xref: /netbsd-src/sys/uvm/uvm_fault.c (revision dc306354b0b29af51801a7632f1e95265a68cd81)
1 /*	$NetBSD: uvm_fault.c,v 1.18 1998/11/20 19:37:06 chuck Exp $	*/
2 
3 /*
4  * XXXCDC: "ROUGH DRAFT" QUALITY UVM PRE-RELEASE FILE!
5  *	   >>>USE AT YOUR OWN RISK, WORK IS NOT FINISHED<<<
6  */
7 /*
8  *
9  * Copyright (c) 1997 Charles D. Cranor and Washington University.
10  * All rights reserved.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *      This product includes software developed by Charles D. Cranor and
23  *      Washington University.
24  * 4. The name of the author may not be used to endorse or promote products
25  *    derived from this software without specific prior written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
28  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
29  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
30  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
31  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
32  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
33  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
34  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
35  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
36  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37  *
38  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
39  */
40 
41 #include "opt_uvmhist.h"
42 
43 /*
44  * uvm_fault.c: fault handler
45  */
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/kernel.h>
50 #include <sys/proc.h>
51 #include <sys/malloc.h>
52 #include <sys/mman.h>
53 #include <sys/user.h>
54 
55 #include <vm/vm.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_kern.h>
58 
59 #include <uvm/uvm.h>
60 
61 /*
62  *
63  * a word on page faults:
64  *
65  * types of page faults we handle:
66  *
67  * CASE 1: upper layer faults                   CASE 2: lower layer faults
68  *
69  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
70  *    read/write1     write>1                  read/write   +-cow_write/zero
71  *         |             |                         |        |
72  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
73  * amap |  V  |       |  ----------->new|          |        | |  ^  |
74  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
75  *                                                 |        |    |
76  *      +-----+       +-----+                   +--|--+     | +--|--+
77  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
78  *      +-----+       +-----+                   +-----+       +-----+
79  *
80  * d/c = don't care
81  *
82  *   case [0]: layerless fault
83  *	no amap or uobj is present.   this is an error.
84  *
85  *   case [1]: upper layer fault [anon active]
86  *     1A: [read] or [write with anon->an_ref == 1]
87  *		I/O takes place in top level anon and uobj is not touched.
88  *     1B: [write with anon->an_ref > 1]
89  *		new anon is alloc'd and data is copied off ["COW"]
90  *
91  *   case [2]: lower layer fault [uobj]
92  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
93  *		I/O takes place directly in object.
94  *     2B: [write to copy_on_write] or [read on NULL uobj]
95  *		data is "promoted" from uobj to a new anon.
96  *		if uobj is null, then we zero fill.
97  *
98  * we follow the standard UVM locking protocol ordering:
99  *
100  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
101  * we hold a PG_BUSY page if we unlock for I/O
102  *
103  *
104  * the code is structured as follows:
105  *
106  *     - init the "IN" params in the ufi structure
107  *   ReFault:
108  *     - do lookups [locks maps], check protection, handle needs_copy
109  *     - check for case 0 fault (error)
110  *     - establish "range" of fault
111  *     - if we have an amap lock it and extract the anons
112  *     - if sequential advice deactivate pages behind us
113  *     - at the same time check pmap for unmapped areas and anon for pages
114  *	 that we could map in (and do map it if found)
115  *     - check object for resident pages that we could map in
116  *     - if (case 2) goto Case2
117  *     - >>> handle case 1
118  *           - ensure source anon is resident in RAM
119  *           - if case 1B alloc new anon and copy from source
120  *           - map the correct page in
121  *   Case2:
122  *     - >>> handle case 2
123  *           - ensure source page is resident (if uobj)
124  *           - if case 2B alloc new anon and copy from source (could be zero
125  *		fill if uobj == NULL)
126  *           - map the correct page in
127  *     - done!
128  *
129  * note on paging:
130  *   if we have to do I/O we place a PG_BUSY page in the correct object,
131  * unlock everything, and do the I/O.   when I/O is done we must reverify
132  * the state of the world before assuming that our data structures are
133  * valid.   [because mappings could change while the map is unlocked]
134  *
135  *  alternative 1: unbusy the page in question and restart the page fault
136  *    from the top (ReFault).   this is easy but does not take advantage
137  *    of the information that we already have from our previous lookup,
138  *    although it is possible that the "hints" in the vm_map will help here.
139  *
140  * alternative 2: the system already keeps track of a "version" number of
141  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
142  *    mapping) you bump the version number up by one...]   so, we can save
143  *    the version number of the map before we release the lock and start I/O.
144  *    then when I/O is done we can relock and check the version numbers
145  *    to see if anything changed.    this might save us some over 1 because
146  *    we don't have to unbusy the page and may be less compares(?).
147  *
148  * alternative 3: put in backpointers or a way to "hold" part of a map
149  *    in place while I/O is in progress.   this could be complex to
150  *    implement (especially with structures like amap that can be referenced
151  *    by multiple map entries, and figuring out what should wait could be
152  *    complex as well...).
153  *
154  * given that we are not currently multiprocessor or multithreaded we might
155  * as well choose alternative 2 now.   maybe alternative 3 would be useful
156  * in the future.    XXX keep in mind for future consideration//rechecking.
157  */
158 
159 /*
160  * local data structures
161  */
162 
163 struct uvm_advice {
164 	int advice;
165 	int nback;
166 	int nforw;
167 };
168 
169 /*
170  * page range array:
171  * note: index in array must match "advice" value
172  * XXX: borrowed numbers from freebsd.   do they work well for us?
173  */
174 
175 static struct uvm_advice uvmadvice[] = {
176 	{ MADV_NORMAL, 3, 4 },
177 	{ MADV_RANDOM, 0, 0 },
178 	{ MADV_SEQUENTIAL, 8, 7},
179 };
180 
181 #define UVM_MAXRANGE 16	/* must be max() of nback+nforw+1 */
182 
183 /*
184  * private prototypes
185  */
186 
187 static void uvmfault_amapcopy __P((struct uvm_faultinfo *));
188 static __inline void uvmfault_anonflush __P((struct vm_anon **, int));
189 
190 /*
191  * inline functions
192  */
193 
194 /*
195  * uvmfault_anonflush: try and deactivate pages in specified anons
196  *
197  * => does not have to deactivate page if it is busy
198  */
199 
200 static __inline void
201 uvmfault_anonflush(anons, n)
202 	struct vm_anon **anons;
203 	int n;
204 {
205 	int lcv;
206 	struct vm_page *pg;
207 
208 	for (lcv = 0 ; lcv < n ; lcv++) {
209 		if (anons[lcv] == NULL)
210 			continue;
211 		simple_lock(&anons[lcv]->an_lock);
212 		pg = anons[lcv]->u.an_page;
213 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
214 			uvm_lock_pageq();
215 			if (pg->wire_count == 0) {
216 				pmap_page_protect(PMAP_PGARG(pg), VM_PROT_NONE);
217 				uvm_pagedeactivate(pg);
218 			}
219 			uvm_unlock_pageq();
220 		}
221 		simple_unlock(&anons[lcv]->an_lock);
222 	}
223 }
224 
225 /*
226  * normal functions
227  */
228 
229 /*
230  * uvmfault_amapcopy: clear "needs_copy" in a map.
