xref: /freebsd/sys/vm/vnode_pager.c (revision 00a91cb7f41bebeca9770e0a76892559f504ec2e)
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  * $FreeBSD$
42  */
43 
44 /*
45  * Page to/from files (vnodes).
46  */
47 
48 /*
49  * TODO:
50  *	Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will
51  *	greatly re-simplify the vnode_pager.
52  */
53 
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/proc.h>
57 #include <sys/vnode.h>
58 #include <sys/mount.h>
59 #include <sys/bio.h>
60 #include <sys/buf.h>
61 #include <sys/vmmeter.h>
62 #include <sys/conf.h>
63 #include <sys/stdint.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_map.h>
70 #include <vm/vnode_pager.h>
71 #include <vm/vm_extern.h>
72 
73 static void vnode_pager_init(void);
74 static vm_offset_t vnode_pager_addr(struct vnode *vp, vm_ooffset_t address,
75 					 int *run);
76 static void vnode_pager_iodone(struct buf *bp);
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 	vnode_pager_init,
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;
95 
96 static void
97 vnode_pager_init(void)
98 {
99 
100 	vnode_pbuf_freecnt = nswbuf / 2 + 1;
101 }
102 
103 /*
104  * Allocate (or lookup) pager for a vnode.
105  * Handle is a vnode pointer.
106  *
107  * MPSAFE
108  */
109 vm_object_t
110 vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
111 		  vm_ooffset_t offset)
112 {
113 	vm_object_t object;
114 	struct vnode *vp;
115 
116 	/*
117 	 * Pageout to vnode, no can do yet.
118 	 */
119 	if (handle == NULL)
120 		return (NULL);
121 
122 	vp = (struct vnode *) handle;
123 
124 	ASSERT_VOP_LOCKED(vp, "vnode_pager_alloc");
125 
126 	mtx_lock(&Giant);
127 	/*
128 	 * Prevent race condition when allocating the object. This
129 	 * can happen with NFS vnodes since the nfsnode isn't locked.
130 	 */
131 	VI_LOCK(vp);
132 	while (vp->v_iflag & VI_OLOCK) {
133 		vp->v_iflag |= VI_OWANT;
134 		msleep(vp, VI_MTX(vp), PVM, "vnpobj", 0);
135 	}
136 	vp->v_iflag |= VI_OLOCK;
137 	VI_UNLOCK(vp);
138 
139 	/*
140 	 * If the object is being terminated, wait for it to
141 	 * go away.
142 	 */
143 	while (((object = vp->v_object) != NULL) &&
144 		(object->flags & OBJ_DEAD)) {
145 		tsleep(object, PVM, "vadead", 0);
146 	}
147 
148 	if (vp->v_usecount == 0)
149 		panic("vnode_pager_alloc: no vnode reference");
150 
151 	if (object == NULL) {
152 		/*
153 		 * And an object of the appropriate size
154 		 */
155 		object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size)));
156 
157 		object->un_pager.vnp.vnp_size = size;
158 
159 		object->handle = handle;
160 		vp->v_object = object;
161 	} else {
162 		object->ref_count++;
163 	}
164 	VI_LOCK(vp);
165 	vp->v_usecount++;
166 	vp->v_iflag &= ~VI_OLOCK;
167 	if (vp->v_iflag & VI_OWANT) {
168 		vp->v_iflag &= ~VI_OWANT;
169 		wakeup(vp);
170 	}
171 	VI_UNLOCK(vp);
172 	mtx_unlock(&Giant);
173 	return (object);
174 }
175 
176 static void
177 vnode_pager_dealloc(object)
178 	vm_object_t object;
179 {
180 	struct vnode *vp = object->handle;
181 
182 	GIANT_REQUIRED;
183 	if (vp == NULL)
184 		panic("vnode_pager_dealloc: pager already dealloced");
185 
186 	vm_object_pip_wait(object, "vnpdea");
187 
188 	object->handle = NULL;
189 	object->type = OBJT_DEAD;
190 	ASSERT_VOP_LOCKED(vp, "vnode_pager_dealloc");
191 	vp->v_object = NULL;
192 	vp->v_vflag &= ~(VV_TEXT | VV_OBJBUF);
193 }
194 
195 static boolean_t
196 vnode_pager_haspage(object, pindex, before, after)
197 	vm_object_t object;
198 	vm_pindex_t pindex;
199 	int *before;
200 	int *after;
201 {
202 	struct vnode *vp = object->handle;
203 	daddr_t bn;
204 	int err;
205 	daddr_t reqblock;
206 	int poff;
207 	int bsize;
208 	int pagesperblock, blocksperpage;
209 
210 	GIANT_REQUIRED;
211 	/*
212 	 * If no vp or vp is doomed or marked transparent to VM, we do not
213 	 * have the page.
