xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 33881f779a77dce6440bdc44610d94de75bebefe)
1 /*	$NetBSD: uvm_fault.c,v 1.219 2020/03/17 18:31:39 ad 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 1.219 2020/03/17 18:31:39 ad 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_write_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 	int error;
280 
281 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
282 	KASSERT(rw_write_held(anon->an_lock));
283 	KASSERT(anon->an_lock == amap->am_lock);
284 
285 	/* Increment the counters.*/
286 	cpu_count(CPU_COUNT_FLTANGET, 1);
287 	if (anon->an_page) {
288 		curlwp->l_ru.ru_minflt++;
289 	} else {
290 		curlwp->l_ru.ru_majflt++;
291 	}
292 	error = 0;
293 
294 	/*
295 	 * Loop until we get the anon data, or fail.
296 	 */
297 
298 	for (;;) {
299 		bool we_own, locked;
300 		/*
301 		 * Note: 'we_own' will become true if we set PG_BUSY on a page.
302 		 */
303 		we_own = false;
304 		pg = anon->an_page;
305 
306 		/*
307 		 * If there is a resident page and it is loaned, then anon
308 		 * may not own it.  Call out to uvm_anon_lockloanpg() to
309 		 * identify and lock the real owner of the page.
310 		 */
311 
312 		if (pg && pg->loan_count)
313 			pg = uvm_anon_lockloanpg(anon);
314 
315 		/*
316 		 * Is page resident?  Make sure it is not busy/released.
317 		 */
318 
319 		if (pg) {
320 
321 			/*
322 			 * at this point, if the page has a uobject [meaning
323 			 * we have it on loan], then that uobject is locked
324 			 * by us!   if the page is busy, we drop all the
325 			 * locks (including uobject) and try again.
326 			 */
327 
328 			if ((pg->flags & PG_BUSY) == 0) {
329 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
330 				return 0;
331 			}
332 			cpu_count(CPU_COUNT_FLTPGWAIT, 1);
333 
334 			/*
335 			 * The last unlock must be an atomic unlock and wait
336 			 * on the owner of page.
337 			 */
338 
339 			if (pg->uobject) {
340 				/* Owner of page is UVM object. */
341 				uvmfault_unlockall(ufi, amap, NULL);
342 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
343 				    0,0,0);
344 				uvm_pagewait(pg, pg->uobject->vmobjlock, "anonget1");
345 			} else {
346 				/* Owner of page is anon. */
347 				uvmfault_unlockall(ufi, NULL, NULL);
348 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
349 				    0,0,0);
350 				uvm_pagewait(pg, anon->an_lock, "anonget2");
351 			}
352 		} else {
353 #if defined(VMSWAP)
354 			/*
355 			 * No page, therefore allocate one.
356 			 */
357 
358 			pg = uvm_pagealloc(NULL,
359 			    ufi != NULL ? ufi->orig_rvaddr : 0,
360 			    anon, ufi != NULL ? UVM_FLAG_COLORMATCH : 0);
361 			if (pg == NULL) {
362 				/* Out of memory.  Wait a little. */
363 				uvmfault_unlockall(ufi, amap, NULL);
364 				cpu_count(CPU_COUNT_FLTNORAM, 1);
365 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
366 				    0,0,0);
367 				if (!uvm_reclaimable()) {
368 					return ENOMEM;
369 				}
370 				uvm_wait("flt_noram1");
371 			} else {
372 				/* PG_BUSY bit is set. */
373 				we_own = true;
374 				uvmfault_unlockall(ufi, amap, NULL);
375 
376 				/*
377 				 * Pass a PG_BUSY+PG_FAKE clean page into
378 				 * the uvm_swap_get() function with all data
379 				 * structures unlocked.  Note that it is OK
380 				 * to read an_swslot here, because we hold
381 				 * PG_BUSY on the page.
382 				 */
383 				cpu_count(CPU_COUNT_PAGEINS, 1);
384 				error = uvm_swap_get(pg, anon->an_swslot,
385 				    PGO_SYNCIO);
386 
387 				/*
388 				 * We clean up after the I/O below in the
389 				 * 'we_own' case.
390 				 */
391 			}
392 #else
393 			panic("%s: no page", __func__);
394 #endif /* defined(VMSWAP) */
395 		}
396 
397 		/*
398 		 * Re-lock the map and anon.
399 		 */
400 
401 		locked = uvmfault_relock(ufi);
402 		if (locked || we_own) {
403 			rw_enter(anon->an_lock, RW_WRITER);
404 		}
405 
406 		/*
407 		 * If we own the page (i.e. we set PG_BUSY), then we need
408 		 * to clean up after the I/O.  There are three cases to
409 		 * consider:
410 		 *
411 		 * 1) Page was released during I/O: free anon and ReFault.
412 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
413 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
414 		 *    case (i.e. drop anon lock if not locked).
415 		 */
416 
417 		if (we_own) {
418 #if defined(VMSWAP)
419 			if (error) {
420 
421 				/*
422 				 * Remove the swap slot from the anon and
423 				 * mark the anon as having no real slot.
424 				 * Do not free the swap slot, thus preventing
425 				 * it from being used again.
426 				 */
427 
428 				if (anon->an_swslot > 0) {
429 					uvm_swap_markbad(anon->an_swslot, 1);
430 				}
431 				anon->an_swslot = SWSLOT_BAD;
432 
433 				if ((pg->flags & PG_RELEASED) != 0) {
434 					goto released;
435 				}
436 
437 				/*
438 				 * Note: page was never !PG_BUSY, so it
439 				 * cannot be mapped and thus no need to
440 				 * pmap_page_protect() it.
441 				 */
442 
443 				uvm_pagefree(pg);
444 
445 				if (locked) {
446 					uvmfault_unlockall(ufi, NULL, NULL);
447 				}
448 				rw_exit(anon->an_lock);
449 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
450 				return error;
451 			}
452 
453 			if ((pg->flags & PG_RELEASED) != 0) {
454 released:
455 				KASSERT(anon->an_ref == 0);
456 
457 				/*
458 				 * Released while we had unlocked amap.
459 				 */
460 
461 				if (locked) {
462 					uvmfault_unlockall(ufi, NULL, NULL);
463 				}
464 				uvm_anon_release(anon);
465 
466 				if (error) {
467 					UVMHIST_LOG(maphist,
468 					    "<- ERROR/RELEASED", 0,0,0,0);
469 					return error;
470 				}
471 
472 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
473 				return ERESTART;
474 			}
475 
476 			/*
477 			 * We have successfully read the page, activate it.
478 			 */
479 
480 			uvm_pagelock(pg);
481 			uvm_pageactivate(pg);
482 			uvm_pagewakeup(pg);
483 			uvm_pageunlock(pg);
484 			pg->flags &= ~(PG_BUSY|PG_FAKE);
485 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN);
486 			UVM_PAGE_OWN(pg, NULL);
487 #else
488 			panic("%s: we_own", __func__);
489 #endif /* defined(VMSWAP) */
490 		}
491 
492 		/*
493 		 * We were not able to re-lock the map - restart the fault.
494 		 */
495 
496 		if (!locked) {
497 			if (we_own) {
498 				rw_exit(anon->an_lock);
499 			}
500 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
501 			return ERESTART;
502 		}
503 
504 		/*
505 		 * Verify that no one has touched the amap and moved
506 		 * the anon on us.
507 		 */
508 
509 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
510 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
511 
512 			uvmfault_unlockall(ufi, amap, NULL);
513 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
514 			return ERESTART;
515 		}
516 
517 		/*
518 		 * Retry..
519 		 */
520 
521 		cpu_count(CPU_COUNT_FLTANRETRY, 1);
522 		continue;
523 	}
524 	/*NOTREACHED*/
525 }
526 
527 /*
528  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
529  *
530  *	1. allocate an anon and a page.
531  *	2. fill its contents.
532  *	3. put it into amap.
533  *
534  * => if we fail (result != 0) we unlock everything.
535  * => on success, return a new locked anon via 'nanon'.
536  *    (*nanon)->an_page will be a resident, locked, dirty page.
537  * => it's caller's responsibility to put the promoted nanon->an_page to the
538  *    page queue.
539  */
540 
541 static int
542 uvmfault_promote(struct uvm_faultinfo *ufi,
543     struct vm_anon *oanon,
544     struct vm_page *uobjpage,
545     struct vm_anon **nanon, /* OUT: allocated anon */
546     struct vm_anon **spare)
547 {
548 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
549 	struct uvm_object *uobj;
550 	struct vm_anon *anon;
551 	struct vm_page *pg;
552 	struct vm_page *opg;
553 	int error;
554 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
555 
556 	if (oanon) {
557 		/* anon COW */
558 		opg = oanon->an_page;
559 		KASSERT(opg != NULL);
560 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
561 	} else if (uobjpage != PGO_DONTCARE) {
562 		/* object-backed COW */
563 		opg = uobjpage;
564 	} else {
565 		/* ZFOD */
566 		opg = NULL;
567 	}
568 	if (opg != NULL) {
569 		uobj = opg->uobject;
570 	} else {
571 		uobj = NULL;
572 	}
573 
574 	KASSERT(amap != NULL);
575 	KASSERT(uobjpage != NULL);
576 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
577 	KASSERT(rw_write_held(amap->am_lock));
578 	KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock);
579 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
580 
581 	if (*spare != NULL) {
582 		anon = *spare;
583 		*spare = NULL;
584 	} else {
585 		anon = uvm_analloc();
586 	}
587 	if (anon) {
588 
589 		/*
590 		 * The new anon is locked.
