xref: /openbsd-src/sys/uvm/uvm_fault.c (revision 58fbf5d6aa35e3d66f2c32c61d2f38824a990e85)
1 /*	$OpenBSD: uvm_fault.c,v 1.115 2021/02/16 09:10:17 mpi Exp $	*/
2 /*	$NetBSD: uvm_fault.c,v 1.51 2000/08/06 00:22:53 thorpej Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  *
28  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
29  */
30 
31 /*
32  * uvm_fault.c: fault handler
33  */
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/percpu.h>
39 #include <sys/proc.h>
40 #include <sys/malloc.h>
41 #include <sys/mman.h>
42 #include <sys/tracepoint.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)
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 struct uvm_advice {
147 	int nback;
148 	int nforw;
149 };
150 
151 /*
152  * page range array: set up in uvmfault_init().
153  */
154 static struct uvm_advice uvmadvice[MADV_MASK + 1];
155 
156 #define UVM_MAXRANGE 16	/* must be max() of nback+nforw+1 */
157 
158 /*
159  * private prototypes
160  */
161 static void uvmfault_amapcopy(struct uvm_faultinfo *);
162 static inline void uvmfault_anonflush(struct vm_anon **, int);
163 void	uvmfault_unlockmaps(struct uvm_faultinfo *, boolean_t);
164 void	uvmfault_update_stats(struct uvm_faultinfo *);
165 
166 /*
167  * inline functions
168  */
169 /*
170  * uvmfault_anonflush: try and deactivate pages in specified anons
171  *
172  * => does not have to deactivate page if it is busy
173  */
174 static inline void
175 uvmfault_anonflush(struct vm_anon **anons, int n)
176 {
177 	int lcv;
178 	struct vm_page *pg;
179 
180 	for (lcv = 0; lcv < n; lcv++) {
181 		if (anons[lcv] == NULL)
182 			continue;
183 		KASSERT(rw_lock_held(anons[lcv]->an_lock));
184 		pg = anons[lcv]->an_page;
185 		if (pg && (pg->pg_flags & PG_BUSY) == 0) {
186 			uvm_lock_pageq();
187 			if (pg->wire_count == 0) {
188 				pmap_page_protect(pg, PROT_NONE);
189 				uvm_pagedeactivate(pg);
190 			}
191 			uvm_unlock_pageq();
192 		}
193 	}
194 }
195 
196 /*
197  * normal functions
198  */
199 /*
200  * uvmfault_init: compute proper values for the uvmadvice[] array.
201  */
202 void
203 uvmfault_init(void)
204 {
205 	int npages;
206 
207 	npages = atop(16384);
208 	if (npages > 0) {
209 		KASSERT(npages <= UVM_MAXRANGE / 2);
210 		uvmadvice[MADV_NORMAL].nforw = npages;
211 		uvmadvice[MADV_NORMAL].nback = npages - 1;
212 	}
213 
214 	npages = atop(32768);
215 	if (npages > 0) {
216 		KASSERT(npages <= UVM_MAXRANGE / 2);
217 		uvmadvice[MADV_SEQUENTIAL].nforw = npages - 1;
218 		uvmadvice[MADV_SEQUENTIAL].nback = npages;
219 	}
220 }
221 
222 /*
223  * uvmfault_amapcopy: clear "needs_copy" in a map.
224  *
225  * => called with VM data structures unlocked (usually, see below)
226  * => we get a write lock on the maps and clear needs_copy for a VA
227  * => if we are out of RAM we sleep (waiting for more)
228  */
229 static void
230 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
231 {
232 	for (;;) {
233 		/* no mapping?  give up. */
234 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
235 			return;
236 
237 		/* copy if needed. */
238 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
239 			amap_copy(ufi->map, ufi->entry, M_NOWAIT,
240 				UVM_ET_ISSTACK(ufi->entry) ? FALSE : TRUE,
241 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
242 
243 		/* didn't work?  must be out of RAM.  sleep. */
244 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
245 			uvmfault_unlockmaps(ufi, TRUE);
246 			uvm_wait("fltamapcopy");
247 			continue;
248 		}
249 
250 		/* got it! */
251 		uvmfault_unlockmaps(ufi, TRUE);
252 		return;
253 	}
254 	/*NOTREACHED*/
255 }
256 
257 /*
258  * uvmfault_anonget: get data in an anon into a non-busy, non-released
259  * page in that anon.
260  *
261  * => Map, amap and thus anon should be locked by caller.
262  * => If we fail, we unlock everything and error is returned.
263  * => If we are successful, return with everything still locked.
264  * => We do not move the page on the queues [gets moved later].  If we
265  *    allocate a new page [we_own], it gets put on the queues.  Either way,
266  *    the result is that the page is on the queues at return time
267  */
268 int
269 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
270     struct vm_anon *anon)
271 {
272 	struct vm_page *pg;
273 	int error;
274 
275 	KASSERT(rw_lock_held(anon->an_lock));
276 	KASSERT(anon->an_lock == amap->am_lock);
277 
278 	/* Increment the counters.*/
279 	counters_inc(uvmexp_counters, flt_anget);
280 	if (anon->an_page) {
281 		curproc->p_ru.ru_minflt++;
282 	} else {
283 		curproc->p_ru.ru_majflt++;
284 	}
285 	error = 0;
286 
287 	/*
288 	 * Loop until we get the anon data, or fail.
289 	 */
290 	for (;;) {
291 		boolean_t we_own, locked;
292 		/*
293 		 * Note: 'we_own' will become true if we set PG_BUSY on a page.
294 		 */
295 		we_own = FALSE;
296 		pg = anon->an_page;
297 
298 		/*
299 		 * Is page resident?  Make sure it is not busy/released.
300 		 */
301 		if (pg) {
302 			KASSERT(pg->pg_flags & PQ_ANON);
303 			KASSERT(pg->uanon == anon);
304 
305 			/*
306 			 * if the page is busy, we drop all the locks and
307 			 * try again.
308 			 */
309 			if ((pg->pg_flags & (PG_BUSY|PG_RELEASED)) == 0)
310 				return (VM_PAGER_OK);
311 			atomic_setbits_int(&pg->pg_flags, PG_WANTED);
312 			counters_inc(uvmexp_counters, flt_pgwait);
313 
314 			/*
315 			 * The last unlock must be an atomic unlock and wait
316 			 * on the owner of page.
317 			 */
318 			if (pg->uobject) {
319 				/* Owner of page is UVM object. */
320 				uvmfault_unlockall(ufi, amap, NULL);
321 				tsleep_nsec(pg, PVM, "anonget1", INFSLP);
322 			} else {
323 				/* Owner of page is anon. */
324 				uvmfault_unlockall(ufi, NULL, NULL);
325 				rwsleep_nsec(pg, anon->an_lock, PVM | PNORELOCK,
326 				    "anonget2", INFSLP);
327 			}
328 		} else {
329 			/*
330 			 * No page, therefore allocate one.
