xref: /netbsd-src/sys/uvm/uvm_fault.c (revision afab4e300d3a9fb07dd8c80daf53d0feb3345706)
1 /*	$NetBSD: uvm_fault.c,v 1.232 2023/04/09 09:00:56 riastradh 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 1.232 2023/04/09 09:00:56 riastradh Exp $");
36 
37 #include "opt_uvmhist.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/atomic.h>
42 #include <sys/kernel.h>
43 #include <sys/mman.h>
44 
45 #include <uvm/uvm.h>
46 #include <uvm/uvm_pdpolicy.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 	{ UVM_ADV_NORMAL, 3, 4 },
163 	{ UVM_ADV_RANDOM, 0, 0 },
164 	{ UVM_ADV_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 		KASSERT(rw_lock_held(anons[lcv]->an_lock));
193 		pg = anons[lcv]->an_page;
194 		if (pg && (pg->flags & PG_BUSY) == 0) {
195 			uvm_pagelock(pg);
196 			uvm_pagedeactivate(pg);
197 			uvm_pageunlock(pg);
198 		}
199 	}
200 }
201 
202 /*
203  * normal functions
204  */
205 
206 /*
207  * uvmfault_amapcopy: clear "needs_copy" in a map.
208  *
209  * => called with VM data structures unlocked (usually, see below)
210  * => we get a write lock on the maps and clear needs_copy for a VA
211  * => if we are out of RAM we sleep (waiting for more)
212  */
213 
214 static void
215 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
216 {
217 	for (;;) {
218 
219 		/*
220 		 * no mapping?  give up.
221 		 */
222 
223 		if (uvmfault_lookup(ufi, true) == false)
224 			return;
225 
226 		/*
227 		 * copy if needed.
228 		 */
229 
230 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
231 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
232 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
233 
234 		/*
235 		 * didn't work?  must be out of RAM.   unlock and sleep.
236 		 */
237 
238 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
239 			uvmfault_unlockmaps(ufi, true);
240 			uvm_wait("fltamapcopy");
241 			continue;
242 		}
243 
244 		/*
245 		 * got it!   unlock and return.
246 		 */
247 
248 		uvmfault_unlockmaps(ufi, true);
249 		return;
250 	}
251 	/*NOTREACHED*/
252 }
253 
254 /*
255  * uvmfault_anonget: get data in an anon into a non-busy, non-released
256  * page in that anon.
257  *
258  * => Map, amap and thus anon should be locked by caller.
259  * => If we fail, we unlock everything and error is returned.
260  * => If we are successful, return with everything still locked.
261  * => We do not move the page on the queues [gets moved later].  If we
262  *    allocate a new page [we_own], it gets put on the queues.  Either way,
263  *    the result is that the page is on the queues at return time
264  * => For pages which are on loan from a uvm_object (and thus are not owned
265  *    by the anon): if successful, return with the owning object locked.
266  *    The caller must unlock this object when it unlocks everything else.
267  */
268 
269 int
270 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
271     struct vm_anon *anon)
272 {
273 	struct vm_page *pg;
274 	krw_t lock_type;
275 	int error;
276 
277 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
278 	KASSERT(rw_lock_held(anon->an_lock));
279 	KASSERT(anon->an_lock == amap->am_lock);
280 
281 	/* Increment the counters.*/
282 	cpu_count(CPU_COUNT_FLTANGET, 1);
283 	if (anon->an_page) {
284 		curlwp->l_ru.ru_minflt++;
285 	} else {
286 		curlwp->l_ru.ru_majflt++;
287 	}
288 	error = 0;
289 
290 	/*
291 	 * Loop until we get the anon data, or fail.
292 	 */
293 
294 	for (;;) {
295 		bool we_own, locked;
296 		/*
297 		 * Note: 'we_own' will become true if we set PG_BUSY on a page.
298 		 */
299 		we_own = false;
300 		pg = anon->an_page;
301 
302 		/*
303 		 * If there is a resident page and it is loaned, then anon
304 		 * may not own it.  Call out to uvm_anon_lockloanpg() to
305 		 * identify and lock the real owner of the page.
306 		 */
307 
308 		if (pg && pg->loan_count)
309 			pg = uvm_anon_lockloanpg(anon);
310 
311 		/*
312 		 * Is page resident?  Make sure it is not busy/released.
313 		 */
314 
315 		lock_type = rw_lock_op(anon->an_lock);
316 		if (pg) {
317 
318 			/*
319 			 * at this point, if the page has a uobject [meaning
320 			 * we have it on loan], then that uobject is locked
321 			 * by us!   if the page is busy, we drop all the
322 			 * locks (including uobject) and try again.
323 			 */
324 
325 			if ((pg->flags & PG_BUSY) == 0) {
326 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
327 				return 0;
328 			}
329 			cpu_count(CPU_COUNT_FLTPGWAIT, 1);
330 
331 			/*
332 			 * The last unlock must be an atomic unlock and wait
333 			 * on the owner of page.
334 			 */
335 
336 			if (pg->uobject) {
337 				/* Owner of page is UVM object. */
338 				uvmfault_unlockall(ufi, amap, NULL);
339 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
340 				    0,0,0);
341 				uvm_pagewait(pg, pg->uobject->vmobjlock, "anonget1");
342 			} else {
343 				/* Owner of page is anon. */
344 				uvmfault_unlockall(ufi, NULL, NULL);
345 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
346 				    0,0,0);
347 				uvm_pagewait(pg, anon->an_lock, "anonget2");
348 			}
349 		} else {
350 #if defined(VMSWAP)
351 			/*
352 			 * No page, therefore allocate one.  A write lock is
353 			 * required for this.  If the caller didn't supply
354 			 * one, fail now and have them retry.
355 			 */
356 
357 			if (lock_type == RW_READER) {
358 				return ENOLCK;
359 			}
360 			pg = uvm_pagealloc(NULL,
361 			    ufi != NULL ? ufi->orig_rvaddr : 0,
362 			    anon, ufi != NULL ? UVM_FLAG_COLORMATCH : 0);
363 			if (pg == NULL) {
364 				/* Out of memory.  Wait a little. */
365 				uvmfault_unlockall(ufi, amap, NULL);
366 				cpu_count(CPU_COUNT_FLTNORAM, 1);
367 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
368 				    0,0,0);
369 				if (!uvm_reclaimable()) {
370 					return ENOMEM;
371 				}
372 				uvm_wait("flt_noram1");
373 			} else {
374 				/* PG_BUSY bit is set. */
375 				we_own = true;
376 				uvmfault_unlockall(ufi, amap, NULL);
377 
378 				/*
379 				 * Pass a PG_BUSY+PG_FAKE clean page into
380 				 * the uvm_swap_get() function with all data
381 				 * structures unlocked.  Note that it is OK
382 				 * to read an_swslot here, because we hold
383 				 * PG_BUSY on the page.
384 				 */
385 				cpu_count(CPU_COUNT_PAGEINS, 1);
386 				error = uvm_swap_get(pg, anon->an_swslot,
387 				    PGO_SYNCIO);
388 
389 				/*
390 				 * We clean up after the I/O below in the
391 				 * 'we_own' case.
392 				 */
393 			}
394 #else
395 			panic("%s: no page", __func__);
396 #endif /* defined(VMSWAP) */
397 		}
398 
399 		/*
400 		 * Re-lock the map and anon.
401 		 */
402 
403 		locked = uvmfault_relock(ufi);
404 		if (locked || we_own) {
405 			rw_enter(anon->an_lock, lock_type);
406 		}
407 
408 		/*
409 		 * If we own the page (i.e. we set PG_BUSY), then we need
410 		 * to clean up after the I/O.  There are three cases to
411 		 * consider:
412 		 *
413 		 * 1) Page was released during I/O: free anon and ReFault.
414 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
415 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
416 		 *    case (i.e. drop anon lock if not locked).
417 		 */
418 
419 		if (we_own) {
420 			KASSERT(lock_type == RW_WRITER);
421 #if defined(VMSWAP)
422 			if (error) {
423 
424 				/*
425 				 * Remove the swap slot from the anon and
426 				 * mark the anon as having no real slot.
427 				 * Do not 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 				}
434 				anon->an_swslot = SWSLOT_BAD;
435 
436 				if ((pg->flags & PG_RELEASED) != 0) {
437 					goto released;
438 				}
439 
440 				/*
441 				 * Note: page was never !PG_BUSY, so it
442 				 * cannot be mapped and thus no need to
443 				 * pmap_page_protect() it.
444 				 */
445 
446 				uvm_pagefree(pg);
447 
448 				if (locked) {
449 					uvmfault_unlockall(ufi, NULL, NULL);
450 				}
451 				rw_exit(anon->an_lock);
452 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
453 				return error;
454 			}
455 
456 			if ((pg->flags & PG_RELEASED) != 0) {
457 released:
458 				KASSERT(anon->an_ref == 0);
459 
460 				/*
461 				 * Released while we had unlocked amap.
462 				 */
463 
464 				if (locked) {
465 					uvmfault_unlockall(ufi, NULL, NULL);
466 				}
467 				uvm_anon_release(anon);
468 
469 				if (error) {
470 					UVMHIST_LOG(maphist,
471 					    "<- ERROR/RELEASED", 0,0,0,0);
472 					return error;
473 				}
474 
475 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
476 				return ERESTART;
477 			}
478 
479 			/*
480 			 * We have successfully read the page, activate it.
481 			 */
482 
483 			uvm_pagelock(pg);
484 			uvm_pageactivate(pg);
485 			uvm_pagewakeup(pg);
486 			uvm_pageunlock(pg);
487 			pg->flags &= ~(PG_BUSY|PG_FAKE);
488 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN);
489 			UVM_PAGE_OWN(pg, NULL);
490 #else
491 			panic("%s: we_own", __func__);
492 #endif /* defined(VMSWAP) */
493 		}
494 
495 		/*
496 		 * We were not able to re-lock the map - restart the fault.
497 		 */
498 
499 		if (!locked) {
500 			if (we_own) {
501 				rw_exit(anon->an_lock);
502 			}
503 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
504 			return ERESTART;
505 		}
506 
507 		/*
508 		 * Verify that no one has touched the amap and moved
509 		 * the anon on us.
510 		 */
511 
512 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
513 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
514 
515 			uvmfault_unlockall(ufi, amap, NULL);
516 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
517 			return ERESTART;
518 		}
519 
520 		/*
521 		 * Retry..
