xref: /freebsd/sys/vm/vnode_pager.c (revision 52ec752989b2e6d4e9a59a8ff25d8ff596d85e62)
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 	/*
126 	 * Prevent race condition when allocating the object. This
127 	 * can happen with NFS vnodes since the nfsnode isn't locked.
128 	 */
129 	VI_LOCK(vp);
130 	while (vp->v_iflag & VI_OLOCK) {
131 		vp->v_iflag |= VI_OWANT;
132 		msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0);
133 	}
134 	vp->v_iflag |= VI_OLOCK;
135 	VI_UNLOCK(vp);
136 
137 	/*
138 	 * If the object is being terminated, wait for it to
139 	 * go away.
140 	 */
141 	while ((object = vp->v_object) != NULL) {
142 		VM_OBJECT_LOCK(object);
143 		if ((object->flags & OBJ_DEAD) == 0)
144 			break;
145 		msleep(object, VM_OBJECT_MTX(object), PDROP | 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 		VM_OBJECT_UNLOCK(object);
164 	}
165 	VI_LOCK(vp);
166 	vp->v_usecount++;
167 	vp->v_iflag &= ~VI_OLOCK;
168 	if (vp->v_iflag & VI_OWANT) {
169 		vp->v_iflag &= ~VI_OWANT;
170 		wakeup(vp);
171 	}
172 	VI_UNLOCK(vp);
173 	return (object);
174 }
175 
176 /*
177  *	The object must be locked.
178  */
179 static void
180 vnode_pager_dealloc(object)
181 	vm_object_t object;
182 {
183 	struct vnode *vp = object->handle;
184 
185 	if (vp == NULL)
186 		panic("vnode_pager_dealloc: pager already dealloced");
187 
188 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
189 	vm_object_pip_wait(object, "vnpdea");
190 
191 	object->handle = NULL;
192 	object->type = OBJT_DEAD;
193 	ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
194 	vp->v_object = NULL;
195 	vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF);
196 }
197 
198 static boolean_t
199 vnode_pager_haspage(object, pindex, before, after)
200 	vm_object_t object;
201 	vm_pindex_t pindex;
202 	int *before;
203 	int *after;
204 {
205 	struct vnode *vp = object->handle;
206 	daddr_t bn;
207 	int err;
208 	daddr_t reqblock;
209 	int poff;
210 	int bsize;
211 	int pagesperblock, blocksperpage;
212 
213 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
214 	/*
215 	 * If no vp or vp is doomed or marked transparent to VM, we do not
216 	 * have the page.
217 	 */
218 	if (vp == NULL)
219 		return FALSE;
220 
221 	VI_LOCK(vp);
222 	if (vp->v_iflag & VI_DOOMED) {
223 		VI_UNLOCK(vp);
224 		return FALSE;
225 	}
226 	VI_UNLOCK(vp);
227 	/*
228 	 * If filesystem no longer mounted or offset beyond end of file we do
229 	 * not have the page.
230 	 */
231 	if ((vp->v_mount == NULL) ||
232 	    (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
233 		return FALSE;
234 
235 	bsize = vp->v_mount->mnt_stat.f_iosize;
236 	pagesperblock = bsize / PAGE_SIZE;
237 	blocksperpage = 0;
238 	if (pagesperblock > 0) {
239 		reqblock = pindex / pagesperblock;
240 	} else {
241 		blocksperpage = (PAGE_SIZE / bsize);
242 		reqblock = pindex * blocksperpage;
243 	}
244 	VM_OBJECT_UNLOCK(object);
245 	err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before);
246 	VM_OBJECT_LOCK(object);
247 	if (err)
248 		return TRUE;
249 	if (bn == -1)
250 		return FALSE;
251 	if (pagesperblock > 0) {
252 		poff = pindex - (reqblock * pagesperblock);
253 		if (before) {
254 			*before *= pagesperblock;
255 			*before += poff;
256 		}
257 		if (after) {
258 			int numafter;
259 			*after *= pagesperblock;
260 			numafter = pagesperblock - (poff + 1);
261 			if (IDX_TO_OFF(pindex + numafter) >
262 			    object->un_pager.vnp.vnp_size) {
263 				numafter =
264 		    		    OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
265 				    pindex;
266 			}
267 			*after += numafter;
268 		}
269 	} else {
270 		if (before) {
271 			*before /= blocksperpage;
272 		}
273 
274 		if (after) {
275 			*after /= blocksperpage;
276 		}
277 	}
278 	return TRUE;
279 }
280 
281 /*
282  * Lets the VM system know about a change in size for a file.
