xref: /openbsd-src/sys/uvm/uvm_fault.c (revision 25c4e8bd056e974b28f4a0ffd39d76c190a56013)
1 /*	$OpenBSD: uvm_fault.c,v 1.131 2022/06/28 10:45:55 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 
393 			/*
394 			 * if we were RELEASED during I/O, then our anon is
395 			 * no longer part of an amap.   we need to free the
396 			 * anon and try again.
397 			 */
398 			if (pg->pg_flags & PG_RELEASED) {
399 				pmap_page_protect(pg, PROT_NONE);
400 				KASSERT(anon->an_ref == 0);
401 				/*
402 				 * Released while we had unlocked amap.
403 				 */
404 				if (locked)
405 					uvmfault_unlockall(ufi, NULL, NULL);
406 				uvm_anon_release(anon);	/* frees page for us */
407 				counters_inc(uvmexp_counters, flt_pgrele);
408 				return (VM_PAGER_REFAULT);	/* refault! */
409 			}
410 
411 			if (error != VM_PAGER_OK) {
412 				KASSERT(error != VM_PAGER_PEND);
413 
414 				/* remove page from anon */
415 				anon->an_page = NULL;
416 
417 				/*
418 				 * Remove the swap slot from the anon and
419 				 * mark the anon as having no real slot.
420 				 * Do not free the swap slot, thus preventing
421 				 * it from being used again.
422 				 */
423 				uvm_swap_markbad(anon->an_swslot, 1);
424 				anon->an_swslot = SWSLOT_BAD;
425 
426 				/*
427 				 * Note: page was never !PG_BUSY, so it
428 				 * cannot be mapped and thus no need to
429 				 * pmap_page_protect() it.
430 				 */
431 				uvm_lock_pageq();
432 				uvm_pagefree(pg);
433 				uvm_unlock_pageq();
434 
435 				if (locked) {
436 					uvmfault_unlockall(ufi, NULL, NULL);
437 				}
438 				rw_exit(anon->an_lock);
439 				return (VM_PAGER_ERROR);
440 			}
441 
442 			/*
443 			 * We have successfully read the page, activate it.
444 			 */
445 			pmap_clear_modify(pg);
446 			uvm_lock_pageq();
447 			uvm_pageactivate(pg);
448 			uvm_unlock_pageq();
449 			atomic_clearbits_int(&pg->pg_flags,
450 			    PG_WANTED|PG_BUSY|PG_FAKE);
451 			UVM_PAGE_OWN(pg, NULL);
452 		}
453 
454 		/*
455 		 * We were not able to re-lock the map - restart the fault.
456 		 */
457 		if (!locked) {
458 			if (we_own) {
459 				rw_exit(anon->an_lock);
460 			}
461 			return (VM_PAGER_REFAULT);
462 		}
463 
464 		/*
465 		 * Verify that no one has touched the amap and moved
466 		 * the anon on us.
467 		 */
468 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
469 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
470 
471 			uvmfault_unlockall(ufi, amap, NULL);
472 			return (VM_PAGER_REFAULT);
473 		}
474 
475 		/*
476 		 * Retry..
477 		 */
478 		counters_inc(uvmexp_counters, flt_anretry);
479 		continue;
480 
481 	}
482 	/*NOTREACHED*/
483 }
484 
485 /*
486  * Update statistics after fault resolution.
487  * - maxrss
488  */
489 void
490 uvmfault_update_stats(struct uvm_faultinfo *ufi)
491 {
492 	struct vm_map		*map;
493 	struct proc		*p;
494 	vsize_t			 res;
495 
496 	map = ufi->orig_map;
497 
498 	/*
499 	 * If this is a nested pmap (eg, a virtual machine pmap managed
500 	 * by vmm(4) on amd64/i386), don't do any updating, just return.
501 	 *
502 	 * pmap_nested() on other archs is #defined to 0, so this is a
503 	 * no-op.
504 	 */
505 	if (pmap_nested(map->pmap))
506 		return;
507 
508 	/* Update the maxrss for the process. */
509 	if (map->flags & VM_MAP_ISVMSPACE) {
510 		p = curproc;
511 		KASSERT(p != NULL && &p->p_vmspace->vm_map == map);
512 
513 		res = pmap_resident_count(map->pmap);
514 		/* Convert res from pages to kilobytes. */
515 		res <<= (PAGE_SHIFT - 10);
516 
517 		if (p->p_ru.ru_maxrss < res)
518 			p->p_ru.ru_maxrss = res;
519 	}
520 }
521 
522 /*
523  *   F A U L T   -   m a i n   e n t r y   p o i n t
524  */
525 
526 /*
527  * uvm_fault: page fault handler
528  *
529  * => called from MD code to resolve a page fault
530  * => VM data structures usually should be unlocked.   however, it is
531  *	possible to call here with the main map locked if the caller
532  *	gets a write lock, sets it recursive, and then calls us (c.f.
533  *	uvm_map_pageable).   this should be avoided because it keeps
534  *	the map locked off during I/O.
535  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
536  */
537 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
538 			 ~PROT_WRITE : PROT_MASK)
539 struct uvm_faultctx {
540 	/*
541 	 * the following members are set up by uvm_fault_check() and
542 	 * read-only after that.