522 		 */
523 
524 		cpu_count(CPU_COUNT_FLTANRETRY, 1);
525 		continue;
526 	}
527 	/*NOTREACHED*/
528 }
529 
530 /*
531  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
532  *
533  *	1. allocate an anon and a page.
534  *	2. fill its contents.
535  *	3. put it into amap.
536  *
537  * => if we fail (result != 0) we unlock everything.
538  * => on success, return a new locked anon via 'nanon'.
539  *    (*nanon)->an_page will be a resident, locked, dirty page.
540  * => it's caller's responsibility to put the promoted nanon->an_page to the
541  *    page queue.
542  */
543 
544 static int
545 uvmfault_promote(struct uvm_faultinfo *ufi,
546     struct vm_anon *oanon,
547     struct vm_page *uobjpage,
548     struct vm_anon **nanon, /* OUT: allocated anon */
549     struct vm_anon **spare)
550 {
551 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
552 	struct uvm_object *uobj;
553 	struct vm_anon *anon;
554 	struct vm_page *pg;
555 	struct vm_page *opg;
556 	int error;
557 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
558 
559 	if (oanon) {
560 		/* anon COW */
561 		opg = oanon->an_page;
562 		KASSERT(opg != NULL);
563 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
564 	} else if (uobjpage != PGO_DONTCARE) {
565 		/* object-backed COW */
566 		opg = uobjpage;
567 		KASSERT(rw_lock_held(opg->uobject->vmobjlock));
568 	} else {
569 		/* ZFOD */
570 		opg = NULL;
571 	}
572 	if (opg != NULL) {
573 		uobj = opg->uobject;
574 	} else {
575 		uobj = NULL;
576 	}
577 
578 	KASSERT(amap != NULL);
579 	KASSERT(uobjpage != NULL);
580 	KASSERT(rw_write_held(amap->am_lock));
581 	KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock);
582 	KASSERT(uobj == NULL || rw_lock_held(uobj->vmobjlock));
583 
584 	if (*spare != NULL) {
585 		anon = *spare;
586 		*spare = NULL;
587 	} else {
588 		anon = uvm_analloc();
589 	}
590 	if (anon) {
591 
592 		/*
593 		 * The new anon is locked.
594 		 *
595 		 * if opg == NULL, we want a zero'd, dirty page,
596 		 * so have uvm_pagealloc() do that for us.
597 		 */
598 
599 		KASSERT(anon->an_lock == NULL);
600 		anon->an_lock = amap->am_lock;
601 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
602 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
603 		if (pg == NULL) {
604 			anon->an_lock = NULL;
605 		}
606 	} else {
607 		pg = NULL;
608 	}
609 
610 	/*
611 	 * out of memory resources?
612 	 */
613 
614 	if (pg == NULL) {
615 		/* save anon for the next try. */
616 		if (anon != NULL) {
617 			*spare = anon;
618 		}
619 
620 		/* unlock and fail ... */
621 		uvmfault_unlockall(ufi, amap, uobj);
622 		if (!uvm_reclaimable()) {
623 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
624 			cpu_count(CPU_COUNT_FLTNOANON, 1);
625 			error = ENOMEM;
626 			goto done;
627 		}
628 
629 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
630 		cpu_count(CPU_COUNT_FLTNORAM, 1);
631 		uvm_wait("flt_noram5");
632 		error = ERESTART;
633 		goto done;
634 	}
635 
636 	/* copy page [pg now dirty] */
637 	if (opg) {
638 		uvm_pagecopy(opg, pg);
639 	}
640 	KASSERT(uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_DIRTY);
641 
642 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
643 	    oanon != NULL);
644 
645 	/*
646 	 * from this point on am_lock won't be dropped until the page is
647 	 * entered, so it's safe to unbusy the page up front.
648 	 *
649 	 * uvm_fault_{upper,lower}_done will activate or enqueue the page.
650 	 */
651 
652 	pg = anon->an_page;
653 	pg->flags &= ~(PG_BUSY|PG_FAKE);
654 	UVM_PAGE_OWN(pg, NULL);
655 
656 	*nanon = anon;
657 	error = 0;
658 done:
659 	return error;
660 }
661 
662 /*
663  * Update statistics after fault resolution.
664  * - maxrss
665  */
666 void
667 uvmfault_update_stats(struct uvm_faultinfo *ufi)
668 {
669 	struct vm_map		*map;
670 	struct vmspace 		*vm;
671 	struct proc		*p;
672 	vsize_t			 res;
673 
674 	map = ufi->orig_map;
675 
676 	p = curproc;
677 	KASSERT(p != NULL);
678 	vm = p->p_vmspace;
679 
680 	if (&vm->vm_map != map)
681 		return;
682 
683 	res = pmap_resident_count(map->pmap);
684 	if (vm->vm_rssmax < res)
685 		vm->vm_rssmax = res;
686 }
687 
688 /*
689  *   F A U L T   -   m a i n   e n t r y   p o i n t
690  */
691 
692 /*
693  * uvm_fault: page fault handler
694  *
695  * => called from MD code to resolve a page fault
696  * => VM data structures usually should be unlocked.   however, it is
697  *	possible to call here with the main map locked if the caller
698  *	gets a write lock, sets it recursive, and then calls us (c.f.
699  *	uvm_map_pageable).   this should be avoided because it keeps
700  *	the map locked off during I/O.
701  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
702  */
703 
704 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
705 			 ~VM_PROT_WRITE : VM_PROT_ALL)
706 
707 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
708 #define UVM_FAULT_WIRE		(1 << 0)
709 #define UVM_FAULT_MAXPROT	(1 << 1)
710 
711 struct uvm_faultctx {
712 
713 	/*
714 	 * the following members are set up by uvm_fault_check() and
715 	 * read-only after that.
716 	 *
717 	 * note that narrow is used by uvm_fault_check() to change
718 	 * the behaviour after ERESTART.
719 	 *
720 	 * most of them might change after RESTART if the underlying
721 	 * map entry has been changed behind us.  an exception is
722 	 * wire_paging, which does never change.
723 	 */
724 	vm_prot_t access_type;
725 	vaddr_t startva;
726 	int npages;
727 	int centeridx;
728 	bool narrow;		/* work on a single requested page only */
729 	bool wire_mapping;	/* request a PMAP_WIRED mapping
730 				   (UVM_FAULT_WIRE or VM_MAPENT_ISWIRED) */
731 	bool wire_paging;	/* request uvm_pagewire
732 				   (true for UVM_FAULT_WIRE) */
733 	bool cow_now;		/* VM_PROT_WRITE is actually requested
734 				   (ie. should break COW and page loaning) */
735 
736 	/*
737 	 * enter_prot is set up by uvm_fault_check() and clamped
738 	 * (ie. drop the VM_PROT_WRITE bit) in various places in case
739 	 * of !cow_now.
740 	 */
741 	vm_prot_t enter_prot;	/* prot at which we want to enter pages in */
742 
743 	/*
744 	 * the following member is for uvmfault_promote() and ERESTART.
745 	 */
746 	struct vm_anon *anon_spare;
747 
748 	/*
749 	 * the following is actually a uvm_fault_lower() internal.
750 	 * it's here merely for debugging.
751 	 * (or due to the mechanical separation of the function?)
752 	 */
753 	bool promote;
754 
755 	/*
756 	 * type of lock to acquire on objects in both layers.
757 	 */
758 	krw_t lower_lock_type;
759 	krw_t upper_lock_type;
760 };
761 
762 static inline int	uvm_fault_check(
763 			    struct uvm_faultinfo *, struct uvm_faultctx *,
764 			    struct vm_anon ***, bool);
765 
766 static int		uvm_fault_upper(
767 			    struct uvm_faultinfo *, struct uvm_faultctx *,
768 			    struct vm_anon **);
769 static inline int	uvm_fault_upper_lookup(
770 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
771 			    struct vm_anon **, struct vm_page **);
772 static inline void	uvm_fault_upper_neighbor(
773 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
774 			    vaddr_t, struct vm_page *, bool);
775 static inline int	uvm_fault_upper_loan(
776 			    struct uvm_faultinfo *, struct uvm_faultctx *,
777 			    struct vm_anon *, struct uvm_object **);
778 static inline int	uvm_fault_upper_promote(
779 			    struct uvm_faultinfo *, struct uvm_faultctx *,
780 			    struct uvm_object *, struct vm_anon *);
781 static inline int	uvm_fault_upper_direct(
782 			    struct uvm_faultinfo *, struct uvm_faultctx *,
783 			    struct uvm_object *, struct vm_anon *);
784 static int		uvm_fault_upper_enter(
785 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
786 			    struct uvm_object *, struct vm_anon *,
787 			    struct vm_page *, struct vm_anon *);
788 static inline void	uvm_fault_upper_done(
789 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
790 			    struct vm_anon *, struct vm_page *);
791 
792 static int		uvm_fault_lower(
793 			    struct uvm_faultinfo *, struct uvm_faultctx *,
794 			    struct vm_page **);
795 static inline void	uvm_fault_lower_lookup(
796 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
797 			    struct vm_page **);
798 static inline void	uvm_fault_lower_neighbor(
799 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
800 			    vaddr_t, struct vm_page *);
801 static inline int	uvm_fault_lower_io(
802 			    struct uvm_faultinfo *, struct uvm_faultctx *,
803 			    struct uvm_object **, struct vm_page **);
804 static inline int	uvm_fault_lower_direct(
805 			    struct uvm_faultinfo *, struct uvm_faultctx *,
806 			    struct uvm_object *, struct vm_page *);
807 static inline int	uvm_fault_lower_direct_loan(
808 			    struct uvm_faultinfo *, struct uvm_faultctx *,
809 			    struct uvm_object *, struct vm_page **,
810 			    struct vm_page **);
811 static inline int	uvm_fault_lower_promote(
812 			    struct uvm_faultinfo *, struct uvm_faultctx *,
813 			    struct uvm_object *, struct vm_page *);
814 static int		uvm_fault_lower_enter(
815 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
816 			    struct uvm_object *,
817 			    struct vm_anon *, struct vm_page *);
818 static inline void	uvm_fault_lower_done(
819 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
820 			    struct uvm_object *, struct vm_page *);
821 
822 int
823 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
824     vm_prot_t access_type, int fault_flag)
825 {
826 	struct uvm_faultinfo ufi;
827 	struct uvm_faultctx flt = {
828 		.access_type = access_type,
829 
830 		/* don't look for neighborhood * pages on "wire" fault */
831 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
832 
833 		/* "wire" fault causes wiring of both mapping and paging */
834 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
835 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
836 
837 		/*
838 		 * default lock type to acquire on upper & lower layer
839 		 * objects: reader.  this can be upgraded at any point
840 		 * during the fault from read -> write and uvm_faultctx
841 		 * changed to match, but is never downgraded write -> read.
