xref: /dflybsd-src/sys/vm/vnode_pager.c (revision 884717e1debcf4b08bda1d29d01b0c8a34b86a59)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1990 University of Utah.
5  * Copyright (c) 1991 The Regents of the University of California.
6  * All rights reserved.
7  * Copyright (c) 1993, 1994 John S. Dyson
8  * Copyright (c) 1995, David Greenman
9  *
10  * This code is derived from software contributed to Berkeley by
11  * the Systems Programming Group of the University of Utah Computer
12  * Science Department.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vnode_pager.c	7.5 (Berkeley) 4/20/91
43  * $FreeBSD: src/sys/vm/vnode_pager.c,v 1.116.2.7 2002/12/31 09:34:51 dillon Exp $
44  * $DragonFly: src/sys/vm/vnode_pager.c,v 1.43 2008/06/19 23:27:39 dillon Exp $
45  */
46 
47 /*
48  * Page to/from files (vnodes).
49  */
50 
51 /*
52  * TODO:
53  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
54  *	greatly re-simplify the vnode_pager.
55  */
56 
57 #include <sys/param.h>
58 #include <sys/systm.h>
59 #include <sys/kernel.h>
60 #include <sys/proc.h>
61 #include <sys/vnode.h>
62 #include <sys/mount.h>
63 #include <sys/buf.h>
64 #include <sys/vmmeter.h>
65 #include <sys/conf.h>
66 
67 #include <cpu/lwbuf.h>
68 
69 #include <vm/vm.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_page.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_map.h>
74 #include <vm/vnode_pager.h>
75 #include <vm/swap_pager.h>
76 #include <vm/vm_extern.h>
77 
78 #include <sys/thread2.h>
79 #include <vm/vm_page2.h>
80 
81 static void vnode_pager_dealloc (vm_object_t);
82 static int vnode_pager_getpage (vm_object_t, vm_page_t *, int);
83 static void vnode_pager_putpages (vm_object_t, vm_page_t *, int, boolean_t, int *);
84 static boolean_t vnode_pager_haspage (vm_object_t, vm_pindex_t);
85 
86 struct pagerops vnodepagerops = {
87 	vnode_pager_dealloc,
88 	vnode_pager_getpage,
89 	vnode_pager_putpages,
90 	vnode_pager_haspage
91 };
92 
93 static struct krate vbadrate = { 1 };
94 static struct krate vresrate = { 1 };
95 
96 int vnode_pbuf_freecnt = -1;	/* start out unlimited */
97 
98 /*
99  * Allocate a VM object for a vnode, typically a regular file vnode.
100  *
101  * Some additional information is required to generate a properly sized
102  * object which covers the entire buffer cache buffer straddling the file
103  * EOF.  Userland does not see the extra pages as the VM fault code tests
104  * against v_filesize.
105  */
106 vm_object_t
107 vnode_pager_alloc(void *handle, off_t length, vm_prot_t prot, off_t offset,
108 		  int blksize, int boff)
109 {
110 	vm_object_t object;
111 	struct vnode *vp;
112 	off_t loffset;
113 	vm_pindex_t lsize;
114 
115 	/*
116 	 * Pageout to vnode, no can do yet.
117 	 */
118 	if (handle == NULL)
119 		return (NULL);
120 
121 	/*
122 	 * XXX hack - This initialization should be put somewhere else.
123 	 */
124 	if (vnode_pbuf_freecnt < 0) {
125 	    vnode_pbuf_freecnt = nswbuf / 2 + 1;
126 	}
127 
128 	/*
129 	 * Serialize potential vnode/object teardowns and interlocks
130 	 */
131 	vp = (struct vnode *)handle;
132 	lwkt_gettoken(&vp->v_token);
133 
134 	/*
135 	 * Prevent race condition when allocating the object. This
136 	 * can happen with NFS vnodes since the nfsnode isn't locked.
137 	 */
138 	while (vp->v_flag & VOLOCK) {
139 		vsetflags(vp, VOWANT);
140 		tsleep(vp, 0, "vnpobj", 0);
141 	}
142 	vsetflags(vp, VOLOCK);
143 	lwkt_reltoken(&vp->v_token);
144 
145 	/*
146 	 * If the object is being terminated, wait for it to
147 	 * go away.