543 	 */
544 	vm_prot_t enter_prot;
545 	vm_prot_t access_type;
546 	vaddr_t startva;
547 	int npages;
548 	int centeridx;
549 	boolean_t narrow;
550 	boolean_t wired;
551 	paddr_t pa_flags;
552 };
553 
554 int		uvm_fault_check(
555 		    struct uvm_faultinfo *, struct uvm_faultctx *,
556 		    struct vm_anon ***);
557 
558 int		uvm_fault_upper(
559 		    struct uvm_faultinfo *, struct uvm_faultctx *,
560 		    struct vm_anon **, vm_fault_t);
561 boolean_t	uvm_fault_upper_lookup(
562 		    struct uvm_faultinfo *, const struct uvm_faultctx *,
563 		    struct vm_anon **, struct vm_page **);
564 
565 int		uvm_fault_lower(
566 		    struct uvm_faultinfo *, struct uvm_faultctx *,
567 		    struct vm_page **, vm_fault_t);
568 
569 int
570 uvm_fault(vm_map_t orig_map, vaddr_t vaddr, vm_fault_t fault_type,
571     vm_prot_t access_type)
572 {
573 	struct uvm_faultinfo ufi;
574 	struct uvm_faultctx flt;
575 	boolean_t shadowed;
576 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
577 	struct vm_page *pages[UVM_MAXRANGE];
578 	int error;
579 
580 	counters_inc(uvmexp_counters, faults);
581 	TRACEPOINT(uvm, fault, vaddr, fault_type, access_type, NULL);
582 
583 	/*
584 	 * init the IN parameters in the ufi
585 	 */
586 	ufi.orig_map = orig_map;
587 	ufi.orig_rvaddr = trunc_page(vaddr);
588 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
589 	if (fault_type == VM_FAULT_WIRE)
590 		flt.narrow = TRUE;	/* don't look for neighborhood
591 					 * pages on wire */
592 	else
593 		flt.narrow = FALSE;	/* normal fault */
594 	flt.access_type = access_type;
595 
596 
597 	error = ERESTART;
598 	while (error == ERESTART) { /* ReFault: */
599 		anons = anons_store;
600 
601 		error = uvm_fault_check(&ufi, &flt, &anons);
602 		if (error != 0)
603 			continue;
604 
605 		/* True if there is an anon at the faulting address */
606 		shadowed = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
607 		if (shadowed == TRUE) {
608 			/* case 1: fault on an anon in our amap */
609 			error = uvm_fault_upper(&ufi, &flt, anons, fault_type);
610 		} else {
611 			struct uvm_object *uobj = ufi.entry->object.uvm_obj;
612 
613 			/*
614 			 * if the desired page is not shadowed by the amap and
615 			 * we have a backing object, then we check to see if
616 			 * the backing object would prefer to handle the fault
617 			 * itself (rather than letting us do it with the usual
618 			 * pgo_get hook).  the backing object signals this by
619 			 * providing a pgo_fault routine.
620 			 */
621 			if (uobj != NULL && uobj->pgops->pgo_fault != NULL) {
622 				KERNEL_LOCK();
623 				rw_enter(uobj->vmobjlock, RW_WRITE);
624 				error = uobj->pgops->pgo_fault(&ufi,
625 				    flt.startva, pages, flt.npages,
626 				    flt.centeridx, fault_type, flt.access_type,
627 				    PGO_LOCKED);
628 				KERNEL_UNLOCK();
629 
630 				if (error == VM_PAGER_OK)
631 					error = 0;
632 				else if (error == VM_PAGER_REFAULT)
633 					error = ERESTART;
634 				else
635 					error = EACCES;
636 			} else {
637 				/* case 2: fault on backing obj or zero fill */
638 				error = uvm_fault_lower(&ufi, &flt, pages,
639 				    fault_type);
640 			}
641 		}
642 	}
643 
644 	return error;
645 }
646 
647 /*
648  * uvm_fault_check: check prot, handle needs-copy, etc.
649  *
650  *	1. lookup entry.
651  *	2. check protection.
652  *	3. adjust fault condition (mainly for simulated fault).
653  *	4. handle needs-copy (lazy amap copy).
654  *	5. establish range of interest for neighbor fault (aka pre-fault).
655  *	6. look up anons (if amap exists).
656  *	7. flush pages (if MADV_SEQUENTIAL)
657  *
658  * => called with nothing locked.
659  * => if we fail (result != 0) we unlock everything.
660  * => initialize/adjust many members of flt.
661  */
662 int
663 uvm_fault_check(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
664     struct vm_anon ***ranons)
665 {
666 	struct vm_amap *amap;
667 	struct uvm_object *uobj;
668 	int nback, nforw;
669 
670 	/*
671 	 * lookup and lock the maps
672 	 */
673 	if (uvmfault_lookup(ufi, FALSE) == FALSE) {
674 		return EFAULT;
675 	}
676 	/* locked: maps(read) */
677 
678 #ifdef DIAGNOSTIC
679 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0)
680 		panic("uvm_fault: fault on non-pageable map (%p, 0x%lx)",
681 		    ufi->map, ufi->orig_rvaddr);
682 #endif
683 
684 	/*
685 	 * check protection
686 	 */
687 	if ((ufi->entry->protection & flt->access_type) != flt->access_type) {
688 		uvmfault_unlockmaps(ufi, FALSE);
689 		return EACCES;
690 	}
691 
692 	/*
693 	 * "enter_prot" is the protection we want to enter the page in at.
694 	 * for certain pages (e.g. copy-on-write pages) this protection can
695 	 * be more strict than ufi->entry->protection.  "wired" means either
696 	 * the entry is wired or we are fault-wiring the pg.
697 	 */
698 
699 	flt->enter_prot = ufi->entry->protection;
700 	flt->pa_flags = UVM_ET_ISWC(ufi->entry) ? PMAP_WC : 0;
701 	flt->wired = VM_MAPENT_ISWIRED(ufi->entry) || (flt->narrow == TRUE);
702 	if (flt->wired)
703 		flt->access_type = flt->enter_prot; /* full access for wired */
704 
705 	/* handle "needs_copy" case. */
706 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
707 		if ((flt->access_type & PROT_WRITE) ||
708 		    (ufi->entry->object.uvm_obj == NULL)) {
709 			/* need to clear */
710 			uvmfault_unlockmaps(ufi, FALSE);
711 			uvmfault_amapcopy(ufi);
712 			counters_inc(uvmexp_counters, flt_amcopy);
713 			return ERESTART;
714 		} else {
715 			/*
716 			 * ensure that we pmap_enter page R/O since
717 			 * needs_copy is still true
718 			 */
719 			flt->enter_prot &= ~PROT_WRITE;
720 		}
721 	}
722 
723 	/*
724 	 * identify the players
725 	 */
726 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
727 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
728 
729 	/*
730 	 * check for a case 0 fault.  if nothing backing the entry then
731 	 * error now.
732 	 */
733 	if (amap == NULL && uobj == NULL) {
734 		uvmfault_unlockmaps(ufi, FALSE);
735 		return EFAULT;
736 	}
737 
738 	/*
739 	 * for a case 2B fault waste no time on adjacent pages because
740 	 * they are likely already entered.
741 	 */
742 	if (uobj != NULL && amap != NULL &&
743 	    (flt->access_type & PROT_WRITE) != 0) {
744 		/* wide fault (!narrow) */
745 		flt->narrow = TRUE;
746 	}
747 
748 	/*
749 	 * establish range of interest based on advice from mapper
750 	 * and then clip to fit map entry.   note that we only want
751 	 * to do this the first time through the fault.   if we
752 	 * ReFault we will disable this by setting "narrow" to true.