842 		 */
843 #ifdef __HAVE_UNLOCKED_PMAP /* XXX temporary */
844 		.upper_lock_type = RW_WRITER,
845 		.lower_lock_type = RW_WRITER,
846 #else
847 		.upper_lock_type = RW_READER,
848 		.lower_lock_type = RW_READER,
849 #endif
850 	};
851 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
852 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
853 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
854 	int error;
855 
856 	UVMHIST_FUNC(__func__);
857 	UVMHIST_CALLARGS(maphist, "(map=%#jx, vaddr=%#jx, at=%jd, ff=%jd)",
858 	      (uintptr_t)orig_map, vaddr, access_type, fault_flag);
859 
860 	/* Don't count anything until user interaction is possible */
861 	kpreempt_disable();
862 	if (__predict_true(start_init_exec)) {
863 		struct cpu_info *ci = curcpu();
864 		CPU_COUNT(CPU_COUNT_NFAULT, 1);
865 		/* Don't flood RNG subsystem with samples. */
866 		if (++(ci->ci_faultrng) == 503) {
867 			ci->ci_faultrng = 0;
868 			rnd_add_uint32(&curcpu()->ci_data.cpu_uvm->rs,
869 			    sizeof(vaddr_t) == sizeof(uint32_t) ?
870 			    (uint32_t)vaddr : sizeof(vaddr_t) ==
871 			    sizeof(uint64_t) ?
872 			    (uint32_t)vaddr :
873 			    (uint32_t)ci->ci_counts[CPU_COUNT_NFAULT]);
874 		}
875 	}
876 	kpreempt_enable();
877 
878 	/*
879 	 * init the IN parameters in the ufi
880 	 */
881 
882 	ufi.orig_map = orig_map;
883 	ufi.orig_rvaddr = trunc_page(vaddr);
884 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
885 
886 	error = ERESTART;
887 	while (error == ERESTART) { /* ReFault: */
888 		anons = anons_store;
889 		pages = pages_store;
890 
891 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
892 		if (error != 0)
893 			continue;
894 
895 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
896 		if (error != 0)
897 			continue;
898 
899 		if (pages[flt.centeridx] == PGO_DONTCARE)
900 			error = uvm_fault_upper(&ufi, &flt, anons);
901 		else {
902 			struct uvm_object * const uobj =
903 			    ufi.entry->object.uvm_obj;
904 
905 			if (uobj && uobj->pgops->pgo_fault != NULL) {
906 				/*
907 				 * invoke "special" fault routine.
908 				 */
909 				rw_enter(uobj->vmobjlock, RW_WRITER);
910 				/* locked: maps(read), amap(if there), uobj */
911 				error = uobj->pgops->pgo_fault(&ufi,
912 				    flt.startva, pages, flt.npages,
913 				    flt.centeridx, flt.access_type,
914 				    PGO_LOCKED|PGO_SYNCIO);
915 
916 				/*
917 				 * locked: nothing, pgo_fault has unlocked
918 				 * everything
919 				 */
920 
921 				/*
922 				 * object fault routine responsible for
923 				 * pmap_update().
924 				 */
925 
926 				/*
927 				 * Wake up the pagedaemon if the fault method
928 				 * failed for lack of memory but some can be
929 				 * reclaimed.
930 				 */
931 				if (error == ENOMEM && uvm_reclaimable()) {
932 					uvm_wait("pgo_fault");
933 					error = ERESTART;
934 				}
935 			} else {
936 				error = uvm_fault_lower(&ufi, &flt, pages);
937 			}
938 		}
939 	}
940 
941 	if (flt.anon_spare != NULL) {
942 		flt.anon_spare->an_ref--;
943 		KASSERT(flt.anon_spare->an_ref == 0);
944 		KASSERT(flt.anon_spare->an_lock == NULL);
945 		uvm_anfree(flt.anon_spare);
946 	}
947 	return error;
948 }
949 
950 /*
951  * uvm_fault_check: check prot, handle needs-copy, etc.
952  *
953  *	1. lookup entry.
954  *	2. check protection.
955  *	3. adjust fault condition (mainly for simulated fault).
956  *	4. handle needs-copy (lazy amap copy).
957  *	5. establish range of interest for neighbor fault (aka pre-fault).
958  *	6. look up anons (if amap exists).
959  *	7. flush pages (if MADV_SEQUENTIAL)
960  *
961  * => called with nothing locked.
962  * => if we fail (result != 0) we unlock everything.
963  * => initialize/adjust many members of flt.
964  */
965 
966 static int
967 uvm_fault_check(
968 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
969 	struct vm_anon ***ranons, bool maxprot)
970 {
971 	struct vm_amap *amap;
972 	struct uvm_object *uobj;
973 	vm_prot_t check_prot;
974 	int nback, nforw;
975 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
976 
977 	/*
978 	 * lookup and lock the maps
979 	 */
980 
981 	if (uvmfault_lookup(ufi, false) == false) {
982 		UVMHIST_LOG(maphist, "<- no mapping @ %#jx", ufi->orig_rvaddr,
983 		    0,0,0);
984 		return EFAULT;
985 	}
986 	/* locked: maps(read) */
987 
988 #ifdef DIAGNOSTIC
989 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
990 		printf("Page fault on non-pageable map:\n");
991 		printf("ufi->map = %p\n", ufi->map);
992 		printf("ufi->orig_map = %p\n", ufi->orig_map);
993 		printf("ufi->orig_rvaddr = %#lx\n", (u_long) ufi->orig_rvaddr);
994 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
995 	}
996 #endif
997 
998 	/*
999 	 * check protection
1000 	 */
1001 
1002 	check_prot = maxprot ?
1003 	    ufi->entry->max_protection : ufi->entry->protection;
1004 	if ((check_prot & flt->access_type) != flt->access_type) {
1005 		UVMHIST_LOG(maphist,
1006 		    "<- protection failure (prot=%#jx, access=%#jx)",
1007 		    ufi->entry->protection, flt->access_type, 0, 0);
1008 		uvmfault_unlockmaps(ufi, false);
1009 		return EFAULT;
1010 	}
1011 
1012 	/*
1013 	 * "enter_prot" is the protection we want to enter the page in at.
1014 	 * for certain pages (e.g. copy-on-write pages) this protection can
1015 	 * be more strict than ufi->entry->protection.  "wired" means either
1016 	 * the entry is wired or we are fault-wiring the pg.
1017 	 */
1018 
1019 	flt->enter_prot = ufi->entry->protection;
1020 	if (VM_MAPENT_ISWIRED(ufi->entry)) {
1021 		flt->wire_mapping = true;
1022 		flt->wire_paging = true;
1023 		flt->narrow = true;
1024 	}
1025 
1026 	if (flt->wire_mapping) {
1027 		flt->access_type = flt->enter_prot; /* full access for wired */
1028 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
1029 	} else {
1030 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
1031 	}
1032 
1033 	if (flt->wire_paging) {
1034 		/* wiring pages requires a write lock. */
1035 		flt->upper_lock_type = RW_WRITER;
1036 		flt->lower_lock_type = RW_WRITER;
1037 	}
1038 
1039 	flt->promote = false;
1040 
1041 	/*
1042 	 * handle "needs_copy" case.   if we need to copy the amap we will
1043 	 * have to drop our readlock and relock it with a write lock.  (we
1044 	 * need a write lock to change anything in a map entry [e.g.
1045 	 * needs_copy]).
1046 	 */
1047 
1048 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
1049 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
1050 			KASSERT(!maxprot);
1051 			/* need to clear */
1052 			UVMHIST_LOG(maphist,
1053 			    "  need to clear needs_copy and refault",0,0,0,0);
1054 			uvmfault_unlockmaps(ufi, false);
1055 			uvmfault_amapcopy(ufi);
1056 			cpu_count(CPU_COUNT_FLTAMCOPY, 1);
1057 			return ERESTART;
1058 
1059 		} else {
1060 
1061 			/*
1062 			 * ensure that we pmap_enter page R/O since
1063 			 * needs_copy is still true
1064 			 */
1065 
1066 			flt->enter_prot &= ~VM_PROT_WRITE;
1067 		}
1068 	}
1069 
1070 	/*
1071 	 * identify the players
1072 	 */
1073 
1074 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
1075 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
1076 
1077 	/*
1078 	 * check for a case 0 fault.  if nothing backing the entry then
1079 	 * error now.
1080 	 */
1081 
1082 	if (amap == NULL && uobj == NULL) {
1083 		uvmfault_unlockmaps(ufi, false);
1084 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
1085 		return EFAULT;
1086 	}
1087 
1088 	/*
1089 	 * for a case 2B fault waste no time on adjacent pages because
1090 	 * they are likely already entered.
1091 	 */
1092 
1093 	if (uobj != NULL && amap != NULL &&
1094 	    (flt->access_type & VM_PROT_WRITE) != 0) {
1095 		/* wide fault (!narrow) */
1096 		flt->narrow = true;
1097 	}
1098 
1099 	/*
1100 	 * establish range of interest based on advice from mapper
1101 	 * and then clip to fit map entry.   note that we only want
1102 	 * to do this the first time through the fault.   if we
1103 	 * ReFault we will disable this by setting "narrow" to true.
1104 	 */
1105 
1106 	if (flt->narrow == false) {
1107 
1108 		/* wide fault (!narrow) */
1109 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
1110 			 ufi->entry->advice);
1111 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
1112 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
1113 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
1114 		/*
1115 		 * note: "-1" because we don't want to count the
1116 		 * faulting page as forw
1117 		 */
1118 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1119 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1120 			     PAGE_SHIFT) - 1);
1121 		flt->npages = nback + nforw + 1;
1122 		flt->centeridx = nback;
1123 
1124 		flt->narrow = true;	/* ensure only once per-fault */
1125 
1126 	} else {
1127 
1128 		/* narrow fault! */
1129 		nback = nforw = 0;
1130 		flt->startva = ufi->orig_rvaddr;
1131 		flt->npages = 1;
1132 		flt->centeridx = 0;
1133 
1134 	}
1135 	/* offset from entry's start to pgs' start */
1136 	const voff_t eoff = flt->startva - ufi->entry->start;
1137 
1138 	/* locked: maps(read) */
1139 	UVMHIST_LOG(maphist, "  narrow=%jd, back=%jd, forw=%jd, startva=%#jx",
1140 		    flt->narrow, nback, nforw, flt->startva);
1141 	UVMHIST_LOG(maphist, "  entry=%#jx, amap=%#jx, obj=%#jx",
1142 	    (uintptr_t)ufi->entry, (uintptr_t)amap, (uintptr_t)uobj, 0);
1143 
1144 	/*
1145 	 * guess at the most suitable lock types to acquire.