331 			 */
332 			pg = uvm_pagealloc(NULL, 0, anon, 0);
333 			if (pg == NULL) {
334 				/* Out of memory.  Wait a little. */
335 				uvmfault_unlockall(ufi, amap, NULL);
336 				counters_inc(uvmexp_counters, flt_noram);
337 				uvm_wait("flt_noram1");
338 			} else {
339 				/* PG_BUSY bit is set. */
340 				we_own = TRUE;
341 				uvmfault_unlockall(ufi, amap, NULL);
342 
343 				/*
344 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
345 				 * page into the uvm_swap_get function with
346 				 * all data structures unlocked.  note that
347 				 * it is ok to read an_swslot here because
348 				 * we hold PG_BUSY on the page.
349 				 */
350 				counters_inc(uvmexp_counters, pageins);
351 				error = uvm_swap_get(pg, anon->an_swslot,
352 				    PGO_SYNCIO);
353 
354 				/*
355 				 * We clean up after the I/O below in the
356 				 * 'we_own' case.
357 				 */
358 			}
359 		}
360 
361 		/*
362 		 * Re-lock the map and anon.
363 		 */
364 		locked = uvmfault_relock(ufi);
365 		if (locked || we_own) {
366 			rw_enter(anon->an_lock, RW_WRITE);
367 		}
368 
369 		/*
370 		 * If we own the page (i.e. we set PG_BUSY), then we need
371 		 * to clean up after the I/O.  There are three cases to
372 		 * consider:
373 		 *
374 		 * 1) Page was released during I/O: free anon and ReFault.
375 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
376 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
377 		 *    case (i.e. drop anon lock if not locked).
378 		 */
379 		if (we_own) {
380 			if (pg->pg_flags & PG_WANTED) {
381 				wakeup(pg);
382 			}
383 			/* un-busy! */
384 			atomic_clearbits_int(&pg->pg_flags,
385 			    PG_WANTED|PG_BUSY|PG_FAKE);
386 			UVM_PAGE_OWN(pg, NULL);
387 
388 			/*
389 			 * if we were RELEASED during I/O, then our anon is
390 			 * no longer part of an amap.   we need to free the
391 			 * anon and try again.
392 			 */
393 			if (pg->pg_flags & PG_RELEASED) {
394 				pmap_page_protect(pg, PROT_NONE);
395 				KASSERT(anon->an_ref == 0);
396 				if (locked)
397 					uvmfault_unlockall(ufi, NULL, NULL);
398 				uvm_anon_release(anon);	/* frees page for us */
399 				counters_inc(uvmexp_counters, flt_pgrele);
400 				return (VM_PAGER_REFAULT);	/* refault! */
401 			}
402 
403 			if (error != VM_PAGER_OK) {
404 				KASSERT(error != VM_PAGER_PEND);
405 
406 				/* remove page from anon */
407 				anon->an_page = NULL;
408 
409 				/*
410 				 * Remove the swap slot from the anon and
411 				 * mark the anon as having no real slot.
412 				 * Do not free the swap slot, thus preventing
413 				 * it from being used again.
414 				 */
415 				uvm_swap_markbad(anon->an_swslot, 1);
416 				anon->an_swslot = SWSLOT_BAD;
417 
418 				/*
419 				 * Note: page was never !PG_BUSY, so it
420 				 * cannot be mapped and thus no need to
421 				 * pmap_page_protect it...
422 				 */
423 				uvm_lock_pageq();
424 				uvm_pagefree(pg);
425 				uvm_unlock_pageq();
426 
427 				if (locked) {
428 					uvmfault_unlockall(ufi, NULL, NULL);
429 				}
430 				rw_exit(anon->an_lock);
431 				return (VM_PAGER_ERROR);
432 			}
433 
434 			/*
435 			 * must be OK, clear modify (already PG_CLEAN)
436 			 * and activate
437 			 */
438 			pmap_clear_modify(pg);
439 			uvm_lock_pageq();
440 			uvm_pageactivate(pg);
441 			uvm_unlock_pageq();
442 		}
443 
444 		/*
445 		 * We were not able to re-lock the map - restart the fault.
446 		 */
447 		if (!locked) {
448 			if (we_own) {
449 				rw_exit(anon->an_lock);
450 			}
451 			return (VM_PAGER_REFAULT);
452 		}
453 
454 		/*
455 		 * Verify that no one has touched the amap and moved
456 		 * the anon on us.
457 		 */
458 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
459 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
460 
461 			uvmfault_unlockall(ufi, amap, NULL);
462 			return (VM_PAGER_REFAULT);
463 		}
464 
465 		/*
466 		 * Retry..
467 		 */
468 		counters_inc(uvmexp_counters, flt_anretry);
469 		continue;
470 
471 	}
472 	/*NOTREACHED*/
473 }
474 
475 /*
476  * Update statistics after fault resolution.
477  * - maxrss
478  */
479 void
480 uvmfault_update_stats(struct uvm_faultinfo *ufi)
481 {
482 	struct vm_map		*map;
483 	struct proc		*p;
484 	vsize_t			 res;
485 
486 	map = ufi->orig_map;
487 
488 	/*
489 	 * If this is a nested pmap (eg, a virtual machine pmap managed
490 	 * by vmm(4) on amd64/i386), don't do any updating, just return.
491 	 *
492 	 * pmap_nested() on other archs is #defined to 0, so this is a
493 	 * no-op.
494 	 */
495 	if (pmap_nested(map->pmap))
496 		return;
497 
498 	/* Update the maxrss for the process. */
499 	if (map->flags & VM_MAP_ISVMSPACE) {
500 		p = curproc;
501 		KASSERT(p != NULL && &p->p_vmspace->vm_map == map);
502 
503 		res = pmap_resident_count(map->pmap);
504 		/* Convert res from pages to kilobytes. */
505 		res <<= (PAGE_SHIFT - 10);
506 
507 		if (p->p_ru.ru_maxrss < res)
508 			p->p_ru.ru_maxrss = res;
509 	}
510 }
511 
512 /*
513  *   F A U L T   -   m a i n   e n t r y   p o i n t
514  */
515 
516 /*
517  * uvm_fault: page fault handler
518  *
519  * => called from MD code to resolve a page fault
520  * => VM data structures usually should be unlocked.   however, it is
521  *	possible to call here with the main map locked if the caller
522  *	gets a write lock, sets it recursive, and then calls us (c.f.
523  *	uvm_map_pageable).   this should be avoided because it keeps
524  *	the map locked off during I/O.
525  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
526  */
527 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
528 			 ~PROT_WRITE : PROT_MASK)
529 struct uvm_faultctx {
530 	/*
531 	 * the following members are set up by uvm_fault_check() and
532 	 * read-only after that.
