xref: /openbsd-src/sys/uvm/uvm_fault.c (revision a0747c9f67a4ae71ccb71e62a28d1ea19e06a63c)
1 /*	$OpenBSD: uvm_fault.c,v 1.123 2021/12/17 14:18:15 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 		/*
234 		 * no mapping?  give up.
235 		 */
236 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
237 			return;
238 
239 		/*
240 		 * copy if needed.
241 		 */
242 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
243 			amap_copy(ufi->map, ufi->entry, M_NOWAIT,
244 				UVM_ET_ISSTACK(ufi->entry) ? FALSE : TRUE,
245 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
246 
247 		/*
248 		 * didn't work?  must be out of RAM.   unlock and sleep.
249 		 */
250 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
251 			uvmfault_unlockmaps(ufi, TRUE);
252 			uvm_wait("fltamapcopy");
253 			continue;
254 		}
255 
256 		/*
257 		 * got it!   unlock and return.
258 		 */
259 		uvmfault_unlockmaps(ufi, TRUE);
260 		return;
261 	}
262 	/*NOTREACHED*/
263 }
264 
265 /*
266  * uvmfault_anonget: get data in an anon into a non-busy, non-released
267  * page in that anon.
268  *
269  * => Map, amap and thus anon should be locked by caller.
270  * => If we fail, we unlock everything and error is returned.
271  * => If we are successful, return with everything still locked.
272  * => We do not move the page on the queues [gets moved later].  If we
273  *    allocate a new page [we_own], it gets put on the queues.  Either way,
274  *    the result is that the page is on the queues at return time
275  */
276 int
277 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
278     struct vm_anon *anon)
279 {
280 	struct vm_page *pg;
281 	int error;
282 
283 	KASSERT(rw_lock_held(anon->an_lock));
284 	KASSERT(anon->an_lock == amap->am_lock);
285 
286 	/* Increment the counters.*/
287 	counters_inc(uvmexp_counters, flt_anget);
288 	if (anon->an_page) {
289 		curproc->p_ru.ru_minflt++;
290 	} else {
291 		curproc->p_ru.ru_majflt++;
292 	}
293 	error = 0;
294 
295 	/*
296 	 * Loop until we get the anon data, or fail.
297 	 */
298 	for (;;) {
299 		boolean_t we_own, locked;
300 		/*
301 		 * Note: 'we_own' will become true if we set PG_BUSY on a page.
302 		 */
303 		we_own = FALSE;
304 		pg = anon->an_page;
305 
306 		/*
307 		 * Is page resident?  Make sure it is not busy/released.
308 		 */
309 		if (pg) {
310 			KASSERT(pg->pg_flags & PQ_ANON);
311 			KASSERT(pg->uanon == anon);
312 
313 			/*
314 			 * if the page is busy, we drop all the locks and
315 			 * try again.
316 			 */
317 			if ((pg->pg_flags & (PG_BUSY|PG_RELEASED)) == 0)
318 				return (VM_PAGER_OK);
319 			atomic_setbits_int(&pg->pg_flags, PG_WANTED);
320 			counters_inc(uvmexp_counters, flt_pgwait);
321 
322 			/*
323 			 * The last unlock must be an atomic unlock and wait
324 			 * on the owner of page.
325 			 */
326 			if (pg->uobject) {
327 				/* Owner of page is UVM object. */
328 				uvmfault_unlockall(ufi, amap, NULL);
329 				rwsleep_nsec(pg, pg->uobject->vmobjlock,
330 				    PVM | PNORELOCK, "anonget1", INFSLP);
331 			} else {
332 				/* Owner of page is anon. */
333 				uvmfault_unlockall(ufi, NULL, NULL);
334 				rwsleep_nsec(pg, anon->an_lock, PVM | PNORELOCK,
335 				    "anonget2", INFSLP);
336 			}
337 		} else {
338 			/*
339 			 * No page, therefore allocate one.
340 			 */
341 			pg = uvm_pagealloc(NULL, 0, anon, 0);
342 			if (pg == NULL) {
343 				/* Out of memory.  Wait a little. */
344 				uvmfault_unlockall(ufi, amap, NULL);
345 				counters_inc(uvmexp_counters, flt_noram);
346 				uvm_wait("flt_noram1");
347 			} else {
348 				/* PG_BUSY bit is set. */
349 				we_own = TRUE;
350 				uvmfault_unlockall(ufi, amap, NULL);
351 
352 				/*
353 				 * Pass a PG_BUSY+PG_FAKE+PG_CLEAN page into
354 				 * the uvm_swap_get() function with all data
355 				 * structures unlocked.  Note that it is OK
356 				 * to read an_swslot here, because we hold
357 				 * PG_BUSY on the page.
358 				 */
359 				counters_inc(uvmexp_counters, pageins);
360 				error = uvm_swap_get(pg, anon->an_swslot,
361 				    PGO_SYNCIO);
362 
363 				/*
364 				 * We clean up after the I/O below in the
365 				 * 'we_own' case.
366 				 */
367 			}
368 		}
369 
370 		/*
371 		 * Re-lock the map and anon.
372 		 */
373 		locked = uvmfault_relock(ufi);
374 		if (locked || we_own) {
375 			rw_enter(anon->an_lock, RW_WRITE);
376 		}
377 
378 		/*
379 		 * If we own the page (i.e. we set PG_BUSY), then we need
380 		 * to clean up after the I/O.  There are three cases to
381 		 * consider:
382 		 *
383 		 * 1) Page was released during I/O: free anon and ReFault.
384 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
385 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
386 		 *    case (i.e. drop anon lock if not locked).
387 		 */
388 		if (we_own) {
389 			if (pg->pg_flags & PG_WANTED) {
390 				wakeup(pg);
391 			}
392 			/* un-busy! */
393 			atomic_clearbits_int(&pg->pg_flags,
394 			    PG_WANTED|PG_BUSY|PG_FAKE);
395 			UVM_PAGE_OWN(pg, NULL);
396 
397 			/*
398 			 * if we were RELEASED during I/O, then our anon is
399 			 * no longer part of an amap.   we need to free the
400 			 * anon and try again.
401 			 */
402 			if (pg->pg_flags & PG_RELEASED) {
403 				pmap_page_protect(pg, PROT_NONE);
404 				KASSERT(anon->an_ref == 0);
405 				/*
406 				 * Released while we had unlocked amap.
407 				 */
408 				if (locked)
409 					uvmfault_unlockall(ufi, NULL, NULL);
410 				uvm_anon_release(anon);	/* frees page for us */
411 				counters_inc(uvmexp_counters, flt_pgrele);
412 				return (VM_PAGER_REFAULT);	/* refault! */
413 			}
414 
415 			if (error != VM_PAGER_OK) {
416 				KASSERT(error != VM_PAGER_PEND);
417 
418 				/* remove page from anon */
419 				anon->an_page = NULL;
420 
421 				/*
422 				 * Remove the swap slot from the anon and
423 				 * mark the anon as having no real slot.
424 				 * Do not free the swap slot, thus preventing
425 				 * it from being used again.
