xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 5dd36a3bc8bf2a9dec29ceb6349550414570c447)
1 /*	$NetBSD: uvm_fault.c,v 1.217 2020/02/24 12:38:57 rin 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 1.217 2020/02/24 12:38:57 rin 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 			pg->flags |= PG_WANTED;
333 			cpu_count(CPU_COUNT_FLTPGWAIT, 1);
334 
335 			/*
336 			 * The last unlock must be an atomic unlock and wait
337 			 * on the owner of page.
338 			 */
339 
340 			if (pg->uobject) {
341 				/* Owner of page is UVM object. */
342 				uvmfault_unlockall(ufi, amap, NULL);
343 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
344 				    0,0,0);
345 				UVM_UNLOCK_AND_WAIT_RW(pg,
346 				    pg->uobject->vmobjlock,
347 				    false, "anonget1", 0);
348 			} else {
349 				/* Owner of page is anon. */
350 				uvmfault_unlockall(ufi, NULL, NULL);
351 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
352 				    0,0,0);
353 				UVM_UNLOCK_AND_WAIT_RW(pg, anon->an_lock,
354 				    false, "anonget2", 0);
355 			}
356 		} else {
357 #if defined(VMSWAP)
358 			/*
359 			 * No page, therefore allocate one.
360 			 */
361 
362 			pg = uvm_pagealloc(NULL,
363 			    ufi != NULL ? ufi->orig_rvaddr : 0,
364 			    anon, ufi != NULL ? UVM_FLAG_COLORMATCH : 0);
365 			if (pg == NULL) {
366 				/* Out of memory.  Wait a little. */
367 				uvmfault_unlockall(ufi, amap, NULL);
368 				cpu_count(CPU_COUNT_FLTNORAM, 1);
369 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
370 				    0,0,0);
371 				if (!uvm_reclaimable()) {
372 					return ENOMEM;
373 				}
374 				uvm_wait("flt_noram1");
375 			} else {
376 				/* PG_BUSY bit is set. */
377 				we_own = true;
378 				uvmfault_unlockall(ufi, amap, NULL);
379 
380 				/*
381 				 * Pass a PG_BUSY+PG_FAKE clean page into
382 				 * the uvm_swap_get() function with all data
383 				 * structures unlocked.  Note that it is OK
384 				 * to read an_swslot here, because we hold
385 				 * PG_BUSY on the page.
386 				 */
387 				cpu_count(CPU_COUNT_PAGEINS, 1);
388 				error = uvm_swap_get(pg, anon->an_swslot,
389 				    PGO_SYNCIO);
390 
391 				/*
392 				 * We clean up after the I/O below in the
393 				 * 'we_own' case.
394 				 */
395 			}
396 #else
397 			panic("%s: no page", __func__);
398 #endif /* defined(VMSWAP) */
399 		}
400 
401 		/*
402 		 * Re-lock the map and anon.
403 		 */
404 
405 		locked = uvmfault_relock(ufi);
406 		if (locked || we_own) {
407 			rw_enter(anon->an_lock, RW_WRITER);
408 		}
409 
410 		/*
411 		 * If we own the page (i.e. we set PG_BUSY), then we need
412 		 * to clean up after the I/O.  There are three cases to
413 		 * consider:
414 		 *
415 		 * 1) Page was released during I/O: free anon and ReFault.
416 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
417 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
418 		 *    case (i.e. drop anon lock if not locked).
419 		 */
420 
421 		if (we_own) {
422 #if defined(VMSWAP)
423 			if (pg->flags & PG_WANTED) {
424 				wakeup(pg);
425 			}
426 			if (error) {
427 
428 				/*
429 				 * Remove the swap slot from the anon and
430 				 * mark the anon as having no real slot.
431 				 * Do not free the swap slot, thus preventing
432 				 * it from being used again.
433 				 */
434 
435 				if (anon->an_swslot > 0) {
436 					uvm_swap_markbad(anon->an_swslot, 1);
437 				}
438 				anon->an_swslot = SWSLOT_BAD;
439 
440 				if ((pg->flags & PG_RELEASED) != 0) {
441 					goto released;
442 				}
443 
444 				/*
445 				 * Note: page was never !PG_BUSY, so it
446 				 * cannot be mapped and thus no need to
447 				 * pmap_page_protect() it.
448 				 */
449 
450 				uvm_pagefree(pg);
451 
452 				if (locked) {
453 					uvmfault_unlockall(ufi, NULL, NULL);
454 				}
455 				rw_exit(anon->an_lock);
456 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
457 				return error;
458 			}
459 
460 			if ((pg->flags & PG_RELEASED) != 0) {
461 released:
462 				KASSERT(anon->an_ref == 0);
463 
464 				/*
465 				 * Released while we had unlocked amap.
466 				 */
467 
468 				if (locked) {
469 					uvmfault_unlockall(ufi, NULL, NULL);
470 				}
471 				uvm_anon_release(anon);
472 
473 				if (error) {
474 					UVMHIST_LOG(maphist,
475 					    "<- ERROR/RELEASED", 0,0,0,0);
476 					return error;
477 				}
478 
479 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
480 				return ERESTART;
481 			}
482 
483 			/*
484 			 * We have successfully read the page, activate it.
485 			 */
486 
487 			uvm_pagelock(pg);
488 			uvm_pageactivate(pg);
489 			uvm_pageunlock(pg);
490 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
491 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_UNKNOWN);
492 			UVM_PAGE_OWN(pg, NULL);
493 #else
494 			panic("%s: we_own", __func__);
495 #endif /* defined(VMSWAP) */
496 		}
497 
498 		/*
499 		 * We were not able to re-lock the map - restart the fault.
500 		 */
501 
502 		if (!locked) {
503 			if (we_own) {
504 				rw_exit(anon->an_lock);
505 			}
506 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
507 			return ERESTART;
508 		}
509 
510 		/*
511 		 * Verify that no one has touched the amap and moved
512 		 * the anon on us.
513 		 */
514 
515 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
516 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
517 
518 			uvmfault_unlockall(ufi, amap, NULL);
519 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
520 			return ERESTART;
521 		}
522 
523 		/*
524 		 * Retry..
525 		 */
526 
527 		cpu_count(CPU_COUNT_FLTANRETRY, 1);
528 		continue;
529 	}
530 	/*NOTREACHED*/
531 }
532 
533 /*
534  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
535  *
536  *	1. allocate an anon and a page.
537  *	2. fill its contents.
538  *	3. put it into amap.
539  *
540  * => if we fail (result != 0) we unlock everything.
541  * => on success, return a new locked anon via 'nanon'.
542  *    (*nanon)->an_page will be a resident, locked, dirty page.
543  * => it's caller's responsibility to put the promoted nanon->an_page to the
544  *    page queue.
545  */
546 
547 static int
548 uvmfault_promote(struct uvm_faultinfo *ufi,
549     struct vm_anon *oanon,
550     struct vm_page *uobjpage,
551     struct vm_anon **nanon, /* OUT: allocated anon */
552     struct vm_anon **spare)
553 {
554 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
555 	struct uvm_object *uobj;
556 	struct vm_anon *anon;
557 	struct vm_page *pg;
558 	struct vm_page *opg;
559 	int error;
560 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
561 
562 	if (oanon) {
563 		/* anon COW */
564 		opg = oanon->an_page;
565 		KASSERT(opg != NULL);
566 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
567 	} else if (uobjpage != PGO_DONTCARE) {
568 		/* object-backed COW */
569 		opg = uobjpage;
570 	} else {
571 		/* ZFOD */
572 		opg = NULL;
573 	}
574 	if (opg != NULL) {
575 		uobj = opg->uobject;
576 	} else {
577 		uobj = NULL;
578 	}
579 
580 	KASSERT(amap != NULL);
581 	KASSERT(uobjpage != NULL);
582 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
583 	KASSERT(rw_write_held(amap->am_lock));
584 	KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock);
585 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
586 
587 	if (*spare != NULL) {
588 		anon = *spare;
589 		*spare = NULL;
590 	} else {
591 		anon = uvm_analloc();
592 	}
593 	if (anon) {
594 
595 		/*
596 		 * The new anon is locked.
597 		 *
598 		 * if opg == NULL, we want a zero'd, dirty page,
599 		 * so have uvm_pagealloc() do that for us.
600 		 */
601 
602 		KASSERT(anon->an_lock == NULL);
603 		anon->an_lock = amap->am_lock;
604 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
605 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
606 		if (pg == NULL) {
607 			anon->an_lock = NULL;
608 		}
609 	} else {
610 		pg = NULL;
611 	}
612 
613 	/*
614 	 * out of memory resources?