533 	 */
534 	vm_prot_t enter_prot;
535 	vm_prot_t access_type;
536 	vaddr_t startva;
537 	int npages;
538 	int centeridx;
539 	boolean_t narrow;
540 	boolean_t wired;
541 	paddr_t pa_flags;
542 };
543 
544 int		uvm_fault_check(
545 		    struct uvm_faultinfo *, struct uvm_faultctx *,
546 		    struct vm_anon ***);
547 
548 int		uvm_fault_upper(
549 		    struct uvm_faultinfo *, struct uvm_faultctx *,
550 		    struct vm_anon **, vm_fault_t);
551 boolean_t	uvm_fault_upper_lookup(
552 		    struct uvm_faultinfo *, const struct uvm_faultctx *,
553 		    struct vm_anon **, struct vm_page **);
554 
555 int		uvm_fault_lower(
556 		    struct uvm_faultinfo *, struct uvm_faultctx *,
557 		    struct vm_page **, vm_fault_t);
558 
559 int
560 uvm_fault(vm_map_t orig_map, vaddr_t vaddr, vm_fault_t fault_type,
561     vm_prot_t access_type)
562 {
563 	struct uvm_faultinfo ufi;
564 	struct uvm_faultctx flt;
565 	boolean_t shadowed;
566 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
567 	struct vm_page *pages[UVM_MAXRANGE];
568 	int error;
569 
570 	counters_inc(uvmexp_counters, faults);
571 	TRACEPOINT(uvm, fault, vaddr, fault_type, access_type, NULL);
572 
573 	/*
574 	 * init the IN parameters in the ufi
575 	 */
576 	ufi.orig_map = orig_map;
577 	ufi.orig_rvaddr = trunc_page(vaddr);
578 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
579 	if (fault_type == VM_FAULT_WIRE)
580 		flt.narrow = TRUE;	/* don't look for neighborhood
581 					 * pages on wire */
582 	else
583 		flt.narrow = FALSE;	/* normal fault */
584 	flt.access_type = access_type;
585 
586 
587 	error = ERESTART;
588 	while (error == ERESTART) { /* ReFault: */
589 		anons = anons_store;
590 
591 		error = uvm_fault_check(&ufi, &flt, &anons);
592 		if (error != 0)
593 			continue;
594 
595 		/* True if there is an anon at the faulting address */
596 		shadowed = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
597 		if (shadowed == TRUE) {
598 			/* case 1: fault on an anon in our amap */
599 			error = uvm_fault_upper(&ufi, &flt, anons, fault_type);
600 		} else {
601 			/* case 2: fault on backing object or zero fill */
602 			KERNEL_LOCK();
603 			error = uvm_fault_lower(&ufi, &flt, pages, fault_type);
604 			KERNEL_UNLOCK();
605 		}
606 	}
607 
608 	return error;
609 }
610 
611 /*
612  * uvm_fault_check: check prot, handle needs-copy, etc.
613  *
614  *	1. lookup entry.
615  *	2. check protection.
616  *	3. adjust fault condition (mainly for simulated fault).
617  *	4. handle needs-copy (lazy amap copy).
618  *	5. establish range of interest for neighbor fault (aka pre-fault).
619  *	6. look up anons (if amap exists).
620  *	7. flush pages (if MADV_SEQUENTIAL)
621  *
622  * => called with nothing locked.
623  * => if we fail (result != 0) we unlock everything.
624  * => initialize/adjust many members of flt.
625  */
626 int
627 uvm_fault_check(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
628     struct vm_anon ***ranons)
629 {
630 	struct vm_amap *amap;
631 	struct uvm_object *uobj;
632 	int nback, nforw;
633 
634 	/*
635 	 * lookup and lock the maps
636 	 */
637 	if (uvmfault_lookup(ufi, FALSE) == FALSE) {
638 		return EFAULT;
639 	}
640 	/* locked: maps(read) */
641 
642 #ifdef DIAGNOSTIC
643 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0)
644 		panic("uvm_fault: fault on non-pageable map (%p, 0x%lx)",
645 		    ufi->map, ufi->orig_rvaddr);
646 #endif
647 
648 	/*
649 	 * check protection
650 	 */
651 	if ((ufi->entry->protection & flt->access_type) != flt->access_type) {
652 		uvmfault_unlockmaps(ufi, FALSE);
653 		return EACCES;
654 	}
655 
656 	/*
657 	 * "enter_prot" is the protection we want to enter the page in at.
658 	 * for certain pages (e.g. copy-on-write pages) this protection can
659 	 * be more strict than ufi->entry->protection.  "wired" means either
660 	 * the entry is wired or we are fault-wiring the pg.
661 	 */
662 
663 	flt->enter_prot = ufi->entry->protection;
664 	flt->pa_flags = UVM_ET_ISWC(ufi->entry) ? PMAP_WC : 0;
665 	flt->wired = VM_MAPENT_ISWIRED(ufi->entry) || (flt->narrow == TRUE);
666 	if (flt->wired)
667 		flt->access_type = flt->enter_prot; /* full access for wired */
668 
669 	/* handle "needs_copy" case. */
670 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
671 		if ((flt->access_type & PROT_WRITE) ||
672 		    (ufi->entry->object.uvm_obj == NULL)) {
673 			/* need to clear */
674 			uvmfault_unlockmaps(ufi, FALSE);
675 			uvmfault_amapcopy(ufi);
676 			counters_inc(uvmexp_counters, flt_amcopy);
677 			return ERESTART;
678 		} else {
679 			/*
680 			 * ensure that we pmap_enter page R/O since
681 			 * needs_copy is still true
682 			 */
683 			flt->enter_prot &= ~PROT_WRITE;
684 		}
685 	}
686 
687 	/*
688 	 * identify the players
689 	 */
690 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
691 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
692 
693 	/*
694 	 * check for a case 0 fault.  if nothing backing the entry then
695 	 * error now.
696 	 */
697 	if (amap == NULL && uobj == NULL) {
698 		uvmfault_unlockmaps(ufi, FALSE);
699 		return EFAULT;
700 	}
701 
702 	/*
703 	 * establish range of interest based on advice from mapper
704 	 * and then clip to fit map entry.   note that we only want
705 	 * to do this the first time through the fault.   if we
706 	 * ReFault we will disable this by setting "narrow" to true.
707 	 */
708 	if (flt->narrow == FALSE) {
709 
710 		/* wide fault (!narrow) */
711 		nback = min(uvmadvice[ufi->entry->advice].nback,
712 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
713 		flt->startva = ufi->orig_rvaddr - ((vsize_t)nback << PAGE_SHIFT);
714 		nforw = min(uvmadvice[ufi->entry->advice].nforw,
715 		    ((ufi->entry->end - ufi->orig_rvaddr) >> PAGE_SHIFT) - 1);
716 		/*
717 		 * note: "-1" because we don't want to count the
718 		 * faulting page as forw
719 		 */
720 		flt->npages = nback + nforw + 1;
721 		flt->centeridx = nback;
722 
723 		flt->narrow = TRUE;	/* ensure only once per-fault */
724 	} else {
725 		/* narrow fault! */
726 		nback = nforw = 0;
727 		flt->startva = ufi->orig_rvaddr;
728 		flt->npages = 1;
729 		flt->centeridx = 0;
730 	}
731 
732 	/*
733 	 * if we've got an amap then lock it and extract current anons.
734 	 */
735 	if (amap) {
736 		amap_lock(amap);
737 		amap_lookups(&ufi->entry->aref,
738 		    flt->startva - ufi->entry->start, *ranons, flt->npages);
739 	} else {
740 		*ranons = NULL;	/* to be safe */
741 	}
742 
743 	/*
744 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
745 	 * now and then forget about them (for the rest of the fault).
