xref: /freebsd/sys/vm/vnode_pager.c (revision 7773002178c8dbc52b44e4d705f07706409af8e4)
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  */
42 
43 /*
44  * Page to/from files (vnodes).
45  */
46 
47 /*
48  * TODO:
49  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
50  *	greatly re-simplify the vnode_pager.
51  */
52 
53 #include <sys/cdefs.h>
54 __FBSDID("$FreeBSD$");
55 
56 #include <sys/param.h>
57 #include <sys/systm.h>
58 #include <sys/proc.h>
59 #include <sys/vnode.h>
60 #include <sys/mount.h>
61 #include <sys/bio.h>
62 #include <sys/buf.h>
63 #include <sys/vmmeter.h>
64 #include <sys/conf.h>
65 
66 #include <vm/vm.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_map.h>
71 #include <vm/vnode_pager.h>
72 #include <vm/vm_extern.h>
73 
74 static void vnode_pager_init(void);
75 static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
76 					 int *run);
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 	.pgo_init =	vnode_pager_init,
86 	.pgo_alloc =	vnode_pager_alloc,
87 	.pgo_dealloc =	vnode_pager_dealloc,
88 	.pgo_getpages =	vnode_pager_getpages,
89 	.pgo_putpages =	vnode_pager_putpages,
90 	.pgo_haspage =	vnode_pager_haspage,
91 };
92 
93 int vnode_pbuf_freecnt;
94 
95 static void
96 vnode_pager_init(void)
97 {
98 
99 	vnode_pbuf_freecnt = nswbuf / 2 + 1;
100 }
101 
102 /*
103  * Allocate (or lookup) pager for a vnode.
104  * Handle is a vnode pointer.
105  *
106  * MPSAFE
107  */
108 vm_object_t
109 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
110 		  vm_ooffset_t offset)
111 {
112 	vm_object_t object;
113 	struct vnode *vp;
114 
115 	/*
116 	 * Pageout to vnode, no can do yet.
117 	 */
118 	if (handle == NULL)
119 		return (NULL);
120 
121 	vp = (struct vnode *) handle;
122 
123 	ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
124 
125 	mtx_lock(&Giant);
126 	/*
127 	 * Prevent race condition when allocating the object. This
128 	 * can happen with NFS vnodes since the nfsnode isn't locked.
129 	 */
130 	VI_LOCK(vp);
131 	while (vp->v_iflag & VI_OLOCK) {
132 		vp->v_iflag |= VI_OWANT;
133 		msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0);
134 	}
135 	vp->v_iflag |= VI_OLOCK;
136 	VI_UNLOCK(vp);
137 
138 	/*
139 	 * If the object is being terminated, wait for it to
140 	 * go away.
141 	 */
142 	while ((object = vp->v_object) != NULL) {
143 		VM_OBJECT_LOCK(object);
144 		if ((object->flags & OBJ_DEAD) == 0)
145 			break;
146 		msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0);
147 	}
148 
149 	if (vp->v_usecount == 0)
150 		panic("vnode_pager_alloc: no vnode reference");
151 
152 	if (object == NULL) {
153 		/*
154 		 * And an object of the appropriate size
155 		 */
156 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
157 
158 		object->un_pager.vnp.vnp_size = size;
159 
160 		object->handle = handle;
161 		vp->v_object = object;
162 	} else {
163 		object->ref_count++;
164 		VM_OBJECT_UNLOCK(object);
165 	}
166 	VI_LOCK(vp);
167 	vp->v_usecount++;
168 	vp->v_iflag &= ~VI_OLOCK;
169 	if (vp->v_iflag & VI_OWANT) {
170 		vp->v_iflag &= ~VI_OWANT;
171 		wakeup(vp);
172 	}
173 	VI_UNLOCK(vp);
174 	mtx_unlock(&Giant);
175 	return (object);
176 }
177 
178 /*
179  *	The object must be locked.
180  */
181 static void
182 vnode_pager_dealloc(object)
183 	vm_object_t object;
184 {
185 	struct vnode *vp = object->handle;
186 
187 	if (vp == NULL)
188 		panic("vnode_pager_dealloc: pager already dealloced");
189 
190 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
191 	vm_object_pip_wait(object, "vnpdea");
192 
193 	object->handle = NULL;
194 	object->type = OBJT_DEAD;
195 	ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
196 	vp->v_object = NULL;
197 	vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF);
198 }
199 
200 static boolean_t
201 vnode_pager_haspage(object, pindex, before, after)
202 	vm_object_t object;
203 	vm_pindex_t pindex;
204 	int *before;
205 	int *after;
206 {
207 	struct vnode *vp = object->handle;
208 	daddr_t bn;
209 	int err;
210 	daddr_t reqblock;
211 	int poff;
212 	int bsize;
213 	int pagesperblock, blocksperpage;
214 
215 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
216 	/*
217 	 * If no vp or vp is doomed or marked transparent to VM, we do not
218 	 * have the page.