615 	 */
616 
617 	if (pg == NULL) {
618 		/* save anon for the next try. */
619 		if (anon != NULL) {
620 			*spare = anon;
621 		}
622 
623 		/* unlock and fail ... */
624 		uvm_page_unbusy(&uobjpage, 1);
625 		uvmfault_unlockall(ufi, amap, uobj);
626 		if (!uvm_reclaimable()) {
627 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
628 			cpu_count(CPU_COUNT_FLTNOANON, 1);
629 			error = ENOMEM;
630 			goto done;
631 		}
632 
633 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
634 		cpu_count(CPU_COUNT_FLTNORAM, 1);
635 		uvm_wait("flt_noram5");
636 		error = ERESTART;
637 		goto done;
638 	}
639 
640 	/* copy page [pg now dirty] */
641 	if (opg) {
642 		uvm_pagecopy(opg, pg);
643 	}
644 	KASSERT(uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_DIRTY);
645 
646 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
647 	    oanon != NULL);
648 
649 	*nanon = anon;
650 	error = 0;
651 done:
652 	return error;
653 }
654 
655 /*
656  * Update statistics after fault resolution.
657  * - maxrss
658  */
659 void
660 uvmfault_update_stats(struct uvm_faultinfo *ufi)
661 {
662 	struct vm_map		*map;
663 	struct vmspace 		*vm;
664 	struct proc		*p;
665 	vsize_t			 res;
666 
667 	map = ufi->orig_map;
668 
669 	p = curproc;
670 	KASSERT(p != NULL);
671 	vm = p->p_vmspace;
672 
673 	if (&vm->vm_map != map)
674 		return;
675 
676 	res = pmap_resident_count(map->pmap);
677 	if (vm->vm_rssmax < res)
678 		vm->vm_rssmax = res;
679 }
680 
681 /*
682  *   F A U L T   -   m a i n   e n t r y   p o i n t
683  */
684 
685 /*
686  * uvm_fault: page fault handler
687  *
688  * => called from MD code to resolve a page fault
689  * => VM data structures usually should be unlocked.   however, it is
690  *	possible to call here with the main map locked if the caller
691  *	gets a write lock, sets it recusive, and then calls us (c.f.
692  *	uvm_map_pageable).   this should be avoided because it keeps
693  *	the map locked off during I/O.
694  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
695  */
696 
697 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
698 			 ~VM_PROT_WRITE : VM_PROT_ALL)
699 
700 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
701 #define UVM_FAULT_WIRE		(1 << 0)
702 #define UVM_FAULT_MAXPROT	(1 << 1)
703 
704 struct uvm_faultctx {
705 
706 	/*
707 	 * the following members are set up by uvm_fault_check() and
708 	 * read-only after that.
709 	 *
710 	 * note that narrow is used by uvm_fault_check() to change
711 	 * the behaviour after ERESTART.
712 	 *
713 	 * most of them might change after RESTART if the underlying
714 	 * map entry has been changed behind us.  an exception is
715 	 * wire_paging, which does never change.
716 	 */
717 	vm_prot_t access_type;
718 	vaddr_t startva;
719 	int npages;
720 	int centeridx;
721 	bool narrow;		/* work on a single requested page only */
722 	bool wire_mapping;	/* request a PMAP_WIRED mapping
723 				   (UVM_FAULT_WIRE or VM_MAPENT_ISWIRED) */
724 	bool wire_paging;	/* request uvm_pagewire
725 				   (true for UVM_FAULT_WIRE) */
726 	bool cow_now;		/* VM_PROT_WRITE is actually requested
727 				   (ie. should break COW and page loaning) */
728 
729 	/*
730 	 * enter_prot is set up by uvm_fault_check() and clamped
731 	 * (ie. drop the VM_PROT_WRITE bit) in various places in case
732 	 * of !cow_now.
733 	 */
734 	vm_prot_t enter_prot;	/* prot at which we want to enter pages in */
735 
736 	/*
737 	 * the following member is for uvmfault_promote() and ERESTART.
738 	 */
739 	struct vm_anon *anon_spare;
740 
741 	/*
742 	 * the folloing is actually a uvm_fault_lower() internal.
743 	 * it's here merely for debugging.
744 	 * (or due to the mechanical separation of the function?)
745 	 */
746 	bool promote;
747 };
748 
749 static inline int	uvm_fault_check(
750 			    struct uvm_faultinfo *, struct uvm_faultctx *,
751 			    struct vm_anon ***, bool);
752 
753 static int		uvm_fault_upper(
754 			    struct uvm_faultinfo *, struct uvm_faultctx *,
755 			    struct vm_anon **);
756 static inline int	uvm_fault_upper_lookup(
757 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
758 			    struct vm_anon **, struct vm_page **);
759 static inline void	uvm_fault_upper_neighbor(
760 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
761 			    vaddr_t, struct vm_page *, bool);
762 static inline int	uvm_fault_upper_loan(
763 			    struct uvm_faultinfo *, struct uvm_faultctx *,
764 			    struct vm_anon *, struct uvm_object **);
765 static inline int	uvm_fault_upper_promote(
766 			    struct uvm_faultinfo *, struct uvm_faultctx *,
767 			    struct uvm_object *, struct vm_anon *);
768 static inline int	uvm_fault_upper_direct(
769 			    struct uvm_faultinfo *, struct uvm_faultctx *,
770 			    struct uvm_object *, struct vm_anon *);
771 static int		uvm_fault_upper_enter(
772 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
773 			    struct uvm_object *, struct vm_anon *,
774 			    struct vm_page *, struct vm_anon *);
775 static inline void	uvm_fault_upper_done(
776 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
777 			    struct vm_anon *, struct vm_page *);
778 
779 static int		uvm_fault_lower(
780 			    struct uvm_faultinfo *, struct uvm_faultctx *,
781 			    struct vm_page **);
782 static inline void	uvm_fault_lower_lookup(
783 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
784 			    struct vm_page **);
785 static inline void	uvm_fault_lower_neighbor(
786 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
787 			    vaddr_t, struct vm_page *);
788 static inline int	uvm_fault_lower_io(
789 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
790 			    struct uvm_object **, struct vm_page **);
791 static inline int	uvm_fault_lower_direct(
792 			    struct uvm_faultinfo *, struct uvm_faultctx *,
793 			    struct uvm_object *, struct vm_page *);
794 static inline int	uvm_fault_lower_direct_loan(
795 			    struct uvm_faultinfo *, struct uvm_faultctx *,
796 			    struct uvm_object *, struct vm_page **,
797 			    struct vm_page **);
798 static inline int	uvm_fault_lower_promote(
799 			    struct uvm_faultinfo *, struct uvm_faultctx *,
800 			    struct uvm_object *, struct vm_page *);
801 static int		uvm_fault_lower_enter(
802 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
803 			    struct uvm_object *,
804 			    struct vm_anon *, struct vm_page *);
805 static inline void	uvm_fault_lower_done(
806 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
807 			    struct uvm_object *, struct vm_page *);
808 
809 int
810 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
811     vm_prot_t access_type, int fault_flag)
812 {
813 	struct uvm_faultinfo ufi;
814 	struct uvm_faultctx flt = {
815 		.access_type = access_type,
816 
817 		/* don't look for neighborhood * pages on "wire" fault */
818 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
819 
820 		/* "wire" fault causes wiring of both mapping and paging */
821 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
822 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
823 	};
824 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
825 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
826 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
827 	int error;
828 
829 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
830 
831 	UVMHIST_LOG(maphist, "(map=%#jx, vaddr=%#jx, at=%jd, ff=%jd)",
832 	      (uintptr_t)orig_map, vaddr, access_type, fault_flag);
833 
834 	/* Don't count anything until user interaction is possible */
835 	kpreempt_disable();
836 	if (__predict_true(start_init_exec)) {
837 		struct cpu_info *ci = curcpu();
838 		CPU_COUNT(CPU_COUNT_NFAULT, 1);
839 		/* Don't flood RNG subsystem with samples. */
840 		if (++(ci->ci_faultrng) == 503) {
841 			ci->ci_faultrng = 0;
842 			rnd_add_uint32(&curcpu()->ci_data.cpu_uvm->rs,
843 			    sizeof(vaddr_t) == sizeof(uint32_t) ?
844 			    (uint32_t)vaddr : sizeof(vaddr_t) ==
845 			    sizeof(uint64_t) ?