746 	 */
747 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
748 		/* flush back-page anons? */
749 		if (amap)
750 			uvmfault_anonflush(*ranons, nback);
751 
752 		/* flush object? */
753 		if (uobj) {
754 			voff_t uoff;
755 
756 			uoff = (flt->startva - ufi->entry->start) + ufi->entry->offset;
757 			KERNEL_LOCK();
758 			(void) uobj->pgops->pgo_flush(uobj, uoff, uoff +
759 			    ((vsize_t)nback << PAGE_SHIFT), PGO_DEACTIVATE);
760 			KERNEL_UNLOCK();
761 		}
762 
763 		/* now forget about the backpages */
764 		if (amap)
765 			*ranons += nback;
766 		flt->startva += ((vsize_t)nback << PAGE_SHIFT);
767 		flt->npages -= nback;
768 		flt->centeridx = 0;
769 	}
770 
771 	return 0;
772 }
773 
774 /*
775  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
776  *
777  * iterate range of interest:
778  *	1. check if h/w mapping exists.  if yes, we don't care
779  *	2. check if anon exists.  if not, page is lower.
780  *	3. if anon exists, enter h/w mapping for neighbors.
781  *
782  * => called with amap locked (if exists).
783  */
784 boolean_t
785 uvm_fault_upper_lookup(struct uvm_faultinfo *ufi,
786     const struct uvm_faultctx *flt, struct vm_anon **anons,
787     struct vm_page **pages)
788 {
789 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
790 	struct vm_anon *anon;
791 	boolean_t shadowed;
792 	vaddr_t currva;
793 	paddr_t pa;
794 	int lcv;
795 
796 	/* locked: maps(read), amap(if there) */
797 	KASSERT(amap == NULL ||
798 	    rw_write_held(amap->am_lock));
799 
800 	/*
801 	 * map in the backpages and frontpages we found in the amap in hopes
802 	 * of preventing future faults.    we also init the pages[] array as
803 	 * we go.
804 	 */
805 	currva = flt->startva;
806 	shadowed = FALSE;
807 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
808 		/*
809 		 * dont play with VAs that are already mapped
810 		 * except for center)
811 		 */
812 		if (lcv != flt->centeridx &&
813 		    pmap_extract(ufi->orig_map->pmap, currva, &pa)) {
814 			pages[lcv] = PGO_DONTCARE;
815 			continue;
816 		}
817 
818 		/*
819 		 * unmapped or center page. check if any anon at this level.
820 		 */
821 		if (amap == NULL || anons[lcv] == NULL) {
822 			pages[lcv] = NULL;
823 			continue;
824 		}
825 
826 		/*
827 		 * check for present page and map if possible.
828 		 */
829 		pages[lcv] = PGO_DONTCARE;
830 		if (lcv == flt->centeridx) {	/* save center for later! */
831 			shadowed = TRUE;
832 			continue;
833 		}
834 		anon = anons[lcv];
835 		KASSERT(anon->an_lock == amap->am_lock);
836 		if (anon->an_page &&
837 		    (anon->an_page->pg_flags & (PG_RELEASED|PG_BUSY)) == 0) {
838 			uvm_lock_pageq();
839 			uvm_pageactivate(anon->an_page);	/* reactivate */
840 			uvm_unlock_pageq();
841 			counters_inc(uvmexp_counters, flt_namap);
842 
843 			/*
844 			 * Since this isn't the page that's actually faulting,
845 			 * ignore pmap_enter() failures; it's not critical
846 			 * that we enter these right now.
847 			 */
848 			(void) pmap_enter(ufi->orig_map->pmap, currva,
849 			    VM_PAGE_TO_PHYS(anon->an_page) | flt->pa_flags,
850 			    (anon->an_ref > 1) ?
851 			    (flt->enter_prot & ~PROT_WRITE) : flt->enter_prot,
852 			    PMAP_CANFAIL |
853 			     (VM_MAPENT_ISWIRED(ufi->entry) ? PMAP_WIRED : 0));
854 		}
855 	}
856 	if (flt->npages > 1)
857 		pmap_update(ufi->orig_map->pmap);
858 
859 	return shadowed;
860 }
861 
862 /*
863  * uvm_fault_upper: handle upper fault.
864  *
865  *	1. acquire anon lock.
866  *	2. get anon.  let uvmfault_anonget do the dirty work.
867  *	3. if COW, promote data to new anon
868  *	4. enter h/w mapping
869  */
870 int
871 uvm_fault_upper(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
872    struct vm_anon **anons, vm_fault_t fault_type)
873 {
874 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
875 	struct vm_anon *oanon, *anon = anons[flt->centeridx];
876 	struct vm_page *pg = NULL;
877 	int error, ret;
878 
879 	KASSERT(rw_write_held(amap->am_lock));
880 	KASSERT(anon->an_lock == amap->am_lock);
881 
882 	/*
883 	 * no matter if we have case 1A or case 1B we are going to need to
884 	 * have the anon's memory resident.   ensure that now.
885 	 */
886 	/*
887 	 * let uvmfault_anonget do the dirty work.
888 	 * if it fails (!OK) it will unlock everything for us.
889 	 * if it succeeds, locks are still valid and locked.
890 	 * also, if it is OK, then the anon's page is on the queues.
891 	 * if the page is on loan from a uvm_object, then anonget will
892 	 * lock that object for us if it does not fail.
893 	 */
894 	error = uvmfault_anonget(ufi, amap, anon);
895 	switch (error) {
896 	case VM_PAGER_OK:
897 		break;
898 
899 	case VM_PAGER_REFAULT:
900 		return ERESTART;
901 
902 	case VM_PAGER_ERROR:
903 		/*
904 		 * An error occured while trying to bring in the
905 		 * page -- this is the only error we return right
906 		 * now.
907 		 */
908 		return EACCES;	/* XXX */
909 	default:
910 #ifdef DIAGNOSTIC
911 		panic("uvm_fault: uvmfault_anonget -> %d", error);
912 #else
913 		return EACCES;
914 #endif
915 	}
916 
917 	KASSERT(rw_write_held(amap->am_lock));
918 	KASSERT(anon->an_lock == amap->am_lock);
919 
920 	/*
921 	 * if we are case 1B then we will need to allocate a new blank
922 	 * anon to transfer the data into.   note that we have a lock
923 	 * on anon, so no one can busy or release the page until we are done.
924 	 * also note that the ref count can't drop to zero here because
925 	 * it is > 1 and we are only dropping one ref.
926 	 *
927 	 * in the (hopefully very rare) case that we are out of RAM we
928 	 * will unlock, wait for more RAM, and refault.
929 	 *
930 	 * if we are out of anon VM we wait for RAM to become available.
