xref: /freebsd/sys/vm/vnode_pager.c (revision 5773cccf19ef7b97e56c1101aa481c43149224da)
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 #include <sys/stdint.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_map.h>
70 #include <vm/vnode_pager.h>
71 #include <vm/vm_extern.h>
72 
73 static void vnode_pager_init(void);
74 static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
75 					 int *run);
76 static void vnode_pager_iodone(struct buf *bp);
77 static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m);
78 static int vnode_pager_input_old(vm_object_t object, vm_page_t m);
79 static void vnode_pager_dealloc(vm_object_t);
80 static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int);
81 static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
82 static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *);
83 
84 struct pagerops vnodepagerops = {
85 	vnode_pager_init,
86 	vnode_pager_alloc,
87 	vnode_pager_dealloc,
88 	vnode_pager_getpages,
89 	vnode_pager_putpages,
90 	vnode_pager_haspage,
91 	NULL
92 };
93 
94 int vnode_pbuf_freecnt;
95 
96 static void
97 vnode_pager_init(void)
98 {
99 
100 	vnode_pbuf_freecnt = nswbuf / 2 + 1;
101 }
102 
103 /*
104  * Allocate (or lookup) pager for a vnode.
105  * Handle is a vnode pointer.
106  *
107  * MPSAFE
108  */
109 vm_object_t
110 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
111 		  vm_ooffset_t offset)
112 {
113 	vm_object_t object;
114 	struct vnode *vp;
115 
116 	/*
117 	 * Pageout to vnode, no can do yet.
118 	 */
119 	if (handle == NULL)
120 		return (NULL);
121 
122 	vp = (struct vnode *) handle;
123 
124 	ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
125 
126 	mtx_lock(&Giant);
127 	/*
128 	 * Prevent race condition when allocating the object. This
129 	 * can happen with NFS vnodes since the nfsnode isn't locked.
130 	 */
131 	VI_LOCK(vp);
132 	while (vp->v_iflag & VI_OLOCK) {
133 		vp->v_iflag |= VI_OWANT;
134 		msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0);
135 	}
136 	vp->v_iflag |= VI_OLOCK;
137 	VI_UNLOCK(vp);
138 
139 	/*
140 	 * If the object is being terminated, wait for it to
141 	 * go away.
142 	 */
143 	while (((object = vp->v_object) != NULL) &&
144 		(object->flags & OBJ_DEAD)) {
145 		tsleep(object, PVM, "vadead", 0);
146 	}
147 
148 	if (vp->v_usecount == 0)
149 		panic("vnode_pager_alloc: no vnode reference");
150 
151 	if (object == NULL) {
152 		/*
153 		 * And an object of the appropriate size
154 		 */
155 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
156 
157 		object->un_pager.vnp.vnp_size = size;
158 
159 		object->handle = handle;
160 		vp->v_object = object;
161 	} else {
162 		object->ref_count++;
163 	}
164 	VI_LOCK(vp);
165 	vp->v_usecount++;
166 	vp->v_iflag &= ~VI_OLOCK;
167 	if (vp->v_iflag & VI_OWANT) {
168 		vp->v_iflag &= ~VI_OWANT;
169 		wakeup(vp);
170 	}
171 	VI_UNLOCK(vp);
172 	mtx_unlock(&Giant);
173 	return (object);
174 }
175 
176 static void
177 vnode_pager_dealloc(object)
178 	vm_object_t object;
179 {
180 	struct vnode *vp = object->handle;
181 
182 	GIANT_REQUIRED;
183 	if (vp == NULL)
184 		panic("vnode_pager_dealloc: pager already dealloced");
185 
186 	vm_object_pip_wait(object, "vnpdea");
187 
188 	object->handle = NULL;
189 	object->type = OBJT_DEAD;
190 	ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
191 	vp->v_object = NULL;
192 	vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF);
193 }
194 
195 static boolean_t
196 vnode_pager_haspage(object, pindex, before, after)
197 	vm_object_t object;
198 	vm_pindex_t pindex;
199 	int *before;
200 	int *after;
201 {
202 	struct vnode *vp = object->handle;
203 	daddr_t bn;
204 	int err;
205 	daddr_t reqblock;
206 	int poff;
207 	int bsize;
208 	int pagesperblock, blocksperpage;
209 
210 	GIANT_REQUIRED;
211 	/*
212 	 * If no vp or vp is doomed or marked transparent to VM, we do not
213 	 * have the page.