846 			    (uint32_t)vaddr :
847 			    (uint32_t)ci->ci_counts[CPU_COUNT_NFAULT]);
848 		}
849 	}
850 	kpreempt_enable();
851 
852 	/*
853 	 * init the IN parameters in the ufi
854 	 */
855 
856 	ufi.orig_map = orig_map;
857 	ufi.orig_rvaddr = trunc_page(vaddr);
858 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
859 
860 	error = ERESTART;
861 	while (error == ERESTART) { /* ReFault: */
862 		anons = anons_store;
863 		pages = pages_store;
864 
865 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
866 		if (error != 0)
867 			continue;
868 
869 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
870 		if (error != 0)
871 			continue;
872 
873 		if (pages[flt.centeridx] == PGO_DONTCARE)
874 			error = uvm_fault_upper(&ufi, &flt, anons);
875 		else {
876 			struct uvm_object * const uobj =
877 			    ufi.entry->object.uvm_obj;
878 
879 			if (uobj && uobj->pgops->pgo_fault != NULL) {
880 				/*
881 				 * invoke "special" fault routine.
882 				 */
883 				rw_enter(uobj->vmobjlock, RW_WRITER);
884 				/* locked: maps(read), amap(if there), uobj */
885 				error = uobj->pgops->pgo_fault(&ufi,
886 				    flt.startva, pages, flt.npages,
887 				    flt.centeridx, flt.access_type,
888 				    PGO_LOCKED|PGO_SYNCIO);
889 
890 				/*
891 				 * locked: nothing, pgo_fault has unlocked
892 				 * everything
893 				 */
894 
895 				/*
896 				 * object fault routine responsible for
897 				 * pmap_update().
898 				 */
899 
900 				/*
901 				 * Wake up the pagedaemon if the fault method
902 				 * failed for lack of memory but some can be
903 				 * reclaimed.
904 				 */
905 				if (error == ENOMEM && uvm_reclaimable()) {
906 					uvm_wait("pgo_fault");
907 					error = ERESTART;
908 				}
909 			} else {
910 				error = uvm_fault_lower(&ufi, &flt, pages);
911 			}
912 		}
913 	}
914 
915 	if (flt.anon_spare != NULL) {
916 		flt.anon_spare->an_ref--;
917 		KASSERT(flt.anon_spare->an_ref == 0);
918 		KASSERT(flt.anon_spare->an_lock == NULL);
919 		uvm_anon_free(flt.anon_spare);
920 	}
921 	return error;
922 }
923 
924 /*
925  * uvm_fault_check: check prot, handle needs-copy, etc.
926  *
927  *	1. lookup entry.
928  *	2. check protection.
929  *	3. adjust fault condition (mainly for simulated fault).
930  *	4. handle needs-copy (lazy amap copy).
931  *	5. establish range of interest for neighbor fault (aka pre-fault).
932  *	6. look up anons (if amap exists).
933  *	7. flush pages (if MADV_SEQUENTIAL)
934  *
935  * => called with nothing locked.
936  * => if we fail (result != 0) we unlock everything.
937  * => initialize/adjust many members of flt.
938  */
939 
940 static int
941 uvm_fault_check(
942 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
943 	struct vm_anon ***ranons, bool maxprot)
944 {
945 	struct vm_amap *amap;
946 	struct uvm_object *uobj;
947 	vm_prot_t check_prot;
948 	int nback, nforw;
949 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
950 
951 	/*
952 	 * lookup and lock the maps
953 	 */
954 
955 	if (uvmfault_lookup(ufi, false) == false) {
956 		UVMHIST_LOG(maphist, "<- no mapping @ %#jx", ufi->orig_rvaddr,
957 		    0,0,0);
958 		return EFAULT;
959 	}
960 	/* locked: maps(read) */
961 
962 #ifdef DIAGNOSTIC
963 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
964 		printf("Page fault on non-pageable map:\n");
965 		printf("ufi->map = %p\n", ufi->map);
966 		printf("ufi->orig_map = %p\n", ufi->orig_map);
967 		printf("ufi->orig_rvaddr = %#lx\n", (u_long) ufi->orig_rvaddr);
968 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
969 	}
970 #endif
971 
972 	/*
973 	 * check protection
974 	 */
975 
976 	check_prot = maxprot ?
977 	    ufi->entry->max_protection : ufi->entry->protection;
978 	if ((check_prot & flt->access_type) != flt->access_type) {
979 		UVMHIST_LOG(maphist,
980 		    "<- protection failure (prot=%#jx, access=%#jx)",
981 		    ufi->entry->protection, flt->access_type, 0, 0);
982 		uvmfault_unlockmaps(ufi, false);
983 		return EFAULT;
984 	}
985 
986 	/*
987 	 * "enter_prot" is the protection we want to enter the page in at.
988 	 * for certain pages (e.g. copy-on-write pages) this protection can
989 	 * be more strict than ufi->entry->protection.  "wired" means either
990 	 * the entry is wired or we are fault-wiring the pg.
991 	 */
992 
993 	flt->enter_prot = ufi->entry->protection;
994 	if (VM_MAPENT_ISWIRED(ufi->entry)) {
995 		flt->wire_mapping = true;
996 		flt->wire_paging = true;
997 		flt->narrow = true;
998 	}
999 
1000 	if (flt->wire_mapping) {
1001 		flt->access_type = flt->enter_prot; /* full access for wired */
1002 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
1003 	} else {
1004 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
1005 	}
1006 
1007 	flt->promote = false;
1008 
1009 	/*
1010 	 * handle "needs_copy" case.   if we need to copy the amap we will
1011 	 * have to drop our readlock and relock it with a write lock.  (we
1012 	 * need a write lock to change anything in a map entry [e.g.
1013 	 * needs_copy]).
1014 	 */
1015 
1016 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
1017 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
1018 			KASSERT(!maxprot);
1019 			/* need to clear */
1020 			UVMHIST_LOG(maphist,
1021 			    "  need to clear needs_copy and refault",0,0,0,0);
1022 			uvmfault_unlockmaps(ufi, false);
1023 			uvmfault_amapcopy(ufi);
1024 			cpu_count(CPU_COUNT_FLTAMCOPY, 1);
1025 			return ERESTART;
1026 
1027 		} else {
1028 
1029 			/*
1030 			 * ensure that we pmap_enter page R/O since
1031 			 * needs_copy is still true
1032 			 */
1033 
1034 			flt->enter_prot &= ~VM_PROT_WRITE;
1035 		}
1036 	}
1037 
1038 	/*
1039 	 * identify the players
1040 	 */
1041 
1042 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
1043 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
1044 
1045 	/*
1046 	 * check for a case 0 fault.  if nothing backing the entry then
1047 	 * error now.
1048 	 */
1049 
1050 	if (amap == NULL && uobj == NULL) {
1051 		uvmfault_unlockmaps(ufi, false);
1052 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
1053 		return EFAULT;
1054 	}
1055 
1056 	/*
1057 	 * establish range of interest based on advice from mapper
1058 	 * and then clip to fit map entry.   note that we only want
1059 	 * to do this the first time through the fault.   if we
1060 	 * ReFault we will disable this by setting "narrow" to true.
1061 	 */
1062 
1063 	if (flt->narrow == false) {
1064 
1065 		/* wide fault (!narrow) */
1066 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
1067 			 ufi->entry->advice);
1068 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
1069 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
1070 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
1071 		/*
1072 		 * note: "-1" because we don't want to count the
1073 		 * faulting page as forw
1074 		 */
1075 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1076 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1077 			     PAGE_SHIFT) - 1);
1078 		flt->npages = nback + nforw + 1;
1079 		flt->centeridx = nback;
1080 
1081 		flt->narrow = true;	/* ensure only once per-fault */
1082 
1083 	} else {
1084 
1085 		/* narrow fault! */
1086 		nback = nforw = 0;
1087 		flt->startva = ufi->orig_rvaddr;
1088 		flt->npages = 1;
1089 		flt->centeridx = 0;
1090 
1091 	}
1092 	/* offset from entry's start to pgs' start */
1093 	const voff_t eoff = flt->startva - ufi->entry->start;
1094 
1095 	/* locked: maps(read) */
1096 	UVMHIST_LOG(maphist, "  narrow=%jd, back=%jd, forw=%jd, startva=%#jx",
1097 		    flt->narrow, nback, nforw, flt->startva);
1098 	UVMHIST_LOG(maphist, "  entry=%#jx, amap=%#jx, obj=%#jx",
1099 	    (uintptr_t)ufi->entry, (uintptr_t)amap, (uintptr_t)uobj, 0);
1100 
1101 	/*
1102 	 * if we've got an amap, lock it and extract current anons.
1103 	 */
1104 
1105 	if (amap) {
1106 		amap_lock(amap, RW_WRITER);
1107 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1108 	} else {
1109 		*ranons = NULL;	/* to be safe */
1110 	}
1111 
1112 	/* locked: maps(read), amap(if there) */
1113 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1114 
1115 	/*
1116 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1117 	 * now and then forget about them (for the rest of the fault).
