xref: /dragonfly/sys/vm/vnode_pager.c (revision 7bc7e232)
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: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
42  * $DragonFly: src/sys/vm/vnode_pager.c,v 1.40 2007/08/28 01:09:07 dillon Exp $
43  */
44 
45 /*
46  * Page to/from files (vnodes).
47  */
48 
49 /*
50  * TODO:
51  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
52  *	greatly re-simplify the vnode_pager.
53  */
54 
55 #include <sys/param.h>
56 #include <sys/systm.h>
57 #include <sys/kernel.h>
58 #include <sys/proc.h>
59 #include <sys/vnode.h>
60 #include <sys/mount.h>
61 #include <sys/buf.h>
62 #include <sys/vmmeter.h>
63 #include <sys/conf.h>
64 #include <sys/sfbuf.h>
65 #include <sys/thread2.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_pager.h>
71 #include <vm/vm_map.h>
72 #include <vm/vnode_pager.h>
73 #include <vm/vm_extern.h>
74 
75 static void vnode_pager_dealloc (vm_object_t);
76 static int vnode_pager_getpages (vm_object_t, vm_page_t *, int, int);
77 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
78 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t, int *, int *);
79 
80 struct pagerops vnodepagerops = {
81 	NULL,
82 	vnode_pager_alloc,
83 	vnode_pager_dealloc,
84 	vnode_pager_getpages,
85 	vnode_pager_putpages,
86 	vnode_pager_haspage,
87 	NULL
88 };
89 
90 static struct krate vbadrate = { 1 };
91 static struct krate vresrate = { 1 };
92 
93 int vnode_pbuf_freecnt = -1;	/* start out unlimited */
94 
95 /*
96  * Allocate (or lookup) pager for a vnode.
97  * Handle is a vnode pointer.
98  */
99 vm_object_t
100 vnode_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
101 {
102 	vm_object_t object;
103 	struct vnode *vp;
104 
105 	/*
106 	 * Pageout to vnode, no can do yet.
107 	 */
108 	if (handle == NULL)
109 		return (NULL);
110 
111 	/*
112 	 * XXX hack - This initialization should be put somewhere else.
113 	 */
114 	if (vnode_pbuf_freecnt < 0) {
115 	    vnode_pbuf_freecnt = nswbuf / 2 + 1;
116 	}
117 
118 	vp = (struct vnode *) handle;
119 
120 	/*
121 	 * Prevent race condition when allocating the object. This
122 	 * can happen with NFS vnodes since the nfsnode isn't locked.
123 	 */
124 	while (vp->v_flag & VOLOCK) {
125 		vp->v_flag |= VOWANT;
126 		tsleep(vp, 0, "vnpobj", 0);
127 	}
128 	vp->v_flag |= VOLOCK;
129 
130 	/*
131 	 * If the object is being terminated, wait for it to
132 	 * go away.
133 	 */
134 	while (((object = vp->v_object) != NULL) &&
135 		(object->flags & OBJ_DEAD)) {
136 		vm_object_dead_sleep(object, "vadead");
137 	}
138 
139 	if (vp->v_sysref.refcnt <= 0)
140 		panic("vnode_pager_alloc: no vnode reference");
141 
142 	if (object == NULL) {
143 		/*
144 		 * And an object of the appropriate size
145 		 */
146 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
147 		object->flags = 0;
148 		object->handle = handle;
149 		vp->v_object = object;
150 		vp->v_filesize = size;
151 	} else {
152 		object->ref_count++;
153 		if (vp->v_filesize != size)
154 			kprintf("vnode_pager_alloc: Warning, filesize mismatch %lld/%lld\n", vp->v_filesize, size);
155 	}
156 	vref(vp);
157 
158 	vp->v_flag &= ~VOLOCK;
159 	if (vp->v_flag & VOWANT) {
160 		vp->v_flag &= ~VOWANT;
161 		wakeup(vp);
162 	}
163 	return (object);
164 }
165 
166 static void
167 vnode_pager_dealloc(vm_object_t object)
168 {
169 	struct vnode *vp = object->handle;
170 
171 	if (vp == NULL)
172 		panic("vnode_pager_dealloc: pager already dealloced");
173 
174 	vm_object_pip_wait(object, "vnpdea");
175 
176 	object->handle = NULL;
177 	object->type = OBJT_DEAD;
178 	vp->v_object = NULL;
179 	vp->v_filesize = NOOFFSET;
180 	vp->v_flag &= ~(VTEXT | VOBJBUF);
181 }
182 
183 /*
184  * Return whether the vnode pager has the requested page.  Return the
185  * number of disk-contiguous pages before and after the requested page,
186  * not including the requested page.