931 	 */
932 
933 	if ((flt->access_type & PROT_WRITE) != 0 && anon->an_ref > 1) {
934 		counters_inc(uvmexp_counters, flt_acow);
935 		oanon = anon;		/* oanon = old */
936 		anon = uvm_analloc();
937 		if (anon) {
938 			anon->an_lock = amap->am_lock;
939 			pg = uvm_pagealloc(NULL, 0, anon, 0);
940 		}
941 
942 		/* check for out of RAM */
943 		if (anon == NULL || pg == NULL) {
944 			uvmfault_unlockall(ufi, amap, NULL);
945 			if (anon == NULL)
946 				counters_inc(uvmexp_counters, flt_noanon);
947 			else {
948 				anon->an_lock = NULL;
949 				anon->an_ref--;
950 				uvm_anfree(anon);
951 				counters_inc(uvmexp_counters, flt_noram);
952 			}
953 
954 			if (uvm_swapisfull())
955 				return ENOMEM;
956 
957 			/* out of RAM, wait for more */
958 			if (anon == NULL)
959 				uvm_anwait();
960 			else
961 				uvm_wait("flt_noram3");
962 			return ERESTART;
963 		}
964 
965 		/* got all resources, replace anon with nanon */
966 		uvm_pagecopy(oanon->an_page, pg);	/* pg now !PG_CLEAN */
967 		/* un-busy! new page */
968 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE);
969 		UVM_PAGE_OWN(pg, NULL);
970 		ret = amap_add(&ufi->entry->aref,
971 		    ufi->orig_rvaddr - ufi->entry->start, anon, 1);
972 		KASSERT(ret == 0);
973 
974 		/* deref: can not drop to zero here by defn! */
975 		oanon->an_ref--;
976 
977 		/*
978 		 * note: anon is _not_ locked, but we have the sole references
979 		 * to in from amap.
980 		 * thus, no one can get at it until we are done with it.
981 		 */
982 	} else {
983 		counters_inc(uvmexp_counters, flt_anon);
984 		oanon = anon;
985 		pg = anon->an_page;
986 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
987 			flt->enter_prot = flt->enter_prot & ~PROT_WRITE;
988 	}
989 
990 	/*
991 	 * now map the page in .
992 	 */
993 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
994 	    VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot,
995 	    flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) {
996 		/*
997 		 * No need to undo what we did; we can simply think of
998 		 * this as the pmap throwing away the mapping information.
999 		 *
1000 		 * We do, however, have to go through the ReFault path,
1001 		 * as the map may change while we're asleep.
1002 		 */
1003 		uvmfault_unlockall(ufi, amap, NULL);
1004 		if (uvm_swapisfull()) {
1005 			/* XXX instrumentation */
1006 			return ENOMEM;
1007 		}
1008 		/* XXX instrumentation */
1009 		uvm_wait("flt_pmfail1");
1010 		return ERESTART;
1011 	}
1012 
1013 	/* ... update the page queues. */
1014 	uvm_lock_pageq();
1015 
1016 	if (fault_type == VM_FAULT_WIRE) {
1017 		uvm_pagewire(pg);
1018 		/*
1019 		 * since the now-wired page cannot be paged out,
1020 		 * release its swap resources for others to use.
1021 		 * since an anon with no swap cannot be PG_CLEAN,
1022 		 * clear its clean flag now.
1023 		 */
1024 		atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1025 		uvm_anon_dropswap(anon);
1026 	} else {
1027 		/* activate it */
1028 		uvm_pageactivate(pg);
1029 	}
1030 
1031 	uvm_unlock_pageq();
1032 
1033 	/*
1034 	 * done case 1!  finish up by unlocking everything and returning success
1035 	 */
1036 	uvmfault_unlockall(ufi, amap, NULL);
1037 	pmap_update(ufi->orig_map->pmap);
1038 	return 0;
1039 }
1040 
1041 /*
1042  * uvm_fault_lower: handle lower fault.
1043  *
1044  */
1045 int
1046 uvm_fault_lower(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1047    struct vm_page **pages, vm_fault_t fault_type)
1048 {
1049 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1050 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1051 	boolean_t promote, locked;
1052 	int result, lcv, gotpages;
1053 	struct vm_page *uobjpage, *pg = NULL;
1054 	struct vm_anon *anon = NULL;
1055 	vaddr_t currva;
1056 	voff_t uoff;
1057 
1058 	/*
1059 	 * if the desired page is not shadowed by the amap and we have a
1060 	 * backing object, then we check to see if the backing object would
1061 	 * prefer to handle the fault itself (rather than letting us do it
1062 	 * with the usual pgo_get hook).  the backing object signals this by
1063 	 * providing a pgo_fault routine.
1064 	 */
1065 	if (uobj != NULL && uobj->pgops->pgo_fault != NULL) {
1066 		result = uobj->pgops->pgo_fault(ufi, flt->startva, pages,
1067 		    flt->npages, flt->centeridx, fault_type, flt->access_type,
1068 		    PGO_LOCKED);
1069 
1070 		if (result == VM_PAGER_OK)
1071 			return (0);		/* pgo_fault did pmap enter */
1072 		else if (result == VM_PAGER_REFAULT)
1073 			return ERESTART;	/* try again! */
1074 		else
1075 			return (EACCES);
1076 	}
1077 
1078 	/*
1079 	 * now, if the desired page is not shadowed by the amap and we have
1080 	 * a backing object that does not have a special fault routine, then
1081 	 * we ask (with pgo_get) the object for resident pages that we care
1082 	 * about and attempt to map them in.  we do not let pgo_get block
1083 	 * (PGO_LOCKED).
1084 	 *
1085 	 * ("get" has the option of doing a pmap_enter for us)
1086 	 */
1087 	if (uobj != NULL) {
1088 		counters_inc(uvmexp_counters, flt_lget);
1089 		gotpages = flt->npages;
1090 		(void) uobj->pgops->pgo_get(uobj, ufi->entry->offset +
1091 				(flt->startva - ufi->entry->start),
1092 				pages, &gotpages, flt->centeridx,
1093 				flt->access_type & MASK(ufi->entry),
1094 				ufi->entry->advice, PGO_LOCKED);
1095 
1096 		/* check for pages to map, if we got any */
1097 		uobjpage = NULL;
1098 		if (gotpages) {
1099 			currva = flt->startva;
1100 			for (lcv = 0 ; lcv < flt->npages ;
1101 			    lcv++, currva += PAGE_SIZE) {
1102 				if (pages[lcv] == NULL ||
1103 				    pages[lcv] == PGO_DONTCARE)
1104 					continue;
1105 
1106 				KASSERT((pages[lcv]->pg_flags & PG_RELEASED) == 0);
1107 
1108 				/*
1109 				 * if center page is resident and not
1110 				 * PG_BUSY, then pgo_get made it PG_BUSY
1111 				 * for us and gave us a handle to it.
1112 				 * remember this page as "uobjpage."
1113 				 * (for later use).
1114 				 */
1115 				if (lcv == flt->centeridx) {
1116 					uobjpage = pages[lcv];
1117 					continue;
1118 				}
1119 
1120 				/*
1121 				 * note: calling pgo_get with locked data
1122 				 * structures returns us pages which are
1123 				 * neither busy nor released, so we don't
1124 				 * need to check for this.   we can just
1125 				 * directly enter the page (after moving it
1126 				 * to the head of the active queue [useful?]).
