xref: /netbsd-src/sys/uvm/uvm_bio.c (revision 3816d47b2c42fcd6e549e3407f842a5b1a1d23ad)
1 /*	$NetBSD: uvm_bio.c,v 1.68 2009/11/07 07:27:49 cegger 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 1.68 2009/11/07 07:27:49 cegger 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/kmem.h>
45 #include <sys/kernel.h>
46 #include <sys/proc.h>
47 #include <sys/vnode.h>
48 
49 #include <uvm/uvm.h>
50 
51 /*
52  * global data structures
53  */
54 
55 /*
56  * local functions
57  */
58 
59 static int	ubc_fault(struct uvm_faultinfo *, vaddr_t, struct vm_page **,
60 			  int, int, vm_prot_t, int);
61 static struct ubc_map *ubc_find_mapping(struct uvm_object *, voff_t);
62 
63 /*
64  * local data structues
65  */
66 
67 #define UBC_HASH(uobj, offset) 						\
68 	(((((u_long)(uobj)) >> 8) + (((u_long)(offset)) >> PAGE_SHIFT)) & \
69 				ubc_object.hashmask)
70 
71 #define UBC_QUEUE(offset)						\
72 	(&ubc_object.inactive[(((u_long)(offset)) >> ubc_winshift) &	\
73 			     (UBC_NQUEUES - 1)])
74 
75 #define UBC_UMAP_ADDR(u)						\
76 	(vaddr_t)(ubc_object.kva + (((u) - ubc_object.umap) << ubc_winshift))
77 
78 
79 #define UMAP_PAGES_LOCKED	0x0001
80 #define UMAP_MAPPING_CACHED	0x0002
81 
82 struct ubc_map
83 {
84 	struct uvm_object *	uobj;		/* mapped object */
85 	voff_t			offset;		/* offset into uobj */
86 	voff_t			writeoff;	/* write offset */
87 	vsize_t			writelen;	/* write len */
88 	int			refcount;	/* refcount on mapping */
89 	int			flags;		/* extra state */
90 	int			advice;
91 
92 	LIST_ENTRY(ubc_map)	hash;		/* hash table */
93 	TAILQ_ENTRY(ubc_map)	inactive;	/* inactive queue */
94 };
95 
96 static struct ubc_object
97 {
98 	struct uvm_object uobj;		/* glue for uvm_map() */
99 	char *kva;			/* where ubc_object is mapped */
100 	struct ubc_map *umap;		/* array of ubc_map's */
101 
102 	LIST_HEAD(, ubc_map) *hash;	/* hashtable for cached ubc_map's */
103 	u_long hashmask;		/* mask for hashtable */
104 
105 	TAILQ_HEAD(ubc_inactive_head, ubc_map) *inactive;
106 					/* inactive queues for ubc_map's */
107 
108 } ubc_object;
109 
110 const struct uvm_pagerops ubc_pager = {
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 = kmem_zalloc(ubc_nwins * sizeof(struct ubc_map),
174 	    KM_SLEEP);
175 	if (ubc_object.umap == NULL)
176 		panic("ubc_init: failed to allocate ubc_map");
177 
178 	if (ubc_winshift < PAGE_SHIFT) {
179 		ubc_winshift = PAGE_SHIFT;
180 	}
181 	va = (vaddr_t)1L;
182 #ifdef PMAP_PREFER
183 	PMAP_PREFER(0, &va, 0, 0);	/* kernel is never topdown */
184 	ubc_nqueues = va >> ubc_winshift;
185 	if (ubc_nqueues == 0) {
186 		ubc_nqueues = 1;
187 	}
188 #endif
189 	ubc_winsize = 1 << ubc_winshift;
190 	ubc_object.inactive = kmem_alloc(UBC_NQUEUES *
191 	    sizeof(struct ubc_inactive_head), KM_SLEEP);
192 	if (ubc_object.inactive == NULL)
193 		panic("ubc_init: failed to allocate inactive queue heads");
194 	for (i = 0; i < UBC_NQUEUES; i++) {