214 	 */
215 	if (vp == NULL)
216 		return FALSE;
217 
218 	VI_LOCK(vp);
219 	if (vp->v_iflag & VI_DOOMED) {
220 		VI_UNLOCK(vp);
221 		return FALSE;
222 	}
223 	VI_UNLOCK(vp);
224 	/*
225 	 * If filesystem no longer mounted or offset beyond end of file we do
226 	 * not have the page.
227 	 */
228 	if ((vp->v_mount == NULL) ||
229 	    (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
230 		return FALSE;
231 
232 	bsize = vp->v_mount->mnt_stat.f_iosize;
233 	pagesperblock = bsize / PAGE_SIZE;
234 	blocksperpage = 0;
235 	if (pagesperblock > 0) {
236 		reqblock = pindex / pagesperblock;
237 	} else {
238 		blocksperpage = (PAGE_SIZE / bsize);
239 		reqblock = pindex * blocksperpage;
240 	}
241 	err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn,
242 		after, before);
243 	if (err)
244 		return TRUE;
245 	if (bn == -1)
246 		return FALSE;
247 	if (pagesperblock > 0) {
248 		poff = pindex - (reqblock * pagesperblock);
249 		if (before) {
250 			*before *= pagesperblock;
251 			*before += poff;
252 		}
253 		if (after) {
254 			int numafter;
255 			*after *= pagesperblock;
256 			numafter = pagesperblock - (poff + 1);
257 			if (IDX_TO_OFF(pindex + numafter) >
258 			    object->un_pager.vnp.vnp_size) {
259 				numafter =
260 		    		    OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
261 				    pindex;
262 			}
263 			*after += numafter;
264 		}
265 	} else {
266 		if (before) {
267 			*before /= blocksperpage;
268 		}
269 
270 		if (after) {
271 			*after /= blocksperpage;
272 		}
273 	}
274 	return TRUE;
275 }
276 
277 /*
278  * Lets the VM system know about a change in size for a file.
279  * We adjust our own internal size and flush any cached pages in
280  * the associated object that are affected by the size change.
281  *
282  * Note: this routine may be invoked as a result of a pager put
283  * operation (possibly at object termination time), so we must be careful.
284  */
285 void
286 vnode_pager_setsize(vp, nsize)
287 	struct vnode *vp;
288 	vm_ooffset_t nsize;
289 {
290 	vm_pindex_t nobjsize;
291 	vm_object_t object = vp->v_object;
292 
293 	GIANT_REQUIRED;
294 
295 	if (object == NULL)
296 		return;
297 
298 	/*
299 	 * Hasn't changed size
300 	 */
301 	if (nsize == object->un_pager.vnp.vnp_size)
302 		return;
303 
304 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
305 
306 	/*
307 	 * File has shrunk. Toss any cached pages beyond the new EOF.
308 	 */
309 	if (nsize < object->un_pager.vnp.vnp_size) {
310 #ifdef ENABLE_VFS_IOOPT
311 		vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size);
312 #endif
313 		if (nobjsize < object->size) {
314 			vm_object_page_remove(object, nobjsize, object->size,
315 				FALSE);
316 		}
317 		/*
318 		 * this gets rid of garbage at the end of a page that is now
319 		 * only partially backed by the vnode.
320 		 *
321 		 * XXX for some reason (I don't know yet), if we take a
322 		 * completely invalid page and mark it partially valid
323 		 * it can screw up NFS reads, so we don't allow the case.
324 		 */
325 		if (nsize & PAGE_MASK) {
326 			vm_page_t m;
327 
328 			m = vm_page_lookup(object, OFF_TO_IDX(nsize));
329 			if (m && m->valid) {
330 				int base = (int)nsize & PAGE_MASK;
331 				int size = PAGE_SIZE - base;
332 
333 				/*
334 				 * Clear out partial-page garbage in case
335 				 * the page has been mapped.
