xref: /dflybsd-src/sys/vm/vnode_pager.c (revision 41871674d0079dec70d55eb824f39d07dc7b3310)
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 1.23 2006/03/27 01:54:18 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 off_t vnode_pager_addr (struct vnode *vp, off_t loffset, int *run);
76 static void vnode_pager_iodone (struct bio *bio);
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 	NULL,
86 	vnode_pager_alloc,
87 	vnode_pager_dealloc,
88 	vnode_pager_getpages,
89 	vnode_pager_putpages,
90 	vnode_pager_haspage,
91 	NULL
92 };
93 
94 int vnode_pbuf_freecnt = -1;	/* start out unlimited */
95 
96 /*
97  * Allocate (or lookup) pager for a vnode.
98  * Handle is a vnode pointer.
99  */
100 vm_object_t
101 vnode_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
102 {
103 	vm_object_t object;
104 	struct vnode *vp;
105 
106 	/*
107 	 * Pageout to vnode, no can do yet.
108 	 */
109 	if (handle == NULL)
110 		return (NULL);
111 
112 	/*
113 	 * XXX hack - This initialization should be put somewhere else.
114 	 */
115 	if (vnode_pbuf_freecnt < 0) {
116 	    vnode_pbuf_freecnt = nswbuf / 2 + 1;
117 	}
118 
119 	vp = (struct vnode *) handle;
120 
121 	/*
122 	 * Prevent race condition when allocating the object. This
123 	 * can happen with NFS vnodes since the nfsnode isn't locked.
124 	 */
125 	while (vp->v_flag & VOLOCK) {
126 		vp->v_flag |= VOWANT;
127 		tsleep(vp, 0, "vnpobj", 0);
128 	}
129 	vp->v_flag |= VOLOCK;
130 
131 	/*
132 	 * If the object is being terminated, wait for it to
133 	 * go away.
134 	 */
135 	while (((object = vp->v_object) != NULL) &&
136 		(object->flags & OBJ_DEAD)) {
137 		tsleep(object, 0, "vadead", 0);
138 	}
139 
140 	if (vp->v_usecount == 0)
141 		panic("vnode_pager_alloc: no vnode reference");
142 
143 	if (object == NULL) {
144 		/*
145 		 * And an object of the appropriate size
146 		 */
147 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
148 		object->flags = 0;
149 		object->handle = handle;
150 		vp->v_object = object;
151 		vp->v_filesize = size;
152 	} else {
153 		object->ref_count++;
154 		if (vp->v_filesize != size)
155 			printf("vnode_pager_alloc: Warning, filesize mismatch %lld/%lld\n", vp->v_filesize, size);
156 	}
157 	vp->v_usecount++;
158 
159 	vp->v_flag &= ~VOLOCK;
160 	if (vp->v_flag & VOWANT) {
161 		vp->v_flag &= ~VOWANT;
162 		wakeup(vp);
163 	}
164 	return (object);
165 }
166 
167 static void
168 vnode_pager_dealloc(vm_object_t object)
169 {
170 	struct vnode *vp = object->handle;
171 
172 	if (vp == NULL)
173 		panic("vnode_pager_dealloc: pager already dealloced");
174 
175 	vm_object_pip_wait(object, "vnpdea");
176 
177 	object->handle = NULL;
178 	object->type = OBJT_DEAD;
179 	vp->v_object = NULL;
180 	vp->v_filesize = NOOFFSET;
181 	vp->v_flag &= ~(VTEXT | VOBJBUF);
182 }
183 
184 /*
185  * Return whether the vnode pager has the requested page.  Return the
186  * number of disk-contiguous pages before and after the requested page,
187  * not including the requested page.
188  */
189 static boolean_t
190 vnode_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
191 		    int *after)
192 {
193 	struct vnode *vp = object->handle;
194 	off_t loffset;
195 	off_t doffset;
196 	int voff;
197 	int bsize;
198 	int error;
199 
200 	/*
201 	 * If no vp or vp is doomed or marked transparent to VM, we do not
202 	 * have the page.
203 	 */
204 	if ((vp == NULL) || (vp->v_flag & VRECLAIMED))
205 		return FALSE;
206 
207 	/*
208 	 * If filesystem no longer mounted or offset beyond end of file we do
209 	 * not have the page.