231  *
232  * => called with VM data structures unlocked (usually, see below)
233  * => we get a write lock on the maps and clear needs_copy for a VA
234  * => if we are out of RAM we sleep (waiting for more)
235  */
236 
237 static void
238 uvmfault_amapcopy(ufi)
239 	struct uvm_faultinfo *ufi;
240 {
241 
242 	/*
243 	 * while we haven't done the job
244 	 */
245 
246 	while (1) {
247 
248 		/*
249 		 * no mapping?  give up.
250 		 */
251 
252 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
253 			return;
254 
255 		/*
256 		 * copy if needed.
257 		 */
258 
259 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
260 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
261 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
262 
263 		/*
264 		 * didn't work?  must be out of RAM.   unlock and sleep.
265 		 */
266 
267 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
268 			uvmfault_unlockmaps(ufi, TRUE);
269 			uvm_wait("fltamapcopy");
270 			continue;
271 		}
272 
273 		/*
274 		 * got it!   unlock and return.
275 		 */
276 
277 		uvmfault_unlockmaps(ufi, TRUE);
278 		return;
279 	}
280 	/*NOTREACHED*/
281 }
282 
283 /*
284  * uvmfault_anonget: get data in an anon into a non-busy, non-released
285  * page in that anon.
286  *
287  * => maps, amap, and anon locked by caller.
288  * => if we fail (result != VM_PAGER_OK) we unlock everything.
289  * => if we are successful, we return with everything still locked.
290  * => we don't move the page on the queues [gets moved later]
291  * => if we allocate a new page [we_own], it gets put on the queues.
292  *    either way, the result is that the page is on the queues at return time
293  * => for pages which are on loan from a uvm_object (and thus are not
294  *    owned by the anon): if successful, we return with the owning object
295  *    locked.   the caller must unlock this object when it unlocks everything
296  *    else.
297  */
298 
299 int uvmfault_anonget(ufi, amap, anon)
300 	struct uvm_faultinfo *ufi;
301 	struct vm_amap *amap;
302 	struct vm_anon *anon;
303 {
304 	boolean_t we_own;	/* we own anon's page? */
305 	boolean_t locked;	/* did we relock? */
306 	struct vm_page *pg;
307 	int result;
308 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
309 
310 	result = 0;		/* XXX shut up gcc */
311 	uvmexp.fltanget++;
312         /* bump rusage counters */
313 	if (anon->u.an_page)
314 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
315 	else
316 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
317 
318 	/*
319 	 * loop until we get it, or fail.
320 	 */
321 
322 	while (1) {
323 
324 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
325 		pg = anon->u.an_page;
326 
327 		/*
328 		 * if there is a resident page and it is loaned, then anon
329 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
330 		 * the real owner of the page has been identified and locked.
331 		 */
332 
333 		if (pg && pg->loan_count)
334 			pg = uvm_anon_lockloanpg(anon);
335 
336 		/*
337 		 * page there?   make sure it is not busy/released.
338 		 */
339 
340 		if (pg) {
341 
342 			/*
343 			 * at this point, if the page has a uobject [meaning
344 			 * we have it on loan], then that uobject is locked
345 			 * by us!   if the page is busy, we drop all the
346 			 * locks (including uobject) and try again.
347 			 */
348 
349 			if ((pg->flags & (PG_BUSY|PG_RELEASED)) == 0) {
350 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
351 				return (VM_PAGER_OK);
352 			}
353 			pg->flags |= PG_WANTED;
354 			uvmexp.fltpgwait++;
355 
356 			/*
357 			 * the last unlock must be an atomic unlock+wait on
358 			 * the owner of page
359 			 */
360 			if (pg->uobject) {	/* owner is uobject ? */
361 				uvmfault_unlockall(ufi, amap, NULL, anon);
362 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
363 				    0,0,0);
364 				UVM_UNLOCK_AND_WAIT(pg,
365 				    &pg->uobject->vmobjlock,
366 				    FALSE, "anonget1",0);
367 			} else {
368 				/* anon owns page */
369 				uvmfault_unlockall(ufi, amap, NULL, NULL);
370 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
371 				    0,0,0);
372 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
373 				    "anonget2",0);
374 			}
375 			/* ready to relock and try again */
376 
377 		} else {
378 
379 			/*
380 			 * no page, we must try and bring it in.
381 			 */
382 			pg = uvm_pagealloc(NULL, 0, anon);
383 
384 			if (pg == NULL) {		/* out of RAM.  */
385 
386 				uvmfault_unlockall(ufi, amap, NULL, anon);
387 				uvmexp.fltnoram++;
388 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
389 				    0,0,0);
390 				uvm_wait("flt_noram1");
391 				/* ready to relock and try again */
392 
393 			} else {
394 
395 				/* we set the PG_BUSY bit */
396 				we_own = TRUE;
397 				uvmfault_unlockall(ufi, amap, NULL, anon);
398 
399 				/*
400 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
401 				 * page into the uvm_swap_get function with
402 				 * all data structures unlocked.  note that
403 				 * it is ok to read an_swslot here because
404 				 * we hold PG_BUSY on the page.
405 				 */
406 				uvmexp.pageins++;
407 				result = uvm_swap_get(pg, anon->an_swslot,
408 				    PGO_SYNCIO);
409 
410 				/*
411 				 * we clean up after the i/o below in the
412 				 * "we_own" case
413 				 */
414 				/* ready to relock and try again */
415 			}
416 		}
417 
418 		/*
419 		 * now relock and try again
420 		 */
421 
422 		locked = uvmfault_relock(ufi);
423 		if (locked) {
424 			simple_lock(&amap->am_l);
425 		}
426 		if (locked || we_own)
427 			simple_lock(&anon->an_lock);
428 
429 		/*
430 		 * if we own the page (i.e. we set PG_BUSY), then we need
431 		 * to clean up after the I/O. there are three cases to
432 		 * consider:
433 		 *   [1] page released during I/O: free anon and ReFault.
434 		 *   [2] I/O not OK.   free the page and cause the fault
435 		 *       to fail.
436 		 *   [3] I/O OK!   activate the page and sync with the
437 		 *       non-we_own case (i.e. drop anon lock if not locked).
438 		 */
439 
440 		if (we_own) {
441 
442 			if (pg->flags & PG_WANTED) {
443 				/* still holding object lock */
444 				thread_wakeup(pg);
445 			}
446 			/* un-busy! */
447 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
448 			UVM_PAGE_OWN(pg, NULL);
449 
450 			/*
451 			 * if we were RELEASED during I/O, then our anon is
452 			 * no longer part of an amap.   we need to free the
453 			 * anon and try again.