214 	 */
215 	if (vp == NULL)
216 		return FALSE;
217 
218 	VI_LOCK(vp);
219 	if (vp->v_iflag & VI_DOOMED) {
220 		VI_UNLOCK(vp);
221 		return FALSE;
222 	}
223 	VI_UNLOCK(vp);
224 	/*
225 	 * If filesystem no longer mounted or offset beyond end of file we do
226 	 * not have the page.
227 	 */
228 	if ((vp->v_mount == NULL) ||
229 	    (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size))
230 		return FALSE;
231 
232 	bsize = vp->v_mount->mnt_stat.f_iosize;
233 	pagesperblock = bsize / PAGE_SIZE;
234 	blocksperpage = 0;
235 	if (pagesperblock > 0) {
236 		reqblock = pindex / pagesperblock;
237 	} else {
238 		blocksperpage = (PAGE_SIZE / bsize);
239 		reqblock = pindex * blocksperpage;
240 	}
241 	err = VOP_BMAP(vp, reqblock, (struct vnode **) 0, &bn,
242 		after, before);
243 	if (err)
244 		return TRUE;
245 	if (bn == -1)
246 		return FALSE;
247 	if (pagesperblock > 0) {
248 		poff = pindex - (reqblock * pagesperblock);
249 		if (before) {
250 			*before *= pagesperblock;
251 			*before += poff;
252 		}
253 		if (after) {
254 			int numafter;
255 			*after *= pagesperblock;
256 			numafter = pagesperblock - (poff + 1);
257 			if (IDX_TO_OFF(pindex + numafter) >
258 			    object->un_pager.vnp.vnp_size) {
259 				numafter =
260 		    		    OFF_TO_IDX(object->un_pager.vnp.vnp_size) -
261 				    pindex;
262 			}
263 			*after += numafter;
264 		}
265 	} else {
266 		if (before) {
267 			*before /= blocksperpage;
268 		}
269 
270 		if (after) {
271 			*after /= blocksperpage;
272 		}
273 	}
274 	return TRUE;
275 }
276 
277 /*
278  * Lets the VM system know about a change in size for a file.
279  * We adjust our own internal size and flush any cached pages in
280  * the associated object that are affected by the size change.
281  *
282  * Note: this routine may be invoked as a result of a pager put
283  * operation (possibly at object termination time), so we must be careful.
284  */
285 void
286 vnode_pager_setsize(vp, nsize)
287 	struct vnode *vp;
288 	vm_ooffset_t nsize;
289 {
290 	vm_pindex_t nobjsize;
291 	vm_object_t object = vp->v_object;
292 
293 	GIANT_REQUIRED;
294 
295 	if (object == NULL)
296 		return;
297 
298 	/*
299 	 * Hasn't changed size
300 	 */
301 	if (nsize == object->un_pager.vnp.vnp_size)
302 		return;
303 
304 	nobjsize = OFF_TO_IDX(nsize + PAGE_MASK);
305 
306 	/*
307 	 * File has shrunk. Toss any cached pages beyond the new EOF.
308 	 */
309 	if (nsize < object->un_pager.vnp.vnp_size) {
310 #ifdef ENABLE_VFS_IOOPT
311 		vm_freeze_copyopts(object, OFF_TO_IDX(nsize), object->size);
312 #endif
313 		if (nobjsize < object->size) {
314 			vm_object_page_remove(object, nobjsize, object->size,
315 				FALSE);
316 		}
317 		/*
318 		 * this gets rid of garbage at the end of a page that is now
319 		 * only partially backed by the vnode.
320 		 *
321 		 * XXX for some reason (I don't know yet), if we take a
322 		 * completely invalid page and mark it partially valid
323 		 * it can screw up NFS reads, so we don't allow the case.
324 		 */
325 		if (nsize & PAGE_MASK) {
326 			vm_page_t m;
327 
328 			m = vm_page_lookup(object, OFF_TO_IDX(nsize));
329 			if (m && m->valid) {
330 				int base = (int)nsize & PAGE_MASK;
331 				int size = PAGE_SIZE - base;
332 
333 				/*
334 				 * Clear out partial-page garbage in case
335 				 * the page has been mapped.
