xref: /netbsd-src/sys/uvm/uvm_bio.c (revision b1c86f5f087524e68db12794ee9c3e3da1ab17a0)
1 /*	$NetBSD: uvm_bio.c,v 1.70 2010/06/22 18:34:50 rmind 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 1.70 2010/06/22 18:34:50 rmind 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_page: helper of ubc_fault to handle a single page.
220  *
221  * => Caller has UVM object locked.
222  */
223 
224 static inline int
225 ubc_fault_page(const struct uvm_faultinfo *ufi, const struct ubc_map *umap,
226     struct vm_page *pg, vm_prot_t prot, vm_prot_t access_type, vaddr_t va)
227 {
228 	struct uvm_object *uobj;
229 	vm_prot_t mask;
230 	int error;
231 	bool rdonly;
232 
233 	uobj = pg->uobject;
234 	KASSERT(mutex_owned(&uobj->vmobjlock));
235 
236 	if (pg->flags & PG_WANTED) {
237 		wakeup(pg);
238 	}
239 	KASSERT((pg->flags & PG_FAKE) == 0);
240 	if (pg->flags & PG_RELEASED) {
241 		mutex_enter(&uvm_pageqlock);
242 		uvm_pagefree(pg);
243 		mutex_exit(&uvm_pageqlock);
244 		return 0;
245 	}
246 	if (pg->loan_count != 0) {
247 
248 		/*
249 		 * Avoid unneeded loan break, if possible.
250 		 */
251 
252 		if ((access_type & VM_PROT_WRITE) == 0) {
253 			prot &= ~VM_PROT_WRITE;
254 		}
255 		if (prot & VM_PROT_WRITE) {
256 			struct vm_page *newpg;
257 
258 			newpg = uvm_loanbreak(pg);
259 			if (newpg == NULL) {
260 				uvm_page_unbusy(&pg, 1);
261 				return ENOMEM;
262 			}
263 			pg = newpg;
264 		}
265 	}
266 
267 	/*
268 	 * Note that a page whose backing store is partially allocated
269 	 * is marked as PG_RDONLY.
270 	 */
271 
272 	KASSERT((pg->flags & PG_RDONLY) == 0 ||
273 	    (access_type & VM_PROT_WRITE) == 0 ||
274 	    pg->offset < umap->writeoff ||
275 	    pg->offset + PAGE_SIZE > umap->writeoff + umap->writelen);
276 
277 	rdonly = ((access_type & VM_PROT_WRITE) == 0 &&
278 	    (pg->flags & PG_RDONLY) != 0) ||
279 	    UVM_OBJ_NEEDS_WRITEFAULT(uobj);
280 	mask = rdonly ? ~VM_PROT_WRITE : VM_PROT_ALL;
281 
282 	error = pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg),
283 	    prot & mask, PMAP_CANFAIL | (access_type & mask));
284 
285 	mutex_enter(&uvm_pageqlock);
286 	uvm_pageactivate(pg);
287 	mutex_exit(&uvm_pageqlock);
288 	pg->flags &= ~(PG_BUSY|PG_WANTED);
289 	UVM_PAGE_OWN(pg, NULL);
290 
291 	return error;
292 }
293 
294 /*
295  * ubc_fault: fault routine for ubc mapping
296  */
297 
298 static int
299 ubc_fault(struct uvm_faultinfo *ufi, vaddr_t ign1, struct vm_page **ign2,
300     int ign3, int ign4, vm_prot_t access_type, int flags)
301 {
302 	struct uvm_object *uobj;
303 	struct ubc_map *umap;
304 	vaddr_t va, eva, ubc_offset, slot_offset;
305 	struct vm_page *pgs[ubc_winsize >> PAGE_SHIFT];
306 	int i, error, npages;
307 	vm_prot_t prot;
308 
309 	UVMHIST_FUNC("ubc_fault"); UVMHIST_CALLED(ubchist);
310 
311 	/*
312 	 * no need to try with PGO_LOCKED...
313 	 * we don't need to have the map locked since we know that
314 	 * no one will mess with it until our reference is released.
315 	 */
316 
317 	if (flags & PGO_LOCKED) {
318 		uvmfault_unlockall(ufi, NULL, &ubc_object.uobj, NULL);
319 		flags &= ~PGO_LOCKED;
320 	}
321 
322 	va = ufi->orig_rvaddr;
323 	ubc_offset = va - (vaddr_t)ubc_object.kva;
324 	umap = &ubc_object.umap[ubc_offset >> ubc_winshift];
325 	KASSERT(umap->refcount != 0);
326 	KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
327 	slot_offset = ubc_offset & (ubc_winsize - 1);
328 
329 	/*
330 	 * some platforms cannot write to individual bytes atomically, so
331 	 * software has to do read/modify/write of larger quantities instead.