283  * We adjust our own internal size and flush any cached pages in
284  * the associated object that are affected by the size change.
285  *
286  * Note: this routine may be invoked as a result of a pager put
287  * operation (possibly at object termination time), so we must be careful.
288  */
289 void
290 vnode_pager_setsize(vp, nsize)
291 	struct vnode *vp;
292 	vm_ooffset_t nsize;
293 {
294 	vm_object_t object;
295 	vm_page_t m;
296 	vm_pindex_t nobjsize;
297 
298 	if ((object = vp->v_object) == NULL)
299 		return;
300 	VM_OBJECT_LOCK(object);
301 	if (nsize == object->un_pager.vnp.vnp_size) {
302 		/*
303 		 * Hasn't changed size
304 		 */
305 		VM_OBJECT_UNLOCK(object);
306 		return;
307 	}
308 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
309 	if (nsize < object->un_pager.vnp.vnp_size) {
310 		/*
311 		 * File has shrunk. Toss any cached pages beyond the new EOF.
312 		 */
313 		if (nobjsize < object->size)
314 			vm_object_page_remove(object, nobjsize, object->size,
315 			    FALSE);
316 		/*
317 		 * this gets rid of garbage at the end of a page that is now
318 		 * only partially backed by the vnode.
319 		 *
320 		 * XXX for some reason (I don't know yet), if we take a
321 		 * completely invalid page and mark it partially valid
322 		 * it can screw up NFS reads, so we don't allow the case.
323 		 */
324 		if ((nsize & PAGE_MASK) &&
325 		    (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL &&
326 		    m->valid != 0) {
327 			int base = (int)nsize & PAGE_MASK;
328 			int size = PAGE_SIZE - base;
329 
330 			/*
331 			 * Clear out partial-page garbage in case
332 			 * the page has been mapped.
333 			 */
334 			pmap_zero_page_area(m, base, size);
335 
336 			/*
337 			 * XXX work around SMP data integrity race
338 			 * by unmapping the page from user processes.
339 			 * The garbage we just cleared may be mapped
340 			 * to a user process running on another cpu
341 			 * and this code is not running through normal
342 			 * I/O channels which handle SMP issues for
343 			 * us, so unmap page to synchronize all cpus.
344 			 *
345 			 * XXX should vm_pager_unmap_page() have
346 			 * dealt with this?
347 			 */
348 			vm_page_lock_queues();
349 			pmap_remove_all(m);
350 
351 			/*
352 			 * Clear out partial-page dirty bits.  This
353 			 * has the side effect of setting the valid
354 			 * bits, but that is ok.  There are a bunch
355 			 * of places in the VM system where we expected
356 			 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
357 			 * case is one of them.  If the page is still
358 			 * partially dirty, make it fully dirty.
359 			 *
360 			 * note that we do not clear out the valid
361 			 * bits.  This would prevent bogus_page
362 			 * replacement from working properly.
