xref: /netbsd-src/sys/uvm/uvm_bio.c (revision 7fa608457b817eca6e0977b37f758ae064f3c99c)
1 /*	$NetBSD: uvm_bio.c,v 1.62 2007/07/27 09:50:37 yamt Exp $	*/
2 
3 /*
4  * Copyright (c) 1998 Chuck Silvers.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. The name of the author may not be used to endorse or promote products
16  *    derived from this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
23  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
25  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
26  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  */
31 
32 /*
33  * uvm_bio.c: buffered i/o object mapping cache
34  */
35 
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: uvm_bio.c,v 1.62 2007/07/27 09:50:37 yamt Exp $");
38 
39 #include "opt_uvmhist.h"
40 #include "opt_ubc.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/malloc.h>
45 #include <sys/kernel.h>
46 #include <sys/proc.h>
47 
48 #include <uvm/uvm.h>
49 
50 /*
51  * global data structures
52  */
53 
54 /*
55  * local functions
56  */
57 
58 static int	ubc_fault(struct uvm_faultinfo *, vaddr_t, struct vm_page **,
59 			  int, int, vm_prot_t, int);
60 static struct ubc_map *ubc_find_mapping(struct uvm_object *, voff_t);
61 
62 /*
63  * local data structues
64  */
65 
66 #define UBC_HASH(uobj, offset) 						\
67 	(((((u_long)(uobj)) >> 8) + (((u_long)(offset)) >> PAGE_SHIFT)) & \
68 				ubc_object.hashmask)
69 
70 #define UBC_QUEUE(offset)						\
71 	(&ubc_object.inactive[(((u_long)(offset)) >> ubc_winshift) &	\
72 			     (UBC_NQUEUES - 1)])
73 
74 #define UBC_UMAP_ADDR(u)						\
75 	(vaddr_t)(ubc_object.kva + (((u) - ubc_object.umap) << ubc_winshift))
76 
77 
78 #define UMAP_PAGES_LOCKED	0x0001
79 #define UMAP_MAPPING_CACHED	0x0002
80 
81 struct ubc_map
82 {
83 	struct uvm_object *	uobj;		/* mapped object */
84 	voff_t			offset;		/* offset into uobj */
85 	voff_t			writeoff;	/* write offset */
86 	vsize_t			writelen;	/* write len */
87 	int			refcount;	/* refcount on mapping */
88 	int			flags;		/* extra state */
89 	int			advice;
90 
91 	LIST_ENTRY(ubc_map)	hash;		/* hash table */
92 	TAILQ_ENTRY(ubc_map)	inactive;	/* inactive queue */
93 };
94 
95 static struct ubc_object
96 {
97 	struct uvm_object uobj;		/* glue for uvm_map() */
98 	char *kva;			/* where ubc_object is mapped */
99 	struct ubc_map *umap;		/* array of ubc_map's */
100 
101 	LIST_HEAD(, ubc_map) *hash;	/* hashtable for cached ubc_map's */
102 	u_long hashmask;		/* mask for hashtable */
103 
104 	TAILQ_HEAD(ubc_inactive_head, ubc_map) *inactive;
105 					/* inactive queues for ubc_map's */
106 
107 } ubc_object;
108 
109 struct uvm_pagerops ubc_pager =
110 {
111 	.pgo_fault = ubc_fault,
112 	/* ... rest are NULL */
113 };
114 
115 int ubc_nwins = UBC_NWINS;
116 int ubc_winshift = UBC_WINSHIFT;
117 int ubc_winsize;
118 #if defined(PMAP_PREFER)
119 int ubc_nqueues;
120 #define UBC_NQUEUES ubc_nqueues
121 #else
122 #define UBC_NQUEUES 1
123 #endif
124 
125 #if defined(UBC_STATS)
126 
127 #define	UBC_EVCNT_DEFINE(name) \
128 struct evcnt ubc_evcnt_##name = \
129 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "ubc", #name); \
130 EVCNT_ATTACH_STATIC(ubc_evcnt_##name);
131 #define	UBC_EVCNT_INCR(name) ubc_evcnt_##name.ev_count++
132 
133 #else /* defined(UBC_STATS) */
134 
135 #define	UBC_EVCNT_DEFINE(name)	/* nothing */
136 #define	UBC_EVCNT_INCR(name)	/* nothing */
137 
138 #endif /* defined(UBC_STATS) */
139 
140 UBC_EVCNT_DEFINE(wincachehit)
141 UBC_EVCNT_DEFINE(wincachemiss)
142 UBC_EVCNT_DEFINE(faultbusy)
143 
144 /*
145  * ubc_init
146  *
147  * init pager private data structures.