1118 	 */
1119 
1120 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1121 
1122 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1123 		    0,0,0,0);
1124 		/* flush back-page anons? */
1125 		if (amap)
1126 			uvmfault_anonflush(*ranons, nback);
1127 
1128 		/* flush object? */
1129 		if (uobj) {
1130 			voff_t uoff;
1131 
1132 			uoff = ufi->entry->offset + eoff;
1133 			rw_enter(uobj->vmobjlock, RW_WRITER);
1134 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1135 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1136 		}
1137 
1138 		/* now forget about the backpages */
1139 		if (amap)
1140 			*ranons += nback;
1141 		flt->startva += (nback << PAGE_SHIFT);
1142 		flt->npages -= nback;
1143 		flt->centeridx = 0;
1144 	}
1145 	/*
1146 	 * => startva is fixed
1147 	 * => npages is fixed
1148 	 */
1149 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1150 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1151 	    flt->startva + (flt->npages << PAGE_SHIFT));
1152 	return 0;
1153 }
1154 
1155 /*
1156  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1157  *
1158  * iterate range of interest:
1159  *	1. check if h/w mapping exists.  if yes, we don't care
1160  *	2. check if anon exists.  if not, page is lower.
1161  *	3. if anon exists, enter h/w mapping for neighbors.
1162  *
1163  * => called with amap locked (if exists).
1164  */
1165 
1166 static int
1167 uvm_fault_upper_lookup(
1168 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1169 	struct vm_anon **anons, struct vm_page **pages)
1170 {
1171 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1172 	int lcv;
1173 	vaddr_t currva;
1174 	bool shadowed __unused;
1175 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1176 
1177 	/* locked: maps(read), amap(if there) */
1178 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1179 
1180 	/*
1181 	 * map in the backpages and frontpages we found in the amap in hopes
1182 	 * of preventing future faults.    we also init the pages[] array as
1183 	 * we go.
1184 	 */
1185 
1186 	currva = flt->startva;
1187 	shadowed = false;
1188 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1189 		/*
1190 		 * don't play with VAs that are already mapped
1191 		 * (except for center)
1192 		 */
1193 		if (lcv != flt->centeridx &&
1194 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1195 			pages[lcv] = PGO_DONTCARE;
1196 			continue;
1197 		}
1198 
1199 		/*
1200 		 * unmapped or center page.   check if any anon at this level.
1201 		 */
1202 		if (amap == NULL || anons[lcv] == NULL) {
1203 			pages[lcv] = NULL;
1204 			continue;
1205 		}
1206 
1207 		/*
1208 		 * check for present page and map if possible.   re-activate it.
1209 		 */
1210 
1211 		pages[lcv] = PGO_DONTCARE;
1212 		if (lcv == flt->centeridx) {	/* save center for later! */
1213 			shadowed = true;
1214 			continue;
1215 		}
1216 
1217 		struct vm_anon *anon = anons[lcv];
1218 		struct vm_page *pg = anon->an_page;
1219 
1220 		KASSERT(anon->an_lock == amap->am_lock);
1221 
1222 		/* Ignore loaned and busy pages. */
1223 		if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0) {
1224 			uvm_fault_upper_neighbor(ufi, flt, currva,
1225 			    pg, anon->an_ref > 1);
1226 		}
1227 	}
1228 
1229 	/* locked: maps(read), amap(if there) */
1230 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1231 	/* (shadowed == true) if there is an anon at the faulting address */
1232 	UVMHIST_LOG(maphist, "  shadowed=%jd, will_get=%jd", shadowed,
1233 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1234 
1235 	/*
1236 	 * note that if we are really short of RAM we could sleep in the above
1237 	 * call to pmap_enter with everything locked.   bad?
1238 	 *
1239 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1240 	 * XXX case.  --thorpej
1241 	 */
1242 
1243 	return 0;
1244 }
1245 
1246 /*
1247  * uvm_fault_upper_neighbor: enter single upper neighbor page.
1248  *
1249  * => called with amap and anon locked.
1250  */
1251 
1252 static void
1253 uvm_fault_upper_neighbor(
1254 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1255 	vaddr_t currva, struct vm_page *pg, bool readonly)
1256 {
1257 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1258 
1259 	/* locked: amap, anon */
1260 
1261 	KASSERT(pg->uobject == NULL);
1262 	KASSERT(pg->uanon != NULL);
1263 	KASSERT(rw_write_held(pg->uanon->an_lock));
1264 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1265 
1266 	uvm_pagelock(pg);
1267 	uvm_pageenqueue(pg);
1268 	uvm_pageunlock(pg);
1269 	UVMHIST_LOG(maphist,
1270 	    "  MAPPING: n anon: pm=%#jx, va=%#jx, pg=%#jx",
1271 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
1272 	cpu_count(CPU_COUNT_FLTNAMAP, 1);
1273 
1274 	/*
1275 	 * Since this page isn't the page that's actually faulting,
1276 	 * ignore pmap_enter() failures; it's not critical that we
1277 	 * enter these right now.
1278 	 */
1279 
1280 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1281 	    VM_PAGE_TO_PHYS(pg),
1282 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1283 	    flt->enter_prot,
1284 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1285 
1286 	pmap_update(ufi->orig_map->pmap);
1287 }
1288 
1289 /*
1290  * uvm_fault_upper: handle upper fault.
1291  *
1292  *	1. acquire anon lock.
1293  *	2. get anon.  let uvmfault_anonget do the dirty work.
1294  *	3. handle loan.
1295  *	4. dispatch direct or promote handlers.
1296  */
1297 
1298 static int
1299 uvm_fault_upper(
1300 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1301 	struct vm_anon **anons)
1302 {
1303 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1304 	struct vm_anon * const anon = anons[flt->centeridx];
1305 	struct uvm_object *uobj;
1306 	int error;
1307 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1308 
1309 	/* locked: maps(read), amap, anon */
1310 	KASSERT(rw_write_held(amap->am_lock));
1311 	KASSERT(anon->an_lock == amap->am_lock);
1312 
1313 	/*
1314 	 * handle case 1: fault on an anon in our amap
1315 	 */
1316 
1317 	UVMHIST_LOG(maphist, "  case 1 fault: anon=%#jx",
1318 	    (uintptr_t)anon, 0, 0, 0);
1319 
1320 	/*
1321 	 * no matter if we have case 1A or case 1B we are going to need to
1322 	 * have the anon's memory resident.   ensure that now.
1323 	 */
1324 
1325 	/*
1326 	 * let uvmfault_anonget do the dirty work.
1327 	 * if it fails (!OK) it will unlock everything for us.
1328 	 * if it succeeds, locks are still valid and locked.
1329 	 * also, if it is OK, then the anon's page is on the queues.
1330 	 * if the page is on loan from a uvm_object, then anonget will
1331 	 * lock that object for us if it does not fail.
1332 	 */
1333 
1334 	error = uvmfault_anonget(ufi, amap, anon);
1335 	switch (error) {
1336 	case 0:
1337 		break;
1338 
1339 	case ERESTART:
1340 		return ERESTART;
1341 
1342 	case EAGAIN:
1343 		kpause("fltagain1", false, hz/2, NULL);
1344 		return ERESTART;
1345 
1346 	default:
1347 		return error;
1348 	}
1349 
1350 	/*
1351 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1352 	 */
1353 
1354 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1355 
1356 	/* locked: maps(read), amap, anon, uobj(if one) */
1357 	KASSERT(rw_write_held(amap->am_lock));
1358 	KASSERT(anon->an_lock == amap->am_lock);
1359 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1360 
1361 	/*
1362 	 * special handling for loaned pages
1363 	 */
1364 
1365 	if (anon->an_page->loan_count) {
1366 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1367 		if (error != 0)
1368 			return error;
1369 	}
1370 
1371 	/*
1372 	 * if we are case 1B then we will need to allocate a new blank
1373 	 * anon to transfer the data into.   note that we have a lock
1374 	 * on anon, so no one can busy or release the page until we are done.
1375 	 * also note that the ref count can't drop to zero here because
1376 	 * it is > 1 and we are only dropping one ref.
1377 	 *
1378 	 * in the (hopefully very rare) case that we are out of RAM we
1379 	 * will unlock, wait for more RAM, and refault.
1380 	 *
1381 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1382 	 */
1383 
1384 	if (flt->cow_now && anon->an_ref > 1) {
1385 		flt->promote = true;
1386 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1387 	} else {
1388 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1389 	}
1390 	return error;
1391 }
1392 
1393 /*
1394  * uvm_fault_upper_loan: handle loaned upper page.
1395  *
1396  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1397  *	2. if cow'ing now, and if ref count is 1, break loan.
1398  */
1399 
1400 static int
1401 uvm_fault_upper_loan(
1402 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1403 	struct vm_anon *anon, struct uvm_object **ruobj)
1404 {
1405 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1406 	int error = 0;
1407 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1408 
1409 	if (!flt->cow_now) {
1410 
1411 		/*
1412 		 * for read faults on loaned pages we just cap the
1413 		 * protection at read-only.