363 			 */
364 			vm_page_set_validclean(m, base, size);
365 			if (m->dirty != 0)
366 				m->dirty = VM_PAGE_BITS_ALL;
367 			vm_page_unlock_queues();
368 		}
369 	}
370 	object->un_pager.vnp.vnp_size = nsize;
371 	object->size = nobjsize;
372 	VM_OBJECT_UNLOCK(object);
373 }
374 
375 /*
376  * calculate the linear (byte) disk address of specified virtual
377  * file address
378  */
379 static vm_offset_t
380 vnode_pager_addr(vp, address, run)
381 	struct vnode *vp;
382 	vm_ooffset_t address;
383 	int *run;
384 {
385 	int rtaddress;
386 	int bsize;
387 	daddr_t block;
388 	int err;
389 	daddr_t vblock;
390 	int voffset;
391 
392 	GIANT_REQUIRED;
393 	if ((int) address < 0)
394 		return -1;
395 
396 	if (vp->v_mount == NULL)
397 		return -1;
398 
399 	bsize = vp->v_mount->mnt_stat.f_iosize;
400 	vblock = address / bsize;
401 	voffset = address % bsize;
402 
403 	err = VOP_BMAP(vp, vblock, NULL, &block, run, NULL);
404 
405 	if (err || (block == -1))
406 		rtaddress = -1;
407 	else {
408 		rtaddress = block + voffset / DEV_BSIZE;
409 		if (run) {
410 			*run += 1;
411 			*run *= bsize/PAGE_SIZE;
412 			*run -= voffset/PAGE_SIZE;
413 		}
414 	}
415 
416 	return rtaddress;
417 }
418 
419 /*
420  * small block filesystem vnode pager input
421  */
422 static int
423 vnode_pager_input_smlfs(object, m)
424 	vm_object_t object;
425 	vm_page_t m;
426 {
427 	int i;
428 	struct vnode *dp, *vp;
429 	struct buf *bp;
430 	vm_offset_t kva;
431 	int fileaddr;
432 	vm_offset_t bsize;
433 	int error = 0;
434 
435 	GIANT_REQUIRED;
436 
437 	vp = object->handle;
438 	if (vp->v_mount == NULL)
439 		return VM_PAGER_BAD;
440 
441 	bsize = vp->v_mount->mnt_stat.f_iosize;
442 
443 	VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
444 
445 	kva = vm_pager_map_page(m);
446 
447 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
448 		vm_ooffset_t address;
449 
450 		if (vm_page_bits(i * bsize, bsize) & m->valid)
451 			continue;
452 
453 		address = IDX_TO_OFF(m->pindex) + i * bsize;
454 		if (address >= object->un_pager.vnp.vnp_size) {
455 			fileaddr = -1;
456 		} else {
457 			fileaddr = vnode_pager_addr(vp, address, NULL);
458 		}
459 		if (fileaddr != -1) {
460 			bp = getpbuf(&vnode_pbuf_freecnt);
461 
462 			/* build a minimal buffer header */
463 			bp->b_iocmd = BIO_READ;
464 			bp->b_iodone = bdone;
465 			KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
466 			KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
467 			bp->b_rcred = crhold(curthread->td_ucred);
468 			bp->b_wcred = crhold(curthread->td_ucred);
469 			bp->b_data = (caddr_t) kva + i * bsize;
470 			bp->b_blkno = fileaddr;
471 			pbgetvp(dp, bp);
472 			bp->b_bcount = bsize;
473 			bp->b_bufsize = bsize;
474 			bp->b_runningbufspace = bp->b_bufsize;
475 			runningbufspace += bp->b_runningbufspace;
476 
477 			/* do the input */
478 			bp->b_iooffset = dbtob(bp->b_blkno);
479 			if (dp->v_type == VCHR)
480 				VOP_SPECSTRATEGY(bp->b_vp, bp);
481 			else
482 				VOP_STRATEGY(bp->b_vp, bp);
483 
484 			/* we definitely need to be at splvm here */
485 
486 			bwait(bp, PVM, "vnsrd");
487 
488 			if ((bp->b_ioflags & BIO_ERROR) != 0)
489 				error = EIO;
490 
491 			/*
492 			 * free the buffer header back to the swap buffer pool
493 			 */
494 			relpbuf(bp, &vnode_pbuf_freecnt);
495 			if (error)
496 				break;
497 
498 			VM_OBJECT_LOCK(object);
499 			vm_page_lock_queues();
500 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
501 			vm_page_unlock_queues();
502 			VM_OBJECT_UNLOCK(object);
503 		} else {
504 			VM_OBJECT_LOCK(object);
505 			vm_page_lock_queues();
506 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
507 			vm_page_unlock_queues();
508 			VM_OBJECT_UNLOCK(object);
509 			bzero((caddr_t) kva + i * bsize, bsize);
510 		}
511 	}
512 	vm_pager_unmap_page(kva);
513 	vm_page_lock_queues();
514 	pmap_clear_modify(m);
515 	vm_page_flag_clear(m, PG_ZERO);
516 	vm_page_unlock_queues();
517 	if (error) {
518 		return VM_PAGER_ERROR;
519 	}
520 	return VM_PAGER_OK;
521 
522 }
523 
524 
525 /*
526  * old style vnode pager output routine
527  */
528 static int
529 vnode_pager_input_old(object, m)
530 	vm_object_t object;
531 	vm_page_t m;
532 {
533 	struct uio auio;
534 	struct iovec aiov;
535 	int error;
536 	int size;
537 	vm_offset_t kva;
538 	struct vnode *vp;
539 
540 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
541 	error = 0;
542 
543 	/*
544 	 * Return failure if beyond current EOF
545 	 */
546 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
547 		return VM_PAGER_BAD;
548 	} else {
549 		size = PAGE_SIZE;
550 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
551 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
552 		vp = object->handle;
553 		VM_OBJECT_UNLOCK(object);
554 
555 		/*
556 		 * Allocate a kernel virtual address and initialize so that
557 		 * we can use VOP_READ/WRITE routines.