148 	 */
149 	while ((object = vp->v_object) != NULL) {
150 		vm_object_hold(object);
151 		if ((object->flags & OBJ_DEAD) == 0)
152 			break;
153 		vm_object_dead_sleep(object, "vadead");
154 		vm_object_drop(object);
155 	}
156 
157 	if (vp->v_sysref.refcnt <= 0)
158 		panic("vnode_pager_alloc: no vnode reference");
159 
160 	/*
161 	 * Round up to the *next* block, then destroy the buffers in question.
162 	 * Since we are only removing some of the buffers we must rely on the
163 	 * scan count to determine whether a loop is necessary.
164 	 *
165 	 * Destroy any pages beyond the last buffer.
166 	 */
167 	if (boff < 0)
168 		boff = (int)(length % blksize);
169 	if (boff)
170 		loffset = length + (blksize - boff);
171 	else
172 		loffset = length;
173 	lsize = OFF_TO_IDX(round_page64(loffset));
174 
175 	if (object == NULL) {
176 		/*
177 		 * And an object of the appropriate size
178 		 */
179 		object = vm_object_allocate_hold(OBJT_VNODE, lsize);
180 		object->flags = 0;
181 		object->handle = handle;
182 		vp->v_object = object;
183 		vp->v_filesize = length;
184 		if (vp->v_mount && (vp->v_mount->mnt_kern_flag & MNTK_NOMSYNC))
185 			vm_object_set_flag(object, OBJ_NOMSYNC);
186 	} else {
187 		object->ref_count++;
188 		if (object->size != lsize) {
189 			kprintf("vnode_pager_alloc: Warning, objsize "
190 				"mismatch %jd/%jd vp=%p obj=%p\n",
191 				(intmax_t)object->size,
192 				(intmax_t)lsize,
193 				vp, object);
194 		}
195 		if (vp->v_filesize != length) {
196 			kprintf("vnode_pager_alloc: Warning, filesize "
197 				"mismatch %jd/%jd vp=%p obj=%p\n",
198 				(intmax_t)vp->v_filesize,
199 				(intmax_t)length,
200 				vp, object);
201 		}
202 	}
203 
204 	vref(vp);
205 	lwkt_gettoken(&vp->v_token);
206 	vclrflags(vp, VOLOCK);
207 	if (vp->v_flag & VOWANT) {
208 		vclrflags(vp, VOWANT);
209 		wakeup(vp);
210 	}
211 	lwkt_reltoken(&vp->v_token);
212 
213 	vm_object_drop(object);
214 
215 	return (object);
216 }
217 
218 /*
219  * Add a ref to a vnode's existing VM object, return the object or
220  * NULL if the vnode did not have one.  This does not create the
221  * object (we can't since we don't know what the proper blocksize/boff
222  * is to match the VFS's use of the buffer cache).
223  */
224 vm_object_t
225 vnode_pager_reference(struct vnode *vp)
226 {
227 	vm_object_t object;
228 
229 	/*
230 	 * Prevent race condition when allocating the object. This
231 	 * can happen with NFS vnodes since the nfsnode isn't locked.
232 	 *
233 	 * Serialize potential vnode/object teardowns and interlocks
234 	 */
235 	lwkt_gettoken(&vp->v_token);
236 	while (vp->v_flag & VOLOCK) {
237 		vsetflags(vp, VOWANT);
238 		tsleep(vp, 0, "vnpobj", 0);
239 	}
240 	vsetflags(vp, VOLOCK);
241 	lwkt_reltoken(&vp->v_token);
242 
243 	/*
244 	 * Prevent race conditions against deallocation of the VM
245 	 * object.
246 	 */
247 	while ((object = vp->v_object) != NULL) {
248 		vm_object_hold(object);
249 		if ((object->flags & OBJ_DEAD) == 0)
250 			break;
251 		vm_object_dead_sleep(object, "vadead");
252 		vm_object_drop(object);
253 	}
254 
255 	/*
256 	 * The object is expected to exist, the caller will handle
257 	 * NULL returns if it does not.