1146 	 * if we've got an amap then lock it and extract current anons.
1147 	 */
1148 
1149 	if (amap) {
1150 		if ((amap_flags(amap) & AMAP_SHARED) == 0) {
1151 			/*
1152 			 * the amap isn't shared.  get a writer lock to
1153 			 * avoid the cost of upgrading the lock later if
1154 			 * needed.
1155 			 *
1156 			 * XXX nice for PostgreSQL, but consider threads.
1157 			 */
1158 			flt->upper_lock_type = RW_WRITER;
1159 		} else if ((flt->access_type & VM_PROT_WRITE) != 0) {
1160 			/*
1161 			 * assume we're about to COW.
1162 			 */
1163 			flt->upper_lock_type = RW_WRITER;
1164 		}
1165 		amap_lock(amap, flt->upper_lock_type);
1166 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1167 	} else {
1168 		if ((flt->access_type & VM_PROT_WRITE) != 0) {
1169 			/*
1170 			 * we are about to dirty the object and that
1171 			 * requires a write lock.
1172 			 */
1173 			flt->lower_lock_type = RW_WRITER;
1174 		}
1175 		*ranons = NULL;	/* to be safe */
1176 	}
1177 
1178 	/* locked: maps(read), amap(if there) */
1179 	KASSERT(amap == NULL ||
1180 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1181 
1182 	/*
1183 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1184 	 * now and then forget about them (for the rest of the fault).
1185 	 */
1186 
1187 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1188 
1189 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1190 		    0,0,0,0);
1191 		/* flush back-page anons? */
1192 		if (amap)
1193 			uvmfault_anonflush(*ranons, nback);
1194 
1195 		/*
1196 		 * flush object?  change lock type to RW_WRITER, to avoid
1197 		 * excessive competition between read/write locks if many
1198 		 * threads doing "sequential access".
1199 		 */
1200 		if (uobj) {
1201 			voff_t uoff;
1202 
1203 			flt->lower_lock_type = RW_WRITER;
1204 			uoff = ufi->entry->offset + eoff;
1205 			rw_enter(uobj->vmobjlock, RW_WRITER);
1206 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1207 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1208 		}
1209 
1210 		/* now forget about the backpages */
1211 		if (amap)
1212 			*ranons += nback;
1213 		flt->startva += (nback << PAGE_SHIFT);
1214 		flt->npages -= nback;
1215 		flt->centeridx = 0;
1216 	}
1217 	/*
1218 	 * => startva is fixed
1219 	 * => npages is fixed
1220 	 */
1221 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1222 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1223 	    flt->startva + (flt->npages << PAGE_SHIFT));
1224 	return 0;
1225 }
1226 
1227 /*
1228  * uvm_fault_upper_upgrade: upgrade upper lock, reader -> writer
1229  */
1230 
1231 static inline int
1232 uvm_fault_upper_upgrade(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1233     struct vm_amap *amap, struct uvm_object *uobj)
1234 {
1235 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1236 
1237 	KASSERT(amap != NULL);
1238 	KASSERT(flt->upper_lock_type == rw_lock_op(amap->am_lock));
1239 
1240 	/*
1241 	 * fast path.
1242 	 */
1243 
1244 	if (__predict_true(flt->upper_lock_type == RW_WRITER)) {
1245 		return 0;
1246 	}
1247 
1248 	/*
1249 	 * otherwise try for the upgrade.  if we don't get it, unlock
1250 	 * everything, restart the fault and next time around get a writer
1251 	 * lock.
1252 	 */
1253 
1254 	flt->upper_lock_type = RW_WRITER;
1255 	if (__predict_false(!rw_tryupgrade(amap->am_lock))) {
1256 		uvmfault_unlockall(ufi, amap, uobj);
1257 		cpu_count(CPU_COUNT_FLTNOUP, 1);
1258 		UVMHIST_LOG(maphist, "  !upgrade upper", 0, 0,0,0);
1259 		return ERESTART;
1260 	}
1261 	cpu_count(CPU_COUNT_FLTUP, 1);
1262 	KASSERT(flt->upper_lock_type == rw_lock_op(amap->am_lock));
1263 	return 0;
1264 }
1265 
1266 /*
1267  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1268  *
1269  * iterate range of interest:
1270  *	1. check if h/w mapping exists.  if yes, we don't care
1271  *	2. check if anon exists.  if not, page is lower.
1272  *	3. if anon exists, enter h/w mapping for neighbors.
1273  *
1274  * => called with amap locked (if exists).
1275  */
1276 
1277 static int
1278 uvm_fault_upper_lookup(
1279 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1280 	struct vm_anon **anons, struct vm_page **pages)
1281 {
1282 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1283 	int lcv;
1284 	vaddr_t currva;
1285 	bool shadowed __unused;
1286 	bool entered;
1287 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1288 
1289 	/* locked: maps(read), amap(if there) */
1290 	KASSERT(amap == NULL ||
1291 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1292 
1293 	/*
1294 	 * map in the backpages and frontpages we found in the amap in hopes
1295 	 * of preventing future faults.    we also init the pages[] array as
1296 	 * we go.
1297 	 */
1298 
1299 	currva = flt->startva;
1300 	shadowed = false;
1301 	entered = false;
1302 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1303 		/*
1304 		 * unmapped or center page.   check if any anon at this level.
1305 		 */
1306 		if (amap == NULL || anons[lcv] == NULL) {
1307 			pages[lcv] = NULL;
1308 			continue;
1309 		}
1310 
1311 		/*
1312 		 * check for present page and map if possible.
1313 		 */
1314 
1315 		pages[lcv] = PGO_DONTCARE;
1316 		if (lcv == flt->centeridx) {	/* save center for later! */
1317 			shadowed = true;
1318 			continue;
1319 		}
1320 
1321 		struct vm_anon *anon = anons[lcv];
1322 		struct vm_page *pg = anon->an_page;
1323 
1324 		KASSERT(anon->an_lock == amap->am_lock);
1325 
1326 		/*
1327 		 * ignore loaned and busy pages.
1328 		 * don't play with VAs that are already mapped.
1329 		 */
1330 
1331 		if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0 &&
1332 		    !pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1333 			uvm_fault_upper_neighbor(ufi, flt, currva,
1334 			    pg, anon->an_ref > 1);
1335 			entered = true;
1336 		}
1337 	}
1338 	if (entered) {
1339 		pmap_update(ufi->orig_map->pmap);
1340 	}
1341 
1342 	/* locked: maps(read), amap(if there) */
1343 	KASSERT(amap == NULL ||
1344 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1345 	/* (shadowed == true) if there is an anon at the faulting address */
1346 	UVMHIST_LOG(maphist, "  shadowed=%jd, will_get=%jd", shadowed,
1347 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1348 
1349 	return 0;
1350 }
1351 
1352 /*
1353  * uvm_fault_upper_neighbor: enter single upper neighbor page.
1354  *
1355  * => called with amap and anon locked.
1356  */
1357 
1358 static void
1359 uvm_fault_upper_neighbor(
1360 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1361 	vaddr_t currva, struct vm_page *pg, bool readonly)
1362 {
1363 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1364 
1365 	/* locked: amap, anon */
1366 
1367 	KASSERT(pg->uobject == NULL);
1368 	KASSERT(pg->uanon != NULL);
1369 	KASSERT(rw_lock_op(pg->uanon->an_lock) == flt->upper_lock_type);
1370 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1371 
1372 	/*
1373 	 * there wasn't a direct fault on the page, so avoid the cost of
1374 	 * activating it.
1375 	 */
1376 
1377 	if (!uvmpdpol_pageisqueued_p(pg) && pg->wire_count == 0) {
1378 		uvm_pagelock(pg);
1379 		uvm_pageenqueue(pg);
1380 		uvm_pageunlock(pg);
1381 	}
1382 
1383 	UVMHIST_LOG(maphist,
1384 	    "  MAPPING: n anon: pm=%#jx, va=%#jx, pg=%#jx",
1385 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
1386 	cpu_count(CPU_COUNT_FLTNAMAP, 1);
1387 
1388 	/*
1389 	 * Since this page isn't the page that's actually faulting,
1390 	 * ignore pmap_enter() failures; it's not critical that we
1391 	 * enter these right now.
1392 	 */
1393 
1394 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1395 	    VM_PAGE_TO_PHYS(pg),
1396 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1397 	    flt->enter_prot,
1398 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1399 }
1400 
1401 /*
1402  * uvm_fault_upper: handle upper fault.
1403  *
1404  *	1. acquire anon lock.
1405  *	2. get anon.  let uvmfault_anonget do the dirty work.
1406  *	3. handle loan.
1407  *	4. dispatch direct or promote handlers.
1408  */
1409 
1410 static int
1411 uvm_fault_upper(
1412 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1413 	struct vm_anon **anons)
1414 {
1415 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1416 	struct vm_anon * const anon = anons[flt->centeridx];
1417 	struct uvm_object *uobj;
1418 	int error;
1419 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1420 
1421 	/* locked: maps(read), amap, anon */
1422 	KASSERT(rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1423 	KASSERT(anon->an_lock == amap->am_lock);
1424 
1425 	/*
1426 	 * handle case 1: fault on an anon in our amap
1427 	 */
1428 
1429 	UVMHIST_LOG(maphist, "  case 1 fault: anon=%#jx",
1430 	    (uintptr_t)anon, 0, 0, 0);
1431 
1432 	/*
1433 	 * no matter if we have case 1A or case 1B we are going to need to
1434 	 * have the anon's memory resident.   ensure that now.
1435 	 */
1436 
1437 	/*
1438 	 * let uvmfault_anonget do the dirty work.
1439 	 * if it fails (!OK) it will unlock everything for us.
1440 	 * if it succeeds, locks are still valid and locked.
1441 	 * also, if it is OK, then the anon's page is on the queues.
1442 	 * if the page is on loan from a uvm_object, then anonget will
1443 	 * lock that object for us if it does not fail.