210 	 */
211 	loffset = IDX_TO_OFF(pindex);
212 
213 	if (vp->v_mount == NULL || loffset >= vp->v_filesize)
214 		return FALSE;
215 
216 	bsize = vp->v_mount->mnt_stat.f_iosize;
217 	voff = loffset % bsize;
218 
219 	error = VOP_BMAP(vp, loffset - voff, NULL, &doffset, after, before);
220 	if (error)
221 		return TRUE;
222 	if (doffset == NOOFFSET)
223 		return FALSE;
224 
225 	if (before) {
226 		*before = (*before + voff) >> PAGE_SHIFT;
227 	}
228 	if (after) {
229 		*after -= voff;
230 		if (loffset + *after > vp->v_filesize)
231 			*after = vp->v_filesize - loffset;
232 		*after >>= PAGE_SHIFT;
233 		if (*after < 0)
234 			*after = 0;
235 	}
236 	return TRUE;
237 }
238 
239 /*
240  * Lets the VM system know about a change in size for a file.
241  * We adjust our own internal size and flush any cached pages in
242  * the associated object that are affected by the size change.
243  *
244  * NOTE: This routine may be invoked as a result of a pager put
245  * operation (possibly at object termination time), so we must be careful.
246  *
247  * NOTE: vp->v_filesize is initialized to NOOFFSET (-1), be sure that
248  * we do not blow up on the case.  nsize will always be >= 0, however.
249  */
250 void
251 vnode_pager_setsize(struct vnode *vp, vm_ooffset_t nsize)
252 {
253 	vm_pindex_t nobjsize;
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 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
266 
267 	/*
268 	 * File has shrunk. Toss any cached pages beyond the new EOF.
269 	 */
270 	if (nsize < vp->v_filesize) {
271 		vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size);
272 		if (nobjsize < object->size) {
273 			vm_object_page_remove(object, nobjsize, object->size,
274 				FALSE);
275 		}
276 		/*
277 		 * This gets rid of garbage at the end of a page that is now
278 		 * only partially backed by the vnode.  Since we are setting
279 		 * the entire page valid & clean after we are done we have
280 		 * to be sure that the portion of the page within the file
281 		 * bounds is already valid.  If it isn't then making it
282 		 * valid would create a corrupt block.
283 		 */
284 		if (nsize & PAGE_MASK) {
285 			vm_offset_t kva;
286 			vm_page_t m;
287 
288 			m = vm_page_lookup(object, OFF_TO_IDX(nsize));
289 			if (m && m->valid) {
290 				int base = (int)nsize & PAGE_MASK;
291 				int size = PAGE_SIZE - base;
292 				struct sf_buf *sf;
293 
294 				/*
295 				 * Clear out partial-page garbage in case
296 				 * the page has been mapped.
297 				 */
298 				sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
299 				kva = sf_buf_kva(sf);
300 				bzero((caddr_t)kva + base, size);
301 				sf_buf_free(sf);
302 
303 				/*
304 				 * XXX work around SMP data integrity race
305 				 * by unmapping the page from user processes.
306 				 * The garbage we just cleared may be mapped
307 				 * to a user process running on another cpu
308 				 * and this code is not running through normal
309 				 * I/O channels which handle SMP issues for
310 				 * us, so unmap page to synchronize all cpus.
311 				 *
312 				 * XXX should vm_pager_unmap_page() have
313 				 * dealt with this?
314 				 */
315 				vm_page_protect(m, VM_PROT_NONE);
316 
317 				/*
318 				 * Clear out partial-page dirty bits.  This
319 				 * has the side effect of setting the valid
320 				 * bits, but that is ok.  There are a bunch
321 				 * of places in the VM system where we expected
322 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
323 				 * case is one of them.  If the page is still
324 				 * partially dirty, make it fully dirty.
325 				 *
326 				 * note that we do not clear out the valid
327 				 * bits.  This would prevent bogus_page
328 				 * replacement from working properly.