558 		 */
559 		kva = vm_pager_map_page(m);
560 
561 		aiov.iov_base = (caddr_t) kva;
562 		aiov.iov_len = size;
563 		auio.uio_iov = &aiov;
564 		auio.uio_iovcnt = 1;
565 		auio.uio_offset = IDX_TO_OFF(m->pindex);
566 		auio.uio_segflg = UIO_SYSSPACE;
567 		auio.uio_rw = UIO_READ;
568 		auio.uio_resid = size;
569 		auio.uio_td = curthread;
570 
571 		error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
572 		if (!error) {
573 			int count = size - auio.uio_resid;
574 
575 			if (count == 0)
576 				error = EINVAL;
577 			else if (count != PAGE_SIZE)
578 				bzero((caddr_t) kva + count, PAGE_SIZE - count);
579 		}
580 		vm_pager_unmap_page(kva);
581 
582 		VM_OBJECT_LOCK(object);
583 	}
584 	vm_page_lock_queues();
585 	pmap_clear_modify(m);
586 	vm_page_undirty(m);
587 	vm_page_flag_clear(m, PG_ZERO);
588 	vm_page_unlock_queues();
589 	if (!error)
590 		m->valid = VM_PAGE_BITS_ALL;
591 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
592 }
593 
594 /*
595  * generic vnode pager input routine
596  */
597 
598 /*
599  * Local media VFS's that do not implement their own VOP_GETPAGES
600  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
601  * to implement the previous behaviour.
602  *
603  * All other FS's should use the bypass to get to the local media
604  * backing vp's VOP_GETPAGES.
605  */
606 static int
607 vnode_pager_getpages(object, m, count, reqpage)
608 	vm_object_t object;
609 	vm_page_t *m;
610 	int count;
611 	int reqpage;
612 {
613 	int rtval;
614 	struct vnode *vp;
615 	int bytes = count * PAGE_SIZE;
616 
617 	vp = object->handle;
618 	VM_OBJECT_UNLOCK(object);
619 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
620 	KASSERT(rtval != EOPNOTSUPP,
621 	    ("vnode_pager: FS getpages not implemented\n"));
622 	VM_OBJECT_LOCK(object);
623 	return rtval;
624 }
625 
626 /*
627  * This is now called from local media FS's to operate against their
628  * own vnodes if they fail to implement VOP_GETPAGES.
629  */
630 int
631 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
632 	struct vnode *vp;
633 	vm_page_t *m;
634 	int bytecount;
635 	int reqpage;
636 {
637 	vm_object_t object;
638 	vm_offset_t kva;
639 	off_t foff, tfoff, nextoff;
640 	int i, j, size, bsize, first, firstaddr;
641 	struct vnode *dp;
642 	int runpg;
643 	int runend;
644 	struct buf *bp;
645 	int count;
646 	int error = 0;
647 
648 	GIANT_REQUIRED;
649 	object = vp->v_object;
650 	count = bytecount / PAGE_SIZE;
651 
652 	if (vp->v_mount == NULL)
653 		return VM_PAGER_BAD;
654 
655 	bsize = vp->v_mount->mnt_stat.f_iosize;
656 
657 	/* get the UNDERLYING device for the file with VOP_BMAP() */
658 
659 	/*
660 	 * originally, we did not check for an error return value -- assuming
661 	 * an fs always has a bmap entry point -- that assumption is wrong!!!
