xref: /freebsd/sys/vm/vnode_pager.c (revision ce4946daa5ce852d28008dac492029500ab2ee95)
1 /*
2  * Copyright (c) 1990 University of Utah.
3  * Copyright (c) 1991 The Regents of the University of California.
4  * All rights reserved.
5  * Copyright (c) 1993, 1994 John S. Dyson
6  * Copyright (c) 1995, David Greenman
7  *
8  * This code is derived from software contributed to Berkeley by
9  * the Systems Programming Group of the University of Utah Computer
10  * Science Department.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	from: @(#)vnode_pager.c	7.5 (Berkeley) 4/20/91
41  * $FreeBSD$
42  */
43 
44 /*
45  * Page to/from files (vnodes).
46  */
47 
48 /*
49  * TODO:
50  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
51  *	greatly re-simplify the vnode_pager.
52  */
53 
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/proc.h>
57 #include <sys/vnode.h>
58 #include <sys/mount.h>
59 #include <sys/bio.h>
60 #include <sys/buf.h>
61 #include <sys/vmmeter.h>
62 #include <sys/conf.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_pager.h>
68 #include <vm/vm_map.h>
69 #include <vm/vnode_pager.h>
70 #include <vm/vm_extern.h>
71 
72 static vm_offset_t vnode_pager_addr __P((struct vnode *vp, vm_ooffset_t address,
73 					 int *run));
74 static void vnode_pager_iodone __P((struct buf *bp));
75 static int vnode_pager_input_smlfs __P((vm_object_t object, vm_page_t m));
76 static int vnode_pager_input_old __P((vm_object_t object, vm_page_t m));
77 static void vnode_pager_dealloc __P((vm_object_t));
78 static int vnode_pager_getpages __P((vm_object_t, vm_page_t *, int, int));
79 static void vnode_pager_putpages __P((vm_object_t, vm_page_t *, int, boolean_t, int *));
80 static boolean_t vnode_pager_haspage __P((vm_object_t, vm_pindex_t, int *, int *));
81 
82 struct pagerops vnodepagerops = {
83 	NULL,
84 	vnode_pager_alloc,
85 	vnode_pager_dealloc,
86 	vnode_pager_getpages,
87 	vnode_pager_putpages,
88 	vnode_pager_haspage,
89 	NULL
90 };
91 
92 int vnode_pbuf_freecnt = -1;	/* start out unlimited */
93 
94 
95 /*
96  * Allocate (or lookup) pager for a vnode.
97  * Handle is a vnode pointer.
98  */
99 vm_object_t
100 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
101 		  vm_ooffset_t offset)
102 {
103 	vm_object_t object;
104 	struct vnode *vp;
105 
106 	/*
107 	 * Pageout to vnode, no can do yet.
108 	 */
109 	if (handle == NULL)
110 		return (NULL);
111 
112 	/*
113 	 * XXX hack - This initialization should be put somewhere else.
114 	 */
115 	if (vnode_pbuf_freecnt < 0) {
116 	    vnode_pbuf_freecnt = nswbuf / 2 + 1;
117 	}
118 
119 	vp = (struct vnode *) handle;
120 
121 	/*
122 	 * Prevent race condition when allocating the object. This
123 	 * can happen with NFS vnodes since the nfsnode isn't locked.
124 	 */
125 	while (vp->v_flag & VOLOCK) {
126 		vp->v_flag |= VOWANT;
127 		tsleep(vp, PVM, "vnpobj", 0);
128 	}
129 	vp->v_flag |= VOLOCK;
130 
131 	/*
132 	 * If the object is being terminated, wait for it to
133 	 * go away.
