulting address */
1157 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
1158 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1159 
1160 	/*
1161 	 * note that if we are really short of RAM we could sleep in the above
1162 	 * call to pmap_enter with everything locked.   bad?
1163 	 *
1164 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1165 	 * XXX case.  --thorpej
1166 	 */
1167 
1168 	return 0;
1169 }
1170 
1171 /*
1172  * uvm_fault_upper_neighbor: enter single lower neighbor page.
1173  *
1174  * => called with amap and anon locked.
1175  */
1176 
1177 static void
1178 uvm_fault_upper_neighbor(
1179 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1180 	vaddr_t currva, struct vm_page *pg, bool readonly)
1181 {
1182 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1183 
1184 	/* locked: amap, anon */
1185 
1186 	mutex_enter(&uvm_pageqlock);
1187 	uvm_pageenqueue(pg);
1188 	mutex_exit(&uvm_pageqlock);
1189 	UVMHIST_LOG(maphist,
1190 	    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
1191 	    ufi->orig_map->pmap, currva, pg, 0);
1192 	uvmexp.fltnamap++;
1193 
1194 	/*
1195 	 * Since this page isn't the page that's actually faulting,
1196 	 * ignore pmap_enter() failures; it's not critical that we
1197 	 * enter these right now.
1198 	 */
1199 
1200 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1201 	    VM_PAGE_TO_PHYS(pg),
1202 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1203 	    flt->enter_prot,
1204 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1205 
1206 	pmap_update(ufi->orig_map->pmap);
1207 }
1208 
1209 /*
1210  * uvm_fault_upper: handle upper fault.
1211  *
1212  *	1. acquire anon lock.
1213  *	2. get anon.  let uvmfault_anonget do the dirty work.
1214  *	3. handle loan.
1215  *	4. dispatch direct or promote handlers.
1216  */
1217 
1218 static int
1219 uvm_fault_upper(
1220 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1221 	struct vm_anon **anons)
1222 {
1223 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1224 	struct vm_anon * const anon = anons[flt->centeridx];
1225 	struct uvm_object *uobj;
1226 	int error;
1227 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1228 
1229 	/* locked: maps(read), amap */
1230 	KASSERT(mutex_owned(&amap->am_l));
1231 
1232 	/*
1233 	 * handle case 1: fault on an anon in our amap
1234 	 */
1235 
1236 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1237 	mutex_enter(&anon->an_lock);
1238 
1239 	/* locked: maps(read), amap, anon */
1240 	KASSERT(mutex_owned(&amap->am_l));
1241 	KASSERT(mutex_owned(&anon->an_lock));
1242 
1243 	/*
1244 	 * no matter if we have case 1A or case 1B we are going to need to
1245 	 * have the anon's memory resident.   ensure that now.
1246 	 */
1247 
1248 	/*
1249 	 * let uvmfault_anonget do the dirty work.
1250 	 * if it fails (!OK) it will unlock everything for us.
1251 	 * if it succeeds, locks are still valid and locked.
1252 	 * also, if it is OK, then the anon's page is on the queues.
1253 	 * if the page is on loan from a uvm_object, then anonget will
1254 	 * lock that object for us if it does not fail.
1255 	 */
1256 
1257 	error = uvmfault_anonget(ufi, amap, anon);
1258 	switch (error) {
1259 	case 0:
1260 		break;
1261 
1262 	case ERESTART:
1263 		return ERESTART;
1264 
1265 	case EAGAIN:
1266 		kpause("fltagain1", false, hz/2, NULL);
1267 		return ERESTART;
1268 
1269 	default:
1270 		return error;
1271 	}
1272 
1273 	/*
1274 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1275 	 */
1276 
1277 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1278 
1279 	/* locked: maps(read), amap, anon, uobj(if one) */
1280 	KASSERT(mutex_owned(&amap->am_l));
1281 	KASSERT(mutex_owned(&anon->an_lock));
1282 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1283 
1284 	/*
1285 	 * special handling for loaned pages
1286 	 */
1287 
1288 	if (anon->an_page->loan_count) {
1289 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1290 		if (error != 0)
1291 			return error;
1292 	}
1293 
1294 	/*
1295 	 * if we are case 1B then we will need to allocate a new blank
1296 	 * anon to transfer the data into.   note that we have a lock
1297 	 * on anon, so no one can busy or release the page until we are done.
1298 	 * also note that the ref count can't drop to zero here because
1299 	 * it is > 1 and we are only dropping one ref.
1300 	 *
1301 	 * in the (hopefully very rare) case that we are out of RAM we
1302 	 * will unlock, wait for more RAM, and refault.
