xref: /netbsd-src/sys/uvm/uvm_fault.c (revision e6c7e151de239c49d2e38720a061ed9d1fa99309)
1 /*	$NetBSD: uvm_fault.c,v 1.224 2020/03/23 10:35:56 skrll Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26  *
27  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
28  */
29 
30 /*
31  * uvm_fault.c: fault handler
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.224 2020/03/23 10:35:56 skrll 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 
47 /*
48  *
49  * a word on page faults:
50  *
51  * types of page faults we handle:
52  *
53  * CASE 1: upper layer faults                   CASE 2: lower layer faults
54  *
55  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
56  *    read/write1     write>1                  read/write   +-cow_write/zero
57  *         |             |                         |        |
58  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
59  * amap |  V  |       |  ---------> new |          |        | |  ^  |
60  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
61  *                                                 |        |    |
62  *      +-----+       +-----+                   +--|--+     | +--|--+
63  * uobj | d/c |       | d/c |                   |  V  |     +----+  |
64  *      +-----+       +-----+                   +-----+       +-----+
65  *
66  * d/c = don't care
67  *
68  *   case [0]: layerless fault
69  *	no amap or uobj is present.   this is an error.
70  *
71  *   case [1]: upper layer fault [anon active]
72  *     1A: [read] or [write with anon->an_ref == 1]
73  *		I/O takes place in upper level anon and uobj is not touched.
74  *     1B: [write with anon->an_ref > 1]
75  *		new anon is alloc'd and data is copied off ["COW"]
76  *
77  *   case [2]: lower layer fault [uobj]
78  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
79  *		I/O takes place directly in object.
80  *     2B: [write to copy_on_write] or [read on NULL uobj]
81  *		data is "promoted" from uobj to a new anon.
82  *		if uobj is null, then we zero fill.
83  *
84  * we follow the standard UVM locking protocol ordering:
85  *
86  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
87  * we hold a PG_BUSY page if we unlock for I/O
88  *
89  *
90  * the code is structured as follows:
91  *
92  *     - init the "IN" params in the ufi structure
93  *   ReFault: (ERESTART returned to the loop in uvm_fault_internal)
94  *     - do lookups [locks maps], check protection, handle needs_copy
95  *     - check for case 0 fault (error)
96  *     - establish "range" of fault
97  *     - if we have an amap lock it and extract the anons
98  *     - if sequential advice deactivate pages behind us
99  *     - at the same time check pmap for unmapped areas and anon for pages
100  *	 that we could map in (and do map it if found)
101  *     - check object for resident pages that we could map in
102  *     - if (case 2) goto Case2
103  *     - >>> handle case 1
104  *           - ensure source anon is resident in RAM
105  *           - if case 1B alloc new anon and copy from source
106  *           - map the correct page in
107  *   Case2:
108  *     - >>> handle case 2
109  *           - ensure source page is resident (if uobj)
110  *           - if case 2B alloc new anon and copy from source (could be zero
111  *		fill if uobj == NULL)
112  *           - map the correct page in
113  *     - done!
114  *
115  * note on paging:
116  *   if we have to do I/O we place a PG_BUSY page in the correct object,
117  * unlock everything, and do the I/O.   when I/O is done we must reverify
118  * the state of the world before assuming that our data structures are
119  * valid.   [because mappings could change while the map is unlocked]
120  *
121  *  alternative 1: unbusy the page in question and restart the page fault
122  *    from the top (ReFault).   this is easy but does not take advantage
123  *    of the information that we already have from our previous lookup,
124  *    although it is possible that the "hints" in the vm_map will help here.
125  *
126  * alternative 2: the system already keeps track of a "version" number of
127  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
128  *    mapping) you bump the version number up by one...]   so, we can save
129  *    the version number of the map before we release the lock and start I/O.
130  *    then when I/O is done we can relock and check the version numbers
131  *    to see if anything changed.    this might save us some over 1 because
132  *    we don't have to unbusy the page and may be less compares(?).
133  *
134  * alternative 3: put in backpointers or a way to "hold" part of a map
135  *    in place while I/O is in progress.   this could be complex to
136  *    implement (especially with structures like amap that can be referenced
137  *    by multiple map entries, and figuring out what should wait could be
138  *    complex as well...).
139  *
140  * we use alternative 2.  given that we are multi-threaded now we may want
141  * to reconsider the choice.
142  */
143 
144 /*
145  * local data structures
146  */
147 
148 struct uvm_advice {
149 	int advice;
150 	int nback;
151 	int nforw;
152 };
153 
154 /*
155  * page range array:
156  * note: index in array must match "advice" value
157  * XXX: borrowed numbers from freebsd.   do they work well for us?
158  */
159 
160 static const struct uvm_advice uvmadvice[] = {
161 	{ UVM_ADV_NORMAL, 3, 4 },
162 	{ UVM_ADV_RANDOM, 0, 0 },
163 	{ UVM_ADV_SEQUENTIAL, 8, 7},
164 };
165 
166 #define UVM_MAXRANGE 16	/* must be MAX() of nback+nforw+1 */
167 
168 /*
169  * private prototypes
170  */
171 
172 /*
173  * externs from other modules
174  */
175 
176 extern int start_init_exec;	/* Is init_main() done / init running? */
177 
178 /*
179  * inline functions
180  */
181 
182 /*
183  * uvmfault_anonflush: try and deactivate pages in specified anons
184  *
185  * => does not have to deactivate page if it is busy
186  */
187 
188 static inline void
189 uvmfault_anonflush(struct vm_anon **anons, int n)
190 {
191 	int lcv;
192 	struct vm_page *pg;
193 
194 	for (lcv = 0; lcv < n; lcv++) {
195 		if (anons[lcv] == NULL)
196 			continue;
197 		KASSERT(rw_lock_held(anons[lcv]->an_lock));
198 		pg = anons[lcv]->an_page;
199 		if (pg && (pg->flags & PG_BUSY) == 0) {
200 			uvm_pagelock(pg);
201 			uvm_pagedeactivate(pg);
202 			uvm_pageunlock(pg);
203 		}
204 	}
205 }
206 
207 /*
208  * normal functions
209  */
210 
211 /*
212  * uvmfault_amapcopy: clear "needs_copy" in a map.
213  *
214  * => called with VM data structures unlocked (usually, see below)
215  * => we get a write lock on the maps and clear needs_copy for a VA
216  * => if we are out of RAM we sleep (waiting for more)
217  */
218 
219 static void
220 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
221 {
222 	for (;;) {
223 
224 		/*
225 		 * no mapping?  give up.
226 		 */
227 
228 		if (uvmfault_lookup(ufi, true) == false)
229 			return;
230 
231 		/*
232 		 * copy if needed.
233 		 */
234 
235 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
236 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
237 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
238 
239 		/*
240 		 * didn't work?  must be out of RAM.   unlock and sleep.
241 		 */
242 
243 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
244 			uvmfault_unlockmaps(ufi, true);
245 			uvm_wait("fltamapcopy");
246 			continue;
247 		}
248 
249 		/*
250 		 * got it!   unlock and return.
251 		 */
252 
253 		uvmfault_unlockmaps(ufi, true);
254 		return;
255 	}
256 	/*NOTREACHED*/
257 }
258 
259 /*
260  * uvmfault_anonget: get data in an anon into a non-busy, non-released
261  * page in that anon.
262  *
263  * => Map, amap and thus anon should be locked by caller.
264  * => If we fail, we unlock everything and error is returned.
265  * => If we are successful, return with everything still locked.
266  * => We do not move the page on the queues [gets moved later].  If we
267  *    allocate a new page [we_own], it gets put on the queues.  Either way,
268  *    the result is that the page is on the queues at return time
269  * => For pages which are on loan from a uvm_object (and thus are not owned
270  *    by the anon): if successful, return with the owning object locked.
271  *    The caller must unlock this object when it unlocks everything else.
272  */
273 
274 int
275 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
276     struct vm_anon *anon)
277 {
278 	struct vm_page *pg;
279 	krw_t lock_type;
280 	int error;
281 
282 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
283 	KASSERT(rw_lock_held(anon->an_lock));
284 	KASSERT(anon->an_lock == amap->am_lock);
285 
286 	/* Increment the counters.*/
287 	cpu_count(CPU_COUNT_FLTANGET, 1);
288 	if (anon->an_page) {
289 		curlwp->l_ru.ru_minflt++;
290 	} else {
291 		curlwp->l_ru.ru_majflt++;
292 	}
293 	error = 0;
294 
295 	/*
296 	 * Loop until we get the anon data, or fail.
297 	 */
298 
299 	for (;;) {
300 		bool we_own, locked;
301 		/*
302 		 * Note: 'we_own' will become true if we set PG_BUSY on a page.
303 		 */
304 		we_own = false;
305 		pg = anon->an_page;
306 
307 		/*
308 		 * If there is a resident page and it is loaned, then anon
309 		 * may not own it.  Call out to uvm_anon_lockloanpg() to
310 		 * identify and lock the real owner of the page.
311 		 */
312 
313 		if (pg && pg->loan_count)
314 			pg = uvm_anon_lockloanpg(anon);
315 
316 		/*
317 		 * Is page resident?  Make sure it is not busy/released.
318 		 */
319 
320 		lock_type = rw_lock_op(anon->an_lock);
321 		if (pg) {
322 
323 			/*
324 			 * at this point, if the page has a uobject [meaning
325 			 * we have it on loan], then that uobject is locked
326 			 * by us!   if the page is busy, we drop all the
327 			 * locks (including uobject) and try again.
328 			 */
329 
330 			if ((pg->flags & PG_BUSY) == 0) {
331 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
332 				return 0;
333 			}
334 			cpu_count(CPU_COUNT_FLTPGWAIT, 1);
335 
336 			/*
337 			 * The last unlock must be an atomic unlock and wait
338 			 * on the owner of page.
339 			 */
340 
341 			if (pg->uobject) {
342 				/* Owner of page is UVM object. */
343 				uvmfault_unlockall(ufi, amap, NULL);
344 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
345 				    0,0,0);
346 				uvm_pagewait(pg, pg->uobject->vmobjlock, "anonget1");
347 			} else {
348 				/* Owner of page is anon. */
349 				uvmfault_unlockall(ufi, NULL, NULL);
350 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
351 				    0,0,0);
352 				uvm_pagewait(pg, anon->an_lock, "anonget2");
353 			}
354 		} else {
355 #if defined(VMSWAP)
356 			/*
357 			 * No page, therefore allocate one.  A write lock is
358 			 * required for this.  If the caller didn't supply
359 			 * one, fail now and have them retry.
360 			 */
361 
362 			if (lock_type == RW_READER) {
363 				return ENOLCK;
364 			}
365 			pg = uvm_pagealloc(NULL,
366 			    ufi != NULL ? ufi->orig_rvaddr : 0,
367 			    anon, ufi != NULL ? UVM_FLAG_COLORMATCH : 0);
368 			if (pg == NULL) {
369 				/* Out of memory.  Wait a little. */
370 				uvmfault_unlockall(ufi, amap, NULL);
371 				cpu_count(CPU_COUNT_FLTNORAM, 1);
372 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
373 				    0,0,0);
374 				if (!uvm_reclaimable()) {
375 					return ENOMEM;
376 				}
377 				uvm_wait("flt_noram1");
378 			} else {
379 				/* PG_BUSY bit is set. */
380 				we_own = true;
381 				uvmfault_unlockall(ufi, amap, NULL);
382 
383 				/*
384 				 * Pass a PG_BUSY+PG_FAKE clean page into
385 				 * the uvm_swap_get() function with all data
386 				 * structures unlocked.  Note that it is OK
387 				 * to read an_swslot here, because we hold
388 				 * PG_BUSY on the page.
