xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 288bb96063654ec504ca8732afc683d3ebc514b5)
1 /*	$NetBSD: uvm_fault.c,v 1.189 2011/07/05 13:47:24 yamt Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26  *
27  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
28  */
29 
30 /*
31  * uvm_fault.c: fault handler
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.189 2011/07/05 13:47:24 yamt Exp $");
36 
37 #include "opt_uvmhist.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/mman.h>
43 
44 #include <uvm/uvm.h>
45 
46 /*
47  *
48  * a word on page faults:
49  *
50  * types of page faults we handle:
51  *
52  * CASE 1: upper layer faults                   CASE 2: lower layer faults
53  *
54  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
55  *    read/write1     write>1                  read/write   +-cow_write/zero
56  *         |             |                         |        |
57  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
58  * amap |  V  |       |  ---------> new |          |        | |  ^  |
59  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
60  *                                                 |        |    |
61  *      +-----+       +-----+                   +--|--+     | +--|--+
62  * uobj | d/c |       | d/c |                   |  V  |     +----+  |
63  *      +-----+       +-----+                   +-----+       +-----+
64  *
65  * d/c = don't care
66  *
67  *   case [0]: layerless fault
68  *	no amap or uobj is present.   this is an error.
69  *
70  *   case [1]: upper layer fault [anon active]
71  *     1A: [read] or [write with anon->an_ref == 1]
72  *		I/O takes place in upper level anon and uobj is not touched.
73  *     1B: [write with anon->an_ref > 1]
74  *		new anon is alloc'd and data is copied off ["COW"]
75  *
76  *   case [2]: lower layer fault [uobj]
77  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
78  *		I/O takes place directly in object.
79  *     2B: [write to copy_on_write] or [read on NULL uobj]
80  *		data is "promoted" from uobj to a new anon.
81  *		if uobj is null, then we zero fill.
82  *
83  * we follow the standard UVM locking protocol ordering:
84  *
85  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
86  * we hold a PG_BUSY page if we unlock for I/O
87  *
88  *
89  * the code is structured as follows:
90  *
91  *     - init the "IN" params in the ufi structure
92  *   ReFault: (ERESTART returned to the loop in uvm_fault_internal)
93  *     - do lookups [locks maps], check protection, handle needs_copy
94  *     - check for case 0 fault (error)
95  *     - establish "range" of fault
96  *     - if we have an amap lock it and extract the anons
97  *     - if sequential advice deactivate pages behind us
98  *     - at the same time check pmap for unmapped areas and anon for pages
99  *	 that we could map in (and do map it if found)
100  *     - check object for resident pages that we could map in
101  *     - if (case 2) goto Case2
102  *     - >>> handle case 1
103  *           - ensure source anon is resident in RAM
104  *           - if case 1B alloc new anon and copy from source
105  *           - map the correct page in
106  *   Case2:
107  *     - >>> handle case 2
108  *           - ensure source page is resident (if uobj)
109  *           - if case 2B alloc new anon and copy from source (could be zero
110  *		fill if uobj == NULL)
111  *           - map the correct page in
112  *     - done!
113  *
114  * note on paging:
115  *   if we have to do I/O we place a PG_BUSY page in the correct object,
116  * unlock everything, and do the I/O.   when I/O is done we must reverify
117  * the state of the world before assuming that our data structures are
118  * valid.   [because mappings could change while the map is unlocked]
119  *
120  *  alternative 1: unbusy the page in question and restart the page fault
121  *    from the top (ReFault).   this is easy but does not take advantage
122  *    of the information that we already have from our previous lookup,
123  *    although it is possible that the "hints" in the vm_map will help here.
124  *
125  * alternative 2: the system already keeps track of a "version" number of
126  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
127  *    mapping) you bump the version number up by one...]   so, we can save
128  *    the version number of the map before we release the lock and start I/O.
129  *    then when I/O is done we can relock and check the version numbers
130  *    to see if anything changed.    this might save us some over 1 because
131  *    we don't have to unbusy the page and may be less compares(?).
132  *
133  * alternative 3: put in backpointers or a way to "hold" part of a map
134  *    in place while I/O is in progress.   this could be complex to
135  *    implement (especially with structures like amap that can be referenced
136  *    by multiple map entries, and figuring out what should wait could be
137  *    complex as well...).
138  *
139  * we use alternative 2.  given that we are multi-threaded now we may want
140  * to reconsider the choice.
141  */
142 
143 /*
144  * local data structures
145  */
146 
147 struct uvm_advice {
148 	int advice;
149 	int nback;
150 	int nforw;
151 };
152 
153 /*
154  * page range array:
155  * note: index in array must match "advice" value
156  * XXX: borrowed numbers from freebsd.   do they work well for us?
157  */
158 
159 static const struct uvm_advice uvmadvice[] = {
160 	{ UVM_ADV_NORMAL, 3, 4 },
161 	{ UVM_ADV_RANDOM, 0, 0 },
162 	{ UVM_ADV_SEQUENTIAL, 8, 7},
163 };
164 
165 #define UVM_MAXRANGE 16	/* must be MAX() of nback+nforw+1 */
166 
167 /*
168  * private prototypes
169  */
170 
171 /*
172  * inline functions
173  */
174 
175 /*
176  * uvmfault_anonflush: try and deactivate pages in specified anons
177  *
178  * => does not have to deactivate page if it is busy
179  */
180 
181 static inline void
182 uvmfault_anonflush(struct vm_anon **anons, int n)
183 {
184 	int lcv;
185 	struct vm_page *pg;
186 
187 	for (lcv = 0; lcv < n; lcv++) {
188 		if (anons[lcv] == NULL)
189 			continue;
190 		KASSERT(mutex_owned(anons[lcv]->an_lock));
191 		pg = anons[lcv]->an_page;
192 		if (pg && (pg->flags & PG_BUSY) == 0) {
193 			mutex_enter(&uvm_pageqlock);
194 			if (pg->wire_count == 0) {
195 				uvm_pagedeactivate(pg);
196 			}
197 			mutex_exit(&uvm_pageqlock);
198 		}
199 	}
200 }
201 
202 /*
203  * normal functions
204  */
205 
206 /*
207  * uvmfault_amapcopy: clear "needs_copy" in a map.
208  *
209  * => called with VM data structures unlocked (usually, see below)
210  * => we get a write lock on the maps and clear needs_copy for a VA
211  * => if we are out of RAM we sleep (waiting for more)
212  */
213 
214 static void
215 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
216 {
217 	for (;;) {
218 
219 		/*
220 		 * no mapping?  give up.
221 		 */
222 
223 		if (uvmfault_lookup(ufi, true) == false)
224 			return;
225 
226 		/*
227 		 * copy if needed.
228 		 */
229 
230 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
231 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
232 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
233 
234 		/*
235 		 * didn't work?  must be out of RAM.   unlock and sleep.
236 		 */
237 
238 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
239 			uvmfault_unlockmaps(ufi, true);
240 			uvm_wait("fltamapcopy");
241 			continue;
242 		}
243 
244 		/*
245 		 * got it!   unlock and return.
246 		 */
247 
248 		uvmfault_unlockmaps(ufi, true);
249 		return;
250 	}
251 	/*NOTREACHED*/
252 }
253 
254 /*
255  * uvmfault_anonget: get data in an anon into a non-busy, non-released
256  * page in that anon.
257  *
258  * => Map, amap and thus anon should be locked by caller.
259  * => If we fail, we unlock everything and error is returned.
260  * => If we are successful, return with everything still locked.
261  * => We do not move the page on the queues [gets moved later].  If we
262  *    allocate a new page [we_own], it gets put on the queues.  Either way,
263  *    the result is that the page is on the queues at return time
264  * => For pages which are on loan from a uvm_object (and thus are not owned
265  *    by the anon): if successful, return with the owning object locked.
266  *    The caller must unlock this object when it unlocks everything else.
267  */
268 
269 int
270 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
271     struct vm_anon *anon)
272 {
273 	struct vm_page *pg;
274 	int error;
275 
276 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
277 	KASSERT(mutex_owned(anon->an_lock));
278 	KASSERT(anon->an_lock == amap->am_lock);
279 
280 	/* Increment the counters.*/
281 	uvmexp.fltanget++;
282 	if (anon->an_page) {
283 		curlwp->l_ru.ru_minflt++;
284 	} else {
285 		curlwp->l_ru.ru_majflt++;
286 	}
287 	error = 0;
288 
289 	/*
290 	 * Loop until we get the anon data, or fail.
291 	 */
292 
293 	for (;;) {
294 		bool we_own, locked;
295 		/*
296 		 * Note: 'we_own' will become true if we set PG_BUSY on a page.
297 		 */
298 		we_own = false;
299 		pg = anon->an_page;
300 
301 		/*
302 		 * If there is a resident page and it is loaned, then anon
303 		 * may not own it.  Call out to uvm_anon_lockloanpg() to
304 		 * identify and lock the real owner of the page.