329 				 */
330 				vm_page_set_validclean(m, base, size);
331 				if (m->dirty != 0)
332 					m->dirty = VM_PAGE_BITS_ALL;
333 			}
334 		}
335 	}
336 	vp->v_filesize = nsize;
337 	object->size = nobjsize;
338 }
339 
340 void
341 vnode_pager_freepage(vm_page_t m)
342 {
343 	vm_page_free(m);
344 }
345 
346 /*
347  * calculate the disk byte address of specified logical byte offset.  The
348  * logical offset will be block-aligned.  Return the number of contiguous
349  * pages that may be read from the underlying block device in *run.  If
350  * *run is non-NULL, it will be set to a value of at least 1.
351  */
352 static off_t
353 vnode_pager_addr(struct vnode *vp, off_t loffset, int *run)
354 {
355 	struct vnode *rtvp;
356 	off_t doffset;
357 	int bsize;
358 	int error;
359 	int voff;
360 
361 	if (loffset < 0)
362 		return -1;
363 
364 	if (vp->v_mount == NULL)
365 		return -1;
366 
367 	/*
368 	 * Align loffset to a block boundary for the BMAP, then adjust the
369 	 * returned disk address appropriately.
370 	 */
371 	bsize = vp->v_mount->mnt_stat.f_iosize;
372 	voff = loffset % bsize;
373 
374 	/*
375 	 * Map the block, adjust the disk offset so it represents the
376 	 * passed loffset rather then the block containing loffset.
377 	 */
378 	error = VOP_BMAP(vp, loffset - voff, &rtvp, &doffset, run, NULL);
379 	if (error || doffset == NOOFFSET) {
380 		doffset = NOOFFSET;
381 	} else {
382 		doffset += voff;
383 
384 		/*
385 		 * When calculating *run, which is the number of pages
386 		 * worth of data which can be read linearly from disk,
387 		 * the minimum return value is 1 page.
388 		 */
389 		if (run) {
390 			*run = (*run - voff) >> PAGE_SHIFT;
391 			if (*run < 1)
392 				*run = 1;
393 		}
394 
395 	}
396 	return (doffset);
397 }
398 
399 /*
400  * interrupt routine for I/O completion
401  */
402 static void
403 vnode_pager_iodone(struct bio *bio)
404 {
405 	struct buf *bp = bio->bio_buf;
406 
407 	bp->b_flags |= B_DONE;
408 	wakeup(bp);
409 }
410 
411 /*
412  * small block file system vnode pager input
413  */
414 static int
415 vnode_pager_input_smlfs(vm_object_t object, vm_page_t m)
416 {
417 	int i;
418 	struct vnode *dp, *vp;
419 	struct buf *bp;
420 	vm_offset_t kva;
421 	struct sf_buf *sf;
422 	off_t doffset;
423 	vm_offset_t bsize;
424 	int error = 0;
425 
426 	vp = object->handle;
427 	if (vp->v_mount == NULL)
428 		return VM_PAGER_BAD;
429 
430 	bsize = vp->v_mount->mnt_stat.f_iosize;
431 
432 
433 	VOP_BMAP(vp, (off_t)0, &dp, NULL, NULL, NULL);
434 
435 	sf = sf_buf_alloc(m, 0);
436 	kva = sf_buf_kva(sf);
437 
438 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
439 		off_t loffset;
440 
441 		if (vm_page_bits(i * bsize, bsize) & m->valid)
442 			continue;
443 
444 		loffset = IDX_TO_OFF(m->pindex) + i * bsize;
445 		if (loffset >= vp->v_filesize) {
446 			doffset = NOOFFSET;
447 		} else {
448 			doffset = vnode_pager_addr(vp, loffset, NULL);
449 		}
450 		if (doffset != NOOFFSET) {
451 			bp = getpbuf(&vnode_pbuf_freecnt);
452 
453 			/* build a minimal buffer header */
454 			bp->b_flags = B_READ;
455 			bp->b_data = (caddr_t) kva + i * bsize;
456 			bp->b_bio1.bio_done = vnode_pager_iodone;
457 			bp->b_bio1.bio_offset = doffset;
458 			pbgetvp(dp, bp);
459 			bp->b_bcount = bsize;
460 			bp->b_bufsize = bsize;
461 			bp->b_runningbufspace = bp->b_bufsize;