662 	 */
663 	foff = IDX_TO_OFF(m[reqpage]->pindex);
664 
665 	/*
666 	 * if we can't bmap, use old VOP code
667 	 */
668 	if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
669 		VM_OBJECT_LOCK(object);
670 		vm_page_lock_queues();
671 		for (i = 0; i < count; i++)
672 			if (i != reqpage)
673 				vm_page_free(m[i]);
674 		vm_page_unlock_queues();
675 		cnt.v_vnodein++;
676 		cnt.v_vnodepgsin++;
677 		error = vnode_pager_input_old(object, m[reqpage]);
678 		VM_OBJECT_UNLOCK(object);
679 		return (error);
680 
681 		/*
682 		 * if the blocksize is smaller than a page size, then use
683 		 * special small filesystem code.  NFS sometimes has a small
684 		 * blocksize, but it can handle large reads itself.
685 		 */
686 	} else if ((PAGE_SIZE / bsize) > 1 &&
687 	    (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
688 		VM_OBJECT_LOCK(object);
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 		VM_OBJECT_UNLOCK(object);
695 		cnt.v_vnodein++;
696 		cnt.v_vnodepgsin++;
697 		return vnode_pager_input_smlfs(object, m[reqpage]);
698 	}
699 
700 	/*
701 	 * If we have a completely valid page available to us, we can
702 	 * clean up and return.  Otherwise we have to re-read the
703 	 * media.
704 	 */
705 	VM_OBJECT_LOCK(object);
706 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
707 		vm_page_lock_queues();
708 		for (i = 0; i < count; i++)
709 			if (i != reqpage)
710 				vm_page_free(m[i]);
711 		vm_page_unlock_queues();
712 		VM_OBJECT_UNLOCK(object);
713 		return VM_PAGER_OK;
714 	}
715 	m[reqpage]->valid = 0;
716 	VM_OBJECT_UNLOCK(object);
717 
718 	/*
719 	 * here on direct device I/O
720 	 */
721 	firstaddr = -1;
722 
723 	/*
724 	 * calculate the run that includes the required page
725 	 */
726 	for (first = 0, i = 0; i < count; i = runend) {
727 		firstaddr = vnode_pager_addr(vp,
728 			IDX_TO_OFF(m[i]->pindex), &runpg);
729 		if (firstaddr == -1) {
730 			VM_OBJECT_LOCK(object);
731 			if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
732 				panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
733 				    firstaddr, (uintmax_t)(foff >> 32),
734 				    (uintmax_t)foff,
735 				    (uintmax_t)
736 				    (object->un_pager.vnp.vnp_size >> 32),
737 				    (uintmax_t)object->un_pager.vnp.vnp_size);
738 			}
739 			vm_page_lock_queues();
740 			vm_page_free(m[i]);
741 			vm_page_unlock_queues();
742 			VM_OBJECT_UNLOCK(object);
743 			runend = i + 1;
744 			first = runend;
745 			continue;
746 		}
747 		runend = i + runpg;
748 		if (runend <= reqpage) {
749 			VM_OBJECT_LOCK(object);
750 			vm_page_lock_queues();
751 			for (j = i; j < runend; j++)
752 				vm_page_free(m[j]);
753 			vm_page_unlock_queues();
754 			VM_OBJECT_UNLOCK(object);
755 		} else {
756 			if (runpg < (count - first)) {
757 				VM_OBJECT_LOCK(object);
758 				vm_page_lock_queues();
759 				for (i = first + runpg; i < count; i++)
760 					vm_page_free(m[i]);
761 				vm_page_unlock_queues();
762 				VM_OBJECT_UNLOCK(object);
763 				count = first + runpg;
764 			}
765 			break;
766 		}
767 		first = runend;
768 	}
769 
770 	/*
771 	 * the first and last page have been calculated now, move input pages
772 	 * to be zero based...