336 				 */
337 				pmap_zero_page_area(m, base, size);
338 
339 				vm_page_lock_queues();
340 				/*
341 				 * XXX work around SMP data integrity race
342 				 * by unmapping the page from user processes.
343 				 * The garbage we just cleared may be mapped
344 				 * to a user process running on another cpu
345 				 * and this code is not running through normal
346 				 * I/O channels which handle SMP issues for
347 				 * us, so unmap page to synchronize all cpus.
348 				 *
349 				 * XXX should vm_pager_unmap_page() have
350 				 * dealt with this?
351 				 */
352 				pmap_remove_all(m);
353 
354 				/*
355 				 * Clear out partial-page dirty bits.  This
356 				 * has the side effect of setting the valid
357 				 * bits, but that is ok.  There are a bunch
358 				 * of places in the VM system where we expected
359 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
360 				 * case is one of them.  If the page is still
361 				 * partially dirty, make it fully dirty.
362 				 *
363 				 * note that we do not clear out the valid
364 				 * bits.  This would prevent bogus_page
365 				 * replacement from working properly.
366 				 */
367 				vm_page_set_validclean(m, base, size);
368 				if (m->dirty != 0)
369 					m->dirty = VM_PAGE_BITS_ALL;
370 				vm_page_unlock_queues();
371 			}
372 		}
373 	}
374 	object->un_pager.vnp.vnp_size = nsize;
375 	object->size = nobjsize;
376 }
377 
378 /*
379  * calculate the linear (byte) disk address of specified virtual
380  * file address
381  */
382 static vm_offset_t
383 vnode_pager_addr(vp, address, run)
384 	struct vnode *vp;
385 	vm_ooffset_t address;
386 	int *run;
387 {
388 	int rtaddress;
389 	int bsize;
390 	daddr_t block;
391 	struct vnode *rtvp;
392 	int err;
393 	daddr_t vblock;
394 	int voffset;
395 
396 	GIANT_REQUIRED;
397 	if ((int) address < 0)
398 		return -1;
399 
400 	if (vp->v_mount == NULL)
401 		return -1;
402 
403 	bsize = vp->v_mount->mnt_stat.f_iosize;
404 	vblock = address / bsize;
405 	voffset = address % bsize;
406 
407 	err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL);
408 
409 	if (err || (block == -1))
410 		rtaddress = -1;
411 	else {
412 		rtaddress = block + voffset / DEV_BSIZE;
413 		if (run) {
414 			*run += 1;
415 			*run *= bsize/PAGE_SIZE;
416 			*run -= voffset/PAGE_SIZE;
417 		}
418 	}
419 
420 	return rtaddress;
421 }
422 
423 /*
424  * interrupt routine for I/O completion
425  */
426 static void
427 vnode_pager_iodone(bp)
428 	struct buf *bp;
429 {
430 	bp->b_flags |= B_DONE;
431 	wakeup(bp);
432 }
433 
434 /*
435  * small block filesystem vnode pager input
436  */
437 static int
438 vnode_pager_input_smlfs(object, m)
439 	vm_object_t object;
440 	vm_page_t m;
441 {
442 	int i;
443 	int s;
444 	struct vnode *dp, *vp;
445 	struct buf *bp;
446 	vm_offset_t kva;
447 	int fileaddr;
448 	vm_offset_t bsize;
449 	int error = 0;
450 
451 	GIANT_REQUIRED;
452 
453 	vp = object->handle;
454 	if (vp->v_mount == NULL)
455 		return VM_PAGER_BAD;
456 
457 	bsize = vp->v_mount->mnt_stat.f_iosize;
458 
459 	VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
460 
461 	kva = vm_pager_map_page(m);
462 
463 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
464 		vm_ooffset_t address;
465 
466 		if (vm_page_bits(i * bsize, bsize) & m->valid)
467 			continue;
468 
469 		address = IDX_TO_OFF(m->pindex) + i * bsize;
470 		if (address >= object->un_pager.vnp.vnp_size) {
471 			fileaddr = -1;
472 		} else {
473 			fileaddr = vnode_pager_addr(vp, address, NULL);
474 		}
475 		if (fileaddr != -1) {