462 			runningbufspace += bp->b_runningbufspace;
463 
464 			/* do the input */
465 			vn_strategy(dp, &bp->b_bio1);
466 
467 			/* we definitely need to be at splvm here */
468 
469 			crit_enter();
470 			while ((bp->b_flags & B_DONE) == 0) {
471 				tsleep(bp, 0, "vnsrd", 0);
472 			}
473 			crit_exit();
474 			if ((bp->b_flags & B_ERROR) != 0)
475 				error = EIO;
476 
477 			/*
478 			 * free the buffer header back to the swap buffer pool
479 			 */
480 			relpbuf(bp, &vnode_pbuf_freecnt);
481 			if (error)
482 				break;
483 
484 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
485 		} else {
486 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
487 			bzero((caddr_t) kva + i * bsize, bsize);
488 		}
489 	}
490 	sf_buf_free(sf);
491 	pmap_clear_modify(m);
492 	vm_page_flag_clear(m, PG_ZERO);
493 	if (error) {
494 		return VM_PAGER_ERROR;
495 	}
496 	return VM_PAGER_OK;
497 
498 }
499 
500 
501 /*
502  * old style vnode pager output routine
503  */
504 static int
505 vnode_pager_input_old(vm_object_t object, vm_page_t m)
506 {
507 	struct uio auio;
508 	struct iovec aiov;
509 	int error;
510 	int size;
511 	vm_offset_t kva;
512 	struct sf_buf *sf;
513 	struct vnode *vp;
514 
515 	error = 0;
516 	vp = object->handle;
517 
518 	/*
519 	 * Return failure if beyond current EOF
520 	 */
521 	if (IDX_TO_OFF(m->pindex) >= vp->v_filesize) {
522 		return VM_PAGER_BAD;
523 	} else {
524 		size = PAGE_SIZE;
525 		if (IDX_TO_OFF(m->pindex) + size > vp->v_filesize)
526 			size = vp->v_filesize - IDX_TO_OFF(m->pindex);
527 
528 		/*
529 		 * Allocate a kernel virtual address and initialize so that
530 		 * we can use VOP_READ/WRITE routines.
531 		 */
532 		sf = sf_buf_alloc(m, 0);
533 		kva = sf_buf_kva(sf);
534 
535 		aiov.iov_base = (caddr_t) kva;
536 		aiov.iov_len = size;
537 		auio.uio_iov = &aiov;
538 		auio.uio_iovcnt = 1;
539 		auio.uio_offset = IDX_TO_OFF(m->pindex);
540 		auio.uio_segflg = UIO_SYSSPACE;
541 		auio.uio_rw = UIO_READ;
542 		auio.uio_resid = size;
543 		auio.uio_td = curthread;
544 
545 		error = VOP_READ(((struct vnode *)object->handle),
546 				&auio, 0, proc0.p_ucred);
547 		if (!error) {
548 			int count = size - auio.uio_resid;
549 
550 			if (count == 0)
551 				error = EINVAL;
552 			else if (count != PAGE_SIZE)
553 				bzero((caddr_t) kva + count, PAGE_SIZE - count);
554 		}
555 		sf_buf_free(sf);
556 	}
557 	pmap_clear_modify(m);
558 	vm_page_undirty(m);
559 	vm_page_flag_clear(m, PG_ZERO);
560 	if (!error)
561 		m->valid = VM_PAGE_BITS_ALL;
562 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
563 }
564 
565 /*
566  * generic vnode pager input routine
567  */
568 
569 /*
570  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
571  * implement their own VOP_GETPAGES, their VOP_GETPAGES should call to
572  * vnode_pager_generic_getpages() to implement the previous behaviour.
573  *
574  * All other FS's should use the bypass to get to the local media
575  * backing vp's VOP_GETPAGES.
576  */
577 static int
578 vnode_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
579 {
580 	int rtval;
581 	struct vnode *vp;
582 	int bytes = count * PAGE_SIZE;
583 
584 	vp = object->handle;
585 	/*
586 	 * XXX temporary diagnostic message to help track stale FS code,
587 	 * Returning EOPNOTSUPP from here may make things unhappy.