1414 		 */
1415 
1416 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1417 
1418 	} else {
1419 		/*
1420 		 * note that we can't allow writes into a loaned page!
1421 		 *
1422 		 * if we have a write fault on a loaned page in an
1423 		 * anon then we need to look at the anon's ref count.
1424 		 * if it is greater than one then we are going to do
1425 		 * a normal copy-on-write fault into a new anon (this
1426 		 * is not a problem).  however, if the reference count
1427 		 * is one (a case where we would normally allow a
1428 		 * write directly to the page) then we need to kill
1429 		 * the loan before we continue.
1430 		 */
1431 
1432 		/* >1 case is already ok */
1433 		if (anon->an_ref == 1) {
1434 			error = uvm_loanbreak_anon(anon, *ruobj);
1435 			if (error != 0) {
1436 				uvmfault_unlockall(ufi, amap, *ruobj);
1437 				uvm_wait("flt_noram2");
1438 				return ERESTART;
1439 			}
1440 			/* if we were a loan receiver uobj is gone */
1441 			if (*ruobj)
1442 				*ruobj = NULL;
1443 		}
1444 	}
1445 	return error;
1446 }
1447 
1448 /*
1449  * uvm_fault_upper_promote: promote upper page.
1450  *
1451  *	1. call uvmfault_promote.
1452  *	2. enqueue page.
1453  *	3. deref.
1454  *	4. pass page to uvm_fault_upper_enter.
1455  */
1456 
1457 static int
1458 uvm_fault_upper_promote(
1459 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1460 	struct uvm_object *uobj, struct vm_anon *anon)
1461 {
1462 	struct vm_anon * const oanon = anon;
1463 	struct vm_page *pg;
1464 	int error;
1465 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1466 
1467 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1468 	cpu_count(CPU_COUNT_FLT_ACOW, 1);
1469 
1470 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1471 	    &flt->anon_spare);
1472 	switch (error) {
1473 	case 0:
1474 		break;
1475 	case ERESTART:
1476 		return ERESTART;
1477 	default:
1478 		return error;
1479 	}
1480 
1481 	KASSERT(anon == NULL || anon->an_lock == oanon->an_lock);
1482 
1483 	pg = anon->an_page;
1484 	/* uvm_fault_upper_done will activate the page */
1485 	uvm_pagelock(pg);
1486 	uvm_pageenqueue(pg);
1487 	uvm_pageunlock(pg);
1488 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1489 	UVM_PAGE_OWN(pg, NULL);
1490 
1491 	/* deref: can not drop to zero here by defn! */
1492 	KASSERT(oanon->an_ref > 1);
1493 	oanon->an_ref--;
1494 
1495 	/*
1496 	 * note: oanon is still locked, as is the new anon.  we
1497 	 * need to check for this later when we unlock oanon; if
1498 	 * oanon != anon, we'll have to unlock anon, too.
1499 	 */
1500 
1501 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1502 }
1503 
1504 /*
1505  * uvm_fault_upper_direct: handle direct fault.
1506  */
1507 
1508 static int
1509 uvm_fault_upper_direct(
1510 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1511 	struct uvm_object *uobj, struct vm_anon *anon)
1512 {
1513 	struct vm_anon * const oanon = anon;
1514 	struct vm_page *pg;
1515 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1516 
1517 	cpu_count(CPU_COUNT_FLT_ANON, 1);
1518 	pg = anon->an_page;
1519 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1520 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1521 
1522 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1523 }
1524 
1525 /*
1526  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1527  */
1528 
1529 static int
1530 uvm_fault_upper_enter(
1531 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1532 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1533 	struct vm_anon *oanon)
1534 {
1535 	struct pmap *pmap = ufi->orig_map->pmap;
1536 	vaddr_t va = ufi->orig_rvaddr;
1537 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1538 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1539 
1540 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1541 	KASSERT(rw_write_held(amap->am_lock));
1542 	KASSERT(anon->an_lock == amap->am_lock);
1543 	KASSERT(oanon->an_lock == amap->am_lock);
1544 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1545 	KASSERT(uvm_pagegetdirty(pg) != UVM_PAGE_STATUS_CLEAN);
1546 
1547 	/*
1548 	 * now map the page in.
1549 	 */
1550 
1551 	UVMHIST_LOG(maphist,
1552 	    "  MAPPING: anon: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
1553 	    (uintptr_t)pmap, va, (uintptr_t)pg, flt->promote);
1554 	if (pmap_enter(pmap, va, VM_PAGE_TO_PHYS(pg),
1555 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1556 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1557 
1558 		/*
1559 		 * If pmap_enter() fails, it must not leave behind an existing
1560 		 * pmap entry.  In particular, a now-stale entry for a different
1561 		 * page would leave the pmap inconsistent with the vm_map.
1562 		 * This is not to imply that pmap_enter() should remove an
1563 		 * existing mapping in such a situation (since that could create
1564 		 * different problems, eg. if the existing mapping is wired),
1565 		 * but rather that the pmap should be designed such that it
1566 		 * never needs to fail when the new mapping is replacing an
1567 		 * existing mapping and the new page has no existing mappings.
1568 		 */
1569 
1570 		KASSERT(!pmap_extract(pmap, va, NULL));
1571 
1572 		/*
1573 		 * No need to undo what we did; we can simply think of
1574 		 * this as the pmap throwing away the mapping information.
1575 		 *
1576 		 * We do, however, have to go through the ReFault path,
1577 		 * as the map may change while we're asleep.
1578 		 */
1579 
1580 		uvmfault_unlockall(ufi, amap, uobj);
1581 		if (!uvm_reclaimable()) {
1582 			UVMHIST_LOG(maphist,
1583 			    "<- failed.  out of VM",0,0,0,0);
1584 			/* XXX instrumentation */
1585 			return ENOMEM;
1586 		}
1587 		/* XXX instrumentation */
1588 		uvm_wait("flt_pmfail1");
1589 		return ERESTART;
1590 	}
1591 
1592 	uvm_fault_upper_done(ufi, flt, anon, pg);
1593 
1594 	/*
1595 	 * done case 1!  finish up by unlocking everything and returning success
1596 	 */
1597 
1598 	pmap_update(pmap);
1599 	uvmfault_unlockall(ufi, amap, uobj);
1600 	return 0;
1601 }
1602 
1603 /*
1604  * uvm_fault_upper_done: queue upper center page.
1605  */
1606 
1607 static void
1608 uvm_fault_upper_done(
1609 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1610 	struct vm_anon *anon, struct vm_page *pg)
1611 {
1612 	const bool wire_paging = flt->wire_paging;
1613 
1614 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1615 
1616 	/*
1617 	 * ... update the page queues.
1618 	 */
1619 
1620 	uvm_pagelock(pg);
1621 	if (wire_paging) {
1622 		uvm_pagewire(pg);
1623 	} else {
1624 		uvm_pageactivate(pg);
1625 	}
1626 	uvm_pageunlock(pg);
1627 
1628 	if (wire_paging) {
1629 		/*
1630 		 * since the now-wired page cannot be paged out,
1631 		 * release its swap resources for others to use.
1632 		 * and since an anon with no swap cannot be clean,
1633 		 * mark it dirty now.
1634 		 */
1635 
1636 		uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
1637 		uvm_anon_dropswap(anon);
1638 	}
1639 }
1640 
1641 /*
1642  * uvm_fault_lower: handle lower fault.
1643  *
1644  *	1. check uobj
1645  *	1.1. if null, ZFOD.
1646  *	1.2. if not null, look up unnmapped neighbor pages.
1647  *	2. for center page, check if promote.
1648  *	2.1. ZFOD always needs promotion.
1649  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1650  *	3. if uobj is not ZFOD and page is not found, do i/o.
1651  *	4. dispatch either direct / promote fault.
1652  */
1653 
1654 static int
1655 uvm_fault_lower(
1656 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1657 	struct vm_page **pages)
1658 {
1659 	struct vm_amap *amap __diagused = ufi->entry->aref.ar_amap;
1660 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1661 	struct vm_page *uobjpage;
1662 	int error;
1663 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1664 
1665 	/*
1666 	 * now, if the desired page is not shadowed by the amap and we have
1667 	 * a backing object that does not have a special fault routine, then
1668 	 * we ask (with pgo_get) the object for resident pages that we care
1669 	 * about and attempt to map them in.  we do not let pgo_get block
1670 	 * (PGO_LOCKED).