773 	 */
774 	if (first != 0) {
775 		for (i = first; i < count; i++) {
776 			m[i - first] = m[i];
777 		}
778 		count -= first;
779 		reqpage -= first;
780 	}
781 
782 	/*
783 	 * calculate the file virtual address for the transfer
784 	 */
785 	foff = IDX_TO_OFF(m[0]->pindex);
786 
787 	/*
788 	 * calculate the size of the transfer
789 	 */
790 	size = count * PAGE_SIZE;
791 	if ((foff + size) > object->un_pager.vnp.vnp_size)
792 		size = object->un_pager.vnp.vnp_size - foff;
793 
794 	/*
795 	 * round up physical size for real devices.
796 	 */
797 	if (dp->v_type == VBLK || dp->v_type == VCHR) {
798 		int secmask = dp->v_rdev->si_bsize_phys - 1;
799 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
800 		size = (size + secmask) & ~secmask;
801 	}
802 
803 	bp = getpbuf(&vnode_pbuf_freecnt);
804 	kva = (vm_offset_t) bp->b_data;
805 
806 	/*
807 	 * and map the pages to be read into the kva
808 	 */
809 	pmap_qenter(kva, m, count);
810 
811 	/* build a minimal buffer header */
812 	bp->b_iocmd = BIO_READ;
813 	bp->b_iodone = bdone;
814 	KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
815 	KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
816 	bp->b_rcred = crhold(curthread->td_ucred);
817 	bp->b_wcred = crhold(curthread->td_ucred);
818 	bp->b_blkno = firstaddr;
819 	pbgetvp(dp, bp);
820 	bp->b_bcount = size;
821 	bp->b_bufsize = size;
822 	bp->b_runningbufspace = bp->b_bufsize;
823 	runningbufspace += bp->b_runningbufspace;
824 
825 	cnt.v_vnodein++;
826 	cnt.v_vnodepgsin += count;
827 
828 	/* do the input */
829 	bp->b_iooffset = dbtob(bp->b_blkno);
830 	if (dp->v_type == VCHR)
831 		VOP_SPECSTRATEGY(bp->b_vp, bp);
832 	else
833 		VOP_STRATEGY(bp->b_vp, bp);
834 
835 	bwait(bp, PVM, "vnread");
836 
837 	if ((bp->b_ioflags & BIO_ERROR) != 0)
838 		error = EIO;
839 
840 	if (!error) {
841 		if (size != count * PAGE_SIZE)
842 			bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
843 	}
844 	pmap_qremove(kva, count);
845 
846 	/*
847 	 * free the buffer header back to the swap buffer pool
848 	 */
849 	relpbuf(bp, &vnode_pbuf_freecnt);
850 
851 	VM_OBJECT_LOCK(object);
852 	vm_page_lock_queues();
853 	for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
854 		vm_page_t mt;
855 
856 		nextoff = tfoff + PAGE_SIZE;
857 		mt = m[i];
858 
859 		if (nextoff <= object->un_pager.vnp.vnp_size) {
860 			/*
861 			 * Read filled up entire page.
862 			 */
863 			mt->valid = VM_PAGE_BITS_ALL;
864 			vm_page_undirty(mt);	/* should be an assert? XXX */
865 			pmap_clear_modify(mt);
866 		} else {
867 			/*
868 			 * Read did not fill up entire page.  Since this
869 			 * is getpages, the page may be mapped, so we have
870 			 * to zero the invalid portions of the page even
871 			 * though we aren't setting them valid.
872 			 *
873 			 * Currently we do not set the entire page valid,
874 			 * we just try to clear the piece that we couldn't
875 			 * read.
876 			 */
877 			vm_page_set_validclean(mt, 0,
878 			    object->un_pager.vnp.vnp_size - tfoff);
879 			/* handled by vm_fault now */
880 			/* vm_page_zero_invalid(mt, FALSE); */
881 		}
882 
883 		vm_page_flag_clear(mt, PG_ZERO);
884 		if (i != reqpage) {
885 
886 			/*
887 			 * whether or not to leave the page activated is up in
888 			 * the air, but we should put the page on a page queue
889 			 * somewhere. (it already is in the object). Result:
890 			 * It appears that empirical results show that
891 			 * deactivating pages is best.