305 		 */
306 
307 		if (pg && pg->loan_count)
308 			pg = uvm_anon_lockloanpg(anon);
309 
310 		/*
311 		 * Is page resident?  Make sure it is not busy/released.
312 		 */
313 
314 		if (pg) {
315 
316 			/*
317 			 * at this point, if the page has a uobject [meaning
318 			 * we have it on loan], then that uobject is locked
319 			 * by us!   if the page is busy, we drop all the
320 			 * locks (including uobject) and try again.
321 			 */
322 
323 			if ((pg->flags & PG_BUSY) == 0) {
324 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
325 				return 0;
326 			}
327 			pg->flags |= PG_WANTED;
328 			uvmexp.fltpgwait++;
329 
330 			/*
331 			 * The last unlock must be an atomic unlock and wait
332 			 * on the owner of page.
333 			 */
334 
335 			if (pg->uobject) {
336 				/* Owner of page is UVM object. */
337 				uvmfault_unlockall(ufi, amap, NULL);
338 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
339 				    0,0,0);
340 				UVM_UNLOCK_AND_WAIT(pg,
341 				    pg->uobject->vmobjlock,
342 				    false, "anonget1", 0);
343 			} else {
344 				/* Owner of page is anon. */
345 				uvmfault_unlockall(ufi, NULL, NULL);
346 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
347 				    0,0,0);
348 				UVM_UNLOCK_AND_WAIT(pg, anon->an_lock,
349 				    false, "anonget2", 0);
350 			}
351 		} else {
352 #if defined(VMSWAP)
353 			/*
354 			 * No page, therefore allocate one.
355 			 */
356 
357 			pg = uvm_pagealloc(NULL,
358 			    ufi != NULL ? ufi->orig_rvaddr : 0,
359 			    anon, ufi != NULL ? UVM_FLAG_COLORMATCH : 0);
360 			if (pg == NULL) {
361 				/* Out of memory.  Wait a little. */
362 				uvmfault_unlockall(ufi, amap, NULL);
363 				uvmexp.fltnoram++;
364 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
365 				    0,0,0);
366 				if (!uvm_reclaimable()) {
367 					return ENOMEM;
368 				}
369 				uvm_wait("flt_noram1");
370 			} else {
371 				/* PG_BUSY bit is set. */
372 				we_own = true;
373 				uvmfault_unlockall(ufi, amap, NULL);
374 
375 				/*
376 				 * Pass a PG_BUSY+PG_FAKE+PG_CLEAN page into
377 				 * the uvm_swap_get() function with all data
378 				 * structures unlocked.  Note that it is OK
379 				 * to read an_swslot here, because we hold
380 				 * PG_BUSY on the page.
381 				 */
382 				uvmexp.pageins++;
383 				error = uvm_swap_get(pg, anon->an_swslot,
384 				    PGO_SYNCIO);
385 
386 				/*
387 				 * We clean up after the I/O below in the
388 				 * 'we_own' case.
389 				 */
390 			}
391 #else
392 			panic("%s: no page", __func__);
393 #endif /* defined(VMSWAP) */
394 		}
395 
396 		/*
397 		 * Re-lock the map and anon.
398 		 */
399 
400 		locked = uvmfault_relock(ufi);
401 		if (locked || we_own) {
402 			mutex_enter(anon->an_lock);
403 		}
404 
405 		/*
406 		 * If we own the page (i.e. we set PG_BUSY), then we need
407 		 * to clean up after the I/O.  There are three cases to
408 		 * consider:
409 		 *
410 		 * 1) Page was released during I/O: free anon and ReFault.
411 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
412 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
413 		 *    case (i.e. drop anon lock if not locked).
414 		 */
415 
416 		if (we_own) {
417 #if defined(VMSWAP)
418 			if (pg->flags & PG_WANTED) {
419 				wakeup(pg);
420 			}
421 			if (error) {
422 
423 				/*
424 				 * Remove the swap slot from the anon and
425 				 * mark the anon as having no real slot.
426 				 * Do not free the swap slot, thus preventing
427 				 * it from being used again.
428 				 */
429 
430 				if (anon->an_swslot > 0) {
431 					uvm_swap_markbad(anon->an_swslot, 1);
432 				}
433 				anon->an_swslot = SWSLOT_BAD;
434 
435 				if ((pg->flags & PG_RELEASED) != 0) {
436 					goto released;
437 				}
438 
439 				/*
440 				 * Note: page was never !PG_BUSY, so it
441 				 * cannot be mapped and thus no need to
442 				 * pmap_page_protect() it.
443 				 */
444 
445 				mutex_enter(&uvm_pageqlock);
446 				uvm_pagefree(pg);
447 				mutex_exit(&uvm_pageqlock);
448 
449 				if (locked) {
450 					uvmfault_unlockall(ufi, NULL, NULL);
451 				}
452 				mutex_exit(anon->an_lock);
453 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
454 				return error;
455 			}
456 
457 			if ((pg->flags & PG_RELEASED) != 0) {
458 released:
459 				KASSERT(anon->an_ref == 0);
460 
461 				/*
462 				 * Released while we had unlocked amap.
463 				 */
464 
465 				if (locked) {
466 					uvmfault_unlockall(ufi, NULL, NULL);
467 				}
468 				uvm_anon_release(anon);
469 
470 				if (error) {
471 					UVMHIST_LOG(maphist,
472 					    "<- ERROR/RELEASED", 0,0,0,0);
473 					return error;
474 				}
475 
476 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
477 				return ERESTART;
478 			}
479 
480 			/*
481 			 * We have successfully read the page, activate it.
482 			 */
483 
484 			mutex_enter(&uvm_pageqlock);
485 			uvm_pageactivate(pg);
486 			mutex_exit(&uvm_pageqlock);
487 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
488 			UVM_PAGE_OWN(pg, NULL);
489 #else
490 			panic("%s: we_own", __func__);
491 #endif /* defined(VMSWAP) */
492 		}
493 
494 		/*
495 		 * We were not able to re-lock the map - restart the fault.
496 		 */
497 
498 		if (!locked) {
499 			if (we_own) {
500 				mutex_exit(anon->an_lock);
501 			}
502 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
503 			return ERESTART;
504 		}
505 
506 		/*
507 		 * Verify that no one has touched the amap and moved
508 		 * the anon on us.
509 		 */
510 
511 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
512 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
513 
514 			uvmfault_unlockall(ufi, amap, NULL);
515 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
516 			return ERESTART;
517 		}
518 
519 		/*
520 		 * Retry..
521 		 */
522 
523 		uvmexp.fltanretry++;
524 		continue;
525 	}
526 	/*NOTREACHED*/
527 }
528 
529 /*
530  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
531  *
532  *	1. allocate an anon and a page.
533  *	2. fill its contents.
534  *	3. put it into amap.
535  *
536  * => if we fail (result != 0) we unlock everything.
537  * => on success, return a new locked anon via 'nanon'.
538  *    (*nanon)->an_page will be a resident, locked, dirty page.
539  * => it's caller's responsibility to put the promoted nanon->an_page to the
540  *    page queue.
541  */
542 
543 static int
544 uvmfault_promote(struct uvm_faultinfo *ufi,
545     struct vm_anon *oanon,
546     struct vm_page *uobjpage,
547     struct vm_anon **nanon, /* OUT: allocated anon */
548     struct vm_anon **spare)
549 {
550 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
551 	struct uvm_object *uobj;
552 	struct vm_anon *anon;
553 	struct vm_page *pg;
554 	struct vm_page *opg;
555 	int error;
556 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
557 
558 	if (oanon) {
559 		/* anon COW */
560 		opg = oanon->an_page;
561 		KASSERT(opg != NULL);
562 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
563 	} else if (uobjpage != PGO_DONTCARE) {
564 		/* object-backed COW */
565 		opg = uobjpage;
566 	} else {
567 		/* ZFOD */
568 		opg = NULL;
569 	}
570 	if (opg != NULL) {
571 		uobj = opg->uobject;
572 	} else {
573 		uobj = NULL;
574 	}
575 
576 	KASSERT(amap != NULL);
577 	KASSERT(uobjpage != NULL);
578 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
579 	KASSERT(mutex_owned(amap->am_lock));
580 	KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock);
581 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
582 
583 	if (*spare != NULL) {
584 		anon = *spare;
585 		*spare = NULL;
586 	} else if (ufi->map != kernel_map) {
587 		anon = uvm_analloc();
588 	} else {
589 		UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0);
590 
591 		/*
592 		 * we can't allocate anons with kernel_map locked.
593 		 */
594 
595 		uvm_page_unbusy(&uobjpage, 1);
596 		uvmfault_unlockall(ufi, amap, uobj);
597 
598 		*spare = uvm_analloc();
599 		if (*spare == NULL) {
600 			goto nomem;
601 		}
602 		KASSERT((*spare)->an_lock == NULL);
603 		error = ERESTART;
604 		goto done;
605 	}
606 	if (anon) {
607 
608 		/*
609 		 * The new anon is locked.
610 		 *
611 		 * if opg == NULL, we want a zero'd, dirty page,
612 		 * so have uvm_pagealloc() do that for us.