454 			 */
455 			if (pg->flags & PG_RELEASED) {
456 				pmap_page_protect(PMAP_PGARG(pg),
457 				    VM_PROT_NONE); /* to be safe */
458 				simple_unlock(&anon->an_lock);
459 				uvm_anfree(anon);	/* frees page for us */
460 				if (locked)
461 				  uvmfault_unlockall(ufi, amap, NULL, NULL);
462 				uvmexp.fltpgrele++;
463 				UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
464 				return (VM_PAGER_REFAULT);	/* refault! */
465 			}
466 
467 			if (result != VM_PAGER_OK) {
468 #ifdef DIAGNOSTIC
469 				if (result == VM_PAGER_PEND)
470 		panic("uvmfault_anonget: got PENDING for non-async I/O");
471 #endif
472 				/* remove page from anon */
473 				anon->u.an_page = NULL;
474 
475 				/*
476 				 * note: page was never !PG_BUSY, so it
477 				 * can't be mapped and thus no need to
478 				 * pmap_page_protect it...
479 				 */
480 				uvm_lock_pageq();
481 				uvm_pagefree(pg);
482 				uvm_unlock_pageq();
483 
484 				if (locked)
485 					uvmfault_unlockall(ufi, amap, NULL,
486 					    anon);
487 				else
488 					simple_unlock(&anon->an_lock);
489 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
490 				return (VM_PAGER_ERROR);
491 			}
492 
493 			/*
494 			 * must be OK, clear modify (already PG_CLEAN)
495 			 * and activate
496 			 */
497 			pmap_clear_modify(PMAP_PGARG(pg));
498 			uvm_lock_pageq();
499 			uvm_pageactivate(pg);
500 			uvm_unlock_pageq();
501 			if (!locked)
502 				simple_unlock(&anon->an_lock);
503 		}
504 
505 		/*
506 		 * we were not able to relock.   restart fault.
507 		 */
508 
509 		if (!locked) {
510 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
511 			return (VM_PAGER_REFAULT);
512 		}
513 
514 		/*
515 		 * verify no one has touched the amap and moved the anon on us.
516 		 */
517 
518 		if (amap_lookup(&ufi->entry->aref,
519 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
520 
521 			uvmfault_unlockall(ufi, amap, NULL, anon);
522 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
523 			return (VM_PAGER_REFAULT);
524 		}
525 
526 		/*
527 		 * try it again!
528 		 */
529 
530 		uvmexp.fltanretry++;
531 		continue;
532 
533 	} /* while (1) */
534 
535 	/*NOTREACHED*/
536 }
537 
538 /*
539  *   F A U L T   -   m a i n   e n t r y   p o i n t
540  */
541 
542 /*
543  * uvm_fault: page fault handler
544  *
545  * => called from MD code to resolve a page fault
546  * => VM data structures usually should be unlocked.   however, it is
547  *	possible to call here with the main map locked if the caller
548  *	gets a write lock, sets it recusive, and then calls us (c.f.
549  *	uvm_map_pageable).   this should be avoided because it keeps
550  *	the map locked off during I/O.
551  */
552 
553 int
554 uvm_fault(orig_map, vaddr, fault_type, access_type)
555 	vm_map_t orig_map;
556 	vaddr_t vaddr;
557 	vm_fault_t fault_type;
558 	vm_prot_t access_type;
559 {
560 	struct uvm_faultinfo ufi;
561 	vm_prot_t enter_prot;
562 	boolean_t wired, narrow, promote, locked, shadowed;
563 	int npages, nback, nforw, centeridx, result, lcv, gotpages;
564 	vaddr_t startva, objaddr, currva, offset;
565 	paddr_t pa;
566 	struct vm_amap *amap;
567 	struct uvm_object *uobj;
568 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
569 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
570 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
571 
572 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
573 	      orig_map, vaddr, fault_type, access_type);
574 
575 	anon = NULL; /* XXX: shut up gcc */
576 
577 	uvmexp.faults++;	/* XXX: locking? */
578 
579 	/*
580 	 * init the IN parameters in the ufi
581 	 */
582 
583 	ufi.orig_map = orig_map;
584 	ufi.orig_rvaddr = trunc_page(vaddr);
585 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
586 	if (fault_type == VM_FAULT_WIRE)
587 		narrow = TRUE;		/* don't look for neighborhood
588 					 * pages on wire */
589 	else
590 		narrow = FALSE;		/* normal fault */
591 
592 	/*
593 	 * "goto ReFault" means restart the page fault from ground zero.
594 	 */
595 ReFault:
596 
597 	/*
598 	 * lookup and lock the maps
599 	 */
600 
601 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
602 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
603 		return (KERN_INVALID_ADDRESS);
604 	}
605 	/* locked: maps(read) */
606 
607 	/*
608 	 * check protection
609 	 */
610 
611 	if ((ufi.entry->protection & access_type) != access_type) {
612 		UVMHIST_LOG(maphist,
613 		    "<- protection failure (prot=0x%x, access=0x%x)",
614 		    ufi.entry->protection, access_type, 0, 0);
615 		uvmfault_unlockmaps(&ufi, FALSE);
616 		return (KERN_PROTECTION_FAILURE);
617 	}
618 
619 	/*
620 	 * "enter_prot" is the protection we want to enter the page in at.
621 	 * for certain pages (e.g. copy-on-write pages) this protection can
622 	 * be more strict than ufi.entry->protection.  "wired" means either
623 	 * the entry is wired or we are fault-wiring the pg.
624 	 */
625 
626 	enter_prot = ufi.entry->protection;
627 	wired = (ufi.entry->wired_count != 0) || (fault_type == VM_FAULT_WIRE);
628 	if (wired)
629 		access_type = enter_prot; /* full access for wired */
630 
631 	/*
632 	 * handle "needs_copy" case.   if we need to copy the amap we will
633 	 * have to drop our readlock and relock it with a write lock.  (we
634 	 * need a write lock to change anything in a map entry [e.g.
635 	 * needs_copy]).
636 	 */
637 
638 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
639 		if ((access_type & VM_PROT_WRITE) ||
640 		    (ufi.entry->object.uvm_obj == NULL)) {
641 			/* need to clear */
642 			UVMHIST_LOG(maphist,
643 			    "  need to clear needs_copy and refault",0,0,0,0);
644 			uvmfault_unlockmaps(&ufi, FALSE);
645 			uvmfault_amapcopy(&ufi);
646 			uvmexp.fltamcopy++;
647 			goto ReFault;
648 
649 		} else {
650 
651 			/*
652 			 * ensure that we pmap_enter page R/O since
653 			 * needs_copy is still true
654 			 */
655 			enter_prot = enter_prot & ~VM_PROT_WRITE;
656 
657 		}
658 	}
659 
660 	/*
661 	 * identify the players
662 	 */
663 
664 	amap = ufi.entry->aref.ar_amap;	/* top layer */
665 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
666 
667 	/*
668 	 * check for a case 0 fault.  if nothing backing the entry then
669 	 * error now.
670 	 */
671 
672 	if (amap == NULL && uobj == NULL) {
673 		uvmfault_unlockmaps(&ufi, FALSE);
674 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
675 		return (KERN_INVALID_ADDRESS);
676 	}
677 
678 	/*
679 	 * establish range of interest based on advice from mapper
680 	 * and then clip to fit map entry.   note that we only want
681 	 * to do this the first time through the fault.   if we
682 	 * ReFault we will disable this by setting "narrow" to true.