591 		 *
592 		 * if opg == NULL, we want a zero'd, dirty page,
593 		 * so have uvm_pagealloc() do that for us.
594 		 */
595 
596 		KASSERT(anon->an_lock == NULL);
597 		anon->an_lock = amap->am_lock;
598 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
599 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
600 		if (pg == NULL) {
601 			anon->an_lock = NULL;
602 		}
603 	} else {
604 		pg = NULL;
605 	}
606 
607 	/*
608 	 * out of memory resources?
609 	 */
610 
611 	if (pg == NULL) {
612 		/* save anon for the next try. */
613 		if (anon != NULL) {
614 			*spare = anon;
615 		}
616 
617 		/* unlock and fail ... */
618 		uvm_page_unbusy(&uobjpage, 1);
619 		uvmfault_unlockall(ufi, amap, uobj);
620 		if (!uvm_reclaimable()) {
621 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
622 			cpu_count(CPU_COUNT_FLTNOANON, 1);
623 			error = ENOMEM;
624 			goto done;
625 		}
626 
627 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
628 		cpu_count(CPU_COUNT_FLTNORAM, 1);
629 		uvm_wait("flt_noram5");
630 		error = ERESTART;
631 		goto done;
632 	}
633 
634 	/* copy page [pg now dirty] */
635 	if (opg) {
636 		uvm_pagecopy(opg, pg);
637 	}
638 	KASSERT(uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_DIRTY);
639 
640 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
641 	    oanon != NULL);
642 
643 	*nanon = anon;
644 	error = 0;
645 done:
646 	return error;
647 }
648 
649 /*
650  * Update statistics after fault resolution.
651  * - maxrss
652  */
653 void
654 uvmfault_update_stats(struct uvm_faultinfo *ufi)
655 {
656 	struct vm_map		*map;
657 	struct vmspace 		*vm;
658 	struct proc		*p;
659 	vsize_t			 res;
660 
661 	map = ufi->orig_map;
662 
663 	p = curproc;
664 	KASSERT(p != NULL);
665 	vm = p->p_vmspace;
666 
667 	if (&vm->vm_map != map)
668 		return;
669 
670 	res = pmap_resident_count(map->pmap);
671 	if (vm->vm_rssmax < res)
672 		vm->vm_rssmax = res;
673 }
674 
675 /*
676  *   F A U L T   -   m a i n   e n t r y   p o i n t
677  */
678 
679 /*
680  * uvm_fault: page fault handler
681  *
682  * => called from MD code to resolve a page fault
683  * => VM data structures usually should be unlocked.   however, it is
684  *	possible to call here with the main map locked if the caller
685  *	gets a write lock, sets it recusive, and then calls us (c.f.
686  *	uvm_map_pageable).   this should be avoided because it keeps
687  *	the map locked off during I/O.
688  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
689  */
690 
691 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
692 			 ~VM_PROT_WRITE : VM_PROT_ALL)
693 
694 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
695 #define UVM_FAULT_WIRE		(1 << 0)
696 #define UVM_FAULT_MAXPROT	(1 << 1)
697 
698 struct uvm_faultctx {
699 
700 	/*
701 	 * the following members are set up by uvm_fault_check() and
702 	 * read-only after that.
703 	 *
704 	 * note that narrow is used by uvm_fault_check() to change
705 	 * the behaviour after ERESTART.
706 	 *
707 	 * most of them might change after RESTART if the underlying
708 	 * map entry has been changed behind us.  an exception is
709 	 * wire_paging, which does never change.
710 	 */
711 	vm_prot_t access_type;
712 	vaddr_t startva;
713 	int npages;
714 	int centeridx;
715 	bool narrow;		/* work on a single requested page only */
716 	bool wire_mapping;	/* request a PMAP_WIRED mapping
717 				   (UVM_FAULT_WIRE or VM_MAPENT_ISWIRED) */
718 	bool wire_paging;	/* request uvm_pagewire
719 				   (true for UVM_FAULT_WIRE) */
720 	bool cow_now;		/* VM_PROT_WRITE is actually requested
721 				   (ie. should break COW and page loaning) */
722 
723 	/*
724 	 * enter_prot is set up by uvm_fault_check() and clamped
725 	 * (ie. drop the VM_PROT_WRITE bit) in various places in case
726 	 * of !cow_now.
727 	 */
728 	vm_prot_t enter_prot;	/* prot at which we want to enter pages in */
729 
730 	/*
731 	 * the following member is for uvmfault_promote() and ERESTART.
732 	 */
733 	struct vm_anon *anon_spare;
734 
735 	/*
736 	 * the folloing is actually a uvm_fault_lower() internal.
737 	 * it's here merely for debugging.
738 	 * (or due to the mechanical separation of the function?)
739 	 */
740 	bool promote;
741 };
742 
743 static inline int	uvm_fault_check(
744 			    struct uvm_faultinfo *, struct uvm_faultctx *,
745 			    struct vm_anon ***, bool);
746 
747 static int		uvm_fault_upper(
748 			    struct uvm_faultinfo *, struct uvm_faultctx *,
749 			    struct vm_anon **);
750 static inline int	uvm_fault_upper_lookup(
751 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
752 			    struct vm_anon **, struct vm_page **);
753 static inline void	uvm_fault_upper_neighbor(
754 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
755 			    vaddr_t, struct vm_page *, bool);
756 static inline int	uvm_fault_upper_loan(
757 			    struct uvm_faultinfo *, struct uvm_faultctx *,
758 			    struct vm_anon *, struct uvm_object **);
759 static inline int	uvm_fault_upper_promote(
760 			    struct uvm_faultinfo *, struct uvm_faultctx *,
761 			    struct uvm_object *, struct vm_anon *);
762 static inline int	uvm_fault_upper_direct(
763 			    struct uvm_faultinfo *, struct uvm_faultctx *,
764 			    struct uvm_object *, struct vm_anon *);
765 static int		uvm_fault_upper_enter(
766 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
767 			    struct uvm_object *, struct vm_anon *,
768 			    struct vm_page *, struct vm_anon *);
769 static inline void	uvm_fault_upper_done(
770 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
771 			    struct vm_anon *, struct vm_page *);
772 
773 static int		uvm_fault_lower(
774 			    struct uvm_faultinfo *, struct uvm_faultctx *,
775 			    struct vm_page **);
776 static inline void	uvm_fault_lower_lookup(
777 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
778 			    struct vm_page **);
779 static inline void	uvm_fault_lower_neighbor(
780 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
781 			    vaddr_t, struct vm_page *);
782 static inline int	uvm_fault_lower_io(
783 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
784 			    struct uvm_object **, struct vm_page **);
785 static inline int	uvm_fault_lower_direct(
786 			    struct uvm_faultinfo *, struct uvm_faultctx *,
787 			    struct uvm_object *, struct vm_page *);
788 static inline int	uvm_fault_lower_direct_loan(
789 			    struct uvm_faultinfo *, struct uvm_faultctx *,
790 			    struct uvm_object *, struct vm_page **,
791 			    struct vm_page **);
792 static inline int	uvm_fault_lower_promote(
793 			    struct uvm_faultinfo *, struct uvm_faultctx *,
794 			    struct uvm_object *, struct vm_page *);
795 static int		uvm_fault_lower_enter(
796 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
797 			    struct uvm_object *,
798 			    struct vm_anon *, struct vm_page *);
799 static inline void	uvm_fault_lower_done(
800 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
801 			    struct uvm_object *, struct vm_page *);
802 
803 int
804 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
805     vm_prot_t access_type, int fault_flag)
806 {
807 	struct uvm_faultinfo ufi;
808 	struct uvm_faultctx flt = {
809 		.access_type = access_type,
810 
811 		/* don't look for neighborhood * pages on "wire" fault */
812 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
813 
814 		/* "wire" fault causes wiring of both mapping and paging */
815 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
816 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
817 	};
818 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
819 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
820 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
821 	int error;
822 
823 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
824 
825 	UVMHIST_LOG(maphist, "(map=%#jx, vaddr=%#jx, at=%jd, ff=%jd)",
826 	      (uintptr_t)orig_map, vaddr, access_type, fault_flag);
827 
828 	/* Don't count anything until user interaction is possible */
829 	kpreempt_disable();
830 	if (__predict_true(start_init_exec)) {
831 		struct cpu_info *ci = curcpu();
832 		CPU_COUNT(CPU_COUNT_NFAULT, 1);
833 		/* Don't flood RNG subsystem with samples. */
834 		if (++(ci->ci_faultrng) == 503) {
835 			ci->ci_faultrng = 0;
836 			rnd_add_uint32(&curcpu()->ci_data.cpu_uvm->rs,
837 			    sizeof(vaddr_t) == sizeof(uint32_t) ?