389 				 */
390 				cpu_count(CPU_COUNT_PAGEINS, 1);
391 				error = uvm_swap_get(pg, anon->an_swslot,
392 				    PGO_SYNCIO);
393 
394 				/*
395 				 * We clean up after the I/O below in the
396 				 * 'we_own' case.
397 				 */
398 			}
399 #else
400 			panic("%s: no page", __func__);
401 #endif /* defined(VMSWAP) */
402 		}
403 
404 		/*
405 		 * Re-lock the map and anon.
406 		 */
407 
408 		locked = uvmfault_relock(ufi);
409 		if (locked || we_own) {
410 			rw_enter(anon->an_lock, lock_type);
411 		}
412 
413 		/*
414 		 * If we own the page (i.e. we set PG_BUSY), then we need
415 		 * to clean up after the I/O.  There are three cases to
416 		 * consider:
417 		 *
418 		 * 1) Page was released during I/O: free anon and ReFault.
419 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
420 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
421 		 *    case (i.e. drop anon lock if not locked).
422 		 */
423 
424 		if (we_own) {
425 			KASSERT(lock_type == RW_WRITER);
426 #if defined(VMSWAP)
427 			if (error) {
428 
429 				/*
430 				 * Remove the swap slot from the anon and
431 				 * mark the anon as having no real slot.
432 				 * Do not free the swap slot, thus preventing
433 				 * it from being used again.
434 				 */
435 
436 				if (anon->an_swslot > 0) {
437 					uvm_swap_markbad(anon->an_swslot, 1);
438 				}
439 				anon->an_swslot = SWSLOT_BAD;
440 
441 				if ((pg->flags & PG_RELEASED) != 0) {
442 					goto released;
443 				}
444 
445 				/*
446 				 * Note: page was never !PG_BUSY, so it
447 				 * cannot be mapped and thus no need to
448 				 * pmap_page_protect() it.
449 				 */
450 
451 				uvm_pagefree(pg);
452 
453 				if (locked) {
454 					uvmfault_unlockall(ufi, NULL, NULL);
455 				}
456 				rw_exit(anon->an_lock);
457 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
458 				return error;
459 			}
460 
461 			if ((pg->flags & PG_RELEASED) != 0) {
462 released:
463 				KASSERT(anon->an_ref == 0);
464 
465 				/*
466 				 * Released while we had unlocked amap.
467 				 */
468 
469 				if (locked) {
470 					uvmfault_unlockall(ufi, NULL, NULL);
471 				}
472 				uvm_anon_release(anon);
473 
474 				if (error) {
475 					UVMHIST_LOG(maphist,
476 					    "<- ERROR/RELEASED", 0,0,0,0);
477 					return error;
478 				}
479 
480 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
481 				return ERESTART;
482 			}
483 
484 			/*
485 			 * We have successfully read the page, activate it.
486 			 */
487 
488 			uvm_pagelock(pg);
489 			uvm_pageactivate(pg);
490 			uvm_pagewakeup(pg);
491 			uvm_pageunlock(pg);
492 			pg->flags &= ~(PG_BUSY|PG_FAKE);
493 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN);
494 			UVM_PAGE_OWN(pg, NULL);
495 #else
496 			panic("%s: we_own", __func__);
497 #endif /* defined(VMSWAP) */
498 		}
499 
500 		/*
501 		 * We were not able to re-lock the map - restart the fault.
502 		 */
503 
504 		if (!locked) {
505 			if (we_own) {
506 				rw_exit(anon->an_lock);
507 			}
508 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
509 			return ERESTART;
510 		}
511 
512 		/*
513 		 * Verify that no one has touched the amap and moved
514 		 * the anon on us.
515 		 */
516 
517 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
518 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
519 
520 			uvmfault_unlockall(ufi, amap, NULL);
521 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
522 			return ERESTART;
523 		}
524 
525 		/*
526 		 * Retry..
527 		 */
528 
529 		cpu_count(CPU_COUNT_FLTANRETRY, 1);
530 		continue;
531 	}
532 	/*NOTREACHED*/
533 }
534 
535 /*
536  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
537  *
538  *	1. allocate an anon and a page.
539  *	2. fill its contents.
540  *	3. put it into amap.
541  *
542  * => if we fail (result != 0) we unlock everything.
543  * => on success, return a new locked anon via 'nanon'.
544  *    (*nanon)->an_page will be a resident, locked, dirty page.
545  * => it's caller's responsibility to put the promoted nanon->an_page to the
546  *    page queue.
547  */
548 
549 static int
550 uvmfault_promote(struct uvm_faultinfo *ufi,
551     struct vm_anon *oanon,
552     struct vm_page *uobjpage,
553     struct vm_anon **nanon, /* OUT: allocated anon */
554     struct vm_anon **spare)
555 {
556 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
557 	struct uvm_object *uobj;
558 	struct vm_anon *anon;
559 	struct vm_page *pg;
560 	struct vm_page *opg;
561 	int error;
562 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
563 
564 	if (oanon) {
565 		/* anon COW */
566 		opg = oanon->an_page;
567 		KASSERT(opg != NULL);
568 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
569 	} else if (uobjpage != PGO_DONTCARE) {
570 		/* object-backed COW */
571 		opg = uobjpage;
572 		if ((uobjpage->flags & PG_BUSY) != 0) {
573 			KASSERT(rw_write_held(opg->uobject->vmobjlock));
574 		} else {
575 			KASSERT(rw_read_held(opg->uobject->vmobjlock));
576 		}
577 	} else {
578 		/* ZFOD */
579 		opg = NULL;
580 	}
581 	if (opg != NULL) {
582 		uobj = opg->uobject;
583 	} else {
584 		uobj = NULL;
585 	}
586 
587 	KASSERT(amap != NULL);
588 	KASSERT(uobjpage != NULL);
589 	KASSERT(rw_write_held(amap->am_lock));
590 	KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock);
591 	KASSERT(uobj == NULL || rw_lock_held(uobj->vmobjlock));
592 
593 	if (*spare != NULL) {
594 		anon = *spare;
595 		*spare = NULL;
596 	} else {
597 		anon = uvm_analloc();
598 	}
599 	if (anon) {
600 
601 		/*
602 		 * The new anon is locked.
603 		 *
604 		 * if opg == NULL, we want a zero'd, dirty page,
605 		 * so have uvm_pagealloc() do that for us.
606 		 */
607 
608 		KASSERT(anon->an_lock == NULL);
609 		anon->an_lock = amap->am_lock;
610 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
611 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
612 		if (pg == NULL) {
613 			anon->an_lock = NULL;
614 		}
615 	} else {
616 		pg = NULL;
617 	}
618 
619 	/*
620 	 * out of memory resources?
621 	 */
622 
623 	if (pg == NULL) {
624 		/* save anon for the next try. */
625 		if (anon != NULL) {
626 			*spare = anon;
627 		}
628 
629 		/* unlock and fail ... */
630 		if (uobjpage != PGO_DONTCARE &&
631 		    (uobjpage->flags & PG_BUSY) != 0) {
632 			uvm_page_unbusy(&uobjpage, 1);
633 		}
634 		uvmfault_unlockall(ufi, amap, uobj);
635 		if (!uvm_reclaimable()) {
636 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
637 			cpu_count(CPU_COUNT_FLTNOANON, 1);
638 			error = ENOMEM;
639 			goto done;
640 		}
641 
642 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
643 		cpu_count(CPU_COUNT_FLTNORAM, 1);
644 		uvm_wait("flt_noram5");
645 		error = ERESTART;
646 		goto done;
647 	}
648 
649 	/* copy page [pg now dirty] */
650 	if (opg) {
651 		uvm_pagecopy(opg, pg);
652 	}
653 	KASSERT(uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_DIRTY);
654 
655 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
656 	    oanon != NULL);
657 
658 	*nanon = anon;
659 	error = 0;
660 done:
661 	return error;
662 }
663 
664 /*
665  * Update statistics after fault resolution.
666  * - maxrss
667  */
668 void
669 uvmfault_update_stats(struct uvm_faultinfo *ufi)
670 {
671 	struct vm_map		*map;
672 	struct vmspace 		*vm;
673 	struct proc		*p;
674 	vsize_t			 res;
675 
676 	map = ufi->orig_map;
677 
678 	p = curproc;
679 	KASSERT(p != NULL);
680 	vm = p->p_vmspace;
681 
682 	if (&vm->vm_map != map)
683 		return;
684 
685 	res = pmap_resident_count(map->pmap);
686 	if (vm->vm_rssmax < res)
687 		vm->vm_rssmax = res;
688 }
689 
690 /*
691  *   F A U L T   -   m a i n   e n t r y   p o i n t
692  */
693 
694 /*
695  * uvm_fault: page fault handler
696  *
697  * => called from MD code to resolve a page fault
698  * => VM data structures usually should be unlocked.   however, it is
699  *	possible to call here with the main map locked if the caller
700  *	gets a write lock, sets it recusive, and then calls us (c.f.
701  *	uvm_map_pageable).   this should be avoided because it keeps
702  *	the map locked off during I/O.
703  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
704  */
705 
706 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
707 			 ~VM_PROT_WRITE : VM_PROT_ALL)
708 
709 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
710 #define UVM_FAULT_WIRE		(1 << 0)
711 #define UVM_FAULT_MAXPROT	(1 << 1)
712 
713 struct uvm_faultctx {
714 
715 	/*
716 	 * the following members are set up by uvm_fault_check() and
717 	 * read-only after that.
718 	 *
719 	 * note that narrow is used by uvm_fault_check() to change
720 	 * the behaviour after ERESTART.
721 	 *
722 	 * most of them might change after RESTART if the underlying
723 	 * map entry has been changed behind us.  an exception is
724 	 * wire_paging, which does never change.
725 	 */
726 	vm_prot_t access_type;
727 	vaddr_t startva;
728 	int npages;
729 	int centeridx;
730 	bool narrow;		/* work on a single requested page only */
731 	bool wire_mapping;	/* request a PMAP_WIRED mapping
732 				   (UVM_FAULT_WIRE or VM_MAPENT_ISWIRED) */
733 	bool wire_paging;	/* request uvm_pagewire
734 				   (true for UVM_FAULT_WIRE) */
735 	bool cow_now;		/* VM_PROT_WRITE is actually requested
736 				   (ie. should break COW and page loaning) */
737 
738 	/*
739 	 * enter_prot is set up by uvm_fault_check() and clamped
740 	 * (ie. drop the VM_PROT_WRITE bit) in various places in case
741 	 * of !cow_now.
742 	 */
743 	vm_prot_t enter_prot;	/* prot at which we want to enter pages in */
744 
745 	/*
746 	 * the following member is for uvmfault_promote() and ERESTART.
747 	 */
748 	struct vm_anon *anon_spare;
749 
750 	/*
751 	 * the folloing is actually a uvm_fault_lower() internal.
752 	 * it's here merely for debugging.
753 	 * (or due to the mechanical separation of the function?)
754 	 */
755 	bool promote;
756 
757 	/*
758 	 * type of lock to acquire on objects in both layers.