683 	 */
684 
685 	if (narrow == FALSE) {
686 
687 		/* wide fault (!narrow) */
688 #ifdef DIAGNOSTIC
689 		if (uvmadvice[ufi.entry->advice].advice != ufi.entry->advice)
690 			panic("fault: advice mismatch!");
691 #endif
692 		nback = min(uvmadvice[ufi.entry->advice].nback,
693 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
694 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
695 		nforw = min(uvmadvice[ufi.entry->advice].nforw,
696 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
697 			     PAGE_SHIFT) - 1);
698 		/*
699 		 * note: "-1" because we don't want to count the
700 		 * faulting page as forw
701 		 */
702 		npages = nback + nforw + 1;
703 		centeridx = nback;
704 
705 		narrow = FALSE;	/* ensure only once per-fault */
706 
707 	} else {
708 
709 		/* narrow fault! */
710 		nback = nforw = 0;
711 		startva = ufi.orig_rvaddr;
712 		npages = 1;
713 		centeridx = 0;
714 
715 	}
716 
717 	/* locked: maps(read) */
718 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
719 		    narrow, nback, nforw, startva);
720 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
721 		    amap, uobj, 0);
722 
723 	/*
724 	 * if we've got an amap, lock it and extract current anons.
725 	 */
726 
727 	if (amap) {
728 		simple_lock(&amap->am_l);
729 		anons = anons_store;
730 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
731 		    anons, npages);
732 	} else {
733 		anons = NULL;	/* to be safe */
734 	}
735 
736 	/* locked: maps(read), amap(if there) */
737 
738 	/*
739 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
740 	 * now and then forget about them (for the rest of the fault).
741 	 */
742 
743 	if (ufi.entry->advice == MADV_SEQUENTIAL) {
744 
745 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
746 		    0,0,0,0);
747 		/* flush back-page anons? */
748 		if (amap)
749 			uvmfault_anonflush(anons, nback);
750 
751 		/* flush object? */
752 		if (uobj) {
753 			objaddr =
754 			    (startva - ufi.entry->start) + ufi.entry->offset;
755 			simple_lock(&uobj->vmobjlock);
756 			(void) uobj->pgops->pgo_flush(uobj, objaddr, objaddr +
757 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
758 			simple_unlock(&uobj->vmobjlock);
759 		}
760 
761 		/* now forget about the backpages */
762 		if (amap)
763 			anons += nback;
764 		startva = startva + (nback << PAGE_SHIFT);
765 		npages -= nback;
766 		nback = centeridx = 0;
767 	}
768 
769 	/* locked: maps(read), amap(if there) */
770 
771 	/*
772 	 * map in the backpages and frontpages we found in the amap in hopes
773 	 * of preventing future faults.    we also init the pages[] array as
774 	 * we go.
775 	 */
776 
777 	currva = startva;
778 	shadowed = FALSE;
779 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
780 
781 		/*
782 		 * dont play with VAs that are already mapped
783 		 * except for center)
784 		 * XXX: return value of pmap_extract disallows PA 0
785 		 */
786 		if (lcv != centeridx) {
787 			pa = pmap_extract(ufi.orig_map->pmap, currva);
788 			if (pa != NULL) {
789 				pages[lcv] = PGO_DONTCARE;
790 				continue;
791 			}
792 		}
793 
794 		/*
795 		 * unmapped or center page.   check if any anon at this level.
796 		 */
797 		if (amap == NULL || anons[lcv] == NULL) {
798 			pages[lcv] = NULL;
799 			continue;
800 		}
801 
802 		/*
803 		 * check for present page and map if possible.   re-activate it.
804 		 */
805 
806 		pages[lcv] = PGO_DONTCARE;
807 		if (lcv == centeridx) {		/* save center for later! */
808 			shadowed = TRUE;
809 			continue;
810 		}
811 		anon = anons[lcv];
812 		simple_lock(&anon->an_lock);
813 		/* ignore loaned pages */
814 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
815 			(anon->u.an_page->flags & (PG_RELEASED|PG_BUSY)) == 0) {
816 			uvm_lock_pageq();
817 			uvm_pageactivate(anon->u.an_page);	/* reactivate */
818 			uvm_unlock_pageq();
819 			UVMHIST_LOG(maphist,
820 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
821 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
822 			uvmexp.fltnamap++;
823 			pmap_enter(ufi.orig_map->pmap, currva,
824 			    VM_PAGE_TO_PHYS(anon->u.an_page),
825 			    (anon->an_ref > 1) ? VM_PROT_READ : enter_prot,
826 			    (ufi.entry->wired_count != 0));
827 		}
828 		simple_unlock(&anon->an_lock);
829 	}
830 
831 	/* locked: maps(read), amap(if there) */
832 	/* (shadowed == TRUE) if there is an anon at the faulting address */
833 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
834 	    (uobj && shadowed == FALSE),0,0);
835 
836 	/*
837 	 * note that if we are really short of RAM we could sleep in the above
838 	 * call to pmap_enter with everything locked.   bad?
839 	 * XXXCDC: this is fixed in PMAP_NEW (no sleep alloc's in pmap)
840 	 */
841 
842 	/*
843 	 * if the desired page is not shadowed by the amap and we have a
844 	 * backing object, then we check to see if the backing object would
845 	 * prefer to handle the fault itself (rather than letting us do it
846 	 * with the usual pgo_get hook).  the backing object signals this by
847 	 * providing a pgo_fault routine.
848 	 */
849 
850 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
851 
852 		simple_lock(&uobj->vmobjlock);
853 
854 		/* locked: maps(read), amap (if there), uobj */
855 		result = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
856 				    centeridx, fault_type, access_type,
857 				    PGO_LOCKED);
858 		/* locked: nothing, pgo_fault has unlocked everything */
859 
860 		if (result == VM_PAGER_OK)
861 			return (KERN_SUCCESS);	/* pgo_fault did pmap enter */
862 		else if (result == VM_PAGER_REFAULT)
863 			goto ReFault;		/* try again! */
864 		else
865 			return (KERN_PROTECTION_FAILURE);
866 	}
867 
868 	/*
869 	 * now, if the desired page is not shadowed by the amap and we have
870 	 * a backing object that does not have a special fault routine, then
871 	 * we ask (with pgo_get) the object for resident pages that we care
872 	 * about and attempt to map them in.  we do not let pgo_get block
873 	 * (PGO_LOCKED).
874 	 *
875 	 * ("get" has the option of doing a pmap_enter for us)
876 	 */
877 
878 	if (uobj && shadowed == FALSE) {
879 		simple_lock(&uobj->vmobjlock);
880 
881 		/* locked (!shadowed): maps(read), amap (if there), uobj */
882 		/*
883 		 * the following call to pgo_get does _not_ change locking state
884 		 */
885 
886 		uvmexp.fltlget++;
887 		gotpages = npages;
888 		result = uobj->pgops->pgo_get(uobj, ufi.entry->offset +
889 				(startva - ufi.entry->start),
890 				pages, &gotpages, centeridx,
891 				UVM_ET_ISCOPYONWRITE(ufi.entry) ?
892 				VM_PROT_READ : access_type,
893 				ufi.entry->advice, PGO_LOCKED);
894 
895 		/*
896 		 * check for pages to map, if we got any
897 		 */
898 
899 		uobjpage = NULL;
900 
901 		if (gotpages) {
902 			currva = startva;
903 			for (lcv = 0 ; lcv < npages ;
904 			    lcv++, currva += PAGE_SIZE) {
905 
906 				if (pages[lcv] == NULL ||
907 				    pages[lcv] == PGO_DONTCARE)
908 					continue;
909 
910 #ifdef DIAGNOSTIC
911 					/*
912 					 * pager sanity check: pgo_get with
913 					 * PGO_LOCKED should never return a
914 					 * released page to us.