892 			 */
893 
894 			/*
895 			 * just in case someone was asking for this page we
896 			 * now tell them that it is ok to use
897 			 */
898 			if (!error) {
899 				if (mt->flags & PG_WANTED)
900 					vm_page_activate(mt);
901 				else
902 					vm_page_deactivate(mt);
903 				vm_page_wakeup(mt);
904 			} else {
905 				vm_page_free(mt);
906 			}
907 		}
908 	}
909 	vm_page_unlock_queues();
910 	VM_OBJECT_UNLOCK(object);
911 	if (error) {
912 		printf("vnode_pager_getpages: I/O read error\n");
913 	}
914 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
915 }
916 
917 /*
918  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
919  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
920  * vnode_pager_generic_putpages() to implement the previous behaviour.
921  *
922  * All other FS's should use the bypass to get to the local media
923  * backing vp's VOP_PUTPAGES.
924  */
925 static void
926 vnode_pager_putpages(object, m, count, sync, rtvals)
927 	vm_object_t object;
928 	vm_page_t *m;
929 	int count;
930 	boolean_t sync;
931 	int *rtvals;
932 {
933 	int rtval;
934 	struct vnode *vp;
935 	struct mount *mp;
936 	int bytes = count * PAGE_SIZE;
937 
938 	GIANT_REQUIRED;
939 	/*
940 	 * Force synchronous operation if we are extremely low on memory
941 	 * to prevent a low-memory deadlock.  VOP operations often need to
942 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
943 	 * operation ).  The swapper handles the case by limiting the amount
944 	 * of asynchronous I/O, but that sort of solution doesn't scale well
945 	 * for the vnode pager without a lot of work.
946 	 *
947 	 * Also, the backing vnode's iodone routine may not wake the pageout
948 	 * daemon up.  This should be probably be addressed XXX.
949 	 */
950 
951 	if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
952 		sync |= OBJPC_SYNC;
953 
954 	/*
955 	 * Call device-specific putpages function
956 	 */
957 	vp = object->handle;
958 	VM_OBJECT_UNLOCK(object);
959 	if (vp->v_type != VREG)
960 		mp = NULL;
961 	(void)vn_start_write(vp, &mp, V_WAIT);
962 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
963 	KASSERT(rtval != EOPNOTSUPP,
964 	    ("vnode_pager: stale FS putpages\n"));
965 	vn_finished_write(mp);
966 	VM_OBJECT_LOCK(object);
967 }
968 
969 
970 /*
971  * This is now called from local media FS's to operate against their
972  * own vnodes if they fail to implement VOP_PUTPAGES.
973  *
974  * This is typically called indirectly via the pageout daemon and
975  * clustering has already typically occured, so in general we ask the
976  * underlying filesystem to write the data out asynchronously rather
977  * then delayed.
978  */
979 int
980 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
981 	struct vnode *vp;
982 	vm_page_t *m;
983 	int bytecount;
984 	int flags;
985 	int *rtvals;
986 {
987 	int i;
988 	vm_object_t object;
989 	int count;
990 
991 	int maxsize, ncount;
992 	vm_ooffset_t poffset;
993 	struct uio auio;
994 	struct iovec aiov;
995 	int error;
996 	int ioflags;
997 
998 	GIANT_REQUIRED;
999 	object = vp->v_object;
1000 	count = bytecount / PAGE_SIZE;
1001 
1002 	for (i = 0; i < count; i++)
1003 		rtvals[i] = VM_PAGER_AGAIN;
1004 
1005 	if ((int) m[0]->pindex < 0) {
1006 		printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n",
1007 			(long)m[0]->pindex, (u_long)m[0]->dirty);
1008 		rtvals[0] = VM_PAGER_BAD;
1009 		return VM_PAGER_BAD;
1010 	}
1011 
1012 	maxsize = count * PAGE_SIZE;
1013 	ncount = count;
1014 
1015 	poffset = IDX_TO_OFF(m[0]->pindex);
1016 
1017 	/*
1018 	 * If the page-aligned write is larger then the actual file we
1019 	 * have to invalidate pages occuring beyond the file EOF.  However,
1020 	 * there is an edge case where a file may not be page-aligned where
1021 	 * the last page is partially invalid.  In this case the filesystem
1022 	 * may not properly clear the dirty bits for the entire page (which
1023 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1024 	 * With the page locked we are free to fix-up the dirty bits here.