258 	 */
259 	if (object) {
260 		object->ref_count++;
261 		vref(vp);
262 	}
263 
264 	lwkt_gettoken(&vp->v_token);
265 	vclrflags(vp, VOLOCK);
266 	if (vp->v_flag & VOWANT) {
267 		vclrflags(vp, VOWANT);
268 		wakeup(vp);
269 	}
270 	lwkt_reltoken(&vp->v_token);
271 	if (object)
272 		vm_object_drop(object);
273 
274 	return (object);
275 }
276 
277 static void
278 vnode_pager_dealloc(vm_object_t object)
279 {
280 	struct vnode *vp = object->handle;
281 
282 	if (vp == NULL)
283 		panic("vnode_pager_dealloc: pager already dealloced");
284 
285 	vm_object_pip_wait(object, "vnpdea");
286 
287 	object->handle = NULL;
288 	object->type = OBJT_DEAD;
289 	vp->v_object = NULL;
290 	vp->v_filesize = NOOFFSET;
291 	vclrflags(vp, VTEXT | VOBJBUF);
292 	swap_pager_freespace_all(object);
293 }
294 
295 /*
296  * Return whether the vnode pager has the requested page.  Return the
297  * number of disk-contiguous pages before and after the requested page,
298  * not including the requested page.
299  */
300 static boolean_t
301 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex)
302 {
303 	struct vnode *vp = object->handle;
304 	off_t loffset;
305 	off_t doffset;
306 	int voff;
307 	int bsize;
308 	int error;
309 
310 	/*
311 	 * If no vp or vp is doomed or marked transparent to VM, we do not
312 	 * have the page.
313 	 */
314 	if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
315 		return FALSE;
316 
317 	/*
318 	 * If filesystem no longer mounted or offset beyond end of file we do
319 	 * not have the page.
320 	 */
321 	loffset = IDX_TO_OFF(pindex);
322 
323 	if (vp->v_mount == NULL || loffset >= vp->v_filesize)
324 		return FALSE;
325 
326 	bsize = vp->v_mount->mnt_stat.f_iosize;
327 	voff = loffset % bsize;
328 
329 	/*
330 	 * XXX
331 	 *
332 	 * BMAP returns byte counts before and after, where after
333 	 * is inclusive of the base page.  haspage must return page
334 	 * counts before and after where after does not include the
335 	 * base page.
336 	 *
337 	 * BMAP is allowed to return a *after of 0 for backwards
338 	 * compatibility.  The base page is still considered valid if
339 	 * no error is returned.
340 	 */
341 	error = VOP_BMAP(vp, loffset - voff, &doffset, NULL, NULL, 0);
342 	if (error)
343 		return TRUE;
344 	if (doffset == NOOFFSET)
345 		return FALSE;
346 	return TRUE;
347 }
348 
349 /*
350  * Lets the VM system know about a change in size for a file.
351  * We adjust our own internal size and flush any cached pages in
352  * the associated object that are affected by the size change.
353  *
354  * NOTE: This routine may be invoked as a result of a pager put
355  * operation (possibly at object termination time), so we must be careful.
356  *
357  * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
358  * we do not blow up on the case.  nsize will always be >= 0, however.
359  */
360 void
361 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
362 {
363 	vm_pindex_t nobjsize;
364 	vm_pindex_t oobjsize;
365 	vm_object_t object;
366 
367 	while ((object = vp->v_object) != NULL) {
368 		vm_object_hold(object);
369 		if (vp->v_object == object)
370 			break;
371 		vm_object_drop(object);
372 	}
373 	if (object == NULL)
374 		return;
375 
376 	/*
377 	 * Hasn't changed size
378 	 */
379 	if (nsize == vp->v_filesize) {
380 		vm_object_drop(object);
381 		return;
382 	}
383 
384 	/*
385 	 * Has changed size.  Adjust the VM object's size and v_filesize
386 	 * before we start scanning pages to prevent new pages from being
387 	 * allocated during the scan.
388 	 */
389 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
390 	oobjsize = object->size;
391 	object->size = nobjsize;
392 
393 	/*
394 	 * File has shrunk. Toss any cached pages beyond the new EOF.
395 	 */
396 	if (nsize < vp->v_filesize) {
397 		vp->v_filesize = nsize;
398 		if (nobjsize < oobjsize) {
399 			vm_object_page_remove(object, nobjsize, oobjsize,
400 					      FALSE);
401 		}
402 		/*
403 		 * This gets rid of garbage at the end of a page that is now
404 		 * only partially backed by the vnode.  Since we are setting
405 		 * the entire page valid & clean after we are done we have
406 		 * to be sure that the portion of the page within the file
407 		 * bounds is already valid.  If it isn't then making it
408 		 * valid would create a corrupt block.