838 			    (uint32_t)vaddr : sizeof(vaddr_t) ==
839 			    sizeof(uint64_t) ?
840 			    (uint32_t)vaddr :
841 			    (uint32_t)ci->ci_counts[CPU_COUNT_NFAULT]);
842 		}
843 	}
844 	kpreempt_enable();
845 
846 	/*
847 	 * init the IN parameters in the ufi
848 	 */
849 
850 	ufi.orig_map = orig_map;
851 	ufi.orig_rvaddr = trunc_page(vaddr);
852 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
853 
854 	error = ERESTART;
855 	while (error == ERESTART) { /* ReFault: */
856 		anons = anons_store;
857 		pages = pages_store;
858 
859 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
860 		if (error != 0)
861 			continue;
862 
863 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
864 		if (error != 0)
865 			continue;
866 
867 		if (pages[flt.centeridx] == PGO_DONTCARE)
868 			error = uvm_fault_upper(&ufi, &flt, anons);
869 		else {
870 			struct uvm_object * const uobj =
871 			    ufi.entry->object.uvm_obj;
872 
873 			if (uobj && uobj->pgops->pgo_fault != NULL) {
874 				/*
875 				 * invoke "special" fault routine.
876 				 */
877 				rw_enter(uobj->vmobjlock, RW_WRITER);
878 				/* locked: maps(read), amap(if there), uobj */
879 				error = uobj->pgops->pgo_fault(&ufi,
880 				    flt.startva, pages, flt.npages,
881 				    flt.centeridx, flt.access_type,
882 				    PGO_LOCKED|PGO_SYNCIO);
883 
884 				/*
885 				 * locked: nothing, pgo_fault has unlocked
886 				 * everything
887 				 */
888 
889 				/*
890 				 * object fault routine responsible for
891 				 * pmap_update().
892 				 */
893 
894 				/*
895 				 * Wake up the pagedaemon if the fault method
896 				 * failed for lack of memory but some can be
897 				 * reclaimed.
898 				 */
899 				if (error == ENOMEM && uvm_reclaimable()) {
900 					uvm_wait("pgo_fault");
901 					error = ERESTART;
902 				}
903 			} else {
904 				error = uvm_fault_lower(&ufi, &flt, pages);
905 			}
906 		}
907 	}
908 
909 	if (flt.anon_spare != NULL) {
910 		flt.anon_spare->an_ref--;
911 		KASSERT(flt.anon_spare->an_ref == 0);
912 		KASSERT(flt.anon_spare->an_lock == NULL);
913 		uvm_anon_free(flt.anon_spare);
914 	}
915 	return error;
916 }
917 
918 /*
919  * uvm_fault_check: check prot, handle needs-copy, etc.
920  *
921  *	1. lookup entry.
922  *	2. check protection.
923  *	3. adjust fault condition (mainly for simulated fault).
924  *	4. handle needs-copy (lazy amap copy).
925  *	5. establish range of interest for neighbor fault (aka pre-fault).
926  *	6. look up anons (if amap exists).
927  *	7. flush pages (if MADV_SEQUENTIAL)
928  *
929  * => called with nothing locked.
930  * => if we fail (result != 0) we unlock everything.
931  * => initialize/adjust many members of flt.
932  */
933 
934 static int
935 uvm_fault_check(
936 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
937 	struct vm_anon ***ranons, bool maxprot)
938 {
939 	struct vm_amap *amap;
940 	struct uvm_object *uobj;
941 	vm_prot_t check_prot;
942 	int nback, nforw;
943 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
944 
945 	/*
946 	 * lookup and lock the maps
947 	 */
948 
949 	if (uvmfault_lookup(ufi, false) == false) {
950 		UVMHIST_LOG(maphist, "<- no mapping @ %#jx", ufi->orig_rvaddr,
951 		    0,0,0);
952 		return EFAULT;
953 	}
954 	/* locked: maps(read) */
955 
956 #ifdef DIAGNOSTIC
957 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
958 		printf("Page fault on non-pageable map:\n");
959 		printf("ufi->map = %p\n", ufi->map);
960 		printf("ufi->orig_map = %p\n", ufi->orig_map);
961 		printf("ufi->orig_rvaddr = %#lx\n", (u_long) ufi->orig_rvaddr);
962 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
963 	}
964 #endif
965 
966 	/*
967 	 * check protection
968 	 */
969 
970 	check_prot = maxprot ?
971 	    ufi->entry->max_protection : ufi->entry->protection;
972 	if ((check_prot & flt->access_type) != flt->access_type) {
973 		UVMHIST_LOG(maphist,
974 		    "<- protection failure (prot=%#jx, access=%#jx)",
975 		    ufi->entry->protection, flt->access_type, 0, 0);
976 		uvmfault_unlockmaps(ufi, false);
977 		return EFAULT;
978 	}
979 
980 	/*
981 	 * "enter_prot" is the protection we want to enter the page in at.
982 	 * for certain pages (e.g. copy-on-write pages) this protection can
983 	 * be more strict than ufi->entry->protection.  "wired" means either
984 	 * the entry is wired or we are fault-wiring the pg.
985 	 */
986 
987 	flt->enter_prot = ufi->entry->protection;
988 	if (VM_MAPENT_ISWIRED(ufi->entry)) {
989 		flt->wire_mapping = true;
990 		flt->wire_paging = true;
991 		flt->narrow = true;
992 	}
993 
994 	if (flt->wire_mapping) {
995 		flt->access_type = flt->enter_prot; /* full access for wired */
996 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
997 	} else {
998 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
999 	}
1000 
1001 	flt->promote = false;
1002 
1003 	/*
1004 	 * handle "needs_copy" case.   if we need to copy the amap we will
1005 	 * have to drop our readlock and relock it with a write lock.  (we
1006 	 * need a write lock to change anything in a map entry [e.g.
1007 	 * needs_copy]).
1008 	 */
1009 
1010 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
1011 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
1012 			KASSERT(!maxprot);
1013 			/* need to clear */
1014 			UVMHIST_LOG(maphist,
1015 			    "  need to clear needs_copy and refault",0,0,0,0);
1016 			uvmfault_unlockmaps(ufi, false);
1017 			uvmfault_amapcopy(ufi);
1018 			cpu_count(CPU_COUNT_FLTAMCOPY, 1);
1019 			return ERESTART;
1020 
1021 		} else {
1022 
1023 			/*
1024 			 * ensure that we pmap_enter page R/O since
1025 			 * needs_copy is still true
1026 			 */
1027 
1028 			flt->enter_prot &= ~VM_PROT_WRITE;
1029 		}
1030 	}
1031 
1032 	/*
1033 	 * identify the players
1034 	 */
1035 
1036 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
1037 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
1038 
1039 	/*
1040 	 * check for a case 0 fault.  if nothing backing the entry then
1041 	 * error now.
1042 	 */
1043 
1044 	if (amap == NULL && uobj == NULL) {
1045 		uvmfault_unlockmaps(ufi, false);
1046 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
1047 		return EFAULT;
1048 	}
1049 
1050 	/*
1051 	 * establish range of interest based on advice from mapper
1052 	 * and then clip to fit map entry.   note that we only want
1053 	 * to do this the first time through the fault.   if we
1054 	 * ReFault we will disable this by setting "narrow" to true.
1055 	 */
1056 
1057 	if (flt->narrow == false) {
1058 
1059 		/* wide fault (!narrow) */
1060 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
1061 			 ufi->entry->advice);
1062 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
1063 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
1064 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
1065 		/*
1066 		 * note: "-1" because we don't want to count the
1067 		 * faulting page as forw
1068 		 */
1069 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1070 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1071 			     PAGE_SHIFT) - 1);
1072 		flt->npages = nback + nforw + 1;
1073 		flt->centeridx = nback;
1074 
1075 		flt->narrow = true;	/* ensure only once per-fault */
1076 
1077 	} else {
1078 
1079 		/* narrow fault! */
1080 		nback = nforw = 0;
1081 		flt->startva = ufi->orig_rvaddr;
1082 		flt->npages = 1;
1083 		flt->centeridx = 0;
1084 
1085 	}
1086 	/* offset from entry's start to pgs' start */
1087 	const voff_t eoff = flt->startva - ufi->entry->start;
1088 
1089 	/* locked: maps(read) */
1090 	UVMHIST_LOG(maphist, "  narrow=%jd, back=%jd, forw=%jd, startva=%#jx",
1091 		    flt->narrow, nback, nforw, flt->startva);
1092 	UVMHIST_LOG(maphist, "  entry=%#jx, amap=%#jx, obj=%#jx",
1093 	    (uintptr_t)ufi->entry, (uintptr_t)amap, (uintptr_t)uobj, 0);
1094 
1095 	/*
1096 	 * if we've got an amap, lock it and extract current anons.