1303 	 *
1304 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1305 	 */
1306 
1307 	if (flt->cow_now && anon->an_ref > 1) {
1308 		flt->promote = true;
1309 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1310 	} else {
1311 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1312 	}
1313 	return error;
1314 }
1315 
1316 /*
1317  * uvm_fault_upper_loan: handle loaned upper page.
1318  *
1319  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1320  *	2. if cow'ing now, and if ref count is 1, break loan.
1321  */
1322 
1323 static int
1324 uvm_fault_upper_loan(
1325 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1326 	struct vm_anon *anon, struct uvm_object **ruobj)
1327 {
1328 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1329 	int error = 0;
1330 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1331 
1332 	if (!flt->cow_now) {
1333 
1334 		/*
1335 		 * for read faults on loaned pages we just cap the
1336 		 * protection at read-only.
1337 		 */
1338 
1339 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1340 
1341 	} else {
1342 		/*
1343 		 * note that we can't allow writes into a loaned page!
1344 		 *
1345 		 * if we have a write fault on a loaned page in an
1346 		 * anon then we need to look at the anon's ref count.
1347 		 * if it is greater than one then we are going to do
1348 		 * a normal copy-on-write fault into a new anon (this
1349 		 * is not a problem).  however, if the reference count
1350 		 * is one (a case where we would normally allow a
1351 		 * write directly to the page) then we need to kill
1352 		 * the loan before we continue.
1353 		 */
1354 
1355 		/* >1 case is already ok */
1356 		if (anon->an_ref == 1) {
1357 			error = uvm_loanbreak_anon(anon, *ruobj);
1358 			if (error != 0) {
1359 				uvmfault_unlockall(ufi, amap, *ruobj, anon);
1360 				uvm_wait("flt_noram2");
1361 				return ERESTART;
1362 			}
1363 			/* if we were a loan reciever uobj is gone */
1364 			if (*ruobj)
1365 				*ruobj = NULL;
1366 		}
1367 	}
1368 	return error;
1369 }
1370 
1371 /*
1372  * uvm_fault_upper_promote: promote upper page.
1373  *
1374  *	1. call uvmfault_promote.
1375  *	2. enqueue page.
1376  *	3. deref.
1377  *	4. pass page to uvm_fault_upper_enter.
1378  */
1379 
1380 static int
1381 uvm_fault_upper_promote(
1382 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1383 	struct uvm_object *uobj, struct vm_anon *anon)
1384 {
1385 	struct vm_anon * const oanon = anon;
1386 	struct vm_page *pg;
1387 	int error;
1388 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1389 
1390 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1391 	uvmexp.flt_acow++;
1392 
1393 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1394 	    &flt->anon_spare);
1395 	switch (error) {
1396 	case 0:
1397 		break;
1398 	case ERESTART:
1399 		return ERESTART;
1400 	default:
1401 		return error;
1402 	}
1403 
1404 	pg = anon->an_page;
1405 	mutex_enter(&uvm_pageqlock);
1406 	uvm_pageenqueue(pg); /* uvm_fault_upper_done will activate the page */
1407 	mutex_exit(&uvm_pageqlock);
1408 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1409 	UVM_PAGE_OWN(pg, NULL);
1410 
1411 	/* deref: can not drop to zero here by defn! */
1412 	KASSERT(oanon->an_ref > 1);
1413 	oanon->an_ref--;
1414 
1415 	/*
1416 	 * note: oanon is still locked, as is the new anon.  we
1417 	 * need to check for this later when we unlock oanon; if
1418 	 * oanon != anon, we'll have to unlock anon, too.
1419 	 */
1420 
1421 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1422 }
1423 
1424 /*
1425  * uvm_fault_upper_direct: handle direct fault.
1426  */
1427 
1428 static int
1429 uvm_fault_upper_direct(
1430 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1431 	struct uvm_object *uobj, struct vm_anon *anon)
1432 {
1433 	struct vm_anon * const oanon = anon;
1434 	struct vm_page *pg;
1435 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1436 
1437 	uvmexp.flt_anon++;
1438 	pg = anon->an_page;
1439 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1440 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1441 
1442 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1443 }
1444 
1445 /*
1446  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1447  */
1448 
1449 static int
1450 uvm_fault_upper_enter(
1451 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1452 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1453 	struct vm_anon *oanon)
1454 {
1455 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1456 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1457 
1458 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1459 	KASSERT(mutex_owned(&amap->am_l));
1460 	KASSERT(mutex_owned(&anon->an_lock));
1461 	KASSERT(mutex_owned(&oanon->an_lock));
1462 
1463 	/*
1464 	 * now map the page in.
1465 	 */
1466 
1467 	UVMHIST_LOG(maphist,
1468 	    "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1469 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
1470 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1471 	    VM_PAGE_TO_PHYS(pg),
1472 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1473 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1474 
1475 		/*
1476 		 * No need to undo what we did; we can simply think of
1477 		 * this as the pmap throwing away the mapping information.