588 	 */
589 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
590 	if (rtval == EOPNOTSUPP) {
591 	    printf("vnode_pager: *** WARNING *** stale FS getpages\n");
592 	    rtval = vnode_pager_generic_getpages( vp, m, bytes, reqpage);
593 	}
594 	return rtval;
595 }
596 
597 
598 /*
599  * This is now called from local media FS's to operate against their
600  * own vnodes if they fail to implement VOP_GETPAGES.
601  */
602 int
603 vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int bytecount,
604     int reqpage)
605 {
606 	vm_object_t object;
607 	vm_offset_t kva;
608 	off_t foff, tfoff, nextoff;
609 	int i, size, bsize, first;
610 	off_t firstaddr;
611 	struct vnode *dp;
612 	int runpg;
613 	int runend;
614 	struct buf *bp;
615 	int count;
616 	int error = 0;
617 
618 	object = vp->v_object;
619 	count = bytecount / PAGE_SIZE;
620 
621 	if (vp->v_mount == NULL)
622 		return VM_PAGER_BAD;
623 
624 	bsize = vp->v_mount->mnt_stat.f_iosize;
625 
626 	/* get the UNDERLYING device for the file with VOP_BMAP() */
627 
628 	/*
629 	 * originally, we did not check for an error return value -- assuming
630 	 * an fs always has a bmap entry point -- that assumption is wrong!!!
631 	 */
632 	foff = IDX_TO_OFF(m[reqpage]->pindex);
633 
634 	/*
635 	 * if we can't bmap, use old VOP code
636 	 */
637 	if (VOP_BMAP(vp, (off_t)0, &dp, NULL, NULL, NULL)) {
638 		for (i = 0; i < count; i++) {
639 			if (i != reqpage) {
640 				vnode_pager_freepage(m[i]);
641 			}
642 		}
643 		mycpu->gd_cnt.v_vnodein++;
644 		mycpu->gd_cnt.v_vnodepgsin++;
645 		return vnode_pager_input_old(object, m[reqpage]);
646 
647 		/*
648 		 * if the blocksize is smaller than a page size, then use
649 		 * special small filesystem code.  NFS sometimes has a small
650 		 * blocksize, but it can handle large reads itself.
651 		 */
652 	} else if ((PAGE_SIZE / bsize) > 1 &&
653 	    (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
654 		for (i = 0; i < count; i++) {
655 			if (i != reqpage) {
656 				vnode_pager_freepage(m[i]);
657 			}
658 		}
659 		mycpu->gd_cnt.v_vnodein++;
660 		mycpu->gd_cnt.v_vnodepgsin++;
661 		return vnode_pager_input_smlfs(object, m[reqpage]);
662 	}
663 
664 	/*
665 	 * If we have a completely valid page available to us, we can
666 	 * clean up and return.  Otherwise we have to re-read the
667 	 * media.
668 	 *
669 	 * Note that this does not work with NFS, so NFS has its own
670 	 * getpages routine.  The problem is that NFS can have partially
671 	 * valid pages associated with the buffer cache due to the piecemeal
672 	 * write support.  If we were to fall through and re-read the media
673 	 * as we do here, dirty data could be lost.
674 	 */
675 
676 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
677 		for (i = 0; i < count; i++) {
678 			if (i != reqpage)
679 				vnode_pager_freepage(m[i]);
680 		}
681 		return VM_PAGER_OK;
682 	}
683 	m[reqpage]->valid = 0;
684 
685 	/*
686 	 * here on direct device I/O
687 	 */
688 
689 	firstaddr = -1;
690 	/*
691 	 * calculate the run that includes the required page
692 	 */
693 	for(first = 0, i = 0; i < count; i = runend) {
694 		firstaddr = vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex),
695 					     &runpg);
696 		if (firstaddr == -1) {
697 			if (i == reqpage && foff < vp->v_filesize) {
698 				/* XXX no %qd in kernel. */
699 				panic("vnode_pager_getpages: unexpected missing page: firstaddr: %012llx, foff: 0x%012llx, v_filesize: 0x%012llx",
700 			   	 firstaddr, foff, vp->v_filesize);
701 			}
702 			vnode_pager_freepage(m[i]);
703 			runend = i + 1;
704 			first = runend;
705 			continue;
706 		}
707 		runend = i + runpg;
708 		if (runend <= reqpage) {
709 			int j;
710 			for (j = i; j < runend; j++) {
711 				vnode_pager_freepage(m[j]);
712 			}
713 		} else {
714 			if (runpg < (count - first)) {
715 				for (i = first + runpg; i < count; i++)
716 					vnode_pager_freepage(m[i]);
717 				count = first + runpg;
718 			}
719 			break;
720 		}
721 		first = runend;
722 	}
723 
724 	/*
725 	 * the first and last page have been calculated now, move input pages
726 	 * to be zero based...