219 	 */
220 	if (vp == NULL)
221 		return FALSE;
222 
223 	VI_LOCK(vp);
224 	if (vp->v_iflag & VI_DOOMED) {
225 		VI_UNLOCK(vp);
226 		return FALSE;
227 	}
228 	VI_UNLOCK(vp);
229 	/*
230 	 * If filesystem no longer mounted or offset beyond end of file we do
231 	 * not have the page.
232 	 */
233 	if ((vp->v_mount == NULL) ||
234 	    (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
235 		return FALSE;
236 
237 	bsize = vp->v_mount->mnt_stat.f_iosize;
238 	pagesperblock = bsize / PAGE_SIZE;
239 	blocksperpage = 0;
240 	if (pagesperblock > 0) {
241 		reqblock = pindex / pagesperblock;
242 	} else {
243 		blocksperpage = (PAGE_SIZE / bsize);
244 		reqblock = pindex * blocksperpage;
245 	}
246 	VM_OBJECT_UNLOCK(object);
247 	err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
248 	VM_OBJECT_LOCK(object);
249 	if (err)
250 		return TRUE;
251 	if (bn == -1)
252 		return FALSE;
253 	if (pagesperblock > 0) {
254 		poff = pindex - (reqblock * pagesperblock);
255 		if (before) {
256 			*before *= pagesperblock;
257 			*before += poff;
258 		}
259 		if (after) {
260 			int numafter;
261 			*after *= pagesperblock;
262 			numafter = pagesperblock - (poff + 1);
263 			if (IDX_TO_OFF(pindex + numafter) >
264 			    object->un_pager.vnp.vnp_size) {
265 				numafter =
266 		    		    OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
267 				    pindex;
268 			}
269 			*after += numafter;
270 		}
271 	} else {
272 		if (before) {
273 			*before /= blocksperpage;
274 		}
275 
276 		if (after) {
277 			*after /= blocksperpage;
278 		}
279 	}
280 	return TRUE;
281 }
282 
283 /*
284  * Lets the VM system know about a change in size for a file.
285  * We adjust our own internal size and flush any cached pages in
286  * the associated object that are affected by the size change.
287  *
288  * Note: this routine may be invoked as a result of a pager put
289  * operation (possibly at object termination time), so we must be careful.
290  */
291 void
292 vnode_pager_setsize(vp, nsize)
293 	struct vnode *vp;
294 	vm_ooffset_t nsize;
295 {
296 	vm_object_t object;
297 	vm_page_t m;
298 	vm_pindex_t nobjsize;
299 
300 	if ((object = vp->v_object) == NULL)
301 		return;
302 	VM_OBJECT_LOCK(object);
303 	if (nsize == object->un_pager.vnp.vnp_size) {
304 		/*
305 		 * Hasn't changed size
306 		 */
307 		VM_OBJECT_UNLOCK(object);
308 		return;
309 	}
310 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
311 	if (nsize < object->un_pager.vnp.vnp_size) {
312 		/*
313 		 * File has shrunk. Toss any cached pages beyond the new EOF.
314 		 */
315 		if (nobjsize < object->size)
316 			vm_object_page_remove(object, nobjsize, object->size,
317 			    FALSE);
318 		/*
319 		 * this gets rid of garbage at the end of a page that is now
320 		 * only partially backed by the vnode.
321 		 *
322 		 * XXX for some reason (I don't know yet), if we take a
323 		 * completely invalid page and mark it partially valid
324 		 * it can screw up NFS reads, so we don't allow the case.
325 		 */
326 		if ((nsize & PAGE_MASK) &&
327 		    (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
328 		    m->valid != 0) {
329 			int base = (int)nsize & PAGE_MASK;
330 			int size = PAGE_SIZE - base;
331 
332 			/*
333 			 * Clear out partial-page garbage in case
334 			 * the page has been mapped.
335 			 */
336 			pmap_zero_page_area(m, base, size);
337 
338 			/*
339 			 * XXX work around SMP data integrity race
340 			 * by unmapping the page from user processes.