753 	 */
754 	if (flt->narrow == FALSE) {
755 
756 		/* wide fault (!narrow) */
757 		nback = min(uvmadvice[ufi->entry->advice].nback,
758 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
759 		flt->startva = ufi->orig_rvaddr - ((vsize_t)nback << PAGE_SHIFT);
760 		nforw = min(uvmadvice[ufi->entry->advice].nforw,
761 		    ((ufi->entry->end - ufi->orig_rvaddr) >> PAGE_SHIFT) - 1);
762 		/*
763 		 * note: "-1" because we don't want to count the
764 		 * faulting page as forw
765 		 */
766 		flt->npages = nback + nforw + 1;
767 		flt->centeridx = nback;
768 
769 		flt->narrow = TRUE;	/* ensure only once per-fault */
770 	} else {
771 		/* narrow fault! */
772 		nback = nforw = 0;
773 		flt->startva = ufi->orig_rvaddr;
774 		flt->npages = 1;
775 		flt->centeridx = 0;
776 	}
777 
778 	/*
779 	 * if we've got an amap then lock it and extract current anons.
780 	 */
781 	if (amap) {
782 		amap_lock(amap);
783 		amap_lookups(&ufi->entry->aref,
784 		    flt->startva - ufi->entry->start, *ranons, flt->npages);
785 	} else {
786 		*ranons = NULL;	/* to be safe */
787 	}
788 
789 	/*
790 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
791 	 * now and then forget about them (for the rest of the fault).
792 	 */
793 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
794 		/* flush back-page anons? */
795 		if (amap)
796 			uvmfault_anonflush(*ranons, nback);
797 
798 		/*
799 		 * flush object?
800 		 */
801 		if (uobj) {
802 			voff_t uoff;
803 
804 			uoff = (flt->startva - ufi->entry->start) + ufi->entry->offset;
805 			rw_enter(uobj->vmobjlock, RW_WRITE);
806 			(void) uobj->pgops->pgo_flush(uobj, uoff, uoff +
807 			    ((vsize_t)nback << PAGE_SHIFT), PGO_DEACTIVATE);
808 			rw_exit(uobj->vmobjlock);
809 		}
810 
811 		/* now forget about the backpages */
812 		if (amap)
813 			*ranons += nback;
814 		flt->startva += ((vsize_t)nback << PAGE_SHIFT);
815 		flt->npages -= nback;
816 		flt->centeridx = 0;
817 	}
818 
819 	return 0;
820 }
821 
822 /*
823  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
824  *
825  * iterate range of interest:
826  *	1. check if h/w mapping exists.  if yes, we don't care
827  *	2. check if anon exists.  if not, page is lower.
828  *	3. if anon exists, enter h/w mapping for neighbors.
829  *
830  * => called with amap locked (if exists).
831  */
832 boolean_t
833 uvm_fault_upper_lookup(struct uvm_faultinfo *ufi,
834     const struct uvm_faultctx *flt, struct vm_anon **anons,
835     struct vm_page **pages)
836 {
837 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
838 	struct vm_anon *anon;
839 	boolean_t shadowed;
840 	vaddr_t currva;
841 	paddr_t pa;
842 	int lcv;
843 
844 	/* locked: maps(read), amap(if there) */
845 	KASSERT(amap == NULL ||
846 	    rw_write_held(amap->am_lock));
847 
848 	/*
849 	 * map in the backpages and frontpages we found in the amap in hopes
850 	 * of preventing future faults.    we also init the pages[] array as
851 	 * we go.
852 	 */
853 	currva = flt->startva;
854 	shadowed = FALSE;
855 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
856 		/*
857 		 * dont play with VAs that are already mapped
858 		 * except for center)
859 		 */
860 		if (lcv != flt->centeridx &&
861 		    pmap_extract(ufi->orig_map->pmap, currva, &pa)) {
862 			pages[lcv] = PGO_DONTCARE;
863 			continue;
864 		}
865 
866 		/*
867 		 * unmapped or center page.   check if any anon at this level.
868 		 */
869 		if (amap == NULL || anons[lcv] == NULL) {
870 			pages[lcv] = NULL;
871 			continue;
872 		}
873 
874 		/*
875 		 * check for present page and map if possible.
876 		 */
877 		pages[lcv] = PGO_DONTCARE;
878 		if (lcv == flt->centeridx) {	/* save center for later! */
879 			shadowed = TRUE;
880 			continue;
881 		}
882 		anon = anons[lcv];
883 		KASSERT(anon->an_lock == amap->am_lock);
884 		if (anon->an_page &&
885 		    (anon->an_page->pg_flags & (PG_RELEASED|PG_BUSY)) == 0) {
886 			uvm_lock_pageq();
887 			uvm_pageactivate(anon->an_page);	/* reactivate */
888 			uvm_unlock_pageq();
889 			counters_inc(uvmexp_counters, flt_namap);
890 
891 			/*
892 			 * Since this isn't the page that's actually faulting,
893 			 * ignore pmap_enter() failures; it's not critical
894 			 * that we enter these right now.
895 			 */
896 			(void) pmap_enter(ufi->orig_map->pmap, currva,
897 			    VM_PAGE_TO_PHYS(anon->an_page) | flt->pa_flags,
898 			    (anon->an_ref > 1) ?
899 			    (flt->enter_prot & ~PROT_WRITE) : flt->enter_prot,
900 			    PMAP_CANFAIL |
901 			     (VM_MAPENT_ISWIRED(ufi->entry) ? PMAP_WIRED : 0));
902 		}
903 	}
904 	if (flt->npages > 1)
905 		pmap_update(ufi->orig_map->pmap);
906 
907 	return shadowed;
908 }
909 
910 /*
911  * uvm_fault_upper: handle upper fault.
912  *
913  *	1. acquire anon lock.
914  *	2. get anon.  let uvmfault_anonget do the dirty work.
915  *	3. if COW, promote data to new anon
916  *	4. enter h/w mapping
917  */
918 int
919 uvm_fault_upper(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
920    struct vm_anon **anons, vm_fault_t fault_type)
921 {
922 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
923 	struct vm_anon *oanon, *anon = anons[flt->centeridx];
924 	struct vm_page *pg = NULL;
925 	int error, ret;
926 
927 	/* locked: maps(read), amap, anon */
928 	KASSERT(rw_write_held(amap->am_lock));
929 	KASSERT(anon->an_lock == amap->am_lock);
930 
931 	/*
932 	 * no matter if we have case 1A or case 1B we are going to need to
933 	 * have the anon's memory resident.   ensure that now.
934 	 */
935 	/*
936 	 * let uvmfault_anonget do the dirty work.