1478 		 *
1479 		 * We do, however, have to go through the ReFault path,
1480 		 * as the map may change while we're asleep.
1481 		 */
1482 
1483 		if (anon != oanon)
1484 			mutex_exit(&anon->an_lock);
1485 		uvmfault_unlockall(ufi, amap, uobj, oanon);
1486 		if (!uvm_reclaimable()) {
1487 			UVMHIST_LOG(maphist,
1488 			    "<- failed.  out of VM",0,0,0,0);
1489 			/* XXX instrumentation */
1490 			return ENOMEM;
1491 		}
1492 		/* XXX instrumentation */
1493 		uvm_wait("flt_pmfail1");
1494 		return ERESTART;
1495 	}
1496 
1497 	uvm_fault_upper_done(ufi, flt, anon, pg);
1498 
1499 	/*
1500 	 * done case 1!  finish up by unlocking everything and returning success
1501 	 */
1502 
1503 	if (anon != oanon) {
1504 		mutex_exit(&anon->an_lock);
1505 	}
1506 	pmap_update(ufi->orig_map->pmap);
1507 	uvmfault_unlockall(ufi, amap, uobj, oanon);
1508 	return 0;
1509 }
1510 
1511 /*
1512  * uvm_fault_upper_done: queue upper center page.
1513  */
1514 
1515 static void
1516 uvm_fault_upper_done(
1517 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1518 	struct vm_anon *anon, struct vm_page *pg)
1519 {
1520 	const bool wire_paging = flt->wire_paging;
1521 
1522 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1523 
1524 	/*
1525 	 * ... update the page queues.
1526 	 */
1527 
1528 	mutex_enter(&uvm_pageqlock);
1529 	if (wire_paging) {
1530 		uvm_pagewire(pg);
1531 
1532 		/*
1533 		 * since the now-wired page cannot be paged out,
1534 		 * release its swap resources for others to use.
1535 		 * since an anon with no swap cannot be PG_CLEAN,
1536 		 * clear its clean flag now.
1537 		 */
1538 
1539 		pg->flags &= ~(PG_CLEAN);
1540 
1541 	} else {
1542 		uvm_pageactivate(pg);
1543 	}
1544 	mutex_exit(&uvm_pageqlock);
1545 
1546 	if (wire_paging) {
1547 		uvm_anon_dropswap(anon);
1548 	}
1549 }
1550 
1551 /*
1552  * uvm_fault_lower: handle lower fault.
1553  *
1554  *	1. check uobj
1555  *	1.1. if null, ZFOD.
1556  *	1.2. if not null, look up unnmapped neighbor pages.
1557  *	2. for center page, check if promote.
1558  *	2.1. ZFOD always needs promotion.
1559  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1560  *	3. if uobj is not ZFOD and page is not found, do i/o.
1561  *	4. dispatch either direct / promote fault.
1562  */
1563 
1564 static int
1565 uvm_fault_lower(
1566 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1567 	struct vm_page **pages)
1568 {
1569 #ifdef DIAGNOSTIC
1570 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1571 #endif
1572 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1573 	struct vm_page *uobjpage;
1574 	int error;
1575 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1576 
1577 	/* locked: maps(read), amap(if there), uobj(if !null) */
1578 
1579 	/*
1580 	 * now, if the desired page is not shadowed by the amap and we have
1581 	 * a backing object that does not have a special fault routine, then
1582 	 * we ask (with pgo_get) the object for resident pages that we care
1583 	 * about and attempt to map them in.  we do not let pgo_get block
1584 	 * (PGO_LOCKED).
1585 	 */
1586 
1587 	if (uobj == NULL) {
1588 		/* zero fill; don't care neighbor pages */
1589 		uobjpage = NULL;
1590 	} else {
1591 		uvm_fault_lower_lookup(ufi, flt, pages);
1592 		uobjpage = pages[flt->centeridx];
1593 	}
1594 
1595 	/*
1596 	 * note that at this point we are done with any front or back pages.
1597 	 * we are now going to focus on the center page (i.e. the one we've
1598 	 * faulted on).  if we have faulted on the upper (anon) layer
1599 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1600 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1601 	 * layer [i.e. case 2] and the page was both present and available,
1602 	 * then we've got a pointer to it as "uobjpage" and we've already
1603 	 * made it BUSY.
1604 	 */
1605 
1606 	/*
1607 	 * locked:
1608 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1609 	 */
1610 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1611 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1612 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1613 
1614 	/*
1615 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1616 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1617 	 * have a backing object, check and see if we are going to promote
1618 	 * the data up to an anon during the fault.