336 				 */
337 				pmap_zero_page_area(m, base, size);
338 
339 				/*
340 				 * XXX work around SMP data integrity race
341 				 * by unmapping the page from user processes.
342 				 * The garbage we just cleared may be mapped
343 				 * to a user process running on another cpu
344 				 * and this code is not running through normal
345 				 * I/O channels which handle SMP issues for
346 				 * us, so unmap page to synchronize all cpus.
347 				 *
348 				 * XXX should vm_pager_unmap_page() have
349 				 * dealt with this?
350 				 */
351 				pmap_page_protect(m, VM_PROT_NONE);
352 
353 				/*
354 				 * Clear out partial-page dirty bits.  This
355 				 * has the side effect of setting the valid
356 				 * bits, but that is ok.  There are a bunch
357 				 * of places in the VM system where we expected
358 				 * m->dirty == VM_PAGE_BITS_ALL.  The file EOF
359 				 * case is one of them.  If the page is still
360 				 * partially dirty, make it fully dirty.
361 				 *
362 				 * note that we do not clear out the valid
363 				 * bits.  This would prevent bogus_page
364 				 * replacement from working properly.
365 				 */
366 				vm_page_set_validclean(m, base, size);
367 				if (m->dirty != 0)
368 					m->dirty = VM_PAGE_BITS_ALL;
369 			}
370 		}
371 	}
372 	object->un_pager.vnp.vnp_size = nsize;
373 	object->size = nobjsize;
374 }
375 
376 /*
377  * calculate the linear (byte) disk address of specified virtual
378  * file address
379  */
380 static vm_offset_t
381 vnode_pager_addr(vp, address, run)
382 	struct vnode *vp;
383 	vm_ooffset_t address;
384 	int *run;
385 {
386 	int rtaddress;
387 	int bsize;
388 	daddr_t block;
389 	struct vnode *rtvp;
390 	int err;
391 	daddr_t vblock;
392 	int voffset;
393 
394 	GIANT_REQUIRED;
395 	if ((int) address < 0)
396 		return -1;
397 
398 	if (vp->v_mount == NULL)
399 		return -1;
400 
401 	bsize = vp->v_mount->mnt_stat.f_iosize;
402 	vblock = address / bsize;
403 	voffset = address % bsize;
404 
405 	err = VOP_BMAP(vp, vblock, &rtvp, &block, run, NULL);
406 
407 	if (err || (block == -1))
408 		rtaddress = -1;
409 	else {
410 		rtaddress = block + voffset / DEV_BSIZE;
411 		if (run) {
412 			*run += 1;
413 			*run *= bsize/PAGE_SIZE;
414 			*run -= voffset/PAGE_SIZE;
415 		}
416 	}
417 
418 	return rtaddress;
419 }
420 
421 /*
422  * interrupt routine for I/O completion
423  */
424 static void
425 vnode_pager_iodone(bp)
426 	struct buf *bp;
427 {
428 	bp->b_flags |= B_DONE;
429 	wakeup(bp);
430 }
431 
432 /*
433  * small block filesystem vnode pager input
434  */
435 static int
436 vnode_pager_input_smlfs(object, m)
437 	vm_object_t object;
438 	vm_page_t m;
439 {
440 	int i;
441 	int s;
442 	struct vnode *dp, *vp;
443 	struct buf *bp;
444 	vm_offset_t kva;
445 	int fileaddr;
446 	vm_offset_t bsize;
447 	int error = 0;
448 
449 	GIANT_REQUIRED;
450 
451 	vp = object->handle;
452 	if (vp->v_mount == NULL)
453 		return VM_PAGER_BAD;
454 
455 	bsize = vp->v_mount->mnt_stat.f_iosize;
456 
457 	VOP_BMAP(vp, 0, &dp, 0, NULL, NULL);
458 
459 	kva = vm_pager_map_page(m);
460 
461 	for (i = 0; i < PAGE_SIZE / bsize; i++) {
462 		vm_ooffset_t address;
463 
464 		if (vm_page_bits(i * bsize, bsize) & m->valid)
465 			continue;
466 
467 		address = IDX_TO_OFF(m->pindex) + i * bsize;
468 		if (address >= object->un_pager.vnp.vnp_size) {
469 			fileaddr = -1;
470 		} else {
471 			fileaddr = vnode_pager_addr(vp, address, NULL);
472 		}
473 		if (fileaddr != -1) {
474 			bp = getpbuf(&vnode_pbuf_freecnt);
475 