1127 				 */
1128 
1129 				uvm_lock_pageq();
1130 				uvm_pageactivate(pages[lcv]);	/* reactivate */
1131 				uvm_unlock_pageq();
1132 				counters_inc(uvmexp_counters, flt_nomap);
1133 
1134 				/*
1135 				 * Since this page isn't the page that's
1136 				 * actually faulting, ignore pmap_enter()
1137 				 * failures; it's not critical that we
1138 				 * enter these right now.
1139 				 */
1140 				(void) pmap_enter(ufi->orig_map->pmap, currva,
1141 				    VM_PAGE_TO_PHYS(pages[lcv]) | flt->pa_flags,
1142 				    flt->enter_prot & MASK(ufi->entry),
1143 				    PMAP_CANFAIL |
1144 				     (flt->wired ? PMAP_WIRED : 0));
1145 
1146 				/*
1147 				 * NOTE: page can't be PG_WANTED because
1148 				 * we've held the lock the whole time
1149 				 * we've had the handle.
1150 				 */
1151 				atomic_clearbits_int(&pages[lcv]->pg_flags,
1152 				    PG_BUSY);
1153 				UVM_PAGE_OWN(pages[lcv], NULL);
1154 			}	/* for "lcv" loop */
1155 			pmap_update(ufi->orig_map->pmap);
1156 		}   /* "gotpages" != 0 */
1157 		/* note: object still _locked_ */
1158 	} else {
1159 		uobjpage = NULL;
1160 	}
1161 
1162 	/*
1163 	 * note that at this point we are done with any front or back pages.
1164 	 * we are now going to focus on the center page (i.e. the one we've
1165 	 * faulted on).  if we have faulted on the bottom (uobj)
1166 	 * layer [i.e. case 2] and the page was both present and available,
1167 	 * then we've got a pointer to it as "uobjpage" and we've already
1168 	 * made it BUSY.
1169 	 */
1170 
1171 	/*
1172 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1173 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1174 	 * have a backing object, check and see if we are going to promote
1175 	 * the data up to an anon during the fault.
1176 	 */
1177 	if (uobj == NULL) {
1178 		uobjpage = PGO_DONTCARE;
1179 		promote = TRUE;		/* always need anon here */
1180 	} else {
1181 		KASSERT(uobjpage != PGO_DONTCARE);
1182 		promote = (flt->access_type & PROT_WRITE) &&
1183 		     UVM_ET_ISCOPYONWRITE(ufi->entry);
1184 	}
1185 
1186 	/*
1187 	 * if uobjpage is not null then we do not need to do I/O to get the
1188 	 * uobjpage.
1189 	 *
1190 	 * if uobjpage is null, then we need to ask the pager to
1191 	 * get the data for us.   once we have the data, we need to reverify
1192 	 * the state the world.   we are currently not holding any resources.
1193 	 */
1194 	if (uobjpage) {
1195 		/* update rusage counters */
1196 		curproc->p_ru.ru_minflt++;
1197 	} else {
1198 		/* update rusage counters */
1199 		curproc->p_ru.ru_majflt++;
1200 
1201 		uvmfault_unlockall(ufi, amap, NULL);
1202 
1203 		counters_inc(uvmexp_counters, flt_get);
1204 		gotpages = 1;
1205 		uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1206 		result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1207 		    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1208 		    PGO_SYNCIO);
1209 
1210 		/* recover from I/O */
1211 		if (result != VM_PAGER_OK) {
1212 			KASSERT(result != VM_PAGER_PEND);
1213 
1214 			if (result == VM_PAGER_AGAIN) {
1215 				tsleep_nsec(&nowake, PVM, "fltagain2",
1216 				    SEC_TO_NSEC(1));
1217 				return ERESTART;
1218 			}
1219 
1220 			if (!UVM_ET_ISNOFAULT(ufi->entry))
1221 				return (EIO);
1222 
1223 			uobjpage = PGO_DONTCARE;
1224 			promote = TRUE;
1225 		}
1226 
1227 		/* re-verify the state of the world.  */
1228 		locked = uvmfault_relock(ufi);
1229 		if (locked && amap != NULL)
1230 			amap_lock(amap);
1231 
1232 		/*
1233 		 * Re-verify that amap slot is still free. if there is
1234 		 * a problem, we clean up.
1235 		 */
1236 		if (locked && amap && amap_lookup(&ufi->entry->aref,
1237 		      ufi->orig_rvaddr - ufi->entry->start)) {
1238 			if (locked)
1239 				uvmfault_unlockall(ufi, amap, NULL);
1240 			locked = FALSE;
1241 		}
1242 
1243 		/* didn't get the lock?   release the page and retry. */
1244 		if (locked == FALSE && uobjpage != PGO_DONTCARE) {
1245 			uvm_lock_pageq();
1246 			/* make sure it is in queues */
1247 			uvm_pageactivate(uobjpage);
1248 			uvm_unlock_pageq();
1249 
1250 			if (uobjpage->pg_flags & PG_WANTED)
1251 				/* still holding object lock */
1252 				wakeup(uobjpage);
1253 			atomic_clearbits_int(&uobjpage->pg_flags,
1254 			    PG_BUSY|PG_WANTED);
1255 			UVM_PAGE_OWN(uobjpage, NULL);
1256 			return ERESTART;
1257 		}
1258 		if (locked == FALSE)
1259 			return ERESTART;
1260 
1261 		/*
1262 		 * we have the data in uobjpage which is PG_BUSY
1263 		 */
1264 	}
1265 
1266 	/*
1267 	 * notes:
1268 	 *  - at this point uobjpage can not be NULL
1269 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1270 	 */
1271 	if (promote == FALSE) {
1272 		/*
1273 		 * we are not promoting.   if the mapping is COW ensure that we
1274 		 * don't give more access than we should (e.g. when doing a read
1275 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1276 		 * R/O even though the entry protection could be R/W).
1277 		 *
1278 		 * set "pg" to the page we want to map in (uobjpage, usually)
1279 		 */
1280 		counters_inc(uvmexp_counters, flt_obj);
1281 		if (UVM_ET_ISCOPYONWRITE(ufi->entry))
1282 			flt->enter_prot &= ~PROT_WRITE;
1283 		pg = uobjpage;		/* map in the actual object */
1284 
1285 		/* assert(uobjpage != PGO_DONTCARE) */
1286 
1287 		/*
1288 		 * we are faulting directly on the page.
1289 		 */
1290 	} else {
1291 		/*
1292 		 * if we are going to promote the data to an anon we
1293 		 * allocate a blank anon here and plug it into our amap.
1294 		 */
1295 #ifdef DIAGNOSTIC
1296 		if (amap == NULL)
1297 			panic("uvm_fault: want to promote data, but no anon");
1298 #endif
1299 
1300 		anon = uvm_analloc();
1301 		if (anon) {
1302 			/*
1303 			 * In `Fill in data...' below, if
1304 			 * uobjpage == PGO_DONTCARE, we want
1305 			 * a zero'd, dirty page, so have
1306 			 * uvm_pagealloc() do that for us.