613 		 */
614 
615 		KASSERT(anon->an_lock == NULL);
616 		anon->an_lock = amap->am_lock;
617 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
618 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
619 		if (pg == NULL) {
620 			anon->an_lock = NULL;
621 		}
622 	} else {
623 		pg = NULL;
624 	}
625 
626 	/*
627 	 * out of memory resources?
628 	 */
629 
630 	if (pg == NULL) {
631 		/* save anon for the next try. */
632 		if (anon != NULL) {
633 			*spare = anon;
634 		}
635 
636 		/* unlock and fail ... */
637 		uvm_page_unbusy(&uobjpage, 1);
638 		uvmfault_unlockall(ufi, amap, uobj);
639 nomem:
640 		if (!uvm_reclaimable()) {
641 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
642 			uvmexp.fltnoanon++;
643 			error = ENOMEM;
644 			goto done;
645 		}
646 
647 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
648 		uvmexp.fltnoram++;
649 		uvm_wait("flt_noram5");
650 		error = ERESTART;
651 		goto done;
652 	}
653 
654 	/* copy page [pg now dirty] */
655 	if (opg) {
656 		uvm_pagecopy(opg, pg);
657 	}
658 
659 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
660 	    oanon != NULL);
661 
662 	*nanon = anon;
663 	error = 0;
664 done:
665 	return error;
666 }
667 
668 
669 /*
670  *   F A U L T   -   m a i n   e n t r y   p o i n t
671  */
672 
673 /*
674  * uvm_fault: page fault handler
675  *
676  * => called from MD code to resolve a page fault
677  * => VM data structures usually should be unlocked.   however, it is
678  *	possible to call here with the main map locked if the caller
679  *	gets a write lock, sets it recusive, and then calls us (c.f.
680  *	uvm_map_pageable).   this should be avoided because it keeps
681  *	the map locked off during I/O.
682  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
683  */
684 
685 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
686 			 ~VM_PROT_WRITE : VM_PROT_ALL)
687 
688 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
689 #define UVM_FAULT_WIRE		(1 << 0)
690 #define UVM_FAULT_MAXPROT	(1 << 1)
691 
692 struct uvm_faultctx {
693 	vm_prot_t access_type;
694 	vm_prot_t enter_prot;
695 	vaddr_t startva;
696 	int npages;
697 	int centeridx;
698 	struct vm_anon *anon_spare;
699 	bool wire_mapping;
700 	bool narrow;
701 	bool wire_paging;
702 	bool cow_now;
703 	bool promote;
704 };
705 
706 static inline int	uvm_fault_check(
707 			    struct uvm_faultinfo *, struct uvm_faultctx *,
708 			    struct vm_anon ***, bool);
709 
710 static int		uvm_fault_upper(
711 			    struct uvm_faultinfo *, struct uvm_faultctx *,
712 			    struct vm_anon **);
713 static inline int	uvm_fault_upper_lookup(
714 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
715 			    struct vm_anon **, struct vm_page **);
716 static inline void	uvm_fault_upper_neighbor(
717 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
718 			    vaddr_t, struct vm_page *, bool);
719 static inline int	uvm_fault_upper_loan(
720 			    struct uvm_faultinfo *, struct uvm_faultctx *,
721 			    struct vm_anon *, struct uvm_object **);
722 static inline int	uvm_fault_upper_promote(
723 			    struct uvm_faultinfo *, struct uvm_faultctx *,
724 			    struct uvm_object *, struct vm_anon *);
725 static inline int	uvm_fault_upper_direct(
726 			    struct uvm_faultinfo *, struct uvm_faultctx *,
727 			    struct uvm_object *, struct vm_anon *);
728 static int		uvm_fault_upper_enter(
729 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
730 			    struct uvm_object *, struct vm_anon *,
731 			    struct vm_page *, struct vm_anon *);
732 static inline void	uvm_fault_upper_done(
733 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
734 			    struct vm_anon *, struct vm_page *);
735 
736 static int		uvm_fault_lower(
737 			    struct uvm_faultinfo *, struct uvm_faultctx *,
738 			    struct vm_page **);
739 static inline void	uvm_fault_lower_lookup(
740 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
741 			    struct vm_page **);
742 static inline void	uvm_fault_lower_neighbor(
743 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
744 			    vaddr_t, struct vm_page *, bool);
745 static inline int	uvm_fault_lower_io(
746 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
747 			    struct uvm_object **, struct vm_page **);
748 static inline int	uvm_fault_lower_direct(
749 			    struct uvm_faultinfo *, struct uvm_faultctx *,
750 			    struct uvm_object *, struct vm_page *);
751 static inline int	uvm_fault_lower_direct_loan(
752 			    struct uvm_faultinfo *, struct uvm_faultctx *,
753 			    struct uvm_object *, struct vm_page **,
754 			    struct vm_page **);
755 static inline int	uvm_fault_lower_promote(
756 			    struct uvm_faultinfo *, struct uvm_faultctx *,
757 			    struct uvm_object *, struct vm_page *);
758 static int		uvm_fault_lower_enter(
759 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
760 			    struct uvm_object *,
761 			    struct vm_anon *, struct vm_page *);
762 static inline void	uvm_fault_lower_done(
763 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
764 			    struct uvm_object *, struct vm_page *);
765 
766 int
767 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
768     vm_prot_t access_type, int fault_flag)
769 {
770 	struct uvm_faultinfo ufi;
771 	struct uvm_faultctx flt = {
772 		.access_type = access_type,
773 
774 		/* don't look for neighborhood * pages on "wire" fault */
775 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
776 
777 		/* "wire" fault causes wiring of both mapping and paging */
778 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
779 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
780 	};
781 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
782 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
783 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
784 	int error;
785 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
786 
787 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)",
788 	      orig_map, vaddr, access_type, fault_flag);
789 
790 	curcpu()->ci_data.cpu_nfault++;
791 
792 	/*
793 	 * init the IN parameters in the ufi
794 	 */
795 
796 	ufi.orig_map = orig_map;
797 	ufi.orig_rvaddr = trunc_page(vaddr);
798 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
799 
800 	error = ERESTART;
801 	while (error == ERESTART) { /* ReFault: */
802 		anons = anons_store;
803 		pages = pages_store;
804 
805 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
806 		if (error != 0)
807 			continue;
808 
809 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
810 		if (error != 0)
811 			continue;
812 
813 		if (pages[flt.centeridx] == PGO_DONTCARE)
814 			error = uvm_fault_upper(&ufi, &flt, anons);
815 		else {
816 			struct uvm_object * const uobj =
817 			    ufi.entry->object.uvm_obj;
818 
819 			if (uobj && uobj->pgops->pgo_fault != NULL) {
820 				/*
821 				 * invoke "special" fault routine.
822 				 */
823 				mutex_enter(uobj->vmobjlock);
824 				/* locked: maps(read), amap(if there), uobj */
825 				error = uobj->pgops->pgo_fault(&ufi,
826 				    flt.startva, pages, flt.npages,
827 				    flt.centeridx, flt.access_type,
828 				    PGO_LOCKED|PGO_SYNCIO);
829 
830 				/*
831 				 * locked: nothing, pgo_fault has unlocked
832 				 * everything
833 				 */
834 
835 				/*
836 				 * object fault routine responsible for
837 				 * pmap_update().
838 				 */
839 			} else {
840 				error = uvm_fault_lower(&ufi, &flt, pages);
841 			}
842 		}
843 	}
844 
845 	if (flt.anon_spare != NULL) {
846 		flt.anon_spare->an_ref--;
847 		KASSERT(flt.anon_spare->an_ref == 0);
848 		KASSERT(flt.anon_spare->an_lock == NULL);
849 		uvm_anfree(flt.anon_spare);
850 	}
851 	return error;
852 }
853 
854 /*
855  * uvm_fault_check: check prot, handle needs-copy, etc.
856  *
857  *	1. lookup entry.
858  *	2. check protection.
859  *	3. adjust fault condition (mainly for simulated fault).
860  *	4. handle needs-copy (lazy amap copy).
861  *	5. establish range of interest for neighbor fault (aka pre-fault).
862  *	6. look up anons (if amap exists).
863  *	7. flush pages (if MADV_SEQUENTIAL)
864  *
865  * => called with nothing locked.
866  * => if we fail (result != 0) we unlock everything.
867  * => initialize/adjust many members of flt.
868  */
869 
870 static int
871 uvm_fault_check(
872 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
873 	struct vm_anon ***ranons, bool maxprot)
874 {
875 	struct vm_amap *amap;
876 	struct uvm_object *uobj;
877 	vm_prot_t check_prot;
878 	int nback, nforw;
879 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
880 
881 	/*
882 	 * lookup and lock the maps
883 	 */
884 
885 	if (uvmfault_lookup(ufi, false) == false) {
886 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", ufi->orig_rvaddr,
887 		    0,0,0);
888 		return EFAULT;
889 	}
890 	/* locked: maps(read) */
891 
892 #ifdef DIAGNOSTIC
893 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
894 		printf("Page fault on non-pageable map:\n");
895 		printf("ufi->map = %p\n", ufi->map);
896 		printf("ufi->orig_map = %p\n", ufi->orig_map);
897 		printf("ufi->orig_rvaddr = 0x%lx\n", (u_long) ufi->orig_rvaddr);
898 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
899 	}
900 #endif
901 
902 	/*
903 	 * check protection
904 	 */
905 
906 	check_prot = maxprot ?