1619 	 */
1620 
1621 	if (uobj == NULL) {
1622 		uobjpage = PGO_DONTCARE;
1623 		flt->promote = true;		/* always need anon here */
1624 	} else {
1625 		KASSERT(uobjpage != PGO_DONTCARE);
1626 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1627 	}
1628 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1629 	    flt->promote, (uobj == NULL), 0,0);
1630 
1631 	/*
1632 	 * if uobjpage is not null then we do not need to do I/O to get the
1633 	 * uobjpage.
1634 	 *
1635 	 * if uobjpage is null, then we need to unlock and ask the pager to
1636 	 * get the data for us.   once we have the data, we need to reverify
1637 	 * the state the world.   we are currently not holding any resources.
1638 	 */
1639 
1640 	if (uobjpage) {
1641 		/* update rusage counters */
1642 		curlwp->l_ru.ru_minflt++;
1643 	} else {
1644 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1645 		if (error != 0)
1646 			return error;
1647 	}
1648 
1649 	/*
1650 	 * locked:
1651 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1652 	 */
1653 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1654 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1655 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1656 
1657 	/*
1658 	 * notes:
1659 	 *  - at this point uobjpage can not be NULL
1660 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1661 	 *  for it above)
1662 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1663 	 */
1664 
1665 	KASSERT(uobjpage != NULL);
1666 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1667 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1668 	    (uobjpage->flags & PG_CLEAN) != 0);
1669 
1670 	if (!flt->promote) {
1671 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1672 	} else {
1673 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1674 	}
1675 	return error;
1676 }
1677 
1678 /*
1679  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1680  *
1681  *	1. get on-memory pages.
1682  *	2. if failed, give up (get only center page later).
1683  *	3. if succeeded, enter h/w mapping of neighbor pages.
1684  */
1685 
1686 static void
1687 uvm_fault_lower_lookup(
1688 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1689 	struct vm_page **pages)
1690 {
1691 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1692 	int lcv, gotpages;
1693 	vaddr_t currva;
1694 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1695 
1696 	mutex_enter(&uobj->vmobjlock);
1697 	/* locked: maps(read), amap(if there), uobj */
1698 	/*
1699 	 * the following call to pgo_get does _not_ change locking state
1700 	 */
1701 
1702 	uvmexp.fltlget++;
1703 	gotpages = flt->npages;
1704 	(void) uobj->pgops->pgo_get(uobj,
1705 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1706 	    pages, &gotpages, flt->centeridx,
1707 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1708 
1709 	/*
1710 	 * check for pages to map, if we got any
1711 	 */
1712 
1713 	if (gotpages == 0) {
1714 		pages[flt->centeridx] = NULL;
1715 		return;
1716 	}
1717 
1718 	currva = flt->startva;
1719 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1720 		struct vm_page *curpg;
1721 
1722 		curpg = pages[lcv];
1723 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1724 			continue;
1725 		}
1726 		KASSERT(curpg->uobject == uobj);
1727 
1728 		/*
1729 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1730 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1731 		 * to it.
1732 		 */
1733 
1734 		if (lcv == flt->centeridx) {
1735 			UVMHIST_LOG(maphist, "  got uobjpage "
1736 			    "(0x%x) with locked get",
1737 			    curpg, 0,0,0);
1738 		} else {
1739 			bool readonly = (curpg->flags & PG_RDONLY)
1740 			    || (curpg->loan_count > 0)
1741 			    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
1742 
1743 			uvm_fault_lower_neighbor(ufi, flt,
1744 			    currva, curpg, readonly);
1745 		}
1746 	}
1747 	pmap_update(ufi->orig_map->pmap);
1748 }
1749 
1750 /*
1751  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1752  */
1753 
1754 static void
1755 uvm_fault_lower_neighbor(
1756 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1757 	vaddr_t currva, struct vm_page *pg, bool readonly)
1758 {
1759 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1760 
1761 	/* locked: maps(read), amap(if there), uobj */
1762 
1763 	/*
1764 	 * calling pgo_get with PGO_LOCKED returns us pages which
1765 	 * are neither busy nor released, so we don't need to check
1766 	 * for this.  we can just directly enter the pages.
1767 	 */
1768 
1769 	mutex_enter(&uvm_pageqlock);
1770 	uvm_pageenqueue(pg);
1771 	mutex_exit(&uvm_pageqlock);
1772 	UVMHIST_LOG(maphist,
1773 	    "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
1774 	    ufi->orig_map->pmap, currva, pg, 0);
1775 	uvmexp.fltnomap++;
1776 
1777 	/*
1778 	 * Since this page isn't the page that's actually faulting,
1779 	 * ignore pmap_enter() failures; it's not critical that we
1780 	 * enter these right now.