759 	 */
760 	krw_t lower_lock_type;
761 	krw_t upper_lock_type;
762 };
763 
764 static inline int	uvm_fault_check(
765 			    struct uvm_faultinfo *, struct uvm_faultctx *,
766 			    struct vm_anon ***, bool);
767 
768 static int		uvm_fault_upper(
769 			    struct uvm_faultinfo *, struct uvm_faultctx *,
770 			    struct vm_anon **);
771 static inline int	uvm_fault_upper_lookup(
772 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
773 			    struct vm_anon **, struct vm_page **);
774 static inline void	uvm_fault_upper_neighbor(
775 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
776 			    vaddr_t, struct vm_page *, bool);
777 static inline int	uvm_fault_upper_loan(
778 			    struct uvm_faultinfo *, struct uvm_faultctx *,
779 			    struct vm_anon *, struct uvm_object **);
780 static inline int	uvm_fault_upper_promote(
781 			    struct uvm_faultinfo *, struct uvm_faultctx *,
782 			    struct uvm_object *, struct vm_anon *);
783 static inline int	uvm_fault_upper_direct(
784 			    struct uvm_faultinfo *, struct uvm_faultctx *,
785 			    struct uvm_object *, struct vm_anon *);
786 static int		uvm_fault_upper_enter(
787 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
788 			    struct uvm_object *, struct vm_anon *,
789 			    struct vm_page *, struct vm_anon *);
790 static inline void	uvm_fault_upper_done(
791 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
792 			    struct vm_anon *, struct vm_page *);
793 
794 static int		uvm_fault_lower(
795 			    struct uvm_faultinfo *, struct uvm_faultctx *,
796 			    struct vm_page **);
797 static inline void	uvm_fault_lower_lookup(
798 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
799 			    struct vm_page **);
800 static inline void	uvm_fault_lower_neighbor(
801 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
802 			    vaddr_t, struct vm_page *);
803 static inline int	uvm_fault_lower_io(
804 			    struct uvm_faultinfo *, struct uvm_faultctx *,
805 			    struct uvm_object **, struct vm_page **);
806 static inline int	uvm_fault_lower_direct(
807 			    struct uvm_faultinfo *, struct uvm_faultctx *,
808 			    struct uvm_object *, struct vm_page *);
809 static inline int	uvm_fault_lower_direct_loan(
810 			    struct uvm_faultinfo *, struct uvm_faultctx *,
811 			    struct uvm_object *, struct vm_page **,
812 			    struct vm_page **);
813 static inline int	uvm_fault_lower_promote(
814 			    struct uvm_faultinfo *, struct uvm_faultctx *,
815 			    struct uvm_object *, struct vm_page *);
816 static int		uvm_fault_lower_enter(
817 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
818 			    struct uvm_object *,
819 			    struct vm_anon *, struct vm_page *);
820 static inline void	uvm_fault_lower_done(
821 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
822 			    struct uvm_object *, struct vm_page *);
823 
824 int
825 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
826     vm_prot_t access_type, int fault_flag)
827 {
828 	struct uvm_faultinfo ufi;
829 	struct uvm_faultctx flt = {
830 		.access_type = access_type,
831 
832 		/* don't look for neighborhood * pages on "wire" fault */
833 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
834 
835 		/* "wire" fault causes wiring of both mapping and paging */
836 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
837 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
838 
839 		/*
840 		 * default lock type to acquire on upper & lower layer
841 		 * objects: reader.  this can be upgraded at any point
842 		 * during the fault from read -> write and uvm_faultctx
843 		 * changed to match, but is never downgraded write -> read.
844 		 */
845 #ifdef __HAVE_UNLOCKED_PMAP /* XXX temporary */
846 		.upper_lock_type = RW_WRITER,
847 		.lower_lock_type = RW_WRITER,
848 #else
849 		.upper_lock_type = RW_READER,
850 		.lower_lock_type = RW_READER,
851 #endif
852 	};
853 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
854 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
855 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
856 	int error;
857 
858 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
859 
860 	UVMHIST_LOG(maphist, "(map=%#jx, vaddr=%#jx, at=%jd, ff=%jd)",
861 	      (uintptr_t)orig_map, vaddr, access_type, fault_flag);
862 
863 	/* Don't count anything until user interaction is possible */
864 	kpreempt_disable();
865 	if (__predict_true(start_init_exec)) {
866 		struct cpu_info *ci = curcpu();
867 		CPU_COUNT(CPU_COUNT_NFAULT, 1);
868 		/* Don't flood RNG subsystem with samples. */
869 		if (++(ci->ci_faultrng) == 503) {
870 			ci->ci_faultrng = 0;
871 			rnd_add_uint32(&curcpu()->ci_data.cpu_uvm->rs,
872 			    sizeof(vaddr_t) == sizeof(uint32_t) ?
873 			    (uint32_t)vaddr : sizeof(vaddr_t) ==
874 			    sizeof(uint64_t) ?
875 			    (uint32_t)vaddr :
876 			    (uint32_t)ci->ci_counts[CPU_COUNT_NFAULT]);
877 		}
878 	}
879 	kpreempt_enable();
880 
881 	/*
882 	 * init the IN parameters in the ufi
883 	 */
884 
885 	ufi.orig_map = orig_map;
886 	ufi.orig_rvaddr = trunc_page(vaddr);
887 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
888 
889 	error = ERESTART;
890 	while (error == ERESTART) { /* ReFault: */
891 		anons = anons_store;
892 		pages = pages_store;
893 
894 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
895 		if (error != 0)
896 			continue;
897 
898 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
899 		if (error != 0)
900 			continue;
901 
902 		if (pages[flt.centeridx] == PGO_DONTCARE)
903 			error = uvm_fault_upper(&ufi, &flt, anons);
904 		else {
905 			struct uvm_object * const uobj =
906 			    ufi.entry->object.uvm_obj;
907 
908 			if (uobj && uobj->pgops->pgo_fault != NULL) {
909 				/*
910 				 * invoke "special" fault routine.
911 				 */
912 				rw_enter(uobj->vmobjlock, RW_WRITER);
913 				/* locked: maps(read), amap(if there), uobj */
914 				error = uobj->pgops->pgo_fault(&ufi,
915 				    flt.startva, pages, flt.npages,
916 				    flt.centeridx, flt.access_type,
917 				    PGO_LOCKED|PGO_SYNCIO);
918 
919 				/*
920 				 * locked: nothing, pgo_fault has unlocked
921 				 * everything
922 				 */
923 
924 				/*
925 				 * object fault routine responsible for
926 				 * pmap_update().
927 				 */
928 
929 				/*
930 				 * Wake up the pagedaemon if the fault method
931 				 * failed for lack of memory but some can be
932 				 * reclaimed.
933 				 */
934 				if (error == ENOMEM && uvm_reclaimable()) {
935 					uvm_wait("pgo_fault");
936 					error = ERESTART;
937 				}
938 			} else {
939 				error = uvm_fault_lower(&ufi, &flt, pages);
940 			}
941 		}
942 	}
943 
944 	if (flt.anon_spare != NULL) {
945 		flt.anon_spare->an_ref--;
946 		KASSERT(flt.anon_spare->an_ref == 0);
947 		KASSERT(flt.anon_spare->an_lock == NULL);
948 		uvm_anfree(flt.anon_spare);
949 	}
950 	return error;
951 }
952 
953 /*
954  * uvm_fault_check: check prot, handle needs-copy, etc.
955  *
956  *	1. lookup entry.
957  *	2. check protection.
958  *	3. adjust fault condition (mainly for simulated fault).
959  *	4. handle needs-copy (lazy amap copy).
960  *	5. establish range of interest for neighbor fault (aka pre-fault).
961  *	6. look up anons (if amap exists).
962  *	7. flush pages (if MADV_SEQUENTIAL)
963  *
964  * => called with nothing locked.
965  * => if we fail (result != 0) we unlock everything.
966  * => initialize/adjust many members of flt.
967  */
968 
969 static int
970 uvm_fault_check(
971 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
972 	struct vm_anon ***ranons, bool maxprot)
973 {
974 	struct vm_amap *amap;
975 	struct uvm_object *uobj;
976 	vm_prot_t check_prot;
977 	int nback, nforw;
978 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
979 
980 	/*
981 	 * lookup and lock the maps
982 	 */
983 
984 	if (uvmfault_lookup(ufi, false) == false) {
985 		UVMHIST_LOG(maphist, "<- no mapping @ %#jx", ufi->orig_rvaddr,
986 		    0,0,0);
987 		return EFAULT;
988 	}
989 	/* locked: maps(read) */
990 
991 #ifdef DIAGNOSTIC
992 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
993 		printf("Page fault on non-pageable map:\n");
994 		printf("ufi->map = %p\n", ufi->map);
995 		printf("ufi->orig_map = %p\n", ufi->orig_map);
996 		printf("ufi->orig_rvaddr = %#lx\n", (u_long) ufi->orig_rvaddr);
997 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
998 	}
999 #endif
1000 
1001 	/*
1002 	 * check protection
1003 	 */
1004 
1005 	check_prot = maxprot ?
1006 	    ufi->entry->max_protection : ufi->entry->protection;
1007 	if ((check_prot & flt->access_type) != flt->access_type) {
1008 		UVMHIST_LOG(maphist,
1009 		    "<- protection failure (prot=%#jx, access=%#jx)",
1010 		    ufi->entry->protection, flt->access_type, 0, 0);
1011 		uvmfault_unlockmaps(ufi, false);
1012 		return EFAULT;
1013 	}
1014 
1015 	/*
1016 	 * "enter_prot" is the protection we want to enter the page in at.
1017 	 * for certain pages (e.g. copy-on-write pages) this protection can
1018 	 * be more strict than ufi->entry->protection.  "wired" means either
1019 	 * the entry is wired or we are fault-wiring the pg.
1020 	 */
1021 
1022 	flt->enter_prot = ufi->entry->protection;
1023 	if (VM_MAPENT_ISWIRED(ufi->entry)) {
1024 		flt->wire_mapping = true;
1025 		flt->wire_paging = true;
1026 		flt->narrow = true;
1027 	}
1028 
1029 	if (flt->wire_mapping) {
1030 		flt->access_type = flt->enter_prot; /* full access for wired */
1031 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
1032 	} else {
1033 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
1034 	}
1035 
1036 	if (flt->wire_paging) {
1037 		/* wiring pages requires a write lock. */
1038 		flt->upper_lock_type = RW_WRITER;
1039 		flt->lower_lock_type = RW_WRITER;
1040 	}
1041 
1042 	flt->promote = false;
1043 
1044 	/*
1045 	 * handle "needs_copy" case.   if we need to copy the amap we will
1046 	 * have to drop our readlock and relock it with a write lock.  (we
1047 	 * need a write lock to change anything in a map entry [e.g.
1048 	 * needs_copy]).
1049 	 */
1050 
1051 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
1052 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
1053 			KASSERT(!maxprot);
1054 			/* need to clear */
1055 			UVMHIST_LOG(maphist,
1056 			    "  need to clear needs_copy and refault",0,0,0,0);
1057 			uvmfault_unlockmaps(ufi, false);
1058 			uvmfault_amapcopy(ufi);
1059 			cpu_count(CPU_COUNT_FLTAMCOPY, 1);
1060 			return ERESTART;
1061 
1062 		} else {
1063 
1064 			/*
1065 			 * ensure that we pmap_enter page R/O since
1066 			 * needs_copy is still true
1067 			 */
1068 
1069 			flt->enter_prot &= ~VM_PROT_WRITE;
1070 		}
1071 	}
1072 
1073 	/*
1074 	 * identify the players
1075 	 */
1076 
1077 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
1078 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
1079 
1080 	/*
1081 	 * check for a case 0 fault.  if nothing backing the entry then
1082 	 * error now.