476 			/* build a minimal buffer header */
477 			bp->b_iocmd = BIO_READ;
478 			bp->b_iodone = vnode_pager_iodone;
479 			KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
480 			KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
481 			bp->b_rcred = crhold(curthread->td_ucred);
482 			bp->b_wcred = crhold(curthread->td_ucred);
483 			bp->b_data = (caddr_t) kva + i * bsize;
484 			bp->b_blkno = fileaddr;
485 			pbgetvp(dp, bp);
486 			bp->b_bcount = bsize;
487 			bp->b_bufsize = bsize;
488 			bp->b_runningbufspace = bp->b_bufsize;
489 			runningbufspace += bp->b_runningbufspace;
490 
491 			/* do the input */
492 			BUF_STRATEGY(bp);
493 
494 			/* we definitely need to be at splvm here */
495 
496 			s = splvm();
497 			while ((bp->b_flags & B_DONE) == 0) {
498 				tsleep(bp, PVM, "vnsrd", 0);
499 			}
500 			splx(s);
501 			if ((bp->b_ioflags & BIO_ERROR) != 0)
502 				error = EIO;
503 
504 			/*
505 			 * free the buffer header back to the swap buffer pool
506 			 */
507 			relpbuf(bp, &vnode_pbuf_freecnt);
508 			if (error)
509 				break;
510 
511 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
512 		} else {
513 			vm_page_set_validclean(m, (i * bsize) & PAGE_MASK, bsize);
514 			bzero((caddr_t) kva + i * bsize, bsize);
515 		}
516 	}
517 	vm_pager_unmap_page(kva);
518 	pmap_clear_modify(m);
519 	vm_page_flag_clear(m, PG_ZERO);
520 	if (error) {
521 		return VM_PAGER_ERROR;
522 	}
523 	return VM_PAGER_OK;
524 
525 }
526 
527 
528 /*
529  * old style vnode pager output routine
530  */
531 static int
532 vnode_pager_input_old(object, m)
533 	vm_object_t object;
534 	vm_page_t m;
535 {
536 	struct uio auio;
537 	struct iovec aiov;
538 	int error;
539 	int size;
540 	vm_offset_t kva;
541 	struct vnode *vp;
542 
543 	GIANT_REQUIRED;
544 	error = 0;
545 
546 	/*
547 	 * Return failure if beyond current EOF
548 	 */
549 	if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) {
550 		return VM_PAGER_BAD;
551 	} else {
552 		size = PAGE_SIZE;
553 		if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size)
554 			size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex);
555 
556 		/*
557 		 * Allocate a kernel virtual address and initialize so that
558 		 * we can use VOP_READ/WRITE routines.
559 		 */
560 		kva = vm_pager_map_page(m);
561 
562 		vp = object->handle;
563 		aiov.iov_base = (caddr_t) kva;
564 		aiov.iov_len = size;
565 		auio.uio_iov = &aiov;
566 		auio.uio_iovcnt = 1;
567 		auio.uio_offset = IDX_TO_OFF(m->pindex);
568 		auio.uio_segflg = UIO_SYSSPACE;
569 		auio.uio_rw = UIO_READ;
570 		auio.uio_resid = size;
571 		auio.uio_td = curthread;
572 
573 		error = VOP_READ(vp, &auio, 0, curthread->td_ucred);
574 		if (!error) {
575 			int count = size - auio.uio_resid;
576 
577 			if (count == 0)
578 				error = EINVAL;
579 			else if (count != PAGE_SIZE)
580 				bzero((caddr_t) kva + count, PAGE_SIZE - count);
581 		}
582 		vm_pager_unmap_page(kva);
583 	}
584 	pmap_clear_modify(m);
585 	vm_page_undirty(m);
586 	vm_page_flag_clear(m, PG_ZERO);
587 	if (!error)
588 		m->valid = VM_PAGE_BITS_ALL;
589 	return error ? VM_PAGER_ERROR : VM_PAGER_OK;
590 }
591 
592 /*
593  * generic vnode pager input routine
594  */
595 
596 /*
597  * Local media VFS's that do not implement their own VOP_GETPAGES
598  * should have their VOP_GETPAGES call to vnode_pager_generic_getpages()
599  * to implement the previous behaviour.
600  *
601  * All other FS's should use the bypass to get to the local media
602  * backing vp's VOP_GETPAGES.