1307 			 */
1308 			anon->an_lock = amap->am_lock;
1309 			pg = uvm_pagealloc(NULL, 0, anon,
1310 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1311 		}
1312 
1313 		/*
1314 		 * out of memory resources?
1315 		 */
1316 		if (anon == NULL || pg == NULL) {
1317 			/* arg!  must unbusy our page and fail or sleep. */
1318 			if (uobjpage != PGO_DONTCARE) {
1319 				uvm_lock_pageq();
1320 				uvm_pageactivate(uobjpage);
1321 				uvm_unlock_pageq();
1322 
1323 				if (uobjpage->pg_flags & PG_WANTED)
1324 					wakeup(uobjpage);
1325 				atomic_clearbits_int(&uobjpage->pg_flags,
1326 				    PG_BUSY|PG_WANTED);
1327 				UVM_PAGE_OWN(uobjpage, NULL);
1328 			}
1329 
1330 			/* unlock and fail ... */
1331 			uvmfault_unlockall(ufi, amap, uobj);
1332 			if (anon == NULL)
1333 				counters_inc(uvmexp_counters, flt_noanon);
1334 			else {
1335 				anon->an_lock = NULL;
1336 				anon->an_ref--;
1337 				uvm_anfree(anon);
1338 				counters_inc(uvmexp_counters, flt_noram);
1339 			}
1340 
1341 			if (uvm_swapisfull())
1342 				return (ENOMEM);
1343 
1344 			/* out of RAM, wait for more */
1345 			if (anon == NULL)
1346 				uvm_anwait();
1347 			else
1348 				uvm_wait("flt_noram5");
1349 			return ERESTART;
1350 		}
1351 
1352 		/* fill in the data */
1353 		if (uobjpage != PGO_DONTCARE) {
1354 			counters_inc(uvmexp_counters, flt_prcopy);
1355 			/* copy page [pg now dirty] */
1356 			uvm_pagecopy(uobjpage, pg);
1357 
1358 			/*
1359 			 * promote to shared amap?  make sure all sharing
1360 			 * procs see it
1361 			 */
1362 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1363 				pmap_page_protect(uobjpage, PROT_NONE);
1364 				}
1365 
1366 			/* dispose of uobjpage. drop handle to uobj as well. */
1367 			if (uobjpage->pg_flags & PG_WANTED)
1368 				wakeup(uobjpage);
1369 			atomic_clearbits_int(&uobjpage->pg_flags,
1370 			    PG_BUSY|PG_WANTED);
1371 			UVM_PAGE_OWN(uobjpage, NULL);
1372 			uvm_lock_pageq();
1373 			uvm_pageactivate(uobjpage);
1374 			uvm_unlock_pageq();
1375 			uobj = NULL;
1376 		} else {
1377 			counters_inc(uvmexp_counters, flt_przero);
1378 			/*
1379 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1380 			 * above.
1381 			 */
1382 		}
1383 
1384 		if (amap_add(&ufi->entry->aref,
1385 		    ufi->orig_rvaddr - ufi->entry->start, anon, 0)) {
1386 			uvmfault_unlockall(ufi, amap, NULL);
1387 			uvm_anfree(anon);
1388 			counters_inc(uvmexp_counters, flt_noamap);
1389 
1390 			if (uvm_swapisfull())
1391 				return (ENOMEM);
1392 
1393 			amap_populate(&ufi->entry->aref,
1394 			    ufi->orig_rvaddr - ufi->entry->start);
1395 			return ERESTART;
1396 		}
1397 	}
1398 
1399 	/* note: pg is either the uobjpage or the new page in the new anon */
1400 	/*
1401 	 * all resources are present.   we can now map it in and free our
1402 	 * resources.
1403 	 */
1404 	if (amap == NULL)
1405 		KASSERT(anon == NULL);
1406 	else {
1407 		KASSERT(rw_write_held(amap->am_lock));
1408 		KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
1409 	}
1410 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1411 	    VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot,
1412 	    flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) {
1413 		/*
1414 		 * No need to undo what we did; we can simply think of
1415 		 * this as the pmap throwing away the mapping information.
1416 		 *
1417 		 * We do, however, have to go through the ReFault path,
1418 		 * as the map may change while we're asleep.
1419 		 */
1420 		if (pg->pg_flags & PG_WANTED)
1421 			wakeup(pg);
1422 
1423 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1424 		UVM_PAGE_OWN(pg, NULL);
1425 		uvmfault_unlockall(ufi, amap, uobj);
1426 		if (uvm_swapisfull()) {
1427 			/* XXX instrumentation */
1428 			return (ENOMEM);
1429 		}
1430 		/* XXX instrumentation */
1431 		uvm_wait("flt_pmfail2");
1432 		return ERESTART;
1433 	}
1434 
1435 	uvm_lock_pageq();
1436 
1437 	if (fault_type == VM_FAULT_WIRE) {
1438 		uvm_pagewire(pg);
1439 		if (pg->pg_flags & PQ_AOBJ) {
1440 			/*
1441 			 * since the now-wired page cannot be paged out,
1442 			 * release its swap resources for others to use.
1443 			 * since an aobj page with no swap cannot be PG_CLEAN,
1444 			 * clear its clean flag now.
1445 			 */
1446 			atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1447 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1448 		}
1449 	} else {
1450 		/* activate it */
1451 		uvm_pageactivate(pg);
1452 	}
1453 	uvm_unlock_pageq();
1454 
1455 	if (pg->pg_flags & PG_WANTED)
1456 		wakeup(pg);
1457 
1458 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1459 	UVM_PAGE_OWN(pg, NULL);
1460 	uvmfault_unlockall(ufi, amap, uobj);
1461 	pmap_update(ufi->orig_map->pmap);
1462 
1463 	return (0);
1464 }
1465 
1466 
1467 /*
1468  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1469  *
1470  * => map may be read-locked by caller, but MUST NOT be write-locked.
1471  * => if map is read-locked, any operations which may cause map to
1472  *	be write-locked in uvm_fault() must be taken care of by
1473  *	the caller.  See uvm_map_pageable().
1474  */
1475 int
1476 uvm_fault_wire(vm_map_t map, vaddr_t start, vaddr_t end, vm_prot_t access_type)
1477 {
1478 	vaddr_t va;
1479 	int rv;
1480 
1481 	/*
1482 	 * now fault it in a page at a time.   if the fault fails then we have
1483 	 * to undo what we have done.   note that in uvm_fault PROT_NONE
1484 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1485 	 */
1486 	for (va = start ; va < end ; va += PAGE_SIZE) {
1487 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1488 		if (rv) {
1489 			if (va != start) {
1490 				uvm_fault_unwire(map, start, va);
1491 			}
1492 			return (rv);
1493 		}
1494 	}
1495 
1496 	return (0);
1497 }
1498 
1499 /*
1500  * uvm_fault_unwire(): unwire range of virtual space.