476 			bp = getpbuf(&vnode_pbuf_freecnt);
477 
478 			/* build a minimal buffer header */
479 			bp->b_iocmd = BIO_READ;
480 			bp->b_iodone = vnode_pager_iodone;
481 			KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
482 			KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
483 			bp->b_rcred = crhold(curthread->td_ucred);
484 			bp->b_wcred = crhold(curthread->td_ucred);
485 			bp->b_data = (caddr_t) kva + i * bsize;
486 			bp->b_blkno = fileaddr;
487 			pbgetvp(dp, bp);
488 			bp->b_bcount = bsize;
489 			bp->b_bufsize = bsize;
490 			bp->b_runningbufspace = bp->b_bufsize;
491 			runningbufspace += bp->b_runningbufspace;
492 
493 			/* do the input */
494 			BUF_STRATEGY(bp);
495 
496 			/* we definitely need to be at splvm here */
497 
498 			s = splvm();
499 			while ((bp->b_flags & B_DONE) == 0) {
500 				tsleep(bp, PVM, "vnsrd", 0);
501 			}
502 			splx(s);
503 			if ((bp->b_ioflags & BIO_ERROR) != 0)
504 				error = EIO;
505 
506 			/*
507 			 * free the buffer header back to the swap buffer pool
508 			 */
509 			relpbuf(bp, &vnode_pbuf_freecnt);
510 			if (error)
511 				break;
512 
513 			vm_page_lock_queues();
514 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
515 			vm_page_unlock_queues();
516 		} else {
517 			vm_page_lock_queues();
518 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
519 			vm_page_unlock_queues();
520 			bzero((caddr_t) kva + i * bsize, bsize);
521 		}
522 	}
523 	vm_pager_unmap_page(kva);
524 	pmap_clear_modify(m);
525 	vm_page_flag_clear(m, PG_ZERO);
526 	if (error) {
527 		return VM_PAGER_ERROR;
528 	}
529 	return VM_PAGER_OK;
530 
531 }
532 
533 
534 /*
535  * old style vnode pager output routine
536  */
537 static int
538 vnode_pager_input_old(object, m)
539 	vm_object_t object;
540 	vm_page_t m;
541 {
542 	struct uio auio;
543 	struct iovec aiov;
544 	int error;
545 	int size;
546 	vm_offset_t kva;
547 	struct vnode *vp;
548 
549 	GIANT_REQUIRED;
550 	error = 0;
551 
552 	/*
553 	 * Return failure if beyond current EOF
554 	 */
555 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
556 		return VM_PAGER_BAD;
557 	} else {
558 		size = PAGE_SIZE;
559 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
560 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
561 
562 		/*
563 		 * Allocate a kernel virtual address and initialize so that
564 		 * we can use VOP_READ/WRITE routines.
565 		 */
566 		kva = vm_pager_map_page(m);
567 
568 		vp = object->handle;
569 		aiov.iov_base = (caddr_t) kva;
570 		aiov.iov_len = size;
571 		auio.uio_iov = &aiov;
572 		auio.uio_iovcnt = 1;
573 		auio.uio_offset = IDX_TO_OFF(m->pindex);
574 		auio.uio_segflg = UIO_SYSSPACE;
575 		auio.uio_rw = UIO_READ;
576 		auio.uio_resid = size;
577 		auio.uio_td = curthread;
578 
579 		error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
580 		if (!error) {
581 			int count = size - auio.uio_resid;
582 
583 			if (count == 0)
584 				error = EINVAL;
585 			else if (count != PAGE_SIZE)
586 				bzero((caddr_t) kva + count, PAGE_SIZE - count);
587 		}
588 		vm_pager_unmap_page(kva);
589 	}
590 	pmap_clear_modify(m);
591 	vm_page_undirty(m);
592 	vm_page_flag_clear(m, PG_ZERO);
593 	if (!error)
594 		m->valid = VM_PAGE_BITS_ALL;
595 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
596 }
597 
598 /*
599  * generic vnode pager input routine
600  */
601 
602 /*
603  * Local media VFS's that do not implement their own VOP_GETPAGES
604  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
605  * to implement the previous behaviour.