1025 	 *
1026 	 * We do not under any circumstances truncate the valid bits, as
1027 	 * this will screw up bogus page replacement.
1028 	 */
1029 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1030 		if (object->un_pager.vnp.vnp_size > poffset) {
1031 			int pgoff;
1032 
1033 			maxsize = object->un_pager.vnp.vnp_size - poffset;
1034 			ncount = btoc(maxsize);
1035 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1036 				vm_page_lock_queues();
1037 				vm_page_clear_dirty(m[ncount - 1], pgoff,
1038 					PAGE_SIZE - pgoff);
1039 				vm_page_unlock_queues();
1040 			}
1041 		} else {
1042 			maxsize = 0;
1043 			ncount = 0;
1044 		}
1045 		if (ncount < count) {
1046 			for (i = ncount; i < count; i++) {
1047 				rtvals[i] = VM_PAGER_BAD;
1048 			}
1049 		}
1050 	}
1051 
1052 	/*
1053 	 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1054 	 * rather then a bdwrite() to prevent paging I/O from saturating
1055 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
1056 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
1057 	 * the system decides how to cluster.
1058 	 */
1059 	ioflags = IO_VMIO;
1060 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
1061 		ioflags |= IO_SYNC;
1062 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
1063 		ioflags |= IO_ASYNC;
1064 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1065 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
1066 
1067 	aiov.iov_base = (caddr_t) 0;
1068 	aiov.iov_len = maxsize;
1069 	auio.uio_iov = &aiov;
1070 	auio.uio_iovcnt = 1;
1071 	auio.uio_offset = poffset;
1072 	auio.uio_segflg = UIO_NOCOPY;
1073 	auio.uio_rw = UIO_WRITE;
1074 	auio.uio_resid = maxsize;
1075 	auio.uio_td = (struct thread *) 0;
1076 	error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1077 	cnt.v_vnodeout++;
1078 	cnt.v_vnodepgsout += ncount;
1079 
1080 	if (error) {
1081 		printf("vnode_pager_putpages: I/O error %d\n", error);
1082 	}
1083 	if (auio.uio_resid) {
1084 		printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1085 		    auio.uio_resid, (u_long)m[0]->pindex);
1086 	}
1087 	for (i = 0; i < ncount; i++) {
1088 		rtvals[i] = VM_PAGER_OK;
1089 	}
1090 	return rtvals[0];
1091 }
1092 
1093 struct vnode *
1094 vnode_pager_lock(vm_object_t first_object)
1095 {
1096 	struct vnode *vp;
1097 	vm_object_t backing_object, object;
1098 
1099 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
1100 	for (object = first_object; object != NULL; object = backing_object) {
1101 		if (object->type != OBJT_VNODE) {
1102 			if ((backing_object = object->backing_object) != NULL)
1103 				VM_OBJECT_LOCK(backing_object);
1104 			if (object != first_object)
1105 				VM_OBJECT_UNLOCK(object);
1106 			continue;
1107 		}
1108 	retry:
1109 		if (object->flags & OBJ_DEAD) {
1110 			if (object != first_object)
1111 				VM_OBJECT_UNLOCK(object);
1112 			return NULL;
1113 		}
1114 		vp = object->handle;
1115 		VI_LOCK(vp);
1116 		VM_OBJECT_UNLOCK(object);
1117 		if (first_object != object)
1118 			VM_OBJECT_UNLOCK(first_object);
1119 		if (vget(vp, LK_CANRECURSE | LK_INTERLOCK | LK_NOPAUSE |
1120 		    LK_RETRY | LK_SHARED, curthread)) {
1121 			VM_OBJECT_LOCK(first_object);
1122 			if (object != first_object)
1123 				VM_OBJECT_LOCK(object);
1124 			if (object->type != OBJT_VNODE) {
1125 				if (object != first_object)
1126 					VM_OBJECT_UNLOCK(object);
1127 				return NULL;
1128 			}
1129 			printf("vnode_pager_lock: retrying\n");
1130 			goto retry;
1131 		}
1132 		VM_OBJECT_LOCK(first_object);
1133 		return (vp);
1134 	}
1135 	return NULL;
1136 }
1137