603  */
604 static int
605 vnode_pager_getpages(object, m, count, reqpage)
606 	vm_object_t object;
607 	vm_page_t *m;
608 	int count;
609 	int reqpage;
610 {
611 	int rtval;
612 	struct vnode *vp;
613 	int bytes = count * PAGE_SIZE;
614 
615 	GIANT_REQUIRED;
616 	vp = object->handle;
617 	rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0);
618 	KASSERT(rtval != EOPNOTSUPP,
619 	    ("vnode_pager: FS getpages not implemented\n"));
620 	return rtval;
621 }
622 
623 /*
624  * This is now called from local media FS's to operate against their
625  * own vnodes if they fail to implement VOP_GETPAGES.
626  */
627 int
628 vnode_pager_generic_getpages(vp, m, bytecount, reqpage)
629 	struct vnode *vp;
630 	vm_page_t *m;
631 	int bytecount;
632 	int reqpage;
633 {
634 	vm_object_t object;
635 	vm_offset_t kva;
636 	off_t foff, tfoff, nextoff;
637 	int i, j, size, bsize, first, firstaddr;
638 	struct vnode *dp;
639 	int runpg;
640 	int runend;
641 	struct buf *bp;
642 	int s;
643 	int count;
644 	int error = 0;
645 
646 	GIANT_REQUIRED;
647 	object = vp->v_object;
648 	count = bytecount / PAGE_SIZE;
649 
650 	if (vp->v_mount == NULL)
651 		return VM_PAGER_BAD;
652 
653 	bsize = vp->v_mount->mnt_stat.f_iosize;
654 
655 	/* get the UNDERLYING device for the file with VOP_BMAP() */
656 
657 	/*
658 	 * originally, we did not check for an error return value -- assuming
659 	 * an fs always has a bmap entry point -- that assumption is wrong!!!
660 	 */
661 	foff = IDX_TO_OFF(m[reqpage]->pindex);
662 
663 	/*
664 	 * if we can't bmap, use old VOP code
665 	 */
666 	if (VOP_BMAP(vp, 0, &dp, 0, NULL, NULL)) {
667 		vm_page_lock_queues();
668 		for (i = 0; i < count; i++)
669 			if (i != reqpage)
670 				vm_page_free(m[i]);
671 		vm_page_unlock_queues();
672 		cnt.v_vnodein++;
673 		cnt.v_vnodepgsin++;
674 		return vnode_pager_input_old(object, m[reqpage]);
675 
676 		/*
677 		 * if the blocksize is smaller than a page size, then use
678 		 * special small filesystem code.  NFS sometimes has a small
679 		 * blocksize, but it can handle large reads itself.
680 		 */
681 	} else if ((PAGE_SIZE / bsize) > 1 &&
682 	    (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) {
683 		vm_page_lock_queues();
684 		for (i = 0; i < count; i++)
685 			if (i != reqpage)
686 				vm_page_free(m[i]);
687 		vm_page_unlock_queues();
688 		cnt.v_vnodein++;
689 		cnt.v_vnodepgsin++;
690 		return vnode_pager_input_smlfs(object, m[reqpage]);
691 	}
692 
693 	/*
694 	 * If we have a completely valid page available to us, we can
695 	 * clean up and return.  Otherwise we have to re-read the
696 	 * media.
697 	 */
698 	if (m[reqpage]->valid == VM_PAGE_BITS_ALL) {
699 		vm_page_lock_queues();
700 		for (i = 0; i < count; i++)
701 			if (i != reqpage)
702 				vm_page_free(m[i]);
703 		vm_page_unlock_queues();
704 		return VM_PAGER_OK;
705 	}
706 	m[reqpage]->valid = 0;
707 
708 	/*
709 	 * here on direct device I/O
710 	 */
711 	firstaddr = -1;
712 
713 	/*
714 	 * calculate the run that includes the required page
715 	 */
716 	for (first = 0, i = 0; i < count; i = runend) {
717 		firstaddr = vnode_pager_addr(vp,
718 			IDX_TO_OFF(m[i]->pindex), &runpg);
719 		if (firstaddr == -1) {
720 			if (i == reqpage && foff < object->un_pager.vnp.vnp_size) {
721 				panic("vnode_pager_getpages: unexpected missing page: firstaddr: %d, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx",
722 				    firstaddr, (uintmax_t)(foff >> 32),
723 				    (uintmax_t)foff,
724 				    (uintmax_t)
725 				    (object->un_pager.vnp.vnp_size >> 32),
726 				    (uintmax_t)object->un_pager.vnp.vnp_size);
727 			}
728 			vm_page_lock_queues();
729 			vm_page_free(m[i]);
730 			vm_page_unlock_queues();
731 			runend = i + 1;
732 			first = runend;
733 			continue;
734 		}
735 		runend = i + runpg;
736 		if (runend <= reqpage) {
737 			vm_page_lock_queues();
738 			for (j = i; j < runend; j++)
739 				vm_page_free(m[j]);
740 			vm_page_unlock_queues();
741 		} else {
742 			if (runpg < (count - first)) {
743 				vm_page_lock_queues();
744 				for (i = first + runpg; i < count; i++)
745 					vm_page_free(m[i]);
746 				vm_page_unlock_queues();
747 				count = first + runpg;
748 			}
749 			break;
750 		}
751 		first = runend;
752 	}
753 
754 	/*
755 	 * the first and last page have been calculated now, move input pages
756 	 * to be zero based...