409 		 */
410 		if (nsize & PAGE_MASK) {
411 			vm_offset_t kva;
412 			vm_page_t m;
413 
414 			m = vm_page_lookup_busy_wait(object, OFF_TO_IDX(nsize),
415 						     TRUE, "vsetsz");
416 
417 			if (m && m->valid) {
418 				int base = (int)nsize & PAGE_MASK;
419 				int size = PAGE_SIZE - base;
420 				struct lwbuf *lwb;
421 				struct lwbuf lwb_cache;
422 
423 				/*
424 				 * Clear out partial-page garbage in case
425 				 * the page has been mapped.
426 				 *
427 				 * This is byte aligned.
428 				 */
429 				lwb = lwbuf_alloc(m, &lwb_cache);
430 				kva = lwbuf_kva(lwb);
431 				bzero((caddr_t)kva + base, size);
432 				lwbuf_free(lwb);
433 
434 				/*
435 				 * XXX work around SMP data integrity race
436 				 * by unmapping the page from user processes.
437 				 * The garbage we just cleared may be mapped
438 				 * to a user process running on another cpu
439 				 * and this code is not running through normal
440 				 * I/O channels which handle SMP issues for
441 				 * us, so unmap page to synchronize all cpus.
442 				 *
443 				 * XXX should vm_pager_unmap_page() have
444 				 * dealt with this?
445 				 */
446 				vm_page_protect(m, VM_PROT_NONE);
447 
448 				/*
449 				 * Clear out partial-page dirty bits.  This
450 				 * has the side effect of setting the valid
451 				 * bits, but that is ok.  There are a bunch
452 				 * of places in the VM system where we expected
453 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
454 				 * case is one of them.  If the page is still
455 				 * partially dirty, make it fully dirty.
456 				 *
457 				 * NOTE: We do not clear out the valid
458 				 * bits.  This would prevent bogus_page
459 				 * replacement from working properly.
460 				 *
461 				 * NOTE: We do not want to clear the dirty
462 				 * bit for a partial DEV_BSIZE'd truncation!
463 				 * This is DEV_BSIZE aligned!
464 				 */
465 				vm_page_clear_dirty_beg_nonincl(m, base, size);
466 				if (m->dirty != 0)
467 					m->dirty = VM_PAGE_BITS_ALL;
468 				vm_page_wakeup(m);
469 			} else if (m) {
470 				vm_page_wakeup(m);
471 			}
472 		}
473 	} else {
474 		vp->v_filesize = nsize;
475 	}
476 	vm_object_drop(object);
477 }
478 
479 /*
480  * Release a page busied for a getpages operation.  The page may have become
481  * wired (typically due to being used by the buffer cache) or otherwise been
482  * soft-busied and cannot be freed in that case.  A held page can still be
483  * freed.
484  */
485 void
486 vnode_pager_freepage(vm_page_t m)
487 {
488 	if (m->busy || m->wire_count) {
489 		vm_page_activate(m);
490 		vm_page_wakeup(m);
491 	} else {
492 		vm_page_free(m);
493 	}
494 }
495 
496 /*
497  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
498  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
499  * vnode_pager_generic_getpages() to implement the previous behaviour.
500  *
501  * All other FS's should use the bypass to get to the local media
502  * backing vp's VOP_GETPAGES.
503  */
504 static int
505 vnode_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
506 {
507 	int rtval;
508 	struct vnode *vp;
509 
510 	vp = object->handle;
511 	rtval = VOP_GETPAGES(vp, mpp, PAGE_SIZE, 0, 0, seqaccess);
512 	if (rtval == EOPNOTSUPP)
513 		panic("vnode_pager: vfs's must implement vop_getpages\n");
514 	return rtval;
515 }
516 
517 /*
518  * This is now called from local media FS's to operate against their
519  * own vnodes if they fail to implement VOP_GETPAGES.
520  *
521  * With all the caching local media devices do these days there is really
522  * very little point to attempting to restrict the I/O size to contiguous
523  * blocks on-disk, especially if our caller thinks we need all the specified
524  * pages.  Just construct and issue a READ.
525  */
526 int
527 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *mpp, int bytecount,
528 			     int reqpage, int seqaccess)
529 {
530 	struct iovec aiov;
531 	struct uio auio;
532 	off_t foff;
533 	int error;
534 	int count;
535 	int i;
536 	int ioflags;
537 
538 	/*
539 	 * Do not do anything if the vnode is bad.
