xref: /dflybsd-src/sys/vm/vnode_pager.c (revision 9b42cabedff59ea3b351e8eccb2da9bab5cefe7e)
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 1.43 2008/06/19 23:27:39 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, 0);
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 			do {
296 				m = vm_page_lookup(object, OFF_TO_IDX(nsize));
297 			} while (m && vm_page_sleep_busy(m, TRUE, "vsetsz"));
298 
299 			if (m && m->valid) {
300 				int base = (int)nsize & PAGE_MASK;
301 				int size = PAGE_SIZE - base;
302 				struct sf_buf *sf;
303 
304 				/*
305 				 * Clear out partial-page garbage in case
306 				 * the page has been mapped.
307 				 */
308 				vm_page_busy(m);
309 				sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
310 				kva = sf_buf_kva(sf);
311 				bzero((caddr_t)kva + base, size);
312 				sf_buf_free(sf);
313 
314 				/*
315 				 * XXX work around SMP data integrity race
316 				 * by unmapping the page from user processes.
317 				 * The garbage we just cleared may be mapped
318 				 * to a user process running on another cpu
319 				 * and this code is not running through normal
320 				 * I/O channels which handle SMP issues for
321 				 * us, so unmap page to synchronize all cpus.
322 				 *
323 				 * XXX should vm_pager_unmap_page() have
324 				 * dealt with this?
325 				 */
326 				vm_page_protect(m, VM_PROT_NONE);
327 
328 				/*
329 				 * Clear out partial-page dirty bits.  This
330 				 * has the side effect of setting the valid
331 				 * bits, but that is ok.  There are a bunch
332 				 * of places in the VM system where we expected
333 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
334 				 * case is one of them.  If the page is still
335 				 * partially dirty, make it fully dirty.
336 				 *
337 				 * note that we do not clear out the valid
338 				 * bits.  This would prevent bogus_page
339 				 * replacement from working properly.
340 				 */
341 				vm_page_set_validclean(m, base, size);
342 				if (m->dirty != 0)
343 					m->dirty = VM_PAGE_BITS_ALL;
344 				vm_page_wakeup(m);
345 			}
346 		}
347 	} else {
348 		vp->v_filesize = nsize;
349 	}
350 }
351 
352 /*
353  * Release a page busied for a getpages operation.  The page may have become
354  * wired (typically due to being used by the buffer cache) or otherwise been
355  * soft-busied and cannot be freed in that case.  A held page can still be
356  * freed.
357  */
358 void
359 vnode_pager_freepage(vm_page_t m)
360 {
361 	if (m->busy || m->wire_count) {
362 		vm_page_activate(m);
363 		vm_page_wakeup(m);
364 	} else {
365 		vm_page_free(m);
366 	}
367 }
368 
369 /*
370  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
371  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
372  * vnode_pager_generic_getpages() to implement the previous behaviour.
373  *
374  * All other FS's should use the bypass to get to the local media
375  * backing vp's VOP_GETPAGES.
376  */
377 static int
378 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
379 {
380 	int rtval;
381 	struct vnode *vp;
382 	int bytes = count * PAGE_SIZE;
383 
384 	vp = object->handle;
385 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
386 	if (rtval == EOPNOTSUPP)
387 		panic("vnode_pager: vfs's must implement vop_getpages\n");
388 	return rtval;
389 }
390 
391 /*
392  * This is now called from local media FS's to operate against their
393  * own vnodes if they fail to implement VOP_GETPAGES.
394  *
395  * With all the caching local media devices do these days there is really
396  * very little point to attempting to restrict the I/O size to contiguous
397  * blocks on-disk, especially if our caller thinks we need all the specified
398  * pages.  Just construct and issue a READ.
399  */
400 int
401 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount,
402 			     int reqpage)
403 {
404 	struct iovec aiov;
405 	struct uio auio;
406 	off_t foff;
407 	int error;
408 	int count;
409 	int i;
410 	int ioflags;
411 
412 	/*
413 	 * Do not do anything if the vnode is bad.