148  */
149 
150 void
151 ubc_init(void)
152 {
153 	struct ubc_map *umap;
154 	vaddr_t va;
155 	int i;
156 
157 	/*
158 	 * Make sure ubc_winshift is sane.
159 	 */
160 	if (ubc_winshift < PAGE_SHIFT)
161 		ubc_winshift = PAGE_SHIFT;
162 
163 	/*
164 	 * init ubc_object.
165 	 * alloc and init ubc_map's.
166 	 * init inactive queues.
167 	 * alloc and init hashtable.
168 	 * map in ubc_object.
169 	 */
170 
171 	UVM_OBJ_INIT(&ubc_object.uobj, &ubc_pager, UVM_OBJ_KERN);
172 
173 	ubc_object.umap = malloc(ubc_nwins * sizeof(struct ubc_map),
174 				 M_TEMP, M_NOWAIT);
175 	if (ubc_object.umap == NULL)
176 		panic("ubc_init: failed to allocate ubc_map");
177 	memset(ubc_object.umap, 0, ubc_nwins * sizeof(struct ubc_map));
178 
179 	if (ubc_winshift < PAGE_SHIFT) {
180 		ubc_winshift = PAGE_SHIFT;
181 	}
182 	va = (vaddr_t)1L;
183 #ifdef PMAP_PREFER
184 	PMAP_PREFER(0, &va, 0, 0);	/* kernel is never topdown */
185 	ubc_nqueues = va >> ubc_winshift;
186 	if (ubc_nqueues == 0) {
187 		ubc_nqueues = 1;
188 	}
189 #endif
190 	ubc_winsize = 1 << ubc_winshift;
191 	ubc_object.inactive = malloc(UBC_NQUEUES *
192 	    sizeof(struct ubc_inactive_head), M_TEMP, M_NOWAIT);
193 	if (ubc_object.inactive == NULL)
194 		panic("ubc_init: failed to allocate inactive queue heads");
195 	for (i = 0; i < UBC_NQUEUES; i++) {
196 		TAILQ_INIT(&ubc_object.inactive[i]);
197 	}
198 	for (i = 0; i < ubc_nwins; i++) {
199 		umap = &ubc_object.umap[i];
200 		TAILQ_INSERT_TAIL(&ubc_object.inactive[i & (UBC_NQUEUES - 1)],
201 				  umap, inactive);
202 	}
203 
204 	ubc_object.hash = hashinit(ubc_nwins, HASH_LIST, M_TEMP, M_NOWAIT,
205 				   &ubc_object.hashmask);
206 	for (i = 0; i <= ubc_object.hashmask; i++) {
207 		LIST_INIT(&ubc_object.hash[i]);
208 	}
209 
210 	if (uvm_map(kernel_map, (vaddr_t *)&ubc_object.kva,
211 		    ubc_nwins << ubc_winshift, &ubc_object.uobj, 0, (vsize_t)va,
212 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
213 				UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) {
214 		panic("ubc_init: failed to map ubc_object");
215 	}
216 	UVMHIST_INIT(ubchist, 300);
217 }
218 
219 /*
220  * ubc_fault: fault routine for ubc mapping
221  */
222 
223 static int
224 ubc_fault(struct uvm_faultinfo *ufi, vaddr_t ign1, struct vm_page **ign2,
225     int ign3, int ign4, vm_prot_t access_type, int flags)
226 {
227 	struct uvm_object *uobj;
228 	struct ubc_map *umap;
229 	vaddr_t va, eva, ubc_offset, slot_offset;
230 	int i, error, npages;
231 	struct vm_page *pgs[ubc_winsize >> PAGE_SHIFT], *pg;
232 	vm_prot_t prot;
233 	UVMHIST_FUNC("ubc_fault");  UVMHIST_CALLED(ubchist);
234 
235 	/*
236 	 * no need to try with PGO_LOCKED...