341 			 * The garbage we just cleared may be mapped
342 			 * to a user process running on another cpu
343 			 * and this code is not running through normal
344 			 * I/O channels which handle SMP issues for
345 			 * us, so unmap page to synchronize all cpus.
346 			 *
347 			 * XXX should vm_pager_unmap_page() have
348 			 * dealt with this?
349 			 */
350 			vm_page_lock_queues();
351 			pmap_remove_all(m);
352 
353 			/*
354 			 * Clear out partial-page dirty bits.  This
355 			 * has the side effect of setting the valid
356 			 * bits, but that is ok.  There are a bunch
357 			 * of places in the VM system where we expected
358 			 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
359 			 * case is one of them.  If the page is still
360 			 * partially dirty, make it fully dirty.
361 			 *
362 			 * note that we do not clear out the valid
363 			 * bits.  This would prevent bogus_page
364 			 * replacement from working properly.
365 			 */
366 			vm_page_set_validclean(m, base, size);
367 			if (m->dirty != 0)
368 				m->dirty = VM_PAGE_BITS_ALL;
369 			vm_page_unlock_queues();
370 		}
371 	}
372 	object->un_pager.vnp.vnp_size = nsize;
373 	object->size = nobjsize;
374 	VM_OBJECT_UNLOCK(object);
375 }
376 
377 /*
378  * calculate the linear (byte) disk address of specified virtual
379  * file address
380  */
381 static vm_offset_t
382 vnode_pager_addr(vp, address, run)
383 	struct vnode *vp;
384 	vm_ooffset_t address;
385 	int *run;
386 {
387 	int rtaddress;
388 	int bsize;
389 	daddr_t block;
390 	int err;
391 	daddr_t vblock;
392 	int voffset;
393 
394 	GIANT_REQUIRED;
395 	if ((int) address < 0)
396 		return -1;
397 
398 	if (vp->v_mount == NULL)
399 		return -1;
400 
401 	bsize = vp->v_mount->mnt_stat.f_iosize;
402 	vblock = address / bsize;
403 	voffset = address % bsize;
404 
405 	err = VOP_BMAP(vp, vblock, NULL, &block, run, NULL);
406 
407 	if (err || (block == -1))
408 		rtaddress = -1;
409 	else {
410 		rtaddress = block + voffset / DEV_BSIZE;
411 		if (run) {
412 			*run += 1;
413 			*run *= bsize/PAGE_SIZE;
414 			*run -= voffset/PAGE_SIZE;
415 		}
416 	}
417 
418 	return rtaddress;
419 }
420 
421 /*
422  * small block filesystem vnode pager input
423  */
424 static int
425 vnode_pager_input_smlfs(object, m)
426 	vm_object_t object;
427 	vm_page_t m;
428 {
429 	int i;
430 	struct vnode *dp, *vp;
431 	struct buf *bp;
432 	vm_offset_t kva;
433 	int fileaddr;
434 	vm_offset_t bsize;
435 	int error = 0;
436 
437 	GIANT_REQUIRED;
438 
439 	vp = object->handle;
440 	if (vp->v_mount == NULL)
441 		return VM_PAGER_BAD;
442 
443 	bsize = vp->v_mount->mnt_stat.f_iosize;
444 
445 	VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
446 
447 	kva = vm_pager_map_page(m);
448 
449 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
450 		vm_ooffset_t address;
451 
452 		if (vm_page_bits(i * bsize, bsize) & m->valid)
453 			continue;
454 
455 		address = IDX_TO_OFF(m->pindex) + i * bsize;
456 		if (address >= object->un_pager.vnp.vnp_size) {
457 			fileaddr = -1;
458 		} else {
459 			fileaddr = vnode_pager_addr(vp, address, NULL);
460 		}
461 		if (fileaddr != -1) {
462 			bp = getpbuf(&vnode_pbuf_freecnt);
463 
464 			/* build a minimal buffer header */
465 			bp->b_iocmd = BIO_READ;
466 			bp->b_iodone = bdone;
467 			KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
468 			KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
469 			bp->b_rcred = crhold(curthread->td_ucred);
470 			bp->b_wcred = crhold(curthread->td_ucred);
471 			bp->b_data = (caddr_t) kva + i * bsize;
472 			bp->b_blkno = fileaddr;
473 			pbgetvp(dp, bp);
474 			bp->b_bcount = bsize;
475 			bp->b_bufsize = bsize;
476 			bp->b_runningbufspace = bp->b_bufsize;
477 			runningbufspace += bp->b_runningbufspace;
478 
479 			/* do the input */
480 			bp->b_iooffset = dbtob(bp->b_blkno);
481 			VOP_SPECSTRATEGY(bp->b_vp, bp);