195 		TAILQ_INIT(&ubc_object.inactive[i]);
196 	}
197 	for (i = 0; i < ubc_nwins; i++) {
198 		umap = &ubc_object.umap[i];
199 		TAILQ_INSERT_TAIL(&ubc_object.inactive[i & (UBC_NQUEUES - 1)],
200 				  umap, inactive);
201 	}
202 
203 	ubc_object.hash = hashinit(ubc_nwins, HASH_LIST, true,
204 	    &ubc_object.hashmask);
205 	for (i = 0; i <= ubc_object.hashmask; i++) {
206 		LIST_INIT(&ubc_object.hash[i]);
207 	}
208 
209 	if (uvm_map(kernel_map, (vaddr_t *)&ubc_object.kva,
210 		    ubc_nwins << ubc_winshift, &ubc_object.uobj, 0, (vsize_t)va,
211 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
212 				UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) {
213 		panic("ubc_init: failed to map ubc_object");
214 	}
215 	UVMHIST_INIT(ubchist, 300);
216 }
217 
218 /*
219  * ubc_fault: fault routine for ubc mapping
220  */
221 
222 static int
223 ubc_fault(struct uvm_faultinfo *ufi, vaddr_t ign1, struct vm_page **ign2,
224     int ign3, int ign4, vm_prot_t access_type, int flags)
225 {
226 	struct uvm_object *uobj;
227 	struct ubc_map *umap;
228 	vaddr_t va, eva, ubc_offset, slot_offset;
229 	int i, error, npages;
230 	struct vm_page *pgs[ubc_winsize >> PAGE_SHIFT], *pg;
231 	vm_prot_t prot;
232 	UVMHIST_FUNC("ubc_fault");  UVMHIST_CALLED(ubchist);
233 
234 	/*
235 	 * no need to try with PGO_LOCKED...
236 	 * we don't need to have the map locked since we know that
237 	 * no one will mess with it until our reference is released.
238 	 */
239 
240 	if (flags & PGO_LOCKED) {
241 		uvmfault_unlockall(ufi, NULL, &ubc_object.uobj, NULL);
242 		flags &= ~PGO_LOCKED;
243 	}
244 
245 	va = ufi->orig_rvaddr;
246 	ubc_offset = va - (vaddr_t)ubc_object.kva;
247 	umap = &ubc_object.umap[ubc_offset >> ubc_winshift];
248 	KASSERT(umap->refcount != 0);
249 	KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
250 	slot_offset = ubc_offset & (ubc_winsize - 1);
251 
252 	/*
253 	 * some platforms cannot write to individual bytes atomically, so
254 	 * software has to do read/modify/write of larger quantities instead.
255 	 * this means that the access_type for "write" operations
256 	 * can be VM_PROT_READ, which confuses us mightily.
257 	 *
258 	 * deal with this by resetting access_type based on the info
259 	 * that ubc_alloc() stores for us.
260 	 */
261 
262 	access_type = umap->writelen ? VM_PROT_WRITE : VM_PROT_READ;
263 	UVMHIST_LOG(ubchist, "va 0x%lx ubc_offset 0x%lx access_type %d",
264 	    va, ubc_offset, access_type, 0);
265 
266 #ifdef DIAGNOSTIC
267 	if ((access_type & VM_PROT_WRITE) != 0) {
268 		if (slot_offset < trunc_page(umap->writeoff) ||
269 		    umap->writeoff + umap->writelen <= slot_offset) {
270 			panic("ubc_fault: out of range write");
271 		}
272 	}
273 #endif
274 
275 	/* no umap locking needed since we have a ref on the umap */
276 	uobj = umap->uobj;
277 
278 	if ((access_type & VM_PROT_WRITE) == 0) {