237 	 * we don't need to have the map locked since we know that
238 	 * no one will mess with it until our reference is released.
239 	 */
240 
241 	if (flags & PGO_LOCKED) {
242 		uvmfault_unlockall(ufi, NULL, &ubc_object.uobj, NULL);
243 		flags &= ~PGO_LOCKED;
244 	}
245 
246 	va = ufi->orig_rvaddr;
247 	ubc_offset = va - (vaddr_t)ubc_object.kva;
248 	umap = &ubc_object.umap[ubc_offset >> ubc_winshift];
249 	KASSERT(umap->refcount != 0);
250 	KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
251 	slot_offset = ubc_offset & (ubc_winsize - 1);
252 
253 	/*
254 	 * some platforms cannot write to individual bytes atomically, so
255 	 * software has to do read/modify/write of larger quantities instead.
256 	 * this means that the access_type for "write" operations
257 	 * can be VM_PROT_READ, which confuses us mightily.
258 	 *
259 	 * deal with this by resetting access_type based on the info
260 	 * that ubc_alloc() stores for us.
261 	 */
262 
263 	access_type = umap->writelen ? VM_PROT_WRITE : VM_PROT_READ;
264 	UVMHIST_LOG(ubchist, "va 0x%lx ubc_offset 0x%lx access_type %d",
265 	    va, ubc_offset, access_type, 0);
266 
267 #ifdef DIAGNOSTIC
268 	if ((access_type & VM_PROT_WRITE) != 0) {
269 		if (slot_offset < trunc_page(umap->writeoff) ||
270 		    umap->writeoff + umap->writelen <= slot_offset) {
271 			panic("ubc_fault: out of range write");
272 		}
273 	}
274 #endif
275 
276 	/* no umap locking needed since we have a ref on the umap */
277 	uobj = umap->uobj;
278 
279 	if ((access_type & VM_PROT_WRITE) == 0) {
280 		npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT;
281 	} else {
282 		npages = (round_page(umap->offset + umap->writeoff +
283 		    umap->writelen) - (umap->offset + slot_offset))
284 		    >> PAGE_SHIFT;
285 		flags |= PGO_PASTEOF;
286 	}
287 
288 again:
289 	memset(pgs, 0, sizeof (pgs));
290 	simple_lock(&uobj->vmobjlock);
291 
292 	UVMHIST_LOG(ubchist, "slot_offset 0x%x writeoff 0x%x writelen 0x%x ",
293 	    slot_offset, umap->writeoff, umap->writelen, 0);
294 	UVMHIST_LOG(ubchist, "getpages uobj %p offset 0x%x npages %d",
295 	    uobj, umap->offset + slot_offset, npages, 0);
296 
297 	error = (*uobj->pgops->pgo_get)(uobj, umap->offset + slot_offset, pgs,
298 	    &npages, 0, access_type, umap->advice, flags | PGO_NOBLOCKALLOC |
299 	    PGO_NOTIMESTAMP);
300 	UVMHIST_LOG(ubchist, "getpages error %d npages %d", error, npages, 0,
301 	    0);
302 
303 	if (error == EAGAIN) {
304 		kpause("ubc_fault", false, hz, NULL);
305 		goto again;
306 	}
307 	if (error) {
308 		return error;
309 	}
310 
311 	va = ufi->orig_rvaddr;
312 	eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT);
313 
314 	UVMHIST_LOG(ubchist, "va 0x%lx eva 0x%lx", va, eva, 0, 0);
315 	for (i = 0; va < eva; i++, va += PAGE_SIZE) {
316 		bool rdonly;
317 		vm_prot_t mask;
318 
319 		/*
320 		 * for virtually-indexed, virtually-tagged caches we should
321 		 * avoid creating writable mappings when we don't absolutely
322 		 * need them, since the "compatible alias" trick doesn't work
323 		 * on such caches.  otherwise, we can always map the pages
324 		 * writable.