540 	 */
541 	if (vp->v_mount == NULL)
542 		return VM_PAGER_BAD;
543 
544 	/*
545 	 * Calculate the number of pages.  Since we are paging in whole
546 	 * pages, adjust bytecount to be an integral multiple of the page
547 	 * size.  It will be clipped to the file EOF later on.
548 	 */
549 	bytecount = round_page(bytecount);
550 	count = bytecount / PAGE_SIZE;
551 
552 	/*
553 	 * We could check m[reqpage]->valid here and shortcut the operation,
554 	 * but doing so breaks read-ahead.  Instead assume that the VM
555 	 * system has already done at least the check, don't worry about
556 	 * any races, and issue the VOP_READ to allow read-ahead to function.
557 	 *
558 	 * This keeps the pipeline full for I/O bound sequentially scanned
559 	 * mmap()'s
560 	 */
561 	/* don't shortcut */
562 
563 	/*
564 	 * Discard pages past the file EOF.  If the requested page is past
565 	 * the file EOF we just leave its valid bits set to 0, the caller
566 	 * expects to maintain ownership of the requested page.  If the
567 	 * entire range is past file EOF discard everything and generate
568 	 * a pagein error.
569 	 */
570 	foff = IDX_TO_OFF(mpp[0]->pindex);
571 	if (foff >= vp->v_filesize) {
572 		for (i = 0; i < count; i++) {
573 			if (i != reqpage)
574 				vnode_pager_freepage(mpp[i]);
575 		}
576 		return VM_PAGER_ERROR;
577 	}
578 
579 	if (foff + bytecount > vp->v_filesize) {
580 		bytecount = vp->v_filesize - foff;
581 		i = round_page(bytecount) / PAGE_SIZE;
582 		while (count > i) {
583 			--count;
584 			if (count != reqpage)
585 				vnode_pager_freepage(mpp[count]);
586 		}
587 	}
588 
589 	/*
590 	 * The size of the transfer is bytecount.  bytecount will be an
591 	 * integral multiple of the page size unless it has been clipped
592 	 * to the file EOF.  The transfer cannot exceed the file EOF.
593 	 *
594 	 * When dealing with real devices we must round-up to the device
595 	 * sector size.
596 	 */
597 	if (vp->v_type == VBLK || vp->v_type == VCHR) {
598 		int secmask = vp->v_rdev->si_bsize_phys - 1;
599 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
600 		bytecount = (bytecount + secmask) & ~secmask;
601 	}
602 
603 	/*
604 	 * Severe hack to avoid deadlocks with the buffer cache
605 	 */
606 	for (i = 0; i < count; ++i) {
607 		vm_page_t mt = mpp[i];
608 
609 		vm_page_io_start(mt);
610 		vm_page_wakeup(mt);
611 	}
612 
613 	/*
614 	 * Issue the I/O with some read-ahead if bytecount > PAGE_SIZE
615 	 */
616 	ioflags = IO_VMIO;
617 	if (seqaccess)
618 		ioflags |= IO_SEQMAX << IO_SEQSHIFT;
619 
620 	aiov.iov_base = NULL;
621 	aiov.iov_len = bytecount;
622 	auio.uio_iov = &aiov;
623 	auio.uio_iovcnt = 1;
624 	auio.uio_offset = foff;
625 	auio.uio_segflg = UIO_NOCOPY;
626 	auio.uio_rw = UIO_READ;
627 	auio.uio_resid = bytecount;
628 	auio.uio_td = NULL;
629 	mycpu->gd_cnt.v_vnodein++;
630 	mycpu->gd_cnt.v_vnodepgsin += count;
631 
632 	error = VOP_READ(vp, &auio, ioflags, proc0.p_ucred);
633 
634 	/*
635 	 * Severe hack to avoid deadlocks with the buffer cache
636 	 */
637 	for (i = 0; i < count; ++i) {
638 		vm_page_busy_wait(mpp[i], FALSE, "getpgs");
639 		vm_page_io_finish(mpp[i]);
640 	}
641 
642 	/*
643 	 * Calculate the actual number of bytes read and clean up the
644 	 * page list.