1671 	 */
1672 
1673 	if (uobj == NULL) {
1674 		/* zero fill; don't care neighbor pages */
1675 		uobjpage = NULL;
1676 	} else {
1677 		uvm_fault_lower_lookup(ufi, flt, pages);
1678 		uobjpage = pages[flt->centeridx];
1679 	}
1680 
1681 	/*
1682 	 * note that at this point we are done with any front or back pages.
1683 	 * we are now going to focus on the center page (i.e. the one we've
1684 	 * faulted on).  if we have faulted on the upper (anon) layer
1685 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1686 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1687 	 * layer [i.e. case 2] and the page was both present and available,
1688 	 * then we've got a pointer to it as "uobjpage" and we've already
1689 	 * made it BUSY.
1690 	 */
1691 
1692 	/*
1693 	 * locked:
1694 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1695 	 */
1696 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1697 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1698 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1699 
1700 	/*
1701 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1702 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1703 	 * have a backing object, check and see if we are going to promote
1704 	 * the data up to an anon during the fault.
1705 	 */
1706 
1707 	if (uobj == NULL) {
1708 		uobjpage = PGO_DONTCARE;
1709 		flt->promote = true;		/* always need anon here */
1710 	} else {
1711 		KASSERT(uobjpage != PGO_DONTCARE);
1712 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1713 	}
1714 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%jd, zfill=%jd",
1715 	    flt->promote, (uobj == NULL), 0,0);
1716 
1717 	/*
1718 	 * if uobjpage is not null then we do not need to do I/O to get the
1719 	 * uobjpage.
1720 	 *
1721 	 * if uobjpage is null, then we need to unlock and ask the pager to
1722 	 * get the data for us.   once we have the data, we need to reverify
1723 	 * the state the world.   we are currently not holding any resources.
1724 	 */
1725 
1726 	if (uobjpage) {
1727 		/* update rusage counters */
1728 		curlwp->l_ru.ru_minflt++;
1729 	} else {
1730 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1731 		if (error != 0)
1732 			return error;
1733 	}
1734 
1735 	/*
1736 	 * locked:
1737 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1738 	 */
1739 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
1740 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1741 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1742 
1743 	/*
1744 	 * notes:
1745 	 *  - at this point uobjpage can not be NULL
1746 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1747 	 *  for it above)
1748 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1749 	 */
1750 
1751 	KASSERT(uobjpage != NULL);
1752 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1753 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1754 	    uvm_pagegetdirty(uobjpage) == UVM_PAGE_STATUS_CLEAN);
1755 
1756 	if (!flt->promote) {
1757 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1758 	} else {
1759 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1760 	}
1761 	return error;
1762 }
1763 
1764 /*
1765  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1766  *
1767  *	1. get on-memory pages.
1768  *	2. if failed, give up (get only center page later).
1769  *	3. if succeeded, enter h/w mapping of neighbor pages.
1770  */
1771 
1772 static void
1773 uvm_fault_lower_lookup(
1774 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1775 	struct vm_page **pages)
1776 {
1777 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1778 	int lcv, gotpages;
1779 	vaddr_t currva;
1780 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1781 
1782 	rw_enter(uobj->vmobjlock, RW_WRITER);
1783 	/* Locked: maps(read), amap(if there), uobj */
1784 
1785 	cpu_count(CPU_COUNT_FLTLGET, 1);
1786 	gotpages = flt->npages;
1787 	(void) uobj->pgops->pgo_get(uobj,
1788 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1789 	    pages, &gotpages, flt->centeridx,
1790 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1791 
1792 	KASSERT(rw_write_held(uobj->vmobjlock));
1793 
1794 	/*
1795 	 * check for pages to map, if we got any
1796 	 */
1797 
1798 	if (gotpages == 0) {
1799 		pages[flt->centeridx] = NULL;
1800 		return;
1801 	}
1802 
1803 	currva = flt->startva;
1804 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1805 		struct vm_page *curpg;
1806 
1807 		curpg = pages[lcv];
1808 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1809 			continue;
1810 		}
1811 		KASSERT(curpg->uobject == uobj);
1812 
1813 		/*
1814 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1815 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1816 		 * to it.
1817 		 */
1818 
1819 		if (lcv == flt->centeridx) {
1820 			UVMHIST_LOG(maphist, "  got uobjpage (%#jx) "
1821 			    "with locked get", (uintptr_t)curpg, 0, 0, 0);
1822 		} else {
1823 			uvm_fault_lower_neighbor(ufi, flt, currva, curpg);
1824 		}
1825 	}
1826 	pmap_update(ufi->orig_map->pmap);
1827 }
1828 
1829 /*
1830  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1831  */
1832 
1833 static void
1834 uvm_fault_lower_neighbor(
1835 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1836 	vaddr_t currva, struct vm_page *pg)
1837 {
1838 	const bool readonly = uvm_pagereadonly_p(pg) || pg->loan_count > 0;
1839 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1840 
1841 	/* locked: maps(read), amap(if there), uobj */
1842 
1843 	/*
1844 	 * calling pgo_get with PGO_LOCKED returns us pages which
1845 	 * are neither busy nor released, so we don't need to check
1846 	 * for this.  we can just directly enter the pages.
1847 	 */
1848 
1849 	uvm_pagelock(pg);
1850 	uvm_pageenqueue(pg);
1851 	uvm_pageunlock(pg);
1852 	UVMHIST_LOG(maphist,
1853 	    "  MAPPING: n obj: pm=%#jx, va=%#jx, pg=%#jx",
1854 	    (uintptr_t)ufi->orig_map->pmap, currva, (uintptr_t)pg, 0);
1855 	cpu_count(CPU_COUNT_FLTNOMAP, 1);
1856 
1857 	/*
1858 	 * Since this page isn't the page that's actually faulting,
1859 	 * ignore pmap_enter() failures; it's not critical that we
1860 	 * enter these right now.
1861 	 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've
1862 	 * held the lock the whole time we've had the handle.
1863 	 */
1864 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1865 	KASSERT((pg->flags & PG_RELEASED) == 0);
1866 	KASSERT((pg->flags & PG_WANTED) == 0);
1867 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) ||
1868 	    uvm_pagegetdirty(pg) == UVM_PAGE_STATUS_CLEAN);
1869 	pg->flags &= ~(PG_BUSY);
1870 	UVM_PAGE_OWN(pg, NULL);
1871 
1872 	KASSERT(rw_write_held(pg->uobject->vmobjlock));
1873 
1874 	const vm_prot_t mapprot =
1875 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1876 	    flt->enter_prot & MASK(ufi->entry);
1877 	const u_int mapflags =
1878 	    PMAP_CANFAIL | (flt->wire_mapping ? (mapprot | PMAP_WIRED) : 0);
1879 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1880 	    VM_PAGE_TO_PHYS(pg), mapprot, mapflags);
1881 }
1882 
1883 /*
1884  * uvm_fault_lower_io: get lower page from backing store.
1885  *
1886  *	1. unlock everything, because i/o will block.
1887  *	2. call pgo_get.
1888  *	3. if failed, recover.
1889  *	4. if succeeded, relock everything and verify things.
1890  */
1891 
1892 static int
1893 uvm_fault_lower_io(
1894 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1895 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1896 {
1897 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1898 	struct uvm_object *uobj = *ruobj;
1899 	struct vm_page *pg;
1900 	bool locked;
1901 	int gotpages;
1902 	int error;
1903 	voff_t uoff;
1904 	vm_prot_t access_type;
1905 	int advice;
1906 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1907 
1908 	/* update rusage counters */
1909 	curlwp->l_ru.ru_majflt++;
1910 
1911 	/* grab everything we need from the entry before we unlock */
1912 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1913 	access_type = flt->access_type & MASK(ufi->entry);
1914 	advice = ufi->entry->advice;
1915 
1916 	/* Locked: maps(read), amap(if there), uobj */
1917 	uvmfault_unlockall(ufi, amap, NULL);
1918 
1919 	/* Locked: uobj */
1920 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
1921 
1922 	cpu_count(CPU_COUNT_FLTGET, 1);
1923 	gotpages = 1;
1924 	pg = NULL;
1925 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1926 	    0, access_type, advice, PGO_SYNCIO);
1927 	/* locked: pg(if no error) */
1928 
1929 	/*
1930 	 * recover from I/O
1931 	 */
1932 
1933 	if (error) {
1934 		if (error == EAGAIN) {
1935 			UVMHIST_LOG(maphist,
1936 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1937 			kpause("fltagain2", false, hz/2, NULL);
1938 			return ERESTART;
1939 		}
1940 
1941 #if 0
1942 		KASSERT(error != ERESTART);
1943 #else
1944 		/* XXXUEBS don't re-fault? */
1945 		if (error == ERESTART)
1946 			error = EIO;
1947 #endif
1948 
1949 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %jd)",
1950 		    error, 0,0,0);
1951 		return error;
1952 	}
1953 
1954 	/*
1955 	 * re-verify the state of the world by first trying to relock
1956 	 * the maps.  always relock the object.