1781 	 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've
1782 	 * held the lock the whole time we've had the handle.
1783 	 */
1784 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1785 	KASSERT((pg->flags & PG_RELEASED) == 0);
1786 	KASSERT((pg->flags & PG_WANTED) == 0);
1787 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) ||
1788 	    (pg->flags & PG_CLEAN) != 0);
1789 	pg->flags &= ~(PG_BUSY);
1790 	UVM_PAGE_OWN(pg, NULL);
1791 
1792 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1793 	    VM_PAGE_TO_PHYS(pg),
1794 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1795 	    flt->enter_prot & MASK(ufi->entry),
1796 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1797 }
1798 
1799 /*
1800  * uvm_fault_lower_io: get lower page from backing store.
1801  *
1802  *	1. unlock everything, because i/o will block.
1803  *	2. call pgo_get.
1804  *	3. if failed, recover.
1805  *	4. if succeeded, relock everything and verify things.
1806  */
1807 
1808 static int
1809 uvm_fault_lower_io(
1810 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1811 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1812 {
1813 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1814 	struct uvm_object *uobj = *ruobj;
1815 	struct vm_page *pg;
1816 	bool locked;
1817 	int gotpages;
1818 	int error;
1819 	voff_t uoff;
1820 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1821 
1822 	/* update rusage counters */
1823 	curlwp->l_ru.ru_majflt++;
1824 
1825 	/* locked: maps(read), amap(if there), uobj */
1826 	uvmfault_unlockall(ufi, amap, NULL, NULL);
1827 	/* locked: uobj */
1828 
1829 	uvmexp.fltget++;
1830 	gotpages = 1;
1831 	pg = NULL;
1832 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1833 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1834 	    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1835 	    PGO_SYNCIO);
1836 	/* locked: pg(if no error) */
1837 
1838 	/*
1839 	 * recover from I/O
1840 	 */
1841 
1842 	if (error) {
1843 		if (error == EAGAIN) {
1844 			UVMHIST_LOG(maphist,
1845 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1846 			kpause("fltagain2", false, hz/2, NULL);
1847 			return ERESTART;
1848 		}
1849 
1850 #if 0
1851 		KASSERT(error != ERESTART);
1852 #else
1853 		/* XXXUEBS don't re-fault? */
1854 		if (error == ERESTART)
1855 			error = EIO;
1856 #endif
1857 
1858 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1859 		    error, 0,0,0);
1860 		return error;
1861 	}
1862 
1863 	/* locked: pg */
1864 
1865 	KASSERT((pg->flags & PG_BUSY) != 0);
1866 
1867 	mutex_enter(&uvm_pageqlock);
1868 	uvm_pageactivate(pg);
1869 	mutex_exit(&uvm_pageqlock);
1870 
1871 	/*
1872 	 * re-verify the state of the world by first trying to relock
1873 	 * the maps.  always relock the object.
1874 	 */
1875 
1876 	locked = uvmfault_relock(ufi);
1877 	if (locked && amap)
1878 		amap_lock(amap);
1879 
1880 	/* might be changed */
1881 	uobj = pg->uobject;
1882 
1883 	mutex_enter(&uobj->vmobjlock);
1884 
1885 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1886 	/* locked(!locked): uobj, pg */
1887 
1888 	/*
1889 	 * verify that the page has not be released and re-verify
1890 	 * that amap slot is still free.   if there is a problem,
1891 	 * we unlock and clean up.
1892 	 */
1893 
1894 	if ((pg->flags & PG_RELEASED) != 0 ||
1895 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1896 	      ufi->orig_rvaddr - ufi->entry->start))) {
1897 		if (locked)
1898 			uvmfault_unlockall(ufi, amap, NULL, NULL);
1899 		locked = false;
1900 	}
1901 
1902 	/*
1903 	 * didn't get the lock?   release the page and retry.
1904 	 */
1905 
1906 	if (locked == false) {
1907 		UVMHIST_LOG(maphist,
1908 		    "  wasn't able to relock after fault: retry",
1909 		    0,0,0,0);
1910 		if (pg->flags & PG_WANTED) {
1911 			wakeup(pg);
1912 		}
1913 		if (pg->flags & PG_RELEASED) {
1914 			uvmexp.fltpgrele++;
1915 			uvm_pagefree(pg);
1916 			mutex_exit(&uobj->vmobjlock);
1917 			return ERESTART;
1918 		}
1919 		pg->flags &= ~(PG_BUSY|PG_WANTED);
1920 		UVM_PAGE_OWN(pg, NULL);
1921 		mutex_exit(&uobj->vmobjlock);
1922 		return ERESTART;
1923 	}
1924 
1925 	/*
1926 	 * we have the data in pg which is busy and
1927 	 * not released.  we are holding object lock (so the page
1928 	 * can't be released on us).