1083 	 */
1084 
1085 	if (amap == NULL && uobj == NULL) {
1086 		uvmfault_unlockmaps(ufi, false);
1087 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
1088 		return EFAULT;
1089 	}
1090 
1091 	/*
1092 	 * establish range of interest based on advice from mapper
1093 	 * and then clip to fit map entry.   note that we only want
1094 	 * to do this the first time through the fault.   if we
1095 	 * ReFault we will disable this by setting "narrow" to true.
1096 	 */
1097 
1098 	if (flt->narrow == false) {
1099 
1100 		/* wide fault (!narrow) */
1101 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
1102 			 ufi->entry->advice);
1103 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
1104 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
1105 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
1106 		/*
1107 		 * note: "-1" because we don't want to count the
1108 		 * faulting page as forw
1109 		 */
1110 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1111 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1112 			     PAGE_SHIFT) - 1);
1113 		flt->npages = nback + nforw + 1;
1114 		flt->centeridx = nback;
1115 
1116 		flt->narrow = true;	/* ensure only once per-fault */
1117 
1118 	} else {
1119 
1120 		/* narrow fault! */
1121 		nback = nforw = 0;
1122 		flt->startva = ufi->orig_rvaddr;
1123 		flt->npages = 1;
1124 		flt->centeridx = 0;
1125 
1126 	}
1127 	/* offset from entry's start to pgs' start */
1128 	const voff_t eoff = flt->startva - ufi->entry->start;
1129 
1130 	/* locked: maps(read) */
1131 	UVMHIST_LOG(maphist, "  narrow=%jd, back=%jd, forw=%jd, startva=%#jx",
1132 		    flt->narrow, nback, nforw, flt->startva);
1133 	UVMHIST_LOG(maphist, "  entry=%#jx, amap=%#jx, obj=%#jx",
1134 	    (uintptr_t)ufi->entry, (uintptr_t)amap, (uintptr_t)uobj, 0);
1135 
1136 	/*
1137 	 * guess at the most suitable lock types to acquire.
1138 	 * if we've got an amap then lock it and extract current anons.
1139 	 */
1140 
1141 	if (amap) {
1142 		if ((amap_flags(amap) & AMAP_SHARED) == 0) {
1143 			/*
1144 			 * the amap isn't shared.  get a writer lock to
1145 			 * avoid the cost of upgrading the lock later if
1146 			 * needed.
1147 			 *
1148 			 * XXX nice for PostgreSQL, but consider threads.
1149 			 */
1150 			flt->upper_lock_type = RW_WRITER;
1151 		} else if ((flt->access_type & VM_PROT_WRITE) != 0) {
1152 			/*
1153 			 * assume we're about to COW.
1154 			 */
1155 			flt->upper_lock_type = RW_WRITER;
1156 		}
1157 		amap_lock(amap, flt->upper_lock_type);
1158 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1159 	} else {
1160 		if ((flt->access_type & VM_PROT_WRITE) != 0) {
1161 			/*
1162 			 * we are about to dirty the object and that
1163 			 * requires a write lock.
1164 			 */
1165 			flt->lower_lock_type = RW_WRITER;
1166 		}
1167 		*ranons = NULL;	/* to be safe */
1168 	}
1169 
1170 	/* locked: maps(read), amap(if there) */
1171 	KASSERT(amap == NULL ||
1172 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1173 
1174 	/*
1175 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1176 	 * now and then forget about them (for the rest of the fault).
1177 	 */
1178 
1179 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1180 
1181 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1182 		    0,0,0,0);
1183 		/* flush back-page anons? */
1184 		if (amap)
1185 			uvmfault_anonflush(*ranons, nback);
1186 
1187 		/* flush object? */
1188 		if (uobj) {
1189 			voff_t uoff;
1190 
1191 			uoff = ufi->entry->offset + eoff;
1192 			rw_enter(uobj->vmobjlock, RW_WRITER);
1193 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1194 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1195 		}
1196 
1197 		/* now forget about the backpages */
1198 		if (amap)
1199 			*ranons += nback;
1200 		flt->startva += (nback << PAGE_SHIFT);
1201 		flt->npages -= nback;
1202 		flt->centeridx = 0;
1203 	}
1204 	/*
1205 	 * => startva is fixed
1206 	 * => npages is fixed
1207 	 */
1208 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1209 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1210 	    flt->startva + (flt->npages << PAGE_SHIFT));
1211 	return 0;
1212 }
1213 
1214 /*
1215  * uvm_fault_upper_upgrade: upgrade upper lock, reader -> writer
1216  */
1217 
1218 static inline int
1219 uvm_fault_upper_upgrade(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1220     struct vm_amap *amap, struct uvm_object *uobj)
1221 {
1222 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1223 
1224 	KASSERT(amap != NULL);
1225 	KASSERT(flt->upper_lock_type == rw_lock_op(amap->am_lock));
1226 
1227 	/*
1228 	 * fast path.
1229 	 */
1230 
1231 	if (__predict_true(flt->upper_lock_type == RW_WRITER)) {
1232 		return 0;
1233 	}
1234 
1235 	/*
1236 	 * otherwise try for the upgrade.  if we don't get it, unlock
1237 	 * everything, restart the fault and next time around get a writer
1238 	 * lock.
1239 	 */
1240 
1241 	flt->upper_lock_type = RW_WRITER;
1242 	if (__predict_false(!rw_tryupgrade(amap->am_lock))) {
1243 		uvmfault_unlockall(ufi, amap, uobj);
1244 		cpu_count(CPU_COUNT_FLTNOUP, 1);
1245 		UVMHIST_LOG(maphist, "  !upgrade upper", 0, 0,0,0);
1246 		return ERESTART;
1247 	}
1248 	cpu_count(CPU_COUNT_FLTUP, 1);
1249 	KASSERT(flt->upper_lock_type == rw_lock_op(amap->am_lock));
1250 	return 0;
1251 }
1252 
1253 /*
1254  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1255  *
1256  * iterate range of interest:
1257  *	1. check if h/w mapping exists.  if yes, we don't care
1258  *	2. check if anon exists.  if not, page is lower.
1259  *	3. if anon exists, enter h/w mapping for neighbors.
1260  *
1261  * => called with amap locked (if exists).
1262  */
1263 
1264 static int
1265 uvm_fault_upper_lookup(
1266 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1267 	struct vm_anon **anons, struct vm_page **pages)
1268 {
1269 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1270 	int lcv;
1271 	vaddr_t currva;
1272 	bool shadowed __unused;
1273 	bool entered;
1274 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1275 
1276 	/* locked: maps(read), amap(if there) */
1277 	KASSERT(amap == NULL ||
1278 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1279 
1280 	/*
1281 	 * map in the backpages and frontpages we found in the amap in hopes
1282 	 * of preventing future faults.    we also init the pages[] array as
1283 	 * we go.
1284 	 */
1285 
1286 	currva = flt->startva;
1287 	shadowed = false;
1288 	entered = false;
1289 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1290 		/*
1291 		 * unmapped or center page.   check if any anon at this level.
1292 		 */
1293 		if (amap == NULL || anons[lcv] == NULL) {
1294 			pages[lcv] = NULL;
1295 			continue;
1296 		}
1297 
1298 		/*
1299 		 * check for present page and map if possible.
1300 		 */
1301 
1302 		pages[lcv] = PGO_DONTCARE;
1303 		if (lcv == flt->centeridx) {	/* save center for later! */
1304 			shadowed = true;
1305 			continue;
1306 		}
1307 
1308 		struct vm_anon *anon = anons[lcv];
1309 		struct vm_page *pg = anon->an_page;
1310 
1311 		KASSERT(anon->an_lock == amap->am_lock);
1312 
1313 		/*
1314 		 * ignore loaned and busy pages.
1315 		 * don't play with VAs that are already mapped.
1316 		 */
1317 
1318 		if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0 &&
1319 		    !pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1320 			uvm_fault_upper_neighbor(ufi, flt, currva,
1321 			    pg, anon->an_ref > 1);
1322 			entered = true;
1323 		}
1324 	}
1325 	if (entered) {
1326 		pmap_update(ufi->orig_map->pmap);
1327 	}
1328 
1329 	/* locked: maps(read), amap(if there) */
1330 	KASSERT(amap == NULL ||
1331 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1332 	/* (shadowed == true) if there is an anon at the faulting address */
1333 	UVMHIST_LOG(maphist, "  shadowed=%jd, will_get=%jd", shadowed,
1334 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1335 
1336 	/*
1337 	 * note that if we are really short of RAM we could sleep in the above
1338 	 * call to pmap_enter with everything locked.   bad?
1339 	 *
1340 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1341 	 * XXX case.  --thorpej
1342 	 */
1343 
1344 	return 0;
1345 }
1346 
1347 /*
1348  * uvm_fault_upper_neighbor: enter single upper neighbor page.
1349  *
1350  * => called with amap and anon locked.
1351  */
1352 
1353 static void
1354 uvm_fault_upper_neighbor(
1355 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1356 	vaddr_t currva, struct vm_page *pg, bool readonly)
1357 {
1358 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1359 
1360 	/* locked: amap, anon */
1361 
1362 	KASSERT(pg->uobject == NULL);
1363 	KASSERT(pg->uanon != NULL);
1364 	KASSERT(rw_lock_op(pg->uanon->an_lock) == flt->upper_lock_type);
1365 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1366 
1367 	/*
1368 	 * in the read-locked case, it's not possible for this to be a new
1369 	 * page, therefore it's enqueued already.  there wasn't a direct
1370 	 * fault on the page, so avoid the cost of re-enqueuing it unless
1371 	 * write-locked.
1372 	 */
1373 
1374 	if (flt->upper_lock_type == RW_WRITER) {
1375 		uvm_pagelock(pg);
1376 		uvm_pageenqueue(pg);
1377 		uvm_pageunlock(pg);
1378 	}
1379 	UVMHIST_LOG(maphist,
1380 	    "  MAPPING: n anon: pm=%#jx, va=%#jx, pg=%#jx",
1381 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
1382 	cpu_count(CPU_COUNT_FLTNAMAP, 1);
1383 
1384 	/*
1385 	 * Since this page isn't the page that's actually faulting,
1386 	 * ignore pmap_enter() failures; it's not critical that we
1387 	 * enter these right now.
1388 	 */
1389 
1390 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1391 	    VM_PAGE_TO_PHYS(pg),
1392 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1393 	    flt->enter_prot,
1394 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1395 }
1396 
1397 /*
1398  * uvm_fault_upper: handle upper fault.
1399  *
1400  *	1. acquire anon lock.
1401  *	2. get anon.  let uvmfault_anonget do the dirty work.
1402  *	3. handle loan.
1403  *	4. dispatch direct or promote handlers.
1404  */
1405 
1406 static int
1407 uvm_fault_upper(
1408 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1409 	struct vm_anon **anons)
1410 {
1411 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1412 	struct vm_anon * const anon = anons[flt->centeridx];
1413 	struct uvm_object *uobj;
1414 	int error;
1415 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1416 
1417 	/* locked: maps(read), amap, anon */
1418 	KASSERT(rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1419 	KASSERT(anon->an_lock == amap->am_lock);
1420 
1421 	/*
1422 	 * handle case 1: fault on an anon in our amap
1423 	 */
1424 
1425 	UVMHIST_LOG(maphist, "  case 1 fault: anon=%#jx",
1426 	    (uintptr_t)anon, 0, 0, 0);
1427 
1428 	/*
1429 	 * no matter if we have case 1A or case 1B we are going to need to
1430 	 * have the anon's memory resident.   ensure that now.