907 	    ufi->entry->max_protection : ufi->entry->protection;
908 	if ((check_prot & flt->access_type) != flt->access_type) {
909 		UVMHIST_LOG(maphist,
910 		    "<- protection failure (prot=0x%x, access=0x%x)",
911 		    ufi->entry->protection, flt->access_type, 0, 0);
912 		uvmfault_unlockmaps(ufi, false);
913 		return EACCES;
914 	}
915 
916 	/*
917 	 * "enter_prot" is the protection we want to enter the page in at.
918 	 * for certain pages (e.g. copy-on-write pages) this protection can
919 	 * be more strict than ufi->entry->protection.  "wired" means either
920 	 * the entry is wired or we are fault-wiring the pg.
921 	 */
922 
923 	flt->enter_prot = ufi->entry->protection;
924 	if (VM_MAPENT_ISWIRED(ufi->entry))
925 		flt->wire_mapping = true;
926 
927 	if (flt->wire_mapping) {
928 		flt->access_type = flt->enter_prot; /* full access for wired */
929 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
930 	} else {
931 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
932 	}
933 
934 	flt->promote = false;
935 
936 	/*
937 	 * handle "needs_copy" case.   if we need to copy the amap we will
938 	 * have to drop our readlock and relock it with a write lock.  (we
939 	 * need a write lock to change anything in a map entry [e.g.
940 	 * needs_copy]).
941 	 */
942 
943 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
944 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
945 			KASSERT(!maxprot);
946 			/* need to clear */
947 			UVMHIST_LOG(maphist,
948 			    "  need to clear needs_copy and refault",0,0,0,0);
949 			uvmfault_unlockmaps(ufi, false);
950 			uvmfault_amapcopy(ufi);
951 			uvmexp.fltamcopy++;
952 			return ERESTART;
953 
954 		} else {
955 
956 			/*
957 			 * ensure that we pmap_enter page R/O since
958 			 * needs_copy is still true
959 			 */
960 
961 			flt->enter_prot &= ~VM_PROT_WRITE;
962 		}
963 	}
964 
965 	/*
966 	 * identify the players
967 	 */
968 
969 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
970 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
971 
972 	/*
973 	 * check for a case 0 fault.  if nothing backing the entry then
974 	 * error now.
975 	 */
976 
977 	if (amap == NULL && uobj == NULL) {
978 		uvmfault_unlockmaps(ufi, false);
979 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
980 		return EFAULT;
981 	}
982 
983 	/*
984 	 * establish range of interest based on advice from mapper
985 	 * and then clip to fit map entry.   note that we only want
986 	 * to do this the first time through the fault.   if we
987 	 * ReFault we will disable this by setting "narrow" to true.
988 	 */
989 
990 	if (flt->narrow == false) {
991 
992 		/* wide fault (!narrow) */
993 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
994 			 ufi->entry->advice);
995 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
996 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
997 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
998 		/*
999 		 * note: "-1" because we don't want to count the
1000 		 * faulting page as forw
1001 		 */
1002 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1003 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1004 			     PAGE_SHIFT) - 1);
1005 		flt->npages = nback + nforw + 1;
1006 		flt->centeridx = nback;
1007 
1008 		flt->narrow = true;	/* ensure only once per-fault */
1009 
1010 	} else {
1011 
1012 		/* narrow fault! */
1013 		nback = nforw = 0;
1014 		flt->startva = ufi->orig_rvaddr;
1015 		flt->npages = 1;
1016 		flt->centeridx = 0;
1017 
1018 	}
1019 	/* offset from entry's start to pgs' start */
1020 	const voff_t eoff = flt->startva - ufi->entry->start;
1021 
1022 	/* locked: maps(read) */
1023 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
1024 		    flt->narrow, nback, nforw, flt->startva);
1025 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi->entry,
1026 		    amap, uobj, 0);
1027 
1028 	/*
1029 	 * if we've got an amap, lock it and extract current anons.
1030 	 */
1031 
1032 	if (amap) {
1033 		amap_lock(amap);
1034 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1035 	} else {
1036 		*ranons = NULL;	/* to be safe */
1037 	}
1038 
1039 	/* locked: maps(read), amap(if there) */
1040 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1041 
1042 	/*
1043 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1044 	 * now and then forget about them (for the rest of the fault).
1045 	 */
1046 
1047 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1048 
1049 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1050 		    0,0,0,0);
1051 		/* flush back-page anons? */
1052 		if (amap)
1053 			uvmfault_anonflush(*ranons, nback);
1054 
1055 		/* flush object? */
1056 		if (uobj) {
1057 			voff_t uoff;
1058 
1059 			uoff = ufi->entry->offset + eoff;
1060 			mutex_enter(uobj->vmobjlock);
1061 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1062 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1063 		}
1064 
1065 		/* now forget about the backpages */
1066 		if (amap)
1067 			*ranons += nback;
1068 		flt->startva += (nback << PAGE_SHIFT);
1069 		flt->npages -= nback;
1070 		flt->centeridx = 0;
1071 	}
1072 	/*
1073 	 * => startva is fixed
1074 	 * => npages is fixed
1075 	 */
1076 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1077 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1078 	    flt->startva + (flt->npages << PAGE_SHIFT));
1079 	return 0;
1080 }
1081 
1082 /*
1083  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1084  *
1085  * iterate range of interest:
1086  *	1. check if h/w mapping exists.  if yes, we don't care
1087  *	2. check if anon exists.  if not, page is lower.
1088  *	3. if anon exists, enter h/w mapping for neighbors.
1089  *
1090  * => called with amap locked (if exists).
1091  */
1092 
1093 static int
1094 uvm_fault_upper_lookup(
1095 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1096 	struct vm_anon **anons, struct vm_page **pages)
1097 {
1098 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1099 	int lcv;
1100 	vaddr_t currva;
1101 	bool shadowed;
1102 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1103 
1104 	/* locked: maps(read), amap(if there) */
1105 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1106 
1107 	/*
1108 	 * map in the backpages and frontpages we found in the amap in hopes
1109 	 * of preventing future faults.    we also init the pages[] array as
1110 	 * we go.
1111 	 */
1112 
1113 	currva = flt->startva;
1114 	shadowed = false;
1115 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1116 		/*
1117 		 * don't play with VAs that are already mapped
1118 		 * (except for center)
1119 		 */
1120 		if (lcv != flt->centeridx &&
1121 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1122 			pages[lcv] = PGO_DONTCARE;
1123 			continue;
1124 		}
1125 
1126 		/*
1127 		 * unmapped or center page.   check if any anon at this level.
1128 		 */
1129 		if (amap == NULL || anons[lcv] == NULL) {
1130 			pages[lcv] = NULL;
1131 			continue;
1132 		}
1133 
1134 		/*
1135 		 * check for present page and map if possible.   re-activate it.
1136 		 */
1137 
1138 		pages[lcv] = PGO_DONTCARE;
1139 		if (lcv == flt->centeridx) {	/* save center for later! */
1140 			shadowed = true;
1141 			continue;
1142 		}
1143 
1144 		struct vm_anon *anon = anons[lcv];
1145 		struct vm_page *pg = anon->an_page;
1146 
1147 		KASSERT(anon->an_lock == amap->am_lock);
1148 
1149 		/* Ignore loaned and busy pages. */
1150 		if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0) {
1151 			uvm_fault_upper_neighbor(ufi, flt, currva,
1152 			    pg, anon->an_ref > 1);
1153 		}
1154 	}
1155 
1156 	/* locked: maps(read), amap(if there) */
1157 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1158 	/* (shadowed == true) if there is an anon at the faulting address */
1159 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
1160 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1161 
1162 	/*
1163 	 * note that if we are really short of RAM we could sleep in the above
1164 	 * call to pmap_enter with everything locked.   bad?
1165 	 *
1166 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1167 	 * XXX case.  --thorpej
1168 	 */
1169 
1170 	return 0;
1171 }
1172 
1173 /*
1174  * uvm_fault_upper_neighbor: enter single lower neighbor page.
1175  *
1176  * => called with amap and anon locked.
1177  */
1178 
1179 static void
1180 uvm_fault_upper_neighbor(
1181 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1182 	vaddr_t currva, struct vm_page *pg, bool readonly)
1183 {
1184 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1185 
1186 	/* locked: amap, anon */
1187 
1188 	mutex_enter(&uvm_pageqlock);
1189 	uvm_pageenqueue(pg);
1190 	mutex_exit(&uvm_pageqlock);
1191 	UVMHIST_LOG(maphist,
1192 	    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
1193 	    ufi->orig_map->pmap, currva, pg, 0);
1194 	uvmexp.fltnamap++;
1195 
1196 	/*
1197 	 * Since this page isn't the page that's actually faulting,
1198 	 * ignore pmap_enter() failures; it's not critical that we
1199 	 * enter these right now.