937 	 * if it fails (!OK) it will unlock everything for us.
938 	 * if it succeeds, locks are still valid and locked.
939 	 * also, if it is OK, then the anon's page is on the queues.
940 	 * if the page is on loan from a uvm_object, then anonget will
941 	 * lock that object for us if it does not fail.
942 	 */
943 	error = uvmfault_anonget(ufi, amap, anon);
944 	switch (error) {
945 	case VM_PAGER_OK:
946 		break;
947 
948 	case VM_PAGER_REFAULT:
949 		return ERESTART;
950 
951 	case VM_PAGER_ERROR:
952 		/*
953 		 * An error occurred while trying to bring in the
954 		 * page -- this is the only error we return right
955 		 * now.
956 		 */
957 		return EACCES;	/* XXX */
958 	default:
959 #ifdef DIAGNOSTIC
960 		panic("uvm_fault: uvmfault_anonget -> %d", error);
961 #else
962 		return EACCES;
963 #endif
964 	}
965 
966 	KASSERT(rw_write_held(amap->am_lock));
967 	KASSERT(anon->an_lock == amap->am_lock);
968 
969 	/*
970 	 * if we are case 1B then we will need to allocate a new blank
971 	 * anon to transfer the data into.   note that we have a lock
972 	 * on anon, so no one can busy or release the page until we are done.
973 	 * also note that the ref count can't drop to zero here because
974 	 * it is > 1 and we are only dropping one ref.
975 	 *
976 	 * in the (hopefully very rare) case that we are out of RAM we
977 	 * will unlock, wait for more RAM, and refault.
978 	 *
979 	 * if we are out of anon VM we wait for RAM to become available.
980 	 */
981 
982 	if ((flt->access_type & PROT_WRITE) != 0 && anon->an_ref > 1) {
983 		counters_inc(uvmexp_counters, flt_acow);
984 		oanon = anon;		/* oanon = old */
985 		anon = uvm_analloc();
986 		if (anon) {
987 			anon->an_lock = amap->am_lock;
988 			pg = uvm_pagealloc(NULL, 0, anon, 0);
989 		}
990 
991 		/* check for out of RAM */
992 		if (anon == NULL || pg == NULL) {
993 			uvmfault_unlockall(ufi, amap, NULL);
994 			if (anon == NULL)
995 				counters_inc(uvmexp_counters, flt_noanon);
996 			else {
997 				anon->an_lock = NULL;
998 				anon->an_ref--;
999 				uvm_anfree(anon);
1000 				counters_inc(uvmexp_counters, flt_noram);
1001 			}
1002 
1003 			if (uvm_swapisfull())
1004 				return ENOMEM;
1005 
1006 			/* out of RAM, wait for more */
1007 			if (anon == NULL)
1008 				uvm_anwait();
1009 			else
1010 				uvm_wait("flt_noram3");
1011 			return ERESTART;
1012 		}
1013 
1014 		/* got all resources, replace anon with nanon */
1015 		uvm_pagecopy(oanon->an_page, pg);	/* pg now !PG_CLEAN */
1016 		/* un-busy! new page */
1017 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE);
1018 		UVM_PAGE_OWN(pg, NULL);
1019 		ret = amap_add(&ufi->entry->aref,
1020 		    ufi->orig_rvaddr - ufi->entry->start, anon, 1);
1021 		KASSERT(ret == 0);
1022 
1023 		/* deref: can not drop to zero here by defn! */
1024 		oanon->an_ref--;
1025 
1026 #ifndef __HAVE_PMAP_MPSAFE_ENTER_COW
1027 		/*
1028 		 * If there are multiple threads, either uvm or the
1029 		 * pmap has to make sure no threads see the old RO
1030 		 * mapping once any have seen the new RW mapping.
1031 		 * uvm does it by inserting the new mapping RO and
1032 		 * letting it fault again.
1033 		 */
1034 		if (P_HASSIBLING(curproc)) {
1035 			flt->enter_prot &= ~PROT_WRITE;
1036 			flt->access_type &= ~PROT_WRITE;
1037 		}
1038 #endif
1039 
1040 		/*
1041 		 * note: anon is _not_ locked, but we have the sole references
1042 		 * to in from amap.
1043 		 * thus, no one can get at it until we are done with it.
1044 		 */
1045 	} else {
1046 		counters_inc(uvmexp_counters, flt_anon);
1047 		oanon = anon;
1048 		pg = anon->an_page;
1049 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1050 			flt->enter_prot = flt->enter_prot & ~PROT_WRITE;
1051 	}
1052 
1053 	/*
1054 	 * now map the page in .
1055 	 */
1056 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1057 	    VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot,
1058 	    flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) {
1059 		/*
1060 		 * No need to undo what we did; we can simply think of
1061 		 * this as the pmap throwing away the mapping information.
1062 		 *
1063 		 * We do, however, have to go through the ReFault path,
1064 		 * as the map may change while we're asleep.
1065 		 */
1066 		uvmfault_unlockall(ufi, amap, NULL);
1067 		if (uvm_swapisfull()) {
1068 			/* XXX instrumentation */
1069 			return ENOMEM;
1070 		}
1071 		/* XXX instrumentation */
1072 		uvm_wait("flt_pmfail1");
1073 		return ERESTART;
1074 	}
1075 
1076 	/*
1077 	 * ... update the page queues.
1078 	 */
1079 	uvm_lock_pageq();
1080 
1081 	if (fault_type == VM_FAULT_WIRE) {
1082 		uvm_pagewire(pg);
1083 		/*
1084 		 * since the now-wired page cannot be paged out,
1085 		 * release its swap resources for others to use.
1086 		 * since an anon with no swap cannot be PG_CLEAN,
1087 		 * clear its clean flag now.
1088 		 */
1089 		atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1090 		uvm_anon_dropswap(anon);
1091 	} else {
1092 		/* activate it */
1093 		uvm_pageactivate(pg);
1094 	}
1095 
1096 	uvm_unlock_pageq();
1097 
1098 	/*
1099 	 * done case 1!  finish up by unlocking everything and returning success
1100 	 */
1101 	uvmfault_unlockall(ufi, amap, NULL);
1102 	pmap_update(ufi->orig_map->pmap);
1103 	return 0;
1104 }
1105 
1106 /*
1107  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1108  *
1109  *	1. get on-memory pages.
1110  *	2. if failed, give up (get only center page later).
1111  *	3. if succeeded, enter h/w mapping of neighbor pages.