332 	 * this means that the access_type for "write" operations
333 	 * can be VM_PROT_READ, which confuses us mightily.
334 	 *
335 	 * deal with this by resetting access_type based on the info
336 	 * that ubc_alloc() stores for us.
337 	 */
338 
339 	access_type = umap->writelen ? VM_PROT_WRITE : VM_PROT_READ;
340 	UVMHIST_LOG(ubchist, "va 0x%lx ubc_offset 0x%lx access_type %d",
341 	    va, ubc_offset, access_type, 0);
342 
343 #ifdef DIAGNOSTIC
344 	if ((access_type & VM_PROT_WRITE) != 0) {
345 		if (slot_offset < trunc_page(umap->writeoff) ||
346 		    umap->writeoff + umap->writelen <= slot_offset) {
347 			panic("ubc_fault: out of range write");
348 		}
349 	}
350 #endif
351 
352 	/* no umap locking needed since we have a ref on the umap */
353 	uobj = umap->uobj;
354 
355 	if ((access_type & VM_PROT_WRITE) == 0) {
356 		npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT;
357 	} else {
358 		npages = (round_page(umap->offset + umap->writeoff +
359 		    umap->writelen) - (umap->offset + slot_offset))
360 		    >> PAGE_SHIFT;
361 		flags |= PGO_PASTEOF;
362 	}
363 
364 again:
365 	memset(pgs, 0, sizeof (pgs));
366 	mutex_enter(&uobj->vmobjlock);
367 
368 	UVMHIST_LOG(ubchist, "slot_offset 0x%x writeoff 0x%x writelen 0x%x ",
369 	    slot_offset, umap->writeoff, umap->writelen, 0);
370 	UVMHIST_LOG(ubchist, "getpages uobj %p offset 0x%x npages %d",
371 	    uobj, umap->offset + slot_offset, npages, 0);
372 
373 	error = (*uobj->pgops->pgo_get)(uobj, umap->offset + slot_offset, pgs,
374 	    &npages, 0, access_type, umap->advice, flags | PGO_NOBLOCKALLOC |
375 	    PGO_NOTIMESTAMP);
376 	UVMHIST_LOG(ubchist, "getpages error %d npages %d", error, npages, 0,
377 	    0);
378 
379 	if (error == EAGAIN) {
380 		kpause("ubc_fault", false, hz >> 2, NULL);
381 		goto again;
382 	}
383 	if (error) {
384 		return error;
385 	}
386 
387 	/*
388 	 * For virtually-indexed, virtually-tagged caches we should avoid
389 	 * creating writable mappings when we do not absolutely need them,
390 	 * since the "compatible alias" trick does not work on such caches.
391 	 * Otherwise, we can always map the pages writable.
392 	 */
393 
394 #ifdef PMAP_CACHE_VIVT
395 	prot = VM_PROT_READ | access_type;
396 #else
397 	prot = VM_PROT_READ | VM_PROT_WRITE;
398 #endif
399 
400 	/*
401 	 * Note: in the common case, all returned pages would have the same
402 	 * UVM object.  However, due to layered file-systems and e.g. tmpfs,
403 	 * returned pages may have different objects.  We "remember" the
404 	 * last object in the loop to reduce locking overhead and to perform
405 	 * pmap_update() before object unlock.
406 	 */
407 	uobj = NULL;
408 
409 	va = ufi->orig_rvaddr;
410 	eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT);
411 
412 	UVMHIST_LOG(ubchist, "va 0x%lx eva 0x%lx", va, eva, 0, 0);
413 	for (i = 0; va < eva; i++, va += PAGE_SIZE) {
414 		struct vm_page *pg;
415 
416 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i], 0, 0);
417 		pg = pgs[i];
418 
419 		if (pg == NULL || pg == PGO_DONTCARE) {
420 			continue;
421 		}
422 		if (__predict_false(pg->uobject != uobj)) {
423 			/* Check for the first iteration and error cases. */
424 			if (uobj != NULL) {
425 				/* Must make VA visible before the unlock. */
426 				pmap_update(ufi->orig_map->pmap);
427 				mutex_exit(&uobj->vmobjlock);
428 			}
429 			uobj = pg->uobject;
430 			mutex_enter(&uobj->vmobjlock);
431 		}
432 		error = ubc_fault_page(ufi, umap, pg, prot, access_type, va);
433 		if (error) {
434 			/*
435 			 * Flush (there might be pages entered), drop the lock,
436 			 * "forget" the object and perform uvm_wait().