606  *
607  * All other FS's should use the bypass to get to the local media
608  * backing vp's VOP_GETPAGES.
609  */
610 static int
611 vnode_pager_getpages(object, m, count, reqpage)
612 	vm_object_t object;
613 	vm_page_t *m;
614 	int count;
615 	int reqpage;
616 {
617 	int rtval;
618 	struct vnode *vp;
619 	int bytes = count * PAGE_SIZE;
620 
621 	GIANT_REQUIRED;
622 	vp = object->handle;
623 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
624 	KASSERT(rtval != EOPNOTSUPP,
625 	    ("vnode_pager: FS getpages not implemented\n"));
626 	return rtval;
627 }
628 
629 /*
630  * This is now called from local media FS's to operate against their
631  * own vnodes if they fail to implement VOP_GETPAGES.
632  */
633 int
634 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
635 	struct vnode *vp;
636 	vm_page_t *m;
637 	int bytecount;
638 	int reqpage;
639 {
640 	vm_object_t object;
641 	vm_offset_t kva;
642 	off_t foff, tfoff, nextoff;
643 	int i, j, size, bsize, first, firstaddr;
644 	struct vnode *dp;
645 	int runpg;
646 	int runend;
647 	struct buf *bp;
648 	int s;
649 	int count;
650 	int error = 0;
651 
652 	GIANT_REQUIRED;
653 	object = vp->v_object;
654 	count = bytecount / PAGE_SIZE;
655 
656 	if (vp->v_mount == NULL)
657 		return VM_PAGER_BAD;
658 
659 	bsize = vp->v_mount->mnt_stat.f_iosize;
660 
661 	/* get the UNDERLYING device for the file with VOP_BMAP() */
662 
663 	/*
664 	 * originally, we did not check for an error return value -- assuming
665 	 * an fs always has a bmap entry point -- that assumption is wrong!!!
666 	 */
667 	foff = IDX_TO_OFF(m[reqpage]->pindex);
668 
669 	/*
670 	 * if we can't bmap, use old VOP code
671 	 */
672 	if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
673 		vm_page_lock_queues();
674 		for (i = 0; i < count; i++)
675 			if (i != reqpage)
676 				vm_page_free(m[i]);
677 		vm_page_unlock_queues();
678 		cnt.v_vnodein++;
679 		cnt.v_vnodepgsin++;
680 		return vnode_pager_input_old(object, m[reqpage]);
681 
682 		/*
683 		 * if the blocksize is smaller than a page size, then use
684 		 * special small filesystem code.  NFS sometimes has a small
685 		 * blocksize, but it can handle large reads itself.
686 		 */
687 	} else if ((PAGE_SIZE / bsize) > 1 &&
688 	    (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
689 		vm_page_lock_queues();
690 		for (i = 0; i < count; i++)
691 			if (i != reqpage)
692 				vm_page_free(m[i]);
693 		vm_page_unlock_queues();
694 		cnt.v_vnodein++;
695 		cnt.v_vnodepgsin++;
696 		return vnode_pager_input_smlfs(object, m[reqpage]);
697 	}
698 
699 	/*
700 	 * If we have a completely valid page available to us, we can
701 	 * clean up and return.  Otherwise we have to re-read the
702 	 * media.
703 	 */
704 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
705 		vm_page_lock_queues();
706 		for (i = 0; i < count; i++)
707 			if (i != reqpage)
708 				vm_page_free(m[i]);
709 		vm_page_unlock_queues();
710 		return VM_PAGER_OK;
711 	}
712 	m[reqpage]->valid = 0;
713 
714 	/*
715 	 * here on direct device I/O
716 	 */
717 	firstaddr = -1;
718 
719 	/*
720 	 * calculate the run that includes the required page
721 	 */
722 	for (first = 0, i = 0; i < count; i = runend) {
723 		firstaddr = vnode_pager_addr(vp,
724 			IDX_TO_OFF(m[i]->pindex), &runpg);
725 		if (firstaddr == -1) {
726 			if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
727 				panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
728 				    firstaddr, (uintmax_t)(foff >> 32),
729 				    (uintmax_t)foff,
730 				    (uintmax_t)
731 				    (object->un_pager.vnp.vnp_size >> 32),
732 				    (uintmax_t)object->un_pager.vnp.vnp_size);
733 			}
734 			vm_page_lock_queues();
735 			vm_page_free(m[i]);
736 			vm_page_unlock_queues();
737 			runend = i + 1;
738 			first = runend;
739 			continue;
740 		}
741 		runend = i + runpg;
742 		if (runend <= reqpage) {
743 			vm_page_lock_queues();
744 			for (j = i; j < runend; j++)
745 				vm_page_free(m[j]);
746 			vm_page_unlock_queues();
747 		} else {
748 			if (runpg < (count - first)) {
749 				vm_page_lock_queues();
750 				for (i = first + runpg; i < count; i++)
751 					vm_page_free(m[i]);
752 				vm_page_unlock_queues();
753 				count = first + runpg;
754 			}
755 			break;
756 		}
757 		first = runend;
758 	}
759 
760 	/*
761 	 * the first and last page have been calculated now, move input pages
762 	 * to be zero based...