414 	 */
415 	if (vp->v_mount == NULL)
416 		return VM_PAGER_BAD;
417 
418 	/*
419 	 * Calculate the number of pages.  Since we are paging in whole
420 	 * pages, adjust bytecount to be an integral multiple of the page
421 	 * size.  It will be clipped to the file EOF later on.
422 	 */
423 	bytecount = round_page(bytecount);
424 	count = bytecount / PAGE_SIZE;
425 
426 	/*
427 	 * If we have a completely valid page available to us, we can
428 	 * clean up and return.  Otherwise we have to re-read the
429 	 * media.
430 	 *
431 	 * Note that this does not work with NFS, so NFS has its own
432 	 * getpages routine.  The problem is that NFS can have partially
433 	 * valid pages associated with the buffer cache due to the piecemeal
434 	 * write support.  If we were to fall through and re-read the media
435 	 * as we do here, dirty data could be lost.
436 	 */
437 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
438 		for (i = 0; i < count; i++) {
439 			if (i != reqpage)
440 				vnode_pager_freepage(m[i]);
441 		}
442 		return VM_PAGER_OK;
443 	}
444 
445 	/*
446 	 * Discard pages past the file EOF.  If the requested page is past
447 	 * the file EOF we just leave its valid bits set to 0, the caller
448 	 * expects to maintain ownership of the requested page.  If the
449 	 * entire range is past file EOF discard everything and generate
450 	 * a pagein error.
451 	 */
452 	foff = IDX_TO_OFF(m[0]->pindex);
453 	if (foff >= vp->v_filesize) {
454 		for (i = 0; i < count; i++) {
455 			if (i != reqpage)
456 				vnode_pager_freepage(m[i]);
457 		}
458 		return VM_PAGER_ERROR;
459 	}
460 
461 	if (foff + bytecount > vp->v_filesize) {
462 		bytecount = vp->v_filesize - foff;
463 		i = round_page(bytecount) / PAGE_SIZE;
464 		while (count > i) {
465 			--count;
466 			if (count != reqpage)
467 				vnode_pager_freepage(m[count]);
468 		}
469 	}
470 
471 	/*
472 	 * The size of the transfer is bytecount.  bytecount will be an
473 	 * integral multiple of the page size unless it has been clipped
474 	 * to the file EOF.  The transfer cannot exceed the file EOF.
475 	 *
476 	 * When dealing with real devices we must round-up to the device
477 	 * sector size.
478 	 */
479 	if (vp->v_type == VBLK || vp->v_type == VCHR) {
480 		int secmask = vp->v_rdev->si_bsize_phys - 1;
481 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
482 		bytecount = (bytecount + secmask) & ~secmask;
483 	}
484 
485 	/*
486 	 * Severe hack to avoid deadlocks with the buffer cache
487 	 */
488 	for (i = 0; i < count; ++i) {
489 		vm_page_t mt = m[i];
490 
491 		vm_page_io_start(mt);
492 		vm_page_wakeup(mt);
493 	}
494 
495 	/*
496 	 * Issue the I/O without any read-ahead
497 	 */
498 	ioflags = IO_VMIO;
499 	/*ioflags |= IO_SEQMAX << IO_SEQSHIFT;*/
500 
501 	aiov.iov_base = (caddr_t) 0;
502 	aiov.iov_len = bytecount;
503 	auio.uio_iov = &aiov;
504 	auio.uio_iovcnt = 1;
505 	auio.uio_offset = foff;
506 	auio.uio_segflg = UIO_NOCOPY;
507 	auio.uio_rw = UIO_READ;
508 	auio.uio_resid = bytecount;
509 	auio.uio_td = NULL;
510 	mycpu->gd_cnt.v_vnodein++;
511 	mycpu->gd_cnt.v_vnodepgsin += count;
512 
513 	error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
514 
515 	/*
516 	 * Severe hack to avoid deadlocks with the buffer cache
517 	 */
518 	for (i = 0; i < count; ++i) {
519 		vm_page_t mt = m[i];
520 
521 		while (vm_page_sleep_busy(mt, FALSE, "getpgs"))
522 			;
523 		vm_page_busy(mt);
524 		vm_page_io_finish(mt);
525 	}
526 
527 	/*
528 	 * Calculate the actual number of bytes read and clean up the
529 	 * page list.