134 	 */
135 	while (((object = vp->v_object) != NULL) &&
136 		(object->flags & OBJ_DEAD)) {
137 		tsleep(object, PVM, "vadead", 0);
138 	}
139 
140 	if (vp->v_usecount == 0)
141 		panic("vnode_pager_alloc: no vnode reference");
142 
143 	if (object == NULL) {
144 		/*
145 		 * And an object of the appropriate size
146 		 */
147 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
148 		object->flags = 0;
149 
150 		object->un_pager.vnp.vnp_size = size;
151 
152 		object->handle = handle;
153 		vp->v_object = object;
154 		vp->v_usecount++;
155 	} else {
156 		object->ref_count++;
157 		vp->v_usecount++;
158 	}
159 
160 	vp->v_flag &= ~VOLOCK;
161 	if (vp->v_flag & VOWANT) {
162 		vp->v_flag &= ~VOWANT;
163 		wakeup(vp);
164 	}
165 	return (object);
166 }
167 
168 static void
169 vnode_pager_dealloc(object)
170 	vm_object_t object;
171 {
172 	register struct vnode *vp = object->handle;
173 
174 	if (vp == NULL)
175 		panic("vnode_pager_dealloc: pager already dealloced");
176 
177 	vm_object_pip_wait(object, "vnpdea");
178 
179 	object->handle = NULL;
180 	object->type = OBJT_DEAD;
181 	vp->v_object = NULL;
182 	vp->v_flag &= ~(VTEXT | VOBJBUF);
183 }
184 
185 static boolean_t
186 vnode_pager_haspage(object, pindex, before, after)
187 	vm_object_t object;
188 	vm_pindex_t pindex;
189 	int *before;
190 	int *after;
191 {
192 	struct vnode *vp = object->handle;
193 	daddr_t bn;
194 	int err;
195 	daddr_t reqblock;
196 	int poff;
197 	int bsize;
198 	int pagesperblock, blocksperpage;
199 
200 	/*
201 	 * If no vp or vp is doomed or marked transparent to VM, we do not
202 	 * have the page.
203 	 */
204 	if ((vp == NULL) || (vp->v_flag & VDOOMED))
205 		return FALSE;
206 
207 	/*
208 	 * If filesystem no longer mounted or offset beyond end of file we do
209 	 * not have the page.
210 	 */
211 	if ((vp->v_mount == NULL) ||
212 		(IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
213 		return FALSE;
214 
215 	bsize = vp->v_mount->mnt_stat.f_iosize;
216 	pagesperblock = bsize / PAGE_SIZE;
217 	blocksperpage = 0;
218 	if (pagesperblock > 0) {
219 		reqblock = pindex / pagesperblock;
220 	} else {
221 		blocksperpage = (PAGE_SIZE / bsize);
222 		reqblock = pindex * blocksperpage;
223 	}
224 	err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn,
225 		after, before);
226 	if (err)
227 		return TRUE;
228 	if ( bn == -1)
229 		return FALSE;
230 	if (pagesperblock > 0) {
231 		poff = pindex - (reqblock * pagesperblock);
232 		if (before) {
233 			*before *= pagesperblock;
234 			*before += poff;
235 		}
236 		if (after) {
237 			int numafter;
238 			*after *= pagesperblock;
239 			numafter = pagesperblock - (poff + 1);
240 			if (IDX_TO_OFF(pindex + numafter) > object->un_pager.vnp.vnp_size) {
241 				numafter = OFF_TO_IDX((object->un_pager.vnp.vnp_size - IDX_TO_OFF(pindex)));
242 			}
243 			*after += numafter;
244 		}
245 	} else {
246 		if (before) {
247 			*before /= blocksperpage;
248 		}
249 
250 		if (after) {
251 			*after /= blocksperpage;
252 		}
253 	}
254 	return TRUE;
255 }
256 
257 /*
258  * Lets the VM system know about a change in size for a file.
259  * We adjust our own internal size and flush any cached pages in
260  * the associated object that are affected by the size change.
261  *
262  * Note: this routine may be invoked as a result of a pager put
263  * operation (possibly at object termination time), so we must be careful.
264  */
265 void
266 vnode_pager_setsize(vp, nsize)
267 	struct vnode *vp;
268 	vm_ooffset_t nsize;
269 {
270 	vm_pindex_t nobjsize;
271 	vm_object_t object = vp->v_object;
272 
273 	if (object == NULL)
274 		return;
275 
276 	/*
277 	 * Hasn't changed size
278 	 */
279 	if (nsize == object->un_pager.vnp.vnp_size)
280 		return;
281 
282 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
283 
284 	/*
285 	 * File has shrunk. Toss any cached pages beyond the new EOF.