279 		npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT;
280 	} else {
281 		npages = (round_page(umap->offset + umap->writeoff +
282 		    umap->writelen) - (umap->offset + slot_offset))
283 		    >> PAGE_SHIFT;
284 		flags |= PGO_PASTEOF;
285 	}
286 
287 again:
288 	memset(pgs, 0, sizeof (pgs));
289 	mutex_enter(&uobj->vmobjlock);
290 
291 	UVMHIST_LOG(ubchist, "slot_offset 0x%x writeoff 0x%x writelen 0x%x ",
292 	    slot_offset, umap->writeoff, umap->writelen, 0);
293 	UVMHIST_LOG(ubchist, "getpages uobj %p offset 0x%x npages %d",
294 	    uobj, umap->offset + slot_offset, npages, 0);
295 
296 	error = (*uobj->pgops->pgo_get)(uobj, umap->offset + slot_offset, pgs,
297 	    &npages, 0, access_type, umap->advice, flags | PGO_NOBLOCKALLOC |
298 	    PGO_NOTIMESTAMP);
299 	UVMHIST_LOG(ubchist, "getpages error %d npages %d", error, npages, 0,
300 	    0);
301 
302 	if (error == EAGAIN) {
303 		kpause("ubc_fault", false, hz, NULL);
304 		goto again;
305 	}
306 	if (error) {
307 		return error;
308 	}
309 
310 	va = ufi->orig_rvaddr;
311 	eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT);
312 
313 	UVMHIST_LOG(ubchist, "va 0x%lx eva 0x%lx", va, eva, 0, 0);
314 	for (i = 0; va < eva; i++, va += PAGE_SIZE) {
315 		bool rdonly;
316 		vm_prot_t mask;
317 
318 		/*
319 		 * for virtually-indexed, virtually-tagged caches we should
320 		 * avoid creating writable mappings when we don't absolutely
321 		 * need them, since the "compatible alias" trick doesn't work
322 		 * on such caches.  otherwise, we can always map the pages
323 		 * writable.
324 		 */
325 
326 #ifdef PMAP_CACHE_VIVT
327 		prot = VM_PROT_READ | access_type;
328 #else
329 		prot = VM_PROT_READ | VM_PROT_WRITE;
330 #endif
331 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i], 0, 0);
332 		pg = pgs[i];
333 
334 		if (pg == NULL || pg == PGO_DONTCARE) {
335 			continue;
336 		}
337 
338 		uobj = pg->uobject;
339 		mutex_enter(&uobj->vmobjlock);
340 		if (pg->flags & PG_WANTED) {
341 			wakeup(pg);
342 		}
343 		KASSERT((pg->flags & PG_FAKE) == 0);
344 		if (pg->flags & PG_RELEASED) {
345 			mutex_enter(&uvm_pageqlock);
346 			uvm_pagefree(pg);
347 			mutex_exit(&uvm_pageqlock);
348 			mutex_exit(&uobj->vmobjlock);
349 			continue;
350 		}
351 		if (pg->loan_count != 0) {
352 
353 			/*
354 			 * avoid unneeded loan break if possible.
355 			 */
356 
357 			if ((access_type & VM_PROT_WRITE) == 0)
358 				prot &= ~VM_PROT_WRITE;
359 
360 			if (prot & VM_PROT_WRITE) {
361 				struct vm_page *newpg;
362 
363 				newpg = uvm_loanbreak(pg);
364 				if (newpg == NULL) {
365 					uvm_page_unbusy(&pg, 1);
366 					mutex_exit(&uobj->vmobjlock);
367 					uvm_wait("ubc_loanbrk");
368 					continue; /* will re-fault */
369 				}
370 				pg = newpg;
371 			}
372 		}
373 
374 		/*
375 		 * note that a page whose backing store is partially allocated
376 		 * is marked as PG_RDONLY.