915 					 */
916 					if (pages[lcv]->flags & PG_RELEASED)
917 		panic("uvm_fault: pgo_get PGO_LOCKED gave us a RELEASED page");
918 #endif
919 
920 					/*
921 					 * if center page is resident and not
922 					 * PG_BUSY|PG_RELEASED then pgo_get
923 					 * made it PG_BUSY for us and gave
924 					 * us a handle to it.   remember this
925 					 * page as "uobjpage." (for later use).
926 					 */
927 
928 					if (lcv == centeridx) {
929 						uobjpage = pages[lcv];
930 	UVMHIST_LOG(maphist, "  got uobjpage (0x%x) with locked get",
931 					    uobjpage, 0,0,0);
932 						continue;
933 				}
934 
935 				/*
936 				 * note: calling pgo_get with locked data
937 				 * structures returns us pages which are
938 				 * neither busy nor released, so we don't
939 				 * need to check for this.   we can just
940 				 * directly enter the page (after moving it
941 				 * to the head of the active queue [useful?]).
942 				 */
943 
944 				uvm_lock_pageq();
945 				uvm_pageactivate(pages[lcv]);	/* reactivate */
946 				uvm_unlock_pageq();
947 				UVMHIST_LOG(maphist,
948 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
949 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
950 				uvmexp.fltnomap++;
951 				pmap_enter(ufi.orig_map->pmap, currva,
952 				    VM_PAGE_TO_PHYS(pages[lcv]),
953 				    UVM_ET_ISCOPYONWRITE(ufi.entry) ?
954 				    VM_PROT_READ : enter_prot, wired);
955 
956 				/*
957 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
958 				 * because we've held the lock the whole time
959 				 * we've had the handle.
960 				 */
961 				pages[lcv]->flags &= ~(PG_BUSY); /* un-busy! */
962 				UVM_PAGE_OWN(pages[lcv], NULL);
963 
964 				/* done! */
965 			}	/* for "lcv" loop */
966 		}   /* "gotpages" != 0 */
967 
968 		/* note: object still _locked_ */
969 	} else {
970 
971 		uobjpage = NULL;
972 
973 	}
974 
975 	/* locked (shadowed): maps(read), amap */
976 	/* locked (!shadowed): maps(read), amap(if there),
977 		 uobj(if !null), uobjpage(if !null) */
978 
979 	/*
980 	 * note that at this point we are done with any front or back pages.
981 	 * we are now going to focus on the center page (i.e. the one we've
982 	 * faulted on).  if we have faulted on the top (anon) layer
983 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
984 	 * not touched it yet).  if we have faulted on the bottom (uobj)
985 	 * layer [i.e. case 2] and the page was both present and available,
986 	 * then we've got a pointer to it as "uobjpage" and we've already
987 	 * made it BUSY.
988 	 */
989 
990 	/*
991 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
992 	 */
993 
994 	/*
995 	 * redirect case 2: if we are not shadowed, go to case 2.
996 	 */
997 
998 	if (shadowed == FALSE)
999 		goto Case2;
1000 
1001 	/* locked: maps(read), amap */
1002 
1003 	/*
1004 	 * handle case 1: fault on an anon in our amap
1005 	 */
1006 
1007 	anon = anons[centeridx];
1008 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1009 	simple_lock(&anon->an_lock);
1010 
1011 	/* locked: maps(read), amap, anon */
1012 
1013 	/*
1014 	 * no matter if we have case 1A or case 1B we are going to need to
1015 	 * have the anon's memory resident.   ensure that now.
1016 	 */
1017 
1018 	/*
1019 	 * let uvmfault_anonget do the dirty work.   if it fails (!OK) it will
1020 	 * unlock for us.   if it is OK, locks are still valid and locked.
1021 	 * also, if it is OK, then the anon's page is on the queues.
1022 	 * if the page is on loan from a uvm_object, then anonget will
1023 	 * lock that object for us if it does not fail.
1024 	 */
1025 
1026 	result = uvmfault_anonget(&ufi, amap, anon);
1027 
1028 	if (result == VM_PAGER_REFAULT)
1029 		goto ReFault;
1030 
1031 	if (result == VM_PAGER_AGAIN) {
1032 		tsleep((caddr_t)&lbolt, PVM, "fltagain1", 0);
1033 		goto ReFault;
1034 	}
1035 
1036 	if (result != VM_PAGER_OK)
1037 		return (KERN_PROTECTION_FAILURE);		/* XXX??? */
1038 
1039 	/*
1040 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1041 	 */
1042 
1043 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1044 
1045 	/* locked: maps(read), amap, anon, uobj(if one) */
1046 
1047 	/*
1048 	 * special handling for loaned pages
1049 	 */
1050 	if (anon->u.an_page->loan_count) {
1051 
1052 		if ((access_type & VM_PROT_WRITE) == 0) {
1053 
1054 			/*
1055 			 * for read faults on loaned pages we just cap the
1056 			 * protection at read-only.
1057 			 */
1058 
1059 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1060 
1061 		} else {
1062 			/*
1063 			 * note that we can't allow writes into a loaned page!
1064 			 *
1065 			 * if we have a write fault on a loaned page in an
1066 			 * anon then we need to look at the anon's ref count.
1067 			 * if it is greater than one then we are going to do
1068 			 * a normal copy-on-write fault into a new anon (this
1069 			 * is not a problem).  however, if the reference count
1070 			 * is one (a case where we would normally allow a
1071 			 * write directly to the page) then we need to kill
1072 			 * the loan before we continue.
1073 			 */
1074 
1075 			/* >1 case is already ok */
1076 			if (anon->an_ref == 1) {
1077 
1078 				/* get new un-owned replacement page */
1079 				pg = uvm_pagealloc(NULL, 0, NULL);
1080 				if (pg == NULL) {
1081 					uvmfault_unlockall(&ufi, amap, uobj,
1082 					    anon);
1083 					uvm_wait("flt_noram2");
1084 					goto ReFault;
1085 				}
1086 
1087 				/*
1088 				 * copy data, kill loan, and drop uobj lock
1089 				 * (if any)
1090 				 */
1091 				/* copy old -> new */
1092 				uvm_pagecopy(anon->u.an_page, pg);
1093 
1094 				/* force reload */
1095 				pmap_page_protect(PMAP_PGARG(anon->u.an_page),
1096 				    VM_PROT_NONE);
1097 				uvm_lock_pageq();	  /* KILL loan */
1098 				if (uobj)
1099 					/* if we were loaning */
1100 					anon->u.an_page->loan_count--;
1101 				anon->u.an_page->uanon = NULL;
1102 				/* in case we owned */
1103 				anon->u.an_page->pqflags &= ~PQ_ANON;
1104 				uvm_unlock_pageq();
1105 				if (uobj) {
1106 					simple_unlock(&uobj->vmobjlock);
1107 					uobj = NULL;
1108 				}
1109 
1110 				/* install new page in anon */
1111 				anon->u.an_page = pg;
1112 				pg->uanon = anon;
1113 				pg->pqflags |= PQ_ANON;
1114 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1115 				UVM_PAGE_OWN(pg, NULL);
1116 
1117 				/* done! */
1118 			}     /* ref == 1 */
1119 		}       /* write fault */
1120 	}         /* loan count */
1121 
1122 	/*
1123 	 * if we are case 1B then we will need to allocate a new blank
1124 	 * anon to transfer the data into.   note that we have a lock
1125 	 * on anon, so no one can busy or release the page until we are done.