1200 	 */
1201 
1202 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1203 	    VM_PAGE_TO_PHYS(pg),
1204 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1205 	    flt->enter_prot,
1206 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1207 
1208 	pmap_update(ufi->orig_map->pmap);
1209 }
1210 
1211 /*
1212  * uvm_fault_upper: handle upper fault.
1213  *
1214  *	1. acquire anon lock.
1215  *	2. get anon.  let uvmfault_anonget do the dirty work.
1216  *	3. handle loan.
1217  *	4. dispatch direct or promote handlers.
1218  */
1219 
1220 static int
1221 uvm_fault_upper(
1222 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1223 	struct vm_anon **anons)
1224 {
1225 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1226 	struct vm_anon * const anon = anons[flt->centeridx];
1227 	struct uvm_object *uobj;
1228 	int error;
1229 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1230 
1231 	/* locked: maps(read), amap, anon */
1232 	KASSERT(mutex_owned(amap->am_lock));
1233 	KASSERT(anon->an_lock == amap->am_lock);
1234 
1235 	/*
1236 	 * handle case 1: fault on an anon in our amap
1237 	 */
1238 
1239 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1240 
1241 	/*
1242 	 * no matter if we have case 1A or case 1B we are going to need to
1243 	 * have the anon's memory resident.   ensure that now.
1244 	 */
1245 
1246 	/*
1247 	 * let uvmfault_anonget do the dirty work.
1248 	 * if it fails (!OK) it will unlock everything for us.
1249 	 * if it succeeds, locks are still valid and locked.
1250 	 * also, if it is OK, then the anon's page is on the queues.
1251 	 * if the page is on loan from a uvm_object, then anonget will
1252 	 * lock that object for us if it does not fail.
1253 	 */
1254 
1255 	error = uvmfault_anonget(ufi, amap, anon);
1256 	switch (error) {
1257 	case 0:
1258 		break;
1259 
1260 	case ERESTART:
1261 		return ERESTART;
1262 
1263 	case EAGAIN:
1264 		kpause("fltagain1", false, hz/2, NULL);
1265 		return ERESTART;
1266 
1267 	default:
1268 		return error;
1269 	}
1270 
1271 	/*
1272 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1273 	 */
1274 
1275 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1276 
1277 	/* locked: maps(read), amap, anon, uobj(if one) */
1278 	KASSERT(mutex_owned(amap->am_lock));
1279 	KASSERT(anon->an_lock == amap->am_lock);
1280 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1281 
1282 	/*
1283 	 * special handling for loaned pages
1284 	 */
1285 
1286 	if (anon->an_page->loan_count) {
1287 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1288 		if (error != 0)
1289 			return error;
1290 	}
1291 
1292 	/*
1293 	 * if we are case 1B then we will need to allocate a new blank
1294 	 * anon to transfer the data into.   note that we have a lock
1295 	 * on anon, so no one can busy or release the page until we are done.
1296 	 * also note that the ref count can't drop to zero here because
1297 	 * it is > 1 and we are only dropping one ref.
1298 	 *
1299 	 * in the (hopefully very rare) case that we are out of RAM we
1300 	 * will unlock, wait for more RAM, and refault.
1301 	 *
1302 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1303 	 */
1304 
1305 	if (flt->cow_now && anon->an_ref > 1) {
1306 		flt->promote = true;
1307 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1308 	} else {
1309 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1310 	}
1311 	return error;
1312 }
1313 
1314 /*
1315  * uvm_fault_upper_loan: handle loaned upper page.
1316  *
1317  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1318  *	2. if cow'ing now, and if ref count is 1, break loan.
1319  */
1320 
1321 static int
1322 uvm_fault_upper_loan(
1323 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1324 	struct vm_anon *anon, struct uvm_object **ruobj)
1325 {
1326 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1327 	int error = 0;
1328 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1329 
1330 	if (!flt->cow_now) {
1331 
1332 		/*
1333 		 * for read faults on loaned pages we just cap the
1334 		 * protection at read-only.
1335 		 */
1336 
1337 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1338 
1339 	} else {
1340 		/*
1341 		 * note that we can't allow writes into a loaned page!
1342 		 *
1343 		 * if we have a write fault on a loaned page in an
1344 		 * anon then we need to look at the anon's ref count.
1345 		 * if it is greater than one then we are going to do
1346 		 * a normal copy-on-write fault into a new anon (this
1347 		 * is not a problem).  however, if the reference count
1348 		 * is one (a case where we would normally allow a
1349 		 * write directly to the page) then we need to kill
1350 		 * the loan before we continue.
1351 		 */
1352 
1353 		/* >1 case is already ok */
1354 		if (anon->an_ref == 1) {
1355 			error = uvm_loanbreak_anon(anon, *ruobj);
1356 			if (error != 0) {
1357 				uvmfault_unlockall(ufi, amap, *ruobj);
1358 				uvm_wait("flt_noram2");
1359 				return ERESTART;
1360 			}
1361 			/* if we were a loan reciever uobj is gone */
1362 			if (*ruobj)
1363 				*ruobj = NULL;
1364 		}
1365 	}
1366 	return error;
1367 }
1368 
1369 /*
1370  * uvm_fault_upper_promote: promote upper page.
1371  *
1372  *	1. call uvmfault_promote.
1373  *	2. enqueue page.
1374  *	3. deref.
1375  *	4. pass page to uvm_fault_upper_enter.
1376  */
1377 
1378 static int
1379 uvm_fault_upper_promote(
1380 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1381 	struct uvm_object *uobj, struct vm_anon *anon)
1382 {
1383 	struct vm_anon * const oanon = anon;
1384 	struct vm_page *pg;
1385 	int error;
1386 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1387 
1388 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1389 	uvmexp.flt_acow++;
1390 
1391 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1392 	    &flt->anon_spare);
1393 	switch (error) {
1394 	case 0:
1395 		break;
1396 	case ERESTART:
1397 		return ERESTART;
1398 	default:
1399 		return error;
1400 	}
1401 
1402 	KASSERT(anon == NULL || anon->an_lock == oanon->an_lock);
1403 
1404 	pg = anon->an_page;
1405 	mutex_enter(&uvm_pageqlock);
1406 	uvm_pageenqueue(pg); /* uvm_fault_upper_done will activate the page */
1407 	mutex_exit(&uvm_pageqlock);
1408 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1409 	UVM_PAGE_OWN(pg, NULL);
1410 
1411 	/* deref: can not drop to zero here by defn! */
1412 	KASSERT(oanon->an_ref > 1);
1413 	oanon->an_ref--;
1414 
1415 	/*
1416 	 * note: oanon is still locked, as is the new anon.  we
1417 	 * need to check for this later when we unlock oanon; if
1418 	 * oanon != anon, we'll have to unlock anon, too.
1419 	 */
1420 
1421 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1422 }
1423 
1424 /*
1425  * uvm_fault_upper_direct: handle direct fault.
1426  */
1427 
1428 static int
1429 uvm_fault_upper_direct(
1430 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1431 	struct uvm_object *uobj, struct vm_anon *anon)
1432 {
1433 	struct vm_anon * const oanon = anon;
1434 	struct vm_page *pg;
1435 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1436 
1437 	uvmexp.flt_anon++;
1438 	pg = anon->an_page;
1439 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1440 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1441 
1442 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1443 }
1444 
1445 /*
1446  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1447  */
1448 
1449 static int
1450 uvm_fault_upper_enter(
1451 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1452 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1453 	struct vm_anon *oanon)
1454 {
1455 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1456 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1457 
1458 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1459 	KASSERT(mutex_owned(amap->am_lock));
1460 	KASSERT(anon->an_lock == amap->am_lock);
1461 	KASSERT(oanon->an_lock == amap->am_lock);
1462 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1463 
1464 	/*
1465 	 * now map the page in.
1466 	 */
1467 
1468 	UVMHIST_LOG(maphist,
1469 	    "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1470 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
1471 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1472 	    VM_PAGE_TO_PHYS(pg),
1473 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1474 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1475 
1476 		/*
1477 		 * No need to undo what we did; we can simply think of
1478 		 * this as the pmap throwing away the mapping information.
1479 		 *
1480 		 * We do, however, have to go through the ReFault path,
1481 		 * as the map may change while we're asleep.
1482 		 */
1483 
1484 		uvmfault_unlockall(ufi, amap, uobj);
1485 		if (!uvm_reclaimable()) {
1486 			UVMHIST_LOG(maphist,
1487 			    "<- failed.  out of VM",0,0,0,0);
1488 			/* XXX instrumentation */
1489 			return ENOMEM;
1490 		}
1491 		/* XXX instrumentation */
1492 		uvm_wait("flt_pmfail1");
1493 		return ERESTART;
1494 	}
1495 
1496 	uvm_fault_upper_done(ufi, flt, anon, pg);
1497 
1498 	/*
1499 	 * done case 1!  finish up by unlocking everything and returning success
1500 	 */
1501 
1502 	pmap_update(ufi->orig_map->pmap);
1503 	uvmfault_unlockall(ufi, amap, uobj);
1504 	return 0;
1505 }
1506 
1507 /*
1508  * uvm_fault_upper_done: queue upper center page.