286 	 */
287 	if (nsize < object->un_pager.vnp.vnp_size) {
288 		vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size);
289 		if (nobjsize < object->size) {
290 			vm_object_page_remove(object, nobjsize, object->size,
291 				FALSE);
292 		}
293 		/*
294 		 * this gets rid of garbage at the end of a page that is now
295 		 * only partially backed by the vnode...
296 		 */
297 		if (nsize & PAGE_MASK) {
298 			vm_offset_t kva;
299 			vm_page_t m;
300 
301 			m = vm_page_lookup(object, OFF_TO_IDX(nsize));
302 			if (m) {
303 				int base = (int)nsize & PAGE_MASK;
304 				int size = PAGE_SIZE - base;
305 
306 				/*
307 				 * Clear out partial-page garbage in case
308 				 * the page has been mapped.
309 				 */
310 				kva = vm_pager_map_page(m);
311 				bzero((caddr_t)kva + base, size);
312 				vm_pager_unmap_page(kva);
313 
314 				/*
315 				 * Clear out partial-page dirty bits.  This
316 				 * has the side effect of setting the valid
317 				 * bits, but that is ok.  There are a bunch
318 				 * of places in the VM system where we expected
319 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
320 				 * case is one of them.  If the page is still
321 				 * partially dirty, make it fully dirty.
322 				 */
323 				vm_page_set_validclean(m, base, size);
324 				if (m->dirty != 0)
325 					m->dirty = VM_PAGE_BITS_ALL;
326 			}
327 		}
328 	}
329 	object->un_pager.vnp.vnp_size = nsize;
330 	object->size = nobjsize;
331 }
332 
333 /*
334  * calculate the linear (byte) disk address of specified virtual
335  * file address
336  */
337 static vm_offset_t
338 vnode_pager_addr(vp, address, run)
339 	struct vnode *vp;
340 	vm_ooffset_t address;
341 	int *run;
342 {
343 	int rtaddress;
344 	int bsize;
345 	daddr_t block;
346 	struct vnode *rtvp;
347 	int err;
348 	daddr_t vblock;
349 	int voffset;
350 
351 	if ((int) address < 0)
352 		return -1;
353 
354 	if (vp->v_mount == NULL)
355 		return -1;
356 
357 	bsize = vp->v_mount->mnt_stat.f_iosize;
358 	vblock = address / bsize;
359 	voffset = address % bsize;
360 
361 	err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL);
362 
363 	if (err || (block == -1))
364 		rtaddress = -1;
365 	else {
366 		rtaddress = block + voffset / DEV_BSIZE;
367 		if( run) {
368 			*run += 1;
369 			*run *= bsize/PAGE_SIZE;
370 			*run -= voffset/PAGE_SIZE;
371 		}
372 	}
373 
374 	return rtaddress;
375 }
376 
377 /*
378  * interrupt routine for I/O completion
379  */
380 static void
381 vnode_pager_iodone(bp)
382 	struct buf *bp;
383 {
384 	bp->b_flags |= B_DONE;
385 	wakeup(bp);
386 }
387 
388 /*
389  * small block file system vnode pager input
390  */
391 static int
392 vnode_pager_input_smlfs(object, m)
393 	vm_object_t object;
394 	vm_page_t m;
395 {
396 	int i;
397 	int s;
398 	struct vnode *dp, *vp;
399 	struct buf *bp;
400 	vm_offset_t kva;
401 	int fileaddr;
402 	vm_offset_t bsize;
403 	int error = 0;
404 
405 	vp = object->handle;
406 	if (vp->v_mount == NULL)
407 		return VM_PAGER_BAD;
408 
409 	bsize = vp->v_mount->mnt_stat.f_iosize;
410 
411 
412 	VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
413 
414 	kva = vm_pager_map_page(m);
415 
416 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
417 
418 		if (vm_page_bits(i * bsize, bsize) & m->valid)
419 			continue;
420 
421 		fileaddr = vnode_pager_addr(vp,
422 			IDX_TO_OFF(m->pindex) + i * bsize, (int *)0);