1126 	 * also note that the ref count can't drop to zero here because
1127 	 * it is > 1 and we are only dropping one ref.
1128 	 *
1129 	 * in the (hopefully very rare) case that we are out of RAM we
1130 	 * will unlock, wait for more RAM, and refault.
1131 	 *
1132 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1133 	 */
1134 
1135 	if ((access_type & VM_PROT_WRITE) != 0 && anon->an_ref > 1) {
1136 
1137 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1138 		uvmexp.flt_acow++;
1139 		oanon = anon;		/* oanon = old, locked anon */
1140 		anon = uvm_analloc();
1141 		if (anon)
1142 			pg = uvm_pagealloc(NULL, 0, anon);
1143 #ifdef __GNUC__
1144 		else
1145 			pg = NULL; /* XXX: gcc */
1146 #endif
1147 
1148 		/* check for out of RAM */
1149 		if (anon == NULL || pg == NULL) {
1150 			if (anon)
1151 				uvm_anfree(anon);
1152 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1153 			if (anon == NULL) {
1154 				UVMHIST_LOG(maphist,
1155 				    "<- failed.  out of VM",0,0,0,0);
1156 				uvmexp.fltnoanon++;
1157 				/* XXX: OUT OF VM, ??? */
1158 				return (KERN_RESOURCE_SHORTAGE);
1159 			}
1160 			uvmexp.fltnoram++;
1161 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1162 			goto ReFault;
1163 		}
1164 
1165 		/* got all resources, replace anon with nanon */
1166 
1167 		uvm_pagecopy(oanon->u.an_page, pg);	/* pg now !PG_CLEAN */
1168 		pg->flags &= ~(PG_BUSY|PG_FAKE);	/* un-busy! new page */
1169 		UVM_PAGE_OWN(pg, NULL);
1170 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1171 		    anon, 1);
1172 
1173 		/* deref: can not drop to zero here by defn! */
1174 		oanon->an_ref--;
1175 
1176 		/*
1177 		 * note: oanon still locked.   anon is _not_ locked, but we
1178 		 * have the sole references to in from amap which _is_ locked.
1179 		 * thus, no one can get at it until we are done with it.
1180 		 */
1181 
1182 	} else {
1183 
1184 		uvmexp.flt_anon++;
1185 		oanon = anon;		/* old, locked anon is same as anon */
1186 		pg = anon->u.an_page;
1187 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1188 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1189 
1190 	}
1191 
1192 	/* locked: maps(read), amap, anon */
1193 
1194 	/*
1195 	 * now map the page in ...
1196 	 * XXX: old fault unlocks object before pmap_enter.  this seems
1197 	 * suspect since some other thread could blast the page out from
1198 	 * under us between the unlock and the pmap_enter.
1199 	 */
1200 
1201 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1202 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1203 	pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1204 	    enter_prot, wired);
1205 
1206 	/*
1207 	 * ... and update the page queues.
1208 	 */
1209 
1210 	uvm_lock_pageq();
1211 
1212 	if (fault_type == VM_FAULT_WIRE) {
1213 		uvm_pagewire(pg);
1214 	} else {
1215 		/* activate it */
1216 		uvm_pageactivate(pg);
1217 
1218 	}
1219 
1220 	uvm_unlock_pageq();
1221 
1222 	/*
1223 	 * done case 1!  finish up by unlocking everything and returning success
1224 	 */
1225 
1226 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1227 	return (KERN_SUCCESS);
1228 
1229 
1230 Case2:
1231 	/*
1232 	 * handle case 2: faulting on backing object or zero fill
1233 	 */
1234 
1235 	/*
1236 	 * locked:
1237 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1238 	 */
1239 
1240 	/*
1241 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1242 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1243 	 * have a backing object, check and see if we are going to promote
1244 	 * the data up to an anon during the fault.
1245 	 */
1246 
1247 	if (uobj == NULL) {
1248 		uobjpage = PGO_DONTCARE;
1249 		promote = TRUE;		/* always need anon here */
1250 	} else {
1251 		/* assert(uobjpage != PGO_DONTCARE) */
1252 		promote = (access_type & VM_PROT_WRITE) &&
1253 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
1254 	}
1255 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1256 	promote, (uobj == NULL), 0,0);
1257 
1258 	/*
1259 	 * if uobjpage is not null then we do not need to do I/O to get the
1260 	 * uobjpage.
1261 	 *
1262 	 * if uobjpage is null, then we need to unlock and ask the pager to
1263 	 * get the data for us.   once we have the data, we need to reverify
1264 	 * the state the world.   we are currently not holding any resources.
1265 	 */
1266 
1267 	if (uobjpage) {
1268 		/* update rusage counters */
1269 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1270 	} else {
1271 		/* update rusage counters */
1272 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1273 
1274 		/* locked: maps(read), amap(if there), uobj */
1275 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1276 		/* locked: uobj */
1277 
1278 		uvmexp.fltget++;
1279 		gotpages = 1;
1280 		result = uobj->pgops->pgo_get(uobj,
1281 		    (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset,
1282 		    &uobjpage, &gotpages, 0,
1283 		    UVM_ET_ISCOPYONWRITE(ufi.entry) ?
1284 			VM_PROT_READ : access_type,
1285 			ufi.entry->advice, 0);
1286 
1287 		/* locked: uobjpage(if result OK) */
1288 
1289 		/*
1290 		 * recover from I/O
1291 		 */
1292 
1293 		if (result != VM_PAGER_OK) {
1294 
1295 #ifdef DIAGNOSTIC
1296 			if (result == VM_PAGER_PEND)
1297 	panic("uvm_fault: pgo_get got PENDing on non-async I/O");
1298 #endif
1299 
1300 			if (result == VM_PAGER_AGAIN) {
1301 	UVMHIST_LOG(maphist, "  pgo_get says TRY AGAIN!",0,0,0,0);
1302 	tsleep((caddr_t)&lbolt, PVM, "fltagain2", 0);
1303 	goto ReFault;
1304 			}
1305 
1306 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1307 			    result, 0,0,0);
1308 			return (KERN_PROTECTION_FAILURE); /* XXX i/o error */
1309 		}
1310 
1311 		/* locked: uobjpage */
1312 
1313 		/*
1314 		 * re-verify the state of the world by first trying to relock
1315 		 * the maps.  always relock the object.