1509  */
1510 
1511 static void
1512 uvm_fault_upper_done(
1513 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1514 	struct vm_anon *anon, struct vm_page *pg)
1515 {
1516 	const bool wire_paging = flt->wire_paging;
1517 
1518 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1519 
1520 	/*
1521 	 * ... update the page queues.
1522 	 */
1523 
1524 	mutex_enter(&uvm_pageqlock);
1525 	if (wire_paging) {
1526 		uvm_pagewire(pg);
1527 
1528 		/*
1529 		 * since the now-wired page cannot be paged out,
1530 		 * release its swap resources for others to use.
1531 		 * since an anon with no swap cannot be PG_CLEAN,
1532 		 * clear its clean flag now.
1533 		 */
1534 
1535 		pg->flags &= ~(PG_CLEAN);
1536 
1537 	} else {
1538 		uvm_pageactivate(pg);
1539 	}
1540 	mutex_exit(&uvm_pageqlock);
1541 
1542 	if (wire_paging) {
1543 		uvm_anon_dropswap(anon);
1544 	}
1545 }
1546 
1547 /*
1548  * uvm_fault_lower: handle lower fault.
1549  *
1550  *	1. check uobj
1551  *	1.1. if null, ZFOD.
1552  *	1.2. if not null, look up unnmapped neighbor pages.
1553  *	2. for center page, check if promote.
1554  *	2.1. ZFOD always needs promotion.
1555  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1556  *	3. if uobj is not ZFOD and page is not found, do i/o.
1557  *	4. dispatch either direct / promote fault.
1558  */
1559 
1560 static int
1561 uvm_fault_lower(
1562 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1563 	struct vm_page **pages)
1564 {
1565 #ifdef DIAGNOSTIC
1566 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1567 #endif
1568 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1569 	struct vm_page *uobjpage;
1570 	int error;
1571 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1572 
1573 	/*
1574 	 * now, if the desired page is not shadowed by the amap and we have
1575 	 * a backing object that does not have a special fault routine, then
1576 	 * we ask (with pgo_get) the object for resident pages that we care
1577 	 * about and attempt to map them in.  we do not let pgo_get block
1578 	 * (PGO_LOCKED).
1579 	 */
1580 
1581 	if (uobj == NULL) {
1582 		/* zero fill; don't care neighbor pages */
1583 		uobjpage = NULL;
1584 	} else {
1585 		uvm_fault_lower_lookup(ufi, flt, pages);
1586 		uobjpage = pages[flt->centeridx];
1587 	}
1588 
1589 	/*
1590 	 * note that at this point we are done with any front or back pages.
1591 	 * we are now going to focus on the center page (i.e. the one we've
1592 	 * faulted on).  if we have faulted on the upper (anon) layer
1593 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1594 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1595 	 * layer [i.e. case 2] and the page was both present and available,
1596 	 * then we've got a pointer to it as "uobjpage" and we've already
1597 	 * made it BUSY.
1598 	 */
1599 
1600 	/*
1601 	 * locked:
1602 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1603 	 */
1604 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1605 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1606 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1607 
1608 	/*
1609 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1610 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1611 	 * have a backing object, check and see if we are going to promote
1612 	 * the data up to an anon during the fault.
1613 	 */
1614 
1615 	if (uobj == NULL) {
1616 		uobjpage = PGO_DONTCARE;
1617 		flt->promote = true;		/* always need anon here */
1618 	} else {
1619 		KASSERT(uobjpage != PGO_DONTCARE);
1620 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1621 	}
1622 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1623 	    flt->promote, (uobj == NULL), 0,0);
1624 
1625 	/*
1626 	 * if uobjpage is not null then we do not need to do I/O to get the
1627 	 * uobjpage.
1628 	 *
1629 	 * if uobjpage is null, then we need to unlock and ask the pager to
1630 	 * get the data for us.   once we have the data, we need to reverify
1631 	 * the state the world.   we are currently not holding any resources.
1632 	 */
1633 
1634 	if (uobjpage) {
1635 		/* update rusage counters */
1636 		curlwp->l_ru.ru_minflt++;
1637 	} else {
1638 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1639 		if (error != 0)
1640 			return error;
1641 	}
1642 
1643 	/*
1644 	 * locked:
1645 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1646 	 */
1647 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1648 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1649 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1650 
1651 	/*
1652 	 * notes:
1653 	 *  - at this point uobjpage can not be NULL
1654 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1655 	 *  for it above)
1656 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1657 	 */
1658 
1659 	KASSERT(uobjpage != NULL);
1660 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1661 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1662 	    (uobjpage->flags & PG_CLEAN) != 0);
1663 
1664 	if (!flt->promote) {
1665 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1666 	} else {
1667 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1668 	}
1669 	return error;
1670 }
1671 
1672 /*
1673  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1674  *
1675  *	1. get on-memory pages.
1676  *	2. if failed, give up (get only center page later).
1677  *	3. if succeeded, enter h/w mapping of neighbor pages.
1678  */
1679 
1680 static void
1681 uvm_fault_lower_lookup(
1682 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1683 	struct vm_page **pages)
1684 {
1685 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1686 	int lcv, gotpages;
1687 	vaddr_t currva;
1688 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1689 
1690 	mutex_enter(uobj->vmobjlock);
1691 	/* Locked: maps(read), amap(if there), uobj */
1692 
1693 	uvmexp.fltlget++;
1694 	gotpages = flt->npages;
1695 	(void) uobj->pgops->pgo_get(uobj,
1696 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1697 	    pages, &gotpages, flt->centeridx,
1698 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1699 
1700 	KASSERT(mutex_owned(uobj->vmobjlock));
1701 
1702 	/*
1703 	 * check for pages to map, if we got any
1704 	 */
1705 
1706 	if (gotpages == 0) {
1707 		pages[flt->centeridx] = NULL;
1708 		return;
1709 	}
1710 
1711 	currva = flt->startva;
1712 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1713 		struct vm_page *curpg;
1714 
1715 		curpg = pages[lcv];
1716 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1717 			continue;
1718 		}
1719 		KASSERT(curpg->uobject == uobj);
1720 
1721 		/*
1722 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1723 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1724 		 * to it.
1725 		 */
1726 
1727 		if (lcv == flt->centeridx) {
1728 			UVMHIST_LOG(maphist, "  got uobjpage "
1729 			    "(0x%x) with locked get",
1730 			    curpg, 0,0,0);
1731 		} else {
1732 			bool readonly = (curpg->flags & PG_RDONLY)
1733 			    || (curpg->loan_count > 0)
1734 			    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
1735 
1736 			uvm_fault_lower_neighbor(ufi, flt,
1737 			    currva, curpg, readonly);
1738 		}
1739 	}
1740 	pmap_update(ufi->orig_map->pmap);
1741 }
1742 
1743 /*
1744  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1745  */
1746 
1747 static void
1748 uvm_fault_lower_neighbor(
1749 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1750 	vaddr_t currva, struct vm_page *pg, bool readonly)
1751 {
1752 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1753 
1754 	/* locked: maps(read), amap(if there), uobj */
1755 
1756 	/*
1757 	 * calling pgo_get with PGO_LOCKED returns us pages which
1758 	 * are neither busy nor released, so we don't need to check
1759 	 * for this.  we can just directly enter the pages.
1760 	 */
1761 
1762 	mutex_enter(&uvm_pageqlock);
1763 	uvm_pageenqueue(pg);
1764 	mutex_exit(&uvm_pageqlock);
1765 	UVMHIST_LOG(maphist,
1766 	    "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
1767 	    ufi->orig_map->pmap, currva, pg, 0);
1768 	uvmexp.fltnomap++;
1769 
1770 	/*
1771 	 * Since this page isn't the page that's actually faulting,
1772 	 * ignore pmap_enter() failures; it's not critical that we
1773 	 * enter these right now.
1774 	 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've
1775 	 * held the lock the whole time we've had the handle.
1776 	 */
1777 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1778 	KASSERT((pg->flags & PG_RELEASED) == 0);
1779 	KASSERT((pg->flags & PG_WANTED) == 0);
1780 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) || (pg->flags & PG_CLEAN) != 0);
1781 	pg->flags &= ~(PG_BUSY);
1782 	UVM_PAGE_OWN(pg, NULL);
1783 
1784 	KASSERT(mutex_owned(pg->uobject->vmobjlock));
1785 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1786 	    VM_PAGE_TO_PHYS(pg),
1787 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1788 	    flt->enter_prot & MASK(ufi->entry),
1789 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1790 }
1791 
1792 /*
1793  * uvm_fault_lower_io: get lower page from backing store.