1112  */
1113 
1114 struct vm_page *
1115 uvm_fault_lower_lookup(
1116 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1117 	struct vm_page **pages)
1118 {
1119 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1120 	struct vm_page *uobjpage = NULL;
1121 	int lcv, gotpages;
1122 	vaddr_t currva;
1123 
1124 	rw_enter(uobj->vmobjlock, RW_WRITE);
1125 
1126 	counters_inc(uvmexp_counters, flt_lget);
1127 	gotpages = flt->npages;
1128 	(void) uobj->pgops->pgo_get(uobj,
1129 	    ufi->entry->offset + (flt->startva - ufi->entry->start),
1130 	    pages, &gotpages, flt->centeridx,
1131 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1132 	    PGO_LOCKED);
1133 
1134 	/*
1135 	 * check for pages to map, if we got any
1136 	 */
1137 	if (gotpages == 0) {
1138 		return NULL;
1139 	}
1140 
1141 	currva = flt->startva;
1142 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1143 		if (pages[lcv] == NULL ||
1144 		    pages[lcv] == PGO_DONTCARE)
1145 			continue;
1146 
1147 		KASSERT((pages[lcv]->pg_flags & PG_RELEASED) == 0);
1148 
1149 		/*
1150 		 * if center page is resident and not
1151 		 * PG_BUSY, then pgo_get made it PG_BUSY
1152 		 * for us and gave us a handle to it.
1153 		 * remember this page as "uobjpage."
1154 		 * (for later use).
1155 		 */
1156 		if (lcv == flt->centeridx) {
1157 			uobjpage = pages[lcv];
1158 			continue;
1159 		}
1160 
1161 		/*
1162 		 * note: calling pgo_get with locked data
1163 		 * structures returns us pages which are
1164 		 * neither busy nor released, so we don't
1165 		 * need to check for this.   we can just
1166 		 * directly enter the page (after moving it
1167 		 * to the head of the active queue [useful?]).
1168 		 */
1169 
1170 		uvm_lock_pageq();
1171 		uvm_pageactivate(pages[lcv]);	/* reactivate */
1172 		uvm_unlock_pageq();
1173 		counters_inc(uvmexp_counters, flt_nomap);
1174 
1175 		/*
1176 		 * Since this page isn't the page that's
1177 		 * actually faulting, ignore pmap_enter()
1178 		 * failures; it's not critical that we
1179 		 * enter these right now.
1180 		 */
1181 		(void) pmap_enter(ufi->orig_map->pmap, currva,
1182 		    VM_PAGE_TO_PHYS(pages[lcv]) | flt->pa_flags,
1183 		    flt->enter_prot & MASK(ufi->entry),
1184 		    PMAP_CANFAIL |
1185 		     (flt->wired ? PMAP_WIRED : 0));
1186 
1187 		/*
1188 		 * NOTE: page can't be PG_WANTED because
1189 		 * we've held the lock the whole time
1190 		 * we've had the handle.
1191 		 */
1192 		atomic_clearbits_int(&pages[lcv]->pg_flags, PG_BUSY);
1193 		UVM_PAGE_OWN(pages[lcv], NULL);
1194 	}
1195 	pmap_update(ufi->orig_map->pmap);
1196 
1197 	return uobjpage;
1198 }
1199 
1200 /*
1201  * uvm_fault_lower: handle lower fault.
1202  *
1203  */
1204 int
1205 uvm_fault_lower(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1206    struct vm_page **pages, vm_fault_t fault_type)
1207 {
1208 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1209 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1210 	boolean_t promote, locked;
1211 	int result;
1212 	struct vm_page *uobjpage, *pg = NULL;
1213 	struct vm_anon *anon = NULL;
1214 	voff_t uoff;
1215 
1216 	/*
1217 	 * now, if the desired page is not shadowed by the amap and we have
1218 	 * a backing object that does not have a special fault routine, then
1219 	 * we ask (with pgo_get) the object for resident pages that we care
1220 	 * about and attempt to map them in.  we do not let pgo_get block
1221 	 * (PGO_LOCKED).
1222 	 */
1223 	if (uobj == NULL) {
1224 		/* zero fill; don't care neighbor pages */
1225 		uobjpage = NULL;
1226 	} else {
1227 		uobjpage = uvm_fault_lower_lookup(ufi, flt, pages);
1228 	}
1229 
1230 	/*
1231 	 * note that at this point we are done with any front or back pages.
1232 	 * we are now going to focus on the center page (i.e. the one we've
1233 	 * faulted on).  if we have faulted on the bottom (uobj)
1234 	 * layer [i.e. case 2] and the page was both present and available,
1235 	 * then we've got a pointer to it as "uobjpage" and we've already
1236 	 * made it BUSY.
1237 	 */
1238 
1239 	/*
1240 	 * locked:
1241 	 */
1242 	KASSERT(amap == NULL ||
1243 	    rw_write_held(amap->am_lock));
1244 	KASSERT(uobj == NULL ||
1245 	    rw_write_held(uobj->vmobjlock));
1246 
1247 	/*
1248 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1249 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1250 	 * have a backing object, check and see if we are going to promote
1251 	 * the data up to an anon during the fault.
1252 	 */
1253 	if (uobj == NULL) {
1254 		uobjpage = PGO_DONTCARE;
1255 		promote = TRUE;		/* always need anon here */
1256 	} else {
1257 		KASSERT(uobjpage != PGO_DONTCARE);
1258 		promote = (flt->access_type & PROT_WRITE) &&
1259 		     UVM_ET_ISCOPYONWRITE(ufi->entry);
1260 	}
1261 
1262 	/*
1263 	 * if uobjpage is not null then we do not need to do I/O to get the
1264 	 * uobjpage.
1265 	 *
1266 	 * if uobjpage is null, then we need to ask the pager to
1267 	 * get the data for us.   once we have the data, we need to reverify
1268 	 * the state the world.   we are currently not holding any resources.