423 		if (fileaddr != -1) {
424 			bp = getpbuf(&vnode_pbuf_freecnt);
425 
426 			/* build a minimal buffer header */
427 			bp->b_iocmd = BIO_READ;
428 			bp->b_iodone = vnode_pager_iodone;
429 			bp->b_rcred = bp->b_wcred = curproc->p_ucred;
430 			if (bp->b_rcred != NOCRED)
431 				crhold(bp->b_rcred);
432 			if (bp->b_wcred != NOCRED)
433 				crhold(bp->b_wcred);
434 			bp->b_data = (caddr_t) kva + i * bsize;
435 			bp->b_blkno = fileaddr;
436 			pbgetvp(dp, bp);
437 			bp->b_bcount = bsize;
438 			bp->b_bufsize = bsize;
439 			bp->b_runningbufspace = bp->b_bufsize;
440 			runningbufspace += bp->b_runningbufspace;
441 
442 			/* do the input */
443 			BUF_STRATEGY(bp);
444 
445 			/* we definitely need to be at splvm here */
446 
447 			s = splvm();
448 			while ((bp->b_flags & B_DONE) == 0) {
449 				tsleep(bp, PVM, "vnsrd", 0);
450 			}
451 			splx(s);
452 			if ((bp->b_ioflags & BIO_ERROR) != 0)
453 				error = EIO;
454 
455 			/*
456 			 * free the buffer header back to the swap buffer pool
457 			 */
458 			relpbuf(bp, &vnode_pbuf_freecnt);
459 			if (error)
460 				break;
461 
462 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
463 		} else {
464 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
465 			bzero((caddr_t) kva + i * bsize, bsize);
466 		}
467 	}
468 	vm_pager_unmap_page(kva);
469 	pmap_clear_modify(m);
470 	vm_page_flag_clear(m, PG_ZERO);
471 	if (error) {
472 		return VM_PAGER_ERROR;
473 	}
474 	return VM_PAGER_OK;
475 
476 }
477 
478 
479 /*
480  * old style vnode pager output routine
481  */
482 static int
483 vnode_pager_input_old(object, m)
484 	vm_object_t object;
485 	vm_page_t m;
486 {
487 	struct uio auio;
488 	struct iovec aiov;
489 	int error;
490 	int size;
491 	vm_offset_t kva;
492 
493 	error = 0;
494 
495 	/*
496 	 * Return failure if beyond current EOF
497 	 */
498 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
499 		return VM_PAGER_BAD;
500 	} else {
501 		size = PAGE_SIZE;
502 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
503 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
504 
505 		/*
506 		 * Allocate a kernel virtual address and initialize so that
507 		 * we can use VOP_READ/WRITE routines.
508 		 */
509 		kva = vm_pager_map_page(m);
510 
511 		aiov.iov_base = (caddr_t) kva;
512 		aiov.iov_len = size;
513 		auio.uio_iov = &aiov;
514 		auio.uio_iovcnt = 1;
515 		auio.uio_offset = IDX_TO_OFF(m->pindex);
516 		auio.uio_segflg = UIO_SYSSPACE;
517 		auio.uio_rw = UIO_READ;
518 		auio.uio_resid = size;
519 		auio.uio_procp = curproc;
520 
521 		error = VOP_READ(object->handle, &auio, 0, curproc->p_ucred);
522 		if (!error) {
523 			register int count = size - auio.uio_resid;
524 
525 			if (count == 0)
526 				error = EINVAL;
527 			else if (count != PAGE_SIZE)
528 				bzero((caddr_t) kva + count, PAGE_SIZE - count);
529 		}
530 		vm_pager_unmap_page(kva);
531 	}
532 	pmap_clear_modify(m);
533 	vm_page_undirty(m);
534 	vm_page_flag_clear(m, PG_ZERO);
535 	if (!error)
536 		m->valid = VM_PAGE_BITS_ALL;
537 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
538 }
539 
540 /*
541  * generic vnode pager input routine
542  */
543 
544 /*
545  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
546  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
547  * vnode_pager_generic_getpages() to implement the previous behaviour.