1431 	 */
1432 
1433 	/*
1434 	 * let uvmfault_anonget do the dirty work.
1435 	 * if it fails (!OK) it will unlock everything for us.
1436 	 * if it succeeds, locks are still valid and locked.
1437 	 * also, if it is OK, then the anon's page is on the queues.
1438 	 * if the page is on loan from a uvm_object, then anonget will
1439 	 * lock that object for us if it does not fail.
1440 	 */
1441  retry:
1442 	error = uvmfault_anonget(ufi, amap, anon);
1443 	switch (error) {
1444 	case 0:
1445 		break;
1446 
1447 	case ERESTART:
1448 		return ERESTART;
1449 
1450 	case EAGAIN:
1451 		kpause("fltagain1", false, hz/2, NULL);
1452 		return ERESTART;
1453 
1454 	case ENOLCK:
1455 		/* it needs a write lock: retry */
1456 		error = uvm_fault_upper_upgrade(ufi, flt, amap, NULL);
1457 		if (error != 0) {
1458 			return error;
1459 		}
1460 		KASSERT(rw_write_held(amap->am_lock));
1461 		goto retry;
1462 
1463 	default:
1464 		return error;
1465 	}
1466 
1467 	/*
1468 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1469 	 */
1470 
1471 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1472 
1473 	/* locked: maps(read), amap, anon, uobj(if one) */
1474 	KASSERT(rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1475 	KASSERT(anon->an_lock == amap->am_lock);
1476 	KASSERT(uobj == NULL ||
1477 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
1478 
1479 	/*
1480 	 * special handling for loaned pages
1481 	 */
1482 
1483 	if (anon->an_page->loan_count) {
1484 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1485 		if (error != 0)
1486 			return error;
1487 	}
1488 
1489 	/*
1490 	 * if we are case 1B then we will need to allocate a new blank
1491 	 * anon to transfer the data into.   note that we have a lock
1492 	 * on anon, so no one can busy or release the page until we are done.
1493 	 * also note that the ref count can't drop to zero here because
1494 	 * it is > 1 and we are only dropping one ref.
1495 	 *
1496 	 * in the (hopefully very rare) case that we are out of RAM we
1497 	 * will unlock, wait for more RAM, and refault.
1498 	 *
1499 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1500 	 */
1501 
1502 	if (flt->cow_now && anon->an_ref > 1) {
1503 		flt->promote = true;
1504 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1505 	} else {
1506 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1507 	}
1508 	return error;
1509 }
1510 
1511 /*
1512  * uvm_fault_upper_loan: handle loaned upper page.
1513  *
1514  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1515  *	2. if cow'ing now, and if ref count is 1, break loan.
1516  */
1517 
1518 static int
1519 uvm_fault_upper_loan(
1520 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1521 	struct vm_anon *anon, struct uvm_object **ruobj)
1522 {
1523 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1524 	int error = 0;
1525 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1526 
1527 	if (!flt->cow_now) {
1528 
1529 		/*
1530 		 * for read faults on loaned pages we just cap the
1531 		 * protection at read-only.
1532 		 */
1533 
1534 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1535 
1536 	} else {
1537 		/*
1538 		 * note that we can't allow writes into a loaned page!
1539 		 *
1540 		 * if we have a write fault on a loaned page in an
1541 		 * anon then we need to look at the anon's ref count.
1542 		 * if it is greater than one then we are going to do
1543 		 * a normal copy-on-write fault into a new anon (this
1544 		 * is not a problem).  however, if the reference count
1545 		 * is one (a case where we would normally allow a
1546 		 * write directly to the page) then we need to kill
1547 		 * the loan before we continue.
1548 		 */
1549 
1550 		/* >1 case is already ok */
1551 		if (anon->an_ref == 1) {
1552 			/* breaking loan requires a write lock. */
1553 			error = uvm_fault_upper_upgrade(ufi, flt, amap, NULL);
1554 			if (error != 0) {
1555 				return error;
1556 			}
1557 			KASSERT(rw_write_held(amap->am_lock));
1558 
1559 			error = uvm_loanbreak_anon(anon, *ruobj);
1560 			if (error != 0) {
1561 				uvmfault_unlockall(ufi, amap, *ruobj);
1562 				uvm_wait("flt_noram2");
1563 				return ERESTART;
1564 			}
1565 			/* if we were a loan receiver uobj is gone */
1566 			if (*ruobj)
1567 				*ruobj = NULL;
1568 		}
1569 	}
1570 	return error;
1571 }
1572 
1573 /*
1574  * uvm_fault_upper_promote: promote upper page.
1575  *
1576  *	1. call uvmfault_promote.
1577  *	2. enqueue page.
1578  *	3. deref.
1579  *	4. pass page to uvm_fault_upper_enter.
1580  */
1581 
1582 static int
1583 uvm_fault_upper_promote(
1584 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1585 	struct uvm_object *uobj, struct vm_anon *anon)
1586 {
1587 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1588 	struct vm_anon * const oanon = anon;
1589 	struct vm_page *pg;
1590 	int error;
1591 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1592 
1593 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1594 	cpu_count(CPU_COUNT_FLT_ACOW, 1);
1595 
1596 	/* promoting requires a write lock. */
1597 	error = uvm_fault_upper_upgrade(ufi, flt, amap, NULL);
1598 	if (error != 0) {
1599 		return error;
1600 	}
1601 	KASSERT(rw_write_held(amap->am_lock));
1602 
1603 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1604 	    &flt->anon_spare);
1605 	switch (error) {
1606 	case 0:
1607 		break;
1608 	case ERESTART:
1609 		return ERESTART;
1610 	default:
1611 		return error;
1612 	}
1613 
1614 	KASSERT(anon->an_lock == oanon->an_lock);
1615 
1616 	/* uvm_fault_upper_done will activate or enqueue the page */
1617 	pg = anon->an_page;
1618 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1619 	UVM_PAGE_OWN(pg, NULL);
1620 
1621 	/* deref: can not drop to zero here by defn! */
1622 	KASSERT(oanon->an_ref > 1);
1623 	oanon->an_ref--;
1624 
1625 	/*
1626 	 * note: oanon is still locked, as is the new anon.  we
1627 	 * need to check for this later when we unlock oanon; if
1628 	 * oanon != anon, we'll have to unlock anon, too.
1629 	 */
1630 
1631 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1632 }
1633 
1634 /*
1635  * uvm_fault_upper_direct: handle direct fault.
1636  */
1637 
1638 static int
1639 uvm_fault_upper_direct(
1640 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1641 	struct uvm_object *uobj, struct vm_anon *anon)
1642 {
1643 	struct vm_anon * const oanon = anon;
1644 	struct vm_page *pg;
1645 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1646 
1647 	cpu_count(CPU_COUNT_FLT_ANON, 1);
1648 	pg = anon->an_page;
1649 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1650 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1651 
1652 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1653 }
1654 
1655 /*
1656  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1657  */
1658 
1659 static int
1660 uvm_fault_upper_enter(
1661 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1662 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1663 	struct vm_anon *oanon)
1664 {
1665 	struct pmap *pmap = ufi->orig_map->pmap;
1666 	vaddr_t va = ufi->orig_rvaddr;
1667 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1668 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1669 
1670 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1671 	KASSERT(rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1672 	KASSERT(anon->an_lock == amap->am_lock);
1673 	KASSERT(oanon->an_lock == amap->am_lock);
1674 	KASSERT(uobj == NULL ||
1675 	    rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
1676 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1677 
1678 	/*
1679 	 * now map the page in.
1680 	 */
1681 
1682 	UVMHIST_LOG(maphist,
1683 	    "  MAPPING: anon: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
1684 	    (uintptr_t)pmap, va, (uintptr_t)pg, flt->promote);
1685 	if (pmap_enter(pmap, va, VM_PAGE_TO_PHYS(pg),
1686 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1687 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1688 
1689 		/*
1690 		 * If pmap_enter() fails, it must not leave behind an existing
1691 		 * pmap entry.  In particular, a now-stale entry for a different
1692 		 * page would leave the pmap inconsistent with the vm_map.
1693 		 * This is not to imply that pmap_enter() should remove an
1694 		 * existing mapping in such a situation (since that could create
1695 		 * different problems, eg. if the existing mapping is wired),
1696 		 * but rather that the pmap should be designed such that it
1697 		 * never needs to fail when the new mapping is replacing an
1698 		 * existing mapping and the new page has no existing mappings.
1699 		 */
1700 
1701 		KASSERT(!pmap_extract(pmap, va, NULL));
1702 
1703 		/*
1704 		 * ensure that the page is queued in the case that
1705 		 * we just promoted.
1706 		 */
1707 
1708 		if (flt->upper_lock_type == RW_WRITER) {
1709 			uvm_pagelock(pg);
1710 			uvm_pageenqueue(pg);
1711 			uvm_pageunlock(pg);
1712 		}
1713 
1714 		/*
1715 		 * No need to undo what we did; we can simply think of
1716 		 * this as the pmap throwing away the mapping information.
1717 		 *
1718 		 * We do, however, have to go through the ReFault path,
1719 		 * as the map may change while we're asleep.
1720 		 */
1721 
1722 		uvmfault_unlockall(ufi, amap, uobj);
1723 		if (!uvm_reclaimable()) {
1724 			UVMHIST_LOG(maphist,
1725 			    "<- failed.  out of VM",0,0,0,0);
1726 			/* XXX instrumentation */
1727 			return ENOMEM;
1728 		}
1729 		/* XXX instrumentation */
1730 		uvm_wait("flt_pmfail1");
1731 		return ERESTART;
1732 	}
1733 
1734 	uvm_fault_upper_done(ufi, flt, anon, pg);
1735 
1736 	/*
1737 	 * done case 1!  finish up by unlocking everything and returning success
1738 	 */
1739 
1740 	pmap_update(pmap);
1741 	uvmfault_unlockall(ufi, amap, uobj);
1742 	return 0;
1743 }
1744 
1745 /*
1746  * uvm_fault_upper_done: queue upper center page.
1747  */
1748 
1749 static void
1750 uvm_fault_upper_done(
1751 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1752 	struct vm_anon *anon, struct vm_page *pg)
1753 {
1754 	const bool wire_paging = flt->wire_paging;
1755 
1756 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1757 
1758 	/*
1759 	 * ... update the page queues.
1760 	 */
1761 
1762 	uvm_pagelock(pg);
1763 	if (wire_paging) {
1764 		uvm_pagewire(pg);
1765 	} else {
1766 		uvm_pageactivate(pg);
1767 	}
1768 	uvm_pageunlock(pg);
1769 
1770 	if (wire_paging) {
1771 		/*
1772 		 * since the now-wired page cannot be paged out,
1773 		 * release its swap resources for others to use.
1774 		 * and since an anon with no swap cannot be clean,
1775 		 * mark it dirty now.
1776 		 */
1777 
1778 		uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
1779 		uvm_anon_dropswap(anon);
1780 	}
1781 }
1782 
1783 /*
1784  * uvm_fault_lower_upgrade: upgrade lower lock, reader -> writer
1785  */
1786 
1787 static inline int
1788 uvm_fault_lower_upgrade(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1789     struct vm_amap *amap, struct uvm_object *uobj, struct vm_page *uobjpage)
1790 {
1791 
1792 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1793 
1794 	KASSERT(uobj != NULL);
1795 	KASSERT(flt->lower_lock_type == rw_lock_op(uobj->vmobjlock));
1796 
1797 	/*
1798 	 * fast path.