1269 	 */
1270 	if (uobjpage) {
1271 		/* update rusage counters */
1272 		curproc->p_ru.ru_minflt++;
1273 	} else {
1274 		int gotpages;
1275 
1276 		/* update rusage counters */
1277 		curproc->p_ru.ru_majflt++;
1278 
1279 		uvmfault_unlockall(ufi, amap, NULL);
1280 
1281 		counters_inc(uvmexp_counters, flt_get);
1282 		gotpages = 1;
1283 		uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1284 		result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1285 		    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1286 		    PGO_SYNCIO);
1287 
1288 		/*
1289 		 * recover from I/O
1290 		 */
1291 		if (result != VM_PAGER_OK) {
1292 			KASSERT(result != VM_PAGER_PEND);
1293 
1294 			if (result == VM_PAGER_AGAIN) {
1295 				tsleep_nsec(&nowake, PVM, "fltagain2",
1296 				    MSEC_TO_NSEC(5));
1297 				return ERESTART;
1298 			}
1299 
1300 			if (!UVM_ET_ISNOFAULT(ufi->entry))
1301 				return (EIO);
1302 
1303 			uobjpage = PGO_DONTCARE;
1304 			uobj = NULL;
1305 			promote = TRUE;
1306 		}
1307 
1308 		/* re-verify the state of the world.  */
1309 		locked = uvmfault_relock(ufi);
1310 		if (locked && amap != NULL)
1311 			amap_lock(amap);
1312 
1313 		/* might be changed */
1314 		if (uobjpage != PGO_DONTCARE) {
1315 			uobj = uobjpage->uobject;
1316 			rw_enter(uobj->vmobjlock, RW_WRITE);
1317 		}
1318 
1319 		/*
1320 		 * Re-verify that amap slot is still free. if there is
1321 		 * a problem, we clean up.
1322 		 */
1323 		if (locked && amap && amap_lookup(&ufi->entry->aref,
1324 		      ufi->orig_rvaddr - ufi->entry->start)) {
1325 			if (locked)
1326 				uvmfault_unlockall(ufi, amap, NULL);
1327 			locked = FALSE;
1328 		}
1329 
1330 		/* didn't get the lock?   release the page and retry. */
1331 		if (locked == FALSE && uobjpage != PGO_DONTCARE) {
1332 			uvm_lock_pageq();
1333 			/* make sure it is in queues */
1334 			uvm_pageactivate(uobjpage);
1335 			uvm_unlock_pageq();
1336 
1337 			if (uobjpage->pg_flags & PG_WANTED)
1338 				/* still holding object lock */
1339 				wakeup(uobjpage);
1340 			atomic_clearbits_int(&uobjpage->pg_flags,
1341 			    PG_BUSY|PG_WANTED);
1342 			UVM_PAGE_OWN(uobjpage, NULL);
1343 		}
1344 
1345 		if (locked == FALSE) {
1346 			if (uobjpage != PGO_DONTCARE)
1347 				rw_exit(uobj->vmobjlock);
1348 			return ERESTART;
1349 		}
1350 
1351 		/*
1352 		 * we have the data in uobjpage which is PG_BUSY
1353 		 */
1354 	}
1355 
1356 	/*
1357 	 * notes:
1358 	 *  - at this point uobjpage can not be NULL
1359 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1360 	 */
1361 	if (promote == FALSE) {
1362 		/*
1363 		 * we are not promoting.   if the mapping is COW ensure that we
1364 		 * don't give more access than we should (e.g. when doing a read
1365 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1366 		 * R/O even though the entry protection could be R/W).
1367 		 *
1368 		 * set "pg" to the page we want to map in (uobjpage, usually)
1369 		 */
1370 		counters_inc(uvmexp_counters, flt_obj);
1371 		if (UVM_ET_ISCOPYONWRITE(ufi->entry))
1372 			flt->enter_prot &= ~PROT_WRITE;
1373 		pg = uobjpage;		/* map in the actual object */
1374 
1375 		/* assert(uobjpage != PGO_DONTCARE) */
1376 
1377 		/*
1378 		 * we are faulting directly on the page.
1379 		 */
1380 	} else {
1381 		/*
1382 		 * if we are going to promote the data to an anon we
1383 		 * allocate a blank anon here and plug it into our amap.
1384 		 */
1385 #ifdef DIAGNOSTIC
1386 		if (amap == NULL)
1387 			panic("uvm_fault: want to promote data, but no anon");
1388 #endif
1389 
1390 		anon = uvm_analloc();
1391 		if (anon) {
1392 			/*
1393 			 * In `Fill in data...' below, if
1394 			 * uobjpage == PGO_DONTCARE, we want
1395 			 * a zero'd, dirty page, so have
1396 			 * uvm_pagealloc() do that for us.
1397 			 */
1398 			anon->an_lock = amap->am_lock;
1399 			pg = uvm_pagealloc(NULL, 0, anon,
1400 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1401 		}
1402 
1403 		/*
1404 		 * out of memory resources?
1405 		 */
1406 		if (anon == NULL || pg == NULL) {
1407 			/*
1408 			 * arg!  must unbusy our page and fail or sleep.
1409 			 */
1410 			if (uobjpage != PGO_DONTCARE) {
1411 				uvm_lock_pageq();
1412 				uvm_pageactivate(uobjpage);
1413 				uvm_unlock_pageq();
1414 
1415 				if (uobjpage->pg_flags & PG_WANTED)
1416 					wakeup(uobjpage);
1417 				atomic_clearbits_int(&uobjpage->pg_flags,
1418 				    PG_BUSY|PG_WANTED);
1419 				UVM_PAGE_OWN(uobjpage, NULL);
1420 			}
1421 
1422 			/* unlock and fail ... */
1423 			uvmfault_unlockall(ufi, amap, uobj);
1424 			if (anon == NULL)
1425 				counters_inc(uvmexp_counters, flt_noanon);
1426 			else {
1427 				anon->an_lock = NULL;
1428 				anon->an_ref--;
1429 				uvm_anfree(anon);
1430 				counters_inc(uvmexp_counters, flt_noram);
1431 			}
1432 
1433 			if (uvm_swapisfull())
1434 				return (ENOMEM);
1435 
1436 			/* out of RAM, wait for more */
1437 			if (anon == NULL)
1438 				uvm_anwait();
1439 			else
1440 				uvm_wait("flt_noram5");
1441 			return ERESTART;
1442 		}
1443 
1444 		/*
1445 		 * fill in the data
1446 		 */
1447 		if (uobjpage != PGO_DONTCARE) {
1448 			counters_inc(uvmexp_counters, flt_prcopy);
1449 			/* copy page [pg now dirty] */
1450 			uvm_pagecopy(uobjpage, pg);
1451 
1452 			/*
1453 			 * promote to shared amap?  make sure all sharing
1454 			 * procs see it
1455 			 */
1456 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1457 				pmap_page_protect(uobjpage, PROT_NONE);
1458 				}
1459 
1460 			/* dispose of uobjpage. drop handle to uobj as well. */
1461 			if (uobjpage->pg_flags & PG_WANTED)
1462 				wakeup(uobjpage);
1463 			atomic_clearbits_int(&uobjpage->pg_flags,
1464 			    PG_BUSY|PG_WANTED);
1465 			UVM_PAGE_OWN(uobjpage, NULL);
1466 			uvm_lock_pageq();
1467 			uvm_pageactivate(uobjpage);
1468 			uvm_unlock_pageq();
1469 			rw_exit(uobj->vmobjlock);
1470 			uobj = NULL;
1471 		} else {
1472 			counters_inc(uvmexp_counters, flt_przero);
1473 			/*
1474 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1475 			 * above.