1097 	 */
1098 
1099 	if (amap) {
1100 		amap_lock(amap, RW_WRITER);
1101 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1102 	} else {
1103 		*ranons = NULL;	/* to be safe */
1104 	}
1105 
1106 	/* locked: maps(read), amap(if there) */
1107 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1108 
1109 	/*
1110 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1111 	 * now and then forget about them (for the rest of the fault).
1112 	 */
1113 
1114 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1115 
1116 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1117 		    0,0,0,0);
1118 		/* flush back-page anons? */
1119 		if (amap)
1120 			uvmfault_anonflush(*ranons, nback);
1121 
1122 		/* flush object? */
1123 		if (uobj) {
1124 			voff_t uoff;
1125 
1126 			uoff = ufi->entry->offset + eoff;
1127 			rw_enter(uobj->vmobjlock, RW_WRITER);
1128 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1129 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1130 		}
1131 
1132 		/* now forget about the backpages */
1133 		if (amap)
1134 			*ranons += nback;
1135 		flt->startva += (nback << PAGE_SHIFT);
1136 		flt->npages -= nback;
1137 		flt->centeridx = 0;
1138 	}
1139 	/*
1140 	 * => startva is fixed
1141 	 * => npages is fixed
1142 	 */
1143 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1144 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1145 	    flt->startva + (flt->npages << PAGE_SHIFT));
1146 	return 0;
1147 }
1148 
1149 /*
1150  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1151  *
1152  * iterate range of interest:
1153  *	1. check if h/w mapping exists.  if yes, we don't care
1154  *	2. check if anon exists.  if not, page is lower.
1155  *	3. if anon exists, enter h/w mapping for neighbors.
1156  *
1157  * => called with amap locked (if exists).
1158  */
1159 
1160 static int
1161 uvm_fault_upper_lookup(
1162 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1163 	struct vm_anon **anons, struct vm_page **pages)
1164 {
1165 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1166 	int lcv;
1167 	vaddr_t currva;
1168 	bool shadowed __unused;
1169 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1170 
1171 	/* locked: maps(read), amap(if there) */
1172 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1173 
1174 	/*
1175 	 * map in the backpages and frontpages we found in the amap in hopes
1176 	 * of preventing future faults.    we also init the pages[] array as
1177 	 * we go.
1178 	 */
1179 
1180 	currva = flt->startva;
1181 	shadowed = false;
1182 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1183 		/*
1184 		 * don't play with VAs that are already mapped
1185 		 * (except for center)
1186 		 */
1187 		if (lcv != flt->centeridx &&
1188 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1189 			pages[lcv] = PGO_DONTCARE;
1190 			continue;
1191 		}
1192 
1193 		/*
1194 		 * unmapped or center page.   check if any anon at this level.
1195 		 */
1196 		if (amap == NULL || anons[lcv] == NULL) {
1197 			pages[lcv] = NULL;
1198 			continue;
1199 		}
1200 
1201 		/*
1202 		 * check for present page and map if possible.   re-activate it.
1203 		 */
1204 
1205 		pages[lcv] = PGO_DONTCARE;
1206 		if (lcv == flt->centeridx) {	/* save center for later! */
1207 			shadowed = true;
1208 			continue;
1209 		}
1210 
1211 		struct vm_anon *anon = anons[lcv];
1212 		struct vm_page *pg = anon->an_page;
1213 
1214 		KASSERT(anon->an_lock == amap->am_lock);
1215 
1216 		/* Ignore loaned and busy pages. */
1217 		if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0) {
1218 			uvm_fault_upper_neighbor(ufi, flt, currva,
1219 			    pg, anon->an_ref > 1);
1220 		}
1221 	}
1222 
1223 	/* locked: maps(read), amap(if there) */
1224 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1225 	/* (shadowed == true) if there is an anon at the faulting address */
1226 	UVMHIST_LOG(maphist, "  shadowed=%jd, will_get=%jd", shadowed,
1227 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1228 
1229 	/*
1230 	 * note that if we are really short of RAM we could sleep in the above
1231 	 * call to pmap_enter with everything locked.   bad?
1232 	 *
1233 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1234 	 * XXX case.  --thorpej
1235 	 */
1236 
1237 	return 0;
1238 }
1239 
1240 /*
1241  * uvm_fault_upper_neighbor: enter single upper neighbor page.
1242  *
1243  * => called with amap and anon locked.
1244  */
1245 
1246 static void
1247 uvm_fault_upper_neighbor(
1248 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1249 	vaddr_t currva, struct vm_page *pg, bool readonly)
1250 {
1251 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1252 
1253 	/* locked: amap, anon */
1254 
1255 	KASSERT(pg->uobject == NULL);
1256 	KASSERT(pg->uanon != NULL);
1257 	KASSERT(rw_write_held(pg->uanon->an_lock));
1258 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1259 
1260 	uvm_pagelock(pg);
1261 	uvm_pageenqueue(pg);
1262 	uvm_pageunlock(pg);
1263 	UVMHIST_LOG(maphist,
1264 	    "  MAPPING: n anon: pm=%#jx, va=%#jx, pg=%#jx",
1265 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
1266 	cpu_count(CPU_COUNT_FLTNAMAP, 1);
1267 
1268 	/*
1269 	 * Since this page isn't the page that's actually faulting,
1270 	 * ignore pmap_enter() failures; it's not critical that we
1271 	 * enter these right now.
1272 	 */
1273 
1274 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1275 	    VM_PAGE_TO_PHYS(pg),
1276 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1277 	    flt->enter_prot,
1278 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1279 
1280 	pmap_update(ufi->orig_map->pmap);
1281 }
1282 
1283 /*
1284  * uvm_fault_upper: handle upper fault.
1285  *
1286  *	1. acquire anon lock.
1287  *	2. get anon.  let uvmfault_anonget do the dirty work.
1288  *	3. handle loan.
1289  *	4. dispatch direct or promote handlers.
1290  */
1291 
1292 static int
1293 uvm_fault_upper(
1294 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1295 	struct vm_anon **anons)
1296 {
1297 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1298 	struct vm_anon * const anon = anons[flt->centeridx];
1299 	struct uvm_object *uobj;
1300 	int error;
1301 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1302 
1303 	/* locked: maps(read), amap, anon */
1304 	KASSERT(rw_write_held(amap->am_lock));
1305 	KASSERT(anon->an_lock == amap->am_lock);
1306 
1307 	/*
1308 	 * handle case 1: fault on an anon in our amap
1309 	 */
1310 
1311 	UVMHIST_LOG(maphist, "  case 1 fault: anon=%#jx",
1312 	    (uintptr_t)anon, 0, 0, 0);
1313 
1314 	/*
1315 	 * no matter if we have case 1A or case 1B we are going to need to
1316 	 * have the anon's memory resident.   ensure that now.
1317 	 */
1318 
1319 	/*
1320 	 * let uvmfault_anonget do the dirty work.
1321 	 * if it fails (!OK) it will unlock everything for us.
1322 	 * if it succeeds, locks are still valid and locked.
1323 	 * also, if it is OK, then the anon's page is on the queues.
1324 	 * if the page is on loan from a uvm_object, then anonget will
1325 	 * lock that object for us if it does not fail.
1326 	 */
1327 
1328 	error = uvmfault_anonget(ufi, amap, anon);
1329 	switch (error) {
1330 	case 0:
1331 		break;
1332 
1333 	case ERESTART:
1334 		return ERESTART;
1335 
1336 	case EAGAIN:
1337 		kpause("fltagain1", false, hz/2, NULL);
1338 		return ERESTART;
1339 
1340 	default:
1341 		return error;
1342 	}
1343 
1344 	/*
1345 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1346 	 */
1347 
1348 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1349 
1350 	/* locked: maps(read), amap, anon, uobj(if one) */
1351 	KASSERT(rw_write_held(amap->am_lock));
1352 	KASSERT(anon->an_lock == amap->am_lock);
1353 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1354 
1355 	/*
1356 	 * special handling for loaned pages
1357 	 */
1358 
1359 	if (anon->an_page->loan_count) {
1360 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1361 		if (error != 0)
1362 			return error;
1363 	}
1364 
1365 	/*
1366 	 * if we are case 1B then we will need to allocate a new blank
1367 	 * anon to transfer the data into.   note that we have a lock
1368 	 * on anon, so no one can busy or release the page until we are done.
1369 	 * also note that the ref count can't drop to zero here because
1370 	 * it is > 1 and we are only dropping one ref.