727 	 */
728 	if (first != 0) {
729 		for (i = first; i < count; i++) {
730 			m[i - first] = m[i];
731 		}
732 		count -= first;
733 		reqpage -= first;
734 	}
735 
736 	/*
737 	 * calculate the file virtual address for the transfer
738 	 */
739 	foff = IDX_TO_OFF(m[0]->pindex);
740 
741 	/*
742 	 * calculate the size of the transfer
743 	 */
744 	size = count * PAGE_SIZE;
745 	if ((foff + size) > vp->v_filesize)
746 		size = vp->v_filesize - foff;
747 
748 	/*
749 	 * round up physical size for real devices.
750 	 */
751 	if (dp->v_type == VBLK || dp->v_type == VCHR) {
752 		int secmask = dp->v_rdev->si_bsize_phys - 1;
753 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
754 		size = (size + secmask) & ~secmask;
755 	}
756 
757 	bp = getpbuf(&vnode_pbuf_freecnt);
758 	kva = (vm_offset_t) bp->b_data;
759 
760 	/*
761 	 * and map the pages to be read into the kva
762 	 */
763 	pmap_qenter(kva, m, count);
764 
765 	/* build a minimal buffer header */
766 	bp->b_flags = B_READ;
767 	bp->b_bio1.bio_done = vnode_pager_iodone;
768 	bp->b_bio1.bio_offset = firstaddr;
769 	pbgetvp(dp, bp);
770 	bp->b_bcount = size;
771 	bp->b_bufsize = size;
772 	bp->b_runningbufspace = bp->b_bufsize;
773 	runningbufspace += bp->b_runningbufspace;
774 
775 	mycpu->gd_cnt.v_vnodein++;
776 	mycpu->gd_cnt.v_vnodepgsin += count;
777 
778 	/* do the input */
779 	vn_strategy(dp, &bp->b_bio1);
780 
781 	crit_enter();
782 	/* we definitely need to be at splvm here */
783 
784 	while ((bp->b_flags & B_DONE) == 0) {
785 		tsleep(bp, 0, "vnread", 0);
786 	}
787 	crit_exit();
788 	if ((bp->b_flags & B_ERROR) != 0)
789 		error = EIO;
790 
791 	if (!error) {
792 		if (size != count * PAGE_SIZE)
793 			bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
794 	}
795 	pmap_qremove(kva, count);
796 
797 	/*
798 	 * free the buffer header back to the swap buffer pool
799 	 */
800 	relpbuf(bp, &vnode_pbuf_freecnt);
801 
802 	for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
803 		vm_page_t mt;
804 
805 		nextoff = tfoff + PAGE_SIZE;
806 		mt = m[i];
807 
808 		if (nextoff <= vp->v_filesize) {
809 			/*
810 			 * Read filled up entire page.
811 			 */
812 			mt->valid = VM_PAGE_BITS_ALL;
813 			vm_page_undirty(mt);	/* should be an assert? XXX */
814 			pmap_clear_modify(mt);
815 		} else {
816 			/*
817 			 * Read did not fill up entire page.  Since this
818 			 * is getpages, the page may be mapped, so we have
819 			 * to zero the invalid portions of the page even
820 			 * though we aren't setting them valid.
821 			 *
822 			 * Currently we do not set the entire page valid,
823 			 * we just try to clear the piece that we couldn't
824 			 * read.