645 	 */
646 	bytecount -= auio.uio_resid;
647 
648 	for (i = 0; i < count; ++i) {
649 		vm_page_t mt = mpp[i];
650 
651 		if (i != reqpage) {
652 			if (error == 0 && mt->valid) {
653 				if (mt->flags & PG_REFERENCED)
654 					vm_page_activate(mt);
655 				else
656 					vm_page_deactivate(mt);
657 				vm_page_wakeup(mt);
658 			} else {
659 				vnode_pager_freepage(mt);
660 			}
661 		} else if (mt->valid == 0) {
662 			if (error == 0) {
663 				kprintf("page failed but no I/O error page "
664 					"%p object %p pindex %d\n",
665 					mt, mt->object, (int) mt->pindex);
666 				/* whoops, something happened */
667 				error = EINVAL;
668 			}
669 		} else if (mt->valid != VM_PAGE_BITS_ALL) {
670 			/*
671 			 * Zero-extend the requested page if necessary (if
672 			 * the filesystem is using a small block size).
673 			 */
674 			vm_page_zero_invalid(mt, TRUE);
675 		}
676 	}
677 	if (error) {
678 		kprintf("vnode_pager_getpage: I/O read error\n");
679 	}
680 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
681 }
682 
683 /*
684  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
685  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
686  * vnode_pager_generic_putpages() to implement the previous behaviour.
687  *
688  * Caller has already cleared the pmap modified bits, if any.
689  *
690  * All other FS's should use the bypass to get to the local media
691  * backing vp's VOP_PUTPAGES.
692  */
693 static void
694 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
695 		     boolean_t sync, int *rtvals)
696 {
697 	int rtval;
698 	struct vnode *vp;
699 	int bytes = count * PAGE_SIZE;
700 
701 	/*
702 	 * Force synchronous operation if we are extremely low on memory
703 	 * to prevent a low-memory deadlock.  VOP operations often need to
704 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
705 	 * operation ).  The swapper handles the case by limiting the amount
706 	 * of asynchronous I/O, but that sort of solution doesn't scale well
707 	 * for the vnode pager without a lot of work.
708 	 *
709 	 * Also, the backing vnode's iodone routine may not wake the pageout
710 	 * daemon up.  This should be probably be addressed XXX.
711 	 */
712 
713 	if ((vmstats.v_free_count + vmstats.v_cache_count) <
714 	    vmstats.v_pageout_free_min) {
715 		sync |= OBJPC_SYNC;
716 	}
717 
718 	/*
719 	 * Call device-specific putpages function
720 	 */
721 	vp = object->handle;
722 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
723 	if (rtval == EOPNOTSUPP) {
724 	    kprintf("vnode_pager: *** WARNING *** stale FS putpages\n");
725 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
726 	}
727 }
728 
729 
730 /*
731  * This is now called from local media FS's to operate against their
732  * own vnodes if they fail to implement VOP_PUTPAGES.
733  *
734  * This is typically called indirectly via the pageout daemon and
735  * clustering has already typically occured, so in general we ask the
736  * underlying filesystem to write the data out asynchronously rather
737  * then delayed.
738  */
739 int
740 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
741 			     int flags, int *rtvals)
742 {
743 	int i;
744 	vm_object_t object;
745 	int maxsize, ncount, count;
746 	vm_ooffset_t poffset;
747 	struct uio auio;
748 	struct iovec aiov;
749 	int error;
750 	int ioflags;
751 
752 	object = vp->v_object;
753 	count = bytecount / PAGE_SIZE;
754 
755 	for (i = 0; i < count; i++)
756 		rtvals[i] = VM_PAGER_AGAIN;
757 
758 	if ((int) m[0]->pindex < 0) {
759 		kprintf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
760 			(long)m[0]->pindex, m[0]->dirty);
761 		rtvals[0] = VM_PAGER_BAD;
762 		return VM_PAGER_BAD;
763 	}
764 
765 	maxsize = count * PAGE_SIZE;
766 	ncount = count;
767 
768 	poffset = IDX_TO_OFF(m[0]->pindex);
769 
770 	/*
771 	 * If the page-aligned write is larger then the actual file we
772 	 * have to invalidate pages occuring beyond the file EOF.
773 	 *
774 	 * If the file EOF resides in the middle of a page we still clear
775 	 * all of that page's dirty bits later on.  If we didn't it would
776 	 * endlessly re-write.
777 	 *
778 	 * We do not under any circumstances truncate the valid bits, as
779 	 * this will screw up bogus page replacement.