1371 	 *
1372 	 * in the (hopefully very rare) case that we are out of RAM we
1373 	 * will unlock, wait for more RAM, and refault.
1374 	 *
1375 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1376 	 */
1377 
1378 	if (flt->cow_now && anon->an_ref > 1) {
1379 		flt->promote = true;
1380 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1381 	} else {
1382 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1383 	}
1384 	return error;
1385 }
1386 
1387 /*
1388  * uvm_fault_upper_loan: handle loaned upper page.
1389  *
1390  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1391  *	2. if cow'ing now, and if ref count is 1, break loan.
1392  */
1393 
1394 static int
1395 uvm_fault_upper_loan(
1396 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1397 	struct vm_anon *anon, struct uvm_object **ruobj)
1398 {
1399 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1400 	int error = 0;
1401 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1402 
1403 	if (!flt->cow_now) {
1404 
1405 		/*
1406 		 * for read faults on loaned pages we just cap the
1407 		 * protection at read-only.
1408 		 */
1409 
1410 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1411 
1412 	} else {
1413 		/*
1414 		 * note that we can't allow writes into a loaned page!
1415 		 *
1416 		 * if we have a write fault on a loaned page in an
1417 		 * anon then we need to look at the anon's ref count.
1418 		 * if it is greater than one then we are going to do
1419 		 * a normal copy-on-write fault into a new anon (this
1420 		 * is not a problem).  however, if the reference count
1421 		 * is one (a case where we would normally allow a
1422 		 * write directly to the page) then we need to kill
1423 		 * the loan before we continue.
1424 		 */
1425 
1426 		/* >1 case is already ok */
1427 		if (anon->an_ref == 1) {
1428 			error = uvm_loanbreak_anon(anon, *ruobj);
1429 			if (error != 0) {
1430 				uvmfault_unlockall(ufi, amap, *ruobj);
1431 				uvm_wait("flt_noram2");
1432 				return ERESTART;
1433 			}
1434 			/* if we were a loan receiver uobj is gone */
1435 			if (*ruobj)
1436 				*ruobj = NULL;
1437 		}
1438 	}
1439 	return error;
1440 }
1441 
1442 /*
1443  * uvm_fault_upper_promote: promote upper page.
1444  *
1445  *	1. call uvmfault_promote.
1446  *	2. enqueue page.
1447  *	3. deref.
1448  *	4. pass page to uvm_fault_upper_enter.
1449  */
1450 
1451 static int
1452 uvm_fault_upper_promote(
1453 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1454 	struct uvm_object *uobj, struct vm_anon *anon)
1455 {
1456 	struct vm_anon * const oanon = anon;
1457 	struct vm_page *pg;
1458 	int error;
1459 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1460 
1461 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1462 	cpu_count(CPU_COUNT_FLT_ACOW, 1);
1463 
1464 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1465 	    &flt->anon_spare);
1466 	switch (error) {
1467 	case 0:
1468 		break;
1469 	case ERESTART:
1470 		return ERESTART;
1471 	default:
1472 		return error;
1473 	}
1474 
1475 	KASSERT(anon == NULL || anon->an_lock == oanon->an_lock);
1476 
1477 	pg = anon->an_page;
1478 	/* uvm_fault_upper_done will activate the page */
1479 	uvm_pagelock(pg);
1480 	uvm_pageenqueue(pg);
1481 	uvm_pageunlock(pg);
1482 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1483 	UVM_PAGE_OWN(pg, NULL);
1484 
1485 	/* deref: can not drop to zero here by defn! */
1486 	KASSERT(oanon->an_ref > 1);
1487 	oanon->an_ref--;
1488 
1489 	/*
1490 	 * note: oanon is still locked, as is the new anon.  we
1491 	 * need to check for this later when we unlock oanon; if
1492 	 * oanon != anon, we'll have to unlock anon, too.
1493 	 */
1494 
1495 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1496 }
1497 
1498 /*
1499  * uvm_fault_upper_direct: handle direct fault.
1500  */
1501 
1502 static int
1503 uvm_fault_upper_direct(
1504 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1505 	struct uvm_object *uobj, struct vm_anon *anon)
1506 {
1507 	struct vm_anon * const oanon = anon;
1508 	struct vm_page *pg;
1509 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1510 
1511 	cpu_count(CPU_COUNT_FLT_ANON, 1);
1512 	pg = anon->an_page;
1513 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1514 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1515 
1516 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1517 }
1518 
1519 /*
1520  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1521  */
1522 
1523 static int
1524 uvm_fault_upper_enter(
1525 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1526 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1527 	struct vm_anon *oanon)
1528 {
1529 	struct pmap *pmap = ufi->orig_map->pmap;
1530 	vaddr_t va = ufi->orig_rvaddr;
1531 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1532 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1533 
1534 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1535 	KASSERT(rw_write_held(amap->am_lock));
1536 	KASSERT(anon->an_lock == amap->am_lock);
1537 	KASSERT(oanon->an_lock == amap->am_lock);
1538 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1539 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1540 
1541 	/*
1542 	 * now map the page in.
1543 	 */
1544 
1545 	UVMHIST_LOG(maphist,
1546 	    "  MAPPING: anon: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
1547 	    (uintptr_t)pmap, va, (uintptr_t)pg, flt->promote);
1548 	if (pmap_enter(pmap, va, VM_PAGE_TO_PHYS(pg),
1549 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1550 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1551 
1552 		/*
1553 		 * If pmap_enter() fails, it must not leave behind an existing
1554 		 * pmap entry.  In particular, a now-stale entry for a different
1555 		 * page would leave the pmap inconsistent with the vm_map.
1556 		 * This is not to imply that pmap_enter() should remove an
1557 		 * existing mapping in such a situation (since that could create
1558 		 * different problems, eg. if the existing mapping is wired),
1559 		 * but rather that the pmap should be designed such that it
1560 		 * never needs to fail when the new mapping is replacing an
1561 		 * existing mapping and the new page has no existing mappings.
1562 		 */
1563 
1564 		KASSERT(!pmap_extract(pmap, va, NULL));
1565 
1566 		/*
1567 		 * No need to undo what we did; we can simply think of
1568 		 * this as the pmap throwing away the mapping information.
1569 		 *
1570 		 * We do, however, have to go through the ReFault path,
1571 		 * as the map may change while we're asleep.
1572 		 */
1573 
1574 		uvmfault_unlockall(ufi, amap, uobj);
1575 		if (!uvm_reclaimable()) {
1576 			UVMHIST_LOG(maphist,
1577 			    "<- failed.  out of VM",0,0,0,0);
1578 			/* XXX instrumentation */
1579 			return ENOMEM;
1580 		}
1581 		/* XXX instrumentation */
1582 		uvm_wait("flt_pmfail1");
1583 		return ERESTART;
1584 	}
1585 
1586 	uvm_fault_upper_done(ufi, flt, anon, pg);
1587 
1588 	/*
1589 	 * done case 1!  finish up by unlocking everything and returning success
1590 	 */
1591 
1592 	pmap_update(pmap);
1593 	uvmfault_unlockall(ufi, amap, uobj);
1594 	return 0;
1595 }
1596 
1597 /*
1598  * uvm_fault_upper_done: queue upper center page.
1599  */
1600 
1601 static void
1602 uvm_fault_upper_done(
1603 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1604 	struct vm_anon *anon, struct vm_page *pg)
1605 {
1606 	const bool wire_paging = flt->wire_paging;
1607 
1608 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1609 
1610 	/*
1611 	 * ... update the page queues.
1612 	 */
1613 
1614 	uvm_pagelock(pg);
1615 	if (wire_paging) {
1616 		uvm_pagewire(pg);
1617 	} else {
1618 		uvm_pageactivate(pg);
1619 	}
1620 	uvm_pageunlock(pg);
1621 
1622 	if (wire_paging) {
1623 		/*
1624 		 * since the now-wired page cannot be paged out,
1625 		 * release its swap resources for others to use.
1626 		 * and since an anon with no swap cannot be clean,
1627 		 * mark it dirty now.
1628 		 */
1629 
1630 		uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
1631 		uvm_anon_dropswap(anon);
1632 	}
1633 }
1634 
1635 /*
1636  * uvm_fault_lower: handle lower fault.
1637  *
1638  *	1. check uobj
1639  *	1.1. if null, ZFOD.
1640  *	1.2. if not null, look up unnmapped neighbor pages.
1641  *	2. for center page, check if promote.
1642  *	2.1. ZFOD always needs promotion.
1643  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1644  *	3. if uobj is not ZFOD and page is not found, do i/o.
1645  *	4. dispatch either direct / promote fault.