1444 	 */
1445  retry:
1446 	error = uvmfault_anonget(ufi, amap, anon);
1447 	switch (error) {
1448 	case 0:
1449 		break;
1450 
1451 	case ERESTART:
1452 		return ERESTART;
1453 
1454 	case EAGAIN:
1455 		kpause("fltagain1", false, hz/2, NULL);
1456 		return ERESTART;
1457 
1458 	case ENOLCK:
1459 		/* it needs a write lock: retry */
1460 		error = uvm_fault_upper_upgrade(ufi, flt, amap, NULL);
1461 		if (error != 0) {
1462 			return error;
1463 		}
1464 		KASSERT(rw_write_held(amap->am_lock));
1465 		goto retry;
1466 
1467 	default:
1468 		return error;
1469 	}
1470 
1471 	/*
1472 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1473 	 */
1474 
1475 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1476 
1477 	/* locked: maps(read), amap, anon, uobj(if one) */
1478 	KASSERT(rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1479 	KASSERT(anon->an_lock == amap->am_lock);
1480 	KASSERT(uobj == NULL ||
1481 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
1482 
1483 	/*
1484 	 * special handling for loaned pages
1485 	 */
1486 
1487 	if (anon->an_page->loan_count) {
1488 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1489 		if (error != 0)
1490 			return error;
1491 	}
1492 
1493 	/*
1494 	 * if we are case 1B then we will need to allocate a new blank
1495 	 * anon to transfer the data into.   note that we have a lock
1496 	 * on anon, so no one can busy or release the page until we are done.
1497 	 * also note that the ref count can't drop to zero here because
1498 	 * it is > 1 and we are only dropping one ref.
1499 	 *
1500 	 * in the (hopefully very rare) case that we are out of RAM we
1501 	 * will unlock, wait for more RAM, and refault.
1502 	 *
1503 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1504 	 */
1505 
1506 	if (flt->cow_now && anon->an_ref > 1) {
1507 		flt->promote = true;
1508 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1509 	} else {
1510 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1511 	}
1512 	return error;
1513 }
1514 
1515 /*
1516  * uvm_fault_upper_loan: handle loaned upper page.
1517  *
1518  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1519  *	2. if cow'ing now, and if ref count is 1, break loan.
1520  */
1521 
1522 static int
1523 uvm_fault_upper_loan(
1524 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1525 	struct vm_anon *anon, struct uvm_object **ruobj)
1526 {
1527 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1528 	int error = 0;
1529 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1530 
1531 	if (!flt->cow_now) {
1532 
1533 		/*
1534 		 * for read faults on loaned pages we just cap the
1535 		 * protection at read-only.
1536 		 */
1537 
1538 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1539 
1540 	} else {
1541 		/*
1542 		 * note that we can't allow writes into a loaned page!
1543 		 *
1544 		 * if we have a write fault on a loaned page in an
1545 		 * anon then we need to look at the anon's ref count.
1546 		 * if it is greater than one then we are going to do
1547 		 * a normal copy-on-write fault into a new anon (this
1548 		 * is not a problem).  however, if the reference count
1549 		 * is one (a case where we would normally allow a
1550 		 * write directly to the page) then we need to kill
1551 		 * the loan before we continue.
1552 		 */
1553 
1554 		/* >1 case is already ok */
1555 		if (anon->an_ref == 1) {
1556 			/* breaking loan requires a write lock. */
1557 			error = uvm_fault_upper_upgrade(ufi, flt, amap, NULL);
1558 			if (error != 0) {
1559 				return error;
1560 			}
1561 			KASSERT(rw_write_held(amap->am_lock));
1562 
1563 			error = uvm_loanbreak_anon(anon, *ruobj);
1564 			if (error != 0) {
1565 				uvmfault_unlockall(ufi, amap, *ruobj);
1566 				uvm_wait("flt_noram2");
1567 				return ERESTART;
1568 			}
1569 			/* if we were a loan receiver uobj is gone */
1570 			if (*ruobj)
1571 				*ruobj = NULL;
1572 		}
1573 	}
1574 	return error;
1575 }
1576 
1577 /*
1578  * uvm_fault_upper_promote: promote upper page.
1579  *
1580  *	1. call uvmfault_promote.
1581  *	2. enqueue page.
1582  *	3. deref.
1583  *	4. pass page to uvm_fault_upper_enter.
1584  */
1585 
1586 static int
1587 uvm_fault_upper_promote(
1588 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1589 	struct uvm_object *uobj, struct vm_anon *anon)
1590 {
1591 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1592 	struct vm_anon * const oanon = anon;
1593 	struct vm_page *pg;
1594 	int error;
1595 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1596 
1597 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1598 	cpu_count(CPU_COUNT_FLT_ACOW, 1);
1599 
1600 	/* promoting requires a write lock. */
1601 	error = uvm_fault_upper_upgrade(ufi, flt, amap, NULL);
1602 	if (error != 0) {
1603 		return error;
1604 	}
1605 	KASSERT(rw_write_held(amap->am_lock));
1606 
1607 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1608 	    &flt->anon_spare);
1609 	switch (error) {
1610 	case 0:
1611 		break;
1612 	case ERESTART:
1613 		return ERESTART;
1614 	default:
1615 		return error;
1616 	}
1617 	pg = anon->an_page;
1618 
1619 	KASSERT(anon->an_lock == oanon->an_lock);
1620 	KASSERT((pg->flags & (PG_BUSY | PG_FAKE)) == 0);
1621 
1622 	/* deref: can not drop to zero here by defn! */
1623 	KASSERT(oanon->an_ref > 1);
1624 	oanon->an_ref--;
1625 
1626 	/*
1627 	 * note: oanon is still locked, as is the new anon.  we
1628 	 * need to check for this later when we unlock oanon; if
1629 	 * oanon != anon, we'll have to unlock anon, too.
1630 	 */
1631 
1632 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1633 }
1634 
1635 /*
1636  * uvm_fault_upper_direct: handle direct fault.
1637  */
1638 
1639 static int
1640 uvm_fault_upper_direct(
1641 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1642 	struct uvm_object *uobj, struct vm_anon *anon)
1643 {
1644 	struct vm_anon * const oanon = anon;
1645 	struct vm_page *pg;
1646 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1647 
1648 	cpu_count(CPU_COUNT_FLT_ANON, 1);
1649 	pg = anon->an_page;
1650 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1651 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1652 
1653 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1654 }
1655 
1656 /*
1657  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1658  */
1659 
1660 static int
1661 uvm_fault_upper_enter(
1662 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1663 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1664 	struct vm_anon *oanon)
1665 {
1666 	struct pmap *pmap = ufi->orig_map->pmap;
1667 	vaddr_t va = ufi->orig_rvaddr;
1668 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1669 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1670 
1671 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1672 	KASSERT(rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1673 	KASSERT(anon->an_lock == amap->am_lock);
1674 	KASSERT(oanon->an_lock == amap->am_lock);
1675 	KASSERT(uobj == NULL ||
1676 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
1677 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1678 
1679 	/*
1680 	 * now map the page in.
1681 	 */
1682 
1683 	UVMHIST_LOG(maphist,
1684 	    "  MAPPING: anon: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
1685 	    (uintptr_t)pmap, va, (uintptr_t)pg, flt->promote);
1686 	if (pmap_enter(pmap, va, VM_PAGE_TO_PHYS(pg),
1687 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1688 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1689 
1690 		/*
1691 		 * If pmap_enter() fails, it must not leave behind an existing
1692 		 * pmap entry.  In particular, a now-stale entry for a different
1693 		 * page would leave the pmap inconsistent with the vm_map.
1694 		 * This is not to imply that pmap_enter() should remove an
1695 		 * existing mapping in such a situation (since that could create
1696 		 * different problems, eg. if the existing mapping is wired),
1697 		 * but rather that the pmap should be designed such that it
1698 		 * never needs to fail when the new mapping is replacing an
1699 		 * existing mapping and the new page has no existing mappings.
1700 		 *
1701 		 * XXX This can't be asserted safely any more because many
1702 		 * LWPs and/or many processes could simultaneously fault on
1703 		 * the same VA and some might succeed.
1704 		 */
1705 
1706 		/* KASSERT(!pmap_extract(pmap, va, NULL)); */
1707 
1708 		/*
1709 		 * ensure that the page is queued in the case that
1710 		 * we just promoted.
1711 		 */
1712 
1713 		uvm_pagelock(pg);
1714 		uvm_pageenqueue(pg);
1715 		uvm_pageunlock(pg);
1716 
1717 		/*
1718 		 * No need to undo what we did; we can simply think of
1719 		 * this as the pmap throwing away the mapping information.
1720 		 *
1721 		 * We do, however, have to go through the ReFault path,
1722 		 * as the map may change while we're asleep.
1723 		 */
1724 
1725 		uvmfault_unlockall(ufi, amap, uobj);
1726 		if (!uvm_reclaimable()) {
1727 			UVMHIST_LOG(maphist,
1728 			    "<- failed.  out of VM",0,0,0,0);
1729 			/* XXX instrumentation */
1730 			return ENOMEM;
1731 		}
1732 		/* XXX instrumentation */
1733 		uvm_wait("flt_pmfail1");
1734 		return ERESTART;
1735 	}
1736 
1737 	uvm_fault_upper_done(ufi, flt, anon, pg);
1738 
1739 	/*
1740 	 * done case 1!  finish up by unlocking everything and returning success
1741 	 */
1742 
1743 	pmap_update(pmap);
1744 	uvmfault_unlockall(ufi, amap, uobj);
1745 	return 0;
1746 }
1747 
1748 /*
1749  * uvm_fault_upper_done: queue upper center page.
1750  */
1751 
1752 static void
1753 uvm_fault_upper_done(
1754 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1755 	struct vm_anon *anon, struct vm_page *pg)
1756 {
1757 	const bool wire_paging = flt->wire_paging;
1758 
1759 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1760 
1761 	/*
1762 	 * ... update the page queues.
1763 	 */
1764 
1765 	if (wire_paging) {
1766 		uvm_pagelock(pg);
1767 		uvm_pagewire(pg);
1768 		uvm_pageunlock(pg);
1769 
1770 		/*
1771 		 * since the now-wired page cannot be paged out,
1772 		 * release its swap resources for others to use.
1773 		 * and since an anon with no swap cannot be clean,
1774 		 * mark it dirty now.
1775 		 */
1776 
1777 		uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
1778 		uvm_anon_dropswap(anon);
1779 	} else if (uvmpdpol_pageactivate_p(pg)) {
1780 		/*
1781 		 * avoid re-activating the page unless needed,
1782 		 * to avoid false sharing on multiprocessor.