426 				 */
427 				uvm_swap_markbad(anon->an_swslot, 1);
428 				anon->an_swslot = SWSLOT_BAD;
429 
430 				/*
431 				 * Note: page was never !PG_BUSY, so it
432 				 * cannot be mapped and thus no need to
433 				 * pmap_page_protect() it.
434 				 */
435 				uvm_lock_pageq();
436 				uvm_pagefree(pg);
437 				uvm_unlock_pageq();
438 
439 				if (locked) {
440 					uvmfault_unlockall(ufi, NULL, NULL);
441 				}
442 				rw_exit(anon->an_lock);
443 				return (VM_PAGER_ERROR);
444 			}
445 
446 			/*
447 			 * We have successfully read the page, activate it.
448 			 */
449 			pmap_clear_modify(pg);
450 			uvm_lock_pageq();
451 			uvm_pageactivate(pg);
452 			uvm_unlock_pageq();
453 		}
454 
455 		/*
456 		 * We were not able to re-lock the map - restart the fault.
457 		 */
458 		if (!locked) {
459 			if (we_own) {
460 				rw_exit(anon->an_lock);
461 			}
462 			return (VM_PAGER_REFAULT);
463 		}
464 
465 		/*
466 		 * Verify that no one has touched the amap and moved
467 		 * the anon on us.
468 		 */
469 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
470 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
471 
472 			uvmfault_unlockall(ufi, amap, NULL);
473 			return (VM_PAGER_REFAULT);
474 		}
475 
476 		/*
477 		 * Retry..
478 		 */
479 		counters_inc(uvmexp_counters, flt_anretry);
480 		continue;
481 
482 	}
483 	/*NOTREACHED*/
484 }
485 
486 /*
487  * Update statistics after fault resolution.
488  * - maxrss
489  */
490 void
491 uvmfault_update_stats(struct uvm_faultinfo *ufi)
492 {
493 	struct vm_map		*map;
494 	struct proc		*p;
495 	vsize_t			 res;
496 
497 	map = ufi->orig_map;
498 
499 	/*
500 	 * If this is a nested pmap (eg, a virtual machine pmap managed
501 	 * by vmm(4) on amd64/i386), don't do any updating, just return.
502 	 *
503 	 * pmap_nested() on other archs is #defined to 0, so this is a
504 	 * no-op.
505 	 */
506 	if (pmap_nested(map->pmap))
507 		return;
508 
509 	/* Update the maxrss for the process. */
510 	if (map->flags & VM_MAP_ISVMSPACE) {
511 		p = curproc;
512 		KASSERT(p != NULL && &p->p_vmspace->vm_map == map);
513 
514 		res = pmap_resident_count(map->pmap);
515 		/* Convert res from pages to kilobytes. */
516 		res <<= (PAGE_SHIFT - 10);
517 
518 		if (p->p_ru.ru_maxrss < res)
519 			p->p_ru.ru_maxrss = res;
520 	}
521 }
522 
523 /*
524  *   F A U L T   -   m a i n   e n t r y   p o i n t
525  */
526 
527 /*
528  * uvm_fault: page fault handler
529  *
530  * => called from MD code to resolve a page fault
531  * => VM data structures usually should be unlocked.   however, it is
532  *	possible to call here with the main map locked if the caller
533  *	gets a write lock, sets it recursive, and then calls us (c.f.
534  *	uvm_map_pageable).   this should be avoided because it keeps
535  *	the map locked off during I/O.
536  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
537  */
538 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
539 			 ~PROT_WRITE : PROT_MASK)
540 struct uvm_faultctx {
541 	/*
542 	 * the following members are set up by uvm_fault_check() and
543 	 * read-only after that.
544 	 */
545 	vm_prot_t enter_prot;
546 	vm_prot_t access_type;
547 	vaddr_t startva;
548 	int npages;
549 	int centeridx;
550 	boolean_t narrow;
551 	boolean_t wired;
552 	paddr_t pa_flags;
553 };
554 
555 int		uvm_fault_check(
556 		    struct uvm_faultinfo *, struct uvm_faultctx *,
557 		    struct vm_anon ***);
558 
559 int		uvm_fault_upper(
560 		    struct uvm_faultinfo *, struct uvm_faultctx *,
561 		    struct vm_anon **, vm_fault_t);
562 boolean_t	uvm_fault_upper_lookup(
563 		    struct uvm_faultinfo *, const struct uvm_faultctx *,
564 		    struct vm_anon **, struct vm_page **);
565 
566 int		uvm_fault_lower(
567 		    struct uvm_faultinfo *, struct uvm_faultctx *,
568 		    struct vm_page **, vm_fault_t);
569 
570 int
571 uvm_fault(vm_map_t orig_map, vaddr_t vaddr, vm_fault_t fault_type,
572     vm_prot_t access_type)
573 {
574 	struct uvm_faultinfo ufi;
575 	struct uvm_faultctx flt;
576 	boolean_t shadowed;
577 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
578 	struct vm_page *pages[UVM_MAXRANGE];
579 	int error;
580 
581 	counters_inc(uvmexp_counters, faults);
582 	TRACEPOINT(uvm, fault, vaddr, fault_type, access_type, NULL);
583 
584 	/*
585 	 * init the IN parameters in the ufi
586 	 */
587 	ufi.orig_map = orig_map;
588 	ufi.orig_rvaddr = trunc_page(vaddr);
589 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
590 	if (fault_type == VM_FAULT_WIRE)
591 		flt.narrow = TRUE;	/* don't look for neighborhood
592 					 * pages on wire */
593 	else
594 		flt.narrow = FALSE;	/* normal fault */
595 	flt.access_type = access_type;
596 
597 
598 	error = ERESTART;
599 	while (error == ERESTART) { /* ReFault: */
600 		anons = anons_store;
601 
602 		error = uvm_fault_check(&ufi, &flt, &anons);
603 		if (error != 0)
604 			continue;
605 
606 		/* True if there is an anon at the faulting address */
607 		shadowed = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
608 		if (shadowed == TRUE) {
609 			/* case 1: fault on an anon in our amap */
610 			error = uvm_fault_upper(&ufi, &flt, anons, fault_type);
611 		} else {
612 			struct uvm_object *uobj = ufi.entry->object.uvm_obj;
613 
614 			/*
615 			 * if the desired page is not shadowed by the amap and
616 			 * we have a backing object, then we check to see if
617 			 * the backing object would prefer to handle the fault
618 			 * itself (rather than letting us do it with the usual
619 			 * pgo_get hook).  the backing object signals this by
620 			 * providing a pgo_fault routine.