1646  */
1647 
1648 static int
1649 uvm_fault_lower(
1650 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1651 	struct vm_page **pages)
1652 {
1653 	struct vm_amap *amap __diagused = ufi->entry->aref.ar_amap;
1654 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1655 	struct vm_page *uobjpage;
1656 	int error;
1657 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1658 
1659 	/*
1660 	 * now, if the desired page is not shadowed by the amap and we have
1661 	 * a backing object that does not have a special fault routine, then
1662 	 * we ask (with pgo_get) the object for resident pages that we care
1663 	 * about and attempt to map them in.  we do not let pgo_get block
1664 	 * (PGO_LOCKED).
1665 	 */
1666 
1667 	if (uobj == NULL) {
1668 		/* zero fill; don't care neighbor pages */
1669 		uobjpage = NULL;
1670 	} else {
1671 		uvm_fault_lower_lookup(ufi, flt, pages);
1672 		uobjpage = pages[flt->centeridx];
1673 	}
1674 
1675 	/*
1676 	 * note that at this point we are done with any front or back pages.
1677 	 * we are now going to focus on the center page (i.e. the one we've
1678 	 * faulted on).  if we have faulted on the upper (anon) layer
1679 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1680 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1681 	 * layer [i.e. case 2] and the page was both present and available,
1682 	 * then we've got a pointer to it as "uobjpage" and we've already
1683 	 * made it BUSY.
1684 	 */
1685 
1686 	/*
1687 	 * locked:
1688 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1689 	 */
1690 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1691 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1692 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1693 
1694 	/*
1695 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1696 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1697 	 * have a backing object, check and see if we are going to promote
1698 	 * the data up to an anon during the fault.
1699 	 */
1700 
1701 	if (uobj == NULL) {
1702 		uobjpage = PGO_DONTCARE;
1703 		flt->promote = true;		/* always need anon here */
1704 	} else {
1705 		KASSERT(uobjpage != PGO_DONTCARE);
1706 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1707 	}
1708 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%jd, zfill=%jd",
1709 	    flt->promote, (uobj == NULL), 0,0);
1710 
1711 	/*
1712 	 * if uobjpage is not null then we do not need to do I/O to get the
1713 	 * uobjpage.
1714 	 *
1715 	 * if uobjpage is null, then we need to unlock and ask the pager to
1716 	 * get the data for us.   once we have the data, we need to reverify
1717 	 * the state the world.   we are currently not holding any resources.
1718 	 */
1719 
1720 	if (uobjpage) {
1721 		/* update rusage counters */
1722 		curlwp->l_ru.ru_minflt++;
1723 	} else {
1724 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1725 		if (error != 0)
1726 			return error;
1727 	}
1728 
1729 	/*
1730 	 * locked:
1731 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1732 	 */
1733 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1734 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1735 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1736 
1737 	/*
1738 	 * notes:
1739 	 *  - at this point uobjpage can not be NULL
1740 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1741 	 *  for it above)
1742 	 *  - at this point uobjpage could be waited on (handle later)
1743 	 */
1744 
1745 	KASSERT(uobjpage != NULL);
1746 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1747 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1748 	    uvm_pagegetdirty(uobjpage) == UVM_PAGE_STATUS_CLEAN);
1749 
1750 	if (!flt->promote) {
1751 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1752 	} else {
1753 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1754 	}
1755 	return error;
1756 }
1757 
1758 /*
1759  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1760  *
1761  *	1. get on-memory pages.
1762  *	2. if failed, give up (get only center page later).
1763  *	3. if succeeded, enter h/w mapping of neighbor pages.
1764  */
1765 
1766 static void
1767 uvm_fault_lower_lookup(
1768 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1769 	struct vm_page **pages)
1770 {
1771 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1772 	int lcv, gotpages;
1773 	vaddr_t currva;
1774 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1775 
1776 	rw_enter(uobj->vmobjlock, RW_WRITER);
1777 	/* Locked: maps(read), amap(if there), uobj */
1778 
1779 	cpu_count(CPU_COUNT_FLTLGET, 1);
1780 	gotpages = flt->npages;
1781 	(void) uobj->pgops->pgo_get(uobj,
1782 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1783 	    pages, &gotpages, flt->centeridx,
1784 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1785 
1786 	KASSERT(rw_write_held(uobj->vmobjlock));
1787 
1788 	/*
1789 	 * check for pages to map, if we got any
1790 	 */
1791 
1792 	if (gotpages == 0) {
1793 		pages[flt->centeridx] = NULL;
1794 		return;
1795 	}
1796 
1797 	currva = flt->startva;
1798 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1799 		struct vm_page *curpg;
1800 
1801 		curpg = pages[lcv];
1802 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1803 			continue;
1804 		}
1805 		KASSERT(curpg->uobject == uobj);
1806 
1807 		/*
1808 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1809 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1810 		 * to it.
1811 		 */
1812 
1813 		if (lcv == flt->centeridx) {
1814 			UVMHIST_LOG(maphist, "  got uobjpage (%#jx) "
1815 			    "with locked get", (uintptr_t)curpg, 0, 0, 0);
1816 		} else {
1817 			uvm_fault_lower_neighbor(ufi, flt, currva, curpg);
1818 		}
1819 	}
1820 	pmap_update(ufi->orig_map->pmap);
1821 }
1822 
1823 /*
1824  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1825  */
1826 
1827 static void
1828 uvm_fault_lower_neighbor(
1829 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1830 	vaddr_t currva, struct vm_page *pg)
1831 {
1832 	const bool readonly = uvm_pagereadonly_p(pg) || pg->loan_count > 0;
1833 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1834 
1835 	/* locked: maps(read), amap(if there), uobj */
1836 
1837 	/*
1838 	 * calling pgo_get with PGO_LOCKED returns us pages which
1839 	 * are neither busy nor released, so we don't need to check
1840 	 * for this.  we can just directly enter the pages.
1841 	 */
1842 
1843 	uvm_pagelock(pg);
1844 	uvm_pageenqueue(pg);
1845 	uvm_pageunlock(pg);
1846 	UVMHIST_LOG(maphist,
1847 	    "  MAPPING: n obj: pm=%#jx, va=%#jx, pg=%#jx",
1848 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
1849 	cpu_count(CPU_COUNT_FLTNOMAP, 1);
1850 
1851 	/*
1852 	 * Since this page isn't the page that's actually faulting,
1853 	 * ignore pmap_enter() failures; it's not critical that we
1854 	 * enter these right now.
1855 	 * NOTE: page can't be waited on or PG_RELEASED because we've
1856 	 * held the lock the whole time we've had the handle.
1857 	 */
1858 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1859 	KASSERT((pg->flags & PG_RELEASED) == 0);
1860 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) ||
1861 	    uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_CLEAN);
1862 	pg->flags &= ~(PG_BUSY);
1863 	UVM_PAGE_OWN(pg, NULL);
1864 
1865 	KASSERT(rw_write_held(pg->uobject->vmobjlock));
1866 
1867 	const vm_prot_t mapprot =
1868 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1869 	    flt->enter_prot & MASK(ufi->entry);
1870 	const u_int mapflags =
1871 	    PMAP_CANFAIL | (flt->wire_mapping ? (mapprot | PMAP_WIRED) : 0);
1872 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1873 	    VM_PAGE_TO_PHYS(pg), mapprot, mapflags);
1874 }
1875 
1876 /*
1877  * uvm_fault_lower_io: get lower page from backing store.
1878  *
1879  *	1. unlock everything, because i/o will block.
1880  *	2. call pgo_get.
1881  *	3. if failed, recover.
1882  *	4. if succeeded, relock everything and verify things.
1883  */
1884 
1885 static int
1886 uvm_fault_lower_io(
1887 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1888 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1889 {
1890 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1891 	struct uvm_object *uobj = *ruobj;
1892 	struct vm_page *pg;
1893 	bool locked;
1894 	int gotpages;
1895 	int error;
1896 	voff_t uoff;
1897 	vm_prot_t access_type;
1898 	int advice;
1899 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1900 
1901 	/* update rusage counters */
1902 	curlwp->l_ru.ru_majflt++;
1903 
1904 	/* grab everything we need from the entry before we unlock */
1905 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1906 	access_type = flt->access_type & MASK(ufi->entry);
1907 	advice = ufi->entry->advice;
1908 
1909 	/* Locked: maps(read), amap(if there), uobj */
1910 	uvmfault_unlockall(ufi, amap, NULL);
1911 
1912 	/* Locked: uobj */
1913 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1914 
1915 	cpu_count(CPU_COUNT_FLTGET, 1);
1916 	gotpages = 1;
1917 	pg = NULL;
1918 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1919 	    0, access_type, advice, PGO_SYNCIO);
1920 	/* locked: pg(if no error) */
1921 
1922 	/*
1923 	 * recover from I/O
1924 	 */
1925 
1926 	if (error) {
1927 		if (error == EAGAIN) {
1928 			UVMHIST_LOG(maphist,
1929 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1930 			kpause("fltagain2", false, hz/2, NULL);
1931 			return ERESTART;
1932 		}
1933 
1934 #if 0
1935 		KASSERT(error != ERESTART);
1936 #else
1937 		/* XXXUEBS don't re-fault? */
1938 		if (error == ERESTART)
1939 			error = EIO;
1940 #endif
1941 
1942 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %jd)",
1943 		    error, 0,0,0);
1944 		return error;
1945 	}
1946 
1947 	/*
1948 	 * re-verify the state of the world by first trying to relock
1949 	 * the maps.  always relock the object.