1794  *
1795  *	1. unlock everything, because i/o will block.
1796  *	2. call pgo_get.
1797  *	3. if failed, recover.
1798  *	4. if succeeded, relock everything and verify things.
1799  */
1800 
1801 static int
1802 uvm_fault_lower_io(
1803 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1804 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1805 {
1806 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1807 	struct uvm_object *uobj = *ruobj;
1808 	struct vm_page *pg;
1809 	bool locked;
1810 	int gotpages;
1811 	int error;
1812 	voff_t uoff;
1813 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1814 
1815 	/* update rusage counters */
1816 	curlwp->l_ru.ru_majflt++;
1817 
1818 	/* Locked: maps(read), amap(if there), uobj */
1819 	uvmfault_unlockall(ufi, amap, NULL);
1820 
1821 	/* Locked: uobj */
1822 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1823 
1824 	uvmexp.fltget++;
1825 	gotpages = 1;
1826 	pg = NULL;
1827 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1828 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1829 	    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1830 	    PGO_SYNCIO);
1831 	/* locked: pg(if no error) */
1832 
1833 	/*
1834 	 * recover from I/O
1835 	 */
1836 
1837 	if (error) {
1838 		if (error == EAGAIN) {
1839 			UVMHIST_LOG(maphist,
1840 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1841 			kpause("fltagain2", false, hz/2, NULL);
1842 			return ERESTART;
1843 		}
1844 
1845 #if 0
1846 		KASSERT(error != ERESTART);
1847 #else
1848 		/* XXXUEBS don't re-fault? */
1849 		if (error == ERESTART)
1850 			error = EIO;
1851 #endif
1852 
1853 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1854 		    error, 0,0,0);
1855 		return error;
1856 	}
1857 
1858 	/*
1859 	 * re-verify the state of the world by first trying to relock
1860 	 * the maps.  always relock the object.
1861 	 */
1862 
1863 	locked = uvmfault_relock(ufi);
1864 	if (locked && amap)
1865 		amap_lock(amap);
1866 
1867 	/* might be changed */
1868 	uobj = pg->uobject;
1869 
1870 	mutex_enter(uobj->vmobjlock);
1871 	KASSERT((pg->flags & PG_BUSY) != 0);
1872 
1873 	mutex_enter(&uvm_pageqlock);
1874 	uvm_pageactivate(pg);
1875 	mutex_exit(&uvm_pageqlock);
1876 
1877 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1878 	/* locked(!locked): uobj, pg */
1879 
1880 	/*
1881 	 * verify that the page has not be released and re-verify
1882 	 * that amap slot is still free.   if there is a problem,
1883 	 * we unlock and clean up.
1884 	 */
1885 
1886 	if ((pg->flags & PG_RELEASED) != 0 ||
1887 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1888 	      ufi->orig_rvaddr - ufi->entry->start))) {
1889 		if (locked)
1890 			uvmfault_unlockall(ufi, amap, NULL);
1891 		locked = false;
1892 	}
1893 
1894 	/*
1895 	 * didn't get the lock?   release the page and retry.
1896 	 */
1897 
1898 	if (locked == false) {
1899 		UVMHIST_LOG(maphist,
1900 		    "  wasn't able to relock after fault: retry",
1901 		    0,0,0,0);
1902 		if (pg->flags & PG_WANTED) {
1903 			wakeup(pg);
1904 		}
1905 		if ((pg->flags & PG_RELEASED) == 0) {
1906 			pg->flags &= ~(PG_BUSY | PG_WANTED);
1907 			UVM_PAGE_OWN(pg, NULL);
1908 		} else {
1909 			uvmexp.fltpgrele++;
1910 			uvm_pagefree(pg);
1911 		}
1912 		mutex_exit(uobj->vmobjlock);
1913 		return ERESTART;
1914 	}
1915 
1916 	/*
1917 	 * we have the data in pg which is busy and
1918 	 * not released.  we are holding object lock (so the page
1919 	 * can't be released on us).
1920 	 */
1921 
1922 	/* locked: maps(read), amap(if !null), uobj, pg */
1923 
1924 	*ruobj = uobj;
1925 	*ruobjpage = pg;
1926 	return 0;
1927 }
1928 
1929 /*
1930  * uvm_fault_lower_direct: fault lower center page
1931  *
1932  *	1. adjust flt->enter_prot.
1933  *	2. if page is loaned, resolve.
1934  */
1935 
1936 int
1937 uvm_fault_lower_direct(
1938 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1939 	struct uvm_object *uobj, struct vm_page *uobjpage)
1940 {
1941 	struct vm_page *pg;
1942 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
1943 
1944 	/*
1945 	 * we are not promoting.   if the mapping is COW ensure that we
1946 	 * don't give more access than we should (e.g. when doing a read
1947 	 * fault on a COPYONWRITE mapping we want to map the COW page in
1948 	 * R/O even though the entry protection could be R/W).
1949 	 *
1950 	 * set "pg" to the page we want to map in (uobjpage, usually)
1951 	 */
1952 
1953 	uvmexp.flt_obj++;
1954 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
1955 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
1956 		flt->enter_prot &= ~VM_PROT_WRITE;
1957 	pg = uobjpage;		/* map in the actual object */
1958 
1959 	KASSERT(uobjpage != PGO_DONTCARE);
1960 
1961 	/*
1962 	 * we are faulting directly on the page.   be careful
1963 	 * about writing to loaned pages...
1964 	 */
1965 
1966 	if (uobjpage->loan_count) {
1967 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
1968 	}
1969 	KASSERT(pg == uobjpage);
1970 
1971 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1972 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
1973 }
1974 
1975 /*
1976  * uvm_fault_lower_direct_loan: resolve loaned page.
1977  *
1978  *	1. if not cow'ing, adjust flt->enter_prot.
1979  *	2. if cow'ing, break loan.
1980  */
1981 
1982 static int
1983 uvm_fault_lower_direct_loan(
1984 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1985 	struct uvm_object *uobj, struct vm_page **rpg,
1986 	struct vm_page **ruobjpage)
1987 {
1988 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1989 	struct vm_page *pg;
1990 	struct vm_page *uobjpage = *ruobjpage;
1991 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
1992 
1993 	if (!flt->cow_now) {
1994 		/* read fault: cap the protection at readonly */
1995 		/* cap! */
1996 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1997 	} else {
1998 		/* write fault: must break the loan here */
1999 
2000 		pg = uvm_loanbreak(uobjpage);
2001 		if (pg == NULL) {
2002 
2003 			/*
2004 			 * drop ownership of page, it can't be released
2005 			 */
2006 
2007 			if (uobjpage->flags & PG_WANTED)
2008 				wakeup(uobjpage);
2009 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2010 			UVM_PAGE_OWN(uobjpage, NULL);
2011 
2012 			uvmfault_unlockall(ufi, amap, uobj);
2013 			UVMHIST_LOG(maphist,
2014 			  "  out of RAM breaking loan, waiting",
2015 			  0,0,0,0);
2016 			uvmexp.fltnoram++;
2017 			uvm_wait("flt_noram4");
2018 			return ERESTART;
2019 		}
2020 		*rpg = pg;
2021 		*ruobjpage = pg;
2022 	}
2023 	return 0;
2024 }
2025 
2026 /*
2027  * uvm_fault_lower_promote: promote lower page.
2028  *
2029  *	1. call uvmfault_promote.
2030  *	2. fill in data.
2031  *	3. if not ZFOD, dispose old page.
2032  */
2033 
2034 int
2035 uvm_fault_lower_promote(
2036 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2037 	struct uvm_object *uobj, struct vm_page *uobjpage)
2038 {
2039 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2040 	struct vm_anon *anon;
2041 	struct vm_page *pg;
2042 	int error;
2043 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2044 
2045 	KASSERT(amap != NULL);
2046 
2047 	/*
2048 	 * If we are going to promote the data to an anon we
2049 	 * allocate a blank anon here and plug it into our amap.
2050 	 */
2051 	error = uvmfault_promote(ufi, NULL, uobjpage,
2052 	    &anon, &flt->anon_spare);
2053 	switch (error) {
2054 	case 0:
2055 		break;
2056 	case ERESTART:
2057 		return ERESTART;
2058 	default:
2059 		return error;
2060 	}
2061 
2062 	pg = anon->an_page;
2063 
2064 	/*
2065 	 * Fill in the data.
2066 	 */
2067 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2068 
2069 	if (uobjpage != PGO_DONTCARE) {
2070 		uvmexp.flt_prcopy++;
2071 
2072 		/*
2073 		 * promote to shared amap?  make sure all sharing
2074 		 * procs see it
2075 		 */
2076 
2077 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2078 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2079 			/*
2080 			 * XXX: PAGE MIGHT BE WIRED!
2081 			 */
2082 		}
2083 
2084 		/*
2085 		 * dispose of uobjpage.  it can't be PG_RELEASED
2086 		 * since we still hold the object lock.