621 			 */
622 			if (uobj != NULL && uobj->pgops->pgo_fault != NULL) {
623 				KERNEL_LOCK();
624 				rw_enter(uobj->vmobjlock, RW_WRITE);
625 				error = uobj->pgops->pgo_fault(&ufi,
626 				    flt.startva, pages, flt.npages,
627 				    flt.centeridx, fault_type, flt.access_type,
628 				    PGO_LOCKED);
629 				KERNEL_UNLOCK();
630 
631 				if (error == VM_PAGER_OK)
632 					error = 0;
633 				else if (error == VM_PAGER_REFAULT)
634 					error = ERESTART;
635 				else
636 					error = EACCES;
637 			} else {
638 				/* case 2: fault on backing obj or zero fill */
639 				KERNEL_LOCK();
640 				error = uvm_fault_lower(&ufi, &flt, pages,
641 				    fault_type);
642 				KERNEL_UNLOCK();
643 			}
644 		}
645 	}
646 
647 	return error;
648 }
649 
650 /*
651  * uvm_fault_check: check prot, handle needs-copy, etc.
652  *
653  *	1. lookup entry.
654  *	2. check protection.
655  *	3. adjust fault condition (mainly for simulated fault).
656  *	4. handle needs-copy (lazy amap copy).
657  *	5. establish range of interest for neighbor fault (aka pre-fault).
658  *	6. look up anons (if amap exists).
659  *	7. flush pages (if MADV_SEQUENTIAL)
660  *
661  * => called with nothing locked.
662  * => if we fail (result != 0) we unlock everything.
663  * => initialize/adjust many members of flt.
664  */
665 int
666 uvm_fault_check(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
667     struct vm_anon ***ranons)
668 {
669 	struct vm_amap *amap;
670 	struct uvm_object *uobj;
671 	int nback, nforw;
672 
673 	/*
674 	 * lookup and lock the maps
675 	 */
676 	if (uvmfault_lookup(ufi, FALSE) == FALSE) {
677 		return EFAULT;
678 	}
679 	/* locked: maps(read) */
680 
681 #ifdef DIAGNOSTIC
682 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0)
683 		panic("uvm_fault: fault on non-pageable map (%p, 0x%lx)",
684 		    ufi->map, ufi->orig_rvaddr);
685 #endif
686 
687 	/*
688 	 * check protection
689 	 */
690 	if ((ufi->entry->protection & flt->access_type) != flt->access_type) {
691 		uvmfault_unlockmaps(ufi, FALSE);
692 		return EACCES;
693 	}
694 
695 	/*
696 	 * "enter_prot" is the protection we want to enter the page in at.
697 	 * for certain pages (e.g. copy-on-write pages) this protection can
698 	 * be more strict than ufi->entry->protection.  "wired" means either
699 	 * the entry is wired or we are fault-wiring the pg.
700 	 */
701 
702 	flt->enter_prot = ufi->entry->protection;
703 	flt->pa_flags = UVM_ET_ISWC(ufi->entry) ? PMAP_WC : 0;
704 	flt->wired = VM_MAPENT_ISWIRED(ufi->entry) || (flt->narrow == TRUE);
705 	if (flt->wired)
706 		flt->access_type = flt->enter_prot; /* full access for wired */
707 
708 	/* handle "needs_copy" case. */
709 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
710 		if ((flt->access_type & PROT_WRITE) ||
711 		    (ufi->entry->object.uvm_obj == NULL)) {
712 			/* need to clear */
713 			uvmfault_unlockmaps(ufi, FALSE);
714 			uvmfault_amapcopy(ufi);
715 			counters_inc(uvmexp_counters, flt_amcopy);
716 			return ERESTART;
717 		} else {
718 			/*
719 			 * ensure that we pmap_enter page R/O since
720 			 * needs_copy is still true
721 			 */
722 			flt->enter_prot &= ~PROT_WRITE;
723 		}
724 	}
725 
726 	/*
727 	 * identify the players
728 	 */
729 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
730 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
731 
732 	/*
733 	 * check for a case 0 fault.  if nothing backing the entry then
734 	 * error now.
735 	 */
736 	if (amap == NULL && uobj == NULL) {
737 		uvmfault_unlockmaps(ufi, FALSE);
738 		return EFAULT;
739 	}
740 
741 	/*
742 	 * establish range of interest based on advice from mapper
743 	 * and then clip to fit map entry.   note that we only want
744 	 * to do this the first time through the fault.   if we
745 	 * ReFault we will disable this by setting "narrow" to true.
746 	 */
747 	if (flt->narrow == FALSE) {
748 
749 		/* wide fault (!narrow) */
750 		nback = min(uvmadvice[ufi->entry->advice].nback,
751 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
752 		flt->startva = ufi->orig_rvaddr - ((vsize_t)nback << PAGE_SHIFT);
753 		nforw = min(uvmadvice[ufi->entry->advice].nforw,
754 		    ((ufi->entry->end - ufi->orig_rvaddr) >> PAGE_SHIFT) - 1);
755 		/*
756 		 * note: "-1" because we don't want to count the
757 		 * faulting page as forw
758 		 */
759 		flt->npages = nback + nforw + 1;
760 		flt->centeridx = nback;
761 
762 		flt->narrow = TRUE;	/* ensure only once per-fault */
763 	} else {
764 		/* narrow fault! */
765 		nback = nforw = 0;
766 		flt->startva = ufi->orig_rvaddr;
767 		flt->npages = 1;
768 		flt->centeridx = 0;
769 	}
770 
771 	/*
772 	 * if we've got an amap then lock it and extract current anons.
773 	 */
774 	if (amap) {
775 		amap_lock(amap);
776 		amap_lookups(&ufi->entry->aref,
777 		    flt->startva - ufi->entry->start, *ranons, flt->npages);
778 	} else {
779 		*ranons = NULL;	/* to be safe */
780 	}
781 
782 	/*
783 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
784 	 * now and then forget about them (for the rest of the fault).
785 	 */
786 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
787 		/* flush back-page anons? */
788 		if (amap)
789 			uvmfault_anonflush(*ranons, nback);
790 
791 		/*
792 		 * flush object?
793 		 */
794 		if (uobj) {
795 			voff_t uoff;
796 
797 			uoff = (flt->startva - ufi->entry->start) + ufi->entry->offset;
798 			rw_enter(uobj->vmobjlock, RW_WRITE);
799 			(void) uobj->pgops->pgo_flush(uobj, uoff, uoff +
800 			    ((vsize_t)nback << PAGE_SHIFT), PGO_DEACTIVATE);
801 			rw_exit(uobj->vmobjlock);
802 		}
803 
804 		/* now forget about the backpages */
805 		if (amap)
806 			*ranons += nback;
807 		flt->startva += ((vsize_t)nback << PAGE_SHIFT);
808 		flt->npages -= nback;
809 		flt->centeridx = 0;
810 	}
811 
812 	return 0;
813 }
814 
815 /*
816  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
817  *
818  * iterate range of interest:
819  *	1. check if h/w mapping exists.  if yes, we don't care
820  *	2. check if anon exists.  if not, page is lower.
821  *	3. if anon exists, enter h/w mapping for neighbors.
822  *
823  * => called with amap locked (if exists).