482 
483 			/* we definitely need to be at splvm here */
484 
485 			bwait(bp, PVM, "vnsrd");
486 
487 			if ((bp->b_ioflags & BIO_ERROR) != 0)
488 				error = EIO;
489 
490 			/*
491 			 * free the buffer header back to the swap buffer pool
492 			 */
493 			relpbuf(bp, &vnode_pbuf_freecnt);
494 			if (error)
495 				break;
496 
497 			VM_OBJECT_LOCK(object);
498 			vm_page_lock_queues();
499 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
500 			vm_page_unlock_queues();
501 			VM_OBJECT_UNLOCK(object);
502 		} else {
503 			VM_OBJECT_LOCK(object);
504 			vm_page_lock_queues();
505 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
506 			vm_page_unlock_queues();
507 			VM_OBJECT_UNLOCK(object);
508 			bzero((caddr_t) kva + i * bsize, bsize);
509 		}
510 	}
511 	vm_pager_unmap_page(kva);
512 	vm_page_lock_queues();
513 	pmap_clear_modify(m);
514 	vm_page_flag_clear(m, PG_ZERO);
515 	vm_page_unlock_queues();
516 	if (error) {
517 		return VM_PAGER_ERROR;
518 	}
519 	return VM_PAGER_OK;
520 
521 }
522 
523 
524 /*
525  * old style vnode pager output routine
526  */
527 static int
528 vnode_pager_input_old(object, m)
529 	vm_object_t object;
530 	vm_page_t m;
531 {
532 	struct uio auio;
533 	struct iovec aiov;
534 	int error;
535 	int size;
536 	vm_offset_t kva;
537 	struct vnode *vp;
538 
539 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
540 	error = 0;
541 
542 	/*
543 	 * Return failure if beyond current EOF
544 	 */
545 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
546 		return VM_PAGER_BAD;
547 	} else {
548 		size = PAGE_SIZE;
549 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
550 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
551 		vp = object->handle;
552 		VM_OBJECT_UNLOCK(object);
553 
554 		/*
555 		 * Allocate a kernel virtual address and initialize so that
556 		 * we can use VOP_READ/WRITE routines.
557 		 */
558 		kva = vm_pager_map_page(m);
559 
560 		aiov.iov_base = (caddr_t) kva;
561 		aiov.iov_len = size;
562 		auio.uio_iov = &aiov;
563 		auio.uio_iovcnt = 1;
564 		auio.uio_offset = IDX_TO_OFF(m->pindex);
565 		auio.uio_segflg = UIO_SYSSPACE;
566 		auio.uio_rw = UIO_READ;
567 		auio.uio_resid = size;
568 		auio.uio_td = curthread;
569 
570 		error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
571 		if (!error) {
572 			int count = size - auio.uio_resid;
573 
574 			if (count == 0)
575 				error = EINVAL;
576 			else if (count != PAGE_SIZE)
577 				bzero((caddr_t) kva + count, PAGE_SIZE - count);
578 		}
579 		vm_pager_unmap_page(kva);
580 
581 		VM_OBJECT_LOCK(object);
582 	}
583 	vm_page_lock_queues();
584 	pmap_clear_modify(m);
585 	vm_page_undirty(m);
586 	vm_page_flag_clear(m, PG_ZERO);
587 	vm_page_unlock_queues();
588 	if (!error)
589 		m->valid = VM_PAGE_BITS_ALL;
590 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
591 }
592 
593 /*
594  * generic vnode pager input routine
595  */
596 
597 /*
598  * Local media VFS's that do not implement their own VOP_GETPAGES
599  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
600  * to implement the previous behaviour.
601  *
602  * All other FS's should use the bypass to get to the local media
603  * backing vp's VOP_GETPAGES.
604  */
605 static int
606 vnode_pager_getpages(object, m, count, reqpage)
607 	vm_object_t object;
608 	vm_page_t *m;
609 	int count;
610 	int reqpage;
611 {
612 	int rtval;
613 	struct vnode *vp;
614 	int bytes = count * PAGE_SIZE;
615 
616 	vp = object->handle;
617 	VM_OBJECT_UNLOCK(object);
618 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
619 	KASSERT(rtval != EOPNOTSUPP,
620 	    ("vnode_pager: FS getpages not implemented\n"));
621 	VM_OBJECT_LOCK(object);
622 	return rtval;
623 }
624 
625 /*
626  * This is now called from local media FS's to operate against their
627  * own vnodes if they fail to implement VOP_GETPAGES.