187  */
188 static boolean_t
189 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
190 		    int *after)
191 {
192 	struct vnode *vp = object->handle;
193 	off_t loffset;
194 	off_t doffset;
195 	int voff;
196 	int bsize;
197 	int error;
198 
199 	/*
200 	 * If no vp or vp is doomed or marked transparent to VM, we do not
201 	 * have the page.
202 	 */
203 	if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
204 		return FALSE;
205 
206 	/*
207 	 * If filesystem no longer mounted or offset beyond end of file we do
208 	 * not have the page.
209 	 */
210 	loffset = IDX_TO_OFF(pindex);
211 
212 	if (vp->v_mount == NULL || loffset >= vp->v_filesize)
213 		return FALSE;
214 
215 	bsize = vp->v_mount->mnt_stat.f_iosize;
216 	voff = loffset % bsize;
217 
218 	error = VOP_BMAP(vp, loffset - voff, &doffset, after, before);
219 	if (error)
220 		return TRUE;
221 	if (doffset == NOOFFSET)
222 		return FALSE;
223 
224 	if (before) {
225 		*before = (*before + voff) >> PAGE_SHIFT;
226 	}
227 	if (after) {
228 		*after -= voff;
229 		if (loffset + *after > vp->v_filesize)
230 			*after = vp->v_filesize - loffset;
231 		*after >>= PAGE_SHIFT;
232 		if (*after < 0)
233 			*after = 0;
234 	}
235 	return TRUE;
236 }
237 
238 /*
239  * Lets the VM system know about a change in size for a file.
240  * We adjust our own internal size and flush any cached pages in
241  * the associated object that are affected by the size change.
242  *
243  * NOTE: This routine may be invoked as a result of a pager put
244  * operation (possibly at object termination time), so we must be careful.
245  *
246  * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
247  * we do not blow up on the case.  nsize will always be >= 0, however.
248  */
249 void
250 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
251 {
252 	vm_pindex_t nobjsize;
253 	vm_pindex_t oobjsize;
254 	vm_object_t object = vp->v_object;
255 
256 	if (object == NULL)
257 		return;
258 
259 	/*
260 	 * Hasn't changed size
261 	 */
262 	if (nsize == vp->v_filesize)
263 		return;
264 
265 	/*
266 	 * Has changed size.  Adjust the VM object's size and v_filesize
267 	 * before we start scanning pages to prevent new pages from being
268 	 * allocated during the scan.
269 	 */
270 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
271 	oobjsize = object->size;
272 	object->size = nobjsize;
273 
274 	/*
275 	 * File has shrunk. Toss any cached pages beyond the new EOF.
276 	 */
277 	if (nsize < vp->v_filesize) {
278 		vp->v_filesize = nsize;
279 		if (nobjsize < oobjsize) {
280 			vm_object_page_remove(object, nobjsize, oobjsize,
281 					      FALSE);
282 		}
283 		/*
284 		 * This gets rid of garbage at the end of a page that is now
285 		 * only partially backed by the vnode.  Since we are setting
286 		 * the entire page valid & clean after we are done we have
287 		 * to be sure that the portion of the page within the file
288 		 * bounds is already valid.  If it isn't then making it
289 		 * valid would create a corrupt block.
290 		 */
291 		if (nsize & PAGE_MASK) {
292 			vm_offset_t kva;
293 			vm_page_t m;
294 
295 			m = vm_page_lookup(object, OFF_TO_IDX(nsize));
296 			if (m && m->valid) {
297 				int base = (int)nsize & PAGE_MASK;
298 				int size = PAGE_SIZE - base;
299 				struct sf_buf *sf;
300 
301 				/*
302 				 * Clear out partial-page garbage in case
303 				 * the page has been mapped.
304 				 */
305 				sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
306 				kva = sf_buf_kva(sf);
307 				bzero((caddr_t)kva + base, size);
308 				sf_buf_free(sf);
309 
310 				/*
311 				 * XXX work around SMP data integrity race
312 				 * by unmapping the page from user processes.
313 				 * The garbage we just cleared may be mapped
314 				 * to a user process running on another cpu
315 				 * and this code is not running through normal
316 				 * I/O channels which handle SMP issues for
317 				 * us, so unmap page to synchronize all cpus.
318 				 *
319 				 * XXX should vm_pager_unmap_page() have
320 				 * dealt with this?
321 				 */
322 				vm_page_protect(m, VM_PROT_NONE);
323 
324 				/*
325 				 * Clear out partial-page dirty bits.  This
326 				 * has the side effect of setting the valid
327 				 * bits, but that is ok.  There are a bunch
328 				 * of places in the VM system where we expected
329 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
330 				 * case is one of them.  If the page is still
331 				 * partially dirty, make it fully dirty.