1929 	 */
1930 
1931 	/* locked: maps(read), amap(if !null), uobj, pg */
1932 
1933 	*ruobj = uobj;
1934 	*ruobjpage = pg;
1935 	return 0;
1936 }
1937 
1938 /*
1939  * uvm_fault_lower_direct: fault lower center page
1940  *
1941  *	1. adjust flt->enter_prot.
1942  *	2. if page is loaned, resolve.
1943  */
1944 
1945 int
1946 uvm_fault_lower_direct(
1947 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1948 	struct uvm_object *uobj, struct vm_page *uobjpage)
1949 {
1950 	struct vm_page *pg;
1951 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
1952 
1953 	/*
1954 	 * we are not promoting.   if the mapping is COW ensure that we
1955 	 * don't give more access than we should (e.g. when doing a read
1956 	 * fault on a COPYONWRITE mapping we want to map the COW page in
1957 	 * R/O even though the entry protection could be R/W).
1958 	 *
1959 	 * set "pg" to the page we want to map in (uobjpage, usually)
1960 	 */
1961 
1962 	uvmexp.flt_obj++;
1963 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
1964 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
1965 		flt->enter_prot &= ~VM_PROT_WRITE;
1966 	pg = uobjpage;		/* map in the actual object */
1967 
1968 	KASSERT(uobjpage != PGO_DONTCARE);
1969 
1970 	/*
1971 	 * we are faulting directly on the page.   be careful
1972 	 * about writing to loaned pages...
1973 	 */
1974 
1975 	if (uobjpage->loan_count) {
1976 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
1977 	}
1978 	KASSERT(pg == uobjpage);
1979 
1980 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1981 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
1982 }
1983 
1984 /*
1985  * uvm_fault_lower_direct_loan: resolve loaned page.
1986  *
1987  *	1. if not cow'ing, adjust flt->enter_prot.
1988  *	2. if cow'ing, break loan.
1989  */
1990 
1991 static int
1992 uvm_fault_lower_direct_loan(
1993 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1994 	struct uvm_object *uobj, struct vm_page **rpg,
1995 	struct vm_page **ruobjpage)
1996 {
1997 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1998 	struct vm_page *pg;
1999 	struct vm_page *uobjpage = *ruobjpage;
2000 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
2001 
2002 	if (!flt->cow_now) {
2003 		/* read fault: cap the protection at readonly */
2004 		/* cap! */
2005 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2006 	} else {
2007 		/* write fault: must break the loan here */
2008 
2009 		pg = uvm_loanbreak(uobjpage);
2010 		if (pg == NULL) {
2011 
2012 			/*
2013 			 * drop ownership of page, it can't be released
2014 			 */
2015 
2016 			if (uobjpage->flags & PG_WANTED)
2017 				wakeup(uobjpage);
2018 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2019 			UVM_PAGE_OWN(uobjpage, NULL);
2020 
2021 			uvmfault_unlockall(ufi, amap, uobj, NULL);
2022 			UVMHIST_LOG(maphist,
2023 			  "  out of RAM breaking loan, waiting",
2024 			  0,0,0,0);
2025 			uvmexp.fltnoram++;
2026 			uvm_wait("flt_noram4");
2027 			return ERESTART;
2028 		}
2029 		*rpg = pg;
2030 		*ruobjpage = pg;
2031 	}
2032 	return 0;
2033 }
2034 
2035 /*
2036  * uvm_fault_lower_promote: promote lower page.
2037  *
2038  *	1. call uvmfault_promote.
2039  *	2. fill in data.
2040  *	3. if not ZFOD, dispose old page.
2041  */
2042 
2043 int
2044 uvm_fault_lower_promote(
2045 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2046 	struct uvm_object *uobj, struct vm_page *uobjpage)
2047 {
2048 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2049 	struct vm_anon *anon;
2050 	struct vm_page *pg;
2051 	int error;
2052 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2053 
2054 	/*
2055 	 * if we are going to promote the data to an anon we
2056 	 * allocate a blank anon here and plug it into our amap.
2057 	 */
2058 #if DIAGNOSTIC
2059 	if (amap == NULL)
2060 		panic("uvm_fault: want to promote data, but no anon");
2061 #endif
2062 	error = uvmfault_promote(ufi, NULL, uobjpage,
2063 	    &anon, &flt->anon_spare);
2064 	switch (error) {
2065 	case 0:
2066 		break;
2067 	case ERESTART:
2068 		return ERESTART;
2069 	default:
2070 		return error;
2071 	}
2072 
2073 	pg = anon->an_page;
2074 
2075 	/*
2076 	 * fill in the data
2077 	 */
2078 
2079 	if (uobjpage != PGO_DONTCARE) {
2080 		uvmexp.flt_prcopy++;
2081 
2082 		/*
2083 		 * promote to shared amap?  make sure all sharing
2084 		 * procs see it
2085 		 */
2086 
2087 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2088 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2089 			/*
2090 			 * XXX: PAGE MIGHT BE WIRED!