825 			 */
826 			vm_page_set_validclean(mt, 0, vp->v_filesize - tfoff);
827 			/* handled by vm_fault now */
828 			/* vm_page_zero_invalid(mt, FALSE); */
829 		}
830 
831 		vm_page_flag_clear(mt, PG_ZERO);
832 		if (i != reqpage) {
833 
834 			/*
835 			 * whether or not to leave the page activated is up in
836 			 * the air, but we should put the page on a page queue
837 			 * somewhere. (it already is in the object). Result:
838 			 * It appears that empirical results show that
839 			 * deactivating pages is best.
840 			 */
841 
842 			/*
843 			 * just in case someone was asking for this page we
844 			 * now tell them that it is ok to use
845 			 */
846 			if (!error) {
847 				if (mt->flags & PG_WANTED)
848 					vm_page_activate(mt);
849 				else
850 					vm_page_deactivate(mt);
851 				vm_page_wakeup(mt);
852 			} else {
853 				vnode_pager_freepage(mt);
854 			}
855 		}
856 	}
857 	if (error) {
858 		printf("vnode_pager_getpages: I/O read error\n");
859 	}
860 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
861 }
862 
863 /*
864  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
865  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
866  * vnode_pager_generic_putpages() to implement the previous behaviour.
867  *
868  * All other FS's should use the bypass to get to the local media
869  * backing vp's VOP_PUTPAGES.
870  */
871 static void
872 vnode_pager_putpages(vm_object_t object, vm_page_t *m, int count,
873     boolean_t sync, int *rtvals)
874 {
875 	int rtval;
876 	struct vnode *vp;
877 	int bytes = count * PAGE_SIZE;
878 
879 	/*
880 	 * Force synchronous operation if we are extremely low on memory
881 	 * to prevent a low-memory deadlock.  VOP operations often need to
882 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
883 	 * operation ).  The swapper handles the case by limiting the amount
884 	 * of asynchronous I/O, but that sort of solution doesn't scale well
885 	 * for the vnode pager without a lot of work.
886 	 *
887 	 * Also, the backing vnode's iodone routine may not wake the pageout
888 	 * daemon up.  This should be probably be addressed XXX.
889 	 */
890 
891 	if ((vmstats.v_free_count + vmstats.v_cache_count) < vmstats.v_pageout_free_min)
892 		sync |= OBJPC_SYNC;
893 
894 	/*
895 	 * Call device-specific putpages function
896 	 */
897 
898 	vp = object->handle;
899 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
900 	if (rtval == EOPNOTSUPP) {
901 	    printf("vnode_pager: *** WARNING *** stale FS putpages\n");
902 	    rtval = vnode_pager_generic_putpages( vp, m, bytes, sync, rtvals);
903 	}
904 }
905 
906 
907 /*
908  * This is now called from local media FS's to operate against their
909  * own vnodes if they fail to implement VOP_PUTPAGES.
910  *
911  * This is typically called indirectly via the pageout daemon and
912  * clustering has already typically occured, so in general we ask the
913  * underlying filesystem to write the data out asynchronously rather
914  * then delayed.
915  */
916 int
917 vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int bytecount,
918     int flags, int *rtvals)
919 {
920 	int i;
921 	vm_object_t object;
922 	int count;
923 
924 	int maxsize, ncount;
925 	vm_ooffset_t poffset;
926 	struct uio auio;
927 	struct iovec aiov;
928 	int error;
929 	int ioflags;
930 
931 	object = vp->v_object;
932 	count = bytecount / PAGE_SIZE;
933 
934 	for (i = 0; i < count; i++)
935 		rtvals[i] = VM_PAGER_AGAIN;
936 
937 	if ((int) m[0]->pindex < 0) {
938 		printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
939 			(long)m[0]->pindex, m[0]->dirty);
940 		rtvals[0] = VM_PAGER_BAD;
941 		return VM_PAGER_BAD;
942 	}
943 
944 	maxsize = count * PAGE_SIZE;
945 	ncount = count;
946 
947 	poffset = IDX_TO_OFF(m[0]->pindex);
948 
949 	/*
950 	 * If the page-aligned write is larger then the actual file we
951 	 * have to invalidate pages occuring beyond the file EOF.  However,
952 	 * there is an edge case where a file may not be page-aligned where
953 	 * the last page is partially invalid.  In this case the filesystem
954 	 * may not properly clear the dirty bits for the entire page (which
955 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
956 	 * With the page locked we are free to fix-up the dirty bits here.