1316 		 */
1317 
1318 		locked = uvmfault_relock(&ufi);
1319 		if (locked && amap)
1320 			simple_lock(&amap->am_l);
1321 		simple_lock(&uobj->vmobjlock);
1322 
1323 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1324 		/* locked(!locked): uobj, uobjpage */
1325 
1326 		/*
1327 		 * verify that the page has not be released and re-verify
1328 		 * that amap slot is still free.   if there is a problem,
1329 		 * we unlock and clean up.
1330 		 */
1331 
1332 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1333 		    (locked && amap &&
1334 		    amap_lookup(&ufi.entry->aref,
1335 		      ufi.orig_rvaddr - ufi.entry->start))) {
1336 			if (locked)
1337 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1338 			locked = FALSE;
1339 		}
1340 
1341 		/*
1342 		 * didn't get the lock?   release the page and retry.
1343 		 */
1344 
1345 		if (locked == FALSE) {
1346 
1347 			UVMHIST_LOG(maphist,
1348 			    "  wasn't able to relock after fault: retry",
1349 			    0,0,0,0);
1350 			if (uobjpage->flags & PG_WANTED)
1351 				/* still holding object lock */
1352 				thread_wakeup(uobjpage);
1353 
1354 			if (uobjpage->flags & PG_RELEASED) {
1355 				uvmexp.fltpgrele++;
1356 #ifdef DIAGNOSTIC
1357 				if (uobj->pgops->pgo_releasepg == NULL)
1358 			panic("uvm_fault: object has no releasepg function");
1359 #endif
1360 				/* frees page */
1361 				if (uobj->pgops->pgo_releasepg(uobjpage,NULL))
1362 					/* unlock if still alive */
1363 					simple_unlock(&uobj->vmobjlock);
1364 				goto ReFault;
1365 			}
1366 
1367 			uvm_lock_pageq();
1368 			/* make sure it is in queues */
1369 			uvm_pageactivate(uobjpage);
1370 
1371 			uvm_unlock_pageq();
1372 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1373 			UVM_PAGE_OWN(uobjpage, NULL);
1374 			simple_unlock(&uobj->vmobjlock);
1375 			goto ReFault;
1376 
1377 		}
1378 
1379 		/*
1380 		 * we have the data in uobjpage which is PG_BUSY and
1381 		 * !PG_RELEASED.  we are holding object lock (so the page
1382 		 * can't be released on us).
1383 		 */
1384 
1385 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1386 
1387 	}
1388 
1389 	/*
1390 	 * locked:
1391 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1392 	 */
1393 
1394 	/*
1395 	 * notes:
1396 	 *  - at this point uobjpage can not be NULL
1397 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1398 	 *  for it above)
1399 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1400 	 */
1401 
1402 	if (promote == FALSE) {
1403 
1404 		/*
1405 		 * we are not promoting.   if the mapping is COW ensure that we
1406 		 * don't give more access than we should (e.g. when doing a read
1407 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1408 		 * R/O even though the entry protection could be R/W).
1409 		 *
1410 		 * set "pg" to the page we want to map in (uobjpage, usually)
1411 		 */
1412 
1413 		uvmexp.flt_obj++;
1414 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1415 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1416 		pg = uobjpage;		/* map in the actual object */
1417 
1418 		/* assert(uobjpage != PGO_DONTCARE) */
1419 
1420 		/*
1421 		 * we are faulting directly on the page.   be careful
1422 		 * about writing to loaned pages...
1423 		 */
1424 		if (uobjpage->loan_count) {
1425 
1426 			if ((access_type & VM_PROT_WRITE) == 0) {
1427 				/* read fault: cap the protection at readonly */
1428 				/* cap! */
1429 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1430 			} else {
1431 				/* write fault: must break the loan here */
1432 
1433 				/* alloc new un-owned page */
1434 				pg = uvm_pagealloc(NULL, 0, NULL);
1435 
1436 				if (pg == NULL) {
1437 					/*
1438 					 * drop ownership of page, it can't
1439 					 * be released
1440 					 * */
1441 					if (uobjpage->flags & PG_WANTED)
1442 						thread_wakeup(uobjpage);
1443 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1444 					UVM_PAGE_OWN(uobjpage, NULL);
1445 
1446 					uvm_lock_pageq();
1447 					/* activate: we will need it later */
1448 					uvm_pageactivate(uobjpage);
1449 
1450 					uvm_unlock_pageq();
1451 					uvmfault_unlockall(&ufi, amap, uobj,
1452 					  NULL);
1453 					UVMHIST_LOG(maphist,
1454 					  "  out of RAM breaking loan, waiting",					  0,0,0,0);
1455 					uvmexp.fltnoram++;
1456 					uvm_wait("flt_noram4");
1457 					goto ReFault;
1458 				}
1459 
1460 				/*
1461 				 * copy the data from the old page to the new
1462 				 * one and clear the fake/clean flags on the
1463 				 * new page (keep it busy).  force a reload
1464 				 * of the old page by clearing it from all
1465 				 * pmaps.  then lock the page queues to
1466 				 * rename the pages.
1467 				 */
1468 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1469 				pg->flags &= ~(PG_FAKE|PG_CLEAN);
1470 				pmap_page_protect(PMAP_PGARG(uobjpage),
1471 				    VM_PROT_NONE);
1472 				if (uobjpage->flags & PG_WANTED)
1473 					thread_wakeup(uobjpage);
1474 				/* uobj still locked */
1475 				uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
1476 				UVM_PAGE_OWN(uobjpage, NULL);
1477 
1478 				uvm_lock_pageq();
1479 				offset = uobjpage->offset;
1480 				/* remove old page */
1481 				uvm_pagerealloc(uobjpage, NULL, 0);
1482 
1483 				/*
1484 				 * at this point we have absolutely no
1485 				 * control over uobjpage
1486 				 */
1487 				/* install new page */
1488 				uvm_pagerealloc(pg, uobj, offset);
1489 				uvm_unlock_pageq();
1490 
1491 				/*
1492 				 * done!  loan is broken and "pg" is
1493 				 * PG_BUSY.   it can now replace uobjpage.
1494 				 */
1495 
1496 				uobjpage = pg;
1497 
1498 			}		/* write fault case */
1499 		}		/* if loan_count */
1500 
1501 	} else {
1502 
1503 		/*
1504 		 * if we are going to promote the data to an anon we
1505 		 * allocate a blank anon here and plug it into our amap.
1506 		 */
1507 #if DIAGNOSTIC
1508 		if (amap == NULL)
1509 			panic("uvm_fault: want to promote data, but no anon");
1510 #endif
1511 
1512 		anon = uvm_analloc();
1513 		if (anon)
1514 			pg = uvm_pagealloc(NULL, 0, anon); /* BUSY+CLEAN+FAKE */
1515 #ifdef __GNUC__
1516 		else
1517 			pg = NULL; /* XXX: gcc */
1518 #endif
1519 
1520 		/*
1521 		 * out of memory resources?
1522 		 */
1523 		if (anon == NULL || pg == NULL) {
1524 
1525 			/*
1526 			 * arg!  must unbusy our page and fail or sleep.