437 			 * Note: page will re-fault.
438 			 */
439 			pmap_update(ufi->orig_map->pmap);
440 			mutex_exit(&uobj->vmobjlock);
441 			uobj = NULL;
442 			uvm_wait("ubc_fault");
443 		}
444 	}
445 	if (__predict_true(uobj != NULL)) {
446 		pmap_update(ufi->orig_map->pmap);
447 		mutex_exit(&uobj->vmobjlock);
448 	}
449 	return 0;
450 }
451 
452 /*
453  * local functions
454  */
455 
456 static struct ubc_map *
457 ubc_find_mapping(struct uvm_object *uobj, voff_t offset)
458 {
459 	struct ubc_map *umap;
460 
461 	LIST_FOREACH(umap, &ubc_object.hash[UBC_HASH(uobj, offset)], hash) {
462 		if (umap->uobj == uobj && umap->offset == offset) {
463 			return umap;
464 		}
465 	}
466 	return NULL;
467 }
468 
469 
470 /*
471  * ubc interface functions
472  */
473 
474 /*
475  * ubc_alloc:  allocate a file mapping window
476  */
477 
478 void *
479 ubc_alloc(struct uvm_object *uobj, voff_t offset, vsize_t *lenp, int advice,
480     int flags)
481 {
482 	vaddr_t slot_offset, va;
483 	struct ubc_map *umap;
484 	voff_t umap_offset;
485 	int error;
486 	UVMHIST_FUNC("ubc_alloc"); UVMHIST_CALLED(ubchist);
487 
488 	UVMHIST_LOG(ubchist, "uobj %p offset 0x%lx len 0x%lx",
489 	    uobj, offset, *lenp, 0);
490 
491 	KASSERT(*lenp > 0);
492 	umap_offset = (offset & ~((voff_t)ubc_winsize - 1));
493 	slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1));
494 	*lenp = MIN(*lenp, ubc_winsize - slot_offset);
495 
496 	/*
497 	 * the object is always locked here, so we don't need to add a ref.
498 	 */
499 
500 again:
501 	mutex_enter(&ubc_object.uobj.vmobjlock);
502 	umap = ubc_find_mapping(uobj, umap_offset);
503 	if (umap == NULL) {
504 		UBC_EVCNT_INCR(wincachemiss);
505 		umap = TAILQ_FIRST(UBC_QUEUE(offset));
506 		if (umap == NULL) {
507 			mutex_exit(&ubc_object.uobj.vmobjlock);
508 			kpause("ubc_alloc", false, hz, NULL);
509 			goto again;
510 		}
511 
512 		/*
513 		 * remove from old hash (if any), add to new hash.
514 		 */
515 
516 		if (umap->uobj != NULL) {
517 			LIST_REMOVE(umap, hash);
518 		}
519 		umap->uobj = uobj;
520 		umap->offset = umap_offset;
521 		LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)],
522 		    umap, hash);
523 		va = UBC_UMAP_ADDR(umap);
524 		if (umap->flags & UMAP_MAPPING_CACHED) {
525 			umap->flags &= ~UMAP_MAPPING_CACHED;
526 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
527 			pmap_update(pmap_kernel());
528 		}
529 	} else {
530 		UBC_EVCNT_INCR(wincachehit);
531 		va = UBC_UMAP_ADDR(umap);
532 	}
533 
534 	if (umap->refcount == 0) {
535 		TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive);
536 	}
537 
538 #ifdef DIAGNOSTIC
539 	if ((flags & UBC_WRITE) && (umap->writeoff || umap->writelen)) {
540 		panic("ubc_alloc: concurrent writes uobj %p", uobj);
541 	}
542 #endif
543 	if (flags & UBC_WRITE) {
544 		umap->writeoff = slot_offset;
545 		umap->writelen = *lenp;
546 	}
547 
548 	umap->refcount++;
549 	umap->advice = advice;
550 	mutex_exit(&ubc_object.uobj.vmobjlock);
551 	UVMHIST_LOG(ubchist, "umap %p refs %d va %p flags 0x%x",
552 	    umap, umap->refcount, va, flags);
553 
554 	if (flags & UBC_FAULTBUSY) {
555 		int npages = (*lenp + PAGE_SIZE - 1) >> PAGE_SHIFT;
556 		struct vm_page *pgs[npages];
557 		int gpflags =
558 		    PGO_SYNCIO|PGO_OVERWRITE|PGO_PASTEOF|PGO_NOBLOCKALLOC|
559 		    PGO_NOTIMESTAMP;
560 		int i;
561 		KDASSERT(flags & UBC_WRITE);
562 		KASSERT(umap->refcount == 1);
563 
564 		UBC_EVCNT_INCR(faultbusy);