325 		 */
326 
327 #ifdef PMAP_CACHE_VIVT
328 		prot = VM_PROT_READ | access_type;
329 #else
330 		prot = VM_PROT_READ | VM_PROT_WRITE;
331 #endif
332 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i], 0, 0);
333 		pg = pgs[i];
334 
335 		if (pg == NULL || pg == PGO_DONTCARE) {
336 			continue;
337 		}
338 
339 		uobj = pg->uobject;
340 		simple_lock(&uobj->vmobjlock);
341 		if (pg->flags & PG_WANTED) {
342 			wakeup(pg);
343 		}
344 		KASSERT((pg->flags & PG_FAKE) == 0);
345 		if (pg->flags & PG_RELEASED) {
346 			uvm_lock_pageq();
347 			uvm_pagefree(pg);
348 			uvm_unlock_pageq();
349 			simple_unlock(&uobj->vmobjlock);
350 			continue;
351 		}
352 		if (pg->loan_count != 0) {
353 
354 			/*
355 			 * avoid unneeded loan break if possible.
356 			 */
357 
358 			if ((access_type & VM_PROT_WRITE) == 0)
359 				prot &= ~VM_PROT_WRITE;
360 
361 			if (prot & VM_PROT_WRITE) {
362 				struct vm_page *newpg;
363 
364 				newpg = uvm_loanbreak(pg);
365 				if (newpg == NULL) {
366 					uvm_page_unbusy(&pg, 1);
367 					simple_unlock(&uobj->vmobjlock);
368 					uvm_wait("ubc_loanbrk");
369 					continue; /* will re-fault */
370 				}
371 				pg = newpg;
372 			}
373 		}
374 
375 		/*
376 		 * note that a page whose backing store is partially allocated
377 		 * is marked as PG_RDONLY.
378 		 */
379 
380 		rdonly = ((access_type & VM_PROT_WRITE) == 0 &&
381 		    (pg->flags & PG_RDONLY) != 0) ||
382 		    UVM_OBJ_NEEDS_WRITEFAULT(uobj);
383 		KASSERT((pg->flags & PG_RDONLY) == 0 ||
384 		    (access_type & VM_PROT_WRITE) == 0 ||
385 		    pg->offset < umap->writeoff ||
386 		    pg->offset + PAGE_SIZE > umap->writeoff + umap->writelen);
387 		mask = rdonly ? ~VM_PROT_WRITE : VM_PROT_ALL;
388 		error = pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg),
389 		    prot & mask, PMAP_CANFAIL | (access_type & mask));
390 		uvm_lock_pageq();
391 		uvm_pageactivate(pg);
392 		uvm_unlock_pageq();
393 		pg->flags &= ~(PG_BUSY|PG_WANTED);
394 		UVM_PAGE_OWN(pg, NULL);
395 		simple_unlock(&uobj->vmobjlock);
396 		if (error) {
397 			UVMHIST_LOG(ubchist, "pmap_enter fail %d",
398 			    error, 0, 0, 0);
399 			uvm_wait("ubc_pmfail");
400 			/* will refault */
401 		}
402 	}
403 	pmap_update(ufi->orig_map->pmap);
404 	return 0;
405 }
406 
407 /*
408  * local functions
409  */
410 
411 static struct ubc_map *
412 ubc_find_mapping(struct uvm_object *uobj, voff_t offset)
413 {
414 	struct ubc_map *umap;
415 
416 	LIST_FOREACH(umap, &ubc_object.hash[UBC_HASH(uobj, offset)], hash) {
417 		if (umap->uobj == uobj && umap->offset == offset) {
418 			return umap;
419 		}
420 	}
421 	return NULL;
422 }
423 
424 
425 /*
426  * ubc interface functions
427  */
428 
429 /*
430  * ubc_alloc:  allocate a file mapping window
431  */
432 
433 void *
434 ubc_alloc(struct uvm_object *uobj, voff_t offset, vsize_t *lenp, int advice,
435     int flags)
436 {
437 	vaddr_t slot_offset, va;
438 	struct ubc_map *umap;
439 	voff_t umap_offset;
440 	int error;
441 	UVMHIST_FUNC("ubc_alloc"); UVMHIST_CALLED(ubchist);
442 
443 	UVMHIST_LOG(ubchist, "uobj %p offset 0x%lx len 0x%lx",
444 	    uobj, offset, *lenp, 0);
445 
446 	KASSERT(*lenp > 0);
447 	umap_offset = (offset & ~((voff_t)ubc_winsize - 1));
448 	slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1));
449 	*lenp = MIN(*lenp, ubc_winsize - slot_offset);
450 
451 	/*
452 	 * the object is always locked here, so we don't need to add a ref.