628  */
629 int
630 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
631 	struct vnode *vp;
632 	vm_page_t *m;
633 	int bytecount;
634 	int reqpage;
635 {
636 	vm_object_t object;
637 	vm_offset_t kva;
638 	off_t foff, tfoff, nextoff;
639 	int i, j, size, bsize, first, firstaddr;
640 	struct vnode *dp;
641 	int runpg;
642 	int runend;
643 	struct buf *bp;
644 	int count;
645 	int error = 0;
646 
647 	GIANT_REQUIRED;
648 	object = vp->v_object;
649 	count = bytecount / PAGE_SIZE;
650 
651 	if (vp->v_mount == NULL)
652 		return VM_PAGER_BAD;
653 
654 	bsize = vp->v_mount->mnt_stat.f_iosize;
655 
656 	/* get the UNDERLYING device for the file with VOP_BMAP() */
657 
658 	/*
659 	 * originally, we did not check for an error return value -- assuming
660 	 * an fs always has a bmap entry point -- that assumption is wrong!!!
661 	 */
662 	foff = IDX_TO_OFF(m[reqpage]->pindex);
663 
664 	/*
665 	 * if we can't bmap, use old VOP code
666 	 */
667 	if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
668 		VM_OBJECT_LOCK(object);
669 		vm_page_lock_queues();
670 		for (i = 0; i < count; i++)
671 			if (i != reqpage)
672 				vm_page_free(m[i]);
673 		vm_page_unlock_queues();
674 		cnt.v_vnodein++;
675 		cnt.v_vnodepgsin++;
676 		error = vnode_pager_input_old(object, m[reqpage]);
677 		VM_OBJECT_UNLOCK(object);
678 		return (error);
679 
680 		/*
681 		 * if the blocksize is smaller than a page size, then use
682 		 * special small filesystem code.  NFS sometimes has a small
683 		 * blocksize, but it can handle large reads itself.
684 		 */
685 	} else if ((PAGE_SIZE / bsize) > 1 &&
686 	    (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
687 		VM_OBJECT_LOCK(object);
688 		vm_page_lock_queues();
689 		for (i = 0; i < count; i++)
690 			if (i != reqpage)
691 				vm_page_free(m[i]);
692 		vm_page_unlock_queues();
693 		VM_OBJECT_UNLOCK(object);
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 	VM_OBJECT_LOCK(object);
705 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
706 		vm_page_lock_queues();
707 		for (i = 0; i < count; i++)
708 			if (i != reqpage)
709 				vm_page_free(m[i]);
710 		vm_page_unlock_queues();
711 		VM_OBJECT_UNLOCK(object);
712 		return VM_PAGER_OK;
713 	}
714 	m[reqpage]->valid = 0;
715 	VM_OBJECT_UNLOCK(object);
716 
717 	/*
718 	 * here on direct device I/O
719 	 */
720 	firstaddr = -1;
721 
722 	/*
723 	 * calculate the run that includes the required page
724 	 */
725 	for (first = 0, i = 0; i < count; i = runend) {
726 		firstaddr = vnode_pager_addr(vp,
727 			IDX_TO_OFF(m[i]->pindex), &runpg);
728 		if (firstaddr == -1) {
729 			VM_OBJECT_LOCK(object);
730 			if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
731 				panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
732 				    firstaddr, (uintmax_t)(foff >> 32),
733 				    (uintmax_t)foff,
734 				    (uintmax_t)
735 				    (object->un_pager.vnp.vnp_size >> 32),
736 				    (uintmax_t)object->un_pager.vnp.vnp_size);
737 			}
738 			vm_page_lock_queues();
739 			vm_page_free(m[i]);
740 			vm_page_unlock_queues();
741 			VM_OBJECT_UNLOCK(object);
742 			runend = i + 1;
743 			first = runend;
744 			continue;
745 		}
746 		runend = i + runpg;
747 		if (runend <= reqpage) {
748 			VM_OBJECT_LOCK(object);
749 			vm_page_lock_queues();
750 			for (j = i; j < runend; j++)
751 				vm_page_free(m[j]);
752 			vm_page_unlock_queues();
753 			VM_OBJECT_UNLOCK(object);
754 		} else {
755 			if (runpg < (count - first)) {
756 				VM_OBJECT_LOCK(object);
757 				vm_page_lock_queues();
758 				for (i = first + runpg; i < count; i++)
759 					vm_page_free(m[i]);
760 				vm_page_unlock_queues();
761 				VM_OBJECT_UNLOCK(object);
762 				count = first + runpg;
763 			}
764 			break;
765 		}
766 		first = runend;
767 	}
768 
769 	/*
770 	 * the first and last page have been calculated now, move input pages
771 	 * to be zero based...