548  *
549  * All other FS's should use the bypass to get to the local media
550  * backing vp's VOP_GETPAGES.
551  */
552 static int
553 vnode_pager_getpages(object, m, count, reqpage)
554 	vm_object_t object;
555 	vm_page_t *m;
556 	int count;
557 	int reqpage;
558 {
559 	int rtval;
560 	struct vnode *vp;
561 	int bytes = count * PAGE_SIZE;
562 
563 	vp = object->handle;
564 	/*
565 	 * XXX temporary diagnostic message to help track stale FS code,
566 	 * Returning EOPNOTSUPP from here may make things unhappy.
567 	 */
568 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
569 	if (rtval == EOPNOTSUPP) {
570 	    printf("vnode_pager: *** WARNING *** stale FS getpages\n");
571 	    rtval = vnode_pager_generic_getpages( vp, m, bytes, reqpage);
572 	}
573 	return rtval;
574 }
575 
576 
577 /*
578  * This is now called from local media FS's to operate against their
579  * own vnodes if they fail to implement VOP_GETPAGES.
580  */
581 int
582 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
583 	struct vnode *vp;
584 	vm_page_t *m;
585 	int bytecount;
586 	int reqpage;
587 {
588 	vm_object_t object;
589 	vm_offset_t kva;
590 	off_t foff, tfoff, nextoff;
591 	int i, size, bsize, first, firstaddr;
592 	struct vnode *dp;
593 	int runpg;
594 	int runend;
595 	struct buf *bp;
596 	int s;
597 	int count;
598 	int error = 0;
599 
600 	object = vp->v_object;
601 	count = bytecount / PAGE_SIZE;
602 
603 	if (vp->v_mount == NULL)
604 		return VM_PAGER_BAD;
605 
606 	bsize = vp->v_mount->mnt_stat.f_iosize;
607 
608 	/* get the UNDERLYING device for the file with VOP_BMAP() */
609 
610 	/*
611 	 * originally, we did not check for an error return value -- assuming
612 	 * an fs always has a bmap entry point -- that assumption is wrong!!!
613 	 */
614 	foff = IDX_TO_OFF(m[reqpage]->pindex);
615 
616 	/*
617 	 * if we can't bmap, use old VOP code
618 	 */
619 	if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
620 		for (i = 0; i < count; i++) {
621 			if (i != reqpage) {
622 				vm_page_free(m[i]);
623 			}
624 		}
625 		cnt.v_vnodein++;
626 		cnt.v_vnodepgsin++;
627 		return vnode_pager_input_old(object, m[reqpage]);
628 
629 		/*
630 		 * if the blocksize is smaller than a page size, then use
631 		 * special small filesystem code.  NFS sometimes has a small
632 		 * blocksize, but it can handle large reads itself.
633 		 */
634 	} else if ((PAGE_SIZE / bsize) > 1 &&
635 	    (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
636 		for (i = 0; i < count; i++) {
637 			if (i != reqpage) {
638 				vm_page_free(m[i]);
639 			}
640 		}
641 		cnt.v_vnodein++;
642 		cnt.v_vnodepgsin++;
643 		return vnode_pager_input_smlfs(object, m[reqpage]);
644 	}
645 
646 	/*
647 	 * If we have a completely valid page available to us, we can
648 	 * clean up and return.  Otherwise we have to re-read the
649 	 * media.