1950 	 */
1951 
1952 	locked = uvmfault_relock(ufi);
1953 	if (locked && amap)
1954 		amap_lock(amap, RW_WRITER);
1955 
1956 	/* might be changed */
1957 	uobj = pg->uobject;
1958 
1959 	rw_enter(uobj->vmobjlock, RW_WRITER);
1960 	KASSERT((pg->flags & PG_BUSY) != 0);
1961 
1962 	uvm_pagelock(pg);
1963 	uvm_pageactivate(pg);
1964 	uvm_pageunlock(pg);
1965 
1966 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1967 	/* locked(!locked): uobj, pg */
1968 
1969 	/*
1970 	 * verify that the page has not be released and re-verify
1971 	 * that amap slot is still free.   if there is a problem,
1972 	 * we unlock and clean up.
1973 	 */
1974 
1975 	if ((pg->flags & PG_RELEASED) != 0 ||
1976 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1977 	      ufi->orig_rvaddr - ufi->entry->start))) {
1978 		if (locked)
1979 			uvmfault_unlockall(ufi, amap, NULL);
1980 		locked = false;
1981 	}
1982 
1983 	/*
1984 	 * didn't get the lock?   release the page and retry.
1985 	 */
1986 
1987 	if (locked == false) {
1988 		UVMHIST_LOG(maphist,
1989 		    "  wasn't able to relock after fault: retry",
1990 		    0,0,0,0);
1991 		if ((pg->flags & PG_RELEASED) == 0) {
1992 			pg->flags &= ~PG_BUSY;
1993 			uvm_pagelock(pg);
1994 			uvm_pagewakeup(pg);
1995 			uvm_pageunlock(pg);
1996 			UVM_PAGE_OWN(pg, NULL);
1997 		} else {
1998 			cpu_count(CPU_COUNT_FLTPGRELE, 1);
1999 			uvm_pagefree(pg);
2000 		}
2001 		rw_exit(uobj->vmobjlock);
2002 		return ERESTART;
2003 	}
2004 
2005 	/*
2006 	 * we have the data in pg which is busy and
2007 	 * not released.  we are holding object lock (so the page
2008 	 * can't be released on us).
2009 	 */
2010 
2011 	/* locked: maps(read), amap(if !null), uobj, pg */
2012 
2013 	*ruobj = uobj;
2014 	*ruobjpage = pg;
2015 	return 0;
2016 }
2017 
2018 /*
2019  * uvm_fault_lower_direct: fault lower center page
2020  *
2021  *	1. adjust flt->enter_prot.
2022  *	2. if page is loaned, resolve.
2023  */
2024 
2025 int
2026 uvm_fault_lower_direct(
2027 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2028 	struct uvm_object *uobj, struct vm_page *uobjpage)
2029 {
2030 	struct vm_page *pg;
2031 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
2032 
2033 	/*
2034 	 * we are not promoting.   if the mapping is COW ensure that we
2035 	 * don't give more access than we should (e.g. when doing a read
2036 	 * fault on a COPYONWRITE mapping we want to map the COW page in
2037 	 * R/O even though the entry protection could be R/W).
2038 	 *
2039 	 * set "pg" to the page we want to map in (uobjpage, usually)
2040 	 */
2041 
2042 	cpu_count(CPU_COUNT_FLT_OBJ, 1);
2043 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
2044 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
2045 		flt->enter_prot &= ~VM_PROT_WRITE;
2046 	pg = uobjpage;		/* map in the actual object */
2047 
2048 	KASSERT(uobjpage != PGO_DONTCARE);
2049 
2050 	/*
2051 	 * we are faulting directly on the page.   be careful
2052 	 * about writing to loaned pages...
2053 	 */
2054 
2055 	if (uobjpage->loan_count) {
2056 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
2057 	}
2058 	KASSERT(pg == uobjpage);
2059 
2060 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2061 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
2062 }
2063 
2064 /*
2065  * uvm_fault_lower_direct_loan: resolve loaned page.
2066  *
2067  *	1. if not cow'ing, adjust flt->enter_prot.
2068  *	2. if cow'ing, break loan.
2069  */
2070 
2071 static int
2072 uvm_fault_lower_direct_loan(
2073 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2074 	struct uvm_object *uobj, struct vm_page **rpg,
2075 	struct vm_page **ruobjpage)
2076 {
2077 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2078 	struct vm_page *pg;
2079 	struct vm_page *uobjpage = *ruobjpage;
2080 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
2081 
2082 	if (!flt->cow_now) {
2083 		/* read fault: cap the protection at readonly */
2084 		/* cap! */
2085 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2086 	} else {
2087 		/* write fault: must break the loan here */
2088 
2089 		pg = uvm_loanbreak(uobjpage);
2090 		if (pg == NULL) {
2091 
2092 			/*
2093 			 * drop ownership of page, it can't be released
2094 			 */
2095 
2096 			uvm_pagelock(uobjpage);
2097 			uvm_pagewakeup(uobjpage);
2098 			uvm_pageunlock(uobjpage);
2099 			uobjpage->flags &= ~PG_BUSY;
2100 			UVM_PAGE_OWN(uobjpage, NULL);
2101 
2102 			uvmfault_unlockall(ufi, amap, uobj);
2103 			UVMHIST_LOG(maphist,
2104 			  "  out of RAM breaking loan, waiting",
2105 			  0,0,0,0);
2106 			cpu_count(CPU_COUNT_FLTNORAM, 1);
2107 			uvm_wait("flt_noram4");
2108 			return ERESTART;
2109 		}
2110 		*rpg = pg;
2111 		*ruobjpage = pg;
2112 	}
2113 	return 0;
2114 }
2115 
2116 /*
2117  * uvm_fault_lower_promote: promote lower page.
2118  *
2119  *	1. call uvmfault_promote.
2120  *	2. fill in data.
2121  *	3. if not ZFOD, dispose old page.
2122  */
2123 
2124 int
2125 uvm_fault_lower_promote(
2126 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2127 	struct uvm_object *uobj, struct vm_page *uobjpage)
2128 {
2129 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2130 	struct vm_anon *anon;
2131 	struct vm_page *pg;
2132 	int error;
2133 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2134 
2135 	KASSERT(amap != NULL);
2136 
2137 	/*
2138 	 * If we are going to promote the data to an anon we
2139 	 * allocate a blank anon here and plug it into our amap.
2140 	 */
2141 	error = uvmfault_promote(ufi, NULL, uobjpage,
2142 	    &anon, &flt->anon_spare);
2143 	switch (error) {
2144 	case 0:
2145 		break;
2146 	case ERESTART:
2147 		return ERESTART;
2148 	default:
2149 		return error;
2150 	}
2151 
2152 	pg = anon->an_page;
2153 
2154 	/*
2155 	 * Fill in the data.
2156 	 */
2157 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2158 
2159 	if (uobjpage != PGO_DONTCARE) {
2160 		cpu_count(CPU_COUNT_FLT_PRCOPY, 1);
2161 
2162 		/*
2163 		 * promote to shared amap?  make sure all sharing
2164 		 * procs see it
2165 		 */
2166 
2167 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2168 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2169 			/*
2170 			 * XXX: PAGE MIGHT BE WIRED!
2171 			 */
2172 		}
2173 
2174 		/*
2175 		 * dispose of uobjpage.  it can't be PG_RELEASED
2176 		 * since we still hold the object lock.
2177 		 */
2178 
2179 		uobjpage->flags &= ~PG_BUSY;
2180 		uvm_pagelock(uobjpage);
2181 		uvm_pagewakeup(uobjpage);
2182 		uvm_pageunlock(uobjpage);
2183 		UVM_PAGE_OWN(uobjpage, NULL);
2184 
2185 		UVMHIST_LOG(maphist,
2186 		    "  promote uobjpage %#jx to anon/page %#jx/%#jx",
2187 		    (uintptr_t)uobjpage, (uintptr_t)anon, (uintptr_t)pg, 0);
2188 
2189 	} else {
2190 		cpu_count(CPU_COUNT_FLT_PRZERO, 1);
2191 
2192 		/*
2193 		 * Page is zero'd and marked dirty by
2194 		 * uvmfault_promote().
2195 		 */
2196 
2197 		UVMHIST_LOG(maphist,"  zero fill anon/page %#jx/%#jx",
2198 		    (uintptr_t)anon, (uintptr_t)pg, 0, 0);
2199 	}
2200 
2201 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2202 }
2203 
2204 /*
2205  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2206  * from the lower page.