530 	 */
531 	bytecount -= auio.uio_resid;
532 
533 	for (i = 0; i < count; ++i) {
534 		vm_page_t mt = m[i];
535 
536 		if (i != reqpage) {
537 			if (error == 0 && mt->valid) {
538 				if (mt->flags & PG_WANTED)
539 					vm_page_activate(mt);
540 				else
541 					vm_page_deactivate(mt);
542 				vm_page_wakeup(mt);
543 			} else {
544 				vnode_pager_freepage(mt);
545 			}
546 		} else if (mt->valid == 0) {
547 			if (error == 0) {
548 				kprintf("page failed but no I/O error page %p object %p pindex %d\n", mt, mt->object, (int) mt->pindex);
549 				/* whoops, something happened */
550 				error = EINVAL;
551 			}
552 		} else if (mt->valid != VM_PAGE_BITS_ALL) {
553 			/*
554 			 * Zero-extend the requested page if necessary (if
555 			 * the filesystem is using a small block size).
556 			 */
557 			vm_page_zero_invalid(mt, TRUE);
558 		}
559 	}
560 	if (error) {
561 		kprintf("vnode_pager_getpages: I/O read error\n");
562 	}
563 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
564 }
565 
566 /*
567  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
568  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
569  * vnode_pager_generic_putpages() to implement the previous behaviour.
570  *
571  * All other FS's should use the bypass to get to the local media
572  * backing vp's VOP_PUTPAGES.
573  */
574 static void
575 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
576     boolean_t sync, int *rtvals)
577 {
578 	int rtval;
579 	struct vnode *vp;
580 	int bytes = count * PAGE_SIZE;
581 
582 	/*
583 	 * Force synchronous operation if we are extremely low on memory
584 	 * to prevent a low-memory deadlock.  VOP operations often need to
585 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
586 	 * operation ).  The swapper handles the case by limiting the amount
587 	 * of asynchronous I/O, but that sort of solution doesn't scale well
588 	 * for the vnode pager without a lot of work.
589 	 *
590 	 * Also, the backing vnode's iodone routine may not wake the pageout
591 	 * daemon up.  This should be probably be addressed XXX.
592 	 */
593 
594 	if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
595 		sync |= OBJPC_SYNC;
596 
597 	/*
598 	 * Call device-specific putpages function
599 	 */
600 
601 	vp = object->handle;
602 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
603 	if (rtval == EOPNOTSUPP) {
604 	    kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
605 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
606 	}
607 }
608 
609 
610 /*
611  * This is now called from local media FS's to operate against their
612  * own vnodes if they fail to implement VOP_PUTPAGES.
613  *
614  * This is typically called indirectly via the pageout daemon and
615  * clustering has already typically occured, so in general we ask the
616  * underlying filesystem to write the data out asynchronously rather
617  * then delayed.
618  */
619 int
620 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
621 			     int flags, int *rtvals)
622 {
623 	int i;
624 	vm_object_t object;
625 	int count;
626 
627 	int maxsize, ncount;
628 	vm_ooffset_t poffset;
629 	struct uio auio;
630 	struct iovec aiov;
631 	int error;
632 	int ioflags;
633 
634 	object = vp->v_object;
635 	count = bytecount / PAGE_SIZE;
636 
637 	for (i = 0; i < count; i++)
638 		rtvals[i] = VM_PAGER_AGAIN;
639 
640 	if ((int) m[0]->pindex < 0) {
641 		kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
642 			(long)m[0]->pindex, m[0]->dirty);
643 		rtvals[0] = VM_PAGER_BAD;
644 		return VM_PAGER_BAD;
645 	}
646 
647 	maxsize = count * PAGE_SIZE;
648 	ncount = count;
649 
650 	poffset = IDX_TO_OFF(m[0]->pindex);
651 
652 	/*
653 	 * If the page-aligned write is larger then the actual file we
654 	 * have to invalidate pages occuring beyond the file EOF.  However,
655 	 * there is an edge case where a file may not be page-aligned where
656 	 * the last page is partially invalid.  In this case the filesystem
657 	 * may not properly clear the dirty bits for the entire page (which
658 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
659 	 * With the page locked we are free to fix-up the dirty bits here.