763 	 */
764 	if (first != 0) {
765 		for (i = first; i < count; i++) {
766 			m[i - first] = m[i];
767 		}
768 		count -= first;
769 		reqpage -= first;
770 	}
771 
772 	/*
773 	 * calculate the file virtual address for the transfer
774 	 */
775 	foff = IDX_TO_OFF(m[0]->pindex);
776 
777 	/*
778 	 * calculate the size of the transfer
779 	 */
780 	size = count * PAGE_SIZE;
781 	if ((foff + size) > object->un_pager.vnp.vnp_size)
782 		size = object->un_pager.vnp.vnp_size - foff;
783 
784 	/*
785 	 * round up physical size for real devices.
786 	 */
787 	if (dp->v_type == VBLK || dp->v_type == VCHR) {
788 		int secmask = dp->v_rdev->si_bsize_phys - 1;
789 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
790 		size = (size + secmask) & ~secmask;
791 	}
792 
793 	bp = getpbuf(&vnode_pbuf_freecnt);
794 	kva = (vm_offset_t) bp->b_data;
795 
796 	/*
797 	 * and map the pages to be read into the kva
798 	 */
799 	pmap_qenter(kva, m, count);
800 
801 	/* build a minimal buffer header */
802 	bp->b_iocmd = BIO_READ;
803 	bp->b_iodone = vnode_pager_iodone;
804 	/* B_PHYS is not set, but it is nice to fill this in */
805 	KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
806 	KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
807 	bp->b_rcred = crhold(curthread->td_ucred);
808 	bp->b_wcred = crhold(curthread->td_ucred);
809 	bp->b_blkno = firstaddr;
810 	pbgetvp(dp, bp);
811 	bp->b_bcount = size;
812 	bp->b_bufsize = size;
813 	bp->b_runningbufspace = bp->b_bufsize;
814 	runningbufspace += bp->b_runningbufspace;
815 
816 	cnt.v_vnodein++;
817 	cnt.v_vnodepgsin += count;
818 
819 	/* do the input */
820 	BUF_STRATEGY(bp);
821 
822 	s = splvm();
823 	/* we definitely need to be at splvm here */
824 
825 	while ((bp->b_flags & B_DONE) == 0) {
826 		tsleep(bp, PVM, "vnread", 0);
827 	}
828 	splx(s);
829 	if ((bp->b_ioflags & BIO_ERROR) != 0)
830 		error = EIO;
831 
832 	if (!error) {
833 		if (size != count * PAGE_SIZE)
834 			bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
835 	}
836 	pmap_qremove(kva, count);
837 
838 	/*
839 	 * free the buffer header back to the swap buffer pool
840 	 */
841 	relpbuf(bp, &vnode_pbuf_freecnt);
842 
843 	vm_page_lock_queues();
844 	for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
845 		vm_page_t mt;
846 
847 		nextoff = tfoff + PAGE_SIZE;
848 		mt = m[i];
849 
850 		if (nextoff <= object->un_pager.vnp.vnp_size) {
851 			/*
852 			 * Read filled up entire page.