1783 		 */
1784 
1785 		uvm_pagelock(pg);
1786 		uvm_pageactivate(pg);
1787 		uvm_pageunlock(pg);
1788 	}
1789 }
1790 
1791 /*
1792  * uvm_fault_lower_upgrade: upgrade lower lock, reader -> writer
1793  */
1794 
1795 static inline int
1796 uvm_fault_lower_upgrade(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1797     struct vm_amap *amap, struct uvm_object *uobj, struct vm_page *uobjpage)
1798 {
1799 
1800 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1801 
1802 	KASSERT(uobj != NULL);
1803 	KASSERT(flt->lower_lock_type == rw_lock_op(uobj->vmobjlock));
1804 
1805 	/*
1806 	 * fast path.
1807 	 */
1808 
1809 	if (__predict_true(flt->lower_lock_type == RW_WRITER)) {
1810 		return 0;
1811 	}
1812 
1813 	/*
1814 	 * otherwise try for the upgrade.  if we don't get it, unlock
1815 	 * everything, restart the fault and next time around get a writer
1816 	 * lock.
1817 	 */
1818 
1819 	flt->lower_lock_type = RW_WRITER;
1820 	if (__predict_false(!rw_tryupgrade(uobj->vmobjlock))) {
1821 		uvmfault_unlockall(ufi, amap, uobj);
1822 		cpu_count(CPU_COUNT_FLTNOUP, 1);
1823 		UVMHIST_LOG(maphist, "  !upgrade lower", 0, 0,0,0);
1824 		return ERESTART;
1825 	}
1826 	cpu_count(CPU_COUNT_FLTUP, 1);
1827 	KASSERT(flt->lower_lock_type == rw_lock_op(uobj->vmobjlock));
1828 	return 0;
1829 }
1830 
1831 /*
1832  * uvm_fault_lower: handle lower fault.
1833  *
1834  *	1. check uobj
1835  *	1.1. if null, ZFOD.
1836  *	1.2. if not null, look up unnmapped neighbor pages.
1837  *	2. for center page, check if promote.
1838  *	2.1. ZFOD always needs promotion.
1839  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1840  *	3. if uobj is not ZFOD and page is not found, do i/o.
1841  *	4. dispatch either direct / promote fault.
1842  */
1843 
1844 static int
1845 uvm_fault_lower(
1846 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1847 	struct vm_page **pages)
1848 {
1849 	struct vm_amap *amap __diagused = ufi->entry->aref.ar_amap;
1850 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1851 	struct vm_page *uobjpage;
1852 	int error;
1853 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1854 
1855 	/*
1856 	 * now, if the desired page is not shadowed by the amap and we have
1857 	 * a backing object that does not have a special fault routine, then
1858 	 * we ask (with pgo_get) the object for resident pages that we care
1859 	 * about and attempt to map them in.  we do not let pgo_get block
1860 	 * (PGO_LOCKED).
1861 	 */
1862 
1863 	if (uobj == NULL) {
1864 		/* zero fill; don't care neighbor pages */
1865 		uobjpage = NULL;
1866 	} else {
1867 		uvm_fault_lower_lookup(ufi, flt, pages);
1868 		uobjpage = pages[flt->centeridx];
1869 	}
1870 
1871 	/*
1872 	 * note that at this point we are done with any front or back pages.
1873 	 * we are now going to focus on the center page (i.e. the one we've
1874 	 * faulted on).  if we have faulted on the upper (anon) layer
1875 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1876 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1877 	 * layer [i.e. case 2] and the page was both present and available,
1878 	 * then we've got a pointer to it as "uobjpage" and we've already
1879 	 * made it BUSY.
1880 	 */
1881 
1882 	/*
1883 	 * locked:
1884 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1885 	 */
1886 	KASSERT(amap == NULL ||
1887 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1888 	KASSERT(uobj == NULL ||
1889 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
1890 
1891 	/*
1892 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1893 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1894 	 * have a backing object, check and see if we are going to promote
1895 	 * the data up to an anon during the fault.
1896 	 */
1897 
1898 	if (uobj == NULL) {
1899 		uobjpage = PGO_DONTCARE;
1900 		flt->promote = true;		/* always need anon here */
1901 	} else {
1902 		KASSERT(uobjpage != PGO_DONTCARE);
1903 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1904 	}
1905 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%jd, zfill=%jd",
1906 	    flt->promote, (uobj == NULL), 0,0);
1907 
1908 	/*
1909 	 * if uobjpage is not null then we do not need to do I/O to get the
1910 	 * uobjpage.
1911 	 *
1912 	 * if uobjpage is null, then we need to unlock and ask the pager to
1913 	 * get the data for us.   once we have the data, we need to reverify
1914 	 * the state the world.   we are currently not holding any resources.
1915 	 */
1916 
1917 	if (uobjpage) {
1918 		/* update rusage counters */
1919 		curlwp->l_ru.ru_minflt++;
1920 	} else {
1921 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1922 		if (error != 0)
1923 			return error;
1924 	}
1925 
1926 	/*
1927 	 * locked:
1928 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1929 	 */
1930 	KASSERT(amap == NULL ||
1931 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1932 	KASSERT(uobj == NULL ||
1933 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
1934 
1935 	/*
1936 	 * notes:
1937 	 *  - at this point uobjpage can not be NULL
1938 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1939 	 *  for it above)
1940 	 *  - at this point uobjpage could be waited on (handle later)
1941 	 *  - uobjpage can be from a different object if tmpfs (vnode vs UAO)
1942 	 */
1943 
1944 	KASSERT(uobjpage != NULL);
1945 	KASSERT(uobj == NULL ||
1946 	    uobjpage->uobject->vmobjlock == uobj->vmobjlock);
1947 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1948 	    uvm_pagegetdirty(uobjpage) == UVM_PAGE_STATUS_CLEAN);
1949 
1950 	if (!flt->promote) {
1951 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1952 	} else {
1953 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1954 	}
1955 	return error;
1956 }
1957 
1958 /*
1959  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1960  *
1961  *	1. get on-memory pages.
1962  *	2. if failed, give up (get only center page later).
1963  *	3. if succeeded, enter h/w mapping of neighbor pages.
1964  */
1965 
1966 static void
1967 uvm_fault_lower_lookup(
1968 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1969 	struct vm_page **pages)
1970 {
1971 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1972 	int lcv, gotpages;
1973 	vaddr_t currva;
1974 	bool entered;
1975 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1976 
1977 	rw_enter(uobj->vmobjlock, flt->lower_lock_type);
1978 
1979 	/*
1980 	 * Locked: maps(read), amap(if there), uobj
1981 	 */
1982 
1983 	cpu_count(CPU_COUNT_FLTLGET, 1);
1984 	gotpages = flt->npages;
1985 	(void) uobj->pgops->pgo_get(uobj,
1986 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1987 	    pages, &gotpages, flt->centeridx,
1988 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1989 	    PGO_LOCKED);
1990 
1991 	KASSERT(rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
1992 
1993 	/*
1994 	 * check for pages to map, if we got any
1995 	 */
1996 
1997 	if (gotpages == 0) {
1998 		pages[flt->centeridx] = NULL;
1999 		return;
2000 	}
2001 
2002 	entered = false;
2003 	currva = flt->startva;
2004 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
2005 		struct vm_page *curpg;
2006 
2007 		curpg = pages[lcv];
2008 		if (curpg == NULL || curpg == PGO_DONTCARE) {
2009 			continue;
2010 		}
2011 
2012 		/*
2013 		 * in the case of tmpfs, the pages might be from a different
2014 		 * uvm_object.  just make sure that they have the same lock.
2015 		 */
2016 
2017 		KASSERT(curpg->uobject->vmobjlock == uobj->vmobjlock);
2018 		KASSERT((curpg->flags & PG_BUSY) == 0);
2019 
2020 		/*
2021 		 * leave the centre page for later.  don't screw with
2022 		 * existing mappings (needless & expensive).
2023 		 */
2024 
2025 		if (lcv == flt->centeridx) {
2026 			UVMHIST_LOG(maphist, "  got uobjpage (%#jx) "
2027 			    "with locked get", (uintptr_t)curpg, 0, 0, 0);
2028 		} else if (!pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
2029 			uvm_fault_lower_neighbor(ufi, flt, currva, curpg);
2030 			entered = true;
2031 		}
2032 	}
2033 	if (entered) {
2034 		pmap_update(ufi->orig_map->pmap);
2035 	}
2036 }
2037 
2038 /*
2039  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
2040  */
2041 
2042 static void
2043 uvm_fault_lower_neighbor(
2044 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2045 	vaddr_t currva, struct vm_page *pg)
2046 {
2047 	const bool readonly = uvm_pagereadonly_p(pg) || pg->loan_count > 0;
2048 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2049 
2050 	/* locked: maps(read), amap(if there), uobj */
2051 
2052 	/*
2053 	 * calling pgo_get with PGO_LOCKED returns us pages which
2054 	 * are neither busy nor released, so we don't need to check
2055 	 * for this.  we can just directly enter the pages.
2056 	 *
2057 	 * there wasn't a direct fault on the page, so avoid the cost of
2058 	 * activating it.
2059 	 */
2060 
2061 	if (!uvmpdpol_pageisqueued_p(pg) && pg->wire_count == 0) {
2062 		uvm_pagelock(pg);
2063 		uvm_pageenqueue(pg);
2064 		uvm_pageunlock(pg);
2065 	}
2066 
2067 	UVMHIST_LOG(maphist,
2068 	    "  MAPPING: n obj: pm=%#jx, va=%#jx, pg=%#jx",
2069 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
2070 	cpu_count(CPU_COUNT_FLTNOMAP, 1);
2071 
2072 	/*
2073 	 * Since this page isn't the page that's actually faulting,
2074 	 * ignore pmap_enter() failures; it's not critical that we
2075 	 * enter these right now.
2076 	 * NOTE: page can't be waited on or PG_RELEASED because we've
2077 	 * held the lock the whole time we've had the handle.
2078 	 */
2079 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
2080 	KASSERT((pg->flags & PG_RELEASED) == 0);
2081 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) ||
2082 	    uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_CLEAN);
2083 	KASSERT((pg->flags & PG_BUSY) == 0);
2084 	KASSERT(rw_lock_op(pg->uobject->vmobjlock) == flt->lower_lock_type);
2085 
2086 	const vm_prot_t mapprot =
2087 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
2088 	    flt->enter_prot & MASK(ufi->entry);
2089 	const u_int mapflags =
2090 	    PMAP_CANFAIL | (flt->wire_mapping ? (mapprot | PMAP_WIRED) : 0);
2091 	(void) pmap_enter(ufi->orig_map->pmap, currva,
2092 	    VM_PAGE_TO_PHYS(pg), mapprot, mapflags);
2093 }
2094 
2095 /*
2096  * uvm_fault_lower_io: get lower page from backing store.