1957 	 */
1958 
1959 	locked = uvmfault_relock(ufi);
1960 	if (locked && amap)
1961 		amap_lock(amap, RW_WRITER);
1962 
1963 	/* might be changed */
1964 	uobj = pg->uobject;
1965 
1966 	rw_enter(uobj->vmobjlock, RW_WRITER);
1967 	KASSERT((pg->flags & PG_BUSY) != 0);
1968 
1969 	uvm_pagelock(pg);
1970 	uvm_pageactivate(pg);
1971 	uvm_pageunlock(pg);
1972 
1973 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1974 	/* locked(!locked): uobj, pg */
1975 
1976 	/*
1977 	 * verify that the page has not be released and re-verify
1978 	 * that amap slot is still free.   if there is a problem,
1979 	 * we unlock and clean up.
1980 	 */
1981 
1982 	if ((pg->flags & PG_RELEASED) != 0 ||
1983 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1984 	      ufi->orig_rvaddr - ufi->entry->start))) {
1985 		if (locked)
1986 			uvmfault_unlockall(ufi, amap, NULL);
1987 		locked = false;
1988 	}
1989 
1990 	/*
1991 	 * didn't get the lock?   release the page and retry.
1992 	 */
1993 
1994 	if (locked == false) {
1995 		UVMHIST_LOG(maphist,
1996 		    "  wasn't able to relock after fault: retry",
1997 		    0,0,0,0);
1998 		if (pg->flags & PG_WANTED) {
1999 			wakeup(pg);
2000 		}
2001 		if ((pg->flags & PG_RELEASED) == 0) {
2002 			pg->flags &= ~(PG_BUSY | PG_WANTED);
2003 			UVM_PAGE_OWN(pg, NULL);
2004 		} else {
2005 			cpu_count(CPU_COUNT_FLTPGRELE, 1);
2006 			uvm_pagefree(pg);
2007 		}
2008 		rw_exit(uobj->vmobjlock);
2009 		return ERESTART;
2010 	}
2011 
2012 	/*
2013 	 * we have the data in pg which is busy and
2014 	 * not released.  we are holding object lock (so the page
2015 	 * can't be released on us).
2016 	 */
2017 
2018 	/* locked: maps(read), amap(if !null), uobj, pg */
2019 
2020 	*ruobj = uobj;
2021 	*ruobjpage = pg;
2022 	return 0;
2023 }
2024 
2025 /*
2026  * uvm_fault_lower_direct: fault lower center page
2027  *
2028  *	1. adjust flt->enter_prot.
2029  *	2. if page is loaned, resolve.
2030  */
2031 
2032 int
2033 uvm_fault_lower_direct(
2034 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2035 	struct uvm_object *uobj, struct vm_page *uobjpage)
2036 {
2037 	struct vm_page *pg;
2038 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
2039 
2040 	/*
2041 	 * we are not promoting.   if the mapping is COW ensure that we
2042 	 * don't give more access than we should (e.g. when doing a read
2043 	 * fault on a COPYONWRITE mapping we want to map the COW page in
2044 	 * R/O even though the entry protection could be R/W).
2045 	 *
2046 	 * set "pg" to the page we want to map in (uobjpage, usually)
2047 	 */
2048 
2049 	cpu_count(CPU_COUNT_FLT_OBJ, 1);
2050 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
2051 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
2052 		flt->enter_prot &= ~VM_PROT_WRITE;
2053 	pg = uobjpage;		/* map in the actual object */
2054 
2055 	KASSERT(uobjpage != PGO_DONTCARE);
2056 
2057 	/*
2058 	 * we are faulting directly on the page.   be careful
2059 	 * about writing to loaned pages...
2060 	 */
2061 
2062 	if (uobjpage->loan_count) {
2063 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
2064 	}
2065 	KASSERT(pg == uobjpage);
2066 
2067 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2068 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
2069 }
2070 
2071 /*
2072  * uvm_fault_lower_direct_loan: resolve loaned page.
2073  *
2074  *	1. if not cow'ing, adjust flt->enter_prot.
2075  *	2. if cow'ing, break loan.
2076  */
2077 
2078 static int
2079 uvm_fault_lower_direct_loan(
2080 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2081 	struct uvm_object *uobj, struct vm_page **rpg,
2082 	struct vm_page **ruobjpage)
2083 {
2084 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2085 	struct vm_page *pg;
2086 	struct vm_page *uobjpage = *ruobjpage;
2087 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
2088 
2089 	if (!flt->cow_now) {
2090 		/* read fault: cap the protection at readonly */
2091 		/* cap! */
2092 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2093 	} else {
2094 		/* write fault: must break the loan here */
2095 
2096 		pg = uvm_loanbreak(uobjpage);
2097 		if (pg == NULL) {
2098 
2099 			/*
2100 			 * drop ownership of page, it can't be released
2101 			 */
2102 
2103 			if (uobjpage->flags & PG_WANTED)
2104 				wakeup(uobjpage);
2105 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2106 			UVM_PAGE_OWN(uobjpage, NULL);
2107 
2108 			uvmfault_unlockall(ufi, amap, uobj);
2109 			UVMHIST_LOG(maphist,
2110 			  "  out of RAM breaking loan, waiting",
2111 			  0,0,0,0);
2112 			cpu_count(CPU_COUNT_FLTNORAM, 1);
2113 			uvm_wait("flt_noram4");
2114 			return ERESTART;
2115 		}
2116 		*rpg = pg;
2117 		*ruobjpage = pg;
2118 	}
2119 	return 0;
2120 }
2121 
2122 /*
2123  * uvm_fault_lower_promote: promote lower page.
2124  *
2125  *	1. call uvmfault_promote.
2126  *	2. fill in data.
2127  *	3. if not ZFOD, dispose old page.
2128  */
2129 
2130 int
2131 uvm_fault_lower_promote(
2132 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2133 	struct uvm_object *uobj, struct vm_page *uobjpage)
2134 {
2135 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2136 	struct vm_anon *anon;
2137 	struct vm_page *pg;
2138 	int error;
2139 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2140 
2141 	KASSERT(amap != NULL);
2142 
2143 	/*
2144 	 * If we are going to promote the data to an anon we
2145 	 * allocate a blank anon here and plug it into our amap.
2146 	 */
2147 	error = uvmfault_promote(ufi, NULL, uobjpage,
2148 	    &anon, &flt->anon_spare);
2149 	switch (error) {
2150 	case 0:
2151 		break;
2152 	case ERESTART:
2153 		return ERESTART;
2154 	default:
2155 		return error;
2156 	}
2157 
2158 	pg = anon->an_page;
2159 
2160 	/*
2161 	 * Fill in the data.
2162 	 */
2163 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2164 
2165 	if (uobjpage != PGO_DONTCARE) {
2166 		cpu_count(CPU_COUNT_FLT_PRCOPY, 1);
2167 
2168 		/*
2169 		 * promote to shared amap?  make sure all sharing
2170 		 * procs see it
2171 		 */
2172 
2173 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2174 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2175 			/*
2176 			 * XXX: PAGE MIGHT BE WIRED!
2177 			 */
2178 		}
2179 
2180 		/*
2181 		 * dispose of uobjpage.  it can't be PG_RELEASED
2182 		 * since we still hold the object lock.
2183 		 */
2184 
2185 		if (uobjpage->flags & PG_WANTED) {
2186 			/* still have the obj lock */
2187 			wakeup(uobjpage);
2188 		}
2189 		uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2190 		UVM_PAGE_OWN(uobjpage, NULL);
2191 
2192 		UVMHIST_LOG(maphist,
2193 		    "  promote uobjpage %#jx to anon/page %#jx/%#jx",
2194 		    (uintptr_t)uobjpage, (uintptr_t)anon, (uintptr_t)pg, 0);
2195 
2196 	} else {
2197 		cpu_count(CPU_COUNT_FLT_PRZERO, 1);
2198 
2199 		/*
2200 		 * Page is zero'd and marked dirty by
2201 		 * uvmfault_promote().
2202 		 */
2203 
2204 		UVMHIST_LOG(maphist,"  zero fill anon/page %#jx/%#jx",
2205 		    (uintptr_t)anon, (uintptr_t)pg, 0, 0);
2206 	}
2207 
2208 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2209 }
2210 
2211 /*
2212  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2213  * from the lower page.