853 			 */
854 			mt->valid = VM_PAGE_BITS_ALL;
855 			vm_page_undirty(mt);	/* should be an assert? XXX */
856 			pmap_clear_modify(mt);
857 		} else {
858 			/*
859 			 * Read did not fill up entire page.  Since this
860 			 * is getpages, the page may be mapped, so we have
861 			 * to zero the invalid portions of the page even
862 			 * though we aren't setting them valid.
863 			 *
864 			 * Currently we do not set the entire page valid,
865 			 * we just try to clear the piece that we couldn't
866 			 * read.
867 			 */
868 			vm_page_set_validclean(mt, 0,
869 			    object->un_pager.vnp.vnp_size - tfoff);
870 			/* handled by vm_fault now */
871 			/* vm_page_zero_invalid(mt, FALSE); */
872 		}
873 
874 		vm_page_flag_clear(mt, PG_ZERO);
875 		if (i != reqpage) {
876 
877 			/*
878 			 * whether or not to leave the page activated is up in
879 			 * the air, but we should put the page on a page queue
880 			 * somewhere. (it already is in the object). Result:
881 			 * It appears that empirical results show that
882 			 * deactivating pages is best.
883 			 */
884 
885 			/*
886 			 * just in case someone was asking for this page we
887 			 * now tell them that it is ok to use
888 			 */
889 			if (!error) {
890 				if (mt->flags & PG_WANTED)
891 					vm_page_activate(mt);
892 				else
893 					vm_page_deactivate(mt);
894 				vm_page_wakeup(mt);
895 			} else {
896 				vm_page_free(mt);
897 			}
898 		}
899 	}
900 	vm_page_unlock_queues();
901 	if (error) {
902 		printf("vnode_pager_getpages: I/O read error\n");
903 	}
904 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
905 }
906 
907 /*
908  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
909  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
910  * vnode_pager_generic_putpages() to implement the previous behaviour.
911  *
912  * All other FS's should use the bypass to get to the local media
913  * backing vp's VOP_PUTPAGES.
914  */
915 static void
916 vnode_pager_putpages(object, m, count, sync, rtvals)
917 	vm_object_t object;
918 	vm_page_t *m;
919 	int count;
920 	boolean_t sync;
921 	int *rtvals;
922 {
923 	int rtval;
924 	struct vnode *vp;
925 	struct mount *mp;
926 	int bytes = count * PAGE_SIZE;
927 
928 	GIANT_REQUIRED;
929 	/*
930 	 * Force synchronous operation if we are extremely low on memory
931 	 * to prevent a low-memory deadlock.  VOP operations often need to
932 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
933 	 * operation ).  The swapper handles the case by limiting the amount
934 	 * of asynchronous I/O, but that sort of solution doesn't scale well
935 	 * for the vnode pager without a lot of work.
936 	 *
937 	 * Also, the backing vnode's iodone routine may not wake the pageout
938 	 * daemon up.  This should be probably be addressed XXX.
939 	 */
940 
941 	if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
942 		sync |= OBJPC_SYNC;
943 
944 	/*
945 	 * Call device-specific putpages function
946 	 */
947 	vp = object->handle;
948 	if (vp->v_type != VREG)
949 		mp = NULL;
950 	(void)vn_start_write(vp, &mp, V_WAIT);
951 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
952 	KASSERT(rtval != EOPNOTSUPP,
953 	    ("vnode_pager: stale FS putpages\n"));
954 	vn_finished_write(mp);
955 }
956 
957 
958 /*
959  * This is now called from local media FS's to operate against their
960  * own vnodes if they fail to implement VOP_PUTPAGES.
961  *
962  * This is typically called indirectly via the pageout daemon and
963  * clustering has already typically occured, so in general we ask the
964  * underlying filesystem to write the data out asynchronously rather
965  * then delayed.