332 				 *
333 				 * note that we do not clear out the valid
334 				 * bits.  This would prevent bogus_page
335 				 * replacement from working properly.
336 				 */
337 				vm_page_set_validclean(m, base, size);
338 				if (m->dirty != 0)
339 					m->dirty = VM_PAGE_BITS_ALL;
340 			}
341 		}
342 	} else {
343 		vp->v_filesize = nsize;
344 	}
345 }
346 
347 /*
348  * Release a page busied for a getpages operation.  The page may have become
349  * wired (typically due to being used by the buffer cache) or otherwise been
350  * soft-busied and cannot be freed in that case.  A held page can still be
351  * freed.
352  */
353 void
354 vnode_pager_freepage(vm_page_t m)
355 {
356 	if (m->busy || m->wire_count) {
357 		vm_page_activate(m);
358 		vm_page_wakeup(m);
359 	} else {
360 		vm_page_free(m);
361 	}
362 }
363 
364 /*
365  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
366  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
367  * vnode_pager_generic_getpages() to implement the previous behaviour.
368  *
369  * All other FS's should use the bypass to get to the local media
370  * backing vp's VOP_GETPAGES.
371  */
372 static int
373 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
374 {
375 	int rtval;
376 	struct vnode *vp;
377 	int bytes = count * PAGE_SIZE;
378 
379 	vp = object->handle;
380 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
381 	if (rtval == EOPNOTSUPP)
382 		panic("vnode_pager: vfs's must implement vop_getpages\n");
383 	return rtval;
384 }
385 
386 /*
387  * This is now called from local media FS's to operate against their
388  * own vnodes if they fail to implement VOP_GETPAGES.
389  *
390  * With all the caching local media devices do these days there is really
391  * very little point to attempting to restrict the I/O size to contiguous
392  * blocks on-disk, especially if our caller thinks we need all the specified
393  * pages.  Just construct and issue a READ.
394  */
395 int
396 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount,
397 			     int reqpage)
398 {
399 	struct iovec aiov;
400 	struct uio auio;
401 	off_t foff;
402 	int error;
403 	int count;
404 	int i;
405 	int ioflags;
406 
407 	/*
408 	 * Do not do anything if the vnode is bad.
409 	 */
410 	if (vp->v_mount == NULL)
411 		return VM_PAGER_BAD;
412 
413 	/*
414 	 * Calculate the number of pages.  Since we are paging in whole
415 	 * pages, adjust bytecount to be an integral multiple of the page
416 	 * size.  It will be clipped to the file EOF later on.
417 	 */
418 	bytecount = round_page(bytecount);
419 	count = bytecount / PAGE_SIZE;
420 
421 	/*
422 	 * If we have a completely valid page available to us, we can
423 	 * clean up and return.  Otherwise we have to re-read the
424 	 * media.
425 	 *
426 	 * Note that this does not work with NFS, so NFS has its own
427 	 * getpages routine.  The problem is that NFS can have partially
428 	 * valid pages associated with the buffer cache due to the piecemeal
429 	 * write support.  If we were to fall through and re-read the media
430 	 * as we do here, dirty data could be lost.
431 	 */
432 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
433 		for (i = 0; i < count; i++) {
434 			if (i != reqpage)
435 				vnode_pager_freepage(m[i]);
436 		}
437 		return VM_PAGER_OK;
438 	}
439 
440 	/*
441 	 * Discard pages past the file EOF.  If the requested page is past
442 	 * the file EOF we just leave its valid bits set to 0, the caller
443 	 * expects to maintain ownership of the requested page.  If the
444 	 * entire range is past file EOF discard everything and generate
445 	 * a pagein error.
446 	 */
447 	foff = IDX_TO_OFF(m[0]->pindex);
448 	if (foff >= vp->v_filesize) {
449 		for (i = 0; i < count; i++) {
450 			if (i != reqpage)
451 				vnode_pager_freepage(m[i]);
452 		}
453 		return VM_PAGER_ERROR;
454 	}
455 
456 	if (foff + bytecount > vp->v_filesize) {
457 		bytecount = vp->v_filesize - foff;
458 		i = round_page(bytecount) / PAGE_SIZE;
459 		while (count > i) {
460 			--count;
461 			if (count != reqpage)
462 				vnode_pager_freepage(m[count]);
463 		}
464 	}
465 
466 	/*
467 	 * The size of the transfer is bytecount.  bytecount will be an
468 	 * integral multiple of the page size unless it has been clipped
469 	 * to the file EOF.  The transfer cannot exceed the file EOF.