453 	 */
454 
455 again:
456 	simple_lock(&ubc_object.uobj.vmobjlock);
457 	umap = ubc_find_mapping(uobj, umap_offset);
458 	if (umap == NULL) {
459 		UBC_EVCNT_INCR(wincachemiss);
460 		umap = TAILQ_FIRST(UBC_QUEUE(offset));
461 		if (umap == NULL) {
462 			simple_unlock(&ubc_object.uobj.vmobjlock);
463 			kpause("ubc_alloc", false, hz, NULL);
464 			goto again;
465 		}
466 
467 		/*
468 		 * remove from old hash (if any), add to new hash.
469 		 */
470 
471 		if (umap->uobj != NULL) {
472 			LIST_REMOVE(umap, hash);
473 		}
474 		umap->uobj = uobj;
475 		umap->offset = umap_offset;
476 		LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)],
477 		    umap, hash);
478 		va = UBC_UMAP_ADDR(umap);
479 		if (umap->flags & UMAP_MAPPING_CACHED) {
480 			umap->flags &= ~UMAP_MAPPING_CACHED;
481 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
482 			pmap_update(pmap_kernel());
483 		}
484 	} else {
485 		UBC_EVCNT_INCR(wincachehit);
486 		va = UBC_UMAP_ADDR(umap);
487 	}
488 
489 	if (umap->refcount == 0) {
490 		TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive);
491 	}
492 
493 #ifdef DIAGNOSTIC
494 	if ((flags & UBC_WRITE) && (umap->writeoff || umap->writelen)) {
495 		panic("ubc_alloc: concurrent writes uobj %p", uobj);
496 	}
497 #endif
498 	if (flags & UBC_WRITE) {
499 		umap->writeoff = slot_offset;
500 		umap->writelen = *lenp;
501 	}
502 
503 	umap->refcount++;
504 	umap->advice = advice;
505 	simple_unlock(&ubc_object.uobj.vmobjlock);
506 	UVMHIST_LOG(ubchist, "umap %p refs %d va %p flags 0x%x",
507 	    umap, umap->refcount, va, flags);
508 
509 	if (flags & UBC_FAULTBUSY) {
510 		int npages = (*lenp + PAGE_SIZE - 1) >> PAGE_SHIFT;
511 		struct vm_page *pgs[npages];
512 		int gpflags =
513 		    PGO_SYNCIO|PGO_OVERWRITE|PGO_PASTEOF|PGO_NOBLOCKALLOC|
514 		    PGO_NOTIMESTAMP;
515 		int i;
516 		KDASSERT(flags & UBC_WRITE);
517 		KASSERT(umap->refcount == 1);
518 
519 		UBC_EVCNT_INCR(faultbusy);
520 		if (umap->flags & UMAP_MAPPING_CACHED) {
521 			umap->flags &= ~UMAP_MAPPING_CACHED;
522 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
523 		}
524 again_faultbusy:
525 		memset(pgs, 0, sizeof(pgs));
526 		simple_lock(&uobj->vmobjlock);
527 		error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs,
528 		    &npages, 0, VM_PROT_READ | VM_PROT_WRITE, advice, gpflags);
529 		UVMHIST_LOG(ubchist, "faultbusy getpages %d", error, 0, 0, 0);
530 		if (error) {
531 			goto out;
532 		}
533 		for (i = 0; i < npages; i++) {
534 			struct vm_page *pg = pgs[i];
535 
536 			KASSERT(pg->uobject == uobj);
537 			if (pg->loan_count != 0) {
538 				simple_lock(&uobj->vmobjlock);
539 				if (pg->loan_count != 0) {
540 					pg = uvm_loanbreak(pg);
541 				}
542 				simple_unlock(&uobj->vmobjlock);