2097  *
2098  *	1. unlock everything, because i/o will block.
2099  *	2. call pgo_get.
2100  *	3. if failed, recover.
2101  *	4. if succeeded, relock everything and verify things.
2102  */
2103 
2104 static int
2105 uvm_fault_lower_io(
2106 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2107 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
2108 {
2109 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2110 	struct uvm_object *uobj = *ruobj;
2111 	struct vm_page *pg;
2112 	bool locked;
2113 	int gotpages;
2114 	int error;
2115 	voff_t uoff;
2116 	vm_prot_t access_type;
2117 	int advice;
2118 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2119 
2120 	/* update rusage counters */
2121 	curlwp->l_ru.ru_majflt++;
2122 
2123 	/* grab everything we need from the entry before we unlock */
2124 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
2125 	access_type = flt->access_type & MASK(ufi->entry);
2126 	advice = ufi->entry->advice;
2127 
2128 	/* Locked: maps(read), amap(if there), uobj */
2129 	KASSERT(rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
2130 
2131 	/* Upgrade to a write lock if needed. */
2132 	error = uvm_fault_lower_upgrade(ufi, flt, amap, uobj, NULL);
2133 	if (error != 0) {
2134 		return error;
2135 	}
2136 	uvmfault_unlockall(ufi, amap, NULL);
2137 
2138 	/* Locked: uobj(write) */
2139 	KASSERT(rw_write_held(uobj->vmobjlock));
2140 
2141 	cpu_count(CPU_COUNT_FLTGET, 1);
2142 	gotpages = 1;
2143 	pg = NULL;
2144 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
2145 	    0, access_type, advice, PGO_SYNCIO);
2146 	/* locked: pg(if no error) */
2147 
2148 	/*
2149 	 * recover from I/O
2150 	 */
2151 
2152 	if (error) {
2153 		if (error == EAGAIN) {
2154 			UVMHIST_LOG(maphist,
2155 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
2156 			kpause("fltagain2", false, hz/2, NULL);
2157 			return ERESTART;
2158 		}
2159 
2160 #if 0
2161 		KASSERT(error != ERESTART);
2162 #else
2163 		/* XXXUEBS don't re-fault? */
2164 		if (error == ERESTART)
2165 			error = EIO;
2166 #endif
2167 
2168 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %jd)",
2169 		    error, 0,0,0);
2170 		return error;
2171 	}
2172 
2173 	/*
2174 	 * re-verify the state of the world by first trying to relock
2175 	 * the maps.  always relock the object.
2176 	 */
2177 
2178 	locked = uvmfault_relock(ufi);
2179 	if (locked && amap)
2180 		amap_lock(amap, flt->upper_lock_type);
2181 
2182 	/* might be changed */
2183 	uobj = pg->uobject;
2184 
2185 	rw_enter(uobj->vmobjlock, flt->lower_lock_type);
2186 	KASSERT((pg->flags & PG_BUSY) != 0);
2187 	KASSERT(flt->lower_lock_type == RW_WRITER);
2188 
2189 	uvm_pagelock(pg);
2190 	uvm_pageactivate(pg);
2191 	uvm_pageunlock(pg);
2192 
2193 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
2194 	/* locked(!locked): uobj, pg */
2195 
2196 	/*
2197 	 * verify that the page has not be released and re-verify
2198 	 * that amap slot is still free.   if there is a problem,
2199 	 * we unlock and clean up.
2200 	 */
2201 
2202 	if ((pg->flags & PG_RELEASED) != 0 ||
2203 	    (locked && amap && amap_lookup(&ufi->entry->aref,
2204 	      ufi->orig_rvaddr - ufi->entry->start))) {
2205 		if (locked)
2206 			uvmfault_unlockall(ufi, amap, NULL);
2207 		locked = false;
2208 	}
2209 
2210 	/*
2211 	 * unbusy/release the page.
2212 	 */
2213 
2214 	if ((pg->flags & PG_RELEASED) == 0) {
2215 		pg->flags &= ~PG_BUSY;
2216 		uvm_pagelock(pg);
2217 		uvm_pagewakeup(pg);
2218 		uvm_pageunlock(pg);
2219 		UVM_PAGE_OWN(pg, NULL);
2220 	} else {
2221 		cpu_count(CPU_COUNT_FLTPGRELE, 1);
2222 		uvm_pagefree(pg);
2223 	}
2224 
2225 	/*
2226 	 * didn't get the lock?   retry.
2227 	 */
2228 
2229 	if (locked == false) {
2230 		UVMHIST_LOG(maphist,
2231 		    "  wasn't able to relock after fault: retry",
2232 		    0,0,0,0);
2233 		rw_exit(uobj->vmobjlock);
2234 		return ERESTART;
2235 	}
2236 
2237 	/*
2238 	 * we have the data in pg.  we are holding object lock (so the page
2239 	 * can't be released on us).
2240 	 */
2241 
2242 	/* locked: maps(read), amap(if !null), uobj */
2243 
2244 	*ruobj = uobj;
2245 	*ruobjpage = pg;
2246 	return 0;
2247 }
2248 
2249 /*
2250  * uvm_fault_lower_direct: fault lower center page
2251  *
2252  *	1. adjust flt->enter_prot.
2253  *	2. if page is loaned, resolve.
2254  */
2255 
2256 int
2257 uvm_fault_lower_direct(
2258 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2259 	struct uvm_object *uobj, struct vm_page *uobjpage)
2260 {
2261 	struct vm_page *pg;
2262 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2263 
2264 	/*
2265 	 * we are not promoting.   if the mapping is COW ensure that we
2266 	 * don't give more access than we should (e.g. when doing a read
2267 	 * fault on a COPYONWRITE mapping we want to map the COW page in
2268 	 * R/O even though the entry protection could be R/W).
2269 	 *
2270 	 * set "pg" to the page we want to map in (uobjpage, usually)
2271 	 */
2272 
2273 	cpu_count(CPU_COUNT_FLT_OBJ, 1);
2274 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
2275 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
2276 		flt->enter_prot &= ~VM_PROT_WRITE;
2277 	pg = uobjpage;		/* map in the actual object */
2278 
2279 	KASSERT(uobjpage != PGO_DONTCARE);
2280 
2281 	/*
2282 	 * we are faulting directly on the page.   be careful
2283 	 * about writing to loaned pages...
2284 	 */
2285 
2286 	if (uobjpage->loan_count) {
2287 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
2288 	}
2289 	KASSERT(pg == uobjpage);
2290 	KASSERT((pg->flags & PG_BUSY) == 0);
2291 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
2292 }
2293 
2294 /*
2295  * uvm_fault_lower_direct_loan: resolve loaned page.
2296  *
2297  *	1. if not cow'ing, adjust flt->enter_prot.
2298  *	2. if cow'ing, break loan.
2299  */
2300 
2301 static int
2302 uvm_fault_lower_direct_loan(
2303 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2304 	struct uvm_object *uobj, struct vm_page **rpg,
2305 	struct vm_page **ruobjpage)
2306 {
2307 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2308 	struct vm_page *pg;
2309 	struct vm_page *uobjpage = *ruobjpage;
2310 	int error;
2311 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2312 
2313 	if (!flt->cow_now) {
2314 		/* read fault: cap the protection at readonly */
2315 		/* cap! */
2316 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2317 	} else {
2318 		/*
2319 		 * write fault: must break the loan here.  to do this
2320 		 * we need a write lock on the object.
2321 		 */
2322 
2323 		error = uvm_fault_lower_upgrade(ufi, flt, amap, uobj, uobjpage);
2324 		if (error != 0) {
2325 			return error;
2326 		}
2327 		KASSERT(rw_write_held(uobj->vmobjlock));
2328 
2329 		pg = uvm_loanbreak(uobjpage);
2330 		if (pg == NULL) {
2331 
2332 			uvmfault_unlockall(ufi, amap, uobj);
2333 			UVMHIST_LOG(maphist,
2334 			  "  out of RAM breaking loan, waiting",
2335 			  0,0,0,0);
2336 			cpu_count(CPU_COUNT_FLTNORAM, 1);
2337 			uvm_wait("flt_noram4");
2338 			return ERESTART;
2339 		}
2340 		*rpg = pg;
2341 		*ruobjpage = pg;
2342 
2343 		/*
2344 		 * drop ownership of page while still holding object lock,
2345 		 * which won't be dropped until the page is entered.
2346 		 */
2347 
2348 		uvm_pagelock(pg);
2349 		uvm_pagewakeup(pg);
2350 		uvm_pageunlock(pg);
2351 		pg->flags &= ~PG_BUSY;
2352 		UVM_PAGE_OWN(pg, NULL);
2353 	}
2354 	return 0;
2355 }
2356 
2357 /*
2358  * uvm_fault_lower_promote: promote lower page.
2359  *
2360  *	1. call uvmfault_promote.
2361  *	2. fill in data.
2362  *	3. if not ZFOD, dispose old page.
2363  */
2364 
2365 int
2366 uvm_fault_lower_promote(
2367 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2368 	struct uvm_object *uobj, struct vm_page *uobjpage)
2369 {
2370 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2371 	struct vm_anon *anon;
2372 	struct vm_page *pg;
2373 	int error;
2374 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2375 
2376 	KASSERT(amap != NULL);
2377 
2378 	/* promoting requires a write lock. */
2379 	error = uvm_fault_upper_upgrade(ufi, flt, amap, uobj);
2380 	if (error != 0) {
2381 		return error;
2382 	}
2383 	KASSERT(rw_write_held(amap->am_lock));
2384 	KASSERT(uobj == NULL ||
2385 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
2386 
2387 	/*
2388 	 * If we are going to promote the data to an anon we
2389 	 * allocate a blank anon here and plug it into our amap.
2390 	 */
2391 	error = uvmfault_promote(ufi, NULL, uobjpage, &anon, &flt->anon_spare);
2392 	switch (error) {
2393 	case 0:
2394 		break;
2395 	case ERESTART:
2396 		return ERESTART;
2397 	default:
2398 		return error;
2399 	}
2400 
2401 	pg = anon->an_page;
2402 
2403 	/*
2404 	 * Fill in the data.