824  */
825 boolean_t
826 uvm_fault_upper_lookup(struct uvm_faultinfo *ufi,
827     const struct uvm_faultctx *flt, struct vm_anon **anons,
828     struct vm_page **pages)
829 {
830 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
831 	struct vm_anon *anon;
832 	boolean_t shadowed;
833 	vaddr_t currva;
834 	paddr_t pa;
835 	int lcv;
836 
837 	/* locked: maps(read), amap(if there) */
838 	KASSERT(amap == NULL ||
839 	    rw_write_held(amap->am_lock));
840 
841 	/*
842 	 * map in the backpages and frontpages we found in the amap in hopes
843 	 * of preventing future faults.    we also init the pages[] array as
844 	 * we go.
845 	 */
846 	currva = flt->startva;
847 	shadowed = FALSE;
848 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
849 		/*
850 		 * dont play with VAs that are already mapped
851 		 * except for center)
852 		 */
853 		if (lcv != flt->centeridx &&
854 		    pmap_extract(ufi->orig_map->pmap, currva, &pa)) {
855 			pages[lcv] = PGO_DONTCARE;
856 			continue;
857 		}
858 
859 		/*
860 		 * unmapped or center page.   check if any anon at this level.
861 		 */
862 		if (amap == NULL || anons[lcv] == NULL) {
863 			pages[lcv] = NULL;
864 			continue;
865 		}
866 
867 		/*
868 		 * check for present page and map if possible.
869 		 */
870 		pages[lcv] = PGO_DONTCARE;
871 		if (lcv == flt->centeridx) {	/* save center for later! */
872 			shadowed = TRUE;
873 			continue;
874 		}
875 		anon = anons[lcv];
876 		KASSERT(anon->an_lock == amap->am_lock);
877 		if (anon->an_page &&
878 		    (anon->an_page->pg_flags & (PG_RELEASED|PG_BUSY)) == 0) {
879 			uvm_lock_pageq();
880 			uvm_pageactivate(anon->an_page);	/* reactivate */
881 			uvm_unlock_pageq();
882 			counters_inc(uvmexp_counters, flt_namap);
883 
884 			/*
885 			 * Since this isn't the page that's actually faulting,
886 			 * ignore pmap_enter() failures; it's not critical
887 			 * that we enter these right now.
888 			 */
889 			(void) pmap_enter(ufi->orig_map->pmap, currva,
890 			    VM_PAGE_TO_PHYS(anon->an_page) | flt->pa_flags,
891 			    (anon->an_ref > 1) ?
892 			    (flt->enter_prot & ~PROT_WRITE) : flt->enter_prot,
893 			    PMAP_CANFAIL |
894 			     (VM_MAPENT_ISWIRED(ufi->entry) ? PMAP_WIRED : 0));
895 		}
896 	}
897 	if (flt->npages > 1)
898 		pmap_update(ufi->orig_map->pmap);
899 
900 	return shadowed;
901 }
902 
903 /*
904  * uvm_fault_upper: handle upper fault.
905  *
906  *	1. acquire anon lock.
907  *	2. get anon.  let uvmfault_anonget do the dirty work.
908  *	3. if COW, promote data to new anon
909  *	4. enter h/w mapping
910  */
911 int
912 uvm_fault_upper(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
913    struct vm_anon **anons, vm_fault_t fault_type)
914 {
915 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
916 	struct vm_anon *oanon, *anon = anons[flt->centeridx];
917 	struct vm_page *pg = NULL;
918 	int error, ret;
919 
920 	/* locked: maps(read), amap, anon */
921 	KASSERT(rw_write_held(amap->am_lock));
922 	KASSERT(anon->an_lock == amap->am_lock);
923 
924 	/*
925 	 * no matter if we have case 1A or case 1B we are going to need to
926 	 * have the anon's memory resident.   ensure that now.
927 	 */
928 	/*
929 	 * let uvmfault_anonget do the dirty work.
930 	 * if it fails (!OK) it will unlock everything for us.
931 	 * if it succeeds, locks are still valid and locked.
932 	 * also, if it is OK, then the anon's page is on the queues.
933 	 * if the page is on loan from a uvm_object, then anonget will
934 	 * lock that object for us if it does not fail.
935 	 */
936 	error = uvmfault_anonget(ufi, amap, anon);
937 	switch (error) {
938 	case VM_PAGER_OK:
939 		break;
940 
941 	case VM_PAGER_REFAULT:
942 		return ERESTART;
943 
944 	case VM_PAGER_ERROR:
945 		/*
946 		 * An error occurred while trying to bring in the
947 		 * page -- this is the only error we return right
948 		 * now.
949 		 */
950 		return EACCES;	/* XXX */
951 	default:
952 #ifdef DIAGNOSTIC
953 		panic("uvm_fault: uvmfault_anonget -> %d", error);
954 #else
955 		return EACCES;
956 #endif
957 	}
958 
959 	KASSERT(rw_write_held(amap->am_lock));
960 	KASSERT(anon->an_lock == amap->am_lock);
961 
962 	/*
963 	 * if we are case 1B then we will need to allocate a new blank
964 	 * anon to transfer the data into.   note that we have a lock
965 	 * on anon, so no one can busy or release the page until we are done.
966 	 * also note that the ref count can't drop to zero here because
967 	 * it is > 1 and we are only dropping one ref.
968 	 *
969 	 * in the (hopefully very rare) case that we are out of RAM we
970 	 * will unlock, wait for more RAM, and refault.
971 	 *
972 	 * if we are out of anon VM we wait for RAM to become available.
973 	 */
974 
975 	if ((flt->access_type & PROT_WRITE) != 0 && anon->an_ref > 1) {
976 		counters_inc(uvmexp_counters, flt_acow);
977 		oanon = anon;		/* oanon = old */
978 		anon = uvm_analloc();
979 		if (anon) {
980 			anon->an_lock = amap->am_lock;
981 			pg = uvm_pagealloc(NULL, 0, anon, 0);
982 		}
983 
984 		/* check for out of RAM */
985 		if (anon == NULL || pg == NULL) {
986 			uvmfault_unlockall(ufi, amap, NULL);
987 			if (anon == NULL)
988 				counters_inc(uvmexp_counters, flt_noanon);
989 			else {
990 				anon->an_lock = NULL;
991 				anon->an_ref--;
992 				uvm_anfree(anon);
993 				counters_inc(uvmexp_counters, flt_noram);
994 			}
995 
996 			if (uvm_swapisfull())
997 				return ENOMEM;
998 
999 			/* out of RAM, wait for more */
1000 			if (anon == NULL)
1001 				uvm_anwait();
1002 			else
1003 				uvm_wait("flt_noram3");
1004 			return ERESTART;
1005 		}
1006 
1007 		/* got all resources, replace anon with nanon */
1008 		uvm_pagecopy(oanon->an_page, pg);	/* pg now !PG_CLEAN */
1009 		/* un-busy! new page */
1010 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE);
1011 		UVM_PAGE_OWN(pg, NULL);
1012 		ret = amap_add(&ufi->entry->aref,
1013 		    ufi->orig_rvaddr - ufi->entry->start, anon, 1);
1014 		KASSERT(ret == 0);
1015 
1016 		/* deref: can not drop to zero here by defn! */
1017 		oanon->an_ref--;
1018 
1019 		/*
1020 		 * note: anon is _not_ locked, but we have the sole references
1021 		 * to in from amap.