1527 			 */
1528 			if (uobjpage != PGO_DONTCARE) {
1529 				if (uobjpage->flags & PG_WANTED)
1530 					/* still holding object lock */
1531 					thread_wakeup(uobjpage);
1532 
1533 				uvm_lock_pageq();
1534 				/* make sure it is in queues */
1535 				uvm_pageactivate(uobjpage);
1536 				uvm_unlock_pageq();
1537 				/* un-busy! (still locked) */
1538 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1539 				UVM_PAGE_OWN(uobjpage, NULL);
1540 			}
1541 
1542 			/* unlock and fail ... */
1543 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1544 			if (anon == NULL) {
1545 				UVMHIST_LOG(maphist, "  promote: out of VM",
1546 				    0,0,0,0);
1547 				uvmexp.fltnoanon++;
1548 				/* XXX: out of VM */
1549 				return (KERN_RESOURCE_SHORTAGE);
1550 			}
1551 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1552 			    0,0,0,0);
1553 			uvm_anfree(anon);
1554 			uvmexp.fltnoram++;
1555 			uvm_wait("flt_noram5");
1556 			goto ReFault;
1557 		}
1558 
1559 		/*
1560 		 * fill in the data
1561 		 */
1562 
1563 		if (uobjpage != PGO_DONTCARE) {
1564 			uvmexp.flt_prcopy++;
1565 			/* copy page [pg now dirty] */
1566 			uvm_pagecopy(uobjpage, pg);
1567 
1568 			/*
1569 			 * promote to shared amap?  make sure all sharing
1570 			 * procs see it
1571 			 */
1572 			if ((amap->am_flags & AMAP_SHARED) != 0) {
1573 				pmap_page_protect(PMAP_PGARG(uobjpage),
1574 				    VM_PROT_NONE);
1575 			}
1576 
1577 			/*
1578 			 * dispose of uobjpage.  it can't be PG_RELEASED
1579 			 * since we still hold the object lock.   drop
1580 			 * handle to uobj as well.
1581 			 */
1582 
1583 			if (uobjpage->flags & PG_WANTED)
1584 				/* still have the obj lock */
1585 				thread_wakeup(uobjpage);
1586 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1587 			UVM_PAGE_OWN(uobjpage, NULL);
1588 			uvm_lock_pageq();
1589 			uvm_pageactivate(uobjpage);	/* put it back */
1590 			uvm_unlock_pageq();
1591 			simple_unlock(&uobj->vmobjlock);
1592 			uobj = NULL;
1593 			UVMHIST_LOG(maphist,
1594 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1595 			    uobjpage, anon, pg, 0);
1596 
1597 		} else {
1598 			uvmexp.flt_przero++;
1599 			uvm_pagezero(pg);	/* zero page [pg now dirty] */
1600 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1601 			    anon, pg, 0, 0);
1602 		}
1603 
1604 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1605 		    anon, 0);
1606 
1607 	}
1608 
1609 	/*
1610 	 * locked:
1611 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1612 	 *
1613 	 * note: pg is either the uobjpage or the new page in the new anon
1614 	 */
1615 
1616 	/*
1617 	 * all resources are present.   we can now map it in and free our
1618 	 * resources.
1619 	 */
1620 
1621 	UVMHIST_LOG(maphist,
1622 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1623 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1624 	pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1625 	    enter_prot, wired);
1626 
1627 	uvm_lock_pageq();
1628 
1629 	if (fault_type == VM_FAULT_WIRE) {
1630 		uvm_pagewire(pg);
1631 	} else {
1632 
1633 		/* activate it */
1634 		uvm_pageactivate(pg);
1635 
1636 	}
1637 
1638 	uvm_unlock_pageq();
1639 
1640 	if (pg->flags & PG_WANTED)
1641 		thread_wakeup(pg);		/* lock still held */
1642 
1643 	/*
1644 	 * note that pg can't be PG_RELEASED since we did not drop the object
1645 	 * lock since the last time we checked.
1646 	 */
1647 
1648 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1649 	UVM_PAGE_OWN(pg, NULL);
1650 	uvmfault_unlockall(&ufi, amap, uobj, NULL);
1651 
1652 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1653 	return (KERN_SUCCESS);
1654 }
1655 
1656 
1657 /*
1658  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1659  *
1660  * => map should be locked by caller?   If so how can we call
1661  *	uvm_fault?   WRONG.
1662  * => XXXCDC: locking here is all screwed up!!!  start with
1663  *	uvm_map_pageable and fix it.
1664  */
1665 
1666 int
1667 uvm_fault_wire(map, start, end)
1668 	vm_map_t map;
1669 	vaddr_t start, end;
1670 {
1671 	vaddr_t va;
1672 	pmap_t  pmap;
1673 	int rv;
1674 
1675 	pmap = vm_map_pmap(map);
1676 
1677 	/*
1678 	 * call pmap pageable: this tells the pmap layer to lock down these
1679 	 * page tables.
1680 	 */
1681 
1682 	pmap_pageable(pmap, start, end, FALSE);
1683 
1684 	/*
1685 	 * now fault it in page at a time.   if the fault fails then we have
1686 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1687 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1688 	 */
1689 
1690 	for (va = start ; va < end ; va += PAGE_SIZE) {
1691 		rv = uvm_fault(map, va, VM_FAULT_WIRE, VM_PROT_NONE);
1692 		if (rv) {
1693 			if (va != start) {
1694 				uvm_fault_unwire(map->pmap, start, va);
1695 			}
1696 			return (rv);
1697 		}
1698 	}
1699 
1700 	return (KERN_SUCCESS);
1701 }
1702 
1703 /*
1704  * uvm_fault_unwire(): unwire range of virtual space.
1705  *
1706  * => caller holds reference to pmap (via its map)
1707  */
1708 
1709 void
1710 uvm_fault_unwire(pmap, start, end)
1711 	struct pmap *pmap;
1712 	vaddr_t start, end;
1713 {
1714 	vaddr_t va;
1715 	paddr_t pa;
1716 	struct vm_page *pg;
1717 
1718 	/*
1719 	 * we assume that the area we are unwiring has actually been wired
1720 	 * in the first place.   this means that we should be able to extract
1721 	 * the PAs from the pmap.   we also lock out the page daemon so that
1722 	 * we can call uvm_pageunwire.
1723 	 */
1724 
1725 	uvm_lock_pageq();
1726 
1727 	for (va = start; va < end ; va += PAGE_SIZE) {
1728 		pa = pmap_extract(pmap, va);
1729 
1730 		/* XXX: assumes PA 0 cannot be in map */
1731 		if (pa == (paddr_t) 0) {
1732 			panic("uvm_fault_unwire: unwiring non-wired memory");
1733 		}
1734 		pmap_change_wiring(pmap, va, FALSE);  /* tell the pmap */
1735 		pg = PHYS_TO_VM_PAGE(pa);
1736 		if (pg)
1737 			uvm_pageunwire(pg);
1738 	}
1739 
1740 	uvm_unlock_pageq();
1741 
1742 	/*
1743 	 * now we call pmap_pageable to let the pmap know that the page tables
1744 	 * in this space no longer need to be wired.
1745 	 */
1746 
1747 	pmap_pageable(pmap, start, end, TRUE);
1748 
1749 }
1750