1476 			 */
1477 		}
1478 
1479 		if (amap_add(&ufi->entry->aref,
1480 		    ufi->orig_rvaddr - ufi->entry->start, anon, 0)) {
1481 			uvmfault_unlockall(ufi, amap, uobj);
1482 			uvm_anfree(anon);
1483 			counters_inc(uvmexp_counters, flt_noamap);
1484 
1485 			if (uvm_swapisfull())
1486 				return (ENOMEM);
1487 
1488 			amap_populate(&ufi->entry->aref,
1489 			    ufi->orig_rvaddr - ufi->entry->start);
1490 			return ERESTART;
1491 		}
1492 	}
1493 
1494 	/* note: pg is either the uobjpage or the new page in the new anon */
1495 	/*
1496 	 * all resources are present.   we can now map it in and free our
1497 	 * resources.
1498 	 */
1499 	if (amap == NULL)
1500 		KASSERT(anon == NULL);
1501 	else {
1502 		KASSERT(rw_write_held(amap->am_lock));
1503 		KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
1504 	}
1505 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1506 	    VM_PAGE_TO_PHYS(pg) | flt->pa_flags, flt->enter_prot,
1507 	    flt->access_type | PMAP_CANFAIL | (flt->wired ? PMAP_WIRED : 0)) != 0) {
1508 		/*
1509 		 * No need to undo what we did; we can simply think of
1510 		 * this as the pmap throwing away the mapping information.
1511 		 *
1512 		 * We do, however, have to go through the ReFault path,
1513 		 * as the map may change while we're asleep.
1514 		 */
1515 		if (pg->pg_flags & PG_WANTED)
1516 			wakeup(pg);
1517 
1518 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1519 		UVM_PAGE_OWN(pg, NULL);
1520 		uvmfault_unlockall(ufi, amap, uobj);
1521 		if (uvm_swapisfull()) {
1522 			/* XXX instrumentation */
1523 			return (ENOMEM);
1524 		}
1525 		/* XXX instrumentation */
1526 		uvm_wait("flt_pmfail2");
1527 		return ERESTART;
1528 	}
1529 
1530 	if (fault_type == VM_FAULT_WIRE) {
1531 		uvm_lock_pageq();
1532 		uvm_pagewire(pg);
1533 		uvm_unlock_pageq();
1534 		if (pg->pg_flags & PQ_AOBJ) {
1535 			/*
1536 			 * since the now-wired page cannot be paged out,
1537 			 * release its swap resources for others to use.
1538 			 * since an aobj page with no swap cannot be clean,
1539 			 * mark it dirty now.
1540 			 *
1541 			 * use pg->uobject here.  if the page is from a
1542 			 * tmpfs vnode, the pages are backed by its UAO and
1543 			 * not the vnode.
1544 			 */
1545 			KASSERT(uobj != NULL);
1546 			KASSERT(uobj->vmobjlock == pg->uobject->vmobjlock);
1547 			atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1548 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1549 		}
1550 	} else {
1551 		/* activate it */
1552 		uvm_lock_pageq();
1553 		uvm_pageactivate(pg);
1554 		uvm_unlock_pageq();
1555 	}
1556 
1557 	if (pg->pg_flags & PG_WANTED)
1558 		wakeup(pg);
1559 
1560 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1561 	UVM_PAGE_OWN(pg, NULL);
1562 	uvmfault_unlockall(ufi, amap, uobj);
1563 	pmap_update(ufi->orig_map->pmap);
1564 
1565 	return (0);
1566 }
1567 
1568 
1569 /*
1570  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1571  *
1572  * => map may be read-locked by caller, but MUST NOT be write-locked.
1573  * => if map is read-locked, any operations which may cause map to
1574  *	be write-locked in uvm_fault() must be taken care of by
1575  *	the caller.  See uvm_map_pageable().
1576  */
1577 int
1578 uvm_fault_wire(vm_map_t map, vaddr_t start, vaddr_t end, vm_prot_t access_type)
1579 {
1580 	vaddr_t va;
1581 	int rv;
1582 
1583 	/*
1584 	 * now fault it in a page at a time.   if the fault fails then we have
1585 	 * to undo what we have done.   note that in uvm_fault PROT_NONE
1586 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1587 	 */
1588 	for (va = start ; va < end ; va += PAGE_SIZE) {
1589 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1590 		if (rv) {
1591 			if (va != start) {
1592 				uvm_fault_unwire(map, start, va);
1593 			}
1594 			return (rv);
1595 		}
1596 	}
1597 
1598 	return (0);
1599 }
1600 
1601 /*
1602  * uvm_fault_unwire(): unwire range of virtual space.
1603  */
1604 void
1605 uvm_fault_unwire(vm_map_t map, vaddr_t start, vaddr_t end)
1606 {
1607 
1608 	vm_map_lock_read(map);
1609 	uvm_fault_unwire_locked(map, start, end);
1610 	vm_map_unlock_read(map);
1611 }
1612 
1613 /*
1614  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1615  *
1616  * => map must be at least read-locked.
1617  */
1618 void
1619 uvm_fault_unwire_locked(vm_map_t map, vaddr_t start, vaddr_t end)
1620 {
1621 	vm_map_entry_t entry, oentry = NULL, next;
1622 	pmap_t pmap = vm_map_pmap(map);
1623 	vaddr_t va;
1624 	paddr_t pa;
1625 	struct vm_page *pg;
1626 
1627 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1628 
1629 	/*
1630 	 * we assume that the area we are unwiring has actually been wired
1631 	 * in the first place.   this means that we should be able to extract
1632 	 * the PAs from the pmap.
1633 	 */
1634 
1635 	/*
1636 	 * find the beginning map entry for the region.
1637 	 */
1638 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1639 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1640 		panic("uvm_fault_unwire_locked: address not in map");
1641 
1642 	for (va = start; va < end ; va += PAGE_SIZE) {
1643 		if (pmap_extract(pmap, va, &pa) == FALSE)
1644 			continue;
1645 
1646 		/*
1647 		 * find the map entry for the current address.