757 	 */
758 	if (first != 0) {
759 		for (i = first; i < count; i++) {
760 			m[i - first] = m[i];
761 		}
762 		count -= first;
763 		reqpage -= first;
764 	}
765 
766 	/*
767 	 * calculate the file virtual address for the transfer
768 	 */
769 	foff = IDX_TO_OFF(m[0]->pindex);
770 
771 	/*
772 	 * calculate the size of the transfer
773 	 */
774 	size = count * PAGE_SIZE;
775 	if ((foff + size) > object->un_pager.vnp.vnp_size)
776 		size = object->un_pager.vnp.vnp_size - foff;
777 
778 	/*
779 	 * round up physical size for real devices.
780 	 */
781 	if (dp->v_type == VBLK || dp->v_type == VCHR) {
782 		int secmask = dp->v_rdev->si_bsize_phys - 1;
783 		KASSERT(secmask < PAGE_SIZE, ("vnode_pager_generic_getpages: sector size %d too large\n", secmask + 1));
784 		size = (size + secmask) & ~secmask;
785 	}
786 
787 	bp = getpbuf(&vnode_pbuf_freecnt);
788 	kva = (vm_offset_t) bp->b_data;
789 
790 	/*
791 	 * and map the pages to be read into the kva
792 	 */
793 	pmap_qenter(kva, m, count);
794 
795 	/* build a minimal buffer header */
796 	bp->b_iocmd = BIO_READ;
797 	bp->b_iodone = vnode_pager_iodone;
798 	/* B_PHYS is not set, but it is nice to fill this in */
799 	KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred"));
800 	KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred"));
801 	bp->b_rcred = crhold(curthread->td_ucred);
802 	bp->b_wcred = crhold(curthread->td_ucred);
803 	bp->b_blkno = firstaddr;
804 	pbgetvp(dp, bp);
805 	bp->b_bcount = size;
806 	bp->b_bufsize = size;
807 	bp->b_runningbufspace = bp->b_bufsize;
808 	runningbufspace += bp->b_runningbufspace;
809 
810 	cnt.v_vnodein++;
811 	cnt.v_vnodepgsin += count;
812 
813 	/* do the input */
814 	BUF_STRATEGY(bp);
815 
816 	s = splvm();
817 	/* we definitely need to be at splvm here */
818 
819 	while ((bp->b_flags & B_DONE) == 0) {
820 		tsleep(bp, PVM, "vnread", 0);
821 	}
822 	splx(s);
823 	if ((bp->b_ioflags & BIO_ERROR) != 0)
824 		error = EIO;
825 
826 	if (!error) {
827 		if (size != count * PAGE_SIZE)
828 			bzero((caddr_t) kva + size, PAGE_SIZE * count - size);
829 	}
830 	pmap_qremove(kva, count);
831 
832 	/*
833 	 * free the buffer header back to the swap buffer pool
834 	 */
835 	relpbuf(bp, &vnode_pbuf_freecnt);
836 
837 	vm_page_lock_queues();
838 	for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) {
839 		vm_page_t mt;
840 
841 		nextoff = tfoff + PAGE_SIZE;
842 		mt = m[i];
843 
844 		if (nextoff <= object->un_pager.vnp.vnp_size) {
845 			/*
846 			 * Read filled up entire page.
847 			 */
848 			mt->valid = VM_PAGE_BITS_ALL;
849 			vm_page_undirty(mt);	/* should be an assert? XXX */
850 			pmap_clear_modify(mt);
851 		} else {
852 			/*
853 			 * Read did not fill up entire page.  Since this
854 			 * is getpages, the page may be mapped, so we have
855 			 * to zero the invalid portions of the page even
856 			 * though we aren't setting them valid.