470 	 *
471 	 * When dealing with real devices we must round-up to the device
472 	 * sector size.
473 	 */
474 	if (vp->v_type == VBLK || vp->v_type == VCHR) {
475 		int secmask = vp->v_rdev->si_bsize_phys - 1;
476 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
477 		bytecount = (bytecount + secmask) & ~secmask;
478 	}
479 
480 	/*
481 	 * Severe hack to avoid deadlocks with the buffer cache
482 	 */
483 	for (i = 0; i < count; ++i) {
484 		vm_page_t mt = m[i];
485 
486 		vm_page_io_start(mt);
487 		vm_page_wakeup(mt);
488 	}
489 
490 	/*
491 	 * Issue the I/O without any read-ahead
492 	 */
493 	ioflags = IO_VMIO;
494 	/*ioflags |= IO_SEQMAX << IO_SEQSHIFT;*/
495 
496 	aiov.iov_base = (caddr_t) 0;
497 	aiov.iov_len = bytecount;
498 	auio.uio_iov = &aiov;
499 	auio.uio_iovcnt = 1;
500 	auio.uio_offset = foff;
501 	auio.uio_segflg = UIO_NOCOPY;
502 	auio.uio_rw = UIO_READ;
503 	auio.uio_resid = bytecount;
504 	auio.uio_td = NULL;
505 	mycpu->gd_cnt.v_vnodein++;
506 	mycpu->gd_cnt.v_vnodepgsin += count;
507 
508 	error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
509 
510 	/*
511 	 * Severe hack to avoid deadlocks with the buffer cache
512 	 */
513 	for (i = 0; i < count; ++i) {
514 		vm_page_t mt = m[i];
515 
516 		while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
517 			;
518 		vm_page_busy(mt);
519 		vm_page_io_finish(mt);
520 	}
521 
522 	/*
523 	 * Calculate the actual number of bytes read and clean up the
524 	 * page list.
525 	 */
526 	bytecount -= auio.uio_resid;
527 
528 	for (i = 0; i < count; ++i) {
529 		vm_page_t mt = m[i];
530 
531 		if (i != reqpage) {
532 			if (error == 0 && mt->valid) {
533 				if (mt->flags & PG_WANTED)
534 					vm_page_activate(mt);
535 				else
536 					vm_page_deactivate(mt);
537 				vm_page_wakeup(mt);
538 			} else {
539 				vnode_pager_freepage(mt);
540 			}
541 		} else if (mt->valid == 0) {
542 			if (error == 0) {
543 				kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
544 				/* whoops, something happened */
545 				error = EINVAL;
546 			}
547 		} else if (mt->valid != VM_PAGE_BITS_ALL) {
548 			/*
549 			 * Zero-extend the requested page if necessary (if
550 			 * the filesystem is using a small block size).
551 			 */
552 			vm_page_zero_invalid(mt, TRUE);
553 		}
554 	}
555 	if (error) {
556 		kprintf("vnode_pager_getpages: I/O read error\n");
557 	}
558 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
559 }
560 
561 /*
562  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
563  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
564  * vnode_pager_generic_putpages() to implement the previous behaviour.
565  *
566  * All other FS's should use the bypass to get to the local media
567  * backing vp's VOP_PUTPAGES.
568  */
569 static void
570 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
571     boolean_t sync, int *rtvals)
572 {
573 	int rtval;
574 	struct vnode *vp;
575 	int bytes = count * PAGE_SIZE;
576 
577 	/*
578 	 * Force synchronous operation if we are extremely low on memory
579 	 * to prevent a low-memory deadlock.  VOP operations often need to
580 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
581 	 * operation ).  The swapper handles the case by limiting the amount
582 	 * of asynchronous I/O, but that sort of solution doesn't scale well
583 	 * for the vnode pager without a lot of work.
584 	 *
585 	 * Also, the backing vnode's iodone routine may not wake the pageout
586 	 * daemon up.  This should be probably be addressed XXX.
587 	 */
588 
589 	if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
590 		sync |= OBJPC_SYNC;
591 
592 	/*
593 	 * Call device-specific putpages function
594 	 */
595 
596 	vp = object->handle;
597 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
598 	if (rtval == EOPNOTSUPP) {
599 	    kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
600 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
601 	}
602 }
603 
604 
605 /*
606  * This is now called from local media FS's to operate against their
607  * own vnodes if they fail to implement VOP_PUTPAGES.
608  *
609  * This is typically called indirectly via the pageout daemon and
610  * clustering has already typically occured, so in general we ask the
611  * underlying filesystem to write the data out asynchronously rather
612  * then delayed.