1022 		 * thus, no one can get at it until we are done with it.
1023 		 */
1024 	} else {
1025 		counters_inc(uvmexp_counters, flt_anon);
1026 		oanon = anon;
1027 		pg = anon->an_page;
1028 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1029 			flt->enter_prot = flt->enter_prot & ~PROT_WRITE;
1030 	}
1031 
1032 	/*
1033 	 * now map the page in .
1034 	 */
1035 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1036 	    VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot,
1037 	    flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) {
1038 		/*
1039 		 * No need to undo what we did; we can simply think of
1040 		 * this as the pmap throwing away the mapping information.
1041 		 *
1042 		 * We do, however, have to go through the ReFault path,
1043 		 * as the map may change while we're asleep.
1044 		 */
1045 		uvmfault_unlockall(ufi, amap, NULL);
1046 		if (uvm_swapisfull()) {
1047 			/* XXX instrumentation */
1048 			return ENOMEM;
1049 		}
1050 		/* XXX instrumentation */
1051 		uvm_wait("flt_pmfail1");
1052 		return ERESTART;
1053 	}
1054 
1055 	/*
1056 	 * ... update the page queues.
1057 	 */
1058 	uvm_lock_pageq();
1059 
1060 	if (fault_type == VM_FAULT_WIRE) {
1061 		uvm_pagewire(pg);
1062 		/*
1063 		 * since the now-wired page cannot be paged out,
1064 		 * release its swap resources for others to use.
1065 		 * since an anon with no swap cannot be PG_CLEAN,
1066 		 * clear its clean flag now.
1067 		 */
1068 		atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1069 		uvm_anon_dropswap(anon);
1070 	} else {
1071 		/* activate it */
1072 		uvm_pageactivate(pg);
1073 	}
1074 
1075 	uvm_unlock_pageq();
1076 
1077 	/*
1078 	 * done case 1!  finish up by unlocking everything and returning success
1079 	 */
1080 	uvmfault_unlockall(ufi, amap, NULL);
1081 	pmap_update(ufi->orig_map->pmap);
1082 	return 0;
1083 }
1084 
1085 /*
1086  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1087  *
1088  *	1. get on-memory pages.
1089  *	2. if failed, give up (get only center page later).
1090  *	3. if succeeded, enter h/w mapping of neighbor pages.
1091  */
1092 
1093 struct vm_page *
1094 uvm_fault_lower_lookup(
1095 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1096 	struct vm_page **pages)
1097 {
1098 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1099 	struct vm_page *uobjpage = NULL;
1100 	int lcv, gotpages;
1101 	vaddr_t currva;
1102 
1103 	rw_enter(uobj->vmobjlock, RW_WRITE);
1104 
1105 	counters_inc(uvmexp_counters, flt_lget);
1106 	gotpages = flt->npages;
1107 	(void) uobj->pgops->pgo_get(uobj,
1108 	    ufi->entry->offset + (flt->startva - ufi->entry->start),
1109 	    pages, &gotpages, flt->centeridx,
1110 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1111 	    PGO_LOCKED);
1112 
1113 	/*
1114 	 * check for pages to map, if we got any
1115 	 */
1116 	if (gotpages == 0) {
1117 		return NULL;
1118 	}
1119 
1120 	currva = flt->startva;
1121 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1122 		if (pages[lcv] == NULL ||
1123 		    pages[lcv] == PGO_DONTCARE)
1124 			continue;
1125 
1126 		KASSERT((pages[lcv]->pg_flags & PG_RELEASED) == 0);
1127 
1128 		/*
1129 		 * if center page is resident and not
1130 		 * PG_BUSY, then pgo_get made it PG_BUSY
1131 		 * for us and gave us a handle to it.
1132 		 * remember this page as "uobjpage."
1133 		 * (for later use).
1134 		 */
1135 		if (lcv == flt->centeridx) {
1136 			uobjpage = pages[lcv];
1137 			continue;
1138 		}
1139 
1140 		/*
1141 		 * note: calling pgo_get with locked data
1142 		 * structures returns us pages which are
1143 		 * neither busy nor released, so we don't
1144 		 * need to check for this.   we can just
1145 		 * directly enter the page (after moving it
1146 		 * to the head of the active queue [useful?]).
1147 		 */
1148 
1149 		uvm_lock_pageq();
1150 		uvm_pageactivate(pages[lcv]);	/* reactivate */
1151 		uvm_unlock_pageq();
1152 		counters_inc(uvmexp_counters, flt_nomap);
1153 
1154 		/*
1155 		 * Since this page isn't the page that's
1156 		 * actually faulting, ignore pmap_enter()
1157 		 * failures; it's not critical that we
1158 		 * enter these right now.
1159 		 */
1160 		(void) pmap_enter(ufi->orig_map->pmap, currva,
1161 		    VM_PAGE_TO_PHYS(pages[lcv]) | flt->pa_flags,
1162 		    flt->enter_prot & MASK(ufi->entry),
1163 		    PMAP_CANFAIL |
1164 		     (flt->wired ? PMAP_WIRED : 0));
1165 
1166 		/*
1167 		 * NOTE: page can't be PG_WANTED because
1168 		 * we've held the lock the whole time
1169 		 * we've had the handle.
1170 		 */
1171 		atomic_clearbits_int(&pages[lcv]->pg_flags, PG_BUSY);
1172 		UVM_PAGE_OWN(pages[lcv], NULL);
1173 	}
1174 	pmap_update(ufi->orig_map->pmap);
1175 
1176 	return uobjpage;
1177 }
1178 
1179 /*
1180  * uvm_fault_lower: handle lower fault.
1181  *
1182  */
1183 int
1184 uvm_fault_lower(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1185    struct vm_page **pages, vm_fault_t fault_type)
1186 {
1187 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1188 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1189 	boolean_t promote, locked;
1190 	int result;
1191 	struct vm_page *uobjpage, *pg = NULL;
1192 	struct vm_anon *anon = NULL;
1193 	voff_t uoff;
1194 
1195 	/*
1196 	 * now, if the desired page is not shadowed by the amap and we have
1197 	 * a backing object that does not have a special fault routine, then
1198 	 * we ask (with pgo_get) the object for resident pages that we care
1199 	 * about and attempt to map them in.  we do not let pgo_get block
1200 	 * (PGO_LOCKED).
1201 	 */
1202 	if (uobj == NULL) {
1203 		/* zero fill; don't care neighbor pages */
1204 		uobjpage = NULL;
1205 	} else {
1206 		uobjpage = uvm_fault_lower_lookup(ufi, flt, pages);
1207 	}
1208 
1209 	/*
1210 	 * note that at this point we are done with any front or back pages.
1211 	 * we are now going to focus on the center page (i.e. the one we've
1212 	 * faulted on).  if we have faulted on the bottom (uobj)
1213 	 * layer [i.e. case 2] and the page was both present and available,
1214 	 * then we've got a pointer to it as "uobjpage" and we've already
1215 	 * made it BUSY.