2405 	 */
2406 
2407 	if (uobjpage != PGO_DONTCARE) {
2408 		cpu_count(CPU_COUNT_FLT_PRCOPY, 1);
2409 
2410 		/*
2411 		 * promote to shared amap?  make sure all sharing
2412 		 * procs see it
2413 		 */
2414 
2415 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2416 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2417 			/*
2418 			 * XXX: PAGE MIGHT BE WIRED!
2419 			 */
2420 		}
2421 
2422 		UVMHIST_LOG(maphist,
2423 		    "  promote uobjpage %#jx to anon/page %#jx/%#jx",
2424 		    (uintptr_t)uobjpage, (uintptr_t)anon, (uintptr_t)pg, 0);
2425 
2426 	} else {
2427 		cpu_count(CPU_COUNT_FLT_PRZERO, 1);
2428 
2429 		/*
2430 		 * Page is zero'd and marked dirty by
2431 		 * uvmfault_promote().
2432 		 */
2433 
2434 		UVMHIST_LOG(maphist,"  zero fill anon/page %#jx/%#jx",
2435 		    (uintptr_t)anon, (uintptr_t)pg, 0, 0);
2436 	}
2437 
2438 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2439 }
2440 
2441 /*
2442  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2443  * from the lower page.
2444  */
2445 
2446 int
2447 uvm_fault_lower_enter(
2448 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2449 	struct uvm_object *uobj,
2450 	struct vm_anon *anon, struct vm_page *pg)
2451 {
2452 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2453 	const bool readonly = uvm_pagereadonly_p(pg);
2454 	int error;
2455 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2456 
2457 	/*
2458 	 * Locked:
2459 	 *
2460 	 *	maps(read), amap(if !null), uobj(if !null),
2461 	 *	anon(if !null), pg(if anon), unlock_uobj(if !null)
2462 	 *
2463 	 * anon must be write locked (promotion).  uobj can be either.
2464 	 *
2465 	 * Note: pg is either the uobjpage or the new page in the new anon.
2466 	 */
2467 
2468 	KASSERT(amap == NULL ||
2469 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
2470 	KASSERT(uobj == NULL ||
2471 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
2472 	KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
2473 
2474 	/*
2475 	 * note that pg can't be PG_RELEASED or PG_BUSY since we did
2476 	 * not drop the object lock since the last time we checked.
2477 	 */
2478 
2479 	KASSERT((pg->flags & PG_RELEASED) == 0);
2480 	KASSERT((pg->flags & PG_BUSY) == 0);
2481 
2482 	/*
2483 	 * all resources are present.   we can now map it in and free our
2484 	 * resources.
2485 	 */
2486 
2487 	UVMHIST_LOG(maphist,
2488 	    "  MAPPING: case2: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
2489 	    (uintptr_t)ufi->orig_map->pmap, ufi->orig_rvaddr,
2490 	    (uintptr_t)pg, flt->promote);
2491 	KASSERTMSG((flt->access_type & VM_PROT_WRITE) == 0 || !readonly,
2492 	    "promote=%u cow_now=%u access_type=%x enter_prot=%x cow=%u "
2493 	    "entry=%p map=%p orig_rvaddr=%p pg=%p",
2494 	    flt->promote, flt->cow_now, flt->access_type, flt->enter_prot,
2495 	    UVM_ET_ISCOPYONWRITE(ufi->entry), ufi->entry, ufi->orig_map,
2496 	    (void *)ufi->orig_rvaddr, pg);
2497 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 || !readonly);
2498 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2499 	    VM_PAGE_TO_PHYS(pg),
2500 	    readonly ? flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2501 	    flt->access_type | PMAP_CANFAIL |
2502 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2503 
2504 		/*
2505 		 * No need to undo what we did; we can simply think of
2506 		 * this as the pmap throwing away the mapping information.
2507 		 *
2508 		 * We do, however, have to go through the ReFault path,
2509 		 * as the map may change while we're asleep.
2510 		 */
2511 
2512 		/*
2513 		 * ensure that the page is queued in the case that
2514 		 * we just promoted the page.
2515 		 */
2516 
2517 		if (anon != NULL) {
2518 			uvm_pagelock(pg);
2519 			uvm_pageenqueue(pg);
2520 			uvm_pagewakeup(pg);
2521 			uvm_pageunlock(pg);
2522 		}
2523 
2524 		uvmfault_unlockall(ufi, amap, uobj);
2525 		if (!uvm_reclaimable()) {
2526 			UVMHIST_LOG(maphist,
2527 			    "<- failed.  out of VM",0,0,0,0);
2528 			/* XXX instrumentation */
2529 			error = ENOMEM;
2530 			return error;
2531 		}
2532 		/* XXX instrumentation */
2533 		uvm_wait("flt_pmfail2");
2534 		return ERESTART;
2535 	}
2536 
2537 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2538 	pmap_update(ufi->orig_map->pmap);
2539 	uvmfault_unlockall(ufi, amap, uobj);
2540 
2541 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2542 	return 0;
2543 }
2544 
2545 /*
2546  * uvm_fault_lower_done: queue lower center page.
2547  */
2548 
2549 void
2550 uvm_fault_lower_done(
2551 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2552 	struct uvm_object *uobj, struct vm_page *pg)
2553 {
2554 
2555 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2556 
2557 	if (flt->wire_paging) {
2558 		uvm_pagelock(pg);
2559 		uvm_pagewire(pg);
2560 		uvm_pageunlock(pg);
2561 		if (pg->flags & PG_AOBJ) {
2562 
2563 			/*
2564 			 * since the now-wired page cannot be paged out,
2565 			 * release its swap resources for others to use.
2566 			 * since an aobj page with no swap cannot be clean,
2567 			 * mark it dirty now.
2568 			 *
2569 			 * use pg->uobject here.  if the page is from a
2570 			 * tmpfs vnode, the pages are backed by its UAO and
2571 			 * not the vnode.
2572 			 */
2573 
2574 			KASSERT(uobj != NULL);
2575 			KASSERT(uobj->vmobjlock == pg->uobject->vmobjlock);
2576 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
2577 			uao_dropswap(pg->uobject, pg->offset >> PAGE_SHIFT);
2578 		}
2579 	} else if (uvmpdpol_pageactivate_p(pg)) {
2580 		/*
2581 		 * avoid re-activating the page unless needed,
2582 		 * to avoid false sharing on multiprocessor.
2583 		 */
2584 
2585 		uvm_pagelock(pg);
2586 		uvm_pageactivate(pg);
2587 		uvm_pageunlock(pg);
2588 	}
2589 }
2590 
2591 
2592 /*
2593  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2594  *
2595  * => map may be read-locked by caller, but MUST NOT be write-locked.
2596  * => if map is read-locked, any operations which may cause map to
2597  *	be write-locked in uvm_fault() must be taken care of by
2598  *	the caller.  See uvm_map_pageable().
2599  */
2600 
2601 int
2602 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2603     vm_prot_t access_type, int maxprot)
2604 {
2605 	vaddr_t va;
2606 	int error;
2607 
2608 	/*
2609 	 * now fault it in a page at a time.   if the fault fails then we have
2610 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2611 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2612 	 */
2613 
2614 	/*
2615 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2616 	 * wiring the last page in the address space, though.
2617 	 */
2618 	if (start > end) {
2619 		return EFAULT;
2620 	}
2621 
2622 	for (va = start; va < end; va += PAGE_SIZE) {
2623 		error = uvm_fault_internal(map, va, access_type,
2624 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2625 		if (error) {
2626 			if (va != start) {
2627 				uvm_fault_unwire(map, start, va);
2628 			}
2629 			return error;
2630 		}
2631 	}
2632 	return 0;
2633 }
2634 
2635 /*
2636  * uvm_fault_unwire(): unwire range of virtual space.
2637  */
2638 
2639 void
2640 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2641 {
2642 	vm_map_lock_read(map);
2643 	uvm_fault_unwire_locked(map, start, end);
2644 	vm_map_unlock_read(map);
2645 }
2646 
2647 /*
2648  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2649  *
2650  * => map must be at least read-locked.
2651  */
2652 
2653 void
2654 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2655 {
2656 	struct vm_map_entry *entry, *oentry;
2657 	pmap_t pmap = vm_map_pmap(map);
2658 	vaddr_t va;
2659 	paddr_t pa;
2660 	struct vm_page *pg;
2661 
2662 	/*
2663 	 * we assume that the area we are unwiring has actually been wired
2664 	 * in the first place.   this means that we should be able to extract
2665 	 * the PAs from the pmap.   we also lock out the page daemon so that
2666 	 * we can call uvm_pageunwire.
2667 	 */
2668 
2669 	/*
2670 	 * find the beginning map entry for the region.
2671 	 */
2672 
2673 	KASSERT(start >= vm_map_min(map));
2674 	KASSERT(end <= vm_map_max(map));
2675 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2676 		panic("uvm_fault_unwire_locked: address not in map");
2677 
2678 	oentry = NULL;
2679 	for (va = start; va < end; va += PAGE_SIZE) {
2680 
2681 		/*
2682 		 * find the map entry for the current address.
2683 		 */
2684 
2685 		KASSERT(va >= entry->start);
2686 		while (va >= entry->end) {
2687 			KASSERT(entry->next != &map->header);
2688 			KASSERT(entry->next->start <= entry->end);
2689 			entry = entry->next;
2690 		}
2691 
2692 		/*
2693 		 * lock it.
2694 		 */
2695 
2696 		if (entry != oentry) {
2697 			if (oentry != NULL) {
2698 				uvm_map_unlock_entry(oentry);
2699 			}
2700 			uvm_map_lock_entry(entry, RW_WRITER);
2701 			oentry = entry;
2702 		}
2703 
2704 		/*
2705 		 * if the entry is no longer wired, tell the pmap.
2706 		 */
2707 
2708 		if (!pmap_extract(pmap, va, &pa))
2709 			continue;
2710 
2711 		if (VM_MAPENT_ISWIRED(entry) == 0)
2712 			pmap_unwire(pmap, va);
2713 
2714 		pg = PHYS_TO_VM_PAGE(pa);
2715 		if (pg) {
2716 			uvm_pagelock(pg);
2717 			uvm_pageunwire(pg);
2718 			uvm_pageunlock(pg);
2719 		}
2720 	}
2721 
2722 	if (oentry != NULL) {
2723 		uvm_map_unlock_entry(entry);
2724 	}
2725 }
2726