957 	 *
958 	 * We do not under any circumstances truncate the valid bits, as
959 	 * this will screw up bogus page replacement.
960 	 */
961 	if (maxsize + poffset > vp->v_filesize) {
962 		if (vp->v_filesize > poffset) {
963 			int pgoff;
964 
965 			maxsize = vp->v_filesize - poffset;
966 			ncount = btoc(maxsize);
967 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
968 				vm_page_clear_dirty(m[ncount - 1], pgoff,
969 					PAGE_SIZE - pgoff);
970 			}
971 		} else {
972 			maxsize = 0;
973 			ncount = 0;
974 		}
975 		if (ncount < count) {
976 			for (i = ncount; i < count; i++) {
977 				rtvals[i] = VM_PAGER_BAD;
978 			}
979 		}
980 	}
981 
982 	/*
983 	 * pageouts are already clustered, use IO_ASYNC to force a bawrite()
984 	 * rather then a bdwrite() to prevent paging I/O from saturating
985 	 * the buffer cache.  Dummy-up the sequential heuristic to cause
986 	 * large ranges to cluster.  If neither IO_SYNC or IO_ASYNC is set,
987 	 * the system decides how to cluster.
988 	 */
989 	ioflags = IO_VMIO;
990 	if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL))
991 		ioflags |= IO_SYNC;
992 	else if ((flags & VM_PAGER_CLUSTER_OK) == 0)
993 		ioflags |= IO_ASYNC;
994 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
995 	ioflags |= IO_SEQMAX << IO_SEQSHIFT;
996 
997 	aiov.iov_base = (caddr_t) 0;
998 	aiov.iov_len = maxsize;
999 	auio.uio_iov = &aiov;
1000 	auio.uio_iovcnt = 1;
1001 	auio.uio_offset = poffset;
1002 	auio.uio_segflg = UIO_NOCOPY;
1003 	auio.uio_rw = UIO_WRITE;
1004 	auio.uio_resid = maxsize;
1005 	auio.uio_td = NULL;
1006 	error = VOP_WRITE(vp, &auio, ioflags, proc0.p_ucred);
1007 	mycpu->gd_cnt.v_vnodeout++;
1008 	mycpu->gd_cnt.v_vnodepgsout += ncount;
1009 
1010 	if (error) {
1011 		printf("vnode_pager_putpages: I/O error %d\n", error);
1012 	}
1013 	if (auio.uio_resid) {
1014 		printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1015 		    auio.uio_resid, (u_long)m[0]->pindex);
1016 	}
1017 	for (i = 0; i < ncount; i++) {
1018 		rtvals[i] = VM_PAGER_OK;
1019 	}
1020 	return rtvals[0];
1021 }
1022 
1023 struct vnode *
1024 vnode_pager_lock(vm_object_t object)
1025 {
1026 	struct thread *td = curthread;	/* XXX */
1027 	int error;
1028 
1029 	for (; object != NULL; object = object->backing_object) {
1030 		if (object->type != OBJT_VNODE)
1031 			continue;
1032 		if (object->flags & OBJ_DEAD)
1033 			return NULL;
1034 
1035 		for (;;) {
1036 			struct vnode *vp = object->handle;
1037 			error = vget(vp, LK_NOPAUSE | LK_SHARED |
1038 					 LK_RETRY | LK_CANRECURSE, td);
1039 			if (error == 0) {
1040 				if (object->handle != vp) {
1041 					vput(vp);
1042 					continue;
1043 				}
1044 				return (vp);
1045 			}
1046 			if ((object->flags & OBJ_DEAD) ||
1047 			    (object->type != OBJT_VNODE)) {
1048 				return NULL;
1049 			}
1050 			printf("vnode_pager_lock: vp %p error %d lockstatus %d, retrying\n", vp, error, lockstatus(&vp->v_lock, td));
1051 			tsleep(object->handle, 0, "vnpgrl", hz);
1052 		}
1053 	}
1054 	return NULL;
1055 }
1056