543 				if (pg == NULL) {
544 					pmap_kremove(va, ubc_winsize);
545 					pmap_update(pmap_kernel());
546 					simple_lock(&uobj->vmobjlock);
547 					uvm_page_unbusy(pgs, npages);
548 					simple_unlock(&uobj->vmobjlock);
549 					uvm_wait("ubc_alloc");
550 					goto again_faultbusy;
551 				}
552 				pgs[i] = pg;
553 			}
554 			pmap_kenter_pa(va + slot_offset + (i << PAGE_SHIFT),
555 			    VM_PAGE_TO_PHYS(pg), VM_PROT_READ | VM_PROT_WRITE);
556 		}
557 		pmap_update(pmap_kernel());
558 		umap->flags |= UMAP_PAGES_LOCKED;
559 	} else {
560 		KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
561 	}
562 
563 out:
564 	return (void *)(va + slot_offset);
565 }
566 
567 /*
568  * ubc_release:  free a file mapping window.
569  */
570 
571 void
572 ubc_release(void *va, int flags)
573 {
574 	struct ubc_map *umap;
575 	struct uvm_object *uobj;
576 	vaddr_t umapva;
577 	bool unmapped;
578 	UVMHIST_FUNC("ubc_release"); UVMHIST_CALLED(ubchist);
579 
580 	UVMHIST_LOG(ubchist, "va %p", va, 0, 0, 0);
581 	umap = &ubc_object.umap[((char *)va - ubc_object.kva) >> ubc_winshift];
582 	umapva = UBC_UMAP_ADDR(umap);
583 	uobj = umap->uobj;
584 	KASSERT(uobj != NULL);
585 
586 	if (umap->flags & UMAP_PAGES_LOCKED) {
587 		int slot_offset = umap->writeoff;
588 		int endoff = umap->writeoff + umap->writelen;
589 		int zerolen = round_page(endoff) - endoff;
590 		int npages = (int)(round_page(umap->writeoff + umap->writelen)
591 				   - trunc_page(umap->writeoff)) >> PAGE_SHIFT;
592 		struct vm_page *pgs[npages];
593 		paddr_t pa;
594 		int i;
595 		bool rv;
596 
597 		KASSERT((umap->flags & UMAP_MAPPING_CACHED) == 0);
598 		if (zerolen) {
599 			memset((char *)umapva + endoff, 0, zerolen);
600 		}
601 		umap->flags &= ~UMAP_PAGES_LOCKED;
602 		uvm_lock_pageq();
603 		for (i = 0; i < npages; i++) {
604 			rv = pmap_extract(pmap_kernel(),
605 			    umapva + slot_offset + (i << PAGE_SHIFT), &pa);
606 			KASSERT(rv);
607 			pgs[i] = PHYS_TO_VM_PAGE(pa);
608 			pgs[i]->flags &= ~(PG_FAKE|PG_CLEAN);
609 			KASSERT(pgs[i]->loan_count == 0);
610 			uvm_pageactivate(pgs[i]);
611 		}
612 		uvm_unlock_pageq();
613 		pmap_kremove(umapva, ubc_winsize);
614 		pmap_update(pmap_kernel());
615 		simple_lock(&uobj->vmobjlock);
616 		uvm_page_unbusy(pgs, npages);
617 		simple_unlock(&uobj->vmobjlock);
618 		unmapped = true;
619 	} else {
620 		unmapped = false;
621 	}
622 
623 	simple_lock(&ubc_object.uobj.vmobjlock);
624 	umap->writeoff = 0;
625 	umap->writelen = 0;
626 	umap->refcount--;
627 	if (umap->refcount == 0) {
628 		if (flags & UBC_UNMAP) {
629 
630 			/*
631 			 * Invalidate any cached mappings if requested.
632 			 * This is typically used to avoid leaving
633 			 * incompatible cache aliases around indefinitely.