857 			 *
858 			 * Currently we do not set the entire page valid,
859 			 * we just try to clear the piece that we couldn't
860 			 * read.
861 			 */
862 			vm_page_set_validclean(mt, 0,
863 			    object->un_pager.vnp.vnp_size - tfoff);
864 			/* handled by vm_fault now */
865 			/* vm_page_zero_invalid(mt, FALSE); */
866 		}
867 
868 		vm_page_flag_clear(mt, PG_ZERO);
869 		if (i != reqpage) {
870 
871 			/*
872 			 * whether or not to leave the page activated is up in
873 			 * the air, but we should put the page on a page queue
874 			 * somewhere. (it already is in the object). Result:
875 			 * It appears that empirical results show that
876 			 * deactivating pages is best.
877 			 */
878 
879 			/*
880 			 * just in case someone was asking for this page we
881 			 * now tell them that it is ok to use
882 			 */
883 			if (!error) {
884 				if (mt->flags & PG_WANTED)
885 					vm_page_activate(mt);
886 				else
887 					vm_page_deactivate(mt);
888 				vm_page_wakeup(mt);
889 			} else {
890 				vm_page_free(mt);
891 			}
892 		}
893 	}
894 	vm_page_unlock_queues();
895 	if (error) {
896 		printf("vnode_pager_getpages: I/O read error\n");
897 	}
898 	return (error ? VM_PAGER_ERROR : VM_PAGER_OK);
899 }
900 
901 /*
902  * EOPNOTSUPP is no longer legal.  For local media VFS's that do not
903  * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to
904  * vnode_pager_generic_putpages() to implement the previous behaviour.
905  *
906  * All other FS's should use the bypass to get to the local media
907  * backing vp's VOP_PUTPAGES.
908  */
909 static void
910 vnode_pager_putpages(object, m, count, sync, rtvals)
911 	vm_object_t object;
912 	vm_page_t *m;
913 	int count;
914 	boolean_t sync;
915 	int *rtvals;
916 {
917 	int rtval;
918 	struct vnode *vp;
919 	struct mount *mp;
920 	int bytes = count * PAGE_SIZE;
921 
922 	GIANT_REQUIRED;
923 	/*
924 	 * Force synchronous operation if we are extremely low on memory
925 	 * to prevent a low-memory deadlock.  VOP operations often need to
926 	 * allocate more memory to initiate the I/O ( i.e. do a BMAP
927 	 * operation ).  The swapper handles the case by limiting the amount
928 	 * of asynchronous I/O, but that sort of solution doesn't scale well
929 	 * for the vnode pager without a lot of work.
930 	 *
931 	 * Also, the backing vnode's iodone routine may not wake the pageout
932 	 * daemon up.  This should be probably be addressed XXX.
933 	 */
934 
935 	if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min)
936 		sync |= OBJPC_SYNC;
937 
938 	/*
939 	 * Call device-specific putpages function
940 	 */
941 	vp = object->handle;
942 	if (vp->v_type != VREG)
943 		mp = NULL;
944 	(void)vn_start_write(vp, &mp, V_WAIT);
945 	rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0);
946 	KASSERT(rtval != EOPNOTSUPP,
947 	    ("vnode_pager: stale FS putpages\n"));
948 	vn_finished_write(mp);
949 }
950 
951 
952 /*
953  * This is now called from local media FS's to operate against their
954  * own vnodes if they fail to implement VOP_PUTPAGES.
955  *
956  * This is typically called indirectly via the pageout daemon and
957  * clustering has already typically occured, so in general we ask the
958  * underlying filesystem to write the data out asynchronously rather
959  * then delayed.