1799 	 */
1800 
1801 	if (__predict_true(flt->lower_lock_type == RW_WRITER)) {
1802 		KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1803 		return 0;
1804 	}
1805 
1806 	/*
1807 	 * otherwise try for the upgrade.  if we don't get it, unlock
1808 	 * everything, restart the fault and next time around get a writer
1809 	 * lock.
1810 	 */
1811 
1812 	flt->lower_lock_type = RW_WRITER;
1813 	if (__predict_false(!rw_tryupgrade(uobj->vmobjlock))) {
1814 		uvmfault_unlockall(ufi, amap, uobj);
1815 		cpu_count(CPU_COUNT_FLTNOUP, 1);
1816 		UVMHIST_LOG(maphist, "  !upgrade lower", 0, 0,0,0);
1817 		return ERESTART;
1818 	}
1819 	cpu_count(CPU_COUNT_FLTUP, 1);
1820 	KASSERT(flt->lower_lock_type == rw_lock_op(uobj->vmobjlock));
1821 
1822 	/*
1823 	 * finally, if a page was supplied, assert that it's not busy
1824 	 * (can't be with a reader lock) and then mark it busy now that
1825 	 * we have a writer lock.
1826 	 */
1827 
1828 	if (uobjpage != NULL) {
1829 		KASSERT((uobjpage->flags & PG_BUSY) == 0);
1830 		uobjpage->flags |= PG_BUSY;
1831 		UVM_PAGE_OWN(uobjpage, "upgrdlwr");
1832 	}
1833 	return 0;
1834 }
1835 
1836 /*
1837  * uvm_fault_lower: handle lower fault.
1838  *
1839  *	1. check uobj
1840  *	1.1. if null, ZFOD.
1841  *	1.2. if not null, look up unnmapped neighbor pages.
1842  *	2. for center page, check if promote.
1843  *	2.1. ZFOD always needs promotion.
1844  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1845  *	3. if uobj is not ZFOD and page is not found, do i/o.
1846  *	4. dispatch either direct / promote fault.
1847  */
1848 
1849 static int
1850 uvm_fault_lower(
1851 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1852 	struct vm_page **pages)
1853 {
1854 	struct vm_amap *amap __diagused = ufi->entry->aref.ar_amap;
1855 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1856 	struct vm_page *uobjpage;
1857 	int error;
1858 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1859 
1860 	/*
1861 	 * now, if the desired page is not shadowed by the amap and we have
1862 	 * a backing object that does not have a special fault routine, then
1863 	 * we ask (with pgo_get) the object for resident pages that we care
1864 	 * about and attempt to map them in.  we do not let pgo_get block
1865 	 * (PGO_LOCKED).
1866 	 */
1867 
1868 	if (uobj == NULL) {
1869 		/* zero fill; don't care neighbor pages */
1870 		uobjpage = NULL;
1871 	} else {
1872 		uvm_fault_lower_lookup(ufi, flt, pages);
1873 		uobjpage = pages[flt->centeridx];
1874 	}
1875 
1876 	/*
1877 	 * note that at this point we are done with any front or back pages.
1878 	 * we are now going to focus on the center page (i.e. the one we've
1879 	 * faulted on).  if we have faulted on the upper (anon) layer
1880 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1881 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1882 	 * layer [i.e. case 2] and the page was both present and available,
1883 	 * then we've got a pointer to it as "uobjpage" and we've already
1884 	 * made it BUSY.
1885 	 */
1886 
1887 	/*
1888 	 * locked:
1889 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1890 	 */
1891 	KASSERT(amap == NULL ||
1892 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1893 	if (flt->lower_lock_type == RW_WRITER) {
1894 		KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1895 		KASSERTMSG(uobjpage == NULL ||
1896 		    (uobjpage->flags & PG_BUSY) != 0,
1897 		    "page %p should be busy", uobjpage);
1898 	} else {
1899 		KASSERT(uobj == NULL || rw_read_held(uobj->vmobjlock));
1900 		KASSERTMSG(uobjpage == NULL ||
1901 		    (uobjpage->flags & PG_BUSY) == 0,
1902 		    "page %p should not be busy", uobjpage);
1903 	}
1904 
1905 	/*
1906 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1907 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1908 	 * have a backing object, check and see if we are going to promote
1909 	 * the data up to an anon during the fault.
1910 	 */
1911 
1912 	if (uobj == NULL) {
1913 		uobjpage = PGO_DONTCARE;
1914 		flt->promote = true;		/* always need anon here */
1915 	} else {
1916 		KASSERT(uobjpage != PGO_DONTCARE);
1917 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1918 	}
1919 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%jd, zfill=%jd",
1920 	    flt->promote, (uobj == NULL), 0,0);
1921 
1922 	/*
1923 	 * if uobjpage is not null then we do not need to do I/O to get the
1924 	 * uobjpage.
1925 	 *
1926 	 * if uobjpage is null, then we need to unlock and ask the pager to
1927 	 * get the data for us.   once we have the data, we need to reverify
1928 	 * the state the world.   we are currently not holding any resources.
1929 	 */
1930 
1931 	if (uobjpage) {
1932 		/* update rusage counters */
1933 		curlwp->l_ru.ru_minflt++;
1934 	} else {
1935 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1936 		if (error != 0)
1937 			return error;
1938 	}
1939 
1940 	/*
1941 	 * locked:
1942 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1943 	 */
1944 	KASSERT(amap == NULL ||
1945 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
1946 	if (flt->lower_lock_type == RW_WRITER) {
1947 		KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1948 		KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1949 	} else {
1950 		KASSERT(uobj == NULL || rw_read_held(uobj->vmobjlock));
1951 		KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) == 0);
1952 	}
1953 
1954 	/*
1955 	 * notes:
1956 	 *  - at this point uobjpage can not be NULL
1957 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1958 	 *  for it above)
1959 	 *  - at this point uobjpage could be waited on (handle later)
1960 	 */
1961 
1962 	KASSERT(uobjpage != NULL);
1963 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1964 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1965 	    uvm_pagegetdirty(uobjpage) == UVM_PAGE_STATUS_CLEAN);
1966 
1967 	if (!flt->promote) {
1968 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1969 	} else {
1970 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1971 	}
1972 	return error;
1973 }
1974 
1975 /*
1976  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1977  *
1978  *	1. get on-memory pages.
1979  *	2. if failed, give up (get only center page later).
1980  *	3. if succeeded, enter h/w mapping of neighbor pages.
1981  */
1982 
1983 static void
1984 uvm_fault_lower_lookup(
1985 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1986 	struct vm_page **pages)
1987 {
1988 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1989 	int lcv, gotpages;
1990 	vaddr_t currva;
1991 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1992 
1993 	rw_enter(uobj->vmobjlock, flt->lower_lock_type);
1994 
1995 	/*
1996 	 * Locked: maps(read), amap(if there), uobj
1997 	 *
1998 	 * if we have a read lock on the object, do a PGO_NOBUSY get, which
1999 	 * will return us pages with PG_BUSY clear.  if a write lock is held
2000 	 * pages will be returned with PG_BUSY set.
2001 	 */
2002 
2003 	cpu_count(CPU_COUNT_FLTLGET, 1);
2004 	gotpages = flt->npages;
2005 	(void) uobj->pgops->pgo_get(uobj,
2006 	    ufi->entry->offset + flt->startva - ufi->entry->start,
2007 	    pages, &gotpages, flt->centeridx,
2008 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice,
2009 	    PGO_LOCKED | (flt->lower_lock_type == RW_WRITER ? 0 : PGO_NOBUSY));
2010 
2011 	KASSERT(rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
2012 
2013 	/*
2014 	 * check for pages to map, if we got any
2015 	 */
2016 
2017 	if (gotpages == 0) {
2018 		pages[flt->centeridx] = NULL;
2019 		return;
2020 	}
2021 
2022 	currva = flt->startva;
2023 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
2024 		struct vm_page *curpg;
2025 
2026 		curpg = pages[lcv];
2027 		if (curpg == NULL || curpg == PGO_DONTCARE) {
2028 			continue;
2029 		}
2030 		KASSERT(curpg->uobject == uobj);
2031 
2032 		if (flt->lower_lock_type == RW_WRITER) {
2033 			KASSERT(rw_write_held(uobj->vmobjlock));
2034 			KASSERTMSG((curpg->flags & PG_BUSY) != 0,
2035 			    "page %p should be busy", curpg);
2036 		} else {
2037 			KASSERT(rw_read_held(uobj->vmobjlock));
2038 			KASSERTMSG((curpg->flags & PG_BUSY) == 0,
2039 			    "page %p should not be busy", curpg);
2040 		}
2041 
2042 		/*
2043 		 * if center page is resident and not PG_BUSY|PG_RELEASED
2044 		 * and !PGO_NOBUSY, then pgo_get made it PG_BUSY for us and
2045 		 * gave us a handle to it.
2046 		 */
2047 
2048 		if (lcv == flt->centeridx) {
2049 			UVMHIST_LOG(maphist, "  got uobjpage (%#jx) "
2050 			    "with locked get", (uintptr_t)curpg, 0, 0, 0);
2051 		} else {
2052 			uvm_fault_lower_neighbor(ufi, flt, currva, curpg);
2053 		}
2054 	}
2055 	pmap_update(ufi->orig_map->pmap);
2056 }
2057 
2058 /*
2059  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
2060  */
2061 
2062 static void
2063 uvm_fault_lower_neighbor(
2064 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2065 	vaddr_t currva, struct vm_page *pg)
2066 {
2067 	const bool readonly = uvm_pagereadonly_p(pg) || pg->loan_count > 0;
2068 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
2069 
2070 	/* locked: maps(read), amap(if there), uobj */
2071 
2072 	/*
2073 	 * calling pgo_get with PGO_LOCKED returns us pages which
2074 	 * are neither busy nor released, so we don't need to check
2075 	 * for this.  we can just directly enter the pages.
2076 	 */
2077 
2078 	/*
2079 	 * in the read-locked case, it's not possible for this to be a new
2080 	 * page.  it must be cached with the object and enqueued already.
2081 	 * there wasn't a direct fault on the page, so avoid the cost of
2082 	 * re-enqueuing it.
2083 	 */
2084 
2085 	if (flt->lower_lock_type == RW_WRITER) {
2086 		uvm_pagelock(pg);
2087 		uvm_pageenqueue(pg);
2088 		uvm_pageunlock(pg);
2089 	}
2090 	UVMHIST_LOG(maphist,
2091 	    "  MAPPING: n obj: pm=%#jx, va=%#jx, pg=%#jx",
2092 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
2093 	cpu_count(CPU_COUNT_FLTNOMAP, 1);
2094 
2095 	/*
2096 	 * Since this page isn't the page that's actually faulting,
2097 	 * ignore pmap_enter() failures; it's not critical that we
2098 	 * enter these right now.
2099 	 * NOTE: page can't be waited on or PG_RELEASED because we've
2100 	 * held the lock the whole time we've had the handle.
2101 	 */
2102 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
2103 	KASSERT((pg->flags & PG_RELEASED) == 0);
2104 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) ||
2105 	    uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_CLEAN);
2106 
2107 	/*
2108 	 * if a write lock was held on the object, the pages have been
2109 	 * busied.  unbusy them now, as we are about to enter and then
2110 	 * forget about them.