2091 			 */
2092 		}
2093 
2094 		/*
2095 		 * dispose of uobjpage.  it can't be PG_RELEASED
2096 		 * since we still hold the object lock.
2097 		 * drop handle to uobj as well.
2098 		 */
2099 
2100 		if (uobjpage->flags & PG_WANTED)
2101 			/* still have the obj lock */
2102 			wakeup(uobjpage);
2103 		uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2104 		UVM_PAGE_OWN(uobjpage, NULL);
2105 		mutex_exit(&uobj->vmobjlock);
2106 		uobj = NULL;
2107 
2108 		UVMHIST_LOG(maphist,
2109 		    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
2110 		    uobjpage, anon, pg, 0);
2111 
2112 	} else {
2113 		uvmexp.flt_przero++;
2114 
2115 		/*
2116 		 * Page is zero'd and marked dirty by
2117 		 * uvmfault_promote().
2118 		 */
2119 
2120 		UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
2121 		    anon, pg, 0, 0);
2122 	}
2123 
2124 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2125 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2126 }
2127 
2128 /*
2129  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2130  * from the lower page.
2131  */
2132 
2133 int
2134 uvm_fault_lower_enter(
2135 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2136 	struct uvm_object *uobj,
2137 	struct vm_anon *anon, struct vm_page *pg)
2138 {
2139 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2140 	int error;
2141 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2142 
2143 	/*
2144 	 * locked:
2145 	 * maps(read), amap(if !null), uobj(if !null),
2146 	 *   anon(if !null), pg(if anon)
2147 	 *
2148 	 * note: pg is either the uobjpage or the new page in the new anon
2149 	 */
2150 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
2151 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
2152 	KASSERT(anon == NULL || mutex_owned(&anon->an_lock));
2153 	KASSERT((pg->flags & PG_BUSY) != 0);
2154 
2155 	/*
2156 	 * all resources are present.   we can now map it in and free our
2157 	 * resources.
2158 	 */
2159 
2160 	UVMHIST_LOG(maphist,
2161 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
2162 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
2163 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 ||
2164 		(pg->flags & PG_RDONLY) == 0);
2165 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2166 	    VM_PAGE_TO_PHYS(pg),
2167 	    (pg->flags & PG_RDONLY) != 0 ?
2168 	    flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2169 	    flt->access_type | PMAP_CANFAIL |
2170 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2171 
2172 		/*
2173 		 * No need to undo what we did; we can simply think of
2174 		 * this as the pmap throwing away the mapping information.
2175 		 *
2176 		 * We do, however, have to go through the ReFault path,
2177 		 * as the map may change while we're asleep.
2178 		 */
2179 
2180 		/*
2181 		 * ensure that the page is queued in the case that
2182 		 * we just promoted the page.
2183 		 */
2184 
2185 		mutex_enter(&uvm_pageqlock);
2186 		uvm_pageenqueue(pg);
2187 		mutex_exit(&uvm_pageqlock);
2188 
2189 		if (pg->flags & PG_WANTED)
2190 			wakeup(pg);
2191 
2192 		/*
2193 		 * note that pg can't be PG_RELEASED since we did not drop
2194 		 * the object lock since the last time we checked.
2195 		 */
2196 		KASSERT((pg->flags & PG_RELEASED) == 0);
2197 
2198 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2199 		UVM_PAGE_OWN(pg, NULL);
2200 
2201 		uvmfault_unlockall(ufi, amap, uobj, anon);
2202 		if (!uvm_reclaimable()) {
2203 			UVMHIST_LOG(maphist,
2204 			    "<- failed.  out of VM",0,0,0,0);
2205 			/* XXX instrumentation */
2206 			error = ENOMEM;
2207 			return error;
2208 		}
2209 		/* XXX instrumentation */
2210 		uvm_wait("flt_pmfail2");
2211 		return ERESTART;
2212 	}
2213 
2214 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2215 
2216 	/*
2217 	 * note that pg can't be PG_RELEASED since we did not drop the object
2218 	 * lock since the last time we checked.
2219 	 */
2220 	KASSERT((pg->flags & PG_RELEASED) == 0);
2221 	if (pg->flags & PG_WANTED)
2222 		wakeup(pg);
2223 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2224 	UVM_PAGE_OWN(pg, NULL);
2225 
2226 	pmap_update(ufi->orig_map->pmap);
2227 	uvmfault_unlockall(ufi, amap, uobj, anon);
2228 
2229 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2230 	return 0;
2231 }
2232 
2233 /*
2234  * uvm_fault_lower_done: queue lower center page.