377 		 */
378 
379 		rdonly = ((access_type & VM_PROT_WRITE) == 0 &&
380 		    (pg->flags & PG_RDONLY) != 0) ||
381 		    UVM_OBJ_NEEDS_WRITEFAULT(uobj);
382 		KASSERT((pg->flags & PG_RDONLY) == 0 ||
383 		    (access_type & VM_PROT_WRITE) == 0 ||
384 		    pg->offset < umap->writeoff ||
385 		    pg->offset + PAGE_SIZE > umap->writeoff + umap->writelen);
386 		mask = rdonly ? ~VM_PROT_WRITE : VM_PROT_ALL;
387 		error = pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg),
388 		    prot & mask, PMAP_CANFAIL | (access_type & mask));
389 		mutex_enter(&uvm_pageqlock);
390 		uvm_pageactivate(pg);
391 		mutex_exit(&uvm_pageqlock);
392 		pg->flags &= ~(PG_BUSY|PG_WANTED);
393 		UVM_PAGE_OWN(pg, NULL);
394 		mutex_exit(&uobj->vmobjlock);
395 		if (error) {
396 			UVMHIST_LOG(ubchist, "pmap_enter fail %d",
397 			    error, 0, 0, 0);
398 			uvm_wait("ubc_pmfail");
399 			/* will refault */
400 		}
401 	}
402 	pmap_update(ufi->orig_map->pmap);
403 	return 0;
404 }
405 
406 /*
407  * local functions
408  */
409 
410 static struct ubc_map *
411 ubc_find_mapping(struct uvm_object *uobj, voff_t offset)
412 {
413 	struct ubc_map *umap;
414 
415 	LIST_FOREACH(umap, &ubc_object.hash[UBC_HASH(uobj, offset)], hash) {
416 		if (umap->uobj == uobj && umap->offset == offset) {
417 			return umap;
418 		}
419 	}
420 	return NULL;
421 }
422 
423 
424 /*
425  * ubc interface functions
426  */
427 
428 /*
429  * ubc_alloc:  allocate a file mapping window
430  */
431 
432 void *
433 ubc_alloc(struct uvm_object *uobj, voff_t offset, vsize_t *lenp, int advice,
434     int flags)
435 {
436 	vaddr_t slot_offset, va;
437 	struct ubc_map *umap;
438 	voff_t umap_offset;
439 	int error;
440 	UVMHIST_FUNC("ubc_alloc"); UVMHIST_CALLED(ubchist);
441 
442 	UVMHIST_LOG(ubchist, "uobj %p offset 0x%lx len 0x%lx",
443 	    uobj, offset, *lenp, 0);
444 
445 	KASSERT(*lenp > 0);
446 	umap_offset = (offset & ~((voff_t)ubc_winsize - 1));
447 	slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1));
448 	*lenp = MIN(*lenp, ubc_winsize - slot_offset);
449 
450 	/*
451 	 * the object is always locked here, so we don't need to add a ref.
452 	 */
453 
454 again:
455 	mutex_enter(&ubc_object.uobj.vmobjlock);
456 	umap = ubc_find_mapping(uobj, umap_offset);
457 	if (umap == NULL) {
458 		UBC_EVCNT_INCR(wincachemiss);
459 		umap = TAILQ_FIRST(UBC_QUEUE(offset));
460 		if (umap == NULL) {
461 			mutex_exit(&ubc_object.uobj.vmobjlock);
462 			kpause("ubc_alloc", false, hz, NULL);
463 			goto again;
464 		}
465 
466 		/*
467 		 * remove from old hash (if any), add to new hash.
468 		 */
469 
470 		if (umap->uobj != NULL) {
471 			LIST_REMOVE(umap, hash);
472 		}
473 		umap->uobj = uobj;
474 		umap->offset = umap_offset;
475 		LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)],
476 		    umap, hash);
477 		va = UBC_UMAP_ADDR(umap);
478 		if (umap->flags & UMAP_MAPPING_CACHED) {
479 			umap->flags &= ~UMAP_MAPPING_CACHED;
480 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
481 			pmap_update(pmap_kernel());
482 		}
483 	} else {
484 		UBC_EVCNT_INCR(wincachehit);
485 		va = UBC_UMAP_ADDR(umap);
486 	}
487 
488 	if (umap->refcount == 0) {
489 		TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive);
490 	}
491 
492 #ifdef DIAGNOSTIC
493 	if ((flags & UBC_WRITE) && (umap->writeoff || umap->writelen)) {
494 		panic("ubc_alloc: concurrent writes uobj %p", uobj);
495 	}