650 	 */
651 
652 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
653 		for (i = 0; i < count; i++) {
654 			if (i != reqpage)
655 				vm_page_free(m[i]);
656 		}
657 		return VM_PAGER_OK;
658 	}
659 	m[reqpage]->valid = 0;
660 
661 	/*
662 	 * here on direct device I/O
663 	 */
664 
665 	firstaddr = -1;
666 	/*
667 	 * calculate the run that includes the required page
668 	 */
669 	for(first = 0, i = 0; i < count; i = runend) {
670 		firstaddr = vnode_pager_addr(vp,
671 			IDX_TO_OFF(m[i]->pindex), &runpg);
672 		if (firstaddr == -1) {
673 			if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
674 				/* XXX no %qd in kernel. */
675 				panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%lx%08lx, vnp_size: 0x%lx%08lx",
676 			   	 firstaddr, (u_long)(foff >> 32),
677 			   	 (u_long)(u_int32_t)foff,
678 				 (u_long)(u_int32_t)
679 				 (object->un_pager.vnp.vnp_size >> 32),
680 				 (u_long)(u_int32_t)
681 				 object->un_pager.vnp.vnp_size);
682 			}
683 			vm_page_free(m[i]);
684 			runend = i + 1;
685 			first = runend;
686 			continue;
687 		}
688 		runend = i + runpg;
689 		if (runend <= reqpage) {
690 			int j;
691 			for (j = i; j < runend; j++) {
692 				vm_page_free(m[j]);
693 			}
694 		} else {
695 			if (runpg < (count - first)) {
696 				for (i = first + runpg; i < count; i++)
697 					vm_page_free(m[i]);
698 				count = first + runpg;
699 			}
700 			break;
701 		}
702 		first = runend;
703 	}
704 
705 	/*
706 	 * the first and last page have been calculated now, move input pages
707 	 * to be zero based...
708 	 */
709 	if (first != 0) {
710 		for (i = first; i < count; i++) {
711 			m[i - first] = m[i];
712 		}
713 		count -= first;
714 		reqpage -= first;
715 	}
716 
717 	/*
718 	 * calculate the file virtual address for the transfer
719 	 */
720 	foff = IDX_TO_OFF(m[0]->pindex);
721 
722 	/*
723 	 * calculate the size of the transfer
724 	 */
725 	size = count * PAGE_SIZE;
726 	if ((foff + size) > object->un_pager.vnp.vnp_size)
727 		size = object->un_pager.vnp.vnp_size - foff;
728 
729 	/*
730 	 * round up physical size for real devices.
731 	 */
732 	if (dp->v_type == VBLK || dp->v_type == VCHR) {
733 		int secmask = dp->v_rdev->si_bsize_phys - 1;
734 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
735 		size = (size + secmask) & ~secmask;
736 	}
737 
738 	bp = getpbuf(&vnode_pbuf_freecnt);
739 	kva = (vm_offset_t) bp->b_data;
740 
741 	/*
742 	 * and map the pages to be read into the kva
743 	 */
744 	pmap_qenter(kva, m, count);
745 
746 	/* build a minimal buffer header */
747 	bp->b_iocmd = BIO_READ;
748 	bp->b_iodone = vnode_pager_iodone;
749 	/* B_PHYS is not set, but it is nice to fill this in */
750 	bp->b_rcred = bp->b_wcred = curproc->p_ucred;
751 	if (bp->b_rcred != NOCRED)
752 		crhold(bp->b_rcred);
753 	if (bp->b_wcred != NOCRED)
754 		crhold(bp->b_wcred);
755 	bp->b_blkno = firstaddr;
756 	pbgetvp(dp, bp);
757 	bp->b_bcount = size;
758 	bp->b_bufsize = size;
759 	bp->b_runningbufspace = bp->b_bufsize;
760 	runningbufspace += bp->b_runningbufspace;
761 
762 	cnt.v_vnodein++;
763 	cnt.v_vnodepgsin += count;
764 
765 	/* do the input */
766 	BUF_STRATEGY(bp);
767 
768 	s = splvm();
769 	/* we definitely need to be at splvm here */
770 
771 	while ((bp->b_flags & B_DONE) == 0) {
772 		tsleep(bp, PVM, "vnread", 0);
773 	}
774 	splx(s);
775 	if ((bp->b_ioflags & BIO_ERROR) != 0)
776 		error = EIO;
777 
778 	if (!error) {
779 		if (size != count * PAGE_SIZE)
780 			bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
781 	}
782 	pmap_qremove(kva, count);
783 
784 	/*
785 	 * free the buffer header back to the swap buffer pool
786 	 */
787 	relpbuf(bp, &vnode_pbuf_freecnt);
788 
789 	for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
790 		vm_page_t mt;
791 
792 		nextoff = tfoff + PAGE_SIZE;
793 		mt = m[i];
794 
795 		if (nextoff <= object->un_pager.vnp.vnp_size) {
796 			/*
797 			 * Read filled up entire page.