1501  */
1502 void
1503 uvm_fault_unwire(vm_map_t map, vaddr_t start, vaddr_t end)
1504 {
1505 
1506 	vm_map_lock_read(map);
1507 	uvm_fault_unwire_locked(map, start, end);
1508 	vm_map_unlock_read(map);
1509 }
1510 
1511 /*
1512  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1513  *
1514  * => map must be at least read-locked.
1515  */
1516 void
1517 uvm_fault_unwire_locked(vm_map_t map, vaddr_t start, vaddr_t end)
1518 {
1519 	vm_map_entry_t entry, next;
1520 	pmap_t pmap = vm_map_pmap(map);
1521 	vaddr_t va;
1522 	paddr_t pa;
1523 	struct vm_page *pg;
1524 
1525 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1526 
1527 	/*
1528 	 * we assume that the area we are unwiring has actually been wired
1529 	 * in the first place.   this means that we should be able to extract
1530 	 * the PAs from the pmap.   we also lock out the page daemon so that
1531 	 * we can call uvm_pageunwire.
1532 	 */
1533 	uvm_lock_pageq();
1534 
1535 	/* find the beginning map entry for the region. */
1536 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1537 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1538 		panic("uvm_fault_unwire_locked: address not in map");
1539 
1540 	for (va = start; va < end ; va += PAGE_SIZE) {
1541 		if (pmap_extract(pmap, va, &pa) == FALSE)
1542 			continue;
1543 
1544 		/* find the map entry for the current address. */
1545 		KASSERT(va >= entry->start);
1546 		while (va >= entry->end) {
1547 			next = RBT_NEXT(uvm_map_addr, entry);
1548 			KASSERT(next != NULL && next->start <= entry->end);
1549 			entry = next;
1550 		}
1551 
1552 		/* if the entry is no longer wired, tell the pmap. */
1553 		if (VM_MAPENT_ISWIRED(entry) == 0)
1554 			pmap_unwire(pmap, va);
1555 
1556 		pg = PHYS_TO_VM_PAGE(pa);
1557 		if (pg)
1558 			uvm_pageunwire(pg);
1559 	}
1560 
1561 	uvm_unlock_pageq();
1562 }
1563 
1564 /*
1565  * uvmfault_unlockmaps: unlock the maps
1566  */
1567 void
1568 uvmfault_unlockmaps(struct uvm_faultinfo *ufi, boolean_t write_locked)
1569 {
1570 	/*
1571 	 * ufi can be NULL when this isn't really a fault,
1572 	 * but merely paging in anon data.
1573 	 */
1574 	if (ufi == NULL) {
1575 		return;
1576 	}
1577 
1578 	uvmfault_update_stats(ufi);
1579 	if (write_locked) {
1580 		vm_map_unlock(ufi->map);
1581 	} else {
1582 		vm_map_unlock_read(ufi->map);
1583 	}
1584 }
1585 
1586 /*
1587  * uvmfault_unlockall: unlock everything passed in.
1588  *
1589  * => maps must be read-locked (not write-locked).
1590  */
1591 void
1592 uvmfault_unlockall(struct uvm_faultinfo *ufi, struct vm_amap *amap,
1593     struct uvm_object *uobj)
1594 {
1595 	if (amap != NULL)
1596 		amap_unlock(amap);
1597 	uvmfault_unlockmaps(ufi, FALSE);
1598 }
1599 
1600 /*
1601  * uvmfault_lookup: lookup a virtual address in a map
1602  *
1603  * => caller must provide a uvm_faultinfo structure with the IN
1604  *	params properly filled in
1605  * => we will lookup the map entry (handling submaps) as we go
1606  * => if the lookup is a success we will return with the maps locked
1607  * => if "write_lock" is TRUE, we write_lock the map, otherwise we only
1608  *	get a read lock.
1609  * => note that submaps can only appear in the kernel and they are
1610  *	required to use the same virtual addresses as the map they
1611  *	are referenced by (thus address translation between the main
1612  *	map and the submap is unnecessary).
1613  */
1614 
1615 boolean_t
1616 uvmfault_lookup(struct uvm_faultinfo *ufi, boolean_t write_lock)
1617 {
1618 	vm_map_t tmpmap;
1619 
1620 	/* init ufi values for lookup. */
1621 	ufi->map = ufi->orig_map;
1622 	ufi->size = ufi->orig_size;
1623 
1624 	/*
1625 	 * keep going down levels until we are done.   note that there can
1626 	 * only be two levels so we won't loop very long.
1627 	 */
1628 	while (1) {
1629 		if (ufi->orig_rvaddr < ufi->map->min_offset ||
1630 		    ufi->orig_rvaddr >= ufi->map->max_offset)
1631 			return(FALSE);
1632 
1633 		/* lock map */
1634 		if (write_lock) {
1635 			vm_map_lock(ufi->map);
1636 		} else {
1637 			vm_map_lock_read(ufi->map);
1638 		}
1639 
1640 		/* lookup */
1641 		if (!uvm_map_lookup_entry(ufi->map, ufi->orig_rvaddr,
1642 		    &ufi->entry)) {
1643 			uvmfault_unlockmaps(ufi, write_lock);
1644 			return(FALSE);
1645 		}
1646 
1647 		/* reduce size if necessary */
1648 		if (ufi->entry->end - ufi->orig_rvaddr < ufi->size)
1649 			ufi->size = ufi->entry->end - ufi->orig_rvaddr;
1650 
1651 		/*
1652 		 * submap?    replace map with the submap and lookup again.
1653 		 * note: VAs in submaps must match VAs in main map.
1654 		 */
1655 		if (UVM_ET_ISSUBMAP(ufi->entry)) {
1656 			tmpmap = ufi->entry->object.sub_map;
1657 			uvmfault_unlockmaps(ufi, write_lock);
1658 			ufi->map = tmpmap;
1659 			continue;
1660 		}
1661 
1662 		/* got it! */
1663 		ufi->mapv = ufi->map->timestamp;
1664 		return(TRUE);
1665 
1666 	}
1667 	/*NOTREACHED*/
1668 }
1669 
1670 /*
1671  * uvmfault_relock: attempt to relock the same version of the map
1672  *
1673  * => fault data structures should be unlocked before calling.
1674  * => if a success (TRUE) maps will be locked after call.
1675  */
1676 boolean_t
1677 uvmfault_relock(struct uvm_faultinfo *ufi)
1678 {
1679 	/*
1680 	 * ufi can be NULL when this isn't really a fault,
1681 	 * but merely paging in anon data.
1682 	 */
1683 	if (ufi == NULL) {
1684 		return TRUE;
1685 	}
1686 
1687 	counters_inc(uvmexp_counters, flt_relck);
1688 
1689 	/*
1690 	 * relock map.   fail if version mismatch (in which case nothing
1691 	 * gets locked).
1692 	 */
1693 	vm_map_lock_read(ufi->map);
1694 	if (ufi->mapv != ufi->map->timestamp) {
1695 		vm_map_unlock_read(ufi->map);
1696 		return(FALSE);
1697 	}
1698 
1699 	counters_inc(uvmexp_counters, flt_relckok);
1700 	return(TRUE);		/* got it! */
1701 }
1702