2111 	 */
2112 
2113 	if (flt->lower_lock_type == RW_WRITER) {
2114 		KASSERT((pg->flags & PG_BUSY) != 0);
2115 		pg->flags &= ~(PG_BUSY);
2116 		UVM_PAGE_OWN(pg, NULL);
2117 	} else {
2118 		KASSERT((pg->flags & PG_BUSY) == 0);
2119 	}
2120 	KASSERT(rw_lock_op(pg->uobject->vmobjlock) == flt->lower_lock_type);
2121 
2122 	const vm_prot_t mapprot =
2123 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
2124 	    flt->enter_prot & MASK(ufi->entry);
2125 	const u_int mapflags =
2126 	    PMAP_CANFAIL | (flt->wire_mapping ? (mapprot | PMAP_WIRED) : 0);
2127 	(void) pmap_enter(ufi->orig_map->pmap, currva,
2128 	    VM_PAGE_TO_PHYS(pg), mapprot, mapflags);
2129 }
2130 
2131 /*
2132  * uvm_fault_lower_io: get lower page from backing store.
2133  *
2134  *	1. unlock everything, because i/o will block.
2135  *	2. call pgo_get.
2136  *	3. if failed, recover.
2137  *	4. if succeeded, relock everything and verify things.
2138  */
2139 
2140 static int
2141 uvm_fault_lower_io(
2142 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2143 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
2144 {
2145 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2146 	struct uvm_object *uobj = *ruobj;
2147 	struct vm_page *pg;
2148 	bool locked;
2149 	int gotpages;
2150 	int error;
2151 	voff_t uoff;
2152 	vm_prot_t access_type;
2153 	int advice;
2154 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
2155 
2156 	/* update rusage counters */
2157 	curlwp->l_ru.ru_majflt++;
2158 
2159 	/* grab everything we need from the entry before we unlock */
2160 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
2161 	access_type = flt->access_type & MASK(ufi->entry);
2162 	advice = ufi->entry->advice;
2163 
2164 	/* Locked: maps(read), amap(if there), uobj */
2165 	KASSERT(rw_lock_op(uobj->vmobjlock) == flt->lower_lock_type);
2166 
2167 	/* Upgrade to a write lock if needed. */
2168 	error = uvm_fault_lower_upgrade(ufi, flt, amap, uobj, NULL);
2169 	if (error != 0) {
2170 		return error;
2171 	}
2172 	uvmfault_unlockall(ufi, amap, NULL);
2173 
2174 	/* Locked: uobj(write) */
2175 	KASSERT(rw_write_held(uobj->vmobjlock));
2176 
2177 	cpu_count(CPU_COUNT_FLTGET, 1);
2178 	gotpages = 1;
2179 	pg = NULL;
2180 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
2181 	    0, access_type, advice, PGO_SYNCIO);
2182 	/* locked: pg(if no error) */
2183 
2184 	/*
2185 	 * recover from I/O
2186 	 */
2187 
2188 	if (error) {
2189 		if (error == EAGAIN) {
2190 			UVMHIST_LOG(maphist,
2191 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
2192 			kpause("fltagain2", false, hz/2, NULL);
2193 			return ERESTART;
2194 		}
2195 
2196 #if 0
2197 		KASSERT(error != ERESTART);
2198 #else
2199 		/* XXXUEBS don't re-fault? */
2200 		if (error == ERESTART)
2201 			error = EIO;
2202 #endif
2203 
2204 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %jd)",
2205 		    error, 0,0,0);
2206 		return error;
2207 	}
2208 
2209 	/*
2210 	 * re-verify the state of the world by first trying to relock
2211 	 * the maps.  always relock the object.
2212 	 */
2213 
2214 	locked = uvmfault_relock(ufi);
2215 	if (locked && amap)
2216 		amap_lock(amap, flt->upper_lock_type);
2217 
2218 	/* might be changed */
2219 	uobj = pg->uobject;
2220 
2221 	rw_enter(uobj->vmobjlock, flt->lower_lock_type);
2222 	KASSERT((pg->flags & PG_BUSY) != 0);
2223 	KASSERT(flt->lower_lock_type == RW_WRITER);
2224 
2225 	uvm_pagelock(pg);
2226 	uvm_pageactivate(pg);
2227 	uvm_pageunlock(pg);
2228 
2229 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
2230 	/* locked(!locked): uobj, pg */
2231 
2232 	/*
2233 	 * verify that the page has not be released and re-verify
2234 	 * that amap slot is still free.   if there is a problem,
2235 	 * we unlock and clean up.
2236 	 */
2237 
2238 	if ((pg->flags & PG_RELEASED) != 0 ||
2239 	    (locked && amap && amap_lookup(&ufi->entry->aref,
2240 	      ufi->orig_rvaddr - ufi->entry->start))) {
2241 		if (locked)
2242 			uvmfault_unlockall(ufi, amap, NULL);
2243 		locked = false;
2244 	}
2245 
2246 	/*
2247 	 * didn't get the lock?   release the page and retry.
2248 	 */
2249 
2250 	if (locked == false) {
2251 		UVMHIST_LOG(maphist,
2252 		    "  wasn't able to relock after fault: retry",
2253 		    0,0,0,0);
2254 		if ((pg->flags & PG_RELEASED) == 0) {
2255 			pg->flags &= ~PG_BUSY;
2256 			uvm_pagelock(pg);
2257 			uvm_pagewakeup(pg);
2258 			uvm_pageunlock(pg);
2259 			UVM_PAGE_OWN(pg, NULL);
2260 		} else {
2261 			cpu_count(CPU_COUNT_FLTPGRELE, 1);
2262 			uvm_pagefree(pg);
2263 		}
2264 		rw_exit(uobj->vmobjlock);
2265 		return ERESTART;
2266 	}
2267 
2268 	/*
2269 	 * we have the data in pg which is busy and
2270 	 * not released.  we are holding object lock (so the page
2271 	 * can't be released on us).
2272 	 */
2273 
2274 	/* locked: maps(read), amap(if !null), uobj, pg */
2275 
2276 	*ruobj = uobj;
2277 	*ruobjpage = pg;
2278 	return 0;
2279 }
2280 
2281 /*
2282  * uvm_fault_lower_direct: fault lower center page
2283  *
2284  *	1. adjust flt->enter_prot.
2285  *	2. if page is loaned, resolve.
2286  */
2287 
2288 int
2289 uvm_fault_lower_direct(
2290 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2291 	struct uvm_object *uobj, struct vm_page *uobjpage)
2292 {
2293 	struct vm_page *pg;
2294 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
2295 
2296 	/*
2297 	 * we are not promoting.   if the mapping is COW ensure that we
2298 	 * don't give more access than we should (e.g. when doing a read
2299 	 * fault on a COPYONWRITE mapping we want to map the COW page in
2300 	 * R/O even though the entry protection could be R/W).
2301 	 *
2302 	 * set "pg" to the page we want to map in (uobjpage, usually)
2303 	 */
2304 
2305 	cpu_count(CPU_COUNT_FLT_OBJ, 1);
2306 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
2307 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
2308 		flt->enter_prot &= ~VM_PROT_WRITE;
2309 	pg = uobjpage;		/* map in the actual object */
2310 
2311 	KASSERT(uobjpage != PGO_DONTCARE);
2312 
2313 	/*
2314 	 * we are faulting directly on the page.   be careful
2315 	 * about writing to loaned pages...
2316 	 */
2317 
2318 	if (uobjpage->loan_count) {
2319 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
2320 	}
2321 	KASSERT(pg == uobjpage);
2322 
2323 	if (flt->lower_lock_type == RW_READER) {
2324 		KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) == 0);
2325 	} else {
2326 		KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2327 	}
2328 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
2329 }
2330 
2331 /*
2332  * uvm_fault_lower_direct_loan: resolve loaned page.
2333  *
2334  *	1. if not cow'ing, adjust flt->enter_prot.
2335  *	2. if cow'ing, break loan.
2336  */
2337 
2338 static int
2339 uvm_fault_lower_direct_loan(
2340 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2341 	struct uvm_object *uobj, struct vm_page **rpg,
2342 	struct vm_page **ruobjpage)
2343 {
2344 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2345 	struct vm_page *pg;
2346 	struct vm_page *uobjpage = *ruobjpage;
2347 	int error;
2348 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
2349 
2350 	if (!flt->cow_now) {
2351 		/* read fault: cap the protection at readonly */
2352 		/* cap! */
2353 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2354 	} else {
2355 		/*
2356 		 * write fault: must break the loan here.  to do this
2357 		 * we need a write lock on the object.
2358 		 */
2359 
2360 		error = uvm_fault_lower_upgrade(ufi, flt, amap, uobj, uobjpage);
2361 		if (error != 0) {
2362 			return error;
2363 		}
2364 		KASSERT(rw_write_held(uobj->vmobjlock));
2365 
2366 		pg = uvm_loanbreak(uobjpage);
2367 		if (pg == NULL) {
2368 
2369 			/*
2370 			 * drop ownership of page, it can't be released
2371 			 */
2372 
2373 			uvm_pagelock(uobjpage);
2374 			uvm_pagewakeup(uobjpage);
2375 			uvm_pageunlock(uobjpage);
2376 			uobjpage->flags &= ~PG_BUSY;
2377 			UVM_PAGE_OWN(uobjpage, NULL);
2378 
2379 			uvmfault_unlockall(ufi, amap, uobj);
2380 			UVMHIST_LOG(maphist,
2381 			  "  out of RAM breaking loan, waiting",
2382 			  0,0,0,0);
2383 			cpu_count(CPU_COUNT_FLTNORAM, 1);
2384 			uvm_wait("flt_noram4");
2385 			return ERESTART;
2386 		}
2387 		*rpg = pg;
2388 		*ruobjpage = pg;
2389 	}
2390 	return 0;
2391 }
2392 
2393 /*
2394  * uvm_fault_lower_promote: promote lower page.
2395  *
2396  *	1. call uvmfault_promote.
2397  *	2. fill in data.
2398  *	3. if not ZFOD, dispose old page.
2399  */
2400 
2401 int
2402 uvm_fault_lower_promote(
2403 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2404 	struct uvm_object *uobj, struct vm_page *uobjpage)
2405 {
2406 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2407 	struct vm_anon *anon;
2408 	struct vm_page *pg;
2409 	int error;
2410 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2411 
2412 	KASSERT(amap != NULL);
2413 
2414 	/* promoting requires a write lock. */
2415 	error = uvm_fault_upper_upgrade(ufi, flt, amap, uobj);
2416 	if (error != 0) {
2417 		return error;
2418 	}
2419 	KASSERT(rw_write_held(amap->am_lock));
2420 
2421 	/*
2422 	 * If we are going to promote the data to an anon we
2423 	 * allocate a blank anon here and plug it into our amap.
2424 	 */
2425 	error = uvmfault_promote(ufi, NULL, uobjpage, &anon, &flt->anon_spare);
2426 	switch (error) {
2427 	case 0:
2428 		break;
2429 	case ERESTART:
2430 		return ERESTART;
2431 	default:
2432 		return error;
2433 	}
2434 
2435 	pg = anon->an_page;
2436 
2437 	/*
2438 	 * Fill in the data.
2439 	 */
2440 	if (flt->lower_lock_type == RW_READER) {
2441 		KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) == 0);
2442 	} else {
2443 		KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2444 	}
2445 
2446 	if (uobjpage != PGO_DONTCARE) {
2447 		cpu_count(CPU_COUNT_FLT_PRCOPY, 1);
2448 
2449 		/*
2450 		 * promote to shared amap?  make sure all sharing
2451 		 * procs see it
2452 		 */
2453 
2454 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2455 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2456 			/*
2457 			 * XXX: PAGE MIGHT BE WIRED!