772 	 */
773 	if (first != 0) {
774 		for (i = first; i < count; i++) {
775 			m[i - first] = m[i];
776 		}
777 		count -= first;
778 		reqpage -= first;
779 	}
780 
781 	/*
782 	 * calculate the file virtual address for the transfer
783 	 */
784 	foff = IDX_TO_OFF(m[0]->pindex);
785 
786 	/*
787 	 * calculate the size of the transfer
788 	 */
789 	size = count * PAGE_SIZE;
790 	if ((foff + size) > object->un_pager.vnp.vnp_size)
791 		size = object->un_pager.vnp.vnp_size - foff;
792 
793 	/*
794 	 * round up physical size for real devices.
795 	 */
796 	if (dp->v_type == VBLK || dp->v_type == VCHR) {
797 		int secmask = dp->v_rdev->si_bsize_phys - 1;
798 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
799 		size = (size + secmask) & ~secmask;
800 	}
801 
802 	bp = getpbuf(&vnode_pbuf_freecnt);
803 	kva = (vm_offset_t) bp->b_data;
804 
805 	/*
806 	 * and map the pages to be read into the kva
807 	 */
808 	pmap_qenter(kva, m, count);
809 
810 	/* build a minimal buffer header */
811 	bp->b_iocmd = BIO_READ;
812 	bp->b_iodone = bdone;
813 	KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
814 	KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
815 	bp->b_rcred = crhold(curthread->td_ucred);
816 	bp->b_wcred = crhold(curthread->td_ucred);
817 	bp->b_blkno = firstaddr;
818 	pbgetvp(dp, bp);
819 	bp->b_bcount = size;
820 	bp->b_bufsize = size;
821 	bp->b_runningbufspace = bp->b_bufsize;
822 	runningbufspace += bp->b_runningbufspace;
823 
824 	cnt.v_vnodein++;
825 	cnt.v_vnodepgsin += count;
826 
827 	/* do the input */
828 	bp->b_iooffset = dbtob(bp->b_blkno);
829 	if (dp->v_type == VCHR)
830 		VOP_SPECSTRATEGY(bp->b_vp, bp);
831 	else
832 		VOP_STRATEGY(bp->b_vp, bp);
833 
834 	bwait(bp, PVM, "vnread");
835 
836 	if ((bp->b_ioflags & BIO_ERROR) != 0)
837 		error = EIO;
838 
839 	if (!error) {
840 		if (size != count * PAGE_SIZE)
841 			bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
842 	}
843 	pmap_qremove(kva, count);
844 
845 	/*
846 	 * free the buffer header back to the swap buffer pool
847 	 */
848 	relpbuf(bp, &vnode_pbuf_freecnt);
849 
850 	VM_OBJECT_LOCK(object);
851 	vm_page_lock_queues();
852 	for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
853 		vm_page_t mt;
854 
855 		nextoff = tfoff + PAGE_SIZE;
856 		mt = m[i];
857 
858 		if (nextoff <= object->un_pager.vnp.vnp_size) {
859 			/*
860 			 * Read filled up entire page.
861 			 */
862 			mt->valid = VM_PAGE_BITS_ALL;
863 			vm_page_undirty(mt);	/* should be an assert? XXX */
864 			pmap_clear_modify(mt);
865 		} else {
866 			/*
867 			 * Read did not fill up entire page.  Since this
868 			 * is getpages, the page may be mapped, so we have
869 			 * to zero the invalid portions of the page even
870 			 * though we aren't setting them valid.
871 			 *
872 			 * Currently we do not set the entire page valid,
873 			 * we just try to clear the piece that we couldn't
874 			 * read.