2207  */
2208 
2209 int
2210 uvm_fault_lower_enter(
2211 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2212 	struct uvm_object *uobj,
2213 	struct vm_anon *anon, struct vm_page *pg)
2214 {
2215 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2216 	const bool readonly = uvm_pagereadonly_p(pg);
2217 	int error;
2218 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2219 
2220 	/*
2221 	 * Locked:
2222 	 *
2223 	 *	maps(read), amap(if !null), uobj(if !null),
2224 	 *	anon(if !null), pg(if anon), unlock_uobj(if !null)
2225 	 *
2226 	 * Note: pg is either the uobjpage or the new page in the new anon.
2227 	 */
2228 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
2229 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
2230 	KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
2231 	KASSERT((pg->flags & PG_BUSY) != 0);
2232 
2233 	/*
2234 	 * all resources are present.   we can now map it in and free our
2235 	 * resources.
2236 	 */
2237 
2238 	UVMHIST_LOG(maphist,
2239 	    "  MAPPING: case2: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
2240 	    (uintptr_t)ufi->orig_map->pmap, ufi->orig_rvaddr,
2241 	    (uintptr_t)pg, flt->promote);
2242 	KASSERTMSG((flt->access_type & VM_PROT_WRITE) == 0 || !readonly,
2243 	    "promote=%u cow_now=%u access_type=%x enter_prot=%x cow=%u "
2244 	    "entry=%p map=%p orig_rvaddr=%p pg=%p",
2245 	    flt->promote, flt->cow_now, flt->access_type, flt->enter_prot,
2246 	    UVM_ET_ISCOPYONWRITE(ufi->entry), ufi->entry, ufi->orig_map,
2247 	    (void *)ufi->orig_rvaddr, pg);
2248 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 || !readonly);
2249 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2250 	    VM_PAGE_TO_PHYS(pg),
2251 	    readonly ? flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2252 	    flt->access_type | PMAP_CANFAIL |
2253 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2254 
2255 		/*
2256 		 * No need to undo what we did; we can simply think of
2257 		 * this as the pmap throwing away the mapping information.
2258 		 *
2259 		 * We do, however, have to go through the ReFault path,
2260 		 * as the map may change while we're asleep.
2261 		 */
2262 
2263 		/*
2264 		 * ensure that the page is queued in the case that
2265 		 * we just promoted the page.
2266 		 */
2267 
2268 		uvm_pagelock(pg);
2269 		uvm_pageenqueue(pg);
2270 		uvm_pagewakeup(pg);
2271 		uvm_pageunlock(pg);
2272 
2273 		/*
2274 		 * note that pg can't be PG_RELEASED since we did not drop
2275 		 * the object lock since the last time we checked.
2276 		 */
2277 		KASSERT((pg->flags & PG_RELEASED) == 0);
2278 		pg->flags &= ~(PG_BUSY|PG_FAKE);
2279 		UVM_PAGE_OWN(pg, NULL);
2280 
2281 		uvmfault_unlockall(ufi, amap, uobj);
2282 		if (!uvm_reclaimable()) {
2283 			UVMHIST_LOG(maphist,
2284 			    "<- failed.  out of VM",0,0,0,0);
2285 			/* XXX instrumentation */
2286 			error = ENOMEM;
2287 			return error;
2288 		}
2289 		/* XXX instrumentation */
2290 		uvm_wait("flt_pmfail2");
2291 		return ERESTART;
2292 	}
2293 
2294 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2295 
2296 	/*
2297 	 * note that pg can't be PG_RELEASED since we did not drop the object
2298 	 * lock since the last time we checked.
2299 	 */
2300 	KASSERT((pg->flags & PG_RELEASED) == 0);
2301 	uvm_pagelock(pg);
2302 	uvm_pagewakeup(pg);
2303 	uvm_pageunlock(pg);
2304 	pg->flags &= ~(PG_BUSY|PG_FAKE);
2305 	UVM_PAGE_OWN(pg, NULL);
2306 
2307 	pmap_update(ufi->orig_map->pmap);
2308 	uvmfault_unlockall(ufi, amap, uobj);
2309 
2310 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2311 	return 0;
2312 }
2313 
2314 /*
2315  * uvm_fault_lower_done: queue lower center page.
2316  */
2317 
2318 void
2319 uvm_fault_lower_done(
2320 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2321 	struct uvm_object *uobj, struct vm_page *pg)
2322 {
2323 	bool dropswap = false;
2324 
2325 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2326 
2327 	uvm_pagelock(pg);
2328 	if (flt->wire_paging) {
2329 		uvm_pagewire(pg);
2330 		if (pg->flags & PG_AOBJ) {
2331 
2332 			/*
2333 			 * since the now-wired page cannot be paged out,
2334 			 * release its swap resources for others to use.
2335 			 * since an aobj page with no swap cannot be clean,
2336 			 * mark it dirty now.
2337 			 */
2338 
2339 			KASSERT(uobj != NULL);
2340 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
2341 			dropswap = true;
2342 		}
2343 	} else {
2344 		uvm_pageactivate(pg);
2345 	}
2346 	uvm_pageunlock(pg);
2347 
2348 	if (dropswap) {
2349 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2350 	}
2351 }
2352 
2353 
2354 /*
2355  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2356  *
2357  * => map may be read-locked by caller, but MUST NOT be write-locked.
2358  * => if map is read-locked, any operations which may cause map to
2359  *	be write-locked in uvm_fault() must be taken care of by
2360  *	the caller.  See uvm_map_pageable().
2361  */
2362 
2363 int
2364 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2365     vm_prot_t access_type, int maxprot)
2366 {
2367 	vaddr_t va;
2368 	int error;
2369 
2370 	/*
2371 	 * now fault it in a page at a time.   if the fault fails then we have
2372 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2373 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2374 	 */
2375 
2376 	/*
2377 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2378 	 * wiring the last page in the address space, though.
2379 	 */
2380 	if (start > end) {
2381 		return EFAULT;
2382 	}
2383 
2384 	for (va = start; va < end; va += PAGE_SIZE) {
2385 		error = uvm_fault_internal(map, va, access_type,
2386 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2387 		if (error) {
2388 			if (va != start) {
2389 				uvm_fault_unwire(map, start, va);
2390 			}
2391 			return error;
2392 		}
2393 	}
2394 	return 0;
2395 }
2396 
2397 /*
2398  * uvm_fault_unwire(): unwire range of virtual space.
2399  */
2400 
2401 void
2402 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2403 {
2404 	vm_map_lock_read(map);
2405 	uvm_fault_unwire_locked(map, start, end);
2406 	vm_map_unlock_read(map);
2407 }
2408 
2409 /*
2410  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2411  *
2412  * => map must be at least read-locked.
2413  */
2414 
2415 void
2416 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2417 {
2418 	struct vm_map_entry *entry, *oentry;
2419 	pmap_t pmap = vm_map_pmap(map);
2420 	vaddr_t va;
2421 	paddr_t pa;
2422 	struct vm_page *pg;
2423 
2424 	/*
2425 	 * we assume that the area we are unwiring has actually been wired
2426 	 * in the first place.   this means that we should be able to extract
2427 	 * the PAs from the pmap.   we also lock out the page daemon so that
2428 	 * we can call uvm_pageunwire.
2429 	 */
2430 
2431 	/*
2432 	 * find the beginning map entry for the region.
2433 	 */
2434 
2435 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2436 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2437 		panic("uvm_fault_unwire_locked: address not in map");
2438 
2439 	oentry = NULL;
2440 	for (va = start; va < end; va += PAGE_SIZE) {
2441 
2442 		/*
2443 		 * find the map entry for the current address.
2444 		 */
2445 
2446 		KASSERT(va >= entry->start);
2447 		while (va >= entry->end) {
2448 			KASSERT(entry->next != &map->header &&
2449 				entry->next->start <= entry->end);
2450 			entry = entry->next;
2451 		}
2452 
2453 		/*
2454 		 * lock it.
2455 		 */
2456 
2457 		if (entry != oentry) {
2458 			if (oentry != NULL) {
2459 				uvm_map_unlock_entry(oentry);
2460 			}
2461 			uvm_map_lock_entry(entry, RW_WRITER);
2462 			oentry = entry;
2463 		}
2464 
2465 		/*
2466 		 * if the entry is no longer wired, tell the pmap.
2467 		 */
2468 
2469 		if (!pmap_extract(pmap, va, &pa))
2470 			continue;
2471 
2472 		if (VM_MAPENT_ISWIRED(entry) == 0)
2473 			pmap_unwire(pmap, va);
2474 
2475 		pg = PHYS_TO_VM_PAGE(pa);
2476 		if (pg) {
2477 			uvm_pagelock(pg);
2478 			uvm_pageunwire(pg);
2479 			uvm_pageunlock(pg);
2480 		}
2481 	}
2482 
2483 	if (oentry != NULL) {
2484 		uvm_map_unlock_entry(entry);
2485 	}
2486 }
2487