798 			 */
799 			mt->valid = VM_PAGE_BITS_ALL;
800 			vm_page_undirty(mt);	/* should be an assert? XXX */
801 			pmap_clear_modify(mt);
802 		} else {
803 			/*
804 			 * Read did not fill up entire page.  Since this
805 			 * is getpages, the page may be mapped, so we have
806 			 * to zero the invalid portions of the page even
807 			 * though we aren't setting them valid.
808 			 *
809 			 * Currently we do not set the entire page valid,
810 			 * we just try to clear the piece that we couldn't
811 			 * read.
812 			 */
813 			vm_page_set_validclean(mt, 0,
814 			    object->un_pager.vnp.vnp_size - tfoff);
815 			/* handled by vm_fault now */
816 			/* vm_page_zero_invalid(mt, FALSE); */
817 		}
818 
819 		vm_page_flag_clear(mt, PG_ZERO);
820 		if (i != reqpage) {
821 
822 			/*
823 			 * whether or not to leave the page activated is up in
824 			 * the air, but we should put the page on a page queue
825 			 * somewhere. (it already is in the object). Result:
826 			 * It appears that empirical results show that
827 			 * deactivating pages is best.
828 			 */
829 
830 			/*
831 			 * just in case someone was asking for this page we
832 			 * now tell them that it is ok to use
833 			 */
834 			if (!error) {
835 				if (mt->flags & PG_WANTED)
836 					vm_page_activate(mt);
837 				else
838 					vm_page_deactivate(mt);
839 				vm_page_wakeup(mt);
840 			} else {
841 				vm_page_free(mt);
842 			}
843 		}
844 	}
845 	if (error) {
846 		printf("vnode_pager_getpages: I/O read error\n");
847 	}
848 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
849 }
850 
851 /*
852  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
853  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
854  * vnode_pager_generic_putpages() to implement the previous behaviour.
855  *
856  * All other FS's should use the bypass to get to the local media
857  * backing vp's VOP_PUTPAGES.
858  */
859 static void
860 vnode_pager_putpages(object, m, count, sync, rtvals)
861 	vm_object_t object;
862 	vm_page_t *m;
863 	int count;
864 	boolean_t sync;
865 	int *rtvals;
866 {
867 	int rtval;
868 	struct vnode *vp;
869 	struct mount *mp;
870 	int bytes = count * PAGE_SIZE;
871 
872 	/*
873 	 * Force synchronous operation if we are extremely low on memory
874 	 * to prevent a low-memory deadlock.  VOP operations often need to
875 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
876 	 * operation ).  The swapper handles the case by limiting the amount
877 	 * of asynchronous I/O, but that sort of solution doesn't scale well
878 	 * for the vnode pager without a lot of work.
879 	 *
880 	 * Also, the backing vnode's iodone routine may not wake the pageout
881 	 * daemon up.  This should be probably be addressed XXX.
882 	 */
883 
884 	if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
885 		sync |= OBJPC_SYNC;
886 
887 	/*
888 	 * Call device-specific putpages function
889 	 */
890 
891 	vp = object->handle;
892 	if (vp->v_type != VREG)
893 		mp = NULL;
894 	(void)vn_start_write(vp, &mp, V_WAIT);
895 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
896 	if (rtval == EOPNOTSUPP) {
897 	    printf("vnode_pager: *** WARNING *** stale FS putpages\n");
898 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
899 	}
900 	vn_finished_write(mp);
901 }
902 
903 
904 /*
905  * This is now called from local media FS's to operate against their
906  * own vnodes if they fail to implement VOP_PUTPAGES.