780 	 *
781 	 * The caller has already read-protected the pages.  The VFS must
782 	 * use the buffer cache to wrap the pages.  The pages might not
783 	 * be immediately flushed by the buffer cache but once under its
784 	 * control the pages themselves can wind up being marked clean
785 	 * and their covering buffer cache buffer can be marked dirty.
786 	 */
787 	if (poffset + maxsize > vp->v_filesize) {
788 		if (poffset < vp->v_filesize) {
789 			maxsize = vp->v_filesize - poffset;
790 			ncount = btoc(maxsize);
791 		} else {
792 			maxsize = 0;
793 			ncount = 0;
794 		}
795 		if (ncount < count) {
796 			for (i = ncount; i < count; i++) {
797 				rtvals[i] = VM_PAGER_BAD;
798 			}
799 		}
800 	}
801 
802 	/*
803 	 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
804 	 * rather then a bdwrite() to prevent paging I/O from saturating
805 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
806 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
807 	 * the system decides how to cluster.
808 	 */
809 	ioflags = IO_VMIO;
810 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
811 		ioflags |= IO_SYNC;
812 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
813 		ioflags |= IO_ASYNC;
814 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
815 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
816 
817 	aiov.iov_base = (caddr_t) 0;
818 	aiov.iov_len = maxsize;
819 	auio.uio_iov = &aiov;
820 	auio.uio_iovcnt = 1;
821 	auio.uio_offset = poffset;
822 	auio.uio_segflg = UIO_NOCOPY;
823 	auio.uio_rw = UIO_WRITE;
824 	auio.uio_resid = maxsize;
825 	auio.uio_td = NULL;
826 	error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
827 	mycpu->gd_cnt.v_vnodeout++;
828 	mycpu->gd_cnt.v_vnodepgsout += ncount;
829 
830 	if (error) {
831 		krateprintf(&vbadrate,
832 			    "vnode_pager_putpages: I/O error %d\n", error);
833 	}
834 	if (auio.uio_resid) {
835 		krateprintf(&vresrate,
836 			    "vnode_pager_putpages: residual I/O %zd at %lu\n",
837 			    auio.uio_resid, (u_long)m[0]->pindex);
838 	}
839 	if (error == 0) {
840 		for (i = 0; i < ncount; i++) {
841 			rtvals[i] = VM_PAGER_OK;
842 			vm_page_undirty(m[i]);
843 		}
844 	}
845 	return rtvals[0];
846 }
847 
848 /*
849  * Run the chain and if the bottom-most object is a vnode-type lock the
850  * underlying vnode.  A locked vnode or NULL is returned.
851  */
852 struct vnode *
853 vnode_pager_lock(vm_object_t object)
854 {
855 	struct vnode *vp = NULL;
856 	vm_object_t lobject;
857 	vm_object_t tobject;
858 	int error;
859 
860 	if (object == NULL)
861 		return(NULL);
862 
863 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
864 	lobject = object;
865 
866 	while (lobject->type != OBJT_VNODE) {
867 		if (lobject->flags & OBJ_DEAD)
868 			break;
869 		tobject = lobject->backing_object;
870 		if (tobject == NULL)
871 			break;
872 		vm_object_hold(tobject);
873 		if (tobject == lobject->backing_object) {
874 			if (lobject != object) {
875 				vm_object_lock_swap();
876 				vm_object_drop(lobject);
877 			}
878 			lobject = tobject;
879 		} else {
880 			vm_object_drop(tobject);
881 		}
882 	}
883 	while (lobject->type == OBJT_VNODE &&
884 	       (lobject->flags & OBJ_DEAD) == 0) {
885 		/*
886 		 * Extract the vp
887 		 */
888 		vp = lobject->handle;
889 		error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
890 		if (error == 0) {
891 			if (lobject->handle == vp)
892 				break;
893 			vput(vp);
894 		} else {
895 			kprintf("vnode_pager_lock: vp %p error %d "
896 				"lockstatus %d, retrying\n",
897 				vp, error,
898 				lockstatus(&vp->v_lock, curthread));
899 			tsleep(object->handle, 0, "vnpgrl", hz);
900 		}
901 		vp = NULL;
902 	}
903 	if (lobject != object)
904 		vm_object_drop(lobject);
905 	return (vp);
906 }
907