660 	 *
661 	 * We do not under any circumstances truncate the valid bits, as
662 	 * this will screw up bogus page replacement.
663 	 *
664 	 * The caller has already read-protected the pages.  The VFS must
665 	 * use the buffer cache to wrap the pages.  The pages might not
666 	 * be immediately flushed by the buffer cache but once under its
667 	 * control the pages themselves can wind up being marked clean
668 	 * and their covering buffer cache buffer can be marked dirty.
669 	 */
670 	if (maxsize + poffset > vp->v_filesize) {
671 		if (vp->v_filesize > poffset) {
672 			int pgoff;
673 
674 			maxsize = vp->v_filesize - poffset;
675 			ncount = btoc(maxsize);
676 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
677 				vm_page_clear_dirty(m[ncount - 1], pgoff,
678 					PAGE_SIZE - pgoff);
679 			}
680 		} else {
681 			maxsize = 0;
682 			ncount = 0;
683 		}
684 		if (ncount < count) {
685 			for (i = ncount; i < count; i++) {
686 				rtvals[i] = VM_PAGER_BAD;
687 			}
688 		}
689 	}
690 
691 	/*
692 	 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
693 	 * rather then a bdwrite() to prevent paging I/O from saturating
694 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
695 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
696 	 * the system decides how to cluster.
697 	 */
698 	ioflags = IO_VMIO;
699 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
700 		ioflags |= IO_SYNC;
701 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
702 		ioflags |= IO_ASYNC;
703 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
704 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
705 
706 	aiov.iov_base = (caddr_t) 0;
707 	aiov.iov_len = maxsize;
708 	auio.uio_iov = &aiov;
709 	auio.uio_iovcnt = 1;
710 	auio.uio_offset = poffset;
711 	auio.uio_segflg = UIO_NOCOPY;
712 	auio.uio_rw = UIO_WRITE;
713 	auio.uio_resid = maxsize;
714 	auio.uio_td = NULL;
715 	error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
716 	mycpu->gd_cnt.v_vnodeout++;
717 	mycpu->gd_cnt.v_vnodepgsout += ncount;
718 
719 	if (error) {
720 		krateprintf(&vbadrate,
721 			    "vnode_pager_putpages: I/O error %d\n", error);
722 	}
723 	if (auio.uio_resid) {
724 		krateprintf(&vresrate,
725 			    "vnode_pager_putpages: residual I/O %d at %lu\n",
726 			    auio.uio_resid, (u_long)m[0]->pindex);
727 	}
728 	for (i = 0; i < ncount; i++)
729 		rtvals[i] = VM_PAGER_OK;
730 	return rtvals[0];
731 }
732 
733 struct vnode *
734 vnode_pager_lock(vm_object_t object)
735 {
736 	struct thread *td = curthread;	/* XXX */
737 	int error;
738 
739 	for (; object != NULL; object = object->backing_object) {
740 		if (object->type != OBJT_VNODE)
741 			continue;
742 		if (object->flags & OBJ_DEAD)
743 			return NULL;
744 
745 		for (;;) {
746 			struct vnode *vp = object->handle;
747 			error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
748 			if (error == 0) {
749 				if (object->handle != vp) {
750 					vput(vp);
751 					continue;
752 				}
753 				return (vp);
754 			}
755 			if ((object->flags & OBJ_DEAD) ||
756 			    (object->type != OBJT_VNODE)) {
757 				return NULL;
758 			}
759 			kprintf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
760 			tsleep(object->handle, 0, "vnpgrl", hz);
761 		}
762 	}
763 	return NULL;
764 }
765