634 			 */
635 
636 			pmap_remove(pmap_kernel(), umapva,
637 				    umapva + ubc_winsize);
638 			umap->flags &= ~UMAP_MAPPING_CACHED;
639 			pmap_update(pmap_kernel());
640 			LIST_REMOVE(umap, hash);
641 			umap->uobj = NULL;
642 			TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap,
643 			    inactive);
644 		} else {
645 			if (!unmapped) {
646 				umap->flags |= UMAP_MAPPING_CACHED;
647 			}
648 			TAILQ_INSERT_TAIL(UBC_QUEUE(umap->offset), umap,
649 			    inactive);
650 		}
651 	}
652 	UVMHIST_LOG(ubchist, "umap %p refs %d", umap, umap->refcount, 0, 0);
653 	simple_unlock(&ubc_object.uobj.vmobjlock);
654 }
655 
656 /*
657  * ubc_uiomove: move data to/from an object.
658  */
659 
660 int
661 ubc_uiomove(struct uvm_object *uobj, struct uio *uio, vsize_t todo, int advice,
662     int flags)
663 {
664 	voff_t off;
665 	const bool overwrite = (flags & UBC_FAULTBUSY) != 0;
666 	int error;
667 
668 	KASSERT(todo <= uio->uio_resid);
669 	KASSERT(((flags & UBC_WRITE) != 0 && uio->uio_rw == UIO_WRITE) ||
670 	    ((flags & UBC_READ) != 0 && uio->uio_rw == UIO_READ));
671 
672 	off = uio->uio_offset;
673 	error = 0;
674 	while (todo > 0) {
675 		vsize_t bytelen = todo;
676 		void *win;
677 
678 		win = ubc_alloc(uobj, off, &bytelen, advice, flags);
679 		if (error == 0) {
680 			error = uiomove(win, bytelen, uio);
681 		}
682 		if (error != 0 && overwrite) {
683 			/*
684 			 * if we haven't initialized the pages yet,
685 			 * do it now.  it's safe to use memset here
686 			 * because we just mapped the pages above.
687 			 */
688 			memset(win, 0, bytelen);
689 		}
690 		ubc_release(win, flags);
691 		off += bytelen;
692 		todo -= bytelen;
693 		if (error != 0 && (flags & UBC_PARTIALOK) != 0) {
694 			break;
695 		}
696 	}
697 
698 	return error;
699 }
700 
701 #if 0 /* notused */
702 /*
703  * removing a range of mappings from the ubc mapping cache.
704  */
705 
706 void
707 ubc_flush(struct uvm_object *uobj, voff_t start, voff_t end)
708 {
709 	struct ubc_map *umap;
710 	vaddr_t va;
711 	UVMHIST_FUNC("ubc_flush");  UVMHIST_CALLED(ubchist);
712 
713 	UVMHIST_LOG(ubchist, "uobj %p start 0x%lx end 0x%lx",
714 		    uobj, start, end, 0);
715 
716 	simple_lock(&ubc_object.uobj.vmobjlock);
717 	for (umap = ubc_object.umap;
718 	     umap < &ubc_object.umap[ubc_nwins];
719 	     umap++) {
720 
721 		if (umap->uobj != uobj || umap->offset < start ||
722 		    (umap->offset >= end && end != 0) ||
723 		    umap->refcount > 0) {
724 			continue;
725 		}
726 
727 		/*
728 		 * remove from hash,
729 		 * move to head of inactive queue.
730 		 */
731 
732 		va = (vaddr_t)(ubc_object.kva +
733 		    ((umap - ubc_object.umap) << ubc_winshift));
734 		pmap_remove(pmap_kernel(), va, va + ubc_winsize);
735 
736 		LIST_REMOVE(umap, hash);
737 		umap->uobj = NULL;
738 		TAILQ_REMOVE(UBC_QUEUE(umap->offset), umap, inactive);
739 		TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap, inactive);
740 	}
741 	pmap_update(pmap_kernel());
742 	simple_unlock(&ubc_object.uobj.vmobjlock);
743 }
744 #endif /* notused */
745