1648 		 */
1649 		KASSERT(va >= entry->start);
1650 		while (entry && va >= entry->end) {
1651 			next = RBT_NEXT(uvm_map_addr, entry);
1652 			entry = next;
1653 		}
1654 
1655 		if (entry == NULL)
1656 			return;
1657 		if (va < entry->start)
1658 			continue;
1659 
1660 		/*
1661 		 * lock it.
1662 		 */
1663 		if (entry != oentry) {
1664 			if (oentry != NULL) {
1665 				uvm_map_unlock_entry(oentry);
1666 			}
1667 			uvm_map_lock_entry(entry);
1668 			oentry = entry;
1669 		}
1670 
1671 		/*
1672 		 * if the entry is no longer wired, tell the pmap.
1673 		 */
1674 		if (VM_MAPENT_ISWIRED(entry) == 0)
1675 			pmap_unwire(pmap, va);
1676 
1677 		pg = PHYS_TO_VM_PAGE(pa);
1678 		if (pg) {
1679 			uvm_lock_pageq();
1680 			uvm_pageunwire(pg);
1681 			uvm_unlock_pageq();
1682 		}
1683 	}
1684 
1685 	if (oentry != NULL) {
1686 		uvm_map_unlock_entry(oentry);
1687 	}
1688 }
1689 
1690 /*
1691  * uvmfault_unlockmaps: unlock the maps
1692  */
1693 void
1694 uvmfault_unlockmaps(struct uvm_faultinfo *ufi, boolean_t write_locked)
1695 {
1696 	/*
1697 	 * ufi can be NULL when this isn't really a fault,
1698 	 * but merely paging in anon data.
1699 	 */
1700 	if (ufi == NULL) {
1701 		return;
1702 	}
1703 
1704 	uvmfault_update_stats(ufi);
1705 	if (write_locked) {
1706 		vm_map_unlock(ufi->map);
1707 	} else {
1708 		vm_map_unlock_read(ufi->map);
1709 	}
1710 }
1711 
1712 /*
1713  * uvmfault_unlockall: unlock everything passed in.
1714  *
1715  * => maps must be read-locked (not write-locked).
1716  */
1717 void
1718 uvmfault_unlockall(struct uvm_faultinfo *ufi, struct vm_amap *amap,
1719     struct uvm_object *uobj)
1720 {
1721 	if (uobj)
1722 		rw_exit(uobj->vmobjlock);
1723 	if (amap != NULL)
1724 		amap_unlock(amap);
1725 	uvmfault_unlockmaps(ufi, FALSE);
1726 }
1727 
1728 /*
1729  * uvmfault_lookup: lookup a virtual address in a map
1730  *
1731  * => caller must provide a uvm_faultinfo structure with the IN
1732  *	params properly filled in
1733  * => we will lookup the map entry (handling submaps) as we go
1734  * => if the lookup is a success we will return with the maps locked
1735  * => if "write_lock" is TRUE, we write_lock the map, otherwise we only
1736  *	get a read lock.
1737  * => note that submaps can only appear in the kernel and they are
1738  *	required to use the same virtual addresses as the map they
1739  *	are referenced by (thus address translation between the main
1740  *	map and the submap is unnecessary).
1741  */
1742 
1743 boolean_t
1744 uvmfault_lookup(struct uvm_faultinfo *ufi, boolean_t write_lock)
1745 {
1746 	vm_map_t tmpmap;
1747 
1748 	/*
1749 	 * init ufi values for lookup.
1750 	 */
1751 	ufi->map = ufi->orig_map;
1752 	ufi->size = ufi->orig_size;
1753 
1754 	/*
1755 	 * keep going down levels until we are done.   note that there can
1756 	 * only be two levels so we won't loop very long.
1757 	 */
1758 	while (1) {
1759 		if (ufi->orig_rvaddr < ufi->map->min_offset ||
1760 		    ufi->orig_rvaddr >= ufi->map->max_offset)
1761 			return FALSE;
1762 
1763 		/* lock map */
1764 		if (write_lock) {
1765 			vm_map_lock(ufi->map);
1766 		} else {
1767 			vm_map_lock_read(ufi->map);
1768 		}
1769 
1770 		/* lookup */
1771 		if (!uvm_map_lookup_entry(ufi->map, ufi->orig_rvaddr,
1772 		    &ufi->entry)) {
1773 			uvmfault_unlockmaps(ufi, write_lock);
1774 			return FALSE;
1775 		}
1776 
1777 		/* reduce size if necessary */
1778 		if (ufi->entry->end - ufi->orig_rvaddr < ufi->size)
1779 			ufi->size = ufi->entry->end - ufi->orig_rvaddr;
1780 
1781 		/*
1782 		 * submap?    replace map with the submap and lookup again.
1783 		 * note: VAs in submaps must match VAs in main map.
1784 		 */
1785 		if (UVM_ET_ISSUBMAP(ufi->entry)) {
1786 			tmpmap = ufi->entry->object.sub_map;
1787 			uvmfault_unlockmaps(ufi, write_lock);
1788 			ufi->map = tmpmap;
1789 			continue;
1790 		}
1791 
1792 		/*
1793 		 * got it!
1794 		 */
1795 		ufi->mapv = ufi->map->timestamp;
1796 		return TRUE;
1797 
1798 	}	/* while loop */
1799 
1800 	/*NOTREACHED*/
1801 }
1802 
1803 /*
1804  * uvmfault_relock: attempt to relock the same version of the map
1805  *
1806  * => fault data structures should be unlocked before calling.
1807  * => if a success (TRUE) maps will be locked after call.
1808  */
1809 boolean_t
1810 uvmfault_relock(struct uvm_faultinfo *ufi)
1811 {
1812 	/*
1813 	 * ufi can be NULL when this isn't really a fault,
1814 	 * but merely paging in anon data.
1815 	 */
1816 	if (ufi == NULL) {
1817 		return TRUE;
1818 	}
1819 
1820 	counters_inc(uvmexp_counters, flt_relck);
1821 
1822 	/*
1823 	 * relock map.   fail if version mismatch (in which case nothing
1824 	 * gets locked).
1825 	 */
1826 	vm_map_lock_read(ufi->map);
1827 	if (ufi->mapv != ufi->map->timestamp) {
1828 		vm_map_unlock_read(ufi->map);
1829 		return FALSE;
1830 	}
1831 
1832 	counters_inc(uvmexp_counters, flt_relckok);
1833 	return TRUE;		/* got it! */
1834 }
1835