875 			 */
876 			vm_page_set_validclean(mt, 0,
877 			    object->un_pager.vnp.vnp_size - tfoff);
878 			/* handled by vm_fault now */
879 			/* vm_page_zero_invalid(mt, FALSE); */
880 		}
881 
882 		vm_page_flag_clear(mt, PG_ZERO);
883 		if (i != reqpage) {
884 
885 			/*
886 			 * whether or not to leave the page activated is up in
887 			 * the air, but we should put the page on a page queue
888 			 * somewhere. (it already is in the object). Result:
889 			 * It appears that empirical results show that
890 			 * deactivating pages is best.
891 			 */
892 
893 			/*
894 			 * just in case someone was asking for this page we
895 			 * now tell them that it is ok to use
896 			 */
897 			if (!error) {
898 				if (mt->flags & PG_WANTED)
899 					vm_page_activate(mt);
900 				else
901 					vm_page_deactivate(mt);
902 				vm_page_wakeup(mt);
903 			} else {
904 				vm_page_free(mt);
905 			}
906 		}
907 	}
908 	vm_page_unlock_queues();
909 	VM_OBJECT_UNLOCK(object);
910 	if (error) {
911 		printf("vnode_pager_getpages: I/O read error\n");
912 	}
913 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
914 }
915 
916 /*
917  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
918  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
919  * vnode_pager_generic_putpages() to implement the previous behaviour.
920  *
921  * All other FS's should use the bypass to get to the local media
922  * backing vp's VOP_PUTPAGES.
923  */
924 static void
925 vnode_pager_putpages(object, m, count, sync, rtvals)
926 	vm_object_t object;
927 	vm_page_t *m;
928 	int count;
929 	boolean_t sync;
930 	int *rtvals;
931 {
932 	int rtval;
933 	struct vnode *vp;
934 	struct mount *mp;
935 	int bytes = count * PAGE_SIZE;
936 
937 	GIANT_REQUIRED;
938 	/*
939 	 * Force synchronous operation if we are extremely low on memory
940 	 * to prevent a low-memory deadlock.  VOP operations often need to
941 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
942 	 * operation ).  The swapper handles the case by limiting the amount
943 	 * of asynchronous I/O, but that sort of solution doesn't scale well
944 	 * for the vnode pager without a lot of work.
945 	 *
946 	 * Also, the backing vnode's iodone routine may not wake the pageout
947 	 * daemon up.  This should be probably be addressed XXX.
948 	 */
949 
950 	if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
951 		sync |= OBJPC_SYNC;
952 
953 	/*
954 	 * Call device-specific putpages function
955 	 */
956 	vp = object->handle;
957 	VM_OBJECT_UNLOCK(object);
958 	if (vp->v_type != VREG)
959 		mp = NULL;
960 	(void)vn_start_write(vp, &mp, V_WAIT);
961 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
962 	KASSERT(rtval != EOPNOTSUPP,
963 	    ("vnode_pager: stale FS putpages\n"));
964 	vn_finished_write(mp);
965 	VM_OBJECT_LOCK(object);
966 }
967 
968 
969 /*
970  * This is now called from local media FS's to operate against their
971  * own vnodes if they fail to implement VOP_PUTPAGES.
972  *
973  * This is typically called indirectly via the pageout daemon and
974  * clustering has already typically occured, so in general we ask the
975  * underlying filesystem to write the data out asynchronously rather
976  * then delayed.
977  */
978 int
979 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
980 	struct vnode *vp;
981 	vm_page_t *m;
982 	int bytecount;
983 	int flags;
984 	int *rtvals;
985 {
986 	int i;
987 	vm_object_t object;
988 	int count;
989 
990 	int maxsize, ncount;
991 	vm_ooffset_t poffset;
992 	struct uio auio;
993 	struct iovec aiov;
994 	int error;
995 	int ioflags;
996 
997 	GIANT_REQUIRED;
998 	object = vp->v_object;
999 	count = bytecount / PAGE_SIZE;
1000 
1001 	for (i = 0; i < count; i++)
1002 		rtvals[i] = VM_PAGER_AGAIN;
1003 
1004 	if ((int) m[0]->pindex < 0) {
1005 		printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1006 			(long)m[0]->pindex, (u_long)m[0]->dirty);
1007 		rtvals[0] = VM_PAGER_BAD;
1008 		return VM_PAGER_BAD;
1009 	}
1010 
1011 	maxsize = count * PAGE_SIZE;
1012 	ncount = count;
1013 
1014 	poffset = IDX_TO_OFF(m[0]->pindex);
1015 
1016 	/*
1017 	 * If the page-aligned write is larger then the actual file we
1018 	 * have to invalidate pages occuring beyond the file EOF.  However,
1019 	 * there is an edge case where a file may not be page-aligned where
1020 	 * the last page is partially invalid.  In this case the filesystem
1021 	 * may not properly clear the dirty bits for the entire page (which
1022 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1023 	 * With the page locked we are free to fix-up the dirty bits here.