1216 	 */
1217 
1218 	/*
1219 	 * locked:
1220 	 */
1221 	KASSERT(amap == NULL ||
1222 	    rw_write_held(amap->am_lock));
1223 	KASSERT(uobj == NULL ||
1224 	    rw_write_held(uobj->vmobjlock));
1225 
1226 	/*
1227 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1228 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1229 	 * have a backing object, check and see if we are going to promote
1230 	 * the data up to an anon during the fault.
1231 	 */
1232 	if (uobj == NULL) {
1233 		uobjpage = PGO_DONTCARE;
1234 		promote = TRUE;		/* always need anon here */
1235 	} else {
1236 		KASSERT(uobjpage != PGO_DONTCARE);
1237 		promote = (flt->access_type & PROT_WRITE) &&
1238 		     UVM_ET_ISCOPYONWRITE(ufi->entry);
1239 	}
1240 
1241 	/*
1242 	 * if uobjpage is not null then we do not need to do I/O to get the
1243 	 * uobjpage.
1244 	 *
1245 	 * if uobjpage is null, then we need to ask the pager to
1246 	 * get the data for us.   once we have the data, we need to reverify
1247 	 * the state the world.   we are currently not holding any resources.
1248 	 */
1249 	if (uobjpage) {
1250 		/* update rusage counters */
1251 		curproc->p_ru.ru_minflt++;
1252 	} else {
1253 		int gotpages;
1254 
1255 		/* update rusage counters */
1256 		curproc->p_ru.ru_majflt++;
1257 
1258 		uvmfault_unlockall(ufi, amap, NULL);
1259 
1260 		counters_inc(uvmexp_counters, flt_get);
1261 		gotpages = 1;
1262 		uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1263 		result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1264 		    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1265 		    PGO_SYNCIO);
1266 
1267 		/*
1268 		 * recover from I/O
1269 		 */
1270 		if (result != VM_PAGER_OK) {
1271 			KASSERT(result != VM_PAGER_PEND);
1272 
1273 			if (result == VM_PAGER_AGAIN) {
1274 				tsleep_nsec(&nowake, PVM, "fltagain2",
1275 				    MSEC_TO_NSEC(5));
1276 				return ERESTART;
1277 			}
1278 
1279 			if (!UVM_ET_ISNOFAULT(ufi->entry))
1280 				return (EIO);
1281 
1282 			uobjpage = PGO_DONTCARE;
1283 			uobj = NULL;
1284 			promote = TRUE;
1285 		}
1286 
1287 		/* re-verify the state of the world.  */
1288 		locked = uvmfault_relock(ufi);
1289 		if (locked && amap != NULL)
1290 			amap_lock(amap);
1291 
1292 		/* might be changed */
1293 		if (uobjpage != PGO_DONTCARE) {
1294 			uobj = uobjpage->uobject;
1295 			rw_enter(uobj->vmobjlock, RW_WRITE);
1296 		}
1297 
1298 		/*
1299 		 * Re-verify that amap slot is still free. if there is
1300 		 * a problem, we clean up.
1301 		 */
1302 		if (locked && amap && amap_lookup(&ufi->entry->aref,
1303 		      ufi->orig_rvaddr - ufi->entry->start)) {
1304 			if (locked)
1305 				uvmfault_unlockall(ufi, amap, NULL);
1306 			locked = FALSE;
1307 		}
1308 
1309 		/* didn't get the lock?   release the page and retry. */
1310 		if (locked == FALSE && uobjpage != PGO_DONTCARE) {
1311 			uvm_lock_pageq();
1312 			/* make sure it is in queues */
1313 			uvm_pageactivate(uobjpage);
1314 			uvm_unlock_pageq();
1315 
1316 			if (uobjpage->pg_flags & PG_WANTED)
1317 				/* still holding object lock */
1318 				wakeup(uobjpage);
1319 			atomic_clearbits_int(&uobjpage->pg_flags,
1320 			    PG_BUSY|PG_WANTED);
1321 			UVM_PAGE_OWN(uobjpage, NULL);
1322 		}
1323 
1324 		if (locked == FALSE) {
1325 			if (uobjpage != PGO_DONTCARE)
1326 				rw_exit(uobj->vmobjlock);
1327 			return ERESTART;
1328 		}
1329 
1330 		/*
1331 		 * we have the data in uobjpage which is PG_BUSY
1332 		 */
1333 	}
1334 
1335 	/*
1336 	 * notes:
1337 	 *  - at this point uobjpage can not be NULL
1338 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1339 	 */
1340 	if (promote == FALSE) {
1341 		/*
1342 		 * we are not promoting.   if the mapping is COW ensure that we
1343 		 * don't give more access than we should (e.g. when doing a read
1344 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1345 		 * R/O even though the entry protection could be R/W).
1346 		 *
1347 		 * set "pg" to the page we want to map in (uobjpage, usually)
1348 		 */
1349 		counters_inc(uvmexp_counters, flt_obj);
1350 		if (UVM_ET_ISCOPYONWRITE(ufi->entry))
1351 			flt->enter_prot &= ~PROT_WRITE;
1352 		pg = uobjpage;		/* map in the actual object */
1353 
1354 		/* assert(uobjpage != PGO_DONTCARE) */
1355 
1356 		/*
1357 		 * we are faulting directly on the page.
1358 		 */
1359 	} else {
1360 		/*
1361 		 * if we are going to promote the data to an anon we
1362 		 * allocate a blank anon here and plug it into our amap.
1363 		 */
1364 #ifdef DIAGNOSTIC
1365 		if (amap == NULL)
1366 			panic("uvm_fault: want to promote data, but no anon");
1367 #endif
1368 
1369 		anon = uvm_analloc();
1370 		if (anon) {
1371 			/*
1372 			 * In `Fill in data...' below, if
1373 			 * uobjpage == PGO_DONTCARE, we want
1374 			 * a zero'd, dirty page, so have
1375 			 * uvm_pagealloc() do that for us.
1376 			 */
1377 			anon->an_lock = amap->am_lock;
1378 			pg = uvm_pagealloc(NULL, 0, anon,
1379 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1380 		}
1381 
1382 		/*
1383 		 * out of memory resources?
1384 		 */
1385 		if (anon == NULL || pg == NULL) {
1386 			/*
1387 			 * arg!  must unbusy our page and fail or sleep.