2235  */
2236 
2237 void
2238 uvm_fault_lower_done(
2239 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2240 	struct uvm_object *uobj, struct vm_page *pg)
2241 {
2242 	bool dropswap = false;
2243 
2244 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2245 
2246 	mutex_enter(&uvm_pageqlock);
2247 	if (flt->wire_paging) {
2248 		uvm_pagewire(pg);
2249 		if (pg->pqflags & PQ_AOBJ) {
2250 
2251 			/*
2252 			 * since the now-wired page cannot be paged out,
2253 			 * release its swap resources for others to use.
2254 			 * since an aobj page with no swap cannot be PG_CLEAN,
2255 			 * clear its clean flag now.
2256 			 */
2257 
2258 			KASSERT(uobj != NULL);
2259 			pg->flags &= ~(PG_CLEAN);
2260 			dropswap = true;
2261 		}
2262 	} else {
2263 		uvm_pageactivate(pg);
2264 	}
2265 	mutex_exit(&uvm_pageqlock);
2266 
2267 	if (dropswap) {
2268 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2269 	}
2270 }
2271 
2272 
2273 /*
2274  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2275  *
2276  * => map may be read-locked by caller, but MUST NOT be write-locked.
2277  * => if map is read-locked, any operations which may cause map to
2278  *	be write-locked in uvm_fault() must be taken care of by
2279  *	the caller.  See uvm_map_pageable().
2280  */
2281 
2282 int
2283 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2284     vm_prot_t access_type, int maxprot)
2285 {
2286 	vaddr_t va;
2287 	int error;
2288 
2289 	/*
2290 	 * now fault it in a page at a time.   if the fault fails then we have
2291 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2292 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2293 	 */
2294 
2295 	/*
2296 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2297 	 * wiring the last page in the address space, though.
2298 	 */
2299 	if (start > end) {
2300 		return EFAULT;
2301 	}
2302 
2303 	for (va = start; va < end; va += PAGE_SIZE) {
2304 		error = uvm_fault_internal(map, va, access_type,
2305 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2306 		if (error) {
2307 			if (va != start) {
2308 				uvm_fault_unwire(map, start, va);
2309 			}
2310 			return error;
2311 		}
2312 	}
2313 	return 0;
2314 }
2315 
2316 /*
2317  * uvm_fault_unwire(): unwire range of virtual space.
2318  */
2319 
2320 void
2321 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2322 {
2323 	vm_map_lock_read(map);
2324 	uvm_fault_unwire_locked(map, start, end);
2325 	vm_map_unlock_read(map);
2326 }
2327 
2328 /*
2329  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2330  *
2331  * => map must be at least read-locked.
2332  */
2333 
2334 void
2335 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2336 {
2337 	struct vm_map_entry *entry;
2338 	pmap_t pmap = vm_map_pmap(map);
2339 	vaddr_t va;
2340 	paddr_t pa;
2341 	struct vm_page *pg;
2342 
2343 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
2344 
2345 	/*
2346 	 * we assume that the area we are unwiring has actually been wired
2347 	 * in the first place.   this means that we should be able to extract
2348 	 * the PAs from the pmap.   we also lock out the page daemon so that
2349 	 * we can call uvm_pageunwire.
2350 	 */
2351 
2352 	mutex_enter(&uvm_pageqlock);
2353 
2354 	/*
2355 	 * find the beginning map entry for the region.
2356 	 */
2357 
2358 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2359 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2360 		panic("uvm_fault_unwire_locked: address not in map");
2361 
2362 	for (va = start; va < end; va += PAGE_SIZE) {
2363 		if (pmap_extract(pmap, va, &pa) == false)
2364 			continue;
2365 
2366 		/*
2367 		 * find the map entry for the current address.
2368 		 */
2369 
2370 		KASSERT(va >= entry->start);
2371 		while (va >= entry->end) {
2372 			KASSERT(entry->next != &map->header &&
2373 				entry->next->start <= entry->end);
2374 			entry = entry->next;
2375 		}
2376 
2377 		/*
2378 		 * if the entry is no longer wired, tell the pmap.
2379 		 */
2380 
2381 		if (VM_MAPENT_ISWIRED(entry) == 0)
2382 			pmap_unwire(pmap, va);
2383 
2384 		pg = PHYS_TO_VM_PAGE(pa);
2385 		if (pg)
2386 			uvm_pageunwire(pg);
2387 	}
2388 
2389 	mutex_exit(&uvm_pageqlock);
2390 }
2391