1024 	 *
1025 	 * We do not under any circumstances truncate the valid bits, as
1026 	 * this will screw up bogus page replacement.
1027 	 */
1028 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1029 		if (object->un_pager.vnp.vnp_size > poffset) {
1030 			int pgoff;
1031 
1032 			maxsize = object->un_pager.vnp.vnp_size - poffset;
1033 			ncount = btoc(maxsize);
1034 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1035 				vm_page_lock_queues();
1036 				vm_page_clear_dirty(m[ncount - 1], pgoff,
1037 					PAGE_SIZE - pgoff);
1038 				vm_page_unlock_queues();
1039 			}
1040 		} else {
1041 			maxsize = 0;
1042 			ncount = 0;
1043 		}
1044 		if (ncount < count) {
1045 			for (i = ncount; i < count; i++) {
1046 				rtvals[i] = VM_PAGER_BAD;
1047 			}
1048 		}
1049 	}
1050 
1051 	/*
1052 	 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1053 	 * rather then a bdwrite() to prevent paging I/O from saturating
1054 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
1055 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1056 	 * the system decides how to cluster.
1057 	 */
1058 	ioflags = IO_VMIO;
1059 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1060 		ioflags |= IO_SYNC;
1061 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1062 		ioflags |= IO_ASYNC;
1063 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1064 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1065 
1066 	aiov.iov_base = (caddr_t) 0;
1067 	aiov.iov_len = maxsize;
1068 	auio.uio_iov = &aiov;
1069 	auio.uio_iovcnt = 1;
1070 	auio.uio_offset = poffset;
1071 	auio.uio_segflg = UIO_NOCOPY;
1072 	auio.uio_rw = UIO_WRITE;
1073 	auio.uio_resid = maxsize;
1074 	auio.uio_td = (struct thread *) 0;
1075 	error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1076 	cnt.v_vnodeout++;
1077 	cnt.v_vnodepgsout += ncount;
1078 
1079 	if (error) {
1080 		printf("vnode_pager_putpages: I/O error %d\n", error);
1081 	}
1082 	if (auio.uio_resid) {
1083 		printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1084 		    auio.uio_resid, (u_long)m[0]->pindex);
1085 	}
1086 	for (i = 0; i < ncount; i++) {
1087 		rtvals[i] = VM_PAGER_OK;
1088 	}
1089 	return rtvals[0];
1090 }
1091 
1092 struct vnode *
1093 vnode_pager_lock(vm_object_t first_object)
1094 {
1095 	struct vnode *vp;
1096 	vm_object_t backing_object, object;
1097 
1098 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
1099 	for (object = first_object; object != NULL; object = backing_object) {
1100 		if (object->type != OBJT_VNODE) {
1101 			if ((backing_object = object->backing_object) != NULL)
1102 				VM_OBJECT_LOCK(backing_object);
1103 			if (object != first_object)
1104 				VM_OBJECT_UNLOCK(object);
1105 			continue;
1106 		}
1107 	retry:
1108 		if (object->flags & OBJ_DEAD) {
1109 			if (object != first_object)
1110 				VM_OBJECT_UNLOCK(object);
1111 			return NULL;
1112 		}
1113 		vp = object->handle;
1114 		VI_LOCK(vp);
1115 		VM_OBJECT_UNLOCK(object);
1116 		if (first_object != object)
1117 			VM_OBJECT_UNLOCK(first_object);
1118 		if (vget(vp, LK_CANRECURSE | LK_INTERLOCK | LK_NOPAUSE |
1119 		    LK_RETRY | LK_SHARED, curthread)) {
1120 			VM_OBJECT_LOCK(first_object);
1121 			if (object != first_object)
1122 				VM_OBJECT_LOCK(object);
1123 			if (object->type != OBJT_VNODE) {
1124 				if (object != first_object)
1125 					VM_OBJECT_UNLOCK(object);
1126 				return NULL;
1127 			}
1128 			printf("vnode_pager_lock: retrying\n");
1129 			goto retry;
1130 		}
1131 		VM_OBJECT_LOCK(first_object);
1132 		return (vp);
1133 	}
1134 	return NULL;
1135 }
1136