2214  */
2215 
2216 int
2217 uvm_fault_lower_enter(
2218 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2219 	struct uvm_object *uobj,
2220 	struct vm_anon *anon, struct vm_page *pg)
2221 {
2222 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2223 	const bool readonly = uvm_pagereadonly_p(pg);
2224 	int error;
2225 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2226 
2227 	/*
2228 	 * Locked:
2229 	 *
2230 	 *	maps(read), amap(if !null), uobj(if !null),
2231 	 *	anon(if !null), pg(if anon), unlock_uobj(if !null)
2232 	 *
2233 	 * Note: pg is either the uobjpage or the new page in the new anon.
2234 	 */
2235 	KASSERT(amap == NULL || rw_write_held(amap->am_lock));
2236 	KASSERT(uobj == NULL || rw_write_held(uobj->vmobjlock));
2237 	KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
2238 	KASSERT((pg->flags & PG_BUSY) != 0);
2239 
2240 	/*
2241 	 * all resources are present.   we can now map it in and free our
2242 	 * resources.
2243 	 */
2244 
2245 	UVMHIST_LOG(maphist,
2246 	    "  MAPPING: case2: pm=%#jx, va=%#jx, pg=%#jx, promote=%jd",
2247 	    (uintptr_t)ufi->orig_map->pmap, ufi->orig_rvaddr,
2248 	    (uintptr_t)pg, flt->promote);
2249 	KASSERTMSG((flt->access_type & VM_PROT_WRITE) == 0 || !readonly,
2250 	    "promote=%u cow_now=%u access_type=%x enter_prot=%x cow=%u "
2251 	    "entry=%p map=%p orig_rvaddr=%p pg=%p",
2252 	    flt->promote, flt->cow_now, flt->access_type, flt->enter_prot,
2253 	    UVM_ET_ISCOPYONWRITE(ufi->entry), ufi->entry, ufi->orig_map,
2254 	    (void *)ufi->orig_rvaddr, pg);
2255 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 || !readonly);
2256 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2257 	    VM_PAGE_TO_PHYS(pg),
2258 	    readonly ? flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2259 	    flt->access_type | PMAP_CANFAIL |
2260 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2261 
2262 		/*
2263 		 * No need to undo what we did; we can simply think of
2264 		 * this as the pmap throwing away the mapping information.
2265 		 *
2266 		 * We do, however, have to go through the ReFault path,
2267 		 * as the map may change while we're asleep.
2268 		 */
2269 
2270 		/*
2271 		 * ensure that the page is queued in the case that
2272 		 * we just promoted the page.
2273 		 */
2274 
2275 		uvm_pagelock(pg);
2276 		uvm_pageenqueue(pg);
2277 		uvm_pageunlock(pg);
2278 
2279 		if (pg->flags & PG_WANTED)
2280 			wakeup(pg);
2281 
2282 		/*
2283 		 * note that pg can't be PG_RELEASED since we did not drop
2284 		 * the object lock since the last time we checked.
2285 		 */
2286 		KASSERT((pg->flags & PG_RELEASED) == 0);
2287 
2288 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2289 		UVM_PAGE_OWN(pg, NULL);
2290 
2291 		uvmfault_unlockall(ufi, amap, uobj);
2292 		if (!uvm_reclaimable()) {
2293 			UVMHIST_LOG(maphist,
2294 			    "<- failed.  out of VM",0,0,0,0);
2295 			/* XXX instrumentation */
2296 			error = ENOMEM;
2297 			return error;
2298 		}
2299 		/* XXX instrumentation */
2300 		uvm_wait("flt_pmfail2");
2301 		return ERESTART;
2302 	}
2303 
2304 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2305 
2306 	/*
2307 	 * note that pg can't be PG_RELEASED since we did not drop the object
2308 	 * lock since the last time we checked.
2309 	 */
2310 	KASSERT((pg->flags & PG_RELEASED) == 0);
2311 	if (pg->flags & PG_WANTED)
2312 		wakeup(pg);
2313 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2314 	UVM_PAGE_OWN(pg, NULL);
2315 
2316 	pmap_update(ufi->orig_map->pmap);
2317 	uvmfault_unlockall(ufi, amap, uobj);
2318 
2319 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2320 	return 0;
2321 }
2322 
2323 /*
2324  * uvm_fault_lower_done: queue lower center page.
2325  */
2326 
2327 void
2328 uvm_fault_lower_done(
2329 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2330 	struct uvm_object *uobj, struct vm_page *pg)
2331 {
2332 	bool dropswap = false;
2333 
2334 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2335 
2336 	uvm_pagelock(pg);
2337 	if (flt->wire_paging) {
2338 		uvm_pagewire(pg);
2339 		if (pg->flags & PG_AOBJ) {
2340 
2341 			/*
2342 			 * since the now-wired page cannot be paged out,
2343 			 * release its swap resources for others to use.
2344 			 * since an aobj page with no swap cannot be clean,
2345 			 * mark it dirty now.
2346 			 */
2347 
2348 			KASSERT(uobj != NULL);
2349 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
2350 			dropswap = true;
2351 		}
2352 	} else {
2353 		uvm_pageactivate(pg);
2354 	}
2355 	uvm_pageunlock(pg);
2356 
2357 	if (dropswap) {
2358 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2359 	}
2360 }
2361 
2362 
2363 /*
2364  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2365  *
2366  * => map may be read-locked by caller, but MUST NOT be write-locked.
2367  * => if map is read-locked, any operations which may cause map to
2368  *	be write-locked in uvm_fault() must be taken care of by
2369  *	the caller.  See uvm_map_pageable().
2370  */
2371 
2372 int
2373 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2374     vm_prot_t access_type, int maxprot)
2375 {
2376 	vaddr_t va;
2377 	int error;
2378 
2379 	/*
2380 	 * now fault it in a page at a time.   if the fault fails then we have
2381 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2382 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2383 	 */
2384 
2385 	/*
2386 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2387 	 * wiring the last page in the address space, though.
2388 	 */
2389 	if (start > end) {
2390 		return EFAULT;
2391 	}
2392 
2393 	for (va = start; va < end; va += PAGE_SIZE) {
2394 		error = uvm_fault_internal(map, va, access_type,
2395 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2396 		if (error) {
2397 			if (va != start) {
2398 				uvm_fault_unwire(map, start, va);
2399 			}
2400 			return error;
2401 		}
2402 	}
2403 	return 0;
2404 }
2405 
2406 /*
2407  * uvm_fault_unwire(): unwire range of virtual space.
2408  */
2409 
2410 void
2411 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2412 {
2413 	vm_map_lock_read(map);
2414 	uvm_fault_unwire_locked(map, start, end);
2415 	vm_map_unlock_read(map);
2416 }
2417 
2418 /*
2419  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2420  *
2421  * => map must be at least read-locked.
2422  */
2423 
2424 void
2425 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2426 {
2427 	struct vm_map_entry *entry, *oentry;
2428 	pmap_t pmap = vm_map_pmap(map);
2429 	vaddr_t va;
2430 	paddr_t pa;
2431 	struct vm_page *pg;
2432 
2433 	/*
2434 	 * we assume that the area we are unwiring has actually been wired
2435 	 * in the first place.   this means that we should be able to extract
2436 	 * the PAs from the pmap.   we also lock out the page daemon so that
2437 	 * we can call uvm_pageunwire.
2438 	 */
2439 
2440 	/*
2441 	 * find the beginning map entry for the region.
2442 	 */
2443 
2444 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2445 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2446 		panic("uvm_fault_unwire_locked: address not in map");
2447 
2448 	oentry = NULL;
2449 	for (va = start; va < end; va += PAGE_SIZE) {
2450 
2451 		/*
2452 		 * find the map entry for the current address.
2453 		 */
2454 
2455 		KASSERT(va >= entry->start);
2456 		while (va >= entry->end) {
2457 			KASSERT(entry->next != &map->header &&
2458 				entry->next->start <= entry->end);
2459 			entry = entry->next;
2460 		}
2461 
2462 		/*
2463 		 * lock it.
2464 		 */
2465 
2466 		if (entry != oentry) {
2467 			if (oentry != NULL) {
2468 				uvm_map_unlock_entry(oentry);
2469 			}
2470 			uvm_map_lock_entry(entry, RW_WRITER);
2471 			oentry = entry;
2472 		}
2473 
2474 		/*
2475 		 * if the entry is no longer wired, tell the pmap.
2476 		 */
2477 
2478 		if (!pmap_extract(pmap, va, &pa))
2479 			continue;
2480 
2481 		if (VM_MAPENT_ISWIRED(entry) == 0)
2482 			pmap_unwire(pmap, va);
2483 
2484 		pg = PHYS_TO_VM_PAGE(pa);
2485 		if (pg) {
2486 			uvm_pagelock(pg);
2487 			uvm_pageunwire(pg);
2488 			uvm_pageunlock(pg);
2489 		}
2490 	}
2491 
2492 	if (oentry != NULL) {
2493 		uvm_map_unlock_entry(entry);
2494 	}
2495 }
2496