2087 		 */
2088 
2089 		if (uobjpage->flags & PG_WANTED) {
2090 			/* still have the obj lock */
2091 			wakeup(uobjpage);
2092 		}
2093 		uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2094 		UVM_PAGE_OWN(uobjpage, NULL);
2095 
2096 		UVMHIST_LOG(maphist,
2097 		    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
2098 		    uobjpage, anon, pg, 0);
2099 
2100 	} else {
2101 		uvmexp.flt_przero++;
2102 
2103 		/*
2104 		 * Page is zero'd and marked dirty by
2105 		 * uvmfault_promote().
2106 		 */
2107 
2108 		UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
2109 		    anon, pg, 0, 0);
2110 	}
2111 
2112 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2113 }
2114 
2115 /*
2116  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2117  * from the lower page.
2118  */
2119 
2120 int
2121 uvm_fault_lower_enter(
2122 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2123 	struct uvm_object *uobj,
2124 	struct vm_anon *anon, struct vm_page *pg)
2125 {
2126 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2127 	int error;
2128 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2129 
2130 	/*
2131 	 * Locked:
2132 	 *
2133 	 *	maps(read), amap(if !null), uobj(if !null),
2134 	 *	anon(if !null), pg(if anon), unlock_uobj(if !null)
2135 	 *
2136 	 * Note: pg is either the uobjpage or the new page in the new anon.
2137 	 */
2138 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
2139 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
2140 	KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
2141 	KASSERT((pg->flags & PG_BUSY) != 0);
2142 
2143 	/*
2144 	 * all resources are present.   we can now map it in and free our
2145 	 * resources.
2146 	 */
2147 
2148 	UVMHIST_LOG(maphist,
2149 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
2150 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
2151 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 ||
2152 		(pg->flags & PG_RDONLY) == 0);
2153 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2154 	    VM_PAGE_TO_PHYS(pg),
2155 	    (pg->flags & PG_RDONLY) != 0 ?
2156 	    flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2157 	    flt->access_type | PMAP_CANFAIL |
2158 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2159 
2160 		/*
2161 		 * No need to undo what we did; we can simply think of
2162 		 * this as the pmap throwing away the mapping information.
2163 		 *
2164 		 * We do, however, have to go through the ReFault path,
2165 		 * as the map may change while we're asleep.
2166 		 */
2167 
2168 		/*
2169 		 * ensure that the page is queued in the case that
2170 		 * we just promoted the page.
2171 		 */
2172 
2173 		mutex_enter(&uvm_pageqlock);
2174 		uvm_pageenqueue(pg);
2175 		mutex_exit(&uvm_pageqlock);
2176 
2177 		if (pg->flags & PG_WANTED)
2178 			wakeup(pg);
2179 
2180 		/*
2181 		 * note that pg can't be PG_RELEASED since we did not drop
2182 		 * the object lock since the last time we checked.
2183 		 */
2184 		KASSERT((pg->flags & PG_RELEASED) == 0);
2185 
2186 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2187 		UVM_PAGE_OWN(pg, NULL);
2188 
2189 		uvmfault_unlockall(ufi, amap, uobj);
2190 		if (!uvm_reclaimable()) {
2191 			UVMHIST_LOG(maphist,
2192 			    "<- failed.  out of VM",0,0,0,0);
2193 			/* XXX instrumentation */
2194 			error = ENOMEM;
2195 			return error;
2196 		}
2197 		/* XXX instrumentation */
2198 		uvm_wait("flt_pmfail2");
2199 		return ERESTART;
2200 	}
2201 
2202 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2203 
2204 	/*
2205 	 * note that pg can't be PG_RELEASED since we did not drop the object
2206 	 * lock since the last time we checked.
2207 	 */
2208 	KASSERT((pg->flags & PG_RELEASED) == 0);
2209 	if (pg->flags & PG_WANTED)
2210 		wakeup(pg);
2211 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2212 	UVM_PAGE_OWN(pg, NULL);
2213 
2214 	pmap_update(ufi->orig_map->pmap);
2215 	uvmfault_unlockall(ufi, amap, uobj);
2216 
2217 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2218 	return 0;
2219 }
2220 
2221 /*
2222  * uvm_fault_lower_done: queue lower center page.
2223  */
2224 
2225 void
2226 uvm_fault_lower_done(
2227 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2228 	struct uvm_object *uobj, struct vm_page *pg)
2229 {
2230 	bool dropswap = false;
2231 
2232 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2233 
2234 	mutex_enter(&uvm_pageqlock);
2235 	if (flt->wire_paging) {
2236 		uvm_pagewire(pg);
2237 		if (pg->pqflags & PQ_AOBJ) {
2238 
2239 			/*
2240 			 * since the now-wired page cannot be paged out,
2241 			 * release its swap resources for others to use.
2242 			 * since an aobj page with no swap cannot be PG_CLEAN,
2243 			 * clear its clean flag now.
2244 			 */
2245 
2246 			KASSERT(uobj != NULL);
2247 			pg->flags &= ~(PG_CLEAN);
2248 			dropswap = true;
2249 		}
2250 	} else {
2251 		uvm_pageactivate(pg);
2252 	}
2253 	mutex_exit(&uvm_pageqlock);
2254 
2255 	if (dropswap) {
2256 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2257 	}
2258 }
2259 
2260 
2261 /*
2262  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2263  *
2264  * => map may be read-locked by caller, but MUST NOT be write-locked.
2265  * => if map is read-locked, any operations which may cause map to
2266  *	be write-locked in uvm_fault() must be taken care of by
2267  *	the caller.  See uvm_map_pageable().
2268  */
2269 
2270 int
2271 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2272     vm_prot_t access_type, int maxprot)
2273 {
2274 	vaddr_t va;
2275 	int error;
2276 
2277 	/*
2278 	 * now fault it in a page at a time.   if the fault fails then we have
2279 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2280 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2281 	 */
2282 
2283 	/*
2284 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2285 	 * wiring the last page in the address space, though.
2286 	 */
2287 	if (start > end) {
2288 		return EFAULT;
2289 	}
2290 
2291 	for (va = start; va < end; va += PAGE_SIZE) {
2292 		error = uvm_fault_internal(map, va, access_type,
2293 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2294 		if (error) {
2295 			if (va != start) {
2296 				uvm_fault_unwire(map, start, va);
2297 			}
2298 			return error;
2299 		}
2300 	}
2301 	return 0;
2302 }
2303 
2304 /*
2305  * uvm_fault_unwire(): unwire range of virtual space.
2306  */
2307 
2308 void
2309 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2310 {
2311 	vm_map_lock_read(map);
2312 	uvm_fault_unwire_locked(map, start, end);
2313 	vm_map_unlock_read(map);
2314 }
2315 
2316 /*
2317  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2318  *
2319  * => map must be at least read-locked.
2320  */
2321 
2322 void
2323 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2324 {
2325 	struct vm_map_entry *entry, *oentry;
2326 	pmap_t pmap = vm_map_pmap(map);
2327 	vaddr_t va;
2328 	paddr_t pa;
2329 	struct vm_page *pg;
2330 
2331 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
2332 
2333 	/*
2334 	 * we assume that the area we are unwiring has actually been wired
2335 	 * in the first place.   this means that we should be able to extract
2336 	 * the PAs from the pmap.   we also lock out the page daemon so that
2337 	 * we can call uvm_pageunwire.
2338 	 */
2339 
2340 	/*
2341 	 * find the beginning map entry for the region.
2342 	 */
2343 
2344 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2345 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2346 		panic("uvm_fault_unwire_locked: address not in map");
2347 
2348 	oentry = NULL;
2349 	for (va = start; va < end; va += PAGE_SIZE) {
2350 		if (pmap_extract(pmap, va, &pa) == false)
2351 			continue;
2352 
2353 		/*
2354 		 * find the map entry for the current address.
2355 		 */
2356 
2357 		KASSERT(va >= entry->start);
2358 		while (va >= entry->end) {
2359 			KASSERT(entry->next != &map->header &&
2360 				entry->next->start <= entry->end);
2361 			entry = entry->next;
2362 		}
2363 
2364 		/*
2365 		 * lock it.
2366 		 */
2367 
2368 		if (entry != oentry) {
2369 			if (oentry != NULL) {
2370 				mutex_exit(&uvm_pageqlock);
2371 				uvm_map_unlock_entry(oentry);
2372 			}
2373 			uvm_map_lock_entry(entry);
2374 			mutex_enter(&uvm_pageqlock);
2375 			oentry = entry;
2376 		}
2377 
2378 		/*
2379 		 * if the entry is no longer wired, tell the pmap.
2380 		 */
2381 
2382 		if (VM_MAPENT_ISWIRED(entry) == 0)
2383 			pmap_unwire(pmap, va);
2384 
2385 		pg = PHYS_TO_VM_PAGE(pa);
2386 		if (pg)
2387 			uvm_pageunwire(pg);
2388 	}
2389 
2390 	if (oentry != NULL) {
2391 		mutex_exit(&uvm_pageqlock);
2392 		uvm_map_unlock_entry(entry);
2393 	}
2394 }
2395