1388 			 */
1389 			if (uobjpage != PGO_DONTCARE) {
1390 				uvm_lock_pageq();
1391 				uvm_pageactivate(uobjpage);
1392 				uvm_unlock_pageq();
1393 
1394 				if (uobjpage->pg_flags & PG_WANTED)
1395 					wakeup(uobjpage);
1396 				atomic_clearbits_int(&uobjpage->pg_flags,
1397 				    PG_BUSY|PG_WANTED);
1398 				UVM_PAGE_OWN(uobjpage, NULL);
1399 			}
1400 
1401 			/* unlock and fail ... */
1402 			uvmfault_unlockall(ufi, amap, uobj);
1403 			if (anon == NULL)
1404 				counters_inc(uvmexp_counters, flt_noanon);
1405 			else {
1406 				anon->an_lock = NULL;
1407 				anon->an_ref--;
1408 				uvm_anfree(anon);
1409 				counters_inc(uvmexp_counters, flt_noram);
1410 			}
1411 
1412 			if (uvm_swapisfull())
1413 				return (ENOMEM);
1414 
1415 			/* out of RAM, wait for more */
1416 			if (anon == NULL)
1417 				uvm_anwait();
1418 			else
1419 				uvm_wait("flt_noram5");
1420 			return ERESTART;
1421 		}
1422 
1423 		/*
1424 		 * fill in the data
1425 		 */
1426 		if (uobjpage != PGO_DONTCARE) {
1427 			counters_inc(uvmexp_counters, flt_prcopy);
1428 			/* copy page [pg now dirty] */
1429 			uvm_pagecopy(uobjpage, pg);
1430 
1431 			/*
1432 			 * promote to shared amap?  make sure all sharing
1433 			 * procs see it
1434 			 */
1435 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1436 				pmap_page_protect(uobjpage, PROT_NONE);
1437 				}
1438 
1439 			/* dispose of uobjpage. drop handle to uobj as well. */
1440 			if (uobjpage->pg_flags & PG_WANTED)
1441 				wakeup(uobjpage);
1442 			atomic_clearbits_int(&uobjpage->pg_flags,
1443 			    PG_BUSY|PG_WANTED);
1444 			UVM_PAGE_OWN(uobjpage, NULL);
1445 			uvm_lock_pageq();
1446 			uvm_pageactivate(uobjpage);
1447 			uvm_unlock_pageq();
1448 			rw_exit(uobj->vmobjlock);
1449 			uobj = NULL;
1450 		} else {
1451 			counters_inc(uvmexp_counters, flt_przero);
1452 			/*
1453 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1454 			 * above.
1455 			 */
1456 		}
1457 
1458 		if (amap_add(&ufi->entry->aref,
1459 		    ufi->orig_rvaddr - ufi->entry->start, anon, 0)) {
1460 			uvmfault_unlockall(ufi, amap, uobj);
1461 			uvm_anfree(anon);
1462 			counters_inc(uvmexp_counters, flt_noamap);
1463 
1464 			if (uvm_swapisfull())
1465 				return (ENOMEM);
1466 
1467 			amap_populate(&ufi->entry->aref,
1468 			    ufi->orig_rvaddr - ufi->entry->start);
1469 			return ERESTART;
1470 		}
1471 	}
1472 
1473 	/* note: pg is either the uobjpage or the new page in the new anon */
1474 	/*
1475 	 * all resources are present.   we can now map it in and free our
1476 	 * resources.
1477 	 */
1478 	if (amap == NULL)
1479 		KASSERT(anon == NULL);
1480 	else {
1481 		KASSERT(rw_write_held(amap->am_lock));
1482 		KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
1483 	}
1484 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1485 	    VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot,
1486 	    flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) {
1487 		/*
1488 		 * No need to undo what we did; we can simply think of
1489 		 * this as the pmap throwing away the mapping information.
1490 		 *
1491 		 * We do, however, have to go through the ReFault path,
1492 		 * as the map may change while we're asleep.
1493 		 */
1494 		if (pg->pg_flags & PG_WANTED)
1495 			wakeup(pg);
1496 
1497 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1498 		UVM_PAGE_OWN(pg, NULL);
1499 		uvmfault_unlockall(ufi, amap, uobj);
1500 		if (uvm_swapisfull()) {
1501 			/* XXX instrumentation */
1502 			return (ENOMEM);
1503 		}
1504 		/* XXX instrumentation */
1505 		uvm_wait("flt_pmfail2");
1506 		return ERESTART;
1507 	}
1508 
1509 	if (fault_type == VM_FAULT_WIRE) {
1510 		uvm_lock_pageq();
1511 		uvm_pagewire(pg);
1512 		uvm_unlock_pageq();
1513 		if (pg->pg_flags & PQ_AOBJ) {
1514 			/*
1515 			 * since the now-wired page cannot be paged out,
1516 			 * release its swap resources for others to use.
1517 			 * since an aobj page with no swap cannot be clean,
1518 			 * mark it dirty now.
1519 			 *
1520 			 * use pg->uobject here.  if the page is from a
1521 			 * tmpfs vnode, the pages are backed by its UAO and
1522 			 * not the vnode.
1523 			 */
1524 			KASSERT(uobj != NULL);
1525 			KASSERT(uobj->vmobjlock == pg->uobject->vmobjlock);
1526 			atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1527 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1528 		}
1529 	} else {
1530 		/* activate it */
1531 		uvm_lock_pageq();
1532 		uvm_pageactivate(pg);
1533 		uvm_unlock_pageq();
1534 	}
1535 
1536 	if (pg->pg_flags & PG_WANTED)
1537 		wakeup(pg);
1538 
1539 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1540 	UVM_PAGE_OWN(pg, NULL);
1541 	uvmfault_unlockall(ufi, amap, uobj);
1542 	pmap_update(ufi->orig_map->pmap);
1543 
1544 	return (0);
1545 }
1546 
1547 
1548 /*
1549  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1550  *
1551  * => map may be read-locked by caller, but MUST NOT be write-locked.
1552  * => if map is read-locked, any operations which may cause map to
1553  *	be write-locked in uvm_fault() must be taken care of by
1554  *	the caller.  See uvm_map_pageable().
1555  */
1556 int
1557 uvm_fault_wire(vm_map_t map, vaddr_t start, vaddr_t end, vm_prot_t access_type)
1558 {
1559 	vaddr_t va;
1560 	int rv;
1561 
1562 	/*
1563 	 * now fault it in a page at a time.   if the fault fails then we have
1564 	 * to undo what we have done.   note that in uvm_fault PROT_NONE
1565 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1566 	 */
1567 	for (va = start ; va < end ; va += PAGE_SIZE) {
1568 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1569 		if (rv) {
1570 			if (va != start) {
1571 				uvm_fault_unwire(map, start, va);
1572 			}
1573 			return (rv);
1574 		}
1575 	}
1576 
1577 	return (0);
1578 }
1579 
1580 /*
1581  * uvm_fault_unwire(): unwire range of virtual space.
1582  */
1583 void
1584 uvm_fault_unwire(vm_map_t map, vaddr_t start, vaddr_t end)
1585 {
1586 
1587 	vm_map_lock_read(map);
1588 	uvm_fault_unwire_locked(map, start, end);
1589 	vm_map_unlock_read(map);
1590 }
1591 
1592 /*
1593  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1594  *
1595  * => map must be at least read-locked.
1596  */
1597 void
1598 uvm_fault_unwire_locked(vm_map_t map, vaddr_t start, vaddr_t end)
1599 {
1600 	vm_map_entry_t entry, oentry = NULL, next;
1601 	pmap_t pmap = vm_map_pmap(map);
1602 	vaddr_t va;
1603 	paddr_t pa;
1604 	struct vm_page *pg;
1605 
1606 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1607 
1608 	/*
1609 	 * we assume that the area we are unwiring has actually been wired
1610 	 * in the first place.   this means that we should be able to extract
1611 	 * the PAs from the pmap.