966  */
967 int
968 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
969 	struct vnode *vp;
970 	vm_page_t *m;
971 	int bytecount;
972 	int flags;
973 	int *rtvals;
974 {
975 	int i;
976 	vm_object_t object;
977 	int count;
978 
979 	int maxsize, ncount;
980 	vm_ooffset_t poffset;
981 	struct uio auio;
982 	struct iovec aiov;
983 	int error;
984 	int ioflags;
985 
986 	GIANT_REQUIRED;
987 	object = vp->v_object;
988 	count = bytecount / PAGE_SIZE;
989 
990 	for (i = 0; i < count; i++)
991 		rtvals[i] = VM_PAGER_AGAIN;
992 
993 	if ((int) m[0]->pindex < 0) {
994 		printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
995 			(long)m[0]->pindex, m[0]->dirty);
996 		rtvals[0] = VM_PAGER_BAD;
997 		return VM_PAGER_BAD;
998 	}
999 
1000 	maxsize = count * PAGE_SIZE;
1001 	ncount = count;
1002 
1003 	poffset = IDX_TO_OFF(m[0]->pindex);
1004 
1005 	/*
1006 	 * If the page-aligned write is larger then the actual file we
1007 	 * have to invalidate pages occuring beyond the file EOF.  However,
1008 	 * there is an edge case where a file may not be page-aligned where
1009 	 * the last page is partially invalid.  In this case the filesystem
1010 	 * may not properly clear the dirty bits for the entire page (which
1011 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1012 	 * With the page locked we are free to fix-up the dirty bits here.
1013 	 *
1014 	 * We do not under any circumstances truncate the valid bits, as
1015 	 * this will screw up bogus page replacement.
1016 	 */
1017 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1018 		if (object->un_pager.vnp.vnp_size > poffset) {
1019 			int pgoff;
1020 
1021 			maxsize = object->un_pager.vnp.vnp_size - poffset;
1022 			ncount = btoc(maxsize);
1023 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1024 				vm_page_clear_dirty(m[ncount - 1], pgoff,
1025 					PAGE_SIZE - pgoff);
1026 			}
1027 		} else {
1028 			maxsize = 0;
1029 			ncount = 0;
1030 		}
1031 		if (ncount < count) {
1032 			for (i = ncount; i < count; i++) {
1033 				rtvals[i] = VM_PAGER_BAD;
1034 			}
1035 		}
1036 	}
1037 
1038 	/*
1039 	 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1040 	 * rather then a bdwrite() to prevent paging I/O from saturating
1041 	 * the buffer cache.
1042 	 */
1043 	ioflags = IO_VMIO;
1044 	ioflags |= (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) ? IO_SYNC: IO_ASYNC;
1045 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1046 
1047 	aiov.iov_base = (caddr_t) 0;
1048 	aiov.iov_len = maxsize;
1049 	auio.uio_iov = &aiov;
1050 	auio.uio_iovcnt = 1;
1051 	auio.uio_offset = poffset;
1052 	auio.uio_segflg = UIO_NOCOPY;
1053 	auio.uio_rw = UIO_WRITE;
1054 	auio.uio_resid = maxsize;
1055 	auio.uio_td = (struct thread *) 0;
1056 	error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1057 	cnt.v_vnodeout++;
1058 	cnt.v_vnodepgsout += ncount;
1059 
1060 	if (error) {
1061 		printf("vnode_pager_putpages: I/O error %d\n", error);
1062 	}
1063 	if (auio.uio_resid) {
1064 		printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1065 		    auio.uio_resid, (u_long)m[0]->pindex);
1066 	}
1067 	for (i = 0; i < ncount; i++) {
1068 		rtvals[i] = VM_PAGER_OK;
1069 	}
1070 	return rtvals[0];
1071 }
1072 
1073 struct vnode *
1074 vnode_pager_lock(object)
1075 	vm_object_t object;
1076 {
1077 	struct thread *td = curthread;	/* XXX */
1078 
1079 	GIANT_REQUIRED;
1080 
1081 	for (; object != NULL; object = object->backing_object) {
1082 		if (object->type != OBJT_VNODE)
1083 			continue;
1084 		if (object->flags & OBJ_DEAD) {
1085 			return NULL;
1086 		}
1087 
1088 		/* XXX; If object->handle can change, we need to cache it. */
1089 		while (vget(object->handle,
1090 			LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, td)){
1091 			if ((object->flags & OBJ_DEAD) || (object->type != OBJT_VNODE))
1092 				return NULL;
1093 			printf("vnode_pager_lock: retrying\n");
1094 		}
1095 		return object->handle;
1096 	}
1097 	return NULL;
1098 }
1099