613  */
614 int
615 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
616 			     int flags, int *rtvals)
617 {
618 	int i;
619 	vm_object_t object;
620 	int count;
621 
622 	int maxsize, ncount;
623 	vm_ooffset_t poffset;
624 	struct uio auio;
625 	struct iovec aiov;
626 	int error;
627 	int ioflags;
628 
629 	object = vp->v_object;
630 	count = bytecount / PAGE_SIZE;
631 
632 	for (i = 0; i < count; i++)
633 		rtvals[i] = VM_PAGER_AGAIN;
634 
635 	if ((int) m[0]->pindex < 0) {
636 		kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
637 			(long)m[0]->pindex, m[0]->dirty);
638 		rtvals[0] = VM_PAGER_BAD;
639 		return VM_PAGER_BAD;
640 	}
641 
642 	maxsize = count * PAGE_SIZE;
643 	ncount = count;
644 
645 	poffset = IDX_TO_OFF(m[0]->pindex);
646 
647 	/*
648 	 * If the page-aligned write is larger then the actual file we
649 	 * have to invalidate pages occuring beyond the file EOF.  However,
650 	 * there is an edge case where a file may not be page-aligned where
651 	 * the last page is partially invalid.  In this case the filesystem
652 	 * may not properly clear the dirty bits for the entire page (which
653 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
654 	 * With the page locked we are free to fix-up the dirty bits here.
655 	 *
656 	 * We do not under any circumstances truncate the valid bits, as
657 	 * this will screw up bogus page replacement.
658 	 */
659 	if (maxsize + poffset > vp->v_filesize) {
660 		if (vp->v_filesize > poffset) {
661 			int pgoff;
662 
663 			maxsize = vp->v_filesize - poffset;
664 			ncount = btoc(maxsize);
665 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
666 				vm_page_clear_dirty(m[ncount - 1], pgoff,
667 					PAGE_SIZE - pgoff);
668 			}
669 		} else {
670 			maxsize = 0;
671 			ncount = 0;
672 		}
673 		if (ncount < count) {
674 			for (i = ncount; i < count; i++) {
675 				rtvals[i] = VM_PAGER_BAD;
676 			}
677 		}
678 	}
679 
680 	/*
681 	 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
682 	 * rather then a bdwrite() to prevent paging I/O from saturating
683 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
684 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
685 	 * the system decides how to cluster.
686 	 */
687 	ioflags = IO_VMIO;
688 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
689 		ioflags |= IO_SYNC;
690 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
691 		ioflags |= IO_ASYNC;
692 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
693 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
694 
695 	aiov.iov_base = (caddr_t) 0;
696 	aiov.iov_len = maxsize;
697 	auio.uio_iov = &aiov;
698 	auio.uio_iovcnt = 1;
699 	auio.uio_offset = poffset;
700 	auio.uio_segflg = UIO_NOCOPY;
701 	auio.uio_rw = UIO_WRITE;
702 	auio.uio_resid = maxsize;
703 	auio.uio_td = NULL;
704 	error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
705 	mycpu->gd_cnt.v_vnodeout++;
706 	mycpu->gd_cnt.v_vnodepgsout += ncount;
707 
708 	if (error) {
709 		krateprintf(&vbadrate,
710 			    "vnode_pager_putpages: I/O error %d\n", error);
711 	}
712 	if (auio.uio_resid) {
713 		krateprintf(&vresrate,
714 			    "vnode_pager_putpages: residual I/O %d at %lu\n",
715 			    auio.uio_resid, (u_long)m[0]->pindex);
716 	}
717 	for (i = 0; i < ncount; i++)
718 		rtvals[i] = VM_PAGER_OK;
719 	return rtvals[0];
720 }
721 
722 struct vnode *
723 vnode_pager_lock(vm_object_t object)
724 {
725 	struct thread *td = curthread;	/* XXX */
726 	int error;
727 
728 	for (; object != NULL; object = object->backing_object) {
729 		if (object->type != OBJT_VNODE)
730 			continue;
731 		if (object->flags & OBJ_DEAD)
732 			return NULL;
733 
734 		for (;;) {
735 			struct vnode *vp = object->handle;
736 			error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
737 			if (error == 0) {
738 				if (object->handle != vp) {
739 					vput(vp);
740 					continue;
741 				}
742 				return (vp);
743 			}
744 			if ((object->flags & OBJ_DEAD) ||
745 			    (object->type != OBJT_VNODE)) {
746 				return NULL;
747 			}
748 			kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
749 			tsleep(object->handle, 0, "vnpgrl", hz);
750 		}
751 	}
752 	return NULL;
753 }
754