1612 	 */
1613 
1614 	/*
1615 	 * find the beginning map entry for the region.
1616 	 */
1617 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1618 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1619 		panic("uvm_fault_unwire_locked: address not in map");
1620 
1621 	for (va = start; va < end ; va += PAGE_SIZE) {
1622 		if (pmap_extract(pmap, va, &pa) == FALSE)
1623 			continue;
1624 
1625 		/*
1626 		 * find the map entry for the current address.
1627 		 */
1628 		KASSERT(va >= entry->start);
1629 		while (va >= entry->end) {
1630 			next = RBT_NEXT(uvm_map_addr, entry);
1631 			KASSERT(next != NULL && next->start <= entry->end);
1632 			entry = next;
1633 		}
1634 
1635 		/*
1636 		 * lock it.
1637 		 */
1638 		if (entry != oentry) {
1639 			if (oentry != NULL) {
1640 				uvm_map_unlock_entry(oentry);
1641 			}
1642 			uvm_map_lock_entry(entry);
1643 			oentry = entry;
1644 		}
1645 
1646 		/*
1647 		 * if the entry is no longer wired, tell the pmap.
1648 		 */
1649 		if (VM_MAPENT_ISWIRED(entry) == 0)
1650 			pmap_unwire(pmap, va);
1651 
1652 		pg = PHYS_TO_VM_PAGE(pa);
1653 		if (pg) {
1654 			uvm_lock_pageq();
1655 			uvm_pageunwire(pg);
1656 			uvm_unlock_pageq();
1657 		}
1658 	}
1659 
1660 	if (oentry != NULL) {
1661 		uvm_map_unlock_entry(entry);
1662 	}
1663 }
1664 
1665 /*
1666  * uvmfault_unlockmaps: unlock the maps
1667  */
1668 void
1669 uvmfault_unlockmaps(struct uvm_faultinfo *ufi, boolean_t write_locked)
1670 {
1671 	/*
1672 	 * ufi can be NULL when this isn't really a fault,
1673 	 * but merely paging in anon data.
1674 	 */
1675 	if (ufi == NULL) {
1676 		return;
1677 	}
1678 
1679 	uvmfault_update_stats(ufi);
1680 	if (write_locked) {
1681 		vm_map_unlock(ufi->map);
1682 	} else {
1683 		vm_map_unlock_read(ufi->map);
1684 	}
1685 }
1686 
1687 /*
1688  * uvmfault_unlockall: unlock everything passed in.
1689  *
1690  * => maps must be read-locked (not write-locked).
1691  */
1692 void
1693 uvmfault_unlockall(struct uvm_faultinfo *ufi, struct vm_amap *amap,
1694     struct uvm_object *uobj)
1695 {
1696 	if (uobj)
1697 		rw_exit(uobj->vmobjlock);
1698 	if (amap != NULL)
1699 		amap_unlock(amap);
1700 	uvmfault_unlockmaps(ufi, FALSE);
1701 }
1702 
1703 /*
1704  * uvmfault_lookup: lookup a virtual address in a map
1705  *
1706  * => caller must provide a uvm_faultinfo structure with the IN
1707  *	params properly filled in
1708  * => we will lookup the map entry (handling submaps) as we go
1709  * => if the lookup is a success we will return with the maps locked
1710  * => if "write_lock" is TRUE, we write_lock the map, otherwise we only
1711  *	get a read lock.
1712  * => note that submaps can only appear in the kernel and they are
1713  *	required to use the same virtual addresses as the map they
1714  *	are referenced by (thus address translation between the main
1715  *	map and the submap is unnecessary).
1716  */
1717 
1718 boolean_t
1719 uvmfault_lookup(struct uvm_faultinfo *ufi, boolean_t write_lock)
1720 {
1721 	vm_map_t tmpmap;
1722 
1723 	/*
1724 	 * init ufi values for lookup.
1725 	 */
1726 	ufi->map = ufi->orig_map;
1727 	ufi->size = ufi->orig_size;
1728 
1729 	/*
1730 	 * keep going down levels until we are done.   note that there can
1731 	 * only be two levels so we won't loop very long.
1732 	 */
1733 	while (1) {
1734 		if (ufi->orig_rvaddr < ufi->map->min_offset ||
1735 		    ufi->orig_rvaddr >= ufi->map->max_offset)
1736 			return FALSE;
1737 
1738 		/* lock map */
1739 		if (write_lock) {
1740 			vm_map_lock(ufi->map);
1741 		} else {
1742 			vm_map_lock_read(ufi->map);
1743 		}
1744 
1745 		/* lookup */
1746 		if (!uvm_map_lookup_entry(ufi->map, ufi->orig_rvaddr,
1747 		    &ufi->entry)) {
1748 			uvmfault_unlockmaps(ufi, write_lock);
1749 			return FALSE;
1750 		}
1751 
1752 		/* reduce size if necessary */
1753 		if (ufi->entry->end - ufi->orig_rvaddr < ufi->size)
1754 			ufi->size = ufi->entry->end - ufi->orig_rvaddr;
1755 
1756 		/*
1757 		 * submap?    replace map with the submap and lookup again.
1758 		 * note: VAs in submaps must match VAs in main map.
1759 		 */
1760 		if (UVM_ET_ISSUBMAP(ufi->entry)) {
1761 			tmpmap = ufi->entry->object.sub_map;
1762 			uvmfault_unlockmaps(ufi, write_lock);
1763 			ufi->map = tmpmap;
1764 			continue;
1765 		}
1766 
1767 		/*
1768 		 * got it!
1769 		 */
1770 		ufi->mapv = ufi->map->timestamp;
1771 		return TRUE;
1772 
1773 	}	/* while loop */
1774 
1775 	/*NOTREACHED*/
1776 }
1777 
1778 /*
1779  * uvmfault_relock: attempt to relock the same version of the map
1780  *
1781  * => fault data structures should be unlocked before calling.
1782  * => if a success (TRUE) maps will be locked after call.
1783  */
1784 boolean_t
1785 uvmfault_relock(struct uvm_faultinfo *ufi)
1786 {
1787 	/*
1788 	 * ufi can be NULL when this isn't really a fault,
1789 	 * but merely paging in anon data.
1790 	 */
1791 	if (ufi == NULL) {
1792 		return TRUE;
1793 	}
1794 
1795 	counters_inc(uvmexp_counters, flt_relck);
1796 
1797 	/*
1798 	 * relock map.   fail if version mismatch (in which case nothing
1799 	 * gets locked).
1800 	 */
1801 	vm_map_lock_read(ufi->map);
1802 	if (ufi->mapv != ufi->map->timestamp) {
1803 		vm_map_unlock_read(ufi->map);
1804 		return FALSE;
1805 	}
1806 
1807 	counters_inc(uvmexp_counters, flt_relckok);
1808 	return TRUE;		/* got it! */
1809 }
1810