2458 			 */
2459 		}
2460 
2461 		/*
2462 		 * dispose of uobjpage.  it can't be PG_RELEASED
2463 		 * since we still hold the object lock.
2464 		 */
2465 
2466 		if ((uobjpage->flags & PG_BUSY) != 0) {
2467 			uobjpage->flags &= ~PG_BUSY;
2468 			uvm_pagelock(uobjpage);
2469 			uvm_pagewakeup(uobjpage);
2470 			uvm_pageunlock(uobjpage);
2471 			UVM_PAGE_OWN(uobjpage, NULL);
2472 		}
2473 
2474 		UVMHIST_LOG(maphist,
2475 		    "  promote uobjpage %#jx to anon/page %#jx/%#jx",
2476 		    (uintptr_t)uobjpage, (uintptr_t)anon, (uintptr_t)pg, 0);
2477 
2478 	} else {
2479 		cpu_count(CPU_COUNT_FLT_PRZERO, 1);
2480 
2481 		/*
2482 		 * Page is zero'd and marked dirty by
2483 		 * uvmfault_promote().
2484 		 */
2485 
2486 		UVMHIST_LOG(maphist,"  zero fill anon/page %#jx/%#jx",
2487 		    (uintptr_t)anon, (uintptr_t)pg, 0, 0);
2488 	}
2489 
2490 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2491 }
2492 
2493 /*
2494  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2495  * from the lower page.
2496  */
2497 
2498 int
2499 uvm_fault_lower_enter(
2500 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2501 	struct uvm_object *uobj,
2502 	struct vm_anon *anon, struct vm_page *pg)
2503 {
2504 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2505 	const bool readonly = uvm_pagereadonly_p(pg);
2506 	int error;
2507 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2508 
2509 	/*
2510 	 * Locked:
2511 	 *
2512 	 *	maps(read), amap(if !null), uobj(if !null),
2513 	 *	anon(if !null), pg(if anon), unlock_uobj(if !null)
2514 	 *
2515 	 * anon must be write locked (promotion).  uobj can be either.
2516 	 *
2517 	 * Note: pg is either the uobjpage or the new page in the new anon.
2518 	 */
2519 	KASSERT(amap == NULL ||
2520 	    rw_lock_op(amap->am_lock) == flt->upper_lock_type);
2521 	KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
2522 	if (flt->lower_lock_type == RW_WRITER) {
2523 		KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
2524 		KASSERTMSG((pg->flags & PG_BUSY) != 0,
2525 		    "page %p should be busy", pg);
2526 	} else {
2527 		KASSERT(uobj == NULL || rw_read_held(uobj->vmobjlock));
2528 		KASSERTMSG(anon != NULL || (pg->flags & PG_BUSY) == 0,
2529 		    "page %p should not be busy", pg);
2530 	}
2531 
2532 	/*
2533 	 * all resources are present.   we can now map it in and free our
2534 	 * resources.
2535 	 */
2536 
2537 	UVMHIST_LOG(maphist,
2538 	    "  MAPPING: case2: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
2539 	    (uintptr_t)ufi->orig_map->pmap, ufi->orig_rvaddr,
2540 	    (uintptr_t)pg, flt->promote);
2541 	KASSERTMSG((flt->access_type & VM_PROT_WRITE) == 0 || !readonly,
2542 	    "promote=%u cow_now=%u access_type=%x enter_prot=%x cow=%u "
2543 	    "entry=%p map=%p orig_rvaddr=%p pg=%p",
2544 	    flt->promote, flt->cow_now, flt->access_type, flt->enter_prot,
2545 	    UVM_ET_ISCOPYONWRITE(ufi->entry), ufi->entry, ufi->orig_map,
2546 	    (void *)ufi->orig_rvaddr, pg);
2547 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 || !readonly);
2548 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2549 	    VM_PAGE_TO_PHYS(pg),
2550 	    readonly ? flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2551 	    flt->access_type | PMAP_CANFAIL |
2552 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2553 
2554 		/*
2555 		 * No need to undo what we did; we can simply think of
2556 		 * this as the pmap throwing away the mapping information.
2557 		 *
2558 		 * We do, however, have to go through the ReFault path,
2559 		 * as the map may change while we're asleep.
2560 		 */
2561 
2562 		/*
2563 		 * ensure that the page is queued in the case that
2564 		 * we just promoted the page.
2565 		 */
2566 
2567 		if (anon != NULL || flt->lower_lock_type == RW_WRITER) {
2568 			uvm_pagelock(pg);
2569 			uvm_pageenqueue(pg);
2570 			uvm_pagewakeup(pg);
2571 			uvm_pageunlock(pg);
2572 		} else {
2573 			KASSERT((pg->flags & PG_BUSY) == 0);
2574 		}
2575 
2576 		/*
2577 		 * note that pg can't be PG_RELEASED since we did not drop
2578 		 * the object lock since the last time we checked.
2579 		 */
2580 		KASSERT((pg->flags & PG_RELEASED) == 0);
2581 		if ((pg->flags & PG_BUSY) != 0) {
2582 			pg->flags &= ~(PG_BUSY|PG_FAKE);
2583 			UVM_PAGE_OWN(pg, NULL);
2584 		}
2585 
2586 		uvmfault_unlockall(ufi, amap, uobj);
2587 		if (!uvm_reclaimable()) {
2588 			UVMHIST_LOG(maphist,
2589 			    "<- failed.  out of VM",0,0,0,0);
2590 			/* XXX instrumentation */
2591 			error = ENOMEM;
2592 			return error;
2593 		}
2594 		/* XXX instrumentation */
2595 		uvm_wait("flt_pmfail2");
2596 		return ERESTART;
2597 	}
2598 
2599 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2600 
2601 	/*
2602 	 * note that pg can't be PG_RELEASED since we did not drop the object
2603 	 * lock since the last time we checked.
2604 	 */
2605 	KASSERT((pg->flags & PG_RELEASED) == 0);
2606 	if ((pg->flags & PG_BUSY) != 0) {
2607 		uvm_pagelock(pg);
2608 		uvm_pagewakeup(pg);
2609 		uvm_pageunlock(pg);
2610 		pg->flags &= ~(PG_BUSY|PG_FAKE);
2611 		UVM_PAGE_OWN(pg, NULL);
2612 	}
2613 
2614 	pmap_update(ufi->orig_map->pmap);
2615 	uvmfault_unlockall(ufi, amap, uobj);
2616 
2617 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2618 	return 0;
2619 }
2620 
2621 /*
2622  * uvm_fault_lower_done: queue lower center page.
2623  */
2624 
2625 void
2626 uvm_fault_lower_done(
2627 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2628 	struct uvm_object *uobj, struct vm_page *pg)
2629 {
2630 	bool dropswap = false;
2631 
2632 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2633 
2634 	uvm_pagelock(pg);
2635 	if (flt->wire_paging) {
2636 		uvm_pagewire(pg);
2637 		if (pg->flags & PG_AOBJ) {
2638 
2639 			/*
2640 			 * since the now-wired page cannot be paged out,
2641 			 * release its swap resources for others to use.
2642 			 * since an aobj page with no swap cannot be clean,
2643 			 * mark it dirty now.
2644 			 */
2645 
2646 			KASSERT(uobj != NULL);
2647 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
2648 			dropswap = true;
2649 		}
2650 	} else {
2651 		uvm_pageactivate(pg);
2652 	}
2653 	uvm_pageunlock(pg);
2654 
2655 	if (dropswap) {
2656 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2657 	}
2658 }
2659 
2660 
2661 /*
2662  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2663  *
2664  * => map may be read-locked by caller, but MUST NOT be write-locked.
2665  * => if map is read-locked, any operations which may cause map to
2666  *	be write-locked in uvm_fault() must be taken care of by
2667  *	the caller.  See uvm_map_pageable().
2668  */
2669 
2670 int
2671 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2672     vm_prot_t access_type, int maxprot)
2673 {
2674 	vaddr_t va;
2675 	int error;
2676 
2677 	/*
2678 	 * now fault it in a page at a time.   if the fault fails then we have
2679 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2680 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2681 	 */
2682 
2683 	/*
2684 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2685 	 * wiring the last page in the address space, though.
2686 	 */
2687 	if (start > end) {
2688 		return EFAULT;
2689 	}
2690 
2691 	for (va = start; va < end; va += PAGE_SIZE) {
2692 		error = uvm_fault_internal(map, va, access_type,
2693 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2694 		if (error) {
2695 			if (va != start) {
2696 				uvm_fault_unwire(map, start, va);
2697 			}
2698 			return error;
2699 		}
2700 	}
2701 	return 0;
2702 }
2703 
2704 /*
2705  * uvm_fault_unwire(): unwire range of virtual space.
2706  */
2707 
2708 void
2709 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2710 {
2711 	vm_map_lock_read(map);
2712 	uvm_fault_unwire_locked(map, start, end);
2713 	vm_map_unlock_read(map);
2714 }
2715 
2716 /*
2717  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2718  *
2719  * => map must be at least read-locked.
2720  */
2721 
2722 void
2723 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2724 {
2725 	struct vm_map_entry *entry, *oentry;
2726 	pmap_t pmap = vm_map_pmap(map);
2727 	vaddr_t va;
2728 	paddr_t pa;
2729 	struct vm_page *pg;
2730 
2731 	/*
2732 	 * we assume that the area we are unwiring has actually been wired
2733 	 * in the first place.   this means that we should be able to extract
2734 	 * the PAs from the pmap.   we also lock out the page daemon so that
2735 	 * we can call uvm_pageunwire.
2736 	 */
2737 
2738 	/*
2739 	 * find the beginning map entry for the region.
2740 	 */
2741 
2742 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2743 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2744 		panic("uvm_fault_unwire_locked: address not in map");
2745 
2746 	oentry = NULL;
2747 	for (va = start; va < end; va += PAGE_SIZE) {
2748 
2749 		/*
2750 		 * find the map entry for the current address.
2751 		 */
2752 
2753 		KASSERT(va >= entry->start);
2754 		while (va >= entry->end) {
2755 			KASSERT(entry->next != &map->header &&
2756 				entry->next->start <= entry->end);
2757 			entry = entry->next;
2758 		}
2759 
2760 		/*
2761 		 * lock it.
2762 		 */
2763 
2764 		if (entry != oentry) {
2765 			if (oentry != NULL) {
2766 				uvm_map_unlock_entry(oentry);
2767 			}
2768 			uvm_map_lock_entry(entry, RW_WRITER);
2769 			oentry = entry;
2770 		}
2771 
2772 		/*
2773 		 * if the entry is no longer wired, tell the pmap.
2774 		 */
2775 
2776 		if (!pmap_extract(pmap, va, &pa))
2777 			continue;
2778 
2779 		if (VM_MAPENT_ISWIRED(entry) == 0)
2780 			pmap_unwire(pmap, va);
2781 
2782 		pg = PHYS_TO_VM_PAGE(pa);
2783 		if (pg) {
2784 			uvm_pagelock(pg);
2785 			uvm_pageunwire(pg);
2786 			uvm_pageunlock(pg);
2787 		}
2788 	}
2789 
2790 	if (oentry != NULL) {
2791 		uvm_map_unlock_entry(entry);
2792 	}
2793 }
2794