960  */
961 int
962 vnode_pager_generic_putpages(vp, m, bytecount, flags, rtvals)
963 	struct vnode *vp;
964 	vm_page_t *m;
965 	int bytecount;
966 	int flags;
967 	int *rtvals;
968 {
969 	int i;
970 	vm_object_t object;
971 	int count;
972 
973 	int maxsize, ncount;
974 	vm_ooffset_t poffset;
975 	struct uio auio;
976 	struct iovec aiov;
977 	int error;
978 	int ioflags;
979 
980 	GIANT_REQUIRED;
981 	object = vp->v_object;
982 	count = bytecount / PAGE_SIZE;
983 
984 	for (i = 0; i < count; i++)
985 		rtvals[i] = VM_PAGER_AGAIN;
986 
987 	if ((int) m[0]->pindex < 0) {
988 		printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%x)\n",
989 			(long)m[0]->pindex, m[0]->dirty);
990 		rtvals[0] = VM_PAGER_BAD;
991 		return VM_PAGER_BAD;
992 	}
993 
994 	maxsize = count * PAGE_SIZE;
995 	ncount = count;
996 
997 	poffset = IDX_TO_OFF(m[0]->pindex);
998 
999 	/*
1000 	 * If the page-aligned write is larger then the actual file we
1001 	 * have to invalidate pages occuring beyond the file EOF.  However,
1002 	 * there is an edge case where a file may not be page-aligned where
1003 	 * the last page is partially invalid.  In this case the filesystem
1004 	 * may not properly clear the dirty bits for the entire page (which
1005 	 * could be VM_PAGE_BITS_ALL due to the page having been mmap()d).
1006 	 * With the page locked we are free to fix-up the dirty bits here.
1007 	 *
1008 	 * We do not under any circumstances truncate the valid bits, as
1009 	 * this will screw up bogus page replacement.
1010 	 */
1011 	if (maxsize + poffset > object->un_pager.vnp.vnp_size) {
1012 		if (object->un_pager.vnp.vnp_size > poffset) {
1013 			int pgoff;
1014 
1015 			maxsize = object->un_pager.vnp.vnp_size - poffset;
1016 			ncount = btoc(maxsize);
1017 			if ((pgoff = (int)maxsize & PAGE_MASK) != 0) {
1018 				vm_page_clear_dirty(m[ncount - 1], pgoff,
1019 					PAGE_SIZE - pgoff);
1020 			}
1021 		} else {
1022 			maxsize = 0;
1023 			ncount = 0;
1024 		}
1025 		if (ncount < count) {
1026 			for (i = ncount; i < count; i++) {
1027 				rtvals[i] = VM_PAGER_BAD;
1028 			}
1029 		}
1030 	}
1031 
1032 	/*
1033 	 * pageouts are already clustered, use IO_ASYNC t o force a bawrite()
1034 	 * rather then a bdwrite() to prevent paging I/O from saturating
1035 	 * the buffer cache.
1036 	 */
1037 	ioflags = IO_VMIO;
1038 	ioflags |= (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) ? IO_SYNC: IO_ASYNC;
1039 	ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0;
1040 
1041 	aiov.iov_base = (caddr_t) 0;
1042 	aiov.iov_len = maxsize;
1043 	auio.uio_iov = &aiov;
1044 	auio.uio_iovcnt = 1;
1045 	auio.uio_offset = poffset;
1046 	auio.uio_segflg = UIO_NOCOPY;
1047 	auio.uio_rw = UIO_WRITE;
1048 	auio.uio_resid = maxsize;
1049 	auio.uio_td = (struct thread *) 0;
1050 	error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred);
1051 	cnt.v_vnodeout++;
1052 	cnt.v_vnodepgsout += ncount;
1053 
1054 	if (error) {
1055 		printf("vnode_pager_putpages: I/O error %d\n", error);
1056 	}
1057 	if (auio.uio_resid) {
1058 		printf("vnode_pager_putpages: residual I/O %d at %lu\n",
1059 		    auio.uio_resid, (u_long)m[0]->pindex);
1060 	}
1061 	for (i = 0; i < ncount; i++) {
1062 		rtvals[i] = VM_PAGER_OK;
1063 	}
1064 	return rtvals[0];
1065 }
1066 
1067 struct vnode *
1068 vnode_pager_lock(object)
1069 	vm_object_t object;
1070 {
1071 	struct thread *td = curthread;	/* XXX */
1072 
1073 	GIANT_REQUIRED;
1074 
1075 	for (; object != NULL; object = object->backing_object) {
1076 		if (object->type != OBJT_VNODE)
1077 			continue;
1078 		if (object->flags & OBJ_DEAD) {
1079 			return NULL;
1080 		}
1081 
1082 		/* XXX; If object->handle can change, we need to cache it. */
1083 		while (vget(object->handle,
1084 			LK_NOPAUSE | LK_SHARED | LK_RETRY | LK_CANRECURSE, td)){
1085 			if ((object->flags & OBJ_DEAD) || (object->type != OBJT_VNODE))
1086 				return NULL;
1087 			printf("vnode_pager_lock: retrying\n");
1088 		}
1089 		return object->handle;
1090 	}
1091 	return NULL;
1092 }
1093