565 		if (umap->flags & UMAP_MAPPING_CACHED) {
566 			umap->flags &= ~UMAP_MAPPING_CACHED;
567 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
568 		}
569 again_faultbusy:
570 		memset(pgs, 0, sizeof(pgs));
571 		mutex_enter(&uobj->vmobjlock);
572 		error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs,
573 		    &npages, 0, VM_PROT_READ | VM_PROT_WRITE, advice, gpflags);
574 		UVMHIST_LOG(ubchist, "faultbusy getpages %d", error, 0, 0, 0);
575 		if (error) {
576 			goto out;
577 		}
578 		for (i = 0; i < npages; i++) {
579 			struct vm_page *pg = pgs[i];
580 
581 			KASSERT(pg->uobject == uobj);
582 			if (pg->loan_count != 0) {
583 				mutex_enter(&uobj->vmobjlock);
584 				if (pg->loan_count != 0) {
585 					pg = uvm_loanbreak(pg);
586 				}
587 				mutex_exit(&uobj->vmobjlock);
588 				if (pg == NULL) {
589 					pmap_kremove(va, ubc_winsize);
590 					pmap_update(pmap_kernel());
591 					mutex_enter(&uobj->vmobjlock);
592 					uvm_page_unbusy(pgs, npages);
593 					mutex_exit(&uobj->vmobjlock);
594 					uvm_wait("ubc_alloc");
595 					goto again_faultbusy;
596 				}
597 				pgs[i] = pg;
598 			}
599 			pmap_kenter_pa(va + slot_offset + (i << PAGE_SHIFT),
600 			    VM_PAGE_TO_PHYS(pg),
601 			    VM_PROT_READ | VM_PROT_WRITE, 0);
602 		}
603 		pmap_update(pmap_kernel());
604 		umap->flags |= UMAP_PAGES_LOCKED;
605 	} else {
606 		KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
607 	}
608 
609 out:
610 	return (void *)(va + slot_offset);
611 }
612 
613 /*
614  * ubc_release:  free a file mapping window.
615  */
616 
617 void
618 ubc_release(void *va, int flags)
619 {
620 	struct ubc_map *umap;
621 	struct uvm_object *uobj;
622 	vaddr_t umapva;
623 	bool unmapped;
624 	UVMHIST_FUNC("ubc_release"); UVMHIST_CALLED(ubchist);
625 
626 	UVMHIST_LOG(ubchist, "va %p", va, 0, 0, 0);
627 	umap = &ubc_object.umap[((char *)va - ubc_object.kva) >> ubc_winshift];
628 	umapva = UBC_UMAP_ADDR(umap);
629 	uobj = umap->uobj;
630 	KASSERT(uobj != NULL);
631 
632 	if (umap->flags & UMAP_PAGES_LOCKED) {
633 		int slot_offset = umap->writeoff;
634 		int endoff = umap->writeoff + umap->writelen;
635 		int zerolen = round_page(endoff) - endoff;
636 		int npages = (int)(round_page(umap->writeoff + umap->writelen)
637 				   - trunc_page(umap->writeoff)) >> PAGE_SHIFT;
638 		struct vm_page *pgs[npages];
639 		paddr_t pa;
640 		int i;
641 		bool rv;
642 
643 		KASSERT((umap->flags & UMAP_MAPPING_CACHED) == 0);
644 		if (zerolen) {
645 			memset((char *)umapva + endoff, 0, zerolen);
646 		}
647 		umap->flags &= ~UMAP_PAGES_LOCKED;
648 		mutex_enter(&uvm_pageqlock);
649 		for (i = 0; i < npages; i++) {
650 			rv = pmap_extract(pmap_kernel(),
651 			    umapva + slot_offset + (i << PAGE_SHIFT), &pa);
652 			KASSERT(rv);
653 			pgs[i] = PHYS_TO_VM_PAGE(pa);
654 			pgs[i]->flags &= ~(PG_FAKE|PG_CLEAN);
655 			KASSERT(pgs[i]->loan_count == 0);
656 			uvm_pageactivate(pgs[i]);
657 		}
658 		mutex_exit(&uvm_pageqlock);
659 		pmap_kremove(umapva, ubc_winsize);
660 		pmap_update(pmap_kernel());
661 		mutex_enter(&uobj->vmobjlock);
662 		uvm_page_unbusy(pgs, npages);
663 		mutex_exit(&uobj->vmobjlock);
664 		unmapped = true;
665 	} else {
666 		unmapped = false;
667 	}
668 
669 	mutex_enter(&ubc_object.uobj.vmobjlock);
670 	umap->writeoff = 0;
671 	umap->writelen = 0;
672 	umap->refcount--;
673 	if (umap->refcount == 0) {
674 		if (flags & UBC_UNMAP) {
675 
676 			/*
677 			 * Invalidate any cached mappings if requested.