496 #endif
497 	if (flags & UBC_WRITE) {
498 		umap->writeoff = slot_offset;
499 		umap->writelen = *lenp;
500 	}
501 
502 	umap->refcount++;
503 	umap->advice = advice;
504 	mutex_exit(&ubc_object.uobj.vmobjlock);
505 	UVMHIST_LOG(ubchist, "umap %p refs %d va %p flags 0x%x",
506 	    umap, umap->refcount, va, flags);
507 
508 	if (flags & UBC_FAULTBUSY) {
509 		int npages = (*lenp + PAGE_SIZE - 1) >> PAGE_SHIFT;
510 		struct vm_page *pgs[npages];
511 		int gpflags =
512 		    PGO_SYNCIO|PGO_OVERWRITE|PGO_PASTEOF|PGO_NOBLOCKALLOC|
513 		    PGO_NOTIMESTAMP;
514 		int i;
515 		KDASSERT(flags & UBC_WRITE);
516 		KASSERT(umap->refcount == 1);
517 
518 		UBC_EVCNT_INCR(faultbusy);
519 		if (umap->flags & UMAP_MAPPING_CACHED) {
520 			umap->flags &= ~UMAP_MAPPING_CACHED;
521 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
522 		}
523 again_faultbusy:
524 		memset(pgs, 0, sizeof(pgs));
525 		mutex_enter(&uobj->vmobjlock);
526 		error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs,
527 		    &npages, 0, VM_PROT_READ | VM_PROT_WRITE, advice, gpflags);
528 		UVMHIST_LOG(ubchist, "faultbusy getpages %d", error, 0, 0, 0);
529 		if (error) {
530 			goto out;
531 		}
532 		for (i = 0; i < npages; i++) {
533 			struct vm_page *pg = pgs[i];
534 
535 			KASSERT(pg->uobject == uobj);
536 			if (pg->loan_count != 0) {
537 				mutex_enter(&uobj->vmobjlock);
538 				if (pg->loan_count != 0) {
539 					pg = uvm_loanbreak(pg);
540 				}
541 				mutex_exit(&uobj->vmobjlock);
542 				if (pg == NULL) {
543 					pmap_kremove(va, ubc_winsize);
544 					pmap_update(pmap_kernel());
545 					mutex_enter(&uobj->vmobjlock);
546 					uvm_page_unbusy(pgs, npages);
547 					mutex_exit(&uobj->vmobjlock);
548 					uvm_wait("ubc_alloc");
549 					goto again_faultbusy;
550 				}
551 				pgs[i] = pg;
552 			}
553 			pmap_kenter_pa(va + slot_offset + (i << PAGE_SHIFT),
554 			    VM_PAGE_TO_PHYS(pg),
555 			    VM_PROT_READ | VM_PROT_WRITE, 0);
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 		mutex_enter(&uvm_pageqlock);
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 		mutex_exit(&uvm_pageqlock);
613 		pmap_kremove(umapva, ubc_winsize);
614 		pmap_update(pmap_kernel());
615 		mutex_enter(&uobj->vmobjlock);
616 		uvm_page_unbusy(pgs, npages);
617 		mutex_exit(&uobj->vmobjlock);
618 		unmapped = true;
619 	} else {
620 		unmapped = false;
621 	}
622 
623 	mutex_enter(&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 	mutex_exit(&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 			printf("%s: error=%d\n", __func__, error);
689 			memset(win, 0, bytelen);
690 		}
691 		ubc_release(win, flags);
692 		off += bytelen;
693 		todo -= bytelen;
694 		if (error != 0 && (flags & UBC_PARTIALOK) != 0) {
695 			break;
696 		}
697 	}
698 
699 	return error;
700 }
701 
702 
703 /*
704  * uvm_vnp_zerorange:  set a range of bytes in a file to zero.
705  */
706 
707 void
708 uvm_vnp_zerorange(struct vnode *vp, off_t off, size_t len)
709 {
710 	void *win;
711 	int flags;
712 
713 	/*
714 	 * XXXUBC invent kzero() and use it
715 	 */
716 
717 	while (len) {
718 		vsize_t bytelen = len;
719 
720 		win = ubc_alloc(&vp->v_uobj, off, &bytelen, UVM_ADV_NORMAL,
721 		    UBC_WRITE);
722 		memset(win, 0, bytelen);
723 		flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
724 		ubc_release(win, flags);
725 
726 		off += bytelen;
727 		len -= bytelen;
728 	}
729 }
730