907  *
908  * This is typically called indirectly via the pageout daemon and
909  * clustering has already typically occured, so in general we ask the
910  * underlying filesystem to write the data out asynchronously rather
911  * then delayed.
912  */
913 int
914 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
915 	struct vnode *vp;
916 	vm_page_t *m;
917 	int bytecount;
918 	int flags;
919 	int *rtvals;
920 {
921 	int i;
922 	vm_object_t object;
923 	int count;
924 
925 	int maxsize, ncount;
926 	vm_ooffset_t poffset;
927 	struct uio auio;
928 	struct iovec aiov;
929 	int error;
930 	int ioflags;
931 
932 	object = vp->v_object;
933 	count = bytecount / PAGE_SIZE;
934 
935 	for (i = 0; i < count; i++)
936 		rtvals[i] = VM_PAGER_AGAIN;
937 
938 	if ((int) m[0]->pindex < 0) {
939 		printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
940 			(long)m[0]->pindex, m[0]->dirty);
941 		rtvals[0] = VM_PAGER_BAD;
942 		return VM_PAGER_BAD;
943 	}
944 
945 	maxsize = count * PAGE_SIZE;
946 	ncount = count;
947 
948 	poffset = IDX_TO_OFF(m[0]->pindex);
949 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
950 		if (object->un_pager.vnp.vnp_size > poffset)
951 			maxsize = object->un_pager.vnp.vnp_size - poffset;
952 		else
953 			maxsize = 0;
954 		ncount = btoc(maxsize);
955 		if (ncount < count) {
956 			for (i = ncount; i < count; i++) {
957 				rtvals[i] = VM_PAGER_BAD;
958 			}
959 		}
960 	}
961 
962 	/*
963 	 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
964 	 * rather then a bdwrite() to prevent paging I/O from saturating
965 	 * the buffer cache.
966 	 */
967 	ioflags = IO_VMIO;
968 	ioflags |= (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) ? IO_SYNC: IO_ASYNC;
969 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
970 
971 	aiov.iov_base = (caddr_t) 0;
972 	aiov.iov_len = maxsize;
973 	auio.uio_iov = &aiov;
974 	auio.uio_iovcnt = 1;
975 	auio.uio_offset = poffset;
976 	auio.uio_segflg = UIO_NOCOPY;
977 	auio.uio_rw = UIO_WRITE;
978 	auio.uio_resid = maxsize;
979 	auio.uio_procp = (struct proc *) 0;
980 	error = VOP_WRITE(vp, &auio, ioflags, curproc->p_ucred);
981 	cnt.v_vnodeout++;
982 	cnt.v_vnodepgsout += ncount;
983 
984 	if (error) {
985 		printf("vnode_pager_putpages: I/O error %d\n", error);
986 	}
987 	if (auio.uio_resid) {
988 		printf("vnode_pager_putpages: residual I/O %d at %lu\n",
989 		    auio.uio_resid, (u_long)m[0]->pindex);
990 	}
991 	for (i = 0; i < ncount; i++) {
992 		rtvals[i] = VM_PAGER_OK;
993 	}
994 	return rtvals[0];
995 }
996 
997 struct vnode *
998 vnode_pager_lock(object)
999 	vm_object_t object;
1000 {
1001 	struct proc *p = curproc;	/* XXX */
1002 
1003 	for (; object != NULL; object = object->backing_object) {
1004 		if (object->type != OBJT_VNODE)
1005 			continue;
1006 		if (object->flags & OBJ_DEAD)
1007 			return NULL;
1008 
1009 		while (vget(object->handle,
1010 			LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, p)) {
1011 			if ((object->flags & OBJ_DEAD) || (object->type != OBJT_VNODE))
1012 				return NULL;
1013 			printf("vnode_pager_lock: retrying\n");
1014 		}
1015 		return object->handle;
1016 	}
1017 	return NULL;
1018 }
1019