678 			 * This is typically used to avoid leaving
679 			 * incompatible cache aliases around indefinitely.
680 			 */
681 
682 			pmap_remove(pmap_kernel(), umapva,
683 				    umapva + ubc_winsize);
684 			umap->flags &= ~UMAP_MAPPING_CACHED;
685 			pmap_update(pmap_kernel());
686 			LIST_REMOVE(umap, hash);
687 			umap->uobj = NULL;
688 			TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap,
689 			    inactive);
690 		} else {
691 			if (!unmapped) {
692 				umap->flags |= UMAP_MAPPING_CACHED;
693 			}
694 			TAILQ_INSERT_TAIL(UBC_QUEUE(umap->offset), umap,
695 			    inactive);
696 		}
697 	}
698 	UVMHIST_LOG(ubchist, "umap %p refs %d", umap, umap->refcount, 0, 0);
699 	mutex_exit(&ubc_object.uobj.vmobjlock);
700 }
701 
702 /*
703  * ubc_uiomove: move data to/from an object.
704  */
705 
706 int
707 ubc_uiomove(struct uvm_object *uobj, struct uio *uio, vsize_t todo, int advice,
708     int flags)
709 {
710 	voff_t off;
711 	const bool overwrite = (flags & UBC_FAULTBUSY) != 0;
712 	int error;
713 
714 	KASSERT(todo <= uio->uio_resid);
715 	KASSERT(((flags & UBC_WRITE) != 0 && uio->uio_rw == UIO_WRITE) ||
716 	    ((flags & UBC_READ) != 0 && uio->uio_rw == UIO_READ));
717 
718 	off = uio->uio_offset;
719 	error = 0;
720 	while (todo > 0) {
721 		vsize_t bytelen = todo;
722 		void *win;
723 
724 		win = ubc_alloc(uobj, off, &bytelen, advice, flags);
725 		if (error == 0) {
726 			error = uiomove(win, bytelen, uio);
727 		}
728 		if (error != 0 && overwrite) {
729 			/*
730 			 * if we haven't initialized the pages yet,
731 			 * do it now.  it's safe to use memset here
732 			 * because we just mapped the pages above.
733 			 */
734 			printf("%s: error=%d\n", __func__, error);
735 			memset(win, 0, bytelen);
736 		}
737 		ubc_release(win, flags);
738 		off += bytelen;
739 		todo -= bytelen;
740 		if (error != 0 && (flags & UBC_PARTIALOK) != 0) {
741 			break;
742 		}
743 	}
744 
745 	return error;
746 }
747 
748 
749 /*
750  * uvm_vnp_zerorange:  set a range of bytes in a file to zero.
751  */
752 
753 void
754 uvm_vnp_zerorange(struct vnode *vp, off_t off, size_t len)
755 {
756 	void *win;
757 	int flags;
758 
759 	/*
760 	 * XXXUBC invent kzero() and use it
761 	 */
762 
763 	while (len) {
764 		vsize_t bytelen = len;
765 
766 		win = ubc_alloc(&vp->v_uobj, off, &bytelen, UVM_ADV_NORMAL,
767 		    UBC_WRITE);
768 		memset(win, 0, bytelen);
769 		flags = UBC_WANT_UNMAP(vp) ? UBC_UNMAP : 0;
770 		ubc_release(win, flags);
771 
772 		off += bytelen;
773